Journal of Software - Academy Publisher

Journal of Software 

ISSN 1796-217X 

Volume 6, Number 5, May 2011 

Contents 

REGULAR PAPERS 

User Requirements Notation: The First Ten Years, The Next Ten Years (Invited Paper) 

Daniel Amyot and Gunter Mussbacher 

Stochastic Process Algebra with Value-Passing and Weak Time Restrictions 

Guang Zheng, Jinzhao Wu, and Aiping Lu 

Study on Visual Knowledge Structure Reasoning 

Huimin Lu, Liang Hu, and Gang Liu 

An Automated X-corner Detection Algorithm (AXDA) 

Fuqing Zhao, Chunmiao Wei, Jizhe Wang, and Jianxin Tang 

Research on Dynamic Rescheduling Program Base On Improved Contract Net Protocol 

Fuqing Zhao, Jizhe Wang, and Jianxin Tang 

Multilevel Network Security Monitoring and Evaluation Model 

Jin Yang, Tang Liu, Lingxi Peng, XueJun Li, and Gang Luo 

Research on Family and Shops Real-time Status of 3G Wireless Remote Monitoring System 

Qian Zhao 

Sliding Mode Control of Surface-Mount Permanent-Magnet Synchronous Motor Based on Error 

Model with Unknown Load 

Bao-jun Wang and Jia-jun Wang 

Information Fusion Based Fault Location Technology for Distribution Network 

Qingle Pang 

Study on Remote Aided Diagnosis System of Mental Health Base on Export Knowledge Base 

Xiaoyong Wang and Yuefeng Fang 

Application of Fault Phenomenon Vector Distance Discriminance in Woodworking Machinery 

System Fault Diagnosis 

Yun-Jie Xu, Shu-Dong Xiu, Quan-Sheng Men, and Liang Fang 

A Novel Gray Image Watermarking Scheme 

Yongqiang Chen, Yanqing Zhang, Hanping Hu, and Hefei Ling 

An Efficient Method for Improving Query Efficiency in Data Warehouse 

Zhiwei Ni, Junfeng Guo, Li Wang, and Yazhuo Gao 

Co-simulation Study of Vehicle ESP System Based on ADAMS and MATLAB 

Shengqin Li and Le He 

747 

769 

783 

791 

798 

806 

814 

819 

826 

834 

842 

849 

857 

866

An Improved Fuzzy C-means Clustering Algorithm based on PSO 

Qiang Niu and Xinjian Huang 

Classification of Bio-potential Surface Electrode based on FKCM and SVM 

Hao Liu, Xiaoming Tao, Pengjun Xu, and Guanxiong Qiu 

Consonant Recognition of Dysarthria Based on Wavelet Transform and Fuzzy Support Vector 

Machines 

Zhuo-ming Chen, Wei-xin Ling, Jian-hui Zhao, and Tao-tao Yao 

ELECTRE I Decision Model of Reliability Design Scheme for Computer Numerical Control Machine 

Jihong Pang, Genbao Zhang, and Guohua Chen 

Fractional Modeling Method Research on Education Evaluation 

Chunna Zhao, Yu Zhao, Liming Luo, and Yingshun Li 

Immune Genetic Evolutionary Algorithm of Wavelet Neural Network to Predict the Performance in 

the Centrifugal Compressor and Research 

Shengzhong Huang 

Development of Optimization Design Software for Bevel Gear Based on Integer Serial Number 

Encoding Genetic Algorithm 

Xiaoqin Zhang, Yu Rong, Jingjing Yu, Liling Zhang, and Lina Cui 

Study on Operating Mechanisms and Dynamics Behavior of Agile Supply Chain 

Guohua Chen, Genbao Zhang, and Jihong Pang 

Unified Service Platform for Accessing Grid Resources 

Shaochong Feng, Yuanchang Zhu, and Yanqiang Di 

Research on an Improved Terrain Aided Positioning Model 

Shidan Li, Liguo Sun, Xin Li, and Desheng Wang 

Research on Integrated Information Platform of Agricultural Supply Chain Management Based on 

Internet of Things 

Yan-e Duan 

873 

880 

887 

894 

901 

908 

915 

923 

930 

937 

944

JOURNAL OF SOFTWARE, VOL. 6, NO. 5, MAY 2011 747 

User Requirements Notation: 

The First Ten Years, The Next Ten Years 

Abstract—The User Requirements Notation (URN), 

standardized by the International Telecommunication 

Union in 2008, is used to model and analyze requirements 

with goals and scenarios. This paper describes the first ten 

years of development of URN, and discusses ongoing efforts 

targeting the next ten years. We did a study inspired by the 

systematic literature review approach, querying five major 

search engines and using the existing URN Virtual Library. 

Based on the 281 scientific publications related to URN we 

collected and analyzed, we observe a shift from a more 

conventional use of URN for telecommunications and 

reactive systems to business process management and 

aspect-oriented modeling, with relevant extensions to the 

language being proposed. URN also benefits from a global 

and active research community, although industrial 

contributions are still sparse. URN is now a leading 

language for goal-driven and scenario-oriented modeling 

with a promising future for many application domains. 

Index Terms—Goals, Goal-oriented Requirement Language 

(GRL), modeling, review, scenarios, tools, Use Case Maps 

(UCM), User Requirements Notation (URN) 

I. INTRODUCTION 

The User Requirements Notation (URN) is a modeling 

language that aims to support the elicitation, analysis, 

specification, and validation of requirements. URN is the 

first international standard to address explicitly, in a 

graphical way and in one unified language, goals and 

scenarios, and the links between them [106]. URN 

models can be used to specify and analyze various types 

of reactive systems as well as telecommunications 

standards and business processes. URN allows software 

and requirements engineers as well as business analysts 

to discover and specify requirements for a proposed 

system or process (or evolving ones), and analyze such 

requirements for correctness and completeness. 

The kind of modeling supported by URN is different 

from the detailed specification of “how” functionalities 

are to be supported, as described with languages such as 

UML [146]. Here the modeler is primarily concerned 

with exposing “why” certain choices for behavior and/or 

structure were introduced, combined with an abstract 

view of “what” capabilities and architecture are required. 

The modeler is not yet interested in the operational details 

of internal component behavior or component 

interactions. Omitting these kinds of details during early 

development allows working at a higher level of 

abstraction when modeling a current or future software 

© 2011 ACADEMY PUBLISHER 

doi:10.4304/jsw.6.5.747-768 

(Invited Paper) 

Daniel Amyot and Gunter Mussbacher 

University of Ottawa, Canada 

Email: {damyot, gunterm}@site.uottawa.ca 

system, business process, or standard, and its embedding 

environment. Modeling and answering “why” questions 

leads us to consider the opportunities stakeholders seek 

out and vulnerabilities they try to avoid within their 

environment, whereas modeling and answering “what” 

questions helps identify capabilities, services, and 

architectures required to satisfy stakeholder goals. 

Based on a systematic literature review, this paper 

provides a historical perspective on the development of 

URN together with trends related to future constructs and 

application domains for this notation. Such study is 

important at this point not only to appreciate the richness 

of URN and the substantial body of work that already 

exists, but also to step back, understand current trends, 

and anticipate future needs for evolving the notation in 

the right direction. 

Section II introduces URN’s basic concepts and 

notational elements, together with its standard analysis 

techniques. As it is important to understand why URN 

was created, a historical description of the origins of the 

notation is presented in Section III. Then, Section IV 

summarizes the main results of our literature survey, 

especially with regards to the sources of contributions to 

URN. In Section V, some of the main research 

contributions that have shaped URN in the past decade 

are categorized and reviewed, whereas section VI 

identifies current and future development activities and 

research areas related to URN for the next decade. 

Finally, section VII provides our conclusions. 

II. OVERVIEW OF URN 

The User Requirements Notation standard combines 

two sub-languages [106]: the Goal-oriented Requirement 

Language for modeling actors and their intentions, and 

the Use Case Maps notation for describing scenarios and 

architectures. In this section, we give a brief overview of 

each of these sub-languages, supported by a simple URN 

model example that targets the evaluation of an 

architectural decision about where to put the data and the 

logic of the authorization service of a wireless system. 

A. Goal-oriented Requirement Language (GRL) 

GRL is a visual modeling notation for intentions, 

business goals, and non-functional requirements (NFR) of 

many stakeholders, for alternatives that have to be 

considered, for decisions that were made, and for 

rationales that helped make these decisions.

748 JOURNAL OF SOFTWARE, VOL. 6, NO. 5, MAY 2011 

A GRL goal graph is a connected graph of intentional 

elements that optionally reside within an actor. An actor 

( , e.g., Service Provider, Figure 1.a) represents a 

stakeholder of a system, or the system itself. A goal graph 

shows the non-functional requirements and business 

goals of interest to the system and its stakeholders, as 

well as the alternatives for achieving these high-level 

elements. Actors are holders of intentions; they are the 

active entities in the system or its environment who want 

goals to be achieved, tasks to be performed, resources to 

be available, and softgoals to be satisfied. Softgoals ( , 

e.g., Low Cost) differentiate themselves from goals ( , 

e.g., Determine Data Location) in that there is no clear, 

objective measure of satisfaction for a softgoal whereas a 

goal is quantifiable, often in a binary way. Softgoals are 

often more related to NFR, whereas goals are more 

related to functional requirements. Tasks ( , e.g., Install 

Service Node) represent solutions to (or 

operationalizations of) goals or softgoals. In order to be 

achieved or completed, softgoals, goals, and tasks may 

require resources (�, e.g., Service Node) to be available. 

Various kinds of links connect the elements in a goal 

graph. Decomposition links allow an element to be 

decomposed into sub-elements ( , e.g., High 

Performance is decomposed into Maximum Hardware 

Utilisation and High Throughput). AND, IOR, as well as 

XOR decompositions are supported. Contribution links 

indicate desired impacts of one element on another 

element (→, e.g., Minimum Changes to Infrastructure 

contributes to Low Cost). A contribution link has a 

qualitative contribution type (Figure 1.b) or a quantitative 

contribution (an integer value between -100 and 100). 

Correlation links ( ) are similar in nature, but describe 

side effects rather than desired impacts. Dependency 

links model relationships between actors ( , e.g., 

System depends on Vendor for Service Node). 

GRL supports reasoning about goals and requirements, 

especially NFR and quality attributes, as it shows the 

impact of often conflicting goals and various global 

alternative solutions proposed to achieve the goals. A 

GRL strategy describes a particular configuration of 

alternatives in the GRL model by assigning an initial 

qualitative satisfaction level (Figure 1.c) or a quantitative 

one (an integer value between -100 and 100) to some of 

the intentional elements in the model (indicated by a star 

(*) and a dashed outline), often leaves in the GRL graph. 

An evaluation mechanism propagates these low-level 

decisions regarding alternatives to satisfaction ratings of 

high-level stakeholder goals and NFR. Strategies can 

therefore be compared with each other to help reach the 

most appropriate trade-offs among often conflicting goals 

of stakeholders. A good strategy offers rationale and 

documentation for decisions leading to requirements, thus 

providing better context for systems and software 

engineers while avoiding unnecessary re-evaluations of 

worse alternative strategies. Color coding of the 

intentional elements also reflect their satisfaction level 

(the greener, the more satisfied). 

GRL takes into account that not all high-level goals 

and NFR are equally important to a stakeholder. 


Therefore, GRL supports the definition of an importance 

attribute for intentional elements inside actors (again 

quantitative or qualitative, and shown between 

parentheses, e.g., 50 for Low Cost). This attribute is also 

taken into account when evaluating strategies for the goal 

model, resulting in satisfaction levels measured at the 

actor level (e.g., 32 for the Service Provider). 

The current URN standard does not enforce a specific 

evaluation mechanism as GRL can be used in different 

ways by different modelers, e.g., for qualitative 

evaluations or quantitative ones, but provides three nonnormative 

examples of evaluation algorithms. A hybrid 

algorithm combining qualitative contributions and 

quantitative satisfaction levels is used for one strategy in 

Figure 1.a. A different strategy would lead to different 

results, enabling comparisons and documenting decisions. 

B. Use Case Maps (UCM) 

The UCM visual scenario notation focuses on the 

causal flow of behavior optionally superimposed on a 

structure of components. UCM depict the causal 

interaction of architectural entities while abstracting from 

message and data details. 

The basic elements of the UCM notation are shown 

in Figure 2. A map contains any number of paths and 

components. Paths express causal sequences and may 

contain several types of path nodes. Paths start at start 

points (�, e.g., StartConnection) and end at end points (▌, 

e.g., Done), which capture triggering and resulting 

conditions respectively. Responsibilities (�, e.g., 

LogReject) describe required actions or steps to fulfill a 

scenario. OR-forks ( ), possibly including guarding 

conditions such as [NotOk], and OR-joins ( ) are used 

to show alternatives, while AND-forks ( ) and ANDjoins 

( ) depict concurrency. Loops can be modeled 

implicitly with OR-joins and OR-forks. As the UCM 

notation does not impose any nesting constraints, joins 

and forks may be freely combined and a fork does not 

need to be followed by a join. Waiting places (�) and 

timers ( ) denote locations on the path where the 

scenario stops until a condition is satisfied. 

UCM models can be decomposed using stubs that 

contain sub-maps called plug-in maps (see Figure 2.b and 

c). Plug-in maps are reusable units of behavior and 

structure. Plug-in bindings define the continuation of a 

path on a plug-in map by connecting in-paths and outpaths 

of a stub (IN1 and OUT1 in Figure 2) with start and 

end points of its plug-in maps, respectively. Plug-in 

bindings also describe the relationship of components on 

the parent map with the ones on the plug-in map (e.g., the 

parent component of the plug-in map in Figure 2.c refers 

to a component in the parent map, ControlFunction in this 

example). A stub may be static (�), which means that it 

can have at most one plug-in map, whereas a dynamic 

stub ( , e.g., Authorization) may have many plug-in maps 

that can be selected at runtime according to a selection 

policy. In Figure 2, the two plug-in maps represent 

alternative ways of supporting authorization, with 

different locations for the data and the logic of the service 

(i.e., different allocations of responsibilities to 

components).


Make Help Some Positive Unknown Some Negative Break Hurt 

(b) GRL Contributions Types 

(a) GRL graph for a system with two stakeholders 

Denied 

Weakly 

Denied 

Figure 1 GRL example: Where should the data and the service be located in the system? 

a) Top-level map: Connection request to a mobile switch 

b) Plugin 1: Service in mobile switch, 

data in external service node 

Weakly 

Satisfied 

Satisfied Conflict Unknown None 

(c) GRL Satisfaction Levels 

c) Plugin 2: Service and data in 

mobile switch 

Figure 2 UCM example: Connection scenario (a), with two potential architectural solutions (b and c) for the authorization service. 

© 2011 ACADEMY PUBLISHER


Components (�, e.g., MobileSwitch) are used to 

specify the structural aspects of a system. Map elements 

which reside inside a component are said to be bound to 

it. Components may contain sub-components and have 

various types and characteristics. For example, a 

component of kind object ( , e.g., LocationDB) does not 

have its own thread of control whereas a component of 

kind process (�, e.g., ControlFunction) does. A 

component of kind actor ( , e.g., MobileStation) 

represents someone or something interacting with the 

system under design. 

UCM support the definition of scenarios including preand 

postconditions. A scenario describes a specific path 

through the UCM model where only one alternative at 

any choice point is taken. The UCM notation supports a 

simple but formal data model that can be used to 

formalize the conditions at selection points (e.g., dynamic 

stubs and OR-forks). Responsibilities can also include 

code that modifies the values of the variables used in this 

data model. A scenario definition can hence be expressed 

with initial values for these variables, combined with a 

sequence of start points being triggered. 

Given the definition of a scenario or combination of 

scenarios, a path traversal mechanism can highlight the 

scenario path being simulated. Figure 2 shows in red the 

paths traversed for the scenario where the service logic 

remains in the mobile switch but the service data is 

located in a new external service node (which 

corresponds to the strategy being evaluated for the GRL 

model in Figure 1.a), and where the authorization is OK. 

The traversal mechanism essentially provides the 

operational semantics of the UCM language. It also turns 

the scenario definitions into a test suite for the UCM 

model, which is especially useful for regression testing as 

the model evolves. 

Different elements in a UCM model can also be 

annotated with specific performance information, 

enabling early performance analysis at the requirements 

level. For example, resources can be defined and 

components assigned to them, selection points can 

include probabilities, responsibilities can specify 

demands on resources, and start points can include 

workload definitions. 

The UCM notation enables a seamless transition from 

the informal to the formal by bridging the modeling gap 

between goal models and natural language requirements 

(e.g., use cases) and design artefacts, in an explicit and 

visual way. 

C. Integration of Goals and Scenarios in URN 

Modeling goals and scenarios is complementary and 

may aid in identifying additional or spurious goals and 

scenarios, thus contributing to the completeness and 

accuracy of requirements. In the language, URN links 

( ) can connect any two URN model elements, 

establishing traceability links that further tighten the 

relationship between GRL and UCM models while 

enabling completeness and consistency analysis. 

The URN language also supports the concept of 

metadata in the form of name/value pairs that can be 

associated with any URN model element. This allows for 


domain-specific extensions to be added to URN and 

exploited by specialized tool support. 

III. PRE-URN HISTORY (1990-1999) 

The roots of URN go back to the early 90’s. Use Case 

Maps originate from Carleton University, where Buhr 

used them as a high-level notation in their project Design 

of Object-Oriented Real-time Systems (DOORS ― 

http://www.sce.carleton.ca/rads/doors.html). Vigder’s 

early work on design slices [186] used a scenario-like 

notation with a connection to the LOTOS formal 

specification language [99]. Buhr then coined the term 

timethread as a name for this graphical notation, which 

was used in a few papers and theses until the release, in 

1995, of a seminal book co-authored with Casselman 

where the term Use Case Maps emerged [52]. This book 

focused on the application of UCM to object-oriented 

systems, with an emphasis on role modeling concepts 

developed in Casselman’s thesis [56]. Another important 

milestone for the UCM notation was the publication of a 

revised and more powerful version of the language in a 

major journal [49]. In those years, typical applications 

that were explored with this notation included design 

activities [13][40] (including architecture [46][49] and 

patterns [47][50]), performance analysis [170], and the 

modeling of telecommunication [8][10], agentoriented 

[51][68], and e-commerce [79] systems. Miga 

provided tool support for the creation and analysis of 

UCM models [129], based on an earlier prototype from 

Carrière (UCMEdit, discussed in [113]). This multiplatform 

UCM Navigator tool (UCMNAV) was used in 

academia and industry mainly between 1998 and 

2005 [181]. 

Work on goal modeling for requirements, agents, and 

organizations that was being done at that time at the 

University of Toronto guided the development of the 

GRL language. The syntax of GRL is in fact based on the 

i* framework described in Yu’s thesis [197], which was 

developed for describing strategic relationships in 

organization models. The reasoning mechanisms behind 

another goal-oriented notation, namely the Non- 

Functional Requirements (NFR) Framework (best 

described in the seminal book of Chung, Mylopoulos, 

Nixon, and Yu [62]), also inspired the evaluation and 

propagation mechanisms now found in GRL. Tool 

support for modeling and analyzing goal models (in i*, 

the NFR Framework, and GRL) was then provided by the 

Java-based OME 3, Yu’s Organization Modelling 

Environment [198]. 

The idea of creating a new standard notation was first 

proposed in 1999 by Visser and Hodges from Nortel, as 

they were deeply involved in standardization activities 

with the International Telecommunication Union (ITU-T) 

and with the Wireless Intelligent Network initiative [96]. 

Through collaborative research projects with Logrippo 

(University of Ottawa) and his team [10][21], it was 

observed that UCMs would likely be more appropriate 

than natural language and Message Sequence Charts 

(MSC) [104] for early descriptions of wireless 

telecommunication features. Monkewich (also from


Nortel) brought the idea of creating a Use Case Maps 

standard to the language experts at ITU-T, who then 

suggested renaming it to “User Requirements Notation”. 

This potential standard captured the attention of another 

Ottawa-based company, Mitel, where Pinard, Weiss, 

Gray, and Mankovski had also used UCM for modeling 

telecommunications features. However, they were also 

interested in i* and the NFR Framework for goal-oriented 

and agent-based modeling. They were collaborating with 

the University of Toronto on projects that led to the 

creation of a new goal modeling language by Yu and Liu, 

which became the first version of GRL [126]. 

Gray and the Mitel experts expected great benefits in 

combining goals with scenarios and regarded this 

combination essential for the understanding of highly 

dynamic and reflective systems, and for feature 

personalization. This potential integration led to the 

introduction of dynamic stubs in the UCM notation in the 

mid-90’s. Gray convinced the Nortel experts and other 

stakeholders to revise the URN proposal as a Canadian 

contribution to ITU-T that would include both GRL and 

UCM. This was then accepted as a new work item at 

ITU-T in 2000, and Hodges became the first Rapporteur 

for the URN question. 

IV. SYSTEMATIC LITERATURE REVIEW 

A. Methodology 

Inspired by the work of Kitchenham et al. [43][118], 

we did a systematic literature review targeting the 

following three questions: 

• Who contributed to the development of URN? 

• What research contributed to the development of 

URN? 

• What are the current and future development 

activities and research areas related to URN? 

In July 2010, we used five major search engines for 

publications in computer science and engineering (IEEE 

Xplore, ACM Digital Library, Google Scholar, 

SpringerLink, and Scopus). Our query was simply "User 

Requirements Notation" OR "Use Case Map" OR 

"Goal-oriented Requirement Language", which 

covered the essential keywords. The URN, GRL, and 

UCM acronyms were not included because an early 

assessment led us to believe they were polluting the 

results without really identifying more valid citations. 

Over 700 references were collected in the end, mostly 

coming from Google Scholar. These were combined with 

the references already present in the URN Virtual 

Library [182]. 

We restricted the results to scientific publications 

appearing in journals, conferences, workshops, books, 

and theses. Furthermore, we excluded papers that: 

• Only cited URN (or GRL/UCM) to acknowledge 

its existence or to discuss it in a comparison. 

• Simply used URN (or GRL/UCM) to illustrate 

some requirements or design (e.g., with a few 

diagrams), without discussing the usage of the 

language itself. 


• Focused on the “other” Use Case Map concept 

developed by Constantine and Lockwood [63], 

which is a variant of UML use case diagrams 

used to model the interrelationships among use 

cases (different from URN’s Use Case Maps). 

We finally included seminal work produced prior to 

the use of the terms UCM [56][186] and GRL [62][197]. 

B. Contributors and Contributions 

Our selection led to a total of 281 scientific 

publications related to research on and with URN. More 

specifically, we have found 38 journals papers, 183 

conference and workshop papers, 15 books and book 

chapters, as well as 45 theses (13 Ph.D., 31 Master’s, and 

1 B.Sc.) The URN Virtual Library was updated to include 

the 31 publications that were missing prior to this 

literature review. Figure 3 shows the distribution of our 

four types of publications over the years. 

To answer our first question, this data shows that there 

were 263 different authors involved (with an average of 

2.7 authors per paper, often from different locations). 

Given the origins of URN, it is not surprising to see that 

the majority of the papers (66%) and theses (80%) 

published since 1992 include co-authors from Canada, 

especially from the University of Ottawa and from 

Carleton University (see Table I). Actually, all papers and 

theses prior to 1999 came from Canada. 

TABLE I. 

NUMBER OF CO-AUTHORS PER COUNTRY 

Country Papers Theses Total 

Canada 162 36 198 

U. of Ottawa 103 21 124 

Carleton U. 64 10 74 

Concordia U. 15 4 19 

U. of Toronto 7 1 8 

Other places 19 19 

The Netherlands 10 4 14 

UK 11 1 12 

USA 9 9 

Japan 8 8 

Hungary 6 6 

Norway 4 1 5 

Italy 4 4 

Australia 4 4 

Brazil 4 4 

Germany 3 1 4 

Argentina 3 1 4 

Spain 3 3 

China 3 3 

Belgium 3 3 

Portugal 3 3 

Korea 2 2 

Viet-Nam 2 2 

South Africa 2 2 

United Arab Emirates 2 2 

Switzerland 1 1 2 

Serbia 1 1 

Latvia 1 1 

Poland 1 1 

Venezuela 1 1 

Libya 1 1 

Thailand 1 1 

Sweden 1 1 

Chile 1 1


Number of publications 

40 

35 

30 

25 

20 

15 

10 

5 

0 

Year 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 

Journals Conf. & Workshops Books & Chapters Theses 

Figure 3 Number of URN-related scientific publications per year. 

However, work on URN then started to draw 

international involvement. Between 1999 and July 2010, 

we observed that 43% of the papers had at least one coauthor 

from outside Canada, 34% of the papers had all 

their co-authors from outside Canada, and 23% of the 

theses were from outside Canada. Collaboration between 

Canadian and non-Canadian authors has also increased 

substantially over the past four years. As shown in 

Table I, researchers and industrial participants from over 

29 different countries on all continents have contributed 

publications on URN. Incidently, in our data set, we also 

detected papers written in seven languages other than 

English: Chinese (4), French (3), Spanish (3), Serbian (2), 

Japanese (2), Korean (1), and German (1). 

Although our results support that scientific 

contributions related to URN are numerous and 

international, industrial contributions are still sparse: only 

22 papers (9% of our data set) involved co-authors with 

industrial affiliations, mainly from the telecom industry 

(Nortel, Mitel, Cisco, and others). This may partially be 

explained by the fact that we excluded many papers 

where URN was simply used. 

C. Bias 

The content of the URN Virtual Library and our 

selection of papers are somewhat biased towards UCM 

because the work related on UCM is all included in 

URN’s, whereas prior and subsequent work on the NFR 

Framework and especially on i*, on which GRL is 

initially based, is not included. i* is different from GRL 

and has a community of its own (e.g., four i* workshops 

and 55 research teams are listed so far on the i* 

Wiki [97]). This is also reflected in the numbers of 

references we collected (e.g., from Google Scholar, we 

found 442 for UCM, 294 for URN, and 186 for GRL). 

Consequently, many of the topics discussed in the next 

two sections focus on UCM and fewer will address GRL 

exclusively. 

Note that we have done the collection and filtering of 

the papers ourselves. However, to mitigate internal bias, 

an exhaustive search using Google Scholar and other 

engines (see previous section) was performed. Given that 

we have been involved with URN since its inception, that 

we co-authored about a third of the scientific papers on 


URN, that we co-edited the standard and that we are 

responsible for its evolution at ITU-T, we believe we are 

uniquely positioned to perform a rigorous assessment of 

the past and future research on URN. 

V. THE FIRST TEN YEARS (2000-2009) 

After ITU-T’s approval of a new question on a User 

Requirements Notation, the URN standardization took 

another eight years. Cameron (Nortel) took over from 

Hodges as Rapporteur in 2001, followed by Amyot 

(University of Ottawa) in 2002. The first standard (Z.150, 

in 2003) described the goals and requirements for the 

URN language [105]. In 2008, the definition of the URN 

language itself (including a URN metamodel based on the 

new meta-metamodel of ITU-T Z.111 [103]) finally 

became available [106]. 

This section summarizes some of the most important 

research contributions that have led to this standard, 

together with emerging application domains. 

A. Specification and Validation of Protocols and Services 

One of the main drivers behind the creation of URN 

was to enable standards bodies such as ITU-T to specify 

and perform early validation of new telecommunication 

protocols and services. UCM models, in particular, 

provide a view that abstracts from messages and 

potentially from component architectures, which 

simplifies the description of services. This view also 

helps multiple stakeholders such as vendors and carriers 

(with conflicting agendas and investments in different 

legacy networks) reach consensus on the essence of these 

services. 

There were indeed several services specified with 

UCM and proposed for standardization, in the early 

2000s, to ITU-T, the Telecommunications Industry 

Association (TIA), and the Internet Engineering Task 

Force (IETF). In the literature, we notice the International 

Mobile Telecommunications-2000 (IMT-2000) [27] and 

other mobile wireless protocols [25][26], the Open 

Shortest Path First (OSPF) routing protocol [132], Mobile 

IPv6 [187], Call Name Presentation (CNAP) [196], and 

GPRS mobile group call [20], among others. All of these


specifications helped further shape the notation as well as 

stylistic guidelines for its application to telecom services. 

More recently, URN was revisited in a service 

engineering context. Amyot et al. [12] described a URNbased 

approach for specifying Next Generation Network 

(NGN) services, where service models (combining GRL, 

UCM, and UML views) provide information and 

mechanisms that help dynamic composition and 

adaptation at runtime. In his thesis, Castejón focuses 

more on the concept of collaboration for compositional 

service descriptions, with both UML and UCM [58]. 

These two approaches to service engineering helped 

explore and understand the synergy between URN and 

UML collaborations. 

B. Multi-Agent Systems 

In the past decade, there was quite a bit of attention 

devoted to the use of URN in the domain of agent 

systems, beyond what was already done in the 90’s by 

Buhr and others as part of their High-Level Design and 

Prototyping of Agent Systems project [51][68]. 

Bush et al. [53] introduced their Styx agent 

methodology, where UCM are used to capture high-level 

system processes. A similar use of UCM is discussed in 

the approaches proposed by Araya and Antillanca [28] 

and by Abdelaziz et al. [2], the latter with an interesting 

application to a medical diagnosis system. In his 

thesis [1], Abdelaziz further enhanced his Multi-Agent 

System Development approach by integrating UML use 

case, activity, and sequence diagrams (and to some extent 

GRL dependency models) with the UCM view. 

Lavendelis and Grunspenkis proposed the MASITS tool, 

similar in intent to Buhr’s [51], for capturing several 

views of agent systems, including goal and use case 

views à la URN. Saleh and Al-Zarouni [165] described 

the use of GRL for capturing non-functional requirements 

for mobile agent systems. Amyot et al. [14] specified and 

analyzed, with UCM, an agent-based telecommunication 

system being built by an industrial partner. 

While Billard captured a collection of eight agent 

interaction patterns with UCM, with an analysis of their 

performance [34], Weiss exploited the NFR Framework 

(at the basis of GRL) to describe and exploit the 

relationships between patterns used in the design of agent 

systems [188]. 

C. Web Applications and Web Services 

In the mid-2000s, URN was also used in the context of 

Web applications. Yu and Liu, following seminal work 

where they first proposed an iterative methodology that 

combines GRL and UCM and demonstrate the 

complementary nature of these two views [124], 

successfully specified a Web-based training application 

with URN [125]. Around the same period, Kaewkasi and 

Rivepiboon also proposed a methodology for Web 

application modeling, but this time based on a 

combination of UCM and UML [111]. Around 2005, 

Weiss started to describe patterns (partly with UCM) for 

Web applications [189], while also exploring with others 

the UCM-driven testing of Web applications [22]. 


Web services also captured the attention of URN 

contributors. Weiss and Esfandiari provided preliminary 

results on the analysis of personalized Web services 

where service goals are specified with GRL and service 

functionalities with UCM [191]. van der Raadt explored 

Web services from a business perspective with the 

Business-oriented Approach Supporting web Service Idea 

Exploration (BASSIE) methodology. BASSIE combines 

three types of models: a) i* (instead of GRL) for 

describing strategic goals and the impact of service 

realization alternatives, Gordijn’s e 3 value framework [77] 

(originating from UCM) for evaluating alternatives based 

on their profitability, and UCM for describing other 

details of the services. 

These two approaches to the development of Web 

applications and services helped clarify the relationships 

that exist between GRL and UCM views, with an impact 

on the inclusion of the concept of URN link in the 

standard. 

D. Formalization 

The URN standard describes the URN abstract and 

concrete syntaxes formally, together with wellformedness 

constraints. However, the semantics is 

currently described more informally using traversal 

requirements for UCM (with which many algorithms 

could comply) and with propagation requirements for 

GRL (with, again, many potential evaluation algorithms). 

Hence, these textual requirements do not fully alleviate 

the risk that different tools could implement different 

semantics while still satisfying the standard’s 

requirements. 

It was judged premature to agree on a unique 

semantics to the notation in the standard, although many 

had already been proposed, especially for UCM. One of 

the first attempts was done by Amyot and Logrippo [8], 

and was based on the LOTOS process algebra [99]. The 

mapping from UCM to LOTOS was further explored by 

Guan [82], who also provided a compiler for UCMNav 

models. This mapping was used in 9 theses and 13 

publications, and contributed to the understanding of 

UCM behavior. 

van der Poll et al. proposed an initial, informal 

mapping from UCM to Z in order to formally analyze 

models capturing user interface scenarios [183]. This 

work was further extended by Dongmo [67], who defined 

a framework to derive Object-Z [176] class schemas from 

UCM models. Similarly, Truong et al. [180] proposed a 

mapping from UCM to the B formal method [171] in 

order to support the verification of component behavior 

against the UCM scenario requirements. To facilitate the 

analysis of component interfaces and composition, de 

Bruin explored a mapping from UCM models enhanced 

with interface information to the object-oriented 

programming language BCOOPL [64]. 

Hassine et al. also provided a formal semantics for 

UCM, but this time based on Abstract State Machines 

(ASM, [38]), with tool support for simulation [88]. They 

further investigated the use of quantitative time 

constraints in UCM models with a UCM extension called 

Timed Use Case Maps, for which they provided


additional semantics based on more appropriate 

formalisms, namely timed automata [90], clocked 

transition systems [89], and again ASM [86]. Hassine’s 

thesis [85] is the best document where these extensions 

and semantics are used, and a recent survey provides a 

comparison with related timed scenario languages [92]. 

There is no formal semantics for GRL at this point. 

However, the initial description of GRL [126], which did 

not include a meta-model at the time, was evaluated by 

Heymans et al. [95] from an ontological perspective. 

Some of their conclusions were actually taken into 

consideration in the definition of the URN metamodel. 

Ayala et al. also provided an interesting analysis of GRL 

compared to other goal-oriented modeling languages, but 

again this is based on the original version proposed in 

2001 (not the standard definition). A comparison between 

i* and van Lamsweerde’s KAOS is offered by 

Matulevičius et al. [127] and also by [17] from an 

analysis point of view. The i* Wiki [97] is also an 

interesting source of information on formalization for 

related goal-oriented languages. For comparisons 

between UCM and other scenario languages (too 

numerous to mention them all here), the reader is referred 

to the studies of Saiedian et al. [164], Amyot and 

Eberlein [16], and Mussbacher and Amyot [136]. 

E. Transformations to Design Models 

Scenario models such as those specified with UCM 

represent a good basis for transformations to more 

detailed design representations. Such transformations 

enable the generation of design artifacts with less effort 

and, yet, higher consistency with the requirements. 

Bordeleau, Buhr, and Cameron were among the first to 

explore systematic relationships and transformation 

between UCM models and (High-level) Message 

Sequence Charts [41]. Miga et al. [130] have then 

demonstrated that lengthy scenarios resulting from UCM 

path traversals can be transformed to MSC in order to 

visualize them in a more scalable and linear form. They 

implemented this transformation in the UCMNAV 

tool [181]. The main challenges in this transformation is 

to infer or synthesize necessary messages ensuring that 

causal relationships between responsibilities in different 

components are correctly supported, and to handle the 

well-formedness rules of a linear scenario representation 

like MSC, which are stricter than the general graph 

representation of UCM. New results partly addressing 

these challenges were provided by Amyot et al. [15]. 

However, the best implementation so far is the one now 

found in the jUCMNav tool [110], as provided by Kealey 

in his thesis [114]. Along the way, Kealey redefined and 

greatly improved the power, flexibility, and robustness of 

the path traversal algorithm initially proposed by Miga, 

and this contribution had a major impact on the definition 

of the path traversal rules now found in the 

standard [114]. 

The synthesis of state machines from scenarios was a 

topic of high interest in the 2000’s [16][164]. Bordeleau, 

Corriveau, and Selic were among the first to provide 

guidelines for the transformation of UCM models to 

hierarchical state machines [42]. Sales and Probert also 


proposed transformation guidelines [165], only this time 

the target language was SDL [102]. He et al. [93] 

explored an automatic transformation from UCM to SDL 

via the intermediate generation of MSC from UCM and 

the synthesis of SDL models from these MSC (based on a 

commercial MSC-to-SDL synthesizer). Castajón also 

reported on an experiment on the synthesis of state 

machine behavior from UML collaborations whose 

dependencies are captured with UCM models [57]. 

On the goal side, we notice the combination of GRL 

and a security extension to UML (UMLsec) proposed by 

Saleh and Elshahry to model security requirements across 

goal and design views [166]. Abid et al. also proposed a 

UML profile for GRL, hence enabling the integration of a 

GRL view in UML design models [3]. 

F. Feature Interaction Analysis 

The various formalisms used to analyze URN models, 

as seen by the many transformations and formalizations 

discussed in the previous sections, are important to 

support the detection of undesirable interactions between 

features or service descriptions, a problem well known in 

telecommunications and other domains [55]. 

Amyot et al. used a mapping from UCM to LOTOS, 

combined with a testing approach, to support the rigorous 

detection of interactions between telecommunication 

features [8][14]. Due to the numerous test cases that have 

to be checked for large sets of features, the need for 

identifying situations where interaction tests are needed 

became apparent. Nakamura et al. hence proposed an 

interaction filtering approach based on the stub/plug-in 

structure of UCM models and formalized with stub 

configuration matrices [144]. This technique helped 

reduce the number of test cases needed to detect 

undesirable interactions by focusing on interaction-prone 

combinations. This seminal work led to various 

improvements by Cheng et al. [61] and Zhang and 

Liu [200] in terms of the required pre-conditions, and by 

Leelaprute et al. [121] who added a second phase for the 

generation of error-prone scenarios from the interactionprone 

configurations. Hassine also adapted this filtering 

technique to identify interaction-prone combinations 

targeting LOTOS specifications, which were then 

checked formally using tests and goal-oriented 

executions [84]. In his thesis, Gorse proposed a different 

filtering technique, this time based on a logic 

representation of the feature requirements in Prolog. The 

filtering results are used for testing a LOTOS 

specification that formalizes features modeled with 

UCM [80]. 

Shiri et al. [174] combined UCM with Birkoff’s 

Formal Concept Analysis [35] to assist maintainers in 

identifying feature modification impacts at the 

requirements level, and minimizing the need for 

regression testing. 

Weiss and Esfandiari studied the feature interaction 

problem in terms of functional and non-functional 

interactions [192]. They used GRL to analyze conflicting 

goals, tradeoffs between softgoals, inadequate interfaces, 

ownership and policy issues, and resource contention. 

They also used UCM to analyze concurrency issues,


violations of assumptions, and incorrect invocation 

ordering. This work led to the first classification of 

undesirable interactions for Web services. 

More recently, Mussbacher et al. [143] studied 

semantic-based interactions in aspect-oriented models. 

Their approach differs from the syntactic approaches like 

filtering, and is more lightweight than detection methods 

that rely on the use of underlying formal languages. This 

approach requires the manual annotation of aspects with 

domain-specific markers, and a GRL model that specifies 

how markers from different domains influence each 

other. Automated analysis can then be used both to 

highlight semantic aspect conflicts and to trade-off 

aspects. This approach is demonstrated on academic and 

industrial examples that use aspect-oriented extensions of 

UCM [134] and other languages. 

G. Performance Analysis 

In URN, modelers may supplement UCM elements 

with standard performance annotations to describe 

resources associated with components, demands of 

responsibilities, workloads on scenario start points, 

allocations of UCM components to devices, and 

probabilistic behavior at selection points [106]. These 

annotations are not taken into consideration for the path 

traversal mechanism, but they can be used in 

transformations of UCM models to specialized 

performance models. This enables performance analysis 

from URN requirements models, before serious barriers 

to performance are frozen into the design and 

implementation. 

This part of the standard was strongly influenced by 

Woodside and his research team at Carleton University. 

In his PERFECT method, Scratchley used annotated 

UCM for evaluating concurrency architectures for a 

system that executes a given set of scenarios [169]. His 

annotations included timestamps and response-time 

requirements (implemented in the original UCMNAV 

tool), which were replaced in the URN standard by more 

generic metadata and URN links. 

The generation of Layered Queueing Network (LQN) 

performance models [177] directly from UCM models 

was first explored and prototyped by Petriu [152][153]. 

LQN performance models can be used as a basis for 

exploring the performance solution space of a system. 

Different kinds of analyses (e.g., sensitivity, scalability, 

concurrency, and configuration) can be performed 

through the use of LQN solver and simulation tools. 

Siddiqui et al. [175] improved upon this approach to 

consider the notions of budget and completions in the 

analysis of performance, while Liu focused on a multilevel 

methodology, with application to large presence 

systems [123]. Wu and Woodside explored the hybrid use 

of LQN and generalized stochastic Petri Nets for the 

performance analysis of annotated UCM models [194]. A 

good summary of the UCM-LQN performance 

engineering vision is found in [154]. 

The original annotations influenced the early 

development of the UML profile for schedulability, 

performance, and time [147]. More recent work on the 

development of the Core Scenario Model (CSM) 


representation [155], led to new annotations that are now 

part of the URN standard. These annotations have also 

evolved in synergy with the creation of the new UML 

profile for real time and embedded systems 

(MARTE) [148]. 

CSM’s purpose is to capture the essence of a range of 

scenario notations (e.g., from URN and UML) and enable 

simple transformations to various target formalisms (e.g., 

LQN, regular queueing networks, and stochastic Petri 

Nets), hence reducing the number and complexity of tools 

needed to analyze various aspects of the same system. 

Accordingly, newer URN-based approaches now target 

the generation of CSM models rather than LQN directly. 

A transformation from UCM to CSM was defined by 

Zeng in his thesis [199], with an implementation in 

UCMNAV. This transformation was adapted by 

Sincennes and others and is implemented in jUCMNav. 

One of the main benefits of this approach is that the 

acquisition and release of resources is inferred implicitly 

from UCM models rather than requiring them to be 

defined explicitly as in profiled UML models. This 

simplifies substantially the creation and maintenance of 

models. Transformations from CSM models to LQN 

models and other types of performance models are 

discussed in [155] and are now supported by prototype 

tools. 

Other types of software performance analysis based on 

UCM do not make use of the standard performance 

annotations. Billard used his own queueing simulator to 

analyze the UCM model of an object-oriented operating 

system [33], whereas Hassine used Timed Use Case 

Maps and a mapping to ASM to analyze resource 

allocation, worst-case time execution, and schedulability 

issues in an automatic protection switching feature [87]. 

Cai and Yu [54], on the other hand, investigated a 

GRL-based approach for qualitatively addressing and 

refining performance requirements. Operationalizations 

of such requirements are linked to UCM scenarios. 

H. Architecture Evaluation 

By combining goals with scenarios, URN provides a 

unique perspective on the evaluation of architectures. The 

previous section discussed performance-oriented 

approaches that often require the generation of 

mathematical performance models, where the quantitative 

parameters are difficult to choose and set. de Bruin and 

van Vliet explored a more qualitative approach to 

architecture evaluation, where a feature model describes 

the alternatives and refinements of the problem domain, 

whereas the solution domain is captured with a UCM 

model (with stubs and plug-ins) [65]. Links between the 

two models enables the evaluation and selection of an 

appropriate architecture, with its behavioral description. 

This approach shares similarities with the method of Liu 

and Yu [124], where a GRL goal model is used to capture 

actor intentions and coarse-grain alternatives, whose 

operationalizations are linked to the UCM view. 

Many surveyed approaches also focus on specific 

architectural qualities. For instance, Amyot describes an 

approach where alternative architectures can be evaluated 

on the complexity and cost of the resulting message


exchanges [9] as in the UCM example in Section II. Wu 

and Kelly have explored architecture evaluation from a 

security angle. They proposed a negative scenario 

framework where they explore “deviations” of UCM 

scenarios as potential security issues that impact 

architectural design decisions. Similar work was done by 

Karpati et al. [112], who introduced Misuse Case Maps 

as a modeling technique that is the anti-behavioral 

complement to UCM, which is used to visualize how 

cyber attacks are performed in an architectural context. 

The work of Folmer et al. [70] focuses on the use of 

scenarios (described with UCM and other means) for 

evaluating the usability of architectures before their 

implementation. 

I. System Comprehension and Evolution 

URN models are not just useful in a forward 

engineering development cycle. They can also be used in 

reverse-engineering, program comprehension, and 

evolution contexts to describe existing systems, 

architectures, and services. 

Amyot et al. have proposed a static approach to 

recovering UCM scenarios from code, based on a manual 

tagging approach and a commercial tool [19]. A dynamic 

approach was explored by Hamou-Lhadj et al., where 

execution traces are transformed to UCM scenarios [83]. 

One key step is the identification of utility functions in 

the code, which can be eliminated in order to shorten the 

resulting scenarios without loss of understandability. 

Hewitt and Rilling proposed a lightweight approach to 

identify the impact of requirement changes on a system 

based on the dependencies and potential ripple effects 

that can be inferred from a UCM model [94]. Shiri’s 

thesis expanded on this work with UCM-based 

techniques for impact analysis at the requirements level, 

prediction of regression testing effort, and feature 

interaction analysis, in order to support system evolution 

activities [173]. 

In his thesis, Störmer proposed the Software Quality 

Attribute Analysis by Architecture Reconstruction 

(SQUA3RE) method, where architectures are recovered 

based on a combination of UCM scenarios and timeperformance 

models [179]. The UCM models are built 

manually based on interviews. His study highlights that 

the participants appreciated the intuitiveness of UCM for 

showing flows of events and mappings to architectural 

components, and for decomposing structure and behavior. 

Störmer developed his own UCM tool, called 

Architecture Explorer, with support for timestamps and 

response-time requirements similar to Scratchley’s [169]. 

More recently, Díaz-Pace et al. [66] presented an 

approach called ArchSync (supported by an Eclipsebased 

UCM tool with the same name and initially 

developed by Blech) that helps architects synchronize 

architectural documentation expressed through UCM 

with Java source code, as modifications are being made 

on the code. Execution traces are used as an input, and 

inconsistencies with the architectural UCM model are 

then highlighted. ArchSync is actually complementing 

another tool (FLABot), discussed by Soria et al. [178]. 

FLABot is a fault-localization tool that uses a UCM 


specification and a set of architecture-to-code mappings 

in order to guide the architect in the identification of code 

regions with possible faults. UCMs were used as they “fit 

well with the exploration of cause-effect paths for faults”. 

Note that an interesting study by Ölvingson et al. 

evaluated UCM as a requirements engineering and 

system comprehension technique for the development of 

information systems in inter-organizational public health 

settings [149]. The UCM notation was found to be at a 

suitable level of abstraction and useful in generating 

intuitive requirements. At the time (2002), the authors 

also identified the absence of guidelines on how to use 

the notation as well as the difficulty in distinguishing asis 

models from to-be models as weaknesses that could 

benefit from further attention. 

J. Testing and Verification 

The availability of scenarios in URN models makes 

URN attractive for requirements-based testing. Beyond 

the various analysis techniques discussed so far in the 

sections on formalization, transformations, and feature 

interactions, we distinguish three main categories of 

approaches for the generation of test purposes from UCM 

models [23]. 

The first category is based on the usage of UCM 

models as is. For example, Amyot’s thesis defines a 

collection of testing patterns that can be used to manually 

cover a UCM model [8]. Charfi’s thesis also proposes an 

approach that generates test goals (this time, as LOTOS 

processes) by automatically covering the paths in UCM 

models [59]. Feng and Lee take into consideration 

statistical usage at the UCM level in order to guide the 

selection of important test cases for frequent paths [69]. 

The second category exploits standard UCM scenario 

definitions and path traversal algorithms. The techniques 

used for generating MSC scenarios can be reused as is to 

generate test purposes in MSC or in other formats. For 

example, Amyot et al. have used scenario definitions and 

the UCMNAV tool to generate test cases automatically for 

a Web application [22]. 

The third category requires the transformation of the 

UCM model to a formal specification from which 

existing test generation methods can be used. All of the 

mappings discussed in the formalization section are hence 

useful in this context. 

In order to turn test purposes extracted from UCM 

models into executable test cases, several issues must be 

addressed. For instance: UCM models do not include 

domain data, implementation messages and interfaces are 

unknown as UCM abstract from inter-component 

communication, and unfeasible test purposes might be 

selected depending on the chosen coverage strategy. 

While these concepts currently are not first-class URN 

modeling entities, some of them might be modeled 

indirectly with URN metadata and links. 

Although Jaskó et al. suggest that GRL can be used to 

provide rationale for test purposes and test 

strategies [108], Arnold et al. observe that URN in 

general should be augmented with a testable model for 

functional and non-functional requirements, an 

implementation under test, and explicit bindings between


the two views [29]. Through their experience in 

developing a model-based testing environment for .NET 

applications, they show the feasibility of having a URNbased 

testing approach where a URN model can be 

transformed into a testable requirements model from 

which executable test cases can be generated and then 

tested against an instrumented implementation (from 

which additional information is generated at run-time to 

check compliance with non-functional requirements). 

Hassine et al. take a different angle and consider UCM 

as a property specification language rather than a source 

of test purposes [91]. The resulting pattern system is 

mapped to popular temporal logics such as CTL, TCTL 

and ArTCTL (architectural real-time temporal logic, an 

extension to TCTL that provides temporal logics with 

architectural scopes). Properties extracted from UCM 

models can then be verified against designs and 

implementations using model checkers. 

K. Patterns 

As anticipated by Buhr in the 90’s [50], URN had a 

positive impact on the pattern-oriented development 

community. URN also benefited directly from some work 

in that community as well. 

In her thesis, Andrade developed a substantial UCMbased 

pattern language to describe common aspects of 

mobile telecommunication systems [25], whereas Billard 

used UCM to describe patterns of interactions in agent 

systems [34]. 

How to create UCM scenario models and exploit them 

are also the topics of several contributions. For example, 

Mussbacher and Amyot proposed a collection of UCM 

modeling patterns for describing and composing 

telecommunication features [135], whereas Amyot 

described a pattern language to derive test purposes from 

UCM models [8]. 

UCM can help describe the solution space of patterns, 

but GRL can also capture the various forces at play. This 

had already been observed for other goal-oriented 

languages such as the NFR framework, used by Gross 

and Yu to document pattern forces [81]. The GRL-UCM 

combination was exploited by Weiss in the description of 

various patterns for agent systems [188] and for Web 

applications [189]. 

GRL strategies can additionally be used to assess the 

qualitative impact of various solutions to a functional 

goal, in context, enabling users to select appropriate 

solutions. UCM level solutions can also be linked to each 

other with a proper use of stubs and plug-ins. This is at 

the basis of the URN-based formalization of patterns 

done by Mussbacher et al. and illustrated with an 

architectural pattern language [140]. Rather than using 

GRL strategies, Weiss and Mouratidis provided a 

mapping from GRL to Prolog to perform an evaluation of 

goal models describing the trade-offs that exist when 

selecting amongst a number of security patterns [193]. 

More recently, Behnam et al. used URN to formalize a 

pattern-based framework for goal-driven business process 

modeling, which can be used to derive suitable business 

processes (traceable to its objectives) for an organization 

whose context is also formalized with GRL. They 


illustrated the framework with a healthcare example. 

Pourshahid et al. [158] also explored business process 

reengineering patterns that require the combination of 

goals, scenarios (for processes), aspects, and indicators, 

as supported by the Aspect-oriented URN notation [134]. 

They demonstrated the potential of such patterns for 

evolving business processes at run-time. 

URN was also influenced by the literature on 

workflow patterns [162], which led directly to the 

introduction of new types of UCM stubs (synchronization 

and blocking) in the standard. The resulting 

expressiveness of UCM is compared to other scenario 

languages (i.e., BPMN, UML activity diagrams, and 

BPEL4WS) with the help of 43 workflow patterns 

in [136]. 

L. User Interface Engineering 

In our literature survey, we detected an original use of 

URN in the domain of user interface engineering, which 

was unforeseen when the standardization work was 

initiated. 

Folmer et al. [70] proposed a scenario-based usability 

assessment method to evaluate whether a given software 

architecture meets its usability requirements. Their 

Scenario-based Architecture Level UsabiliTy Assessment 

(SALUTA) makes use of UCM in its scenario evaluation 

step. One research issue they identified is the need for 

UCM to express static properties of usability. 

Such properties are actually proposed as UCM 

extensions in Alsumait’s thesis [6]. Alsumait et al. started 

investigating UCM as a medium for integrating task 

analysis (a topic already explored by Lethbridge and 

Singer [122]) and usability into a user-centered 

requirements engineering process [7]. The UCM notation 

was extended with concepts for supporting tasks, dialogs, 

and grouping/layout of user interface elements. They 

observed the potential of their extended UCM models for 

capturing user interface requirements. van der Poll et al. 

extended this work to provide a Z-based interpretation of 

UCM models, which enables formal usability 

analysis [183]. They provided an e-mail system as an 

example. 

Alsumait’s Scenario and Use Case-based for 

Requirements Engineering (SUCRE) framework provides 

the latest details on this work. The extended UCM 

notation for user interface modeling is presented together 

with examples as well as a set of analysis techniques 

based on metrics and on mappings to Z and LOTOS. 

These extensions were not included in the URN 

standard because usability engineering was not one of the 

original objectives. In addition, the extensibility of URN 

in terms of metadata and links enables the support of 

most of the extensions proposed here, except for their 

graphical representation. 

M. E-Business and Business Process Modeling 

For more than a decade, UCM have been used in the 

design of e-business applications. In particular, Gordijn 

and Akkermans, with the help of de Bruin and van Vliet, 

defined UCM extensions and ontological concepts for 

capturing (economic) value that became very successful


over the years [77]. This work resulted in a framework 

for value-based requirements engineering known as e 3 - 

value and detailed in Gordijn’s thesis [78]. e 3 -value, 

which has now become an independent language with its 

own tools and user community (see www.e3value.com), 

is used to model organizations in a value web, 

exchanging things of economic value with each other. 

Relationships between e 3 -value and goal modeling à la 

GRL are explored in [184]. 

Rather than focusing on value exchanges, Lethbridge 

and Singer [122] used UCM as one of their techniques for 

representing the work (i.e., the processes) of software 

engineers after observing it through shadowing. Later, 

Bleinstein et al. [36] proposed to use GRL combined with 

Jackson’s problem frames [107] and role activity 

diagrams in a requirements engineering approach that 

captures both business strategy and process requirements 

for e-business systems. In their approach, projections 

(rather than URN links) are used to connect the views. 

Additional work focusing on business process alignment 

was done in [37]. 

The combination of GRL and UCM for describing 

business objectives and processes/workflows then 

became very apparent. Weiss and Amyot argued that 

URN is a suitable notation for business process modeling, 

business evolution, and business alignment, and they 

illustrated their case with a supply chain management 

example [190]. Pourshahid and Tran have also shown the 

usefulness of URN in the modeling and analysis of trust 

in e-commerce systems [159]. 

In order to better handle business management 

concepts and be able to capture quantities in terms of 

domain-specific units, Pourshahid and Chen have 

extended GRL with the concept of Key Performance 

Indicators (KPI) [157]. A KPI converts a value observed 

in a running business process or context to a satisfaction 

level in the [-100,100] range understood by GRL. Target, 

threshold, and worst-case values are defined in each KPI 

to assist this conversion. In jUCMNav, GRL strategy 

definitions were also extended to access external sources 

of information (e.g., data warehouses, sensors, business 

intelligence application, or performance management 

tools) for online monitoring, management, and runtime 

adaptation of business processes [60]. These extensions 

were used in methodologies applied to several real 

healthcare process examples by, e.g., Pourshahid [156] 

and Kuziemsky et al. [119]. 

N. URN Tools 

Several tools have been developed over the years to 

support GRL, UCM, or URN modeling. On the UCM 

side, the development of the C++, multi-platform 

UCMNav tool [181] ended around 2005 in favor of the 

new Eclipse-based jUCMNav [110][137], a Java tool that 

actually started as an undergraduate student project [117]. 

jUCMNav (see Figure 4) was originally a simple UCM 

tool that prevents the creation of syntactically incorrect 

URN models. As part of his thesis [161], Roy added a 

GRL editor and invented the concept of GRL strategy, 

supported by a hybrid propagation algorithm with color 

feedback, as seen in Figure 1 [160]. He also provided 


goal-scenario traceability management (via URN links) 

and support for GRL catalogues. Kealey, in his 

thesis [114], implemented a flexible UCM traversal 

mechanism with color highlight (see Figure 2), together 

with an MSC export feature [115]. Kealey developed a 

mechanism where GRL evaluation results can influence 

the traversal of UCM paths, and vice-versa. Numerous 

semantic variation points in UCM were also identified for 

further clarification. Many of the contributions by Roy 

and Kealey found their way into the URN standard. Yan 

added a mechanism for the verification of user-defined 

static semantic rules and constraints written in OCL [24], 

whereas Gao recently contributed support for the import 

and export of URN models in the XML-based standard 

interchange format [71]. Other features, some of which 

are discussed in the previous sections, from dozens of 

contributors are also present in this tool. 

Figure 4 Overview of the jUCMNav tool interface. 

jUCMNav is for the moment the only tool that 

supports both goal and scenario modeling and analysis. 

As discussed earlier, other tools with partial and 

specialized support for UCM also exist: Störmer’s 

Architecture Explorer for architecture recovery 

activities [179] and ArchSync for the documentation, 

maintenance and diagnosis of applications written in 

Java [66]. 

On the GRL side, Liu extended Yu’s Organization 

Modelling Environment (OME 3) to support an early 

version of the notation and an interactive propagation 

mechanism [198]. This was the only GRL tool available 

for a long time, and it helped shape GRL as it is known 

today. OME 3 was also deprecated in favor of an Eclipsebased 

version called OpenOME [150], a project led by 

Yu whose major contributors include Ernst, Horkoff, Ng, 

Olinescu, and Y. Yu. OpenOME integrates with other 

platforms (such as Protégé and Visio) to support goaloriented 

and agent-oriented modeling. Strategy-based 

evaluation is not supported, but there are a variety of 

analysis features, including interactive propagation. 

There also exists a Visio-based tool, named Sandrila 

SDL [163], which supports GRL modeling, without 

analysis capabilities. Other goal-oriented modeling tools 

are discussed and evaluated on the i* Wiki [97].


O. Requirements Management and Compliance 

URN models capture only a fraction of the 

requirements of telecommunication standards and 

software products. Accordingly, such models need to be 

used in cooperation with complementary general 

requirements, and both views must be linked in a way 

that supports traceability, navigation, and analysis. The 

proposed URN standard ensures that model elements are 

uniquely identifiable inside a specification, which helps 

supporting such links. However, one important challenge 

that remains is the maintenance of these links as models 

and general requirements evolve. 

Jiang proposed an approach to export UCM scenario 

models to the IBM Rational DOORS requirements 

management system and to maintain relationships as both 

views evolve over time [109]. Originally developed for 

UCMNAV, this functionality is now supported as an 

export filter for jUCMNav, thanks to the efforts of Kim et 

al. [116]. The tool also provides a link auto-completion 

mechanism to minimize the possibly large number of 

links that have to be created manually by DOORS users 

between external and UCM requirements. Roy later 

extended this mechanism to support GRL and URN 

links [161]. 

Ghanavati built on this work to study the compliance 

of organization goals and business processes against laws 

and policies [73]. URN is used both to capture the goals 

and processes of the organization and to model 

legislation. Exporting and linking both views to DOORS 

enables one to assess the legal compliance of business 

processes, as well as maintain it when laws or processes 

evolve [74]. Her original framework was extended to 

exploit contributions of processes to the elements of the 

law/policy model, enabling the measure of partial 

compliance [76]. 

All of the above contributions convinced us that URN 

models can indeed be combined with other types of 

requirements and design artifacts in a requirements 

management context. 

P. Aspect-oriented Modeling 

Over the last decade, aspect-oriented modeling (AOM) 

techniques have been developed for many requirements 

and design notations in order to better address separation 

of concern issues found in complex models. 

The UCM notation’s ability to model aspects was 

identified in the late 90’s by Buhr [48] but received little 

attention since then with the exception of work by de 

Bruin and van Vliet [65]. The top-down approach by de 

Bruin and van Vliet explicitly added a “Pre” stub and a 

“Post” stub for each location on a map that requires a 

change. The stubs allowed behavior to be added before or 

after the location by plugging refinement maps into the 

stubs. 

In 2005, work started on the Aspect-oriented User 

Requirements Notation (AoURN), which is best 

described in Mussbacher’s thesis [134]. Mussbacher 

proposed aspect-oriented extensions to UCM and GRL 

models to unify goal-oriented, scenario-based, and 

aspect-oriented techniques in one modeling framework. 


AoURN allows a concern to be encapsulated even if it is 

crosscutting other concerns, thus leading to 

improvements in the modularity, maintainability, and 

reusability of URN models. The concept of a concern was 

deemed important enough to be included in the URN 

standard. In AoURN, patterns and composition rules are 

described with URN itself, thus allowing for a flexible 

and exhaustive approach that is not limited to a particular 

composition language but can harness the full expressive 

power of URN [141]. In AOM, patterns specify where an 

aspect is to be applied, and composition rules specify 

how an aspect is to be applied at the location identified by 

a pattern. The matching and composition mechanism of 

AoURN goes well beyond typical composition operators 

and includes among others concurrent, loop, and 

interleaved composition. AoURN’s mechanism is further 

enhanced by taking semantic equivalences of URN 

related to hierarchical structuring into account [142]. This 

approach allows common refactoring operations to be 

performed on an AoURN model without breaking 

aspectual specifications. 

AoURN has been used for various applications and 

some of them are discussed here as examples. A large 

challenge problem posed by the aspect-oriented modeling 

community involved a safety-critical, reactive 

system [139]. In the context of business process 

monitoring and improvement, AoURN enabled changes 

to business processes based on business process redesign 

patterns and an assessment of the shortcomings of the 

current business process with the help of goal 

models [158]. Crosscutting concerns were also described 

with AoURN for a SOA-based application and added to 

composite services based on an assessment of nonfunctional 

properties modeled with URN [31]. Finally, 

AoURN has also been applied to model commonalities 

and variabilities in Software Product Lines [138]. 

Q. Support for Standardization 

Several authors have emphasized the role of URN in 

the development of standards and systems, beyond the 

vision described by Hodges and Visser in 1999 [96]. For 

instance, Sales described how UCM can fit in the 

development of IETF protocol standards [167]. Adamis et 

al. briefly compared ITU-T languages (including URN) 

with UML, and illustrated how the former can be used 

together to model systems [4]. Medve also studied the 

ITU-T languages (with an emphasis on UCM) and UML, 

this time in the context of system re-engineering [128]. 

From the perspective of ITU-T standardization 

processes, URN’s capability to model goals and scenarios 

fits well the so-called “stage 1” requirements descriptions 

described in Recommendation I.130 [100]. In the 

telecommunications networks management domain, 

Recommendation M.3020 proposes the description of 

various types of requirements (functional, non-functional, 

administrative, etc.) with textual use case and UML use 

case diagrams [101]. Again, URN models fit nicely in 

such a process as they bring formality and executability 

to the use cases while enabling concrete support for goal 

models, which are useful to derive and analyze nonfunctional 

and administrative requirements.


More recently, the growing interest in standards for 

Next-Generation Networks (NGN) brought new needs for 

improved service description and engineering 

approaches. Ideally, one would like to specify and 

analyze services and standards at a high level of 

abstraction, using modeling concepts close to the user and 

problem domain rather than at the platform and 

implementation domain, and then be able to derive design 

components and implementations from service models 

with a high degree of automation. This is essentially the 

abstraction level targeted by URN, as discussed in [12]. 

GRL goal models offer a holistic view that integrates 

stakeholder goals, non-functional requirements, and 

alternative operational solutions for design time 

decisions, supplemented with indicators that enable 

adaptive behavior at runtime. UCM offer scenarios that 

express variability points explicitly while offering much 

flexibility in ordering activities, which may be bound to 

components or not. The integration of GRL and UCM, 

combined with strategies and scenario definitions, and 

possibly with aspect-oriented concepts, emphasizes the 

importance of enabling dynamic choices in the service 

modeling and design phases in order to take into account 

contextual information and differentiated service 

availability requirements in dynamic service composition, 

which are key aspects of NGN services. 

VI. THE NEXT TEN YEARS 

The previous section proved that much development 

related to URN has happened in the past ten years. Yet, 

we expect the next ten years to be even more exciting in 

terms of the diversity of application domains for URN, 

and of new modeling and analysis features that will be 

emerging. In particular, we predict major developments 

in the following eight areas. 

A. Domain-Specific Profiles 

There will be a need to tailor and extend URN for 

specific application domains. The need for extensions 

was already raised on several occasions in the previous 

section, for instance in the areas of user interface 

engineering and testing. It has been observed for other 

domains as well, including the popular area of software 

product lines [44], or emerging domains like home 

network systems [151]. 

The URN standard already offers mechanisms that, 

when combined, support the profiling of the language to a 

particular domain. 1) Metadata are name-value pairs that 

can be used to tag any URN element, similar to 

stereotypes in UML. 2) URN links can be used to specify 

relationships between any pair of URN elements. 3) URN 

concerns can be used to group any collection of URN 

elements (including other concerns). 4) OCL constraints 

can be defined to restrict the use of the language or of its 

extensions [24]. For example, a URN profile for i* is 

defined and implemented in [18], demonstrating that 

URN can represent concepts and constraints found in 

another language. 

ITU-T is also inviting contributions on the definition 

of UML profiles for all its languages. Abid et al. have 


already defined a tool-supported UML profile for 

GRL [3] (based on ITU-T guidelines). Molina et al. also 

proposed a UML profile for measurable goal modeling 

(i.e., with indicators), with the GRL syntax as its concrete 

notation [131]. However, much work remains to be done 

to cover URN entirely. Progress in this direction may also 

cause the URN standard to include the capability of 

creating profiles (including changing the shapes of some 

elements) as first-class entities. 

B. Enhanced Workflow Executions 

The URN standard defines advanced workflow 

operators (e.g., blocking stubs with threshold and 

replication factors) together with the corresponding 

traversal rules. However, at this time, no tool is currently 

supporting them at the analysis/simulation level. 

URN is also missing important constructs to specify 

properly cancellation scenarios and situations akin to 

exception handling. Mussbacher proposed the addition of 

failure points and failure start points, which would enable 

the concise representation and analysis of cancellation 

situations [134]. This topic was actually adopted in a 

revised version of Recommendation Z.150 in February 

2011. 

These new operators will require proper tool support to 

be useful. In addition, tools could consider supporting 

other execution modes like debuggers, which would be 

useful during the analysis of workflow models. 

C. Performance Management 

The concept of indicator (KPI) is another addition 

being considered in the short term for the URN standard 

and recently adopted in the revised Recommendation 

Z.150. We have had over three years of experience with 

KPI in jUCMNav and numerous models [60][157], and 

they have quickly become an essential part of any 

description of business process or adaptive system in a 

performance management context. 

Dynamic adaptation of systems based on KPI is an 

exciting area of research that is expected to benefit from 

URN’s simultaneous support for goal modeling and 

scenario modeling. With KPI, real-world values may 

influence the evaluation of goal models and drive the 

simulation of what-if scenarios to guide and inform 

dynamic adaptation at run-time. 

There might however be a need for a more flexible 

definition of what a KPI is at the metamodel level. For 

example, the mapping of an observed value to a GRL 

satisfaction level could be done differently than the 

current simple linear correspondence. KPI in a model 

could also influence or contribute to each other, enabling 

the computation of aggregate KPI. Trends computed from 

external data sources could also be integrated in GRL 

models. Such improved KPI definitions are actually being 

explored in the Business Intelligence Model 

language [30], inspired partly from URN. 

D. Compliance Management 

The assessment of compliance of business goals and 

processes with laws and policies is also a domain where 

URN is expected to have an impact in the next ten years.


Following the initial work of Ghanavati et al. [76], there 

are still important challenges to be addressed, including 

the systematic extraction of URN models from textual 

laws and policies, and the prioritization of efforts to 

improve a partial degree of compliance. Other research 

questions include the potential need for deontic 

modalities (e.g., obligations, permissions, and 

interdictions) or Hohfeldian classes of rights in URN 

models of laws and policies [75], as well as the role of 

indicators for measuring compliance [172]. 

E. Formal Semantics 

URN has a formal description of its abstract syntax, 

but only a natural language description of its semantics in 

terms of rules constraining the propagation in GRL and 

the traversal of UCM paths. There is a need to reduce the 

number of semantic variation points (known and 

unknown) in the language. This could be done with a 

complete mapping to an underlying formalism. Many 

partial mappings were discussed in Section V.D, but 

complete mappings are much more difficult to achieve, 

especially if they are to span GRL and UCM. There 

might also be a way to provide a formal description in the 

form of a virtual machine describing the interpretation of 

URN models. 

F. Improved Analysis Techniques 

There are many opportunities to add to the set of 

analysis techniques currently used in URN. For example, 

the current GRL propagation algorithms based on 

strategies are mainly bottom-up, in a way similar to test 

cases. The availability of top-down algorithms would 

provide substantial benefits to many users, who would 

simply ask the model how to optimize one or several 

actors or intentional elements given an initial context. 

The main difficulty is usually that top-down algorithms 

correspond to a kind of search problem and are hence 

much more complex than bottom-up algorithms. Weiss 

and Mouratidis [193] have a mapping from GRL to 

Prolog that might help solve this issue. Okamura et 

al. [151] may also have elements of answers. 

The analysis of aspect-oriented GRL and UCM in 

AoURN is still an open issue [134], and this may even 

have an impact on how strategies and scenarios are 

defined in URN. 

Time extensions and analysis for UCM, as proposed by 

Hassine et al [85], are also relevant to URN and hence 

deserve some attention. 

Finally, from an analysis perspective, there is a need 

for a tighter integration between GRL and UCM, which 

could result in the combination of strategies and scenarios 

in one logical unit in the URN standard. 

G. Improved Model Representations 

Recent analyses of the graphical syntaxes of UCM by 

Genon et al. [72] and of i* by Moody et al. [133] reveal 

that there are many problems with the cognitive fitness of 

the symbols used, and with the completeness of the 

language’s concrete syntax. For instance, performance 

annotations and stub bindings do not have any visual 

representation in URN. The concrete visual syntaxes 


could be reviewed and completed in the standard. 

Additional concrete syntaxes, such as textual or tabular 

for GRL graphs, also deserve to be explored. 

H. Guidelines and Methodologies 

Last but not least, there is currently a lack of guidelines 

and methodologies for URN modeling, both in isolation 

and in combination with other modeling languages. This 

was already raised as an issue for UCM in 2002 by 

Ölvingson et al. [149], and this is unfortunately still true 

to some extent today. A related challenge is the 

automated or semi-automated transformation of URN 

models to other modeling techniques. While many 

transformations have already been investigated in Section 

V.E, there is a need to revisit some of these 

transformations and new transformations to emerging 

modeling techniques in the light of recent technologies 

and applications such as AOM, domain-specific 

languages, dynamic adaptation of systems, and SOAbased 

systems. ITU-T also invites contributions to the 

definition of a URN-based methodology. Better 

guidelines and methodologies could have a strong impact 

on the adoption of URN in industry. 

VII. CONCLUSIONS 

This paper reports on a systematic literature survey 

about the development of the User Requirements 

Notation. Section II first gives an overview of the 

notation and typical analysis techniques, followed by a 

discussion of the origins of URN in the 1990’s. In 

Section IV, the analysis of the 281 papers selected for this 

study reveals that URN is a growing, global language, 

both in terms of contributors and users. 

In section V, we introduce and commented on 17 

categories of contributions to URN during its first ten 

years. This period is mainly characterized as follows: 

• Simple formalization of URN in terms of abstract 

and concrete syntaxes; 

• The definition of analysis techniques such as 

GRL strategies with forward propagation and 

UCM scenario definitions with a path traversal 

mechanism; 

• The completion of a first version of the URN 

standard; 

• Simple combinations of GRL and UCM in 

models, with traceability, completeness, and 

consistency analysis, and with GRL strategies 

and UCM scenarios that can influence each 

others; 

• Emerging techniques for the analysis of feature 

interactions, performance, and architectures; 

• The availability of open-source, Eclipse-based 

tool support (jUCMNav); 

• A multitude of application domains explored, 

mainly related to telecommunication systems and 

reactive systems at the beginning, and later 

mainly related to business processes and aspectoriented 

modeling; 

• Many formalizations and transformations also 

explored.


We believe that the next ten years of URN 

development will be even more active than the first ten 

and will focus on the major topics identified in Section 

VI, including domain-specific profiles, enhanced 

workflow executions, performance and compliance 

management, formal semantics, improved analysis 

techniques and model representations, and guidelines and 

methodologies. 

We hope this survey paper represents a useful one-stop 

document for URN beginners and experts alike. We also 

take the opportunity to invite users and other interested 

parties to get actively involved in future developments of 

the User Requirements Notation. 

ACKNOWLEDGMENT 

The authors wish to thank the many people who have 

contributed to the success of the User Requirements 

Notation over the years, with special thanks to C.M. 

Woodside and T. Gray for comments on an earlier draft 

of this paper. This work was supported in part by the 

Discovery grants and Postgraduate Scholarships 

programs from NSERC (Canada) and by the Ontario 

Graduate Scholarship Program. 

REFERENCES 

[1] T. Abdelaziz, Towards a Comprehensive Agent-Oriented 

Software Engineering Methodology, Doctoral 

Dissertation, Universität Duisburg-Essen, Germany, 

October 2008 

[2] T. Abdelaziz, M. Elammari, and R. Unland, “Visualizing 

a Multiagent-Based Medical Diagnosis System Using a 

Methodology Based on Use Case Maps”, in MATES 

2004: multiagent system technologies, LNCS 3187, 

Springer, pp. 198–212, 2004. doi:10.1007/978-3-540- 

30082-3_15 

[3] M.R. Abid, D. Amyot, S.S. Somé, and G. Mussbacher, “A 

UML Profile for Goal-Oriented Modeling”, in SDL 2009: 

Design for Motes and Mobiles, 14th Int. SDL Forum, 

LNCS 5719, Springer, pp. 133–148, September 2009. 

doi:10.1007/978-3-642-04554-7_9 

[4] G. Adamis, R. Horváth, Z. Pap, and K. Tarnay, 

“Standardized languages for telecommunication systems”. 

Computer Standards & Interfaces, 27(3), Elsevier, pp. 

191–205, March 2005. doi:10.1016/j.csi.2004.09.005 

[5] C.P. Ayala, C. Cares, J.P. Carvallo, G. Grau, M. Haya, G. 

Salazar, X. Franch, E. Mayol, and C. Quer, “A 

Comparative Analysis of i*-Based Goal-Oriented 

Modelling Languages”, in Int. Workshop on Agent- 

Oriented Software Development Methodologies (AOSDM 

@SEKE), Taipei, China, pp. 43–50, July 2005 

[6] A. Alsumait, User Interface Requirements Engineering: A 

Scenario-Based Framework. Ph.D. thesis, Concordia 

University, Canada, August 2004. 

[7] A. Alsumait, A. Seffah, and T. Radhakrishnan, “Use Case 

Maps: A Roadmap for Usability and Software Integrated 

Specification”, in 17th World Computer Congress - TC13 

Stream on Usability, IFIP, pp. 119–131, August 2002, 

[8] D. Amyot, Specification and Validation of Telecommunications 

Systems with Use Case Maps and LOTOS. Ph.D. 

thesis, SITE, University of Ottawa, Canada, Sept. 2001. 

[9] D. Amyot, “Introduction to the User Requirements 

Notation: Learning by Example”. Computer Networks, 


42(3), pp. 285–301, June 2003. doi:10.1016/S1389- 

1286(03)00244-5 

[10] D. Amyot and R. Andrade, “Description of Wireless 

Intelligent Network Services with Use Case Maps”, in 

17th Brazilian Symposium on Computer Networks 

(SBRC'99), Salvador, Brazil, pp. 418–433, May 1999. 

[11] D. Amyot, R. Andrade, L. Logrippo, J. Sincennes, and Z. 

Yi, “Formal Methods for Mobility Standards”, in IEEE 

1999 Emerging Technology Symposium on Wireless 

Communications & Systems, Dallas, USA, pp. 14.1–14.7, 

April 1999. doi:10.1109/ETWCS.1999.897332 

[12] D. Amyot, H. Becha, R. Bræk, and J.E.Y. Rossebø, “Next 

Generation Service Engineering”, in ITU-T Innovations in 

NGN - Kaleidoscope Academic Conference, Geneva, 

Switzerland, pp. 195–202, May 2008. doi:10.1109/ 

KINGN.2008.4542266 

[13] D. Amyot, F. Bordeleau, R.J.A. Buhr, and L. Logrippo, 

“Formal support for design techniques: a Timethreads- 

LOTOS approach”, in FORTE VIII, 8th Int. Conf. on 

Formal Description Techniques, Chapman & Hall, pp. 

57–72, 1995. 

[14] D. Amyot, L. Charfi, N. Gorse, T. Gray, L. Logrippo, J. 

Sincennes, B. Stepien, and T. Ware, “Feature Description 

and Feature Interaction Analysis with Use Case Maps and 

LOTOS”, in Sixth International Workshop on Feature 

Interactions in Telecommunications and Software Systems 

(FIW'00), IOS Press, pp. 274–289, May 2000. 

[15] D. Amyot, D.Y. Cho, X. He, and Y. He, “Generating 

Scenarios from Use Case Map Specifications”, in Third 

Int. Conf. on Quality Software (QSIC'03), IEEE CS, pp. 

108–115, Nov. 2003. doi:10.1109/QSIC.2003.1319092 

[16] D. Amyot and A. Eberlein, “An Evaluation of Scenario 

Notations and Construction Approaches for 

Telecommunication”. Telecommunications Systems 

Journal, 24(1), Kluwer, pp. 61–94, September 2003. 

doi:10.1023/A:1025890110119 

[17] D. Amyot, S. Ghanavati, J. Horkoff, G. Mussbacher, L. 

Peyton, and E. Yu, “Evaluating Goal Models within the 

Goal-oriented Requirement Language”. Int. Journal of 

Intelligent Systems, 25(8), Wiley, pp. 841–877, August 

2010. doi:10.1002/int.20433 

[18] D. Amyot, J. Horkoff, D. Gross, and G. Mussbacher, “A 

Lightweight GRL Profile for i* Modeling”, in 3rd Int. 

Work. on Requirements, Intentions and Goals in 

Conceptual Modeling (RIGiM 2009), LNCS 5833, 

Springer, pp. 254–264, Nov. 2009. doi:10.1002/int.20433 

[19] D. Amyot, M. Mussbacher, and N. Mansurov, 

“Understanding Existing Software with Use Case Map 

Scenarios”, in 3 rd SDL and MSC Workshop (SAM02), 

LNCS 2599, Springer, pp. 124–140, June 2002. 

doi:10.1007/3-540-36573-7_9 

[20] D. Amyot and L. Logrippo, “Use Case Maps and LOTOS 

for the Prototyping and Validation of a Mobile Group Call 

System”. Computer Communications, 23(12), pp. 1135– 

1157, July 2000. doi:10.1016/S0140-3664(99)00242-X 

[21] D. Amyot, L. Logrippo, R.J.A. Buhr, and T. Gray, “Use 

Case Maps for the Capture and Validation of Distributed 

Systems Requirements”, in Fourth Int. Symposium on 

Requirements Engineering (RE'99), IEEE CS, pp. 44–53, 

June 1999. doi:10.1109/ISRE.1999.777984 

[22] D. Amyot, J.-F. Roy, and M. Weiss, “UCM-Driven 

Testing of Web Applications”, in 12th SDL Forum (SDL 

2005), LNCS 3530, Springer, pp. 247–264, June 2005. 

doi:10.1007/11506843_18 

[23] D. Amyot, M. Weiss, and L. Logrippo, “Generation of 

Test Purposes from Use Case Maps”. Computer


Networks, 49(5), Elsevier, pp. 643–660, December 2005. 

doi:10.1016/j.comnet.2005.05.006 

[24] D. Amyot and J.B. Yan, “Flexible Verification of User- 

Defined Semantic Constraints in Modelling Tools”, in 

18th Int. Conf. of Computer Science and Software 

Engineering (CASCON), IBM CAS, October 2008. 

doi:10.1145/1463788.1463798 

[25] R. Andrade, Capture, Reuse, and Validation of 

Requirements and Analysis Patterns for Mobile Systems. 

Ph.D. thesis, SITE, Univ. of Ottawa, Canada, May 2001. 

[26] R. Andrade and L. Logrippo, “Reusability at the Early 

Development Stages of the Mobile Wireless 

Communication Systems”, in 4th World Multiconference 

on Systemics, Cybernetics and Informatics (SCI 2000), 

IIIS, Orlando, USA, pp. 11–16, July 2000. 

[27] R. Andrade, W. Viana, and D.P. Menezes, “A high-level 

application framework for mobile system development: 

IMT-2000 case study”, in 9th Int. Conf. on 

Telecommunications, IEEE, Beijing, China, pp. 321–325, 

June 2002. 

[28] P. Araya and H. Antillanca, “Una metodología de agents”, 

in 1er. Workshop Chileno de Ingeniería de Software, 

Punta Arenas, Chile, November 2001. 

[29] D. Arnold, J.-P. Corriveau, and W. Shi, “Scenario-Based 

Validation: Beyond the User Requirements Notation”, in 

21st Australian Software Engineering Conf. (ASWEC 

2010), IEEE CS, pp 75–84, April 2010. doi:10.1109/ 

ASWEC.2010.29 

[30] D. Barone, E. Yu, J. Won, L. Jiang, and J. Mylopoulos, 

“Enterprise Modeling for Business Intelligence”, in The 

Practice of Enterprise Modeling, LNBIP 68, Springer, pp. 

31–45, 2010. doi:10.1007/978-3-642-16782-9_3 

[31] H. Becha, G. Mussbacher, and D. Amyot, “Modeling and 

Analyzing Non-Functional Requirements in Service 

Oriented Architecture with the User Requirements 

Notation”. Non-functional Properties in Service Oriented 

Architecture: Requirements, Models and Methods, IGI 

Global, USA, pp. 48–72, 2011. doi:10.4018/978-1-60566- 

794-2.ch003 

[32] S.A. Behnam, D. Amyot, and G. Mussbacher, “Towards a 

Pattern-Based Framework for Goal-Driven Business 

Process Modeling”, in 8th Int. Conf. on Software 

Engineering Research, Management and Applications 

(SERA2010), IEEE CS, pp. 137–145, May 2010. 

doi:10.1109/SERA.2010.27 

[33] E.A. Billard, “Operating system scenarios as Use Case 

Maps”, in Fourth Int. Work. on Software and 

Performance (WOSP 2004), ACM Press, pp. 266–277, 

January 2004. doi:10.1145/974044.974087 

[34] E.A. Billard, “Patterns of agent interaction scenarios as 

Use Case Maps”. IEEE Transactions on Systems, Man 

and Cybernetics, 24B:4, pp. 1933–1939, August 2004. 

doi:10.1109/TSMCB.2004.828192 

[35] G. Birkhoff, Lattice theory. American Mathematical 

Society, 1967. 

[36] S.J. Bleistein, K. Cox, and J. Verner, “Requirements 

Engineering for e-Business Systems: Integrating Jackson 

Problem Diagrams with Goal Modeling and BPM”, in 

11th Asia Pacific Software Engineering Conference 

(APSEC 2004), IEEE CS, pp. 410–417, November 2004. 

doi:10.1109/APSEC.2004.84 

[37] S.J. Bleistein, K., Cox, J. Verner, and K.T. Phalp, 

“Requirements engineering for e-business advantage”, 

Requirements Engineering, 11(1), pp. 4–16, March 2006. 

doi:10.1007/s00766-005-0012-7 


[38] E. Börger and R. Stärk, Abstract State Machines: A 

Method for High-Level System Design and Analysis. 

Springer-Verlag, 2003. 

[39] F. Bordeleau, A Systematic and Traceable Progression 

from Scenario Models to Communicating Hierarchical 

State Machines. Ph.D. thesis, SCE Dept., Carleton 

University, Canada, December 1999. 

[40] F. Bordeleau and R.J.A. Buhr, “The UCM-ROOM Design 

Method: from Use Case Maps to Communicating State 

Machines”, in Conf. on the Engineering of Computer- 

Based Systems (ECBS), pp. 167–179, March 1997. 

doi:10.1109/ECBS.1997.581850 

[41] F. Bordeleau and D. Cameron, “On the Relationship 

between Use Case Maps and Message Sequence Charts”, 

in 2nd Workshop on SDL and MSC (SAM 2000), 

Grenoble, France, pp. 123–138, June 2000. 

[42] F. Bordeleau, J.-P. Corriveau, and B. Selic, “A Scenario- 

Based Approach to Hierarchical State Machine Design”, 

in ISORC 2000: 3rd IEEE Int. Symp. on Object-Oriented 

Real-time distributed Computing, IEEE CS, pp. 78–85, 

March 2000. doi:10.1109/ISORC.2000.839514 

[43] P. Brereton, B.A. Kitchenham, D. Budgen, M. Turner, 

and M. Khalil, “Lessons from applying the systematic 

literature review process within the software engineering 

domain”. J. of Systems and Software, 80(4), pp. 571–583, 

April 2007. doi:10.1016/j.jss.2006.07.009 

[44] J. Brown, R. Gawley, I. Spence, P. Kilpatrick, C. Gillan, 

and R Bashroush, “Requirements Modelling and Design 

Notations for Software Product Lines”, in First Int. 

Workshop on Variability Modelling of Software-intensive 

Systems (VaMoS), Limerick, Ireland, January 2007. 

[45] H. de Bruin, “A Grey-Box Approach to Component 

Composition”, in Generative and Component-Based 

Software Engineering (GCSE 99), pp. 195–209, 1999. 

[46] R.J.A. Buhr, “Use Case Maps for Attributing Behaviour 

to System Architecture”, in Fourth Int. Work. on Parallel 

and Distributed Real Time Systems (WPDRTS), pp. 3–10, 

1996. 

[47] R.J.A. Buhr, “Design Patterns at Different Scales”, in 

Pattern Languages of Programs (PLoP96), June 1996. 

[48] R.J.A. Buhr, “A Possible Design Notation for Aspect 

Oriented Programming”, in ECOOP Workshop on Aspect 

Oriented Programming, Brussels, Belgium, July 1998. 

[49] R.J.A. Buhr, “Use Case Maps as Architectural Entities for 

Complex Systems”. IEEE Transactions on Software 

Engineering, 24(12), pp. 1131–1155, December 1998. 

doi:10.1109/32.738343 

[50] R.J.A. Buhr, “Understanding Macroscopic Behaviour 

Patterns in Object-Oriented Frameworks, with Use Case 

Maps (chapter 18)”. Building Application Frameworks: 

Object-Oriented Foundations of Framework Design, 

Wiley, pp. 415–440, September 1999. 

[51] R.J.A. Buhr, D. Amyot, M. Elammari, D. Quesnel, T. 

Gray, and S. Mankovski, “Feature-Interaction Visualization 

and Resolution in an Agent Environment”, in Fifth 

Int. Work. on Feature Interactions in Telecommunications 

and Software Systems (FIW'98), IOS Press, pp. 135–149, 

July 1998. 

[52] R.J.A. Buhr and R.S. Casselman, Use Case Maps for 

Object-Oriented Systems. Prentice-Hall, November 1995. 

[53] G. Bush, S. Cranefield, and M.K. Purvis, “The Styx agent 

methodology”. Information Science Discussion Paper 

Series, 2001/02, University of Otago, New Zealand, 2001. 

[54] Z. Cai and E.Yu, “Addressing Performance Requirements 

Using a Goal and Scenario-Oriented Approach”, in 

CAISE'02: 14th Int. Conf. on Advanced Information


Systems Engineering, LNCS 2348, Springer, pp. 706–710, 

May 2002. doi:10.1007/3-540-47961-9 

[55] M. Calder, M. Kolberg, E.H. Magill, and S. Reiff- 

Marganiec, “Feature interaction: a critical review and 

considered forecast”. Computer Networks, 41, pp. 115– 

141, 2003. 

[56] R.S. Casselman, A Role-Based Architectural Model 

Applied to Object-Oriented Systems. Master’s thesis, SCE 

Dept., Carleton University, Canada, August 1993. 

[57] H.N. Castejón Martínez, “Synthesizing State-Machine 

Behaviour from UML Collaborations and Use Case 

Maps”, in 12th SDL Forum (SDL 2005), LNCS 3530, 

Springer, pp. 339–359, June 2005. doi:10.1007/ 

11506843_24 

[58] H.N. Castejón, Collaborations in Service Engineering: 

Modeling, Analysis and Execution. Ph.D. thesis, Dept. of 

Telematics, NTNU, Norway, November 2008. 

[59] L. Charfi, Formal Modeling and Test Generation 

Automation with Use Case Maps and LOTOS. M.Sc. 

thesis, SITE, University of Ottawa, Canada, Feb. 2001. 

[60] P. Chen, Goal-Oriented Business Process Monitoring: An 

Approach based on User Requirement Notation combined 

with Business Intelligence and Web Services. M.Sc. 

thesis, SCS Dept., Carleton University, Canada, 

December 2007. 

[61] J. Cheng, L. Yang, Y.-J. Kuai, and D.-F. Zhang, “Nondeterministic 

feature interaction filtering method based on 

scenarios with Use Case Map” (基于用例图呼叫处理场景的不 

确定性冲突过滤方法), Hunan Daxue Xuebao / Journal of 

Hunan University Natural Sciences, 32(2), pp. 104–109, 

April 2005. 

[62] L. Chung, B.A. Nixon, E. Yu, and J. Mylopoulos, Non- 

Functional Requirements in Software Engineering. 

Kluwer Academic Publisher, 2000. 

[63] L. Constantine and L. Lockwood, Software for Use: A 

Practical Guide to the Models and Methods of User- 

Centered Design. Addison-Wesley, April 1999. 

[64] H. de Bruin, “Scenario-Based Analysis of Component 

Compositions”, in Generative and Component-Based 

Software Engineering (GCSE'00), LNCS 2177, Springer, 

pp. 129–146, Oct. 2000. doi:10.1007/3-540-44815-2_10 

[65] H. de Bruin and H. van Vliet, “Quality-Driven Software 

Architecture Composition”. Journal of Systems and 

Software, 66(3), Elsevier, pp. 269–284, June 2003. 

doi:10.1016/S0164-1212(02)00079-1 

[66] J.A. Díaz-Pace, J.P. Carlino, M. Blech, A. Soria, and 

M.R. Campo, “Assisting the Synchronization of UCMbased 

Architectural Documentation with 

Implementation”, in IEEE/IFIP Conf. on Software 

Architecture and European Conference on Software 

Architecture (WICSA/ECSA 2009), IEEE CS, pp. 151– 

160, September 2009. doi:10.1109/WICSA.2009.5290801 

[67] C. Dongmo and J. A. van der Poll, “Use Case Maps as an 

Aid in the Construction of a Formal Specification”, in 7th 

Int. Workshop on Modelling, Simulation, Verification and 

Validation of Enterprise Information Systems (MSVVEIS- 

2009), INSTICC Press, pp. 3–13, May 2009. 

[68] M. Elammari and W. Lalonde, “An agent-oriented 

methodology: High-level and intermediate models”, in 

Proc. 1st Int. Workshop on Agent-Oriented Information 

Systems, Seattle, USA, June 1999. 

[69] Y. Feng and L.-S. Lee, “The Importance Analysis of Use 

Case Map with Markov Chains”. Int. J. of Computer 

Science and Information Security (IJCSIS), 7(1), pp. 55– 

62, January 2010. arXiv:1002.1692v1 

[70] E. Folmer, J. van Gurp, and J. Bosch, “Scenario-based 

Assessment of Software Architecture Usability”, in Work. 


on Bridging the Gaps Between Software Engineering and 

Human-Computer Interaction (SE-HCI), IFIP, pp. 61–68, 

May 2003. 

[71] Y. Gao, Import/Export of URN Models in Z.151 XML File 

Format with jUCMNav. M.Sc. project, SITE, University 

of Ottawa, Canada, January 2010. 

[72] N. Genon, D. Amyot, and P. Heymans, “Analysing the 

Cognitive Effectiveness of the UCM Visual Notation”, in 

6th Workshop on System Analysis and Modelling (SAM 

2010), LNCS, Springer, October 2010 (to appear). 

[73] S. Ghanavati, A Compliance Framework for Business 

Processes Based on URN. M.Sc. thesis, SYS, University 

of Ottawa, Canada, May 2007. 

[74] S. Ghanavati, D. Amyot, and L. Peyton, “Towards a 

Framework for Tracking Legal Compliance in 

Healthcare”, in 19th Int. Conf. on Advanced Information 

Systems Engineering (CAiSE'07), LNCS 44495, Springer, 

pp. 218–232, June 2007. doi:10.1007/978-3-540-72988- 

4_16 

[75] S. Ghanavati, D. Amyot, A. Siena, A. Perini, and A. Susi, 

“Towards a Framework for Business Process 

Compliance”, in Int. Workshop on Goal-based Business 

Process Engineering (WGBP 2010), IEEE CS, pp. 330– 

334, October 2010. doi:10.1109/EDOCW.2010.46 

[76] S. Ghanavati, A. Siena, D. Amyot, A. Perini, L. Peyton, 

and A. Susi, “Integrating Business Strategies with 

Requirement Models of Legal Compliance”. Int. J. of 

Electronic Business, Inderscience Publishers, pp. 260– 

280, 2010. doi:10.1504/IJEB.2010.034171 

[77] J. Gordijn and J.M. Akkermans, “Value-based 

Requirements Engineering Exploring Innovative e- 

Commerce Ideas”. Requirements Engineering, 8(2), 

Springer, pp. 114–134, 2003. doi:10.1007/s00766-003- 

0169-x 

[78] J. Gordijn, Value-based Requirements Engineering 

Exploring Innovative e-Commerce Ideas. Ph.D. thesis, 

Vrije Universiteit, Amsterdam, The Netherlands, June 

2002. doi:10.1007/s00766-003-0169-x 

[79] J. Gordijn and J.C. van Vliet, “Integral Design of E- 

Commerce Systems: Aligning the Business with Software 

Architecture through Scenarios”, in ICT-Architecture in 

the BeNeLux (ICT 1999), 1999. 

[80] N. Gorse, The Feature Interaction Problem: Automatic 

Filtering of Incoherences & Generation of Validation Test 

Suites at the Design Stage. M.Sc. thesis, SITE, University 

of Ottawa, Canada, September 2001. 

[81] D. Gross and E.S.K. Yu, “From Non-Functional 

Requirements to Design through Patterns”. Requirements 

Engineering, 6(1), Springer, pp. 18–36, 2001. 

[82] R. Guan, From Requirements to Scenarios through 

Specifications: A Translation Procedure from Use Case 

Maps to LOTOS. M.Sc. thesis, SITE, University of 

Ottawa, Canada, September 2002. 

[83] A. Hamou-Lhadj, E. Braun, D. Amyot, and T. Lethbridge, 

“Recovering Behavioral Design Models from Execution 

Traces”, in 9 th European Conf. on Software Maintenance 

and Reengineering (CSMR), IEEE CS, pp. 112–121, 

March 2005. doi:10.1109/CSMR.2005.46 

[84] J. Hassine, Feature Interaction Filtering and Detection 

with Use Case Maps and LOTOS. M.Sc. thesis, SITE, 

University of Ottawa, Canada, February 2001. 

[85] J. Hassine, Formal Semantics and Verification of Use 

Case Maps. Ph.D. thesis, CSCE dept., Concordia 

University, Canada, April 2008. 

[86] J. Hassine, “AsmL-Based Concurrency Semantic 

Variations for Timed Use Case Maps”, in Abstract State 

Machines, Alloy, B and Z (ABZ 2010), LNCS 5977,


Springer, pp. 34–36, 2010. doi:10.1007/978-3-642-11811- 

1_4 

[87] J. Hassine, “Early Schedulability Analysis with Timed 

Use Case Maps”, in SDL 2009: Design for Motes and 

Mobiles, 14th Int. SDL Forum, LNCS 5719, Springer, pp. 

98–114, Sept. 2009. doi:10.1007/978-3-642-04554-7_7 

[88] J. Hassine, J. Rilling, and R. Dssouli, “An ASM 

Operational Semantics for Use Case Maps”, in 13th IEEE 

Int. Requirement Engineering Conf. (RE05), IEEE CS, pp. 

467–468, September 2005. doi:10.1109/RE.2005.10 

[89] J. Hassine, J. Rilling, and R. Dssouli, “Timed Use Case 

Maps”, in SAM 2006: Language Profiles - Fifth 

Workshop on System Analysis and Modelling, LNCS 

4320, Springer, pp. 99–114, 2006. doi:10.1007/ 

11951148_7 

[90] J. Hassine, J. Rilling and R. Dssouli, “Formal Verification 

of Use Case Maps with Real Time Extensions”, in SDL- 

Forum 2007, LNCS 4745, Springer, pp. 225–241, 2007. 

doi:10.1007/978-3-540-74984-4_14 

[91] J. Hassine, J. Rilling, and R. Dssouli, “Use Case Maps as 

a property specification language”, Software and Systems 

Modeling, 8(2), pp. 205–220, 2009. doi:10.1007/s10270- 

007-0076-6 

[92] J. Hassine, J. Rilling, and R. Dssouli, “An evaluation of 

timed scenario notations”. Journal of Systems and 

Software, 83(2), pp. 326–350, 2010. doi:10.1016/ 

j.jss.2009.09.014 

[93] Y. He, D. Amyot, and A. Williams, “Synthesizing SDL 

from Use Case Maps: An Experiment”, in 11th SDL 

Forum (SDL'03), LNCS 2708, Springer, pp. 117–136, 

July 2003. doi:10.1007/3-540-45075-0_7 

[94] J. Hewitt and J. Rilling, “A Light-Weight Proactive 

Software Change Impact Analysis Using Use Case 

Maps”, in IEEE Int. Workshop on Software Evolvability, 

IEEE CS, pp. 41–46, 2005. doi:10.1109/IWSE.2005.1 

[95] P. Heymans, G. Saval, G. Dallons, and I. Pollet, “Chapter 

VIII: A Template-Based Analysis of GRL”. Advanced 

Topics in Database Research, IGI Publishing, pp. 124– 

147, 2006. doi:10.4018/978-1-59140-935-9.ch008 

[96] J. Hodges and J. Visser, “Accelerating Wireless 

Intelligent Network Standards Through Formal 

Techniques”, in IEEE 1999 Vehicular Technology 

Conference, IEEE CS, pp. 737–742, 1999. 

doi:10.1109/VETEC.1999.778276 

[97] i* Wiki, http://istar.rwth-aachen.de/ (last accessed: July 

12, 2010). 

[98] IBM, Rational DOORS, USA, November 2010. 

http://www.ibm.com/software/awdtools/doors/ 

[99] ISO, Information Processing Systems, Open Systems 

Interconnection, LOTOS — A Formal Description 

Technique Based on the Temporal Ordering of 

Observational Behaviour. IS 8807, 1989. 

[100] ITU-T, Recommendation I.130 (11/88), Method for the 

characterization of telecommunication services supported 

by an ISDN and network capabilities of an ISDN. 

November 1988. 

[101] ITU-T, Recommendation M.3020 (07/07), Management 

interface specification methodology. July 2007. 

[102] ITU-T, Recommendation Z.100 (11/07), Specification and 

Description Language. November 2007. 

[103] ITU-T, Recommendation Z.111 (11/08), Notations to 

Define ITU-T Languages. November 2008. 

[104] ITU-T, Recommendation Z.120 (04/04), Message 

Sequence Chart (MSC). April 2004. 

[105] ITU-T, Recommendation Z.150 (02/03), User 

Requirements Notation (URN) – Language Requirements 

and Framework. February 2003. 


[106] ITU-T, Recommendation Z.151 (11/08), User 

Requirements Notation (URN) – Language definition. 

November 2008. http://www.itu.int/rec/T-REC-Z.151/en 

[107] M. Jackson, Problem Frames: Analyzing and Structuring 

Software Development Problem. Addison-Wesley, 2001. 

[108] S. Jaskó, T. Dulaia, D. Muhia, and K. Tarnaya, “Test 

aspect of requirement specification”. Computer Standards 

& Interfaces, 32(1-2), pp. 1–9, January 2010. 

doi:10.1016/j.csi.2008.12.005 

[109] B. Jiang, Combining Graphical Scenarios with a 

Requirements Management System. M.Sc. thesis, SITE, 

University of Ottawa, Canada, June 2005. 

[110] jUCMNav 4.3, University of Ottawa, Canada, September 

2010. http://jucmnav.softwareengineering.ca/jucmnav 

[111] C. Kaewkasi and W. Rivepiboon, “WWM: a practical 

methodology for Web application modeling”, in 26th 

Annual Int. Computer Software and Applications Conf. 

(COMPSAC 2002), IEEE CS, pp. 603–608, August 2002. 

doi:10.1109/CMPSAC.2002.1045070 

[112] P. Karpati, G. Sindre and A.L. Opdahl, “Visualizing 

Cyber Attacks with Misuse Case Maps”, in 16th Int. 

Working Conf. on Requirements Engineering: Foundation 

for Software Quality (REFSQ 2010), LNCS 6182, 

Springer, pp. 262–275, June 2010. doi:10.1007/978-3- 

642-14192-8_24 

[113] R. Kazman and S.J. Carrière, “Playing Detective: 

Reconstructing Software Architecture from Available 

Evidence”. Automated Software Engineering, 6(2), pp. 

107–138, 1999. doi:10.1023/A:1008781513258 

[114] J. Kealey, Enhanced Use Case Map Analysis and 

Transformation Tooling. M.Sc. thesis, SITE, University 

of Ottawa, Canada, September 2007. 

[115] J. Kealey and D. Amyot, “Enhanced Use Case Map 

Traversal Semantics”, in 13th SDL Forum (SDL'07), 

LNCS 4745, Springer, pp. 133–149, September 2007. 

doi:10.1007/978-3-540-74984-4_9 

[116] J. Kealey, Y. Kim, D. Amyot, and G. Mussbacher, 

“Integrating an Eclipse-Based Scenario Modeling 

Environment with a Requirements Management System”, 

in 2006 IEEE Canadian Conf. on Electrical and 

Computer Engineering (CCECE06), IEEE CS, pp. 2432– 

2435, May 2006. 

[117] J. Kealey, E. Tremblay, J.-P. Daigle, J. McManus, O. 

Clift-Noël, and D. Amyot, “jUCMNav: une nouvelle 

plateforme ouverte pour l'édition et l'analyse de modèles 

UCM”, in 5ième Nouvelles Technologies de la Répartition 

(NOTERE 2005), Gatineau, Canada, pp. 215–222, August 

2005. 

[118] B. Kitchenham, O. Pearl Brereton, D. Budgen, M. Turner, 

J. Bailey, and S. Linkman, “Systematic literature reviews 

in software engineering - A systematic literature review”. 

Inf. Softw. Technol. 51, 1, pp. 7–15, Jan. 2009. 

doi:10.1016/j.infsof.2008.09.009 

[119] C. Kuziemsky, X. Liu, and L. Peyton, “Leveraging Goal 

Models and Performance Indicators to Assess Health Care 

Information Systems”. 7th Int. Conf. on the Quality of 

Information and Communications Technology (QUATIC 

2010), IEEE CS, Porto, Portugal, September 2010. 

doi:10.1109/QUATIC.2010.37 

[120] E. Lavendelis and J. Grundspenkis, “MASITS - A Tool 

for Multi-Agent Based Intelligent Tutoring System 

Development”, in 7th Int. Conf. on Practical Applications 

of Agents and Multi-Agent Systems (PAAMS 2009), 

Advances in Soft. Computing Vol. 55, Springer, pp. 490– 

500, March 2009. doi:10.1007/978-3-642-00487-2_52 

[121] P. Leelaprute, M. Nakamura, K. Matsumoto, and T. 

Kikuno, “Design and Evaluation of Feature Interaction


Filtering with Use Case Maps”. NECTEC Technical 

Journal, 5(16) pp. 581–597, December 2005. 

[122] T. Lethbridge and J. Singer, “Studies of the Work 

Practices of Software Engineers”. Advances in Software 

Engineering: Comprehension, Evaluation and Evolution, 

Springer-Verlag, pp. 51–72, 2002. 

[123] H. Liu, Multilevel Performance Analysis of Scenario 

Specification for a Presence System. M.Sc. thesis, SCE 

Dept., Carleton University, Canada, October 2002. 

[124] L. Liu and E. Yu, “From requirements to architectural 

design—using goals and scenarios”, in ICSE-2001 

Workshop: From Software Requirements to Architectures 

(STRAW 2001), Toronto, Canada, pp.22–30, May 2001. 

[125] L. Liu and E. Yu, “Designing Information Systems in 

Social Context: A Goal and Scenario Modelling 

Approach”. Information Systems, pp. 187–203, April 

2004. doi:10.1016/S0306-4379(03)00052-8 

[126] L. Liu and E. Yu, GRL - Goal-oriented Requirement 

Language, 2000. http://www.cs.toronto.edu/km/GRL/ 

[127] R. Matulevičius, P. Heymans, and A. L. Opdahl, 

“Comparing GRL and KAOS using the UEML 

Approach”. Enterprise Interoperability II: New 

Challenges and Approaches, August 2007, pp. 77-88 

[128] A. Medve, “Advanced steps with standardized languages 

in the re-engineering process”. Computer Standards & 

Interfaces, 30(5), Elsevier, p. 315–322, July 2008. 

doi:10.1016/j.csi.2007.09.004 

[129] A. Miga, Application of Use Case Maps to System Design 

With Tool Support. Master’s thesis, SCE Dept., Carleton 

University, Canada, October 1998. 

[130] A. Miga, D. Amyot, F. Bordeleau, D. Cameron, and M. 

Woodside, “Deriving Message Sequence Charts from Use 

Case Maps Scenario Specifications”, in Meeting UML - 

Tenth SDL Forum (SDL'01), LNCS 2078, Springer, pp. 

268–287, June 2001. doi:10.1007/3-540-48213-X_17 

[131] F. Molina, J. Pardillo, C. Cachero, and A. Toval, “An 

MDE Modeling Framework for Measurable Goal- 

Oriented Requirements”. Int. J. of Intelligent Systems, 

25(8), Wiley, pp. 757–783, August 2010. doi:10.1002/ 

int.20430 

[132] O. Monkewich, I. Sales, and R. L. Probert, “OSPF 

Efficient LSA Refreshment Function in SDL”, in Tenth 

SDL Forum (SDL'01), LNCS 2078, Springer, June 2001, 

pp. 300–315. doi:10.1007/3-540-48213-X_19 

[133] D.L. Moody, P. Heymans, and R. Matulevičius, “Visual 

syntax does matter: improving the cognitive effectiveness 

of the i* visual notation”. Requirements Engineering, 

15(2), Springer, pp.141–175, 2010. doi:10.1007/s00766- 

010-0100-1 

[134] G. Mussbacher, Aspect-oriented User Requirements 

Notation. Ph.D. thesis, SITE, University of Ottawa, 

Canada, November 2010. 

[135] G. Mussbacher and D. Amyot, “A Collection of Patterns 

for Use Case Maps”, in First Latin American Conf. on 

Pattern Languages of Programming (SugarLoafPLoP), 

UERJ - Série Informática, Special Edition, pp. 57–82, 

June 2002. 

[136] G. Mussbacher and D. Amyot, “Assessing the 

Applicability of Use Case Maps for Business Process and 

Workflow Description”, in 3rd Int. MCeTech Conference 

on eTechnologies, IEEE CS, pp. 219–222, January 2008. 

doi:10.1109/MCETECH.2008.18 

[137] G. Mussbacher and D. Amyot, “Goal and Scenario 

Modeling, Analysis, and Transformation with 

jUCMNav”, in 31st Int. Conf. on Software Engineering 

(ICSE-Companion), ACM, Canada, pp. 431–432, May 

2009. doi:10.1109/ICSE-COMPANION.2009.5071047 


[138] G. Mussbacher, D. Amyot, J. Araújo, and A. Moreira, 

“Modeling Software Product Lines with AoURN”, in 

Early Aspects Workshop @ AOSD08, ACM, March 2008. 

doi:10.1145/1404946.1404948 

[139] G. Mussbacher, D. Amyot, J. Araújo, and A. Moreira, 

“Requirements Modeling with the Aspect-oriented User 

Requirements Notation (AoURN): A Case Study”. 

Transactions on Aspect-Oriented Software Development 

VII, LNCS 6210, Springer, pp. 23–68, 2010. 

doi:10.1007/978-3-642-16086-8_2 

[140] G. Mussbacher, D. Amyot, and M. Weiss, “Formalizing 

Patterns with the User Requirements Notation”. Design 

patterns formalization techniques, IGI Global, pp. 302– 

322, 2007. doi:10.4018/978-1-59904-219-0.ch014 

[141] G. Mussbacher, D. Amyot, and M. Weiss, “Visualizing 

Early Aspects with Use Case Maps”. LNCS Journal on 

Transactions on Aspect-Oriented Software Development, 

LNCS 4620, Springer, p. 105–143, November 2007. 

doi:10.1007/978-3-540-75162-5_5 

[142] G. Mussbacher, D. Amyot, and J. Whittle, “Refactoring- 

Safe Modeling of Aspect-Oriented Scenarios”, in 12th Int. 

Conf. on Model Driven Eng. Languages and Systems 

(MODELS 2009), LNCS 5795, Springer, pp. 286–300, 

October 2009. doi:10.1007/978-3-642-04425-0_21 

[143] G. Mussbacher, J. Whittle, and D. Amyot, “Modeling and 

Detecting Semantic-Based Interactions in Aspect- 

Oriented Scenarios”. Requirements Engineering, 15(2), 

Springer, pp.197-214, 2010. doi:10.1007/s00766-010- 

0098-4 

[144] M. Nakamura, T. Kikuno, J. Hassine, and L. Logrippo, 

“Feature Interaction Filtering with Use Case Maps at 

Requirements Stage”, in Sixth International Workshop on 

Feature Interactions in Telecommunications and Software 

Systems (FIW'00), IOS Press, pp. 163–178, May 2000. 

[145] Object Management Group, BPMN 1.2 Specification, 

formal/2009-01-03, January 2009. 

[146] Object Management Group, UML 2.2 Specification, 

formal/2009-02-04, February 2009. 

[147] Object Management Group, UML Profile for 

Schedulability, Performance and Time, v1.0, formal/03- 

09-01, September 2003. 

[148] Object Management Group, UML Profile for Modeling 

and Analysis of Real-time and Embedded Systems 

(MARTE), v1.0, formal 2009-11-02, November 2009. 

[149] C. Ölvingson, N. Hallberg, T. Timpka, and K. Lindqvist, 

“Requirements Engineering for Inter-Organizational 

Health Information Systems with Functions for Spatial 

Analyses: Modeling a WHO Safe Community Applying 

Use Case Maps”. Methods of Information in Medicine, 

Schattauer Gmb H, 4/2002, pp. 299–304, 2002. 

[150] OpenOME, an open-source requirements engineering tool, 

University of Toronto, Canada, November 2010. 

https://se.cs.toronto.edu/trac/ome 

[151] T. Okamura, M. Nakamura, and H. Igaki, “Finding 

Optimal Energy-Saving Operations in Home Network 

System Based on Effects between Appliances and 

Environment”, in 8th Asia-Pacific Symp. on Information 

and Telecommunication Technologies (APSITT 2010), 

IEEE CS, Kuching, Malaysia, , pp. 1–6, June 2010. 

[152] D.B. Petriu, Layered Software Performance Models 

Constructed from Use Case Map Specifications. M.Eng. 

thesis, SCE Dept., Carleton University, Canada, May 

2001. 

[153] D.B. Petriu, D. Amyot, and C.M. Woodside, “Scenario- 

Based Performance Engineering with UCMNav”, in 11th 

SDL Forum (SDL'03), LNCS 2708, Springer, pp. 18–35, 

July 2003. doi:10.1007/3-540-45075-0


[154] D.B. Petriu and C.M. Woodside, “Software performance 

models from system scenarios”. Performance Evaluation, 

61(1), Elsevier, pp. 65–89, June 2005. doi:10.1016/ 

j.peva.2004.09.005 

[155] D.B. Petriu and C.M. Woodside, “An intermediate 

metamodel with scenarios and resources for generating 

performance models from UML designs”. Software and 

Systems Modeling, 6(2), Springer, pp. 163–184, June 

2007. doi:10.1007/s10270-006-0026-8 

[156] A. Pourshahid, A URN-Based Methodology for Business 

Process Monitoring, M.Sc. thesis, EBT, University of 

Ottawa, Canada, March 2008. 

[157] A. Pourshahid, P. Chen, D. Amyot, A.J. Forster, S. 

Ghanavati, L. Peyton, and M. Weiss, “Business Process 

Management with the User Requirements Notation”. 

Electronic Commerce Research, 9(4), Springer, pp. 269– 

316, December 2009. doi:10.1007/s10660-009-9039-z 

[158] A. Pourshahid, G. Mussbacher, D. Amyot, and M. Weiss, 

“Toward an Aspect-Oriented Framework for Business 

Process Improvement”. Int. J. of Electronic Business, 

8(3), Inderscience Publisers, pp. 233–259, 2010. 

doi:10.1504/IJEB.2010.034170 

[159] A. Pourshahid and T. Tran, “Toward an Effective Trust 

Management System for E-Commerce: Modeling Trust 

Components and Processes Using URN”. Journal of 

Business and Technology (JBT), 2(2), Atlantic Academic 

Press, pp. 37–46, 2007. 

[160] J.-F. Roy, J. Kealey, and D. Amyot, “Towards Integrated 

Tool Support for the User Requirements Notation”, in 

SAM 2006: Language Profiles - Fifth Workshop on 

System Analysis and Modelling, LNCS 4320, Springer, 

pp. 198–215, May 2006. doi:10.1007/11951148_13 

[161] J.-F. Roy, Requirement Engineering with URN: 

Integrating Goals and Scenarios. M.Sc. thesis, SITE, 

University of Ottawa, Canada, March 2007 

[162] N. Russell, A.H.M. ter Hofstede, W.M.P. van der Aalst, 

and N. Mulyar, Workflow Control-Flow Patterns: A Revised 

View. BPM Center Report BPM-06-22, 2006. 

http://workflowpatterns.com/ 

[163] Sandrila Ltd., Sandrila SDL. UK, November 2010. 

http://www.sandrila.co.uk 

[164] H. Saiedian, P. Kumarakulasingam, and M. Anan, 

“Scenario-Based Requirements Analysis Techniques for 

Real-Time Software Systems: A Comparative 

Evaluation”. Requirements Engineering, 10(1), Springer, 

pp. 22–33, January 2005. doi:10.1007/s00766-004-0192-6 

[165] K. Saleh and A. Al-Zarouni, “Capturing Non-Functional 

Software Requirements using the User Requirements 

Notation”, in 2004 Int. Research Conf. on Innovation in 

Information Technology (IIT'04), Dubai, pp. 222–230, 

October 2004. 

[166] K. Saleh and G. Elshahry, “Modeling Security 

Requirements for Trustworthy Systems”. Encyclopedia of 

Information Science and Technology, 2nd edition, IGI 

Global, pp. 2657–2664, 2009. doi:10.4018/978-1-60566- 

026-4.ch424 

[167] I. Sales, A Bridging Methodology for Internet Protocols 

Standards Development. M.Sc. thesis, SITE, University of 

Ottawa, Canada, August 2001. 

[168] I.S. Sales and R.L. Probert, “From High-Level Behaviour 

to High-Level Design: Use Case Maps to Specification 

and Description Language”, in 18th Brazilian Symp. on 

Computer Networks (SBRC2000), Brazil, May 2000. 

[169] W.C. Scratchley, Evaluation and Diagnosis of 

Concurrency Architectures. Ph.D. thesis, SCE Dept., 

Carleton University, Canada, July 2000. 


[170] W.C. Scratchley and C.M. Woodside, “Evaluating 

Concurrency Options in Software Specifications”, in 

Seventh Int. Symp. on Modelling, Analysis and Simulation 

of Computer and Telecom. Systems (MASCOTS'99), 

College Park, USA, pp. 330–338, October 1999. 

doi:10.1109/MASCOT.1999.805071 

[171] S. Schneider, The B-Method: An Introduction, Palgrave, 

Cornerstones of Computing series, 2001. 

[172] A. Shamsaei, A. Pourshahid, and D. Amyot, “Business 

Process Compliance Tracking Using Key Performance 

Indicators”, in 6th Int. Workshop on Business Process 

Design (BPD 2010), LNBIP 66, Springer, pp. 73–84, 

September 2010. 

[173] M. Shiri, Supporting UCM Requirements Evolution by 

Means of Formal Concept Analysis. M.Sc. thesis, CSCE 

dept., Concordia University, Canada, February 2008. 

[174] M. Shiri, J. Hassine, and J. Rilling, “Feature Interaction 

Analysis: A Maintenance Perspective”, in 22nd 

IEEE/ACM Int. Conf. on Automated Software Engineering 

(ASE), ACM Press, pp. 437–440, November 2007. 

doi:10.1145/1321631.1321703 

[175] K.H. Siddiqui and C. M. Woodside, “Performance Aware 

Software Development (PASD) Using Resource Demand 

Budgets”, in WOSP 2002: Third Int. Work. on Software 

and Performance, ACM Press, pp. 275-285, July 2002. 

doi:10.1145/584369.584412 

[176] G. Smith, The Object-Z Specification Language. 

Advances in Formal Methods Series, Kluwer Academic 

Publishers, 2000. 

[177] Software Performance Research Group, Layered 

Queueing Research Resource Page. Carleton University, 

Canada, November 2010. http://www.layeredqueues.org/ 

[178] A. Soria, J.A. Díaz-Pace, and M.R. Campo, “Tool Support 

for Fault Localization Using Architectural Models”, 13th 

European Conf. on Software Maintenance and 

Reengineering (CSMR), IEEE CS, March 2009, pp. 59– 

68. doi:10.1109/CSMR.2009.42 

[179] C.H. Störmer, Software Quality Attribute Analysis by 

Architecture Reconstruction (SQUA3RE), Ph.D. thesis, 

Vrije Universiteit, The Netherlands, March 2007. 

[180] N.-T. Truong, T.M.T. Tran, V.-K. To, and V.H. Nguyen, 

“Checking the Consistency between UCM and PSM 

Using a Graph-Based Method”, in 1st Asian Conf. on 

Intelligent Information and Database System (ACIIDS 

09), IEEE CS, pp. 190–195, April 2009. 

doi:10.1109/ACIIDS.2009.66 

[181] Use Case Map Navigator (UCMNAV) 2.3, July 2005. 

http://jucmnav.softwareengineering.ca/ucm/bin/view/UC 

M/UcmNav 

[182] URN Virtual Library, http://www.UseCaseMaps.org/pub 

(last accessed: July 27, 2010). 

[183] J.A. van der Poll, P. Kotze, A. Seffah, T. Radhakrishnan, 

and E. Alsumait, “Combining UCMs and Formal Methods 

for Representing and Checking the Validity of Scenarios 

as User Requirements”, in 2003 annual research conf. of 

the South African Institute of Computer Scientists and 

Information Technologists (SAICSIT 2003), 

Johannesburg, South Africa, pp. 59–68, September 2003. 

[184] B. van der Raadt, Business-Oriented Exploration of Web 

Services Ideas - Combining Goal-Oriented and Value- 

Based Approaches. MSc. Thesis, Vrije Universiteit, 

Amsterdam, The Netherlands, February 2005. 

[185] A. van Lamsweerde, Requirements engineering: From 

System Goals to UML Models to Software Specifications. 

John Wiley & Sons, 2009.


[186] M. Vigder, Applying Formal Techniques to the Design of 

Concurrent Systems. Ph.D. thesis, SCE Dept., Carleton 

University, Canada, July 1992. 

[187] M. Vinje, An Auditing Framework for Service Provision 

in Mobile IPv6 Networks. Diploma Thesis, EPF Zurich, 

Switzerland, August 2002. 

[188] M. Weiss, “Pattern-Driven Design of Agent Systems: 

Approach and Case Study”, in 15th Conf. on Advanced 

Information Systems Engineering (CAiSE'03), LNCS 

2681, Springer, pp. 711–723, June 2003. doi:10.1007/3- 

540-45017-3 

[189] M. Weiss, “More Patterns for Web Applications”, in 

Tenth European Conf. on Pattern Languages of Programs 

(Euro PLoP 2005), Irsee, Germany, pp. 21–34, July 2005. 

[190] M. Weiss and D. Amyot, “Business process modeling 

with URN,” Int. J. of E-Business Research, 1(3), pp. 63– 

90, 2005. doi:10.4018/jebr.2005070104 

[191] M. Weiss and B. Esfandiari, “On Feature Interactions 

among Web Services”. Int. J. of Web Services Research, 

2(4) pp. 22–47, October 2005. doi:10.4018/jwsr. 

2005100102 

[192] M. Weiss, B. Esfandiari, and Y. Luo, “Towards a 

classification of Web service feature interactions”. 

Computer Networks, 51(2), Elsevier, pp. 359–381, 

February 2007. doi:10.1016/j.comnet.2006.08.003 

[193] M. Weiss and H. Mouratidis, “Selecting Security Patterns 

that Fulfill Security Requirements”, in 16th IEEE Int. 

Requirements Engineering Conf. (RE'08), IEEE CS, pp. 

169–172, September 2008. doi:10.1109/RE.2008.32 

[194] P. Wu and C.M. Woodside, “An Aggregation Approach to 

Constructing Hybrid Layered Queueing Models”, in 7th 

Int. Workshop on Performability Modeling of Computer 

and Communication Systems (PMCCS7), Torino, Italy, 

September 2005. 

[195] W. Wu and T.P. Kelly, “Managing Architectural Design 

Decisions for Safety-Critical Software Systems”, in 2nd 

Int. Conf. on the Quality of Software Architectures (QoSA 

2006), LNCS 4126, Springer, pp. 59–77, June 2006. 

doi:10.1007/11921998_9 

[196] Z. Yi, CNAP Specification and Validation: A Design 

Methodology Using LOTOS and UCM. M.Sc. thesis, 

SITE, University of Ottawa, Canada, January 2000. 

[197] E.S.-K. Yu, Modelling strategic relationships for process 

reengineering. Ph.D. thesis, Dept. of Computer Science, 

University of Toronto, Canada, 1995. 

[198] E. Yu, Y. Yu, and L. Liu, OME — Organization 

Modelling Environment, University of Toronto, 2000. 

http://www.cs.toronto.edu/km/ome/ 

[199] X.Y. Zeng, Transforming Use Case Maps to the Core 

Scenario Model Representation. M.Sc. thesis, SITE, 

University of Ottawa, Canada, June 2005 

[200] R. Zhang and X.-X. Liu, “Feature Interaction Filtering 

Method Based on URN (基于 URN 的特征冲突过滤方法)”, 

Computer Engineering (计算机工程), 35(21), pp. 45–47, 



Daniel Amyot received both his Ph.D. 

(2001) and M.Sc. (1994) degrees in 

computer science from the University of 

Ottawa. The research topic was related 

to the specification and validation of 

telecommunication systems with Use 

Case Maps and LOTOS. 

After working for Mitel Networks as 

a senior researcher, he joined the School 

of Information Technology and Engineering of the University of 

Ottawa, where he is now Associate Professor in software 

engineering. His research interests include goal-oriented and 

scenario-based software engineering, requirements engineering, 

business process modeling, aspect-oriented modeling, and 

healthcare informatics. He has published over 90 papers in 

various conferences and in journals such as Requirements 

Engineering, Computer Networks, and the International Journal 

of Electronic Business. 

Dr. Amyot is a member of ACM, IEEE Computer Society, 

and APIIQ, and he is a professional engineer in the province of 

Québec (Canada). He is also Associate Rapporteur for 

requirements languages at the International Telecommunication 

Union, where he leads the evolution of the User Requirements 

Notation. 

Gunter Mussbacher received a M.Sc. 

degree in computer science from Simon 

Fraser University in 1999, and a Ph.D. in 

computer science from the University of 

Ottawa in 2010. In his thesis, he 

developed the Aspect-oriented User 

Requirements Notation (AoURN), a 

framework that enables goal-oriented, 

scenario-based, and aspect-oriented 

modeling in a unified way. 

After his M.Sc., he worked as a research engineer for the 

Strategic Technology department of Mitel Networks, where he 

applied and taught URN concepts. He has published in the 

Requirements Engineering Journal (REJ) and in the 

Transactions on Aspect-Oriented Software Development 

(TAOSD), and co-edited with Daniel Amyot the URN standard 

(ITU Recommendation Z.151 11/2008). He is also teaching 

software engineering undergraduate courses as well as URN and 

AoURN tutorials for industry and at international conferences. 

His general research interests lie in requirements engineering, 

URN, aspect-oriented modeling, and patterns. 

Dr. Mussbacher is an organizer and program committee 

member of Early Aspects (EA), Aspect-oriented Modeling 

(AOM), and Systems Analysis and Modelling (SAM) 

workshops since 2008.


Stochastic Process Algebra with Value-Passing 

and Weak Time Restrictions 

Guang Zheng 1,2 , Jinzhao Wu 3,4 , and Aiping Lu 2,∗ 

1 Information Science and Engineering School, Lanzhou University, 

Lanzhou 730000, China, Email: forzhengguang@163.com 

2 Institute of Basic Research in Clinical Medicine, China Academy of Chinese Medical Sciences, 

Dongzhimen, Beijing 100700, China 

3 Guangxi University for Nationalities, China Academy of Chinese Medical Sciences, 

Nanning, 530004, China 

4 School of Computer and Information Technology, Beijing Jiaotong University, 

Beijing 100044, China 

∗ Corresponding author: lap64067611@126.com 

Abstract— Process algebra provides essential tools for studying 

distributed and concurrent systems. Stochastic process 

algebra (i.e., YAWN ) enhances the process algebra with 

stochastic extensions which is perfect to analyze phenomena 

of process with executing durations in the real world. What’s 

more, in system runs, value passing is tightly bounded with 

their processes. However, stochastic process algebras lack 

value passing can limit their expressiveness. Based on this, 

we propose a process algebra of stochastic process algebra 

with value passing. This new process algebra can specify 

the behaviors of systems in a more clear and accurate way. 

In dealing with relationship of bisimulations, we introduce 

a new policy of weak time comparison between processes 

in bisimulation which is more convenient and doable in 

practice. 

Index Terms— stochastic process algebra, value passing, 

equivalence, bisimulation, weak time restriction. 


Process algebra is a widely accepted language of specifying 

distributed and concurrent systems. The fundamental 

work is done by Milner in CCS [23], Hoare in CSP 

[13] and Hilston in ACP [2]. 

Stochastic Process Algebras (SPAs) [3], [6], [7], [16], 

[18] have been invented in the early 90’s, the main idea of 

stochastic process algebras is to incorporate quantitative 

information in a qualitative process algebra model. In 

these approaches proposed so far, the quantitative information 

is given in terms of distribution functions or 

random variables. These variables denote the duration of 

actions, and these durations are specified together with 

actions. 

The basic activity of a process is action. In stochastic 

process algebras, actions are equipped with stochastic 

distribution functions which describes the execution time 

of the actions stochastically. SPAs are suitable to describe 

This work was partially supported by the National Eleventh Five Year 

Support Project of China (2006BAI04A10), the Innovative Methodology 

Project supported by MOST of China(2008IM020900, 2009ZX09502- 

019). National Science Foundation of China (No. 30902003, 30973975, 

90709007, and 81072982). 


doi:10.4304/jsw.6.5.769-782 

functional as well as stochastic behaviors in one single 

specification. 

The actions in SPA give the framework specification 

of system’s behaviors. During the execution of SPAs, 

value passing occurs intuitively and naturally. Processes 

cooperate with each other by exchanging messages [10], 

[11], [27], it can be happened in typical operators like: 

• Sequential composition, where the prefix action 

might pass the value to the following action for further 

execution. This happens commonly in programs; 

• Parallel composition, where a synchronous communication 

event can be executed. This is the type of 

communication in SPAs; 

• Recursive operation, this operator is useful in dealing 

with repeated actions with certain rules. 

SPAs can be models for describing phenomena of the 

real world in an abstract level. Value passing can enhance 

SPAs with the abilities to describing phenomena in a 

more detailed level, more intuitive to understand and more 

doable to put into use. This can be demonstrated by the 

above examples. In other systems, (for example, traffic 

control, weather forecasting, scientific calculation, and 

stock markets), there are values passing with all processes 

running all time long. Value passing exists at any moment 

in system runs. With value passing, we can get a inner 

sight into the phenomena (e.g., under certain situation, 

values are final results). So, it is necessary and intuitively 

for us to equip the language of SPAs with value passing. 

By doing this, we can get a better understanding of the 

phenomena described in SPAs. 

The main idea of SPA with value passing is to enhance 

actions with notions of duration and value passing during 

their executions. Durations are described stochastically by 

means of distribution functions. Values passing during 

executions can simulate the key parameters in system 

runs. In Markovian SPA, only exponential distributions 

are considered as delay parameters. As for value passing, 

we assume that only valid values are permitted during the 

execution, and the invalid values will trigger an exception.


An exponential distributions is λ and its mean value is 

1/λ. 

One of the most attractive features of process algebras 

is their compositional nature. But it is not the only one, 

another important aspect of the formalism is the definition 

of equivalence relations i.e., strong and weak bisimulation 

[22], [23]. These equivalence relations can be used to 

compare agents (model verification) and to replace one 

agent by another which exhibits an equivalent behavior 

but has a simpler representation (model simplification). 

Such notions of equivalence are considered part of the 

semantics of the language, and therefore their definition 

is an integral part of its development. 

One important class of equivalence relations in process 

algebras are bisimulations. Most SPAs [3], [6], [7], [16], 

[18] provide bisimulation relations both on action and 

time. They are characterized by the exponential distributions, 

i.e., λ, µ and so on. Bisimulations of this kind 

is an extension of the classic bisimulations (strong and 

weak). However, in practice, even weak bisimulation is 

too strict. In real world, under certain situations, we might 

be more tolerant in the execution time under bisimulation 

relations. Based on this, we propose a weak time restriction 

bisimulation called time restricted bisimulation. 

This bisimulation relationship is more tolerant in time 

restrictions when comparing two processes with criterion 

of bisimulations. We will prove that this time restricted 

bisimulation relationship can be preserved over all the 

operators in the language of SPA. 

Another use of equivalence relations is over the states 

within a model. When a set of states are found to 

have equivalent behaviors, we can analyze them by these 

relationships to partition the state space and considering 

only one representative of relation to partition these states. 

Then, only compare one representative of each partition 

(model aggregation). This is an important way in state 

space reduction. 

This paper is organized as follows. Section 2 introduces 

the language of YAWN with value passing, including 

syntax and its meanings. Section 3 introduces generalized 

Markovian transition systems with value passing which 

can be used as models to express the semantics of the 

language YAWN with value passing. Section 4 shows 

operational semantics of the languageYAWN with value 

passing. Section 5 introduces the axioms of operators 

in YAWN with value passing. Section 6 shows some 

equivalence relations of the language YAWN with value 

passing, including strong bisimulation, weak bisimulation, 

expansion law and time restricted bisimulations. Section 

7 concludes the paper. Section 8 lists the acknowledgements. 

II. LANGUAGE OF YAWN WITH VALUE 

Now, we define the language in the style of YAMN 

with value passing. We first define the set L of all process 

algebra expressions with value passing. An expression 

P ∈ L is said to be closed if and only if every process 

variable, say X, occurring in P occurs within the scope 


of the recX operator, and if every process constant is 

defined by a defining equation. 

Definition 2.1 (LYAMN ) Let L be the language with 

value passing defined by the following grammar: 

P :: = 0 � � X � � A � � av.P � � [λ].P � � if b then P � � P;P � � 

P +P � � recX : P � � P \H � � P||SP 

av :: = i � � c?x � � c!e 

We use av to stand for the generalized form of actions 

with value in situation no more specification is needed. i 

is un-observable actions which likes the τ in CCS; c?x 

for input action with value x on channel c; and c!e stands 

for output action with value e on channel c. When it is 

necessary, we will use i, c?x and c!e to specify actions 

under different situations. 

0 is an empty process which cannot perform any 

actions. It can also be taken as “STOP” in some literature. 

av.P is action prefixing. After executing value passing 

action av, the process av.P will behave as P . 

[λ].P is prefix delay. This term means there is a 

time delay before the execution of process P . The time 

is characterized by the stochastic variable λ. λ is an 

exponential distribution parameter, and the mean time of 

λ is 1/λ. We use t = 1/λ as the label to stand for the 

time transition, then we have transition in the form of 

[λ].P i −→ P . 

P;Q is the sequential composition of two processes P 

and Q. After the execution of process P , the system P;Q 

behaves as Q. 

P + Q is the choice composition of two processes P 

andQ. If processP is selected for execution, then process 

Q is dropped and have no chance for further execution. 

A is used to express CONST . We use CONST to 

express process constants. A constant C ∈ CONST is 

assigned a process with value by means of a defining 

equation C def 

= av.C ′ . The defining equation A def 

= av.A 

is an example. Intuitively,A is supposed to be the process 

that can execute infinite number of action a with value v. 

recX : P stands for recursive expression of processes. 

With the sequential and choice operators, only finite 

behavior can be described. As for some reactive systems 

that generally never terminate, there should be a way to 

describe them. recX : P is selected to stands for it. 

In the above example, recX : P behaves as P[recX : 

P/X], where P[recX : P/X] is the process term where 

simultaneously all occurrences ofX inP are syntactically 

replaced by recX : P . 

P||SQ is parallel composition of processP and process 

Q. Actions in P or Q which are not in set S can 

be executed independently at the same time without 

synchronization. However, actions of P or Q that are in 

set S can only be executed by synchronization. 

Example 2.2 Consider the processes of P def 

= 

ava.bvb.cvc.0 and Q def 

= dvd.bvb.eve.0, we know that 

R def 

= P||bQ denotes a process in which bothP andQcan 

perform the actions ava and dvd independently. What’s 

more, action bvb can only be proceed in the synchronize


way. After the synchronization, P and Q can proceed 

again independently, i.e., they can perform actions of cvc 

and eve respectively. 

Example 2.3 If we consider the process (P||bS)||bR, 

where S def 

= R def 

= Q(Q in Example 2.2), then, all 

three processes can start independently at the same time. 

However, P , S, and R can only take part in the synchronization 

over bvb before they can execute their respective 

last action. 

P \H is hiding operator. The purpose of this operator 

is to mark the scope of actions which should never 

again take part in synchronization. To do this, a special 

action is introduced, which is often denoted as τ or i: 

the internal action. Reconsider Example 2.3, we can see 

that process R could be inhibited from participating in 

the synchronization over bvb that P and S are already 

involved in. The effect of the hiding operator is that all 

actions in H are hidden away: they are no longer visible 

from outside. Then, a process which is synchronized by 

P and S can be executed, and R proceeds independently 

from both can be expressed as (P||bS)\{b}|| {∅}R. 

if b then P is the one-armed condition if b then in the 

language. With the help of+, the conventional two-armed 

if b then else expression can be defined by 

if b then P else Q = if b then P + if ¬b then Q 

In what follows, we will use the if b then else 

construction freely without further comments. 

We assume that the operators have the following precedence: 

prefix > recursion > hiding > choice > parallel 

composition, i.e., prefix has precedence over recursion, 

recursion over hiding, etc. Parentheses can be used to circumvent 

these rules. If we have more than two processes 

combined (i.e., for example, in P+Q+R or P||SQ||S ′R 

for P,Q,R ∈ LYAWN ) then we assume a left-associative 

evaluation order: P +Q+R and P||SQ||S ′R are assume 

to be equal to (P+Q)+R and(P||SQ)||S ′R respectively. 

These rules determine a unique evaluation order, which 

later will become especially important for the application 

of SOS rules. 

Please note that the YAMN language comes with 

bells and whistles: we allow to define recursion by means 

of process constants, and by recX operators with process 

variables. The only reason for this is to have a more 

convenient syntax for YAWN . 

Frequently, we have to compare elements of the 

YAWN language with value passing syntactically. For 

two terms P,Q ∈ LYAWN , we define P ≡ Q if and 

only if P and Q are syntactically equal for the value 

assignment for all executing actions with value passing 

of the same equivalent class leading to the result also of 

the same equivalent class. 

III. OPERATIONAL SEMANTICS 

In classic process algebras, Labeled Transition System 

(LTS) is used to demonstrate the operational semantics of 

the language. In this section, we will give out the definition 

of transition systems with value passing that will 


demonstrate the operational semantics for the LYAWN 

with value. 

A. Transition Systems with Value 

The semantics of YAMN processes is given in terms 

of transition systems. So, in order to introduce the operational 

semantics of the language, we introduce the 

generalized Markovian transition systems. 

Definition 3.1 A generalized Markovian transition system 

with value (GMTSV ) is a tuple(S,AV,T,R), where 

• S is a set of states; 

• AV is a set of labels with value; 

• T ⊆ S ×AV ×S is a set of labeled transitions; 

• R : T → R∪{∞} 

Typical elements of S are s,s ′ ,s ′′ ,s1,s2,···, and 

typical elements of T are t,t ′ ,t ′′ ,t1,t2,···. Transitions 

labeled with t are meant to be exponentially distributed 

time delays. The function R specifies the rates of the 

distributions. A GMTS is said to be properly timed, if 

whenever t ∈ T with t = (s,a,s ′ ) and a ∈ Act (i.e., 

∀a,a �= t), then R(t) = ∞. Hence, all internal or visible 

actions are considered to have no durations, which is 

expressed by assigning them infinite rates. 

Definition 3.2 We define a GMTS with value passing 

as (S,AV,T,R) together with a state s ∈ S (starting 

state) a generalized Markovian process (GMP). We denote 

a GMP by a five-tuple (S,AV,T,R,s) where AV is the 

label of transition with value. 

B. Operational semantics 

Based on the language discussed in the previous section 

and the informal explanation of the syntax, we know 

that our language can describe the behavior of systems 

with stochastic actions with value passing. Through the 

Markovian Transition System (MTS), we know that it is 

convenient to express the semantics of the behavior of 

such systems. Now, it is ready for us to give out the 

formalize rules of the language LYAWN with operators 

described in the previous section in table III-B. 

Rule (1) expresses the action prefix. Process term av.P 

executes action av first, then behave as P . 

Rule (2) expresses the delay prefix. Process term [λ].P 

delays time t and then executes as P . As we restrict the 

distribution of λ as exponential distribution, it is clear that 

the mean time of the delay t is 1/λ. 

Rule (3a) and (3c) express the choice composition 

between two processes with actions influenced by the 

environment. 

Rule (3b) and (3d) express the choice of two delays. 

Processes P and Q have delays characterized by µ and 

ν respectively. We do not compare µ and ν, and we 

know that only if the delay reaches 1/µ, and the process 

P will continue its execution. Similar, when the delay 

reaches1/ν, processQwill continue, we will have further 

explanations later by example.


(1) 

(3a) 

(3c) 

(3e) 

(4a) 

(4c) 

(4e) 

(4g) 

(4i) 

(5a) 

(5c) 

(6a) 

(7a) 

av.P av 

−−→ P 

P av 

−−→ P ′ 

P +Q av 

−−→ P ′ 

Q av 

−−→ Q ′ 

P +Q av 

−−→ Q ′ 

P a −→ P ′ , Q [λ] 

−−→ Q ′ 

P +Q av 

−−→ P ′ 

P av 

−−→ P ′ 

P||SQ av 

−−→ P ′ ||SQ 

Q av 

−−→ Q ′ 

(2) 

(3b) 

(3d) 

(3f) 

av �∈ S (4b) 

P||SQ av 

av �∈ S (4d) 

−−→ P||SQ ′ 

P av 

−−→ P ′ , Q av 

−−→ Q ′ 

P||SQ i −→ P ′ ||SQ ′ 

P av 

−−→ P ′ , Q [λ] 

−−→ Q ′ 

P||SQ av 

−−→ P ′ ||SQ 

av 

P −−→ P ′ , Q [λ] 

−−→ Q ′ 

P||SQ t −→ P||SQ ′ 

P av 

−−→ P ′ 

P \H av 

−−→ P ′ \H 

P av 

−−→ P ′ 

P \H i −→ P ′ \H 

P{recX : P/X} av 

−−→ P ′ 

( av ∈ S, 

recX : P av 

−−→ P ′ 

av ∈ S (4f) 

av �∈ S (4h) 

t = 1 ) (4j) 

λ 

av �∈ H (5b) 

a ∈ H (5d) 

(6b) 

P av 

−−→ P ′ 

A av 

def 

A = P (7b) 

−−→ P ′ 

Rule (3e) and (3f) express the choice between action 

and delay proposed by two processes. Under the assumption 

of maximal execution, we propose this rule to execute 

action and left the delay alone. 

Rule (4a) and (4c) express the execution of parallel 

composition of processes where the executing action is 

not within the scope of synchronization. Under this situation, 

the executing process just continues its execution 

and the other processes just waiting for their turns. 

Rule (4b) and (4d) express how the paralleled processes 

trait their delays: each process waits for its time to end 

the delay and continue its further executions. 

Rule (4g) and (4h) express how the paralleled processes 

with action (no synchronization with others) and delay 

trait their behavior. The system executes the action, while 


[λ].P t −→ P 

t = 1/λ 

P [µ] 

−−→ P ′ , Q [ν] 

−−→ Q ′ 

P +Q t −→ P ′ 

P [µ] 

−−→ P ′ , Q [ν] 

−−→ Q ′ 

P +Q t −→ Q ′ 

P [λ] 

−−→ P ′ , Q av 

−−→ Q ′ 

P +Q av 

−−→ Q ′ 

P [ν] 

−−→ P ′ , Q [µ] 

−−→ Q ′ 

P||SQ t −→ P ′ ||SQ 

P [ν] 

−−→ P ′ , Q [µ] 

−−→ Q ′ 

P||SQ t −→ P||SQ ′ 

P [ν] 

−−→ P ′ , Q [µ] 

−−→ Q ′ 

P||SQ t −→ P||SQ ′ 

t = 1 

µ 

t = 1 

ν 

t = 1 

ν 

t = 1 

µ 

t = f(ν,µ) 

P [λ] 

−−→ P ′ , Q av 

−−→ Q ′ 

P||SQ av 

av �∈ S 

−−→ P||SQ ′ 

P [λ] 

−−→ P ′ , Q av 

−−→ Q ′ 

P||SQ t −→ P ′ ||SQ 

P [λ] 

−−→ P ′ 

P \H t −→ P ′ \H 

av ∈ S,t = 1 

λ 

t = 1 

λ 

P [µ] 

−−→ P ′ , P ′ av 

−−→ P ′′ ,P ′′ 

[ν] 

−−→ P ′′′ 

P t −→ P ′′′ 

P{recX : P/X} [λ] 

−−→ P ′ 

recX : P t −→ P ′ 

P [λ] 

−−→ P ′ 

A t −→ P ′ 

TABLE I. 

OPERATIONAL SEMANTICS OF YAWNV 

� 

a ∈ H, 

t = 1 

µ + 1 

� 

ν 

t = 1 

λ 

A def 

= P, t = 1 

λ 

waiting for the delay at the same time. 

Rule (4i) and (4j) express how the paralleled processes 

with action (synchronize with other process) and delay 

trait their behavior. The system can not execute the 

synchronization, it just wait for the end of delay if 

no synchronization available. Then continue its further 

executions. 

Rule (5a) express how a process with hiding actions 

executes un-hidden actions, which is intuitive and do not 

need further explanation. 

Rule (5b) express how a process with hiding actions 

traits delay: it just waits to the end of the delay and then 

continues its executions. 

Rule (5c) express how a process dealing with hiding 

actions. It just executes the action, however, the execution


cannot be observed from outside. So, according to the 

definition, we name the action as i. 

Rule (5d) tells us the delay of execution of a hiding 

action. There can be a sum of two delays of the hiding 

action: before µ and after ν. So, we have the result t = 

1 

µ + 1 

ν . 

Rule (6a) and (6b) express how the recursive terms 

traits their actions and delays. It is rather intuitive base 

on explanations of the rules above. 

Rule (7a) and (7b) express how a process term assigns 

to a constant A. They are also intuitive and easy to 

understand. 

Some literatures trait action in SPAs with duration 

in the form of ([λ].av), which is rather intuitive in the 

understanding of the execution. We separate them in 

our language of YAWNV as action av which do not 

have durations, and delay [λ] which characterize the time 

between two actions. There is no difference in the essence 

of the two kinds of expressions. The latter form is more 

flexible and compact, so we adapt it here. 

IV. AXIOMS 

In this section, we propose the axioms of operators 

in the SPA language with value passing. It is based on 

the study of operational semantics and the equivalences 

relations as strong bisimulation and weak bisimulation. 

A. Value 

As data play an important role in the language of 

YAWN with value passing, axiomatization for such 

process operators must involve dealing with data domain. 

However, it turns out that, we can factor out data reasoning 

from process reasoning by employing conditional 

equations [10] of the form 

b⊲P = Q 

where P and Q are process terms and b is a boolean 

expression representing the condition on the data domain 

under which P and Q are equal. An example of a proof 

rule is: 

b ′ ∧b⊲P = Q, b ′ ∧¬b⊲0 = Q 

b ′ ⊲ if b then P = Q 

It captures the intuitive meaning of the conditional construct: 

if b then P behaves like P when b is true, and 

like 0 otherwise. In this rule, all we need to know about is 

construct if then when manipulating syntactical terms. 

From a “goal-directed” point of view, it moves the parts 

involving data (b) from the process term (if b then P) to 

the conditional guard part. Such conditions can be used 

to discharge constructs involving data when some other 

inference rules are applied. 

Reasoning about YAWN with value passing will 

inevitably involve the reasoning about data. However, 

instead of inventing rules for all possible data domains, 

we would like to factor out reasoning about data from 

reasoning about processes as much as possible. Therefore 


our proof system will be parameterized over data reasoning 

of the form b |= b ′ , with the intuitive meaning that 

whenever b is true then so is b ′ . 

Now, we present the axioms of data in the language of 

LYAWN in Table II: 

This set of inference rules we put forward in Table II 

can be taken as a natural generalization of pure equational 

reasoning. For each construct in our language, there is a 

corresponding introduction rule with a set of axioms. 

In this paper, we introduce value passing into the language 

of LYAWN . We try to focus on the core meaning 

of the value passing and not of the kind and quantity of 

the value. 

Example 4.1 There are two testing systems guarded 

by scores. When the score is greater than 60, the system 

s1 would respond message “PASS”. For system s2 with 

the same value, it shows color “GREEN” as a respond. 

Both the systems obey the same rules, and output results 

with different kinds of values, and this is very popular in 

scoring systems in the real world. We treat them as equal 

in our language for they obey the same rules which can 

be described as 

b |= rules1 = rules2 = if score ≥ 60 then true 

and the result is also of the same equal class that can be 

described as 

b ′ � 

trues1 = PASS 

|= 

trues2 = GREEN 

We omit the input action in the design of the systems 

designed above. System s1 and s2 are simple, they 

can deal with value satisfying condition if score ≥ 

60 then true. These systems ignore other score and 

respond nothing according to the condition b. 

B. Sequential Composition 

Essentially, axioms of prefix and sequential composition 

are of the same class. They are all sequential 

operators, and they obey rules (1) and (2). 

(S1) P.0 = P 

(S2) P.0.Q = P 

(S3) (P.Q).R = P.(Q.R) 

Axioms of Sequential Composition:AS 

A process will STOP when it encounters with 0. So, the 

execution of process will stop at the point of 0, and left 

the other actions aside. Thus, we know that S1 and S2 

are right. As to S3, it is rather intuitive to understand, for 

the parenthesis do not influence the execution sequences 

of P.Q.R. 

C. Choice Composition 

In this section, we present the axioms of choice composition. 

They are based on the rules of (3a), (3b), (3c),


and (3d). 

(D1) α−Conversion 

c?x.P = c?y.Q[y/x] 

(D2) Premise true⊲P=Q 

(D3) Input 

(D4) Output 

(D5) Choice 

(D6) Partition 

(D7) Condition 

(D7) Parallel 

(D8) Hiding 

(C1) α.P +β.Q = β.Q+α.P 

(C2) α.P +[λ].Q = α.P 

(C3) [µ].P +[ν].P = 1/(µ+ν).P 

(C4) P +(Q+R) = (P +Q)+R 

(C5) P +0 = P 

Axioms of Choice Composition:AC 

The axioms of choice composition deal with actions 

and delays separately. C1 shows that the exchange of 

position in choice composition does not affect the execution. 

C1 left the choice for the outside environment. 

C2 shows that the execution policy of maximal processes 

during the execution of choice composition: the system 

does not wait if there is an action ready for execution.C3 

shows that when there is a choice between two identical 

processes with different (exponential) delays, the system 

would delay as the sum of the two stochastic variables. 

C4 shows that the choice composition among processes 

with parenthesis does not affect the executing policy of 

choices. C5 shows that the system would select P under 

the situation of P + 0, which means the system can do 

nothing but P . This is intuitive, for 0 means STOP of 

the execution. If P +0 = 0 means the system is out of 

control, and STOPs at wrong point. 

D. Internal Action i 

We present axioms of internal action (i.e., τ in classic 

process algebra). They are based on the rules of (5c) and 


P = Q 

b⊲P = Q 

b⊲c?x.P = c?x.Q 

b |= e = e ′ , b⊲P = Q 

b⊲c!e.P = c!e ′ .Q 

b⊲P = Q 

b⊲P +R = Q+R 

y �∈ fv(t) 

b |= b1 ∨b2, b1 ⊲P = Q, b2 ⊲P = Q 

b⊲P = Q 

b ′ ∧b⊲P = Q, b ′ ∧¬b⊲0 = Q 

b ′ ⊲ if b then P = Q 

b⊲P = Q 

b⊲P||SR = Q||SR 

b⊲P = Q 

b⊲P \H = Q\H 

TABLE II. 

AXIOMS OF VALUE UNDER CONDITIONAD 

(5d). 

(I1) α.i.P = α.P 

(I2) P +i.P = i.P 

(I3) α.(P +i.Q)+i.Q = α.(P +i.Q) 

Axioms of Internal Action:Ai 

The axioms of internal actions are designed for the 

observable equivalences. When the system is executing 

an internal action, the action being executed cannot be 

observed from outside. This is what I1 means. I2 and I3 

are rather intuitive: as we cannot tell if there are internal 

actions being executed, we assume there are internal 

executions in system runs. 

E. Parallel Composition 

We present axioms of parallel composition here. They 

are based on the rules of (4a), (4b), (4c), (4d), (4e), and 

(4f). 

(P1) P||S0 = P 

(P2) P||SQ = Q||SP 

(P3) (P||SQ)||TR = P||S(Q||TR) 

Axioms of Parallel Composition:AP 

Axiom P1 means the same as C5 ( P+0 = P ): when 

a process P is paralleled with an empty process, it just 

executes as P . P2 shows that the execution of paralleled 

processes do not care about the position under parallel 

composition. That is, communication under parallel composition 

is preserved. P3 shows that the parenthesis of 

paralleled processes do not affect the execution when it 

is paralleled with other processes.


F. Hiding Operation 

We present axioms of hiding operation here. They are 

based on the rules of (5a), (5b), (5c), and (5d). 

(H1) P \L = P if Act(P) �∈ L 

(H2) P \K \L = P \(K ∪L) 

(H3) (P||SQ)\L = P||S∪LQ 

(H4) (P +Q)\L = P \L+Q\L 

Axioms of Hiding Operation:AH 

Hiding operator is useful in the software engineering. 

We can take single programs as processes, and the 

composition of all associated programs so as to form 

a system which can complete designed functions. The 

program/process with certain sub-functions usually with 

input, output, and some other kind of control. When they 

are compiled into one system (sometime one executive 

file), most of the programs’ input, output, and control are 

transformed into internal communications which cannot 

be observed outside. 

H1 means process P with hiding action set L which 

contains no action during the execution ofP , so the hiding 

operator cannot affects the P . H2 means the function 

of composition of more than one sets of hiding actions 

into one hiding action set. H3 means that hiding set in 

parallel composition can be added to the synchronization 

set. In this rule, the hiding action set is the set that process 

P and Q will synchronized. H4 means that the hiding 

operator can distribute through the choice composition of 

processes. 

G. Recursive Operation 

We present axioms of recursive operator here, which 

are based on the rules of (6a), (6b), (7a), and (7b). 

(R1) rec(X : P) = P{recX : P/X} 

(R2) If P = E{P/X} then P = recX : E/X 

(R3) recX : (X = X +P) = recX : P 

(R4) recX : (X = i.X +P) = recX : (i.P) 

(R5) recX : (X = i.(X +P)+Q = 

recX : (i.X +P +Q) 

Axioms of Recursive operator:AR 

R1 is rather intuitive, the unwind of rec(X : P) 

is the substitution of variable X with P such form 

the recursive expression. R2 shows how to define the 

recursive expression. R3, R4, and R5 are rather intuitive 

based on the understanding of the rules. 

V. EQUIVALENT RELATIONS 

From the very beginning, an essential part of process 

algebra theory has been devoted to the development 

of equivalence notions. The starting point of all 

process algebraic equivalences is the observation that 

different processes may exhibit the same behavior. R.J. 

van Glabbeek has extensively studies different notions of 


an experiment that interacts with an interactive process in 

order to determine its behavior [8], [9]. We consider so 

called “strong” equivalence, where internal and external 

actions are treated in the same way. Afterward, we discuss 

“weak” equivalence, which aims to abstract away internal 

state/action as much as possible. 

In this section, we define YAMN processes to be 

equivalent (and substitutive) and their requirements. Since 

GMP are very similar to IMC transition systems, we adopt 

the definitions from [14]. 

The congruences we are going to define are strong 

Markovian bisimulation and weak Markovian congruence. 

A. Strong Markovian Bisimulation 

To define strong Markovian bisimulation, we first need 

a function γM that sums up all rates from transitions that 

start in a single state s and end in some state in a set C. 

Definition 5.1 (γM) Let (S,AV,T,R) be a GMTS and 

for s ∈ S and C ⊆ S, let 

T s C = {t|t ∈ {s}×{t}×C}. 

Then the function γM is defined as 

⎧ 

⎪⎨ S ×2 

γM : 

⎪⎩ 

S → R 

(s,C) ↦→ � 

R(t) 

t∈T s C 

Example 5.2 Consider a GMTS with states s, s1, s2, 

s3, s4, s5 and states sets of C1 = {s1,s2,s4} and C2 = 

{s3,s5}. In Fig.1 (a), we see transitions going from s to 

si for i = 1,··· ,5. Then, after the cumulative rate of (a), 

we get 

γM(s,C1) = 2λ+ν and γM(s,C2) = µ+κ. 

which is (b). 

Definition 5.3 An equivalence relation R ⊆ LYAWN× 

LYAWN is a strong Markovian bisimulation. It is a family 

of symmetric relations R = {Rb | b ∈ BExp} which 

satisfies: if and only if PRbQ implies for all a ∈ Actt 

and all equivalence classes C of Rb : 

1) If P b1,av 

−−−→ P ′ with bv(a) ∩ fv(b,P,Q) = ∅, 

then there is a b ∧ b1-partition B with 

fv(B) ⊆ fv(b,P,Q) such that for each b ′ ∈ B 

there exist b2, a ′ and u ′ with b ′ |= b2, a = b′ a ′ , 

Q b2,a′ 

−−−→ Q ′′ and P ′ RbQ ′ ; 

2) If P � i −→ then γM(P,C) = γM(Q,C). 

bv(a) is the variable through which the value can be 

carried for execution by action a, i.e., bv(c?x) = {x} 

and bv(i) = bv(c!e) = ∅. fv(a) is the value which can 

be used by action a during its execution. 

Two processes P and Q are strongly bisimilar (P ∼ Q) 

if they are contained in strong bisimulation. 

This definition amalgamates strong bisimilarity for 

stochastic processes with value passing during their executions. 

In order to compare the stochastic timing behavior, 

the cumulative rate function γM is used. What’s


s 

λ 

λ 

�� 

�� 

�� 

�� 

�� 

� 

�� 

� 

�� 

s 

�� 

� 

� 

� 

ν 

µ �� 

�� 

�� 

� κ 

�� 

�� 

� � 

�� 

� �� 

�� 

�� 

2λ+ν 

�� 

� �� 

� 

� 

� 

� 

� 

� 

�� 

� 

�µ+κ 

��s1�� 

��s2�� s3�� 

� 

� 

� 

� 

�� 

� ��s4�� s5�� 

� 

�� 

� 

� 

� 

� 

� 

� 

�� 

� 

� 

�� 

�� 

� 

� 

� 

� 

�� 

� 

�� 

�� 

�� 

�� 

�� 

��C1�� 

��C2�� C1�� C2�� (a) (b) 

Figure 1. Illustration of Example 5.2 

�� 

S3 

µ0 

true � 

�� 

� 

S0 

�� 

µ1,Readfp 

�� 

S1 

µ2,Identifyfp �� 

more, maximal progress is realized because the stochastic 

timing behavior is irrelevant for unstable expressions. 

Example 5.4 We assume that under certain situation, 

people have to register themselves either with a card 

or with their fingerprint to identify their identity so as 

to get their permissions. There are two register systems 

available, one system is equipped with a fingerprint reader 

(short for Sysfp), and the other is equipped with a card 

reader (short for Syscrd). The systems Sysfp can be 

specified as: 

And the formalized description of the Sysfp is: 

S2 

µ3,Log fp 

�� 

�� 

�� 

� 

¬true �� 

µ7 

Figure 2. Illustration of Sysfp 

[µ0].Sysfp := ([µ1].Readfp).Sys ′ fp 

Sys ′ fp := (([µ2].Identifyfp)||([µ3].Logfp)).Sys ′′ 

fp 

Sys ′′ 

fp := ((true).([µ4].Msg OK )+ 

(¬true).([µ5].Msg Again )).Sys ′′′ 

fp 

Sys ′′′ 

fp := ([µ6].Done).[µ7].Sysfp 

The systems Syscrd can be specified in Figure 1. 

And the formalized description of the Syscrd is: 

[ν0].Syscrd := ([ν1].Readcrd).Sys ′ crd 

Sys ′ crd := (([ν2].Identifycrd)||([ν3].Logcrd)).Sys ′′ 

Sys ′′ 

crd := ((true).([ν4].Msg PASS )+ 

(¬true).([ν5].Msg WrongCard )).Sys ′′′ 

crd 

Sys ′′′ 

crd := ([ν6].Done).[ν7].Syscrd 

From the Fig.2 and Fig.3, it is intuitive that Syscrd ∼ 

Syscrd during their execution when νi = νi for i = 

0,1,...,7. 

This example also show us that the data of the same action 

can be different in systems which in the equivalence 

relation of strong bisimulation, i.e., Readcrd can take 


crd 

S4 

�� 

�� 

� 

µ 4 ,Msg OK 

µ 5 ,Msg Again 

�� 

�� 

�� 

�� 

S5 

µ6,Done 

�� 

S6 

value passing of card while the Readfp can take value 

passing of fingerprint. They belong to different kinds of 

data, however, they identify the same person and achieve 

the same goal as well. 

Check executing actions, it is clear that 

Sysfp ∼ Syscrd. What’s more, from the point of 

value passing under condition of value(Readfp)/bfp 

and value(Readcrd)/bcrd, we know that the core 

“value” of them are equal: value(Readfp)/bfp = 

value(Readcrd)/bcrd. 

B. Expansion Law 

The following law expresses the most basic principle 

of the operational semantics of process algebras. It states 

that for each parallel composition of “sums” of processes 

P (where the choice operator takes the role of the sum 

here) there exists a process P ′ such that P ∼ P ′ and P ′ 

is the “sum” of parallel compositions. This means that 

parallelism is not represented explicitly, but encoded by 

the choice operator. 

Definition 5.5 (Expansion Law) 

Let 

P = � 

[λi].Pi + � 

(bj,pj).Pj 

and 

I 

Q = � 

[µk].Qk + � 

(bl,ql).Ql 

K 

where i,j,k,l range over the respective index sets 

I,J,K,L, bj and bl stands for the condition of value 

under which the action pj and ql can perform their 

J 

L


S ′ 0 

ν0 �� 

�� 

�� 

� 

�� 

ν1,Readcrd �� 

′ S 1 

executions. Let S ⊆ Act. Then 

P||SQ ∼ � 

[λi].Pi + � 

(bj,pj).Pj + 

I 

J 

� 

[µk].Qk + � 

(bl,ql).Ql + 

K 

� 

A∩B∩S 

(b j ,p j )=(b l ,q l ) 

L 

(bj,pj).(Pj||SQl) 

where A = {(bj,pj)|j ∈ J}, B = {(bl,ql)|l ∈ L} 

and 

Example 5.6 Revisit Example 5.3, we assume that 

P = (true).([ν4].Msg PASS )+ 

(¬true).([ν5].Msg WrongCard )) 

Q = (true).([ν4].Msg OK )+ 

(¬true).([ν5].Msg Again )) 

ν2,Identifycrd �� 

as time delays characterized by µ4, µ5, ν4 and ν5 have no 

influences with boolean expression true and ¬true. We 

can exchange the position between delays and boolean 

expressions by shorten true to bp and bq, Msg PASS 

to MP , Msg WrongCard to MW , Msg OK to MO and 

Msg Again to MA. Then, we have 

and 

P = [ν4].(bp,MP)+[ν5].(¬bp,MW) 

Q = [µ4].(bq,MO)+[µ5].(¬bq,MA) 

as there is no action for P and Q to synchronize, then it 

can be expanded as 

P|| ∅Q = ([ν4].(bp,MP)+[ν5].(¬bp,MW))|| ∅ 

([µ4].(bq,MO)+[µ5].(¬bq,MA)) 

∼ [ν4].(bp,MP)+[ν5].(¬bp,MW)+ 

[µ4].(bq,MO)+[µ5].(¬bq,MA) 

We know that ν4, ν5, µ4 and µ5 are variables of exponential 

distributions, by the “memoryless” property of 

Markovian process, at any time point, the time passed cannot 

influence them. So it is reasonable for[ν4].(bp,MP)+ 

[ν5].(¬bp,MW)+[µ4].(bq,MO)+[µ5].(¬bq,MA) to simulate 

the execution of P|| ∅Q. 


S ′ 2 

ν3,Log crd 

�� 

true �� 

�� 

�� 

� 

�� 

� ¬true �� 

ν7 

Figure 3. Illustration of Syscrd 

S ′ 3 

S ′ 4 

�� 

�� 

� 

ν 4 ,Msg PASS 

�� 

�� 

�� 

�� 

ν 5 ,Msg Wrong Card 

C. Weak Markovian Congruence 

S ′ 5 

ν6,Done 

�� 

′ S 6 

The weak Markovian congruence abstracts away internal 

actions. To treat internal transitions properly, we need 

the following definition. 

Definition 5.7 (Weak Markovian Bisimulation) An 

equivalence relation R with R = LYAWN × LYAWN 

is called weak Markovian bisimulation is a family of 

symmetric relationsR = {R b | b ∈ BExp}, and satisfies: 

iff PR b Q implies for all a ∈ Act and all equivalence 

classes C of R b : 

1) If P b1,av 





Q b2,â′ 

−−−→ Q ′′ and such that 

• If a ≡ c?x then there is a b ′ -partition B ′ such 

that for each b ′′ ∈ B there are b ′ 2 and Q′′ with 

b ′′ |= b ′ b′ 

2 , Q′ −−→ Q ′′ and P ′ Rb′′ Q ′′ ; 

• otherwise P ′ Rb′ Q ′ . 

2 ,i 

2) P i −→ P ′ and P ′ � i −→ imply γM(p,C i ) = γM(q,C i ) 

P andQare called weakly Markovian bisimulation equivalent 

(P ≈ Q) if there is a weak Markovian bisimulation 

R such that PRQ. 

Example 5.8 Revisit Example 5.3, we can abstract 

away some internal actions to form a system as which can 

be observed by the outside observers. First we simplify 

Sysfp in Figure 4. 

Based on Definition 5.7, we know that system Sysfp 

and Syscrd in the relation of weakly Markovian bisimulation 

in Fig.4 iff they satisfy: µi = νi for i = 1,2,3,4. 

In [14], Hermanns proved that ≈ is a congruence for 

all IMC operators (and hence also for YAMN ) except 

choice operator. The reasons for this are well known due 

to Milner [23], and the deficiency is fixed with the follow 

definition: 

Definition 5.9 (Weak Markovian Congruence)P and 

Q are said to be weakly Markovian congruent(P ≃ Q) is 

a family of symmetric relations R = {R b | b ∈ BExp}, 

if and only if for all a ∈ Act, all C ∈ YAMN/ ≈:


[µ 

S0 

′ 1 ],Readfp �� 

S1 

�� 

[µ ′ 

4 ],Done 

[µ ′ 2 ],Msg OK 

[µ ′ 

3 ],Msg Again 

1) If P b1,av 


then there is a b ∧ b1-partition B with fv(B) ⊆ 

fv(b,P,Q) such that for each b ′ ∈ B there exist 

b2, a ′ and u ′ with b ′ |= b2, a = b′ 

and such that 

�� 

� 

� S5 

S ′ [ν 

0 

′ 

1 ],Readcrd � 

�� 

� S ′ 1 

[ν ′ 2 ],Msg PASS 

[ν ′ 

3 ],Msg Wrong Card 

[ν ′ 

4 ],Done 

Sysfp Syscrd 

Figure 4. Illustration of simplified Sysfp and Syscrd 

a ′ , Q b2,â′ 

−−−→ Q ′′ 



b′ 

′ 2 ,i 

b ′′ |= b ′ 2, Q −−→ Q ′′ and P ′ Rb′′ Q ′′ ; 


2) P stable ⇒ γM(P,C) = γM(Q,C); 

3) P stable ⇔ Q stable. 

We use P ∼ = Q to stands for the weak Markovian 

congruence. 

Lemma 5.10 IfP ,Q, andRare processes, andP ∼ = Q, 

then: 

1) a.P ∼ = a.Q, and [λ].P ∼ = [λ].Q; 

2) P +R ∼ = Q+R, and R+P ∼ = R+Q; 

3) P||SR ∼ = Q||SR, and R||SP ∼ = R||SQ; 

4) recX : P ∼ = recX : Q. 

Proof: All the proofs are alike, and we prove the 

parallel composition as representation of them. 

We all know that the weak bisimulation of CCS [10], 

[22], which only restrict the bisimulation with pure action 

and states during the execution. In the definition above, 

we add another restriction based on the exponential distribution. 

The restricted exponential distribution can be 

used to calculate the mean time of the delay or duration 

of executions. 

⇒: From P ∼ = Q to P||SR ∼ = Q||SR, there are several 

action types available, and we will discuss them one by 

one: 

• For action a ∈ S, then a ∈ Act(P) and a ∈ Act(Q) 

are changed into internal action i, and we know that 

(P||SR ∼ = Q||SR) \ a is still P||SR ∼ = Q||SR for 

a ∈ S; 

• For a ∈ Act(R), R ∼ = R and R b⊲av 

−−−→ R ′ , it is easy 

to know that P||SR ′ ∼ = Q||SR ′ ; 

• For action a �∈ S ∪Act(R): 

– For input action, as a?x.P ′ ∼ = i.a?x.i.Q ′ and 

P ′ ∼ = Q ′ , we know that γ(P,P ′ ) = γ(Q,Q ′ ), 

though there are internal action is duration the 

execution of Q, there is no difference between 

the execution of a?x both in P and Q, then, we 

know that P||SR ∼ = Q||SR; 


�� 

�� 

S ′ 5 

– For output action, as c!e.P ′ ∼ = i.c!e ′ .i.Q ′ and 

P ′ ∼ = Q ′ , we know that γ(P,P ′ ) = γ(Q,Q ′ ), 



the execution of c!e in P and c!e ′ in Q, then, 

we know that P||SR ∼ = Q||SR; 

– For internal action i, it could not be observed 

from outside, and there is no change inP||SR ∼ = 

Q||SR, and of course it equals with itself. 

⇐: We know that P||SR ∼ = Q||SR, and there are 

several kinds of actions during their execution. We also 

figure them out one by one: 

• For action a ∈ Act(R) and R b⊲av 

−−−→ R ′ , the situation 

have nothing to do with P ∼ = Q; 

• For action a ∈ S, it is an internal action i, and 

P||SR ∼ = Q||SR evolves into P ′ ||SR ′ ∼ = Q ′ ||SR ′ , 

as we know that (P,P ′ ) ∈ C and (Q,Q ′ ) ∈ C, and 

there is no way to calculate the delay or duration, 

so, it is still unable to know that P ∼ = Q on action 

i; 

• For action a �∈ S ∪Act(R) 

– For input action, as a?x.P ′ ||SR ∼ = 

i.a?x.i.Q ′ ||SR, we know that 

γ(P||SR,P ′ ||SR) = γ(Q||SR,Q ′ ||SR), 



the execution of a?x both in P||SR and Q||SR, 

then, get ride of R, we get that P ∼ = Q; 

– For output action, as c!e.P ′ ||SR ∼ = 

i.c!e ′ .i.Q ′ ||SR, we know that 

γ(P||SR,P ′ ||SR) = γ(Q||SR,Q ′ ||SR), 



the execution of c!e in P and c!e ′ in Q, then, 

we know that P ′ ||SR ∼ = Q ′ ||SR, since there is 

no change on R, we have P ∼ = Q; 

– For internal action i, it could not be observed 

from outside, and there is no change inP||SR ∼ = 

Q||SR, and of course it equals with itself. 

Based on the analyze above, we complete the proof. 

� 

Example 5.11 Revisit Example 5.8. We know that 

under the condition that µi = νi < ∞ for i = 1,2,3,4, 

system Sysfp is in weak congruence with Syscrd according 

to the definition 5.7.


D. Time Restricted Markovian Bisimulation 

For bisimulation relationships, we think that Examples 

of 5.3 and 5.8 are perfect for it. Even though, it is hard 

to keep the system in bisimulation relations in real world 

systems. We have reason to assume that the subprocess 

Identifyfp in Sysfp does not have the same time as 

Identifycrd in Syscrd, because they use different kinds 

of devices to get their information. So, it is reasonable 

for us to assume that µ2 in Figure 2 does not equal with 

ν2 in Figure 3. From this point of view, it is hard to build 

even a weak bisimulation over Sysfp and Syscrd. 

In real world systems, we do not distinguish the two 

systems if they function well. Then, what makes the distance 

between bisimulation relations from the real world? 

The key reason is time, which appears in the definition 

of bisimulations both strong and weak in Markovian 

relations. 

Here, we give out another definition of bisimulation. It 

is built on the opinion of time restriction tr. tr means 

there is a time restriction/limitation for the execution 

of certain kind of actions. Usually, we use tr(P,Q) to 

show the time restriction for process P to involve into 

1 

Q duration a serial action(s). Use formula 

γM(P,Q) 

to describe the mean time of process P involves into 

Q during a serial action(s). If the execution time is 

within the restriction, we call it normal. Otherwise, we 

call it abnormal. If two systems constructed with the 

same description of actions, but with different execution 

durations, we call them bisimulation if their execution 

times are all within the restriction. In other words, both of 

the systems can satisfy the requirements on both actions 

and on time restrictions. 

According to the bisimulations equivalences of strong 

and weak, we give out the definitions of strong and weak 

bisimulation equivalences with time restrictions. 

Definition 5.12 (Time Restricted Strong Bisimulation 

with Value) An equivalence relationR ⊆ LYAWN× 

LYAWN is a time restricted strong Markovian bisimulation, 

is a family of symmetric relations R = {Rb | b ∈ 

BExp}, and satisfies: iff PRbQ implies for all a ∈ Actt, 

time restriction tr and all equivalence classes C of Rb : 

1) If P b1,av 





Q b2,a′ 

−−−→ Q ′′ and P ′ RbQ ′ ; 

2) If P � i 1 

−→ then ≤ tr(P,C) and 

γM(P,C) 

1 

≤ tr(Q,C). 

γM(Q,C) 

Two processP and Q are strongly bisimilar (P ∼tr Q) 

if they are contained in some strong bisimulation. 

Example 5.13 Revisit Example 5.3. We know that 

for i = 0,1,...,7, strong bisimulation Sysfp ∼ Syscrd 

means µi = νi. From the point of practice, we know 

that this condition is too hard to satisfy. However, to full 


fill the design requirements of these two systems, we can 

restrict the total responding time to no more than TR. 

That is, systems can accomplish its requirements. Thus, 

we might specify the systems step by step to split TR 

into tri (i ∈ N). Then, we might design Sysfp (Syscrd) 

under condition that µi ≤ tri (νi ≤ tri) for i = 0,1,...,7. 

The time restriction of the two systems can be illustrated 

by Fig.5 where tri for i = 0,1,...,7 are time restrictions 

for the action above the arrow. 

Under the time restrictions specified by Fig.5, we know 

that Sysfp ∼tr Syscrd iff µi ≤ tri and νi ≤ tri for 

i = 0,1,...,7 (which means that (µi,νi) ∈ Ci where 

Ci is a serious of equivalent class). This might loosen 

the definition of strong Markovian bisimulation 5.1 as 

µi = νi for i = 0,1,...,7. However, this change can meet 

the needs in practice, and it is more practical and easy to 

control. 

This definition can also be explained as follows: all 

the systems satisfying their design requirements can be 

taken as equal. When system Sysfp and syscrd working 

independently, all of them can function well according 

to the design requirements. That is, one system can 

take the place of another in practice according to the 

design requirements, this can also be taken as a kind of 

equivalence in both algebra and practice. 

Definition 5.14 (Time Restricted Weak Markovian 

Bisimulation) An equivalence relation R with R ⊆ 

LYAWN×LYAWN is called time restricted weak Markovian 

bisimulation, is a family of symmetric relations 

R = {Rb | b ∈ BExp}. It satisfies: iff PRbQ implies 

for all a ∈ Act, time restriction tr and all equivalence 

classes C of Rb 1) If P b1,av 





and such that 

a ′ , Q b2,â′ 

−−−→ Q ′′ 


that for each b ′′ ∈ B there are b ′ 2 and Q ′′ with 

b ′′ |= b ′ 2 , Q′ b′ 2 ,i 

−−→ Q ′′ and P ′ Rb′′ Q ′′ ; 


2) P i −→ P ′ and P ′ � i −→ imply 

and 

1 

γM(q,C i ) ≤ tr(p,Ci ). 

1 

γM(p,C i ) ≤ tr(p,Ci ) 

P and Q are called time restricted weak Markovian 

bisimulation equivalent (P ≈tr Q) if there is a weak 

Markovian bisimulation R such that PRQ. 

Example 5.15 When we make clear of the Example 

5.13, it is easy to understand this one. Time restricted 

weak Markovian bisimulation take input action Readfp 

and Readcrd as a special case. It is easy to understand 

in Fig.4 of Example 5.8. Fingerprint might take different 

steps in number to get its information as card reader. It is 

reasonable to assume that card reader takes less steps to 

get its value than fingerprint. Because it might take more 

steps to calculate the value of fingerprint, thus fingerprint 

has more internal action i than card reader. As we know


�� S3 

��Initial 

T �� 

tr0 

�� 

�� 

S0 

�� 

Read �� 

S1 

tr1 

that either Readfp in Sysfp or Readcrd in Syscrd is 

restricted by executing time to less than tr1 (Fig.5). This 

satisfies condition 2 in definition 5.14. 

Checking the execution of Sysfp and Syscrd step by 

step, we know that they belong to the time restricted weak 

Markovian bisimulation. 

For any different abstract levels of the description of 

systems, there are different atomic actions. The higher 

the abstract level, the more abstract atomic actions are 

required for that level. Thus, atomic actions of higher 

abstract level contain more internal actions. Another way 

to turn normal action (observable) into internal action is 

the composition of compositions into a larger system. 

The output of one component is the input of another, 

thus, at least two normal actions are abstracted away by 

composition. 

Lemma 5.16 Time restricted congruence is a congruence 

with respect to all operators of LYAWN . If P , Q 

and R are expressions of LYAWN and a ∈ Act, λ ∈ R 

and X ∈ Var, then 

• P ≈tr Q implies a.P ≈tr a.Q; 

• P ≈tr Q implies [λ].P ≈tr [λ].Q; 

• P ≈tr Q implies P +R ≈tr Q+R and R+P ≈tr 

R+Q; 

• P ≈tr Q implies P||SR ≈tr Q||SR and R||SP ≈tr 

R||SQ; 

• P ≈tr Q implies rexX : P ≈tr rexX : Q. 

Proof: All the proofs are alike, and we prove the choice 

composition as representation of them. 

⇒: We suggest that the executing time of P as trP ≤ 

trP , Q as trQ ≤ trQ, and R as trR as we know 

that P ≈tr Q. So, we get trP = trQ, according 

to the definition 5.14. We have the executing time of 

P + Q is max(trP,trQ), the executing time of R + 

Q is max(trR,trQ). Then, we get max(trR,trP) = 

max(trR,trQ). As to the pure actions, it is easy to know 

that P = Q. Then we have P+R = Q+R. As the choice 

composition does not distinguish the position under the 

summation, then we have R + P = R + Q. Put the 

executing time of actions and pure action together, we 

have P +R ≈tr Q+R. 

⇐: We suggest that the executing time of P + Q 

is max(trP,trQ). The executing time of R + Q is 

max(trR,trQ). As we know that P +R ≈tr Q+R, then 

we have max(trR,trP) = max(trR,trQ). Now, there 


Identify 

�� 

S2 

tr2 

Log 

tr3 

�� 

�� 

�� 

�� 

F � 

S4 

Re−Initial 

tr7 

Figure 5. Illustration of Sysfp 

�� 

Msg(T) 

tr4 

Msg(F) 

�� tr5 

�� 

�� 

�� 

�� 

�� 

are four situations: 

S5 

Done �� 

S6 

tr6 

1) max(trP,trR) = trP and max(trQ,trR) = trQ, 

then we have trP = trQ, i.e., we have P ≈tr Q as 

needed; 

2) max(trP,trR) = trP and max(trQ,trR) = trR, 

then we have trP ≥ trR and trR ≥ trQ which is 

conflict with max(trR,trP) = max(trR,trQ). So, 

this condition is impossible; 

3) max(trP,trR) = trR and max(trQ,trR) = trQ, 

then we have trR ≥ trP and trQ ≥ trR, which is 

conflict with max(trR,trP) = max(trR,trQ). So 

this condition is impossible; 

4) max(trP,trR) = trR and max(trQ,trR) = trR. 

We know that max(trP,trQ) ≤ trR, this means 

that the executing time of P and Q are less than 

trR. As we know that trR is assumed to be under 

the time restriction of the requirements, so we have 

P ≈tr Q. 

Based on the above analyze, we get the proof done. 

� 

Based on the assumption that 1 

≤ tri ( 1 

≤ tri) 

for i = 0,1,...,7. We know that either the time delay 

before next action or the duration of an action is no more 

than the design requirements which is a serious of real 

numbers. That is 1 

≤ tri < ∞ ( 1 

≤ tri < ∞) for 

µi 

i = 0,1,...,7. In other words, the variables characterizing 

executing time in processes cannot be infinite. From the 

point of Markovian chains, all the chains of this kind is 

stable. This is a bridge to fill the gap between time restricted 

weak Markovian bisimulation and time restricted 

weak Markovian congruence which will be defined in the 

following. 

Definition 5.17 (Time Restricted Weak Markovian 

Congruence) P and Q are said to be weakly Markovian 

congruent (P ≃tr Q) is a family of symmetric relations 

R = {R b | b ∈ BExp}. If and only if ∀a ∈ Act, time 

restriction tr and all C ∈ YAMN/ ≈tr: 

1) If P b1,av 




µi 

νi 


and such that 

νi 

a ′ , Q b2,â′ 

−−−→ Q ′′ 

• If a ≡ c?x then there is a b ′ -partition B ′ such


S0 

�� 

[tr1],Read 

�� 

S1 

[tr4],Done 

[tr2],Msg(T) 

�� 

[tr3],Msg(F) �� 

S5 

Figure 6. Illustration of simplified Systr 


b′ 

′ 2 ,i 

b ′′ |= b ′ 2, Q −−→ Q ′′ and P ′ Rb′′ Q ′′ ; 


1 

2) ≤ tr(P,C), and P is stable; 

γM(P,C) 

1 

3) ≤ tr(P,C), and Q is stable. 

γM(Q,C) 

Example 5.18 Revisit Example 5.8 and 5.11, we know 

that system Sysfp and Syscrd obey the same time restriction 

which can be illustrated in Fig.6. 

Fig.6 illustrates the observable actions and durations 

of the execution of system Sysfp and Syscrd. We know 

that system Sysfp ≈ Syscrd in Example 5.8. Sysfp ≈tr 

Syscrd satisfies the condition max( 1 

, 

µi 

1 

) ≤ tri for i = 

νi 

1,2,3,4. 

Example 5.19 If we take the Example 5.8, 5.11, 5.15 

and 5.18 as the abstract levels of observations. Take 

Example 5.3 as a refined level of observation, we can 

build a time restriction on execution between these two 

different levels. 

If we want Fig. 1 in Example 5.3 to obey the time 

restriction in Fig. 6, it is intuitive for us to know that for 

Sysfp as Fig.2 with time restriction as Fig.6 satisfy the 

following conditions: 

max( 1 

, 

µ0 

1 

+ 

µ6 

1 

,ν0, 

µ7 

1 

+ 

ν6 

1 

) ≤ tr4; 

ν7 

max(µ1,ν1) ≤ tr1; 

max( 1 

, 

µ3 

1 

+ 

ν2 

1 

) ≤ tr2; 

ν4 

max( 1 

, 

µ3 

1 

+ 

ν2 

1 

) ≤ tr3. 

ν5 

and for Syscrd as Fig.3 with time restriction in Fig.6 

satisfies the following conditions: 

max( 1 

, 

ν0 

1 

+ 

ν6 

1 

,ν0, 

ν7 

1 

+ 

ν6 

1 

) ≤ tr4; 

ν7 

max(ν1,ν1) ≤ tr1; 

max( 1 

, 

ν3 

1 

+ 

ν2 

1 

) ≤ tr2; 

ν4 

max( 1 

, 

ν3 

1 

+ 

ν2 

1 

) ≤ tr3. 

ν5 

Based on the above conditions, we know that 

Sysfp ∼ =tr Syscrd and Sysfp ≈tr Syscrd. It is intuitive 

that ∼ =tr and ≈tr under the time restricted weak 

Markovian bisimulation and congruence are of the same 

equivalence class ( ∼ =tr,≈tr) ∈≡. 

Theorem 5.20 P ≈tr Q ⇔ P ∼ =tr Q. 


Proof sketch: The only difference between P ≈tr Q 

and P ∼ =tr Q lies in the definitions of 5.14 and 5.17. It 

is based on whether the Markovian chain is stable or not. 

Based on the assumption that the mean time of delays 

and executing durations are no more than time restriction 

tr and tr < ∞, we know that all actions in Markovian 

chains can full fill their time restrictions. They terminate 

within time limitations (i.e., no more than tr). Both 

time restricted Markovian bisimulation and time restricted 

Markovian congruence are based on the assumption that 

executing time is no more than time restriction. 

� 

Lemma 5.21 Time restricted congruence is a congruence 

with respect to all operators of LYAWN . If P , Q 

and R are expressions of LYAWN and a ∈ Act, λ ∈ R 

and X ∈ Var, then 

• P ∼ =tr Q implies a.P ∼ =tr a.Q; 

• P ∼ =tr Q implies [λ].P ∼ =tr [λ].Q; 

• P ∼ =tr Q implies P +R ∼ =tr Q+R and R+P ∼ =tr 

R+Q; 

• P ∼ =tr Q implies P||SR ∼ =tr Q||SR and R||SP ∼ =tr 

R||SQ; 

• P ∼ =tr Q implies rexX : P ∼ =tr rexX : Q. 

Proof: Similar with the proof of Lemma 5.16. 

� 

VI. CONCLUSION 

In this paper, we introduced the language of YAWN 

with value passing which is perfect to describe the 

stochastic phenomena in real world. The supporting 

model of YAWN is continuous time Markovian chains, 

which are frequently used in modeling manufacturing 

system, computer networks, communication systems and 

so on. 

In analyzing the behaviors of complex systems, i.e., 

computer networks and operating systems, it is inevitable 

to deal with value passing (i.e., data of all kinds and forms 

and control based on values). During the analyzing of 

actions in processes, we introduced value passing into 

the language of YAWN with value passing. Thus, there 

are several kinds of actions with value passing including 

input, output, internal action, and generalized form. 

After we had a general view of the language dealing 

with stochastic processes, we gave out the syntax of 

YAWN with its informal descriptions. In order to define 

the semantics more clearly, we gave out the formal meaning 

of the language YAWN . Then, we introduced generalized 

Markovian transition system (short for GMTS) as 

the model of LYAWN . Based on GMTS, we gave out the 

operational semantics of the languageYAWN with value 

passing. Based on the theory of classic process algebras, 

we showed axioms of the operators in the language of 

YAWN . 

The axioms gave out the basic equivalence relations 

of the basic operators in YAWN with value passing. 

One important task for process algebras is to build the 

equivalent relationships between processes. We treated 

them in two different ways: one way is to treat both action


and time (delay or duration) strictly. This policy introduce 

the equivalent relations as strong Markovian bisimulation, 

weak Markovian congruence and expansion law. 

Another way is to treat action and time with different 

policy: treating action strictly while treating time loosely 

with a duration (i.e., within time limitations). This means 

that executing time within time limitation can be considered 

as equal in the comparison of two processes. 

This policy produces time restricted strong bisimulation, 

time restricted weak Markovian bisimulation, and time 

restricted weak Markovian congruence. 

When all the bisimulation relations are defined, we 

proved that they can be applied to all the operators inside 

the language of YAMN with value passing. 

REFERENCES 

[1] Luca de Alfaro. Stochastic transition systems. In David 

Sangiorgi and Robert de Simone, editos, CONCUR ’98: 

Concurrency Theory (Proceedings), volume 1466 of Lecture 

Notes in Computer Science., Springer Verlag, September 

2001. 

[2] J. A. Bergstra and J.W. Klop, Algebra of Communitating 

Processes with Abstraction, TCS 37,1,pp. 77-121, 1985. 

[3] M. Bernardo and R. Gorrieri. Extended Markovian Process 

Algebra. In Ugo Montanari and Vladimiro Sassone, editors, 

CONCUR’96: Concurrency Theory (7th International Conference, 

Pisa, Italy, August 1996), volume 1119 of Lecture 

Notes in Computer Science. Springer, 1996. 

[4] M. Bravetti, M. Bernardo, R. Gorrieri, Generalized Semi 

Markovian Process Algebra, Technical Report UBLCS-97- 

9, University of Bologna (Italy), October 1997 

[5] Ed Brinksma Holger Hermanns, Process algebra and 

Markov chains. In Ed Brinksma, Holger Hermanns, and 

Joost-Pieter Katoen, editors, Lectures on Formal Methods 

and Performance Analysis, volume 2090 of Lecture Notes 

in Computer Science, page 183-231, Springer Verlag, 2001 

[6] P. Buchholz, Markovian Process Algebra: Composition and 

Equivalence. In U. Herzog and M. Rettelbach, editors, Pro. 

of the 2nd Workshop on Process Algebras and Performance 

Modelling , Erlangen-Regensberg, July 1994. IMMD, Universität 

Erlangen-Nürnberg. 

[7] Norbert Götz, Stochastische Prozeßalgebren – Integration 

von funktionalem Entwurf and Leistungsbewertung Verteilter 

Systeme. PhD thesis. Univerität Erlangen-Nürnberg, 

Germany, 1994. 

[8] R.J. van Glabbeek, The linear time - branching time 

spectrum (extended abstract). CWI Amsterdam Report CS- 

R9029. 

[9] R.J. van Glabbeek, The linear time - branching time spectrum 

II: The semantics of sequential systems with silent 

moves (Extended Abstract) In Eike Best, editor, Fourth International 

Conference on Concurrency Theory (CONCUR 

’93, Hildesheim, Germany), volume 715 of LNCS, pp. 66- 

81, Springer, 1993. 

[10] A. Ingólfsdóttir and H. Lin. A symbolic Approach to 

value passing Processes. In J.A. Bergstra, A. Ponse, and 

S.A. Smolka, eds., Handbook of Process Algebra, 427-478, 

Elsevier, 2001. 

[11] M. Hennessy, H. Lin (1996) Proof systems for messagepassing 

process algebras Formal Aspects of Computing, 379- 

407, Volume 8, 1996, Springer London 

[12] H. Hermanns and M. Ribaudo, Syntax, Semantics, Equivalences, 

and Axioms for MTIPP. In U. Herzog and M. 

Rettelbach, editors, Pro. of the 2nd Workshop on Process Algebras 

and Performance Modelling , Erlangen-Regensberg, 

July 1994. IMMD, Universität Erlangen-Nürnberg. 


[13] C. A. R. Hoare, Communicating Sequential Processes, 

Prentice Hall International, 1985. 

[14] A Holger Hermanns. Interactive Markov Chains. PhD 

thesis, Universität Erlangen- Nürnberg, Germany, 1998. 

[15] Holger Hermanns and Michael Rettelbach, Towards a 

superset of LOTOS for performance prediction, In Ribaudo 

[25], page 77-94, 1996. 

[16] H. Hermanns, M. Rettelbach, Syntax, Semantics, Equivalences, 

and Axioms for MTIPP, in Proc. of PAPM ’94, pp. 

71-87, Erlangen, 1994 

[17] H. Hermanns, U. Herzog, and J.-P. Katoen. Process algebra 

for performance evaluation. Theoretical Computer 

Science, 274(1-2):43–87, 2002. 

[18] Jane Hillston, A Compositional Approach to Performance 

Modelling, PhD thesis, University of Edinburgh, 1994. 

[19] Ronald A. Howard. Dynamic Probabilistic Systems. volume 

2: Semimarkov and Decision Processes, John Wiley & 

Sons, 1971. 

[20] J.P. Katoen. Concepts, Algorithms and Tools for Model 

Checking. Erlangen: Institut f. Mathematische Maschinen 

und Datenverarbeitung, 1999. 188-255. 

[21] J-P. Katoen. Quantitative and qualitative extensions of 

event structures, PhD thesis, University of Twente, 1996. 

[22] Robin Milner, A Calculus of Communicating Systems, 

volume 92 of Lecture Notes in Computer Science. Prentice 

Hall International, 1989. 

[23] Robin Milner, Communication and Concurrency, Prentice 

Hall, 1989. 

[24] Martin L. Puterman. Markovian Decision Processes. John 

Wiley & Sons, 1994. 

[25] Marina Ribaudo, editor. Proceedings of the fourth workshop 

on process algebras and performance modelling. Edizione 

C.L.U.T. Torino, 1996. 

[26] Moshe Y. Vardi. Automatic verification of probabilistic 

concurrent finite-state systems. In 26th Annual Symposium 

on Foundations of Computer Science (FOCS ’85), pages 

327-338. IEEE Computer Society Press, October 1985. 

[27] Guang Zheng, Shaorong Li, Jinzhao Wu, and Lian Li. A 

Non-interleaving Denotational Semantics of Value Passing 

CCS with Action Refinement. FAW 2007, LNCS 4613, pp. 

178-190, 2007., Springer-Verlag Berlin Heidelberg 2007.


Study on Visual Knowledge Structure Reasoning 

Huimin Lu 

1 College of Software Engineering, Changchun University of Technology, Changchun, China 

2 College of Computer Science and Technology, Jilin University, Changchun, China 

Email: luhm.cc@gmail.com 

Abstract—Intelligent Topic Map (ITM) embodies the multilevel, 

multi-granularity and the inherent relevant 

characteristics of knowledge. With ITM as infrastructure, 

this paper presents a visual knowledge structure reasoning 

method integrates the logic-based knowledge reasoning and 

the structure-based knowledge reasoning. The logic-based 

knowledge reasoning implements knowledge consistency 

checking and the implicit associations reasoning between 

knowledge points, it can help us obtain the optimal 

description of knowledge. In order to construct the complete 

knowledge structure, a Knowledge Unit Circle Search 

strategy for structure-based knowledge reasoning is 

proposed, by which more detailed semantic association of 

knowledge is provided and the inherent relevant 

characteristics of knowledge is obtained. The knowledge 

reasoning results are visualized by ITM, which provides a 

visual knowledge map. It is available for users to acquire the 

knowledge and associations among them. A prototype 

system has been implemented and applied to the massive 

knowledge organization, management and service for 

education. 

Index Terms—topic map, intelligent topic map, knowledge 

reasoning, knowledge visualization 


Knowledge reasoning mainly includes two types: the 

logic-based knowledge reasoning and the structure-based 

knowledge reasoning. The logic-based knowledge 

reasoning often used to describe knowledge 

representation and reasoning based on the logic. It is 

rigorous, flexible and with a strict formal definition, but 

the lack of structure constraint. The structure-based 

knowledge reasoning constructs knowledge based on 

some data structure, such as vector space, tree, graph, etc. 

It bodes well for knowledge and the relations between 

them. Knowledge doesn’t exist by itself, since knowledge 

always has all kinds of relations with other knowledge. 

According to constructivism theory and cognitive load 

theory perspective, the inner relevance of knowledge can 

contribute to achieving consistent with the person’s own 

cognitive pattern, and thereby the cognitive efficiency 

Manuscript received April 3, 2010; accepted February 10, 2011. 

Copyright corresponding author: Huimin Lu. 


doi:10.4304/jsw.6.5.783-790 

2 Liang Hu and Gang Liu 1 

Email: hul@jlu.edu.cn, liug8818@mail.ccut.edu.cn 

can be increased [1], but knowledge reasoning can not 

guarantee as effective as logical representation. So, a 

knowledge representation model should be built to 

integrate these two types of knowledge reasoning in order 

to obtain the satisfactory knowledge reasoning results [2]. 

Moreover, the reasoning results should be displayed by 

visual knowledge structure. Its goal is to transfer and 

create new knowledge through using visualizations. 

Topic Map(TM) is an ISO standard (ISO/IEC 13250) 

that describes knowledge structures and associates them 

with information resources [3] [4]. Topic map constructs 

a structured semantic network above the knowledge 

resources. It describes the concepts and the semantic 

relations between them, and can locate the resources 

which are associated with the concepts and realize the 

concrete objects to be joined with abstract concepts. It 

provides a visual knowledge map, which is available for 

users to acquire knowledge and associations among them. 

However, the conventional topic map can not provide 

users with efficient knowledge navigation, and we unable 

to acquire the implicit knowledge for it lack of reasoning 

abilities. So, we extend the conventional topic map in 

structure and enhance the reasoning functions, which is 

defined Intelligent Topic Map (ITM) [5]. EXTM 

(Extended XTM) extended the syntax and semantics of 

XTM (XML for Topic Maps) [6] so that it can describe 

ITM elements (such as clusters, topics, knowledge 

elements), and provides a model and grammar for 

representing the structure of ITM and defining reasoning 

rules. EXTM makes XML extend to the semantic field. It 

defines an abstract, graphics-based knowledge 

association model and allows the logic-based knowledge 

reasoning to discover new knowledge. 

We propose a novel method of visual knowledge 

structure reasoning with the intelligent topic map as 

infrastructure, which can efficiently implement both the 

structure-based knowledge reasoning and the logic-based 

knowledge reasoning. The reasoning results are 

visualized by ITM. It provides a visual knowledge map, 

which is available for users to acquire the knowledge and 

associations among them. Visualization navigation 

capabilities of exploiting the created knowledge 

structures are based on hyperbolic geometry concepts and 

provide users with intuitive access mechanisms to the 

required knowledge.


II. RELATED WORKS 

The knowledge representation model which is able to 

integrate logic reasoning and structure reasoning includes 

XML, RDF, ontology, etc. XML provides a flexible, 

general, rich structured information representation and 

convenient for the cooperative processing of 

heterogeneous knowledge [7]. RDF is an effective means 

of semantic information description [8]. Ontology 

establishes a classified hierarchy by defining the concepts 

and the relevance between them, and thus to build the 

semantic space of concepts [9]. However, they are not in 

an intuitive and graphical way to display knowledge, and 

there is no relationship between the resources and the 

related concepts contained. The structure of topic map 

composed of Topics, Associations and Occurrences 

(TAO) [10], which describes the concepts and the 

semantic relationships between them and can locate the 

resource which are associated with the concept. TM 

establishes a structured semantic web above the resources 

level and implements the semantic organization and 

joining between the physical resource entities and the 

abstract concepts. Topic maps are dubbed “the GPS of 

the information universe”. TM can be applied to crosssystem 

since the XTM (XML for Topic Maps) syntax is 

based on XML and is an exchangeable data standard. The 

greatest advantage of TM is the discovery and 

visualization of knowledge architecture [11] [12]. 

Graphic display based on topic map is more perceivable, 

it can provide visual knowledge navigation mechanism. 

Topic map inherits the characteristics of knowledge 

organization methods such as index, glossary, thesaurus, 

taxonomy, concept map, ontology, etc. Consequently, 

topic map adapts to knowledge logical organization and 

becomes the state-of-art semantic technologies, such as 

the application of topic maps technology in context of elearning 

environment, especially based on analyses of 

topic relative semantic structure, and used topic maps to 

represent learning resources and associated semantics 

such as metadata [13][14][15]. H. Lu, et al proposed a 

novel concept of intelligent topic map for knowledge 

organization and knowledge services, which embodies 

the multi-level, multi-granularity and inherent relevant 

characteristics of knowledge and realizes knowledge 

reasoning [16]. 

III. ITM DESCRIPTION 

A. Overview of ITM Structure 

The structure of topic map is shown in Fig. 1. It 

composed of Topics, Associations and Occurrences 

(TAO). In order to overcome the drawbacks of topic map, 

we add a clustering level and a knowledge element level 

in ITM, which depicts the hierarchical relation of “cluster 

- topic - knowledge element - occurrence”. The structure 

of ITM is shown in Fig. 2. 

Cluster: Each cluster contains several closely related 

topics so that the topics in the same cluster are similar in 

some sense. Clusters provide the effective navigation and 

browsing mechanism for users. 


Figure 1. The structure of conventional topic map. 

Figure 2. The structure of intelligent topic map. 

Definition 1: When given an ITM, a cluster (c) is 

defined as following two tuples: 

c = ( Nc, Tc) 

Nc —the name of cluster 

Tc —the set of all topics in the c 

Topic: It can be any “thing” (such as a person, an 

entity, a concept, really anything) — regardless of 

whether it exists or has any other specific characteristics. 

Definition 2: When given an ITM, a topic (t) is defined 

as following six tuples: 

t = ( Nt, At, Dt, E, g, f ) 

Nt —the name of topic 

At = { at 

1 

, at 

2 

,..., atn} — a set of associations with 

topic Nt 

Dt = { dt 

1 

, dt 

2 

,..., dtm} — a set of topic association 

types ( m≤ n ) 

E = { e 

1 

, e 

2 

,..., en} —a set of elements relevant to Nt , 

the element is cluster, topic or knowledge element 

Function g : At → E —given a association relevant to 

element 

Function f : At → Dt —given a association relevant to 

type 

Definition 3: When given an ITM, a knowledge 

element (ke) is defined as following six tuples: 

ke = ( Nke, Ake, Dke, E, g, f ) 

Nke —the name of knowledge element 

Ake = { ake 

1 

, ake 

2 

,..., aken} —a set of associations with 

knowledge element Nke


Dke = { dke 

1 

, dke 

2 

,..., dkem} — a set of knowledge 

element association types ( ≤ ) 

m n 

E = { e 

1 

, e 

2 

,..., en} — a set of elements relevant to 

Nke 

Function g : At → E —given a association relevant to 

element 

Function f : At → Dt —given a association relevant 

to type 

Occurrence: representing information resources 

relevant to a particular topic. An occurrence can be a 

document, a picture or video depicting the topic, a simple 

mention of the topic in the context of something else. 

Association: A topic association asserts a relationship 

between two or more topics. 

Definition 4: When given an ITM, an association (a) is 

defined as following three tuples: 

a = ( e 

1 

, e 

2 

, d) 

e 

1 

, e 2 —the elements of ITM 

d—the association type 

ITM provides strong paradigm and concept for the 

semantic structuring of linked networks. It can establish 

the relations among unstructured information resources, 

thereby allowing to link heterogeneous, unmodified 

resources of information semantically by creating a 

semantic web and implement concrete objects to be 

joined with abstract concepts. It lays a foundation for 

high-quality structure-based knowledge reasoning. 

B. EXTM 

XTM was proposed by Newcomb and Biezunsk. It 

provides a model and grammar for representing the 

structure of information resources used to define the 

topics and their associations. Moreover, we enhance the 

reasoning functions in ITM. We establish corresponding 

logical reasoning rules and grammar, and then realize 

knowledge representation and knowledge reasoning. 

::= 

'' 

'' 

 

'' 

'' 

{} 1-n 

'' 

'' 

{} 1-n 

'' 

'' 

::=| 

::=| 

::= 

::= '' 

::={AND |OR |NOT} 

::='' '' 


::='' 

::={partOf | subClassOf | instanceOf | 

propertyOf | reasonOf | preconditionOf| caseOf | referenceOf, 

and so on} 

::='' 

'' 

::=''{}1-n '' 

::='' 

IV. VISUAL KNOWLEDGE STRUCTURE REASONING 

The visual knowledge structure reasoning method 

using ITM includes three parts: the logic-based 

knowledge reasoning, the structure-based knowledge 

reasoning and visualization of reasoning results. The topdown 

method is adopted to define the abstract workflow 

as following: 

Step 1: Defining the top-level composite processes. As 

shown in Fig. 3, three composite processes which named 

“LogicKnowledgeReasoning”, 

“StructureKnowledgeReasoning” and 

“VisualizationDisplay” are defined, respectively. “Join” 

denotes the former processes must be finished before the 

last one is started. The input of process 

“VisualizationDisplay” is the reasoning results while the 

outputs of it is the visual knowledge structure. 

Figure 3. The top-level definition of composite processes. 

Step 2: Refining the definition of process 

“LogicKnowledgeReasoning” as shown in Fig. 4, it 

includes two processes: the knowledge consistency 

checking and the implicit associations reasoning. 

Knowledge 

Consistency Checking 

Implicit Associations 

Reasoning 

Join 

Logic Knowledge 

Reasoning Results 

Figure 4. The definition of “LogicKnowledgeReasoning” 

A. The Knowledge Consistency Checking 

In the process of ITM constructing, conflicts can be 

caused by many reasons, like the differences of people’s 

understanding, the marking of knowledge resources, and 

the constructing of knowledge organization. These 

conflicts cause information redundancies, contradictions 

and mistakes. The knowledge consistency checking can 

eliminate them and can help us obtain the optimal 

description of ITM. It includes the reflexivity checking,


loop transitivity checking, knowledge redundancy 

checking and knowledge contradiction checking. 

Reflexivity checking: If an element (topic or knowledge 

element) of ITM is associated with itself, there exists 

reflexivity conflict. It is defined as follows: 

∃e∈ITM, eAe 

When the reflexivity conflict is detected, the 

association between the same elements would be deleted. 

Loop transitivity checking: If there is an association 

loop between the two directly related elements of ITM, 

there exists a loop transitivity conflict. It is defined as 

follows: 

∃e1∈ITM , ∃e2 ∈ITM , e1 A e2 ∧e2 A e1 

When the transitivity conflict is detected, one of the 

associations between the elements would be deleted. 

Knowledge redundancy checking: There exists 

redundancy if have the same elements (topics or 

knowledge elements) in an ITM. 

∃e∈ITM , ∃e∈ ITM , e = e 

1 2 1 

Though knowledge redundancy is not a mistake on 

semantics, it would be resolved when it is detected for 

ensuring certainty and uniqueness. 

Knowledge redundancy checking includes two steps: 

the same elements searching and merging. 

First, we adopt a similarity measure algorithm for 

topics (or knowledge elements) which called 

Comprehensive Information-based Similarity Measure 

Algorithm (CISMA) [17]. This algorithm describes how 

similar the related topics (or knowledge elements) are. 

The process used in the similarity algorithm consists of 

syntactic matching, semantic matching, and pragmatic 

matching. For an element pair (e1, e2), we calculate the 

similarity as follows: 

( 1 

, 

2) = 

1 Syntax ( 1 

, 

2) + 

2 Semantics ( 1 

, 

2) 

w SIM ( e , e ) 

SIM e e w SIM e e w SIM e e 

3 Pragmatics 1 2 

SIMSyntax(e1, e2): denotes syntactic matching. It is used 

to compute the syntactic similarity by analyzing the 

character composition of elements. 

SIMSemantics(e1, e2): denotes semantic matching. It 

analyses the static semantic similarity with aspect to 

synonyms. 

SIMPragmatics(e1, e2): denotes pragmatic matching. It 

computes dynamic semantic similarity, which resolves 

the problem of polysemy. 

w is weight. 

Second, merging the same elements adopt the 

following rules. 

Rule 1: Attribute Merging (AM). When given a 

merging element, AM is defined as following five tuples: 

AM = ( Ne, Na, D, V , θ 

I ) 

Ne —the name of element 

Na —the name of attribute 


2 

(1) 

(2) 

(3) 

(4) 

D —the values range of Na 

VI = { I 

1 

, I 

2 

, ..., In} —a set of Na values in range of D 

θ —merging operator 

If given a question about attribute 

merging AM ( Ne, Na, D, VI , θ) 

defined as follows: 

= , its solution K a is 

Ka = ( Ne, Na, D, θ ( I 

1 

, I 

2 

, ..., In)) 

Rule 2: Element Merging (EM). If element e1 has high 

similarity with e2 in ITM, the two elements would be 

merged into one element (e1 or e2). Element merging is 

defined as following four tuples: 

( , , , ) 

EM = NE E 

A 

E 

AI 

Eθ 

NE { ne 

1 

, ne 

2 

,..., ne 

k 

} 

= —a set of the element name 

{ 1 

, 

2 

, ..., } 

{ 1 

, 

2 

,..., } 

{ θ, θ , θ ,..., θ} 

E A A A n 

A = —a set of all EM attributes 

= —a set of all attribute values 

E 

AI 

E 

I 

E 

I 

EIn Eθ = 

1 2 n —a set of merging operators for 

each attribute used 

If given a question about elements 

merging EM = ( NE, E 

A 

, E 

AI 

, Eθ 

) , its solution Kea is 

defined as follows: 

Kea = ( θ( ne 

1 

, ne 

2 

, ..., ne 

k 

), EA, θ1 ( E 

I1 ) ∪θ2 ( E 

I2 

), ..., θn( 

EIn 

)) 

Rule 3: Association Merging (AssM). When two 

elements are merged, the association merging would be 

considered. It is defined as following three tuples: 

AssM = ( NE, ER , θ) 

NE { ne 

1 

, ne 

2 

, ..., ne 

k 

} 

= —a set of the element name 

{ ( , ) , ( , ) ,..., ( , ) } 

E 

R 

= R 

S1 R 

O1 R 

S2 NE 

R 

O2 R 

Sn 

R 

On 

of elements related to 

R 

Sn —association type 

R On —association object 

(5) 

(6) 

— a set 

θ —merging operator 

Through knowledge consistency checking, we can 

obtain an ideal ITM description. It lays a foundation for 

the structure-based knowledge reasoning. 

B. The Implicit Associations Reasoning 

The implicit associations reasoning can discover new 

associations between elements and can help us obtain 

new knowledge. In this paper, we mainly discuss the 

association of subClassOf, instanceOf, memberOf, 

precorderOf, and postorderOf. 

subClassOf: When given element ta and tb, subClassOf 

(ta, tb) indicates topic ta is a subclass of tb, ta is called subtopic 

and tb is called the relevant parent-topic. Knowledge 

reasoning rules based on subClassOf is as follows: 

( ) ∧ 

( 

( , ) 

subClassOf t , t subClassOf t , t 

→ subClassOf t t 

a b b c 

a c 

) 

(7)


( ) ∧ 

( 

( , ) 

subClassOf t , t hasAttribute t , A 

→ hasAttribute t A 

a b b 

a 

( ) ∧ ( 

( , ) 

subClassOf t , t instanceOf i, t 

→ instanceOf i t 

a b a 

b 

) 

) 

(8) 

(9) 

instanceOf . For the element e and its instance set I e , 

i i I instanceOf i , e 

the association between ( ∈ ) 

( ) 

denotes i is an instance of e . Knowledge reasoning rule 

based on instanceOf is as follows: 

( , ) ∧ 

( , ) 

( , ) 

instanceOf i e hasProperty e P 

→ hasProperty i P 

e 

(10) 

memberOf : memberOf ( M , W ) denotes M is a 

member of W . memberOf and instanceOf are two 

kinds of completely different associations, it emphasizes 

on the association between elements. 

preorderOf and postorderOf : The preorderOf 

represents that one elements B is comes out before 

another element A , denoted as preorderOf ( B, A) 

. The 

postorderOf represents that A is comes out after B , 

denoted as postorderOf 

( A, B) 

. Knowledge reasoning 

rules based on the preorderOf and postorderOf 

associations are as follows: 

( , ) ∧ 

( 

( , ) 

preorderOf B A preorderOf A, C 

→ preorderOf B C 

( ) ( 

( , ) 

postorderOf A, B ∧ postorderOf B, C 

→ postorderOf A C 

preorderOf ( B, A) → postorderOf ( A, B) 

) 

) 

) 

(11) 

(12) 

Inverse relation between preorderOf and 

postorderOf : 

( ) → 

( 

postorderOf A, B preorderOf B, A 

(13) 

(14) 

In addition to the above association types, there are 

causalOf, referenceOf, exampleOf, and so on. 


“StructureKnowledgeReasoning” as shown in Fig. 5, it 

includes two processes: Get user interest node and 

Structure reasoning method. 

Structure reasoning method: Since knowledge is 

highly correlated with each other, in order to acquire the 

complete knowledge structure, we must implement the 

semantic implication extension, the semantic relevant 


Figure 5. The definition of “StructureKnowledgeReasoning” 

extension and the semantic class belonging confirmation. 

According to the characteristics of ITM, we propose an 

extended algorithm based on knowledge unit circle, 

named Knowledge Unit Circle Search (KUCS) strategy. 

Before discussing what can be reasoned based on 

knowledge structure in ITM, we would like to define 

three concepts: knowledge path and knowledge radius. 

Definition 1: Knowledge path. In ITM, if there is a 

sequence e , e , e ,..., e , e , and there are association 

p 

1 2 

m q 

( , ) , ( , ) ,..., ( , ) 

between epe1 e1 e2 

em eq 

respectively in ITM, 

then we said that there exists a knowledge path between 

concept ep and e q . 

Definition 2: Knowledge radius. A knowledge path is a 

sequence of consecutive elements in ITM, and the 

knowledge radius is the minimum number of elements 

traversed in a knowledge path, i.e., the length of the path. 

KUCS is described as follows: 

r = 1 ; // r is knowledge radius 

for ∀t∈ T do //T is the set of topic 

if associationOf ( t _ po int, t ) =true then 

set _ T ⇐ t ; HashSet ⇐ t ; 

else 

set _ T ⇐ t ; 

end 

while r ≤ R do 

for ∀t h 

∈ HashSet do 

for ∀t∈ T do 

if associationOf ( t 

h 

, t) 

=true then 

set _ T ⇐ t ; HashSet1 ⇐ t ; 

end 

end 

r=r+1; HashSet = HashSet1 

; 

end 

for ∀t∈ set _ T do 

if associationOf ( t, ke ) =true then 

set _ KE ⇐ ke ; 

if associationOf ( t, c ) =true then 

set _ C = set _ C ∪ { c} 

; 

end 

ETM_building();


Through the structure-based knowledge reasoning, we 

can obtain all the knowledge elements, topics, cluster, 

and resource occurrence which are associated with the 

knowledge point within a certain knowledge radius. 


“VisualizationDisplay” is shown as follows: 

Based on the ITM logical representation of knowledge, 

the visual knowledge map constructing tool is designed, it 

is free software coded by Java applet, to assist users in 

sharing, and navigating the domain knowledge. The ITM 

document is visually displayed as a double-layer network, 

the schematic diagram is shown in Fig. 6. 

Figure 6. The schematic diagram of visual Knowledge map 

constructing. 

Clusters, topics and topic associations are represented 

in the upper layer in which fillet rectangular node is 

regarded as a topic. The dark node is regarded as the 

knowledge point. Each edge is regarded as an association 

of topics. When user clicking the edge, it will display the 

association type. Knowledge elements and their 

associations are in the lower layer in which ellipse node 

is regarded as a knowledge element. Each edge is 

regarded as an association of knowledge elements. When 

user clicking the edge, it will display the association type. 

When clicking the nodes in the knowledge element layer, 

it will display the occurrences which are associated with 

the knowledge element. 

V. EMPIRICAL EVALUATION 

A. The Experimental Data 

We built the corpus of Computer Network, which 

includes 34007 topics, 3307 knowledge elements, 4317 

associations between topics, 2214 associations between 

knowledge elements, 1872 associations between topic 

and knowledge element and 7031 domain-specific terms. 

B. The Logic Knowledge Reasoning Experiment 

We implement the knowledge consistency checking 

and the implicit relations reasoning experiment 

respectively. The knowledge consistency checking 

includes the reflexivity checking and loop transitivity 

checking, knowledge redundancy checking and 

contradiction checking. The implicit relations reasoning 


can discover the new associations between elements. The 

results are shown in Table 1. 

TABLE I. 

LOGIC KNOWLEDGE REASONING RESULTS 

Checking item Statistics 

Reflexivity checking 72 

Transitivity checking 216 

Redundancy checking 161 

Contradiction checking 19 

New associations 

New associations between topics 516 

New associations 

knowledge elements 

between 

312 

The main conflict type is transitivity conflict, which 

makes up 52% of total conflicts, knowledge redundancy 

conflict type makes up 34% of total conflicts, and 

knowledge reflexivity conflict and knowledge transitivity 

conflict make up 14% of total conflicts. Conflicts can be 

caused by many reasons. The ITM corpus construction is 

a process that needs many people’s collaboration and 

many times of revision, and the local ITM to be reused, 

they first need to be merged or aligned to one another to 

produce a single integrated and reconciled global ITM 

that deals with a larger domain of interest. In the process 

of building, conflicts can be caused by many reasons, so 

the consistency checking is a key component of 

knowledge reasoning strategy. The implicit relations 

reasoning can reason out new associations between topics 

(or knowledge elements), provide knowledge structure 

more detailed semantic association and provide inherent 

relevant characteristics of knowledge to constructing the 

complete knowledge structure, but we find that some 

reasoning relations between topics (or knowledge 

elements) are not tight enough. 

C. The Knowledge Structure Reasoning Experiment 

We select a topic “TCP/IP protocol” as knowledge 

point and different knowledge radius to carry out the 

structure-based knowledge reasoning experiment. It 

returns all the knowledge elements and topics which are 

associated with the knowledge point within a certain 

knowledge radius. The structure-based knowledge 

reasoning results is shown in Fig. 7. With the knowledge 

radius increasing, the number of topics, knowledge 

elements and relations continuously increase. When 

knowledge radius is equal to 2, the structure-based 

knowledge reasoning results include ten topics (such as 

“IP protocol”, “TCP/IP protocol”, “TCP protocol”, etc.) 

and twelve associations between the topics, six 

knowledge elements (“TCP protocol definition”, “IP 

protocol definition”, “TCP/IP protocol definition”, etc.) 

and five associations between the knowledge elements, 

and six relations between the topic and knowledge 

element. The knowledge structure is depicted in Fig. 8.


Figure 7. The structure-based knowledge reasoning results. 

VI. CONCLUSIONS 

The proposed visual knowledge structure reasoning 

model provides us a means to organize, discovery and 

display knowledge. Visual knowledge structure reasoning 

based on ITM not only achieves the better structure-based 

knowledge reasoning results and provides users with 

intuitive access mechanisms for the required knowledge. 

Knowledge has been provided by a stereo knowledge 

map and hence overcomes the shortcoming of linear 

display. The ongoing work is knowledge organization, 

knowledge search and knowledge reasoning can be 

carried out by computing cloud with huge computing 

ability and storage capacity distributed and parallel. We 

hope that the real visual knowledge structure reasoning 

system will be widely deployed in the future. 


This work is supported in part by Northeast Asia 

Chinese International Promotion Information Platform 

(Hanban). This work was also supported in part by the 

National High-Tech Research and Development Plan of 

China under Grant No. 2008AA01Z131. 

REFERENCES 

[1] J. van Merriënboer and P. Ayres, “Research on cognitive 

load theory and its design implications for e-learning,” 

Educational Technology Research and Development, vol. 

53, no. 3, pp. 5-13, 2005. 

[2] Q. Wang, L. Rong, and K. Yu, “Visual knowledge 

reasoning on typed categorical structure,” Proc. 5th 

International Conference on Fuzzy Systems and 

Knowledge Discovery (FSKD-08), pp. 684-688, 2008. 


[3] “ISO/IEC 13250 Topic Maps Second Edition,” 

Information Technology Document Description and 

Processing Languages, 19 May 2002. 

[4] “ISO/IEC JTC 1/SC 34, ISO/IEC 13250-2: Information 

Technology-Topic Maps-Part 2: Data Model,” 

http://www.isotopicmaps.org/sam/sam-model/datamodel.pdf, 

2008. 

[5] H. Lu and B. Feng, “The intelligent topic map-based multiresource 

knowledge services system,” Journal of 

Information and Computational Science, Binary 

Information Press, United States, vol. 7, no. 3, pp. 657-665, 

June 2010. 

[6] S. Pepper and G. Moore, “XML Topic Maps (XTM) 1.0,” 

http://www.topicmaps.org/xtm/1.0/index.html, 2002. 

[7] C. Baru, A. Gupta, B. Ludäscher, R. Marciano, Y. 

Papakonstantinou, P. Velikhov, and V. Chu, “XML-based 

information mediation with MIX,” ACM SIGMOD Record, 

vol. 28, pp. 597-599, 1999. 

[8] P. A. Silva, C. M. F. A. Ribeiro, and U. Schiel, 

“Formalizing ontology reconciliation techniques as a basis 

for meaningful mediation in service related tasks,” Proc. 

ACM first Ph.D. workshop in CIKM’07, pp. 147-154, 

2007. 

[9] J. L. Seng and I. L. Kong, “A schema and ontology-aided 

intelligent information integration,” Expert Systems with 

Applications, vol. 36, pp. 10538-10550, 2009. 

[10] S. Pepper, “The TAO of topic maps-finding the way in the 

age of infoglut,” 

http://www.gca.org/papers/xmleurope2000/papers/s11- 

01.html 

[11] H. Lu, B. Feng, and X. Chen, “Extended topic map: 

knowledge collaborative building for distributed 

knowledge resource,” Proceedings of the 12th IEEE/IFIP 

Network Operations and Management Symposium (NOMS 

2010), IEEE Communication Society, pp. 128-135, Osaka, 

Japan, April 2010. 

[12] A. Korthaus, M. Aleksy, and S. Henke, “A distributed 

knowledge management infrastructure based on a Topic 

Map grid,” International Journal of High Performance 

Computing and Networking, vol. 6, pp. 66-80, 2009. 

[13] K. Olsevicova, “Application of topic maps in e-learning 

environment,” ACM SIGCSE Bulletin, vol. 37, pp. 363, 

2005. 

[14] J. Qiu, Y. Yao, Y. Wang and X. Wang, “Research of egovernment 

knowledge navigation system based on 

XTM,” Proceedings of the 2006 IEEE/WIC/ACM 

International Conference on Web Intelligence and 

Intelligent Agent Technology, pp. 586-589, 2006. 

[15] M. Ouziri, “Semantic integration of web-based learning 

resources: a topic maps based approach,” Proceedings of 

the 6th International Conference on Advanced Learning 

Technologies, pp. 875-879, 2006. 

[16] H. Lu and B. Feng, “Intelligent topic map: a new approach 

to multi-resource knowledge service,” Proceedings of the 

2009 IEEE International Conference on Intelligent 

Computing and Intelligent Systems (ICIS 2009), IEEE 

Computer Society, pp. 373-377, November 2009. 

[17] H. Lu, B. Feng, and X. Li, “Novel Similarity Algorithm of 

Extended Topic Maps for Multi-Resource Knowledge 

Fusion,” Journal of Xi’an Jiaotong University, vol. 44, no. 

2, pp. 23-27, February 2010. 

Huimin Lu received the M.S. degree and Ph.D degree in the 

major of computer science and technology from Xi’an Jiaotong 

University, Xi’an, China, in 2005 and 2010, respectively. 

She is working in Changchun University of Technology, 

Changchun, China. She has published 15 papers in referred



journals and international conferences. Her 

current research interests include: 

knowledge science and knowledge 

engineering, topic map. 

Dr. Lu is the member of ACM, IEEE, 

IEICE and CCF. Now she is also working 

in the computer science and technology 

post-doctoral research center of Jilin 

University. 

Liang Hu is a professor at the college of 

computer science and technology, Jilin 

University, Changchun, China. His current 

research interests include: knowledge science, computer 

network. He has published more than fifty research articles in 

referred journals and international conferences. 

Gang Liu is a professor at the college of 

software engineering, Changchun 

University of Technology, Changchun, 

China. His current research interests 

include: software engineering, knowledge 

science. He has published 5 research 

articles in referred journals and 

international conferences.


An Automated X-corner Detection Algorithm 

(AXDA) 

Fuqing Zhao 

School of Computer and Communication,Lanzhou University of Technology,Lanzhou,Gansu,P.R.China 

Key Laboratory of Gansu Advanced Control for Industrial Processes 

Email: fzhao2000@hotmail.com 

Chunmiao Wei 


Email: h09310917@126.com 

Jizhe Wang 


Email: wangjizhe2009@mail2.lut.cn 

Jianxin Tang 


Email: tangjianxin2009@mail2.lut.cn 

Abstract—According to the central symmetry and brightdark 

alteration of the four peripheral regions at the Xcorner, 

an automated X-corner detection algorithm (AXDA) 

is presented to camera calibration problem. By detecting the 

gray changes of the image, the algorithm can locate the 

position of X-corner accurately using the minimum 

correlation coefficient of the symmetry regions. Cross points 

of intersection are calculated using the detection points and 

the least square straight line fitting algorithm. The method 

can not only realize the sub-pixel X-corner extraction, but 

also resolve the low automation degree problem of the 

present detection algorithm under complex background. 

Experiment results show that the algorithm is an easilyrealized, 

highly-automated and robust method for rotation 

transform and brightness transform of the X-corner image. 

Index Terms—chessboard, corner detection, camera 

calibration 


Camera calibration[1,2] based on the chessboard 

planar template consists of corner detection and 

parameter calculation. Calibration is to create the 

constraint between the world coordinates and the image 

coordinates of the chessboard feature points. So X-corner 

detection is the essential technology for camera 

calibration. 

At present, X-corner detection algorithm can be 

classified as follows. One is the traditional corner 

detection algorithm, such as Harris operator [3,4]and 

SUSAN [5,6]. These algorithms are versatility and 

Manuscript received December 15, 2010; accepted January 17, 2011. 

This work is financially supported by the National Natural Science 

Foundation of China under Grant No.61064011 


doi:10.4304/jsw.6.5.791-797 

straightforward but are invalid for the vagueness of the 

image. The idea of Harris corner detection is simple and 

the corners extracted by this algorithm distribute 

uniformly. Besides, this algorithm is insensitive to the 

rotation of the image and the alteration of the image gray. 

However, in the X-corner image, the image blurring in the 

corner area usually causes a high Harris respond value of 

one or several points, which makes it difficult to 

determine the position of the corner exactly. SUSAN 

operator is proposed by Smith and Brady. The operator is 

a completely different corner detection method, which 

mainly based on light contrast. Do not have to calculate 

image difference. The calculation speed is so fast that the 

SUSAN are widely used in the higher requirements of 

real-time areas. The other is the curve fitting based 

algorithm. Algorithms based on the two times Radon 

transform [7] are the representative ones. But there will 

be a great error if image distortion exists. And there are 

some other detection algorithms [8,9] aimed at the 

characteristics of the chessboard image. But [8] is 

computational expensive. Although [9] is straightforward 

in computation, great error will generate after 

binarization processing. All the above algorithms are 

liable to be influenced by light and background image in 

practical applications and will inevitably result in missing 

error, detection error or false error, which leads the 

automation detection hard to realize. Although the 

calibration toolbox in Matlab[10] resolves these problems, 

the automation level for camera calibration decreases 

greatly since four points should be selected manually 

before the corner extraction. 

An automated X-corner detection algorithm (AXDA) 

is provided by studying the present methods carefully. 

This algorithm can detect the corner using the unique 

feature of X-corner so as to avoid detection error, false


error as well as the influence of complex background. 

Meanwhile, detection result could achieve sub-pixel level 

with the help of least square straight line fitting and the 

missing error is resolved. This algorithm is 

straightforward, computational cheap and robust for 

rotation and gray alteration. Besides, it can resist the edge 

vagueness and significantly improve the automation level. 

II. X-CORNER DETECTION METHOD FOR CHESSBOARD 

IMAGE 

Chessboard image is shown as Fig. 1. Chessboard is 

made up of black and white square boxes, where the 

connection of two black boxes or two white boxes is the 

X-corner. In order to detect corners in complex 

background without false error, algorithm suitable for the 

characteristic of chessboard image should be used. Image 

at the X-corner is shown as Fig. 2. In Fig.2, there are two 

features. First, brightness of the four regions around the 

X-corner changes alternately. There are gray changes in 

the adjacent regions of part I、II、III and IV. Besides, 

they are light and dark alternation. Second, in the corner 

centered regular region, image is central symmetry. Xcorner 

can be detected accurately in complex background 

using the above two exclusive features of the X-corner. 

A. Corner Detection 

Chessboard image is easy to be affected by the 

brightness of the light and the angle of shooting in the 

time of acquisition, which proposed higher requirements 

for the automated detection algorithm. Taking into 

account various aspects, such as the impact of complex 

background on the chessboard image, it is required that 

the number of extracted corners is not excess. All the 

corners should be extracted with a certain accuracy. So 

the detection algorithm is necessarily required a certain 

Figure 1. Chessboard Image 

Figure 2. Features of Chessboard Image 


specific. In the absence of human intervention, the 

common corner detection algorithms can not meet the 

above requirements. To solve the above problem, we 

should choose the algorithm based on the characteristics 

of chessboard images. The algorithm is described as 

follows: 

Firstly, crude extract. Corners can be roughly extracted 

by using the first features. After the crude extract, the 

point at the region between two Haig rectangular can be 

obtained. 

Secondly, filter the results. Corners extracted by using 

above method should be filtered by using the second 

features, which aim to look for the regional center of 

symmetry as the corner. 

Finally, Corners should be classified by the row 

(column). The corners in the same row (column) are used 

to get curve fitting. The intersection can be calculated by 

the curve fitting. In this way, the final corners can be 

obtained. Further more, there is no missing error. 

B. Algorithm Implementation 

First, roughly extract the X-corner using the light and 

dark alternation characteristic of the four regions around 

the corner. Since there is gray difference between the 

adjacent regions, in the window W(7*7), gray summation 

of this region can be defined as follows: 

∑∑ 

FI ( ) = frc ( , ) x -3 < r< 

x y -3


corner can be selected. The correlation coefficient can be 

calculated as follows (Take a 20 ×20 window as an 

example): 

P= 

20 20 

∑∑ 

( f ( x−ry , −c) − η)( f ( x+ ry , + c) 

−η) 

X 0 0 1 X 0 0 2 

r= 1c= 1 

20 20 20 20 

2 2 

∑∑ ( fX( x0−ry , 0−c) − 1 η) ∑∑ ( f ( x0+ ry , 0+ c) 

−η2) 

X 

r= 1c= 1 r= 1c= 1 

� (6) 

Where (x0,y0) is the current detecting point. fx(x0-r,y0-c) is 

the gray value of (x0-r,y0-c)�, x∈{I,II}. X is the central 

symmetrical region of x. η1�and�η2 are the average gray 

values in the region. 

In the small region where real corner is included, only 

the real corner has the maximum correlation coefficient. 

So the extracted points can be filtered with this 

characteristic so as to get the real corner. In practical 

calculation, it only needs to compute half of the 

correlation coefficients. Namely, compute the region of I 

and III or II and VI is enough. 

C. Missing error solution and sub-pixel corner detection 

realization 

Corner detection algorithms are often influenced by 

image brightness and shooting angle in practical 

application, which would inevitably cause the missing 

error and is inconvenient for automated calibration. 

Besides, in order to pursuit more accurate camera 

calibration, sub-pixel corner coordinates are required. 

Aimed at the above problems, first of all, pattern 

classification is implemented on the extracted points to 

classify those points by rows and columns. Points in the 

same row or in the same column are fitted by least square 

method to solve the intersection point. This approach can 

avoid the missing error and achieve the sub-pixel 

precision corner detection. Meanwhile, it lays the 

foundation for resolving the coefficient automatically in 

the next step since the sequence information of corner 

position is included in the intersection point. It should be 

noted that if there is a strict requirement on precision, 

points in the outside rows and columns should be rejected 

before straight line fitting since they would sometimes 

cause the performance decline of the detection due to 

image distortion. In addition to the above method to 

improve the precision of the corners to sub-pixel level, 

there are also other methods. The most common method 

is gray-scale interpolation and then extracts corners on 

sub-pixel level using the algorithm provided by this paper. 

Another method is proposed in [11]. The main idea is that 

the vector starting from the real corners (q) to any other 

adjacent points (pi) is vertical to the gray gradient at the 

point pi. 

T 

ε =∇() I .( q− p) 

(7) 

i pi i 

Where ▽(I)pi T is the gray gradient at the point pi . εi 

should be zero in theory, but because of the noise, εi may 

not be zero. Therefore, q is the point where εi gets the 


minimum value. We build the iterative function according 

to (7) and calculate the coordinate of the corner on subpixel 

level. Results show that this is a precise algorithm. 

The whole algorithm is as shown in Fig.3. 

Where A is the set, which contains the points detected 

by using the first character. B is the set that contains the 

final pixel level corners. W is the detection window. C is 

the set that contains the final sub-pixel corners. 

III. EXPERIMENTAL ANALYSIS 

In order to test the correctness of the above algorithm, 

a series of experiments were conducted. Images involved 

in the experiments can be obtained on the website of 

http://www.vision.caltech.edu/bouguetj/calib_doc/htmls/c 

alib_example/index.html. Image resolution is 640×480. 

The chessboard template consists of 14×13 boxes, where 

156 internal corners are included. The size of each box is 

30mm×30 mm. 

Figure 3. Program Flow Chart


Figure 4. Image fragment 

Figure 6. Pixel-level image and results 

First, process Fig. 4 using the first method mentioned in 

II where T is equal to 400.The results are marked with red 

box as shown in Fig. 5.. It can be seen that these points are 

distributed in a 3 × 3 small areas. 

Second, we use the second method in II to filter the 

above results. The correlation coefficient of points 

between I and III is shown in Table 1. 

From Table I we can see that the maximum value of the 

correlation coefficient is at (16,15). So we define (16,15) 

as the corner we want. In Fig. 6, the red cross indicates the 

final extracted result. It show that algorithm is correct and 

has a high accuracy 


A B C 

Figure 7. Results extracted by AXDA 

TABLE I. 

EXTRACTED RESULTS 

Point Extracted Results Corner 

Coordinate (15,14) (16,14) (16,15) (17,16) 

P 0.1632 -0.1085 0.6497 0.4043 

Figure 8. Extracted Result 

(16,15) 

A. Comparison with Harris corner detection algorithm 

Red labels in Fig. 7, Fig. 8, Fig. 9and Fig. 10 are 

detection results processed by AXDA and Harris 

operator respectively. Numbers of detected corners are 

listed in Table II. 

From Fig. 7, Fig.8 and Table II we can see that the 

performance of the AXDA is much better than that of the 

Harris operator. The AXDA detects only one corner 

outside the chessboard. As the general algorithm however, 

Harris detects large numbers of non-X-corners in the 

background. These points distribute disorderly and will 

cause great trouble in the automated calibration.


A B C 

Figure 9. Results extracted by Harris 

Fig. 9 and Fig. 10 are the local enlargement result 

images of AXDA and Harris operator respectively. It can 

be seen clearly that the performance of the AXDA 

outweighs the Harris operator. Corners detected by 

Harris in Fig. 9 shifts from the real position and false 

Figure 10. Results extracted by AXDA 

corners are detected. But in Fig. 10 the results are more 

accurate. 

Figure 8. Results extracted by Harris 

TABLE II. 

NUMBER OF POINTS 

Number of Corners 

Algorithm 

Image A Image B Image C 

Harris 443 382 393 

AXDA 157 156 156 


Corners AXDA 

TABLE III. 

SUB-PIXEL CORNERS 

Camera Calibration Toolbox 

for Matlab 

1st (287.5431,116.2525) (287.1477,114.9688) 

2nd (282.3183,152.3035) (281.4477,151.0909) 

3rd (277.3627,186.4973) (276.3348,185.5046) 

4th (272.6387,219.0928) (271.3617,217.9590) 

5th (268.1781,249.8713) (267.0234,248.7857) 

6th (264.0387,278.4330) (262.5504,277.8376) 

7th (260.0439,305.9974) (258.6074,305.3773) 

8th (256.3077,331.7771) (254.5650,331.1984) 

9th (252.7597,356.2584) (251.3006,355.5708) 

10th (249.4266,379.2563) (247.6796,378.6948) 

11th (246.2785,400.9786) (244.7730,400.3725) 

12th (243.3301,421.3226) (241.7335,420.8897) 

B. Precision comparison on sub-pixel level 

Corners detected by intersection of straight line fitting 

and corners detected by Matlab calibration toolbox are 

compared on the sub-pixel level. Take the detection 

results of the fourth column in Fig.7.A as an example, 

comparisons from the top to the bottom are shown as 

Fig.11 and Tab.III.


Figure 11. Image fragment Sub-pixel level results contrast 

In Table III, the sub-pixel corner coordinates, detected 

by AXDA and Camera Calibration Toolbox for Matlab 

were listed. There are about 1-2 pixels difference in the 

results. Those points are projected on the test image, 

shown as in Fig.11.In Fig. 11, red corners are extracted 

by the AXDA and green corners are detected by the 

Matlab calibration toolbox. It can be seen from Fig. 11 

that compared with green labels, red labels marked more 

accurately on the connection of two black boxes. So the 

performance of the AXDA is much better than that of the 

Matlab calibration toolbox. Besides, the AXDA does not 

need human intervention and has a higher degree of 

automation. 

To sum up, it is show that the proposed algorithm can 

not only detected corners accurately, but also has a low 

algorithm complexity .In the process of coarse 


extraction, which uses the first feature, a lot of points are 

excluded. The computation can be reduced effectively in 

the subsequent Essence extracted. In addition to, it is 

feasible for obtaining the missing corners. Furthermore, 

the proposed algorithm can automatic acquisition corner 

position sequence information, which can provide 

coordinates sequence information for further automatic 

camera calibration. 

IV. CONCLUSIONS 

By analyzing the current X-corner detection 

algorithms, a new X-corner extraction approach is 

provided using the characteristic of the central symmetry 

of the gray image as well as the characteristic of the 

bright and dark alteration of the four regions around 

corner. First of all, roughly extract the corners using the 

characteristic of the bright and dark alteration of the four 

regions around corner. Second, accurately extract the 

coordinates of the corner with the symmetry 

characteristic. Then, classify the extracted corners by 

rows and columns. Fit the points with least square 

straight line fitting algorithm. Calculate the intersection 

of the fitting straight lines and finally obtain the precise 

corners. The proposed method effectively settles the 

false error and missing error in practical application and 

is robust for rotation transformation and brightness 

variation. Besides, our algorithm is computational cheap 

and has a high degree of automation, which is beneficial 

to the real-time camera calibration. 


This work is financially supported by the National 

Natural Science Foundation of China under Grant 

No.61064011. And it was also supported by China 

Postdoctoral Science Foundation, Science Foundation for 

The Excellent Youth Scholars of Lanzhou University of 

Technology, and Educational Commission of Gansu 

Province of China under Grant No.20100470088, 

1014ZCX017 and 1014ZTC090, respectively. 

REFERENCES. 

[1] G.J.Zhang, “Machine vision,” Beijing: Science press,pp. 

35-55, 2005. 

[2] Z.Zhang, “A flexible new technique for camera 

calibration,”IEEE Transactions on Pattem Analysis and 

Machine Intelligence, 22(11),pp.1330-1334,2000.doi: 

10.1109/34.888718. 

[3] C.Harris , M.Stephens, “A combined comer and edge 

detector,”Proceedings of the 4th Alvey Vision Conference, 

Manchester, pp.l47-l51,1988.doi:10.1016/SO262- 

8856(97) 

0010-3. 

[4] D.H.Xie,Z.Q.Zhan,W.SH.Jang, “Improving Harris Corner 

Detection,” Journal of Geomatics, vol.28(2),pp.22-23, 

2003.doi:1007-3817(2003)02-0022-02. 

[5] Q.He,Q.H.Li,X.S.Wang, “Corner Features Extraetion of 

Image Based on 0rientation SUSAN 0perator,” Journal of 

Chinese Computer Systems, vol.29(3),pp.508-510. 

2008.doi: 1000-1220(2008)03-0508-03.


[6] H.X.Lv,H.LCH, “On angular—point extracting method 

based on SUSAN and point matching method,”Electronics 

Optics&Control,2008,vol.15(3),pp.45-48.doi:1671-637X( 

2008)03-0045-04. 

[7] H.F Hu, Y.G. Xiong, “A new algorithm for chessboard 

grid corners detection based on two successive 

radonTransform,” Acta Scientirarum Naturalium 

Universitatis Suny Atseni,vol.42(2), pp.23-26, 2003. doi:0 

529-6579(2003)02-0023-04. 

[8] Y. L,F.L. Wang, Y.Q.Chang, “Black and White X-Corner 

Detection Algorithm,” Journal of Northeastern University 

(NaturalScience), vol.28(8),pp.1090-1093,2007.doi: 1005 

-3026(2007)08-1090-04. 

[9] X.J. Tan , Z.H. Guo, Z. Jiang,“Chessboard grid corners 

detection based on geometric symmetry,” Journal of 

Computer Applications , vol.28(6), pp.1540-1542. 

2008,doi: 1001-9081(2008)06-1540-03 

[10] http://www.vision.caltech.edu/bouguetj/calib_doc/. 

[11] Z.M. Liang,H.M.Gao,Z.J.Wang,L.Wang, “Sub-pixels 

corner detection for camera calibration,”Transactions of 

The China Welding Institution, vol.27(2), pp.102-104, 

2006. doi:0253-360x(2006)02-102-03. 


Fuqing Zhao P.h.D., born in Gansu, China, 1977, has got a 

P.h.D. in Dynamic Holonic Manufacturing System, Lanzhou 

University of Technology, Gansu, 2006. He is a Post Doctor in 

Control Theory and Engineering in Xi’an Jiaotong University 

and Visiting Professor of Exeter University. His research work 

includes theory and application of pattern recognition, 

computational Intelligence and its application, where fifteen 

published articles can be found. 

Chunmiao Wei born in Shanxi,China,1984.His research 

interest is the application of pattern recognition, Graphics and 

Image Processing , Computer Vision 

Jizhe Wang born in Henan,China,1986.His research interest 

is the application of pattern recognition, Artificial Intelligence 

Jianxin Tang born in Henan, China, 1985.His research is the 

theory and application of pattern recognition.


Research on Dynamic Rescheduling Program 

Base on Improved Contract Net Protocol 

Fuqing Zhao 

School of Computer and Communication, Lanzhou University of Technology, Lanzhou, Gansu, P.R.China 

Key Laboratory of Gansu Advanced Control for Industrial ProcessesLanzhou, Gansu, P.R.China 

Email: fzhao2000@hotmail.com 

Jizhe Wang 

School of Computer and Communication, Lanzhou University of Technology,Lanzhou, Gansu, P.R.China 

Email: wangjizhe2009@mail2.lut.cn,tangjianxin2009@mail2.lut.cn 

Jianxin Tang 

School of Computer and Communication, Lanzhou University of Technology,Lanzhou, Gansu, P.R.China 

Email: tangjianxin2009@mail2.lut.cn 

Abstract— Dynamic rescheduling of workshop production 

management, with the feature of combinatorial computation 

complexity, is an important and difficult research area, and 

be of significant importance for the dynamic scheduling 

problem. An improved Contract Net Protocol (CNP) with 

the global two-way, Multi-Agent System (MAS) based 

communication model, which incorporated the local 

autonomy of working mutually in consultation by 

negotiation, is presented in this paper. Furthermore, the 

simulation results in dynamic scheduling accompanying 

with its perturbation show that the proposed model and the 

algorithm are effective to the dynamic scheduling problem 

in manufacturing system. 

Index Terms—MAS, Agent, dynamic scheduling, Contract 

Net Protocol 


Present industrial system forward in the direction of 

large complex dynamic changes, Traditional industrial 

systems and technology in a number of key issues are 

serious challenges. Efficient and practical method which 

is used in scheduling and optimization technology is a 

key to plant productivity [1]. Assume that the traditional 

process with a clear schedule and a fixed processing time, 

while the actual processing, there are many uncertain 

factors, for example, changes in processing time, product 

demand, delivery, equipment failure, resources and 

production processes. The dynamic interference of these 

factors make the original dynamic scheduling can not be 

implemented successfully. The rescheduling is occurring 

in the course of events and uncertain response to other 

changes, which is based on the state of the system time 

Manuscript received December 1, 2010; revised January 15, 2011; 

accepted January 17, 2011. 

This work is financially supported by the National Natural Science 

Foundation of China under Grant No.61064011. 


doi:10.4304/jsw.6.5.798-805 

for the next cycle of program activities; therefore it has a 

very high search value [2]. 

Job shop scheduling is a NP-hard problem, Church LK 

[3] studied the rescheduling driven based on cyclical time 

and the two rounds of re-scheduling-driven methods. Jian 

Fang [4] designed the Ministry of Information under full 

dynamic scheduling model rolling horizon procedure, and 

then had analyzed and evaluated. Sanjay V [5] processed 

the way to absorb the random failure of the disturbance 

proposed by the appropriate method of inserting idle time. 

Kim [6] proposed a flexible production environment 

which can handle processing of planning and shop 

scheduling symbiotic genetic algorithm. D.Petrovic [7] 

used the fuzzy method to study the re-scheduling of the 

start. Cheng [8] proposed scheduling based on genetic 

algorithm concept of real-time response. Wang Hui [9] 

put the uncertainty of the impact of events designed as a 

set of random changes in the time period. Goncalves [10] 

proposed a hybrid genetic algorithm for shop scheduling. 

Wong [11] designed a device based on the production of 

multi-Agent system. 

Based Agent which has a high degree of autonomy, in 

this paper, a group of self-government body (Agent) was 

used to solve the effective coordination between the 

complex and dynamic rescheduling problem, and Agent 

in the process of scheduling, we use consultation and 

collaboration between scheduling, therefore have a higher 

real-time [12]. This is the same need in the dynamic shop 

scheduling. So this paper from the Agent communication 

mechanism of individual multi-Agent cooperative 

mechanisms and MAS's system of institutions to start; 

combine the re-scheduling problem of dynamic workshop 

is described in detail. Propose a global two-way dispatch, 

local self-improvement contract net protocol negotiation 

and simulate, then confirm its effectiveness. 

II. MAS-BASED MODEL RESCHEDULING


In dynamic shop scheduling environment, job shop 

problem can be described as: In a processing unit or 

system, n jobs need to be processed on m machines, 

every job i J (1 ≤i ≤ n ) has i n process ij O (1 ≤i ≤ n,1≤j 

≤ ni 

) 

need to processing, Set machine tool with a collection 

of M , Then each process ij O either by the concentration 

of machine tools ij M can be processed in a machine, 

where M ij M ⊆ . If M ij M = , the scheduling problem is a 

completely flexible scheduling problem; if M ij M ⊂ , it is a 

local scheduling problem with flexible [13] . 

Re-scheduling operation set is a machine failure occurs, 

all machines need to re-scheduling of the operation on the 

set. Rescheduling operation set is essentially a set of 

variables constraint satisfaction problem. 

As rescheduling model is the corresponding evolution 

of the initial scheduling model, so the initial problem 

modeling available: 

min max{ cis | i∈ I} 

S.t. sij + 1 ≥ sij+ 1 + pij 

, i∈I, J ∈{1,..., s− 

1} 

(1) 

( mi1j≠mi) ( ) 

2 j ∨ si1j≥ci1j∨si2 j≥ ci2 

j 

(2) 

i1, i2 ∈I, i1 ≠i2, j∈ J 

(3) 

cij = sij + pij 

, i∈I, j∈ J 

(4) 

j−1 

si ≥0, s , , {2,..., } 

1 ij ≥∑ p 

k 1 ik i∈I j∈ s 

(5) 

= 

mij ∈ Rj= { rjl,... rjl | j∈J}, i∈ I 

j 

(6) 

i means workpiece number and i∈ I = {1,..., n} 

, j 

means level number and j∈ J = {1,..., s} 

, jl r means the 

machine number, ij s means the start time of initial 

scheduling, ij m means the start machine, ij p means the 

processing time of the operate workpiece. 

In the above model, (1) shows the optimization goal of 

the scheduling problem is minimum of C max . 

(2) shows the operation of the timing constraints, it is 

said that the workpiece after the end of the previous stage 

to begin the next stage of processing tasks.(3)shows that 

If the two jobs processed on the same machine, then the 

can not doing at the same time.(4)shows processing the 

workpiece can not be interrupted after 

starting.(5)and(6)shows operation started variable time 

and the variable range of processing machinery. 

Then it supposes the machine r jl d d disruptions at the 

time [ tb, t e] 

,so the initial scheduling begins to have a 

change in b t ,and the initial scheduling will change to the 

dynamic rescheduling, 

∑∑wijδ1ij ∑∑vijδ2ij 

i∈I j∈J i∈I j∈J max f = + 

w v 

∑∑ ∑∑ (7) 

ij ij 

i∈I j∈J i∈I j∈J ( m ≠m ) ∨( s ≥c ∨s ≥ c ) 

S.t. ij 1 ij 2 ij 1 ij 1 ij 2 ij 2 

i1, i2 ∈I, i1 ≠i2, j∈ J 

cij = sij + pij 

, i∈I, j∈ J 

m ∈ R = { r ,... r | j∈J}, i∈ I 

ij j j1 jl j 

( m′ ij ≠ rj) ( ), , 

dl ∨ s′ d ij ≥tei∈I j∈ J 

(8) 

s′ ij ≥tb, i∈I, j∈ J 

(9) 


(7) shows the scheduling programs to maximize the 

time before and after adjustment arrangement and the 

total weight assigned to the similarity machine. δ1ij shows 

rescheduling operation of The similarity of the timing 

with workpiece ij o . 

max{min{ c′ ij , cij} − max{ s′ ij , sij},0} 

δ1ij 

= 

pij 

δ 2ij shows the dynamic rescheduling with Before and 

after the operation of the dynamic rescheduling to assign 

the similarity in the workpiece ij o . 

⎧⎪ 1, m′ ij ∈ Mij; 

δ 2ij 

= ⎨ 

⎪⎩ 0, other 

(8) shows the new constraints of mechanical 

failures,(9)shows the beginning of the operation of the 

new range of variable start time. ij o means the stage j of 

workpiece i , wij , v ij mean operating weight of the 

workpiece and the machine time consistency of weight 

respectively, ij s′ means the operating parts of the starting 

time in rescheduling, ij m′ means the machine of operating 

workpiece in rescheduling. 

For the rescheduling problem, the structure of Agent 

can be expressed as: Agent= def < Id, Goal, Act, Rule, L >. 

Agent Id is the identifier which, different Id in 

different Agent. Re-scheduling of the workshop can be 

the Agent of Id from 1 corresponds to the location of the 

machine in the list. The Agent can be used to express 

as agi which Id is i . 

Goal is the Agent of the goal, the goal is that the Agent 

inserts after the current job is still making the current 

optimal or near optimal job queue. The goal can be 

i i i i 

expressed as Goali = ( Cmax, J1 → J 2... 

→ J ni) 

, where 

i i i { J1, J2... J ni } is the machine M i currently operating the 

current sort order queue 1 2 ... 

i i i 

J → J → Jni 

and 

set. 1 2 ... 

i i i 

J → J → J is the priority of the current sequence 

ni 

i 

set on the machine or near optimal order. Cmax is the 

optimal value of the machine M i or similar to the 

corresponding optimal value. 

Act can be said to the action set, in the form on behalf 

of the Agent Act = { act1, act2...... actn} 

can be to complete the 

operation. Each Agent has a communication, 

collaboration features. 

Rule represents Agent on behalf of Agent cooperation 

with other rule sets, In this paper we use the modified 

contract net protocol . 

L is the Agent communication language, different 

Agent use the languages to communicate with L , In this 

paper the rules based on FIPA ACL language. 

III. DYNAMIC SCHEDULING SYSTEM SASED ON MAS 

MODEL 

A.Functional Design Agent 

Traditional rescheduling is generally aided by the 

manual or has operations in accordance with certain re-


allocation algorithm [14]. This paper uses the MAS-based 

intelligent scheduling system Agent mainly through 

collaboration between the nature of making intelligent 

machines, in order to achieve the automation of job rescheduling 

and optimization. The basic structure of 

improved contract net model is unchanged, by the 

Management Agent, Resource Agent, Supervision Agent 

and Work piece Agent composition. 

Management Agent (MA) Management Agent is the 

core of the scheduling system, mainly responsible for 

evaluating and scheduling the task which mandate outside 

received. Specific tasks include the host information and 

the degree of emergency. Then put the information 

submit to the Resource Agent. Management Agent and 

other Agents’ relationship are shown in Fig.1. 

Figure1. Description of Management Agent 

Resource Agent (RA) Resource Agent is responsible 

for receiving and processing plant outside production 

tasks and in accordance with the current processing 

capacity, to determine whether to perform the task 

workshop. In the decomposition of tasks, each Equipment 

Agent releases to the tender, accordance with the rules of 

the agreement to form processing program, and then 

reports to the Supervision Agent, to obtain feedback on 

the various parts after the Agent is responsible for 

scheduling production. Resource Agent internal 

schematic is shown in Fig. 2. 

Figure2. Resource Agent internal schematic. 

Supervision Agent(SA) Supervision Agent mainly 

reports on alternative production plan of Management 

Agent for a simulation, then selects processing route back 

to the Management Agent to comply specifically. And the 

Supervision Agent mainly is responsible for the 

supervision of Agent equipment failure, the addition of 

new equipment and the arrival of other emergency tasks. 

Fig. 3 shows the internal schematic. 


Figure3. Supervision Agent internal schematic. 

Equipment Agent (EA) Actually it can be considered 

as a manufacturing unit. Each Agent self-management of 

each piece of unit, responsible for the appropriate 

operation management, equipment, command transfer 

and information collection. Equipment Agent receives the 

information after Resource Agent, products equipment on 

their assessment of the corresponding , then decide 

whether to tender. According to the equipment cases it 

makes a corresponding quote if tender, feedback on 

whether the production capacity to the Resources Agent 

to complete the task. Internal schematic shows in Fig. 4. 

Figure4. Equipment Agent Internal Schematic. 

Then, the Management Agent send a message to 

Workpiece Agent with the communication primitives 

sample can be expressed as: 

:Sender(managerAgent@abc:1099/jade) 

:Receiver(Equipment@abc:1099/jade) 

:Ontology AMS-ontology 

:Protocol FIPA-contract-net 

:Language FIFA-KQML 

:Content "((Issue (taskid(01),surface 

Type(plane),machining 

Type(drilling),number(8),tolerance(geometic 

Tol:02dimensional tol:01roughness:02), 

deadline(2010.12.01/21:10)))" 

Resources Agent releases from the processing of 

waiting tasks, select processing tasks in sequence, 

according to the form of tender to manufacture parts of 

the process for the Workpiece Agent with issuing the 

request, the communication primitives can be expressed 

as: 

( 

CFP


:Sender(Agent-identifier:name 

resource@abc:abc:1099/jade) 

:Receiver(Agent-identifier:name 

equipment@abc:abc:1099/jade) 

:Content(action 

issues:issuebook:taskli\task01:working procedure\01\) 

:task ready time"2010-12-01 21:20" 

:surface roughness 4:dimensional tolerance 

time\"60"\:deadline\"2010-12-01 21:20"\) 

:Reply-with CFP1 

:in Reply-with PROPOSE1 

:Language FIPA-KQML 

:Ontology scheduling ontology 

:Protocol fipa-contract-net 

) 

According to their capacity and status of the request, 

Workpiece Agent is in a given period of time which gives 

the proposed tender. Agent for the tender parts request 

primitives can be expressed as: 

( 

PROPOSE 

:Sender(Agent-identifier:name 

equipment@abc:1099/jade) 

:Receiver(Agent-identifier:name 

resource@abc:1099/jade) 

:Content"((action(bidbook(bidbook 

:finishtime\2010-12-01\21:30\)):cost:10:equipment 

(Agent-identifier:name equipment@abc:1099/jade)))" 

:Reply-with CFP1 

:in Reply-with PROPOSE1 

:Language FIPA-KQML 

:Ontology scheduling ontology 

:Protocol fipa-contract-net 

) 

B.The contract net protocol based on the improved 

process of rescheduling 

In the planed internal allocation model, Management 

Agent generates the appropriate contract under the task 

order, the final bidding through the contract net protocol 

mechanism to determine the distribution relationship [15]. 

But by given the efficiency of consultation and workshop 

frequent dynamic scheduling, to improve efficiency, the 

global scheduling use a two-way consultation mechanism. 

The workshop is no longer accepted management's 

bidding information passively. It can take the initiative to 

inform the Management Agent on free time, and to have 

rescheduling with Resource Agent and Equipment Agent. 

Shorten the time required for scheduling. Resource Agent 

is no longer the same time with the broadcast model of 

unconditional tender information published to the 

workshop, they test whether Agent scheduling 

applications have been submitted firstly, and then bidding 

between these application workshops, it means Invitation 

to bid model. Through this two-way consultation 

mechanism, the system is greatly reduced communication, 

negotiation efficiency also improved. It is shown in Fig.5. 

In this paper, Scheduling in the local autonomous 

negotiation strategy is used. It focuses primarily on a 

single operating part of the consultation process. 

Management Agent access to the task, the state machine 


select Agent in a particular queue. By the time they run 

the task initiated by notice to select the appropriate Agent 

to negotiate on its mandate. If access to the task at the 

same time, the launch of negotiations on a random 

selection. When the machine authorization of the Agent 

and executes for the task, the machine first notify the 

current Management Agent has completed the task, and 

then update their state, while awaiting transfer to the next 

stage of the job queue. On the other hand, work piece 

Agent change them idle. 

Reschedule for emergency orders : Due to market 

dynamics, new orders appear frequently. at that time the 

Resource Agent running in the system, first use the 

conventional approach to internal Management Agent to 

launch negotiation, if it can not find the time for new 

entrants to the scheduling order, then Resource Agent 

release some of the production plan, while the delivery of 

these orders guarantee to be completed before delivery, 

the scheduling of the Resource Agent released until the 

successful operation of emergency orders, which the 

released orders will reschedule after the emergency 

scheduling. The flow chart is shown on the left of Fig6. 

Failure of the machine: For the failure of the machine, 

immediately terminate the operation, and then issue a 

notice in need of repair, the Equipment Agent timely 

processing the state feedback to the Resource Agent, 

Resource Agent records the current processing situation, 

the task then to be processed back and see if there are 

other Agent can instead of the Equipment Agent, if you 

can replace, then the task will be distributed back out; if 

not, the task is to re-bid, re-scheduling. The flow chart is 

shown in middle of Fig. 6. 

For other exceptions: such as the shortage of raw 

materials, the task can not be completed in the near 

future, Management Agent will recover the 

corresponding tasks, so ahead of the back scheduling, the 

unfinished task will schedule again after the input 

processing conditions. Thus, the autonomy of local 

consultation can be well on the impact of the whole 

system on the strategy eliminating rescheduling. The flow 

Figure5. Dynamic Scheduling Model Based on Multi Agent. 

chart is shown on right of Fig.6.


Resource Agent 

Management Agent 

Find new time 

Y 

Workpiece Agent 

Release order 

Y 

Scheduling released order 

Planned order 

Emergency order 

Report 

Emergency scheduling order 

N 

N 

Stop jobs 

Other Agent 

Malfunction 

Notice 


feedback 

Resource Agent 

record 

Recycling order 

Process failure order 

IV. BASIC SCHEDULING ALGORITHMS IN AGENT 

The dynamic re-scheduling based on MAS composes 

of local scheduling by a multi-stage process. The local 

scheduling of each stage is carried out based on CNP 

model, the basic algorithm is as follows: 

Step 1 Scheduling information received from the 

outside world after the initial Management Agent release 

price PR i . It can be defined: PRi = ( ti Ta Ba Ma) 

t i Means that other Agent where the deadline to 

respond Ta means that the time required to complete the 

task constraints, Ba means space constraints, M a means 

material constraints. 

Then for emergency insertion of the work piece, once 

the original work piece delay, the delay time should to be 

as short as possible, so the time constraint can be 

expressed as: 

Ta = min[( Ts + ti),( Td + Tp)] 

(10) 

S.t. cik − pik + M(1 −aihk ) ≥ cih 

, 

i = 1,2,... n; h, k = 1,2,... m 

cjk − cik + M(1 −xijk ) ≥ pjk 

, 

i, j = 1,2,... n; k = 1,2,... m 

cik ≥ 0 , 

i = 1, 2,... n; k = 1, 2,... m 

find 

Y 

Planned order 

Figure6. Dynamic scheduling flow chart 


N 


Recover task 

Material shortage 

feedback 

Management Agent 

Y 

Planned Order 

N 

Other Agent 

x ijk =0 or 1, 

i, j = 1,2,... n; k = 1,2,... m 

In which (10) indicates time constraints. s T means the 

time that Management Agent make the initial offer, 

Td means the latest time that jobs end, p T means the 

average of extend operating time, which ik c and ik p mean 

that the finished time and the processing time of work 

piece i and machine k . aihk and xijk are coefficient and the 

indicator variable indicating. 

Step 2 Equipment Agent tender offer are given 

counter offer, Equipment Agent first assess their own 

parts to meet the resource constraints, and then give 

counter offer PRj = ( aj Tc, Mc) 

, a j means the commitment 

wait time, Tc means the first beginning time that is 

produced by the Equipment Agent which after assess, if 

the Equipment Agent do not meet a T , a B , M a ,or 

occupied by either the state constraints, then give up 

bidding. If T c is idle, then the Equipment Agent initiate to 

Resource Agent for counter-bids which is in the idle time 

scheduling, to save the scheduling time. 

Step 3 Resource Agent assesses the counter offer and 

then authorize. Management Agent evaluates all bids 

which returned, according to the formula: 

min[( Tc + Tp), Mc] 

.Select the best Equipment Agent to 

authorize, that is, considering the earliest start time of the 

Equipment Agent, the workpiece capacity and efficiency. 

Step 4 Perform an operation process. The Equipment 

Agent authorizes to perform job tasks, in the process of 

failure may occur. If normal, while the Management 

Agent of total consumption statistics, and then back to the 

Equipment Agent. If failure occurs, report to the 

Supervision Agent, to stop the operation and into the fault 

repair process of negotiation. 

V. SIMULATION RESULTS 

For example, to a machine shop, considering the 

equipment of 5parts, 4processes and 8machines; the 

workshop is to complete planning, milling, turning, 

drilling and other processes. There are two multifunctional 

machines: two different specifications of the 

plan and two specifications different lathe, a milling 

machine and a multi-function machine tool. Relationship 

between process machines is shown in Table I(in the 

table, 1 indicates that the machine can complete the 

process, 0 not).Process sets of the workpiece are shown in 

Table II. The processing time is shown in Table III. 

Consider two cases, one piece of the delivery is not 

particularly tense situation, is set to FIFO scheduling the 

delivery of products under the rules of the average 

processing cycle, the other is a more intense delivery time, 

that is 1:1.2. MAS proposed methods will be used in the 

FIFO rules and EDD rules performance comparison, the 

simulation results are shown in Table IV.


ID 

Machine 

Process 

As can be seen from the table IV, Fig.7 and Fig.8(G- 

FIFO means General FIFO;G-EDD means General 

EDD;G-MAS means General MAS;E-FIFO means 

Emergency FIFO;E-MAS means Emergency MAS;G-F- 

FIFO means General Failure FIFO;G-F-MAS means 

General Failure MAS;E-F-FIFO means Emergency 

Failure FIFO;E-F-MAS means Emergency Failure MAS), 

using the proposed consultation mechanism from this 

paper, for reducing the weighted average delay in 

delivery of products, improve product time delivery, has 

an extremely effective results. 

For the problem of equipment failure, assuming a 

daily equipment failure 12h, simulation time for one 

TABLE II. 

EQUIPMENT AND PROCESS 

Piece ID number 1 procedure 2 procedure 3 procedure 4 procedure 

j1 Plane(id:1) Milling (id:2) Diamond (id:4) Car (id:3) 

j2 Car (id:3) Diamond (id:4) Milling (id:2) Plane (id:1) 

j3 Milling (id:2) Plane (id:1) Car (id:3) Diamond (id:4) 

j4 Diamond (id:4) Plane (id:1) Milling (id:2) Car (id:3) 

j5 Milling (id:2) Diamond (id:4) Plane (id:1) Car (id:3) 

TABLE III. 

PROCEDURE PROCESSING TIME 

Machine 

Plane Milling Car Diamond 

A B C C G D E F H 

j1 3 4 6 10 8 2 3 5 9 

j2 7 8 8 5 6 6 7 6 4 

j3 2 3 3 3 5 4 5 3 3 

j4 8 9 9 2 7 11 12 7 7 

j5 3 5 6 5 3 5 6 8 6 

TABLE IV. 

MODEL SIMULATION RESULTS 

Production Line Scheduling 

Time delivery/% Weighted average delay/h 

Status 

rules A B C D E F G H A B C D E F G H 

FIFO 49 48 51 57 50 50 50 51 21 18 18 16 16 20 15 16 

General 

EDD 46 54 55 61 65 50 52 68 

MAS 85 91 94 88 90 84 93 80 3 3 1 1 8 7 9 8 

Emergency FIFO 0 0 0 0 0 0 0 0 55 84 65 59 71 65 84 74 

MAS 100 94 85 89 98 89 92 94 5 7 5 3 10 6 8 6 

Production Line 

Status 

A(Planer1) 

B(Planer2) 

Scheduling rules 

TABLE I. 

RELATIONSHIP BETWEEN THE MACHINE AND PROCESS 

C(Milling) 

D(Lathe1) 

E(Lathe2) 

TABLE V. 

EQUIPMENT FAILURE MODEL SIMULATION 

F(Multifunction) 

G(Lathe) 

H(Multifunction) 

P1 1 1 0 1 0 1 1 0 

P2 0 0 1 0 0 0 0 1 

P3 0 0 0 1 1 0 1 0 

P4 0 0 1 0 0 1 1 0 

Time delivery /% Weighted average delay /h 

A B C D E F G H A B C D E F G H 

General Failure FIFO 55 58 51 55 54 59 50 49 41 20 15 17 19 21 19 20 

MAS 96 92 92 95 90 95 99 94 6 8 3 3 7 6 4 5 

Emergency 

FIFO 0 0 0 0 0 0 0 0 35 33 24 32 29 16 31 121 

Failure 

MAS 99 94 88 89 98 89 92 94 8 6 6 11 11 8 7 8 


month, the same as the other set, FIFO is not processed 

on equipment failure, MAS consultation mechanism with 

self-processing, simulation statistics are shown in Table 

V, Fig.9 and Fig. 10. 

Fault simulation shows that the proposed consultation 

mechanism of local autonomy which reduce equipment 

failures better impact on the production line. 

In the normal fault and the fault of the emergency, 

were increased on-time delivery rate, and reduce the 

average weighted delay.


Avarage Weighted Delay 

Time delivery% 

Delay Time 

Time delivery 

120 

100 

80 

60 

40 

20 

0 

90 

80 

70 

60 

50 

40 

30 

20 

10 

0 

120 

100 

80 

60 

40 

20 

45 

40 

35 

30 

25 

20 

15 

10 

5 

0 

0 

1 2 3 4 5 6 7 8 Process Set 

Figure7. Time Delivery Comparison In Normal State 

1 2 3 4 5 6 7 8 

Process Set 

Figure8. Delay Time Comparison In Normal State 

1 2 3 4 5 6 7 8 Process Set 

Figure9. Time Delivery Comparison In Failure Model 


G-FIFO 

G-EDD 

G-MAS 

E-FIFO 

E-MAS 

G-F-FIFO 

G-F-MAS 

E-F-FIFO 

E-F-MAS 

1 2 3 4 5 6 7 8 Process Set 

Figure10. Average Weighted Delay In Failure Model 

G-F-FIFO 

G-F-MAS 

E-F-FIFO 

E-F-MAS 

Dynamic rescheduling problem is widely used in 

modern production plant. In this paper, it is the first time 

that the improved Contract Net Protocol from MAS is 

used into the rescheduling of the workshop environment, 

to provide a new way of solving the problem in this area. 

Then give full consideration to the workshop production 

of the machine failure and repair process of scheduling, 

the complex dynamic rescheduling process is divided 

into corresponding independent Agent and interactive 

process. The technology extends the behavior of the fault 

and related considerations. The problem of mechanical 

failure was simulated to prove the validity of the model 

based on MAS system simultaneously. 


G-FIFO 

G-MAS 

E-FIFO 

E-MAS 


This work is financially supported by the National 

Natural Science Foundation of China under Grant 

No.61064011. And it was also supported by China 

Postdoctoral Science Foundation, Science Foundation for 

The Excellent Youth Scholars of Lanzhou University of 

Technology, and Educational Commission of Gansu 

Province of China under Grant No.20100470088, 

1014ZCX017 and 1014ZTC090, respectively. 

REFERENCES 

[1] H.Ayton,M.Lawley ,and K.Mckay, “Executing production 

schedules in the face of uncertainties:a review and some 

future direction.”European Journal of Operation Research, 

vol.161(1),pp.86-110, 2005. doi:10.1016/j.ejor. 2003. 08. 

027. 

[2] A.Rossi,and E.Boschi, “A hybrid heuristic to solve the 

parallel machines job-shop scheduling problem.” 

International Journal of Production Research, 

vol.40(2),pp.118-127,2009.doi:10.1016/j.advengsoft.2008. 

03.020. 

[3] L.K.Church,and R.Uzsoy,“Analysis of periodic and event 

driven rescheduling policies in dynamic 

shops.”International Journal of Computer Integrated 

Manufacturing,vol.5(3),pp.153-163,1992,doi:10.1080/095 

11929208944524. 

[4] J.Fang,and Y.G. Xi. “A rolling horizon job shop 

rescheduling strategy in the dynamic environment.”The 

International Journal of Advanced Manufacturing 

Technology,1997,13(3):227-232,doi:10.1007/BF01305874. 

[5] S.V.Mehta,and R.M.Uzsoy, “Predictable Scheduling of a 

Job Shop Subject to Breakdowns.”IEEE Transactions on 

Robotics and Automation, 1998,vol.14(3),pp.365-378,doi: 

10.1109/70.678447. 

[6] Y.K.Kim,K.Park,J.Ko “A symbiotic evolutionary 

algorithm for the integration of process planning and job 

shop scheduling.” Computers and Operations 

Research,2003,vol.30(8),pp.1151-1171. doi:10.1016/S030 

5-0548(02)00063-1. 

[7] D.Petrovic, D. Alejandra, “A fuzzy logic based production 

scheduling/rescheduling in the presence of uncertain 

disruptions.” Fuzzy Sets and Systems, vol.157 

(16),pp.2273-2285,2006.doi:10.1016/j.fss.2006.04. 009. 

[8] S.C. Cheng, D.F. Shiau,and Y.M.Huang “Dynamic hardreal-time 

scheduling using genetic algorithm for 

multiprocessor task with resource and timing 

constraints.”Expert System with Application, vol.36(1), 

pp.852-860,2009. doi:10.1016/j.eswa.2007.10.037. 

[9] H.Wang,D.Xiang, and G.H. Duan. “Assembly planning 

based on semantic modeling appoach.”Computer in 

Industry,vol.58(3),pp.227-239,2009.doi:10.1016/j.compin 

d.2006.05.002. 

[10] J.F.Goncalves, J.J.Mendes,and M.G.Resende “A hybrid 

genetic algorithm for the job shop scheduling problem.” 

European Journal of Operational Research.vol.167(1), 

pp.77-95,2005.doi:10.1016/j.ejor. 2004.03.012. 

[11] T.N.Wong, C.W.Leung, K.L.Mak and R.Y.K. Fung 

“Dynamic shopfloor scheduling in multi-agent 

manufacturing system.” Expert System with 

Applications,vol.31(3),pp.486-494,2006.doi:10.1016/j.esw 

a.2005.09.073. 

[12] M.Ennigrou and K.Ghédira. “New local diversification 

techniques for flexible job shop scheduling problem with a 

multi-agent approach.”Autonomous Agents and Multi-


AgentSystems,vol.17(2),pp.270-287,2006.doi:10.1007/ s10 

458-008-9031-3. 

[13] I.Kacem, S.Hammadi, and P.Borne “Pareto-optimality 

approach for flexible job-shop scheduling problems: 

Hybridization of evolutionary algorithms and fuzzy logic.” 

Mathematics and Computer in Simulation,vol.60(3- 

5),245-276, 2002, doi:10.1016/S0378-4754(02)00019-8. 

[14] H.Z. Jiaa, J.Y.H. Fuh,A.Y.C.Neea and Y.F.Zhang. 

“Integration of genetic algorithm and Ganntt chart for job 

shop scheduling in distributed manufacturing systems.” 

Computer and Industrial Engineering.vol.53(2),pp.313- 

320,2007, doi:10.1016/j.cie.2007.06.024. 

[15] M.K. Lim, Z.Zhang “A multi-agent based manufacturing 

control strategy for responsive manufacturing”. Journal of 

Materials Processing Technology. Vol.139(1-3),pp.379- 

38,2003, doi:10.1016/S0924-0136(03)00535-1. 


Fuqing Zhao P.h.D., born in Gansu, China, 1977, has got a 

P.h.D. in Dynamic Holonic Manufacturing System, Lanzhou 

University of Technology, Gansu, 2006. He is a Post Doctor in 

Control Theory and Engineering in Xi’an Jiaotong University 

and Visiting Professor of Exeter University. His research work 

includes theory and application of pattern recognition, 

computational Intelligence and its application, where fifteen 

published articles can be found. 

Jizhe Wang born in Henan,China,1986.His research interest is 

the application of pattern recognition and Artificial 

Intelligence . 

Jianxin Tang born in Henan, China, 1985. His research is the 

theory and application of pattern recognition.


Multilevel Network Security Monitoring and 

Evaluation Model 

Jin Yang 

Department of Computer Science/LeShan Normal Univ., LeShan, China 

Email:jinnyang@163.com 

Tang Liu 

College of Fundamental Education/Sichuan Normal Univ., Chengdu, China 

Email: 253960818@qq.com 

Lingxi Peng * 

Department of Computer and Education software/Guangzhou Univ., Guangzhou, China 

Email: bigluckboy@163.com 

XueJun Li, Gang Luo 

Department of Computer Science/LeShan Normal Univ., LeShan, China 

Email:279718518@qq.com 

Abstract—Based on the correspondence between the 

artificial immune system antibody and pathogen invasion 

intensity, this paper is to establish a real-time network risk 

evaluation model. According to the network intrusion own 

characteristics and the consequence from service, assets and 

attack, this paper design to build a hierarchical, 

quantitative measurement indicator system, and an unified 

evaluation information base and knowledge base. The paper 

also combines assets evaluation system and network 

integration evaluation system, considering from the 

application layer, the host layer, network layer may be 

factors that affect the network risks. The experimental 

results show that the new model improves the ability of 

intrusion detection and prevention than that of the 

traditional passive intrusion prevention systems. 

Index Terms—network security; artificial immune; 

intrusion detection 


The traditional network security approaches include 

virus detection, frangibility evaluation, and firewall etc, 

e.g., the Intrusion Detection System (IDS) [1]. They rely 

upon collecting and analyzing the viruses’ specimens or 

intrusion signatures with some traditional techniques [2]. 

Moreover, being lack of self-learning and self-adapting 

abilities, they can only prevent those known network 

intrusions, and can do nothing for those variety intrusions. 

Recent years, the artificial immune system has the 

* Corresponding author. 

This work is supported by the National Natural Science Foundation 

of China under Grant (No.61003310) and the Scientific Research Fund 

of Sichuan Provincial Education Department (No. 10ZB005). 


doi:10.4304/jsw.6.5.806-813 

features of dynamic, self-adaptation and diversity [3-6] 

that just meet the constraints derived from the 

characteristics of the grid environment, and mobile agent 

has many same appealing properties as that of artificial 

immune system. Negative Selection Algorithm and the 

concept of computer immunity proposed by Forrest in 

1994 [7-8]. In contrast, the AIS theory adaptively 

generates new immune cells so that it is able to detect 

previously unknown and rapidly evolving harmful 

antigens [9]. However, much theoretical groundwork in 

immunological computation has been taken up, but there 

is a lack of perfectly systems based AIS of dynamical 

immunological surveillance for network security. 

Based on the correspondence between the artificial 

immune system antibody in the artificial immune systems 

and pathogen invasion intensity, this paper is to establish 

a network risk evaluation model. According to the 

network intrusion own characteristics and the 

consequence from service, assets and attack, we design to 

built a hierarchical, quantitative measurement indicator 

system, and an unified evaluation information base and 

knowledge base. This model will help the network 

managers evaluate the possibility and the graveness 

degree of the network dangerous quickly, ease the 

pressure of recognition, to get targeted immediate defense 

strategy of the strength and risk level of the current 

network attacks. 

II. THE EVALUATION OF THE NETWORK DANGER 

The biological immune system can produce antibodies 

to resist pathogens through B cells distributing all over 

the human body. And T cells can regulate the antibody 

concentration. Simulating biological immune system, we 

place a certain amount of immune cells into the network, 

and perceive the surrounding environment. Simulating 

creatural immune system, we place a certain amount of


immune cells into the network, and perceive the 

surrounding environment of the detectors. As soon as the 

immune detectors detect an attack, the detectors begin 

clone and generate a mass of similar detectors in order to 

defend from fiercer network attacks and warn the 

dangerous level of the network [10]. While the network 

danger become abating, the corresponding numbers of 

antibodies will decrease at the same time. The detectors’ 

number and type reflect the attack’s intensity and type 

suffered by the network intrusion. In this model, the 

detectors can be categorized, according to the evolvement 

progress of the detectors themselves, into 3 types, viz. 

immature detectors, mature detectors and memory 

detectors, and the content will be expatiated in detail in 

the following. 

Figure 1. The Framework of danger-detecting in the NDEMAIS 

III. THE DYNAMICAL IMMUNOLOGICAL 

SURVEILLANCE MODEL BASED ON AIS 

A. The Definition of Antigen, Antibody, Self and Non-self 

l 

Definition: Antigens ( Ag , Ag ⊂ U , D = { 0, 

1} 

) are 

fixed-length binary strings extracted from the Internet 

Protocol (IP) packets transferred in the network. The 

antigen consists of the source and destination IP 

addresses, port number, protocol type, IP flags, IP overall 

packet length, TCP/UDP/ICMP fields [11-12], etc. The 

structure of an antibody is the same as that of an antigen. 

For virus detection, the nonself set (Nonself) represents IP 

packets from a computer network attack, while the self 

set (Self) is normal sanctioned network service 

transactions and nonmalicious background clutter. Set 

Ag contains two subsets, Self ⊆ Ag and Nonself ⊆ Ag 

such that 

Self ∪ Nonself = Ag and Self ∩ Nonself = Φ . (1) 

B. The Dynamic Equations of the Mature Detectors 

Tb = { x | x ∈ B, 

∀y 

∈ Self ( < x. 

d, 

y >∉ Match ∧ x. 

count < β)} 

(2) 


T ( t) 

= T ( 0) 

= 0, 

t = 0 

(3) 

T ( t + Δt) 

= T ( t) 

+ Tnew 

( Δt) 

−Tmatch 

_ self ( Δt) 

−Tdead 

( Δt) 

−Tactive 

( Δt) 

T . age( 

t + Δt) 

= T. 

age( 

t) 

+ 1, 

when T. 

age( 

t + Δt) 

< λ (5) 

∂Tmatch 

_ self 

Tmatch 

_ self ( t + Δt) 

= 

T ( t), 

∂xmatch 

_ self 

when f match ( T ( t −1), 

Self ( t −1)) 

= 1 

(4) 

(6) 

T . count( 

t + Δt) 

= T. 

count( 

t) 

+ 1 (7) 

∂Tactive 

Tactive 

( Δ t) 

= ⋅ Δt, 

when T. 

count( 

t + Δt) 

≥ β (8) 

∂x 

active 

∂Tdeath 

Tdead 

( Δt) 

= ⋅ Δt, 

∂xdeath 

when T. 

count( 

t + Δt) 

< β , andT. 

count > λ 

∂Tactive 

⋅ Δt 

= I maturatiio n ( t), 

T. 

ρ( 

Δt) 

= 0, 

∂xactive 

when I. 

age( 

t) 

> α 

(9) 

(10) 

Equation (4) depicts the lifecycle of the mature 

detector, simulating the process that the mature detectors 

evolve into the next generation. All mature detectors have 

a fixed lifecycle (λ). If a mature detector matches enough 

antigens ( ≥ β ) in its lifecycle, it will evolve to a memory 

detector. However, the detector will be eliminated and 

replaced by new generated mature detector if they do not 

match enough antigens in their lifecycle. Tnew (t) 

is the 

generation of new mature detector. Tdead (t) 

is the set of 

detector that haven’t match enough antigens ( ≤ β ) in 

lifecycle or classified self antigens as nonself at time t. 

T ( t + Δt) 

simulates that the mature detector undergo one 

step of evolution. Tdead (t) 

indicates that the mature 

detector are getting older. Tactive (t) 

is the set of the least 

recently used mature detector which degrade into 

Memory detector and be given a new age T > 0 and count 

β > 1. 

Because the degraded memory detector has better 

detection capability than mature detector, it is better to 

form a memory detector. When the same antigens arrive 

again, they will be detected immediately by the memory 

detector. In the mature detector lifecycle, the inefficient 

detectors on classifying antigens are killed through the 

process of clone selection. Therefore, the method can 

enhance detection efficiency when the abnormal network 

behaviors intrude the system again. 

In the course, λ is the threshold of the affinity for the 

activated detectors. The affinity function fmatch ( x, 

y) 

may 

be any kind of Hamming, Manhattan, Euclidean, and rcontinuous 

matching, etc. In this model, we take rcontinuous 

matching algorithm to compute the affinity of


mature detectors. The matching functions utilize the 

following definitions: 

⎧1 

∃i, 

j, 

j − i ≥ r ∧ 0 

⎪ 

fmatch ( x, 

y) 

= ⎨ xi 

= yi 

, xi+ 

1 = yi+ 

1, 

L , x j = y j (11) 

⎪ 

⎩0 

otherwise 

The r-continuous matching is commonly used method 

for measuring the distance between bit strings with the 

goal of producing a better similarity coefficient. 

C. The Dynamic Equation of Memory Detectors 

M ( t) 

= M ( 0) 

= 0, 

t = 0 (12) 

M ( t + Δt) 

= M ( t) 

+ M new( 

Δt) 

+ M from _ other ( Δt) 

− M dead ( Δt), 

when fmatch( 

M ( t), 

Ag( 

t)) 

≠ 1 , t > 1, 

(13) 

M ( t + Δt) 

= M ( t) 

+ Mclone( 

t) 

+ M new( 

Δt) 

+ M from _ other ( Δt) 

− M dead ( Δt), 

when fmatch( 

M ( t), 

Ag( 

t)) 

= 1 

∂M 

clone ∂M 

active 

Mclone( 

t) 

= ⋅ ⋅ Δ( 

t −1), 

∂xclone 

∂xactive 

when fmatch( 

M ( t), 

Ag( 

t)) 

= 1 

(14) 

(15) 

Mclone( t + Δt) 

= Mclone( 

t), 

M. 

ρ( 

t + Δt) 

= M . ρ( 

t) 

+ Vp 

⋅ Δt, 

(16) 

M. 

count( 

t + Δt) 

= M. 

count( 

t) 

+ 1 

1 

M . ρ( 

t + Δt) 

= ⋅ M. 

ρ( 

t), 

M. 

age( 

t + Δt) 

= M. 

age( 

t) 

+ 1, 

2 

when fmatch( 

M ( t), 

Ag( 

t)) 

≠ 1 

M 

new 

∂M 

Δt = new ∂T 

( ) ⋅ Δt 

= ⋅ Δ( 

t −1) 

∂x 

∂x 

new 

(17) 

active , M new . ρ ( t) 

= ρ0 

active 

(18) 

∂M 

death 

M dead ( Δt) = ⋅ Δt, 

when fmatch( 

M ( t −1), 

Self ( t −1)) 

= 1 

∂xdeath 

(19) 

k i 

∂M 

from _ other 

M from _ other ( Δt) = ∑ ( 

⋅ Δt) 

(20) 

i= 

1 ∂x 

from _ other 

Equation (13) depicts the dynamic evolution of 

memory detector. M ( t + Δt) 

simulates the process that the 

memory detector evolve into the next generation ones. 

M new (t) 

is the set of memory detector that are activated 

by antigens lately. These mature detector matched by an 

antigen will be activated immediately and turn to a 

memory detector. M dead (t) 

is the memory detector that 

be deleted if it matches a known self antigen. M clone (t) 

is 

the reproduced memory detector when the detector 

distinguish a antigens. M from _ other ( t) 

is the memory 


detector that transformed from other computers. The k 

indicates that the ID number of the computer. Therefore, 

dynamic model of immune is to generate more antibodies 

and enhance the ability of self-adaptation for the system. 

D. The Dynamic Model of Self 

In a real-network environment some network services 

and activities are often change, which were permitted in 

the past but may be forbidden at the next time. 

I t) 

= I( 

0) 

= { x , x ,..., xn 

}, t = 0 (21) 

( 1 2 

I new match self maturation 

( t + Δt) 

= I( 

t) 

+ I ( Δt) 

− I _ ( Δt) 

− I ⋅ Δt 

(22) 

Imatch _ self 

match 

I . age( 

t + Δt) 

= I. 

age( 

t) 

+ 1 (23) 

( t + Δt) 

= I( 

t), 

when f ( I( 

t −1), 

Self ( t −1)) 

= 1 

(24) 

I n 

maturatiio ( t + Δt) 

= I ( t), 

I. 

age( 

t + Δt) 

> α (25) 

∂Irandom 

∂Iinherit 

Inew( Δ t) 

= ( ξ 1 ⋅ ) ⋅ Δt 

+ ( ξ2 

⋅ ) ⋅ Δt 

(26) 

∂x 

∂x 

Equation (22) stimulates the dynamic evolution of selfantigens, 

where xi ∈ ℜ( 

i ≥1, 

i ∈ N) 

is the initial self 

element defined. Inew is the set of newly defined elements 

at time t, and Imaturation is the set of mutated elements. 

fmatch ( y, 

x) 

is used to classify antigens as either self or 

nonself: if x is a self-antigen, return 0; if x is a nonself 

one, return 1; if x is detected as nonself but was detected 

as a self-antigen before, then it may be a nonself antigen 

(needs to be confirmed), and return 2. There are two 

advantages in this model. (1) Self immune surveillance: 

The model deletes mutated self-antigens (Imaturation) in 

time through surveillance. The false-negative error is 

reduced. (2) The dynamic growth of Self: The model can 

extend the depiction scope of self through adding new 

self-antigens (Inew) into Self. Therefore, the false-positive 

error is prevented. 

E. The Antibody Cross 

In order to keep the variety of individual as well as the 

optimal solution can be achieved, we divide the antibody 

gene to n gene bits set and utilize multi-point cross 

process. For example, we select two gene by random such 

as 

' ' ' ' 

G1 = { g1, 

g2 

, L, 

gi 

, Lg 

n} 

, G2 = { g1, 

g2 

, L, 

gi 

, Lg 

n} 

, 

Select some points randomly, and then form two-point 

pair with some probability (p) to cross operation, to 

generate cross point set, and then to generate new gap of 

' 

set Gnew = { g1, 

g2 

, L, 

gi 

, Lg 

n} 

. Select cross point 

according to binomial distribution 

⎛ n⎞ 

k n− 

k 

P{ 

X = k} 

= ⎜ p ( 1− 

p) 

, k = 0, 

1, 

2, 

K, 

n 

k ⎟ 

. 0 

⎝ ⎠


E ( X ) = np , D( X ) = np( 

1− 

p) 

, where X is the numbers of 

cross points. Then the G 1 and G 2 turn into the offspring 

G new by the cross process. 

F. Antibody Variation 

In order to prevent algorithm from converging 

prematurely, we take variation operation to the Gene set 

G1 = { g1, 

g 2 , L, 

gi 

, L g n} 

after the cross process. Select 

variation point randomly and varied with some variation 

probability ( p m ) to generate new generation 

' 

Gnew = { g1, 

g2, 

L, 

gi 

, Lg 

n} 

. Select variation point 

according to Poisson distribution 

k −λ 

λ e 

P{ 

X = k} 

= , k = 0, 

1, 

2, 

L 

k! 

(28) 

E ( X ) = D( 

X ) = λ > 0 , where X is the numbers of 

variation points. Then the 1 G turn into the offspring G new 

by the variation process. 

G. The Process of Immunological Surveillance 

These response processes can be divided into two 

stages: primary immune response, secondary immune 

response. When the same antigen intrude into the body 

for the second time, because of the organism have some 

antibody to identify the antigen, the immune system can 

respond quickly and come into being a large number of 

homothetic antibodies in order to clear the rapid antigen 

in the body. The rapid process is known as secondary 

immune response. 

Memory detectors Mb(t) express that the system has 

suffering various attacks at moment t. Memory detectors’ 

concentration value of antibody is ρ(t). ρ(t) shows that at 

current time the system is suffering what kind of attacks 

and computes the intensity and categories. It is one of the 

important indicators of reflection the current system 

network intrusion danger level. There are two major 

changes probability of antibody concentration. 

(1) The increase of the antibody concentration: When 

the memory detector antibody captures a particular 

antigen, representing have detecting a kind of invasion, 

we increase the antibody concentration. We use Vρ(t) 

reflects the rate of increase of antibody concentration. 

Therefore, at moment t the antibody concentration ρ(t) of 

Mb(t) is: ρ(t)=ρ(t-1)+ Vρ·Δt. Affected by the antigen, the 

more intensive invasion antigens, the faster at the 

increase rate of antibody concentration. Taking into 

account the emergence of the invasion was a random act, 

such as biodynamic body falling on the ground, such as 

trees fluttering in the wind, the random movement of the 

rate could be subject to Gaussian distribution. Antigen 

number x and the variety speed excitation function of 

antibody concentration Vρ(x) accord with the Gaussian 

distribution with parameters (h, μ), and these nonlinear 

causal relationship can be expressed as follows: (which, x 

is the number of memory detector detecting the invasion 

in period of time.) 


2 

[( x−h) 

⋅u] 

A − 

V ( x) 

= e 2 

ρ , u > 0, 

0 < x < +∞ (29) 

2π 

σ 

In order to avoid detection unlimited cloning, we 

regulate A is the largest concentration of limiting growth. 

Since the each invasion of antigen make different impact 

on the network and host, we define μ parameters to reflect 

the extent of the damage caused by antigens. The greater 

of μ value the faster of the antibody concentration 

increase speed, and also shows the more dangerous of the 

antigens captured by the antibody. Parameters h shows 

that the antigen stimulate the antibody to the limit when 

stimulation course achieve to a certain extent. 

(2) The attenuation of the antibody concentration: In 

organisms these antibodies which stimulated by some 

certain antigens will gradually disappear after a certain 

period of time. In our system, if the memory detector in 

the next step failed to clone once more, we set the 

antibody concentration turn into its attenuation phase, in 

accordance with the following. 

1 

ρ( t ) = ρ( 

t −τ 

) , τ ≤ t ≤ T 

(30) 

2 

After each half life τ, if the detector did not find any 

antigen, the antibody concentration will reduce by half. 

When the antibody concentration decay to ε τ , it will stop 

decaying, showing that the act of invasion has 

disappeared. In the mature-detector lifecycle, the 

inefficient detectors on classifying antigens are killed 

through the process of clone selection. However, the 

efficient detectors on classifying antigens will evolve to 

memory detectors. Therefore, similar antigens 

representing abnormal network behaviors can be detected 

quickly when they intrude the system again as secondary 

immune response. 

Ⅳ. THE EVALUATION OF THE NETWORK 

DANGER 

After we describe the network attacking actions, it is 

necessary to evaluate the dangerous degree of the 

network, and judge the severity of the attacking actions. 

Network dangers are of diversity (because of many 

affecting factors) and of randomicity. Thus, evaluation is 

a process involving numerous complicated factors. The 

essence of evaluation is to hierarchically grade each 

evaluation factor in terms of its weight in order to 

ultimately reflect the dangerous degree of the entire 

network. 

Our model simulates the process that metabolism and 

competition of the cells organism through the use of 

continuous renovation and enrichment process. The 

values of M b reflect the intensity of intrusion in current 

network. Therefore, system evaluates the network 

security by perceiving the danger around of them. Owing 

to the fact that our model relates to enormous factors for 

evaluation, on purpose of reasonably and entirely


measuring the network dangerous status, we classify the 

involved factors as host dangers, area dangers, detectors 

dangers, and special dangers. The host dangers mainly 

arise out of the dangers which are possessed by each host 

computers; the area dangers are divided by regions; the 

detector dangers are those factors which determined in 

terms of antibody consistencies, antibody consistencies 

change speed, and detector types; and particular dangers 

refer to the second corresponding time and responding 

speed etc. Afterwards, we subdivide and arrange all the 

factors which influence the network dangers, in order to 

let them locate on different layers, forming a structure 

model with identify matrix. In the traditional evaluation 

system, the choice usually was made according to few of 

main factors, taking into account other factors merely as 

references, and then simply described the happening 

possibility of the dangers by high, medium and low. 

Hence, many factors which cannot be quantified were 

always ignored, so that the traditional system can not 

evaluate the varying situation where there are various 

factors and conditions involved at the same time, which 

more often than not give rise to the result that evaluation 

decisions are void of comprehensiveness and that the 

outcomes are distorted. 

In our evaluation model we utilize the AHP (Analytic 

Hierarchy Process) Principle provided by operational 

research expert T. L. Saaty to evaluate the network 

dangerous situation. This method can efficiently cope 

with those complex problems which are difficult to be 

solved by quantitative methods. Its characteristics are: 

firstly, it can break complex questions down to some 

gradations, then analyze progressively on the layers much 

simpler than before in order to express and handle the 

decision-makers’ subjective judgments through quantity 

format. Next, it calculates by mathematics the weight of 

the sequence of relative significance of the factors on 

each layer. Via the general permutation among all the 

layers, compute and rank the relative weight of all the 

factors. We synthetically consider the qualitative and 

quantitative factors during the evaluation process, adopt 

the AHP Principle, so as to provide reasonable methods 

and instruments for comprehensive evaluation of the 

network dangerous situation. 

A. Computation the Hiberarchy of the Model 

Owing to the fact that our model relates to enormous 

factors for evaluation, on purpose of reasonably and 

entirely measuring the network dangerous status, we 

classify the involved factors as host dangers, area dangers, 

cells dangers, and special dangers. The host dangers 

mainly arise out of the dangers which are possessed by 

each host computers; the area dangers are divided by 

regions; the cell dangers are those factors which 

determined in terms of cell consistencies, cell-varying 

speed, and cell types; and particular dangers refer to the 

second corresponding time and responding speed etc. 

Afterwards, we subdivide and arrange all the factors 

which influence the network dangers, in order to let them 

locate on different layers, forming a structure model with 

identify matrix. The following danger evaluation model is 


divided in to the general object layer (A), the standard 

layer (B), and the factor layer (C) (see Fig. 2). 

Furthermore, to get the weight values of these factors, 

this model adopts the AHP method. The underlying idea 

is to compare and estimate the eigenvalue λi of the 

special equation of the matrix B by way of seeking a 

solution, then find out the maximum eigenvalue λ max and 

get its corresponding eigenvector X = ( x1, 

x2 

, L, 

xn 

) . 

Finally, we will get relative weight vector 

A = w , w , L, 

w ) after merging all eigenvectors into one. 

( 1 2 n 

In a word, the entire method can be approximately 

reduced to four steps, that is, ① establish hierarchical 

structure model; ② construct evaluation matrices; ③ 

calculate factors’ relative weight under a coincident 

standard; ④ compute the integrated weight of the factors 

on each layer. Hereinafter, the evaluation process will be 

introduced in detail. 

Figure 2. The hierarchy of network danger evaluation model 

B. Computing Single Weight 

1) Construct Identify Matrix: First of all, we must 

construct identify matrix which is result that we 

compared the relative importance of one group of 

elements on next layer with some past layer element 

constraint. That is, it shows the relative importance of any 

pair of factors. In detail, denote b ij the compared result of 

the ith factor and jth one, b ij all together form the identify 

matrix B : 

⎛ b11 

b12 

L b1n 

⎞ 

⎜ 

⎟ 

⎜b21 

b22 

L b2n 

⎟ 

B = ⎜ 

L L L L 

⎟ 

⎜ 

⎟ 

⎜ 

⎟ 

⎝bn1 

bn2 

L bnn 

⎠ 

Where: = 1 

b if j 

ii 

i = and b ij 1/ 

b ji 

= if i ≠ j . 

2) Computing Weights: Next we obtain the weight of 

each factor. According to the identify matrix B , we can 

get the maximum eigenvalue of the matrix λ max . Here, we 

can get the maximum λ max according with the following 

contidion: 

b 

11 

b 

− λ 

21 

L 

b 

n1 

b 

b 

22 

b 

12 

L 

n2 

L 

− λ L 

L 

L 

b 

b 

nn 

1n 

b 

2n 

L 

− λ 

= 0


Work out the corresponding eigenvector of maximum 

eigenvalue of B , X = ( x1, 

x3, 

L, 

xn) 

, let x i to be the 

weight of factor u i , then we can get unitary weights 

denote W i . 

n n 

n 

A = ( W1, 

W2, 

L, 

Wn) 

= ( x1 / ∑ xi 

, x2 

/ ∑xi, L , xn 

/ ∑xi, 

) 

i= 

1 i= 

1 i= 

1 

3) Test of Consistency: Because of complexity of 

evaluation and limit of individual knowledge, the 

individual identify matrix may not be consistent with the 

actual one, or the disagreement of any two identify 

matrixes may result in error of subjective judgment. 

However, we must test the consistency of the matrix B as 

follows: 

① Computing consistency value C ⋅ I 

λmax 

− n 

C ⋅ I = 

(31) 

n − 1 

② Computing consistency ratio C ⋅ R 

C ⋅ I 

C ⋅ R = 

(32) 

R ⋅ I 

Where R ⋅ I is mean consistency value that can be found 

in the reference and forms, we often consider that if C ⋅ R 

is smaller than 0.1, the consistency of matrix is 

acceptable, otherwise we must modify the identify matrix 

B. 

4) Computing the General Weight Order: The general 

weight order means that the weight order comparing the 

elements in the present layer and the highest layer. We 

have got each order of element in rule layer to the object 

layer and the values are W 1, W2 

, L , Wn 

, respectively, we 

also know that order that design layer to the rule layer 

j j j 

and the values are W W , , W 

j ⎛W 

⎜ 1 

j 

j ⎜W2 

V = W W = ⎜ 

⎜ M 

⎜ j 

⎝Wn 

1 , 2 L n , then the general order is 

W 

W 

W 

j 

1 

j 

2 

M 

j 

n 

j 

L W ⎞⎛W 

⎞ ⎛V 

1 1 1 ⎞ 

⎟⎜ 

⎟ ⎜ ⎟ 

j 

L W ⎟⎜W 

⎟ ⎜V 

2 2 2 ⎟ 

⎟⎜ 

⎟ = ⎜ ⎟ 

M M ⎟⎜ 

M 

⎟ ⎜ 

M 

⎟ 

j 

L W ⎟⎜ 

⎟ ⎜ ⎟ 

n ⎠⎝Wn 

⎠ ⎝Vn 

⎠ 

, (33) 

C. Evaluating the danger level 

The entire network of danger level should fully reflect 

the value of each of the host facing attacks. As the host of 

each position is not the same such as running a different 

system for different users and providing different services, 

influencing different economic, affecting different social 

and even political values, they are in possession of 

different essentiality. 

Let nij (t) 

be the numbers of i th computers detect 

attacking at time t. Let β i ( 0 ≤ βi 

≤1) 

be the importance 

coefficient of ith computer in the network and α j 

( 0 ≤ α j ≤ 1) 

be the danger coefficient of the j th kind of 

attack in the network. Then, we can define the attack 

intensity ri (t) 

of the j th kind of attack and the 

corresponding network danger ri (t) 

as follows: 

2 

ri 

( t) 

= 

− 1 

(34) 

1 + e 


− 

∑ 

α jnij 

j 

8 

Let Importancei 

= ∑ ( I k × Wk 

) be the importance 

k= 

1 

coefficient of j th host in the network. Then, we obtain the 

network entire danger level value: R(t) =∑(indicator 

value × indicator weight). Therefore, we can get network 

danger R(t) situation and evaluate network security at real 

time. 

N n 

R( 

t) 

= tanh( ∑( 

∑( 

Hosti 

's 

danger× 

Importancej 

) × LCRS_Weightm 

)) 

m= 

1 i= 

1 

N n 

8 

= tanh( ∑( 

∑( Hosti 

's 

danger× 

∑( 

I j, 

k × Wk 

) ) × LCRS_Weightm 

)) 

m= 

1 i= 1 

k= 

1 

N n 8 

= tanh( ∑( 

∑( ri 

( t) 

× ∑( 

I j, 

k × Wk 

)) × LCRS_Weightm 

)) 

m= 

1 i= 1 k= 

1 

(35) 

The conclusion can be shown that the higher value R(t) 

reaches the more dangerous the network is. 

Ⅴ.EXPERIMENTAL RESULTS AND ANALYSIS 

A. Experimental Environment and Evaluation Indicators 

Experiments of attack simulation were also carried 

out in our Laboratory. Analytic Hierarchy Process is 

applied to our model to evaluate the weights of the 

indicators in the experiments. Considering the preciseness 

and efficiency, we use 12 indicators to evaluate the 

network danger, which include host danger, area danger, 

cells danger, special danger etc. The weights of the 

indicators are evaluated and sorted by orders and the 

mapping of all hierarchies can be seen in Table 1. 

B 1 

0.0969 

TABLE I 

Weight Orders of All Layers 

B 2 

0.0485 

B 3 

0.8253 

B 4 

0.0293 

C 1 1 0.0969 

C 2 

1 0.0485 

C 31 

0.3801 0.3137 

C 32 

0.4029 0.3325 

C3 3 

0.0934 0.0771 

C3 4 

0.1235 0.1020 

C 41 

0.3333 0.0097 

C 42 

0.6667 0.0195 

Evaluate the above indicators: C.R = 0.036 < 0.1. It 

has a satisfactory consistency. Consequently, the weights 

of the indicators in the NAIMAI can be evaluated as the 

following: host dangers: 0.0969, area dangers: 0.0485, 

cells dangers: 0.8253, special dangers: 0.0293 and others 

dangers: 0.0969, 0.0485, 0.3137, 0.3325, 0.0771, 0.1020, 

0.0097, and 0.0195. The conclusion can be draw that the 

higher values R(t) reaches , the more dangerous the 

network is. Otherwise, the lower the R(t) is, the safer the 

network is. 

W


An antigen was defined as a fixed length binary string 

composed of the source/destination IP address, port 

number, protocol type, IP flags, IP overall packet length, 

TCP/UDP/ICMP fields, and etc. The network was 

attacked by 20 kinds of attacks, such as Syn Flood, Land, 

Smurf, and Teardrop. A total of 20 computers in a 

network were under surveillance. The task aimed to 

detect network attacks. Here are the coefficients for the 

model. We use r-contiguous bits matching rule (r=8) for 

computing the affinity, n=40 (the size of initial self set), 

and ξ =4 (the number of new generated immature 

detectors). The activation threshold is β ; tolerance period 

is λ . 

B. Results and Analysis 

Figure 3 illustrates the syn attacks. Figure 4 depicts the 

evaluation of the network danger in our model. 

Packets/s 

30000 

20000 

10000 

0 

0 20 40 60 80 100 120 

Time 

Figure 3. The network suffering from the syn incursions for instance 

Attack intensity 

1 

0.8 

0.6 

0.4 

0.2 

0 

0 20 40 60 80 100 120 

Time 

Figure 4. The line of the network dangers obtained by our model at 

these incursions 

As is shown in Figure 4, R(t) changes when attack 

levels changes. The rise in attack levels is accompanied 

by a corresponding increase in R(t), as implies the bad 

network security. On the other hand, if attack levels 

decline, R(t) decreases accordingly after seconds of delay. 

Therefore, the network can stays on guard even when the 

attacks occur once again during a very short time. 


Ⅵ.CONCLUSIONS 

This paper combines the risk evaluation methods with 

application security engineering principles, and can 

change current passive defense situation using traditional 

network security approaches, and is helpful to establish 

new generation proactive defense theories and realization 

techniques. At the same time, the work is of not only 

theoretic values to design proactive defense systems 

which have intrusion tolerant ability and survivability in 

any complex network circumstances, but also very 

significant to protect network infrastructure. The 

experimental results show that the proposed model has 

the features of real-time processing that provide a good 

solution for network surveillance. 


This work is supported by the National Natural 

Science Foundation of China under Grant (No.61003310) 

and the Scientific Research Fund of Sichuan Provincial 

Education Department (No. 10ZB005). 

REFERENCES 

[1] Thomas, Ciza, Balakrishnan, N. Performance enhancement 

of Intrusion Detection Systems using advances in sensor 

fusion. Information Fusion, 2008 11 th International 

Conference on June 30 , (2008):1-7 

[2] Abdoul Karim Ganame, Julien Bourgeois, Renaud Bidou, 

Francois Spies. A global security architecture for intrusion 

detection on computer networks. Computers & Security, 

27(1), 2008:30-47 

[3] Vasilios Katos. Network intrusion detection: Evaluating 

cluster, discriminant, and logit analysis. Information 

Sciences, 177(15), (2007), 3060-3073. 

[4] Agustín Orfila, Javier Carbó, Arturo Ribagorda. 

Autonomous decision on intrusion detection with trained 

BDI agents. Computer Communications, 31(9), 

(2008),1803-1813. 

[5] Vincent Toubiana, Houda Labiod, Laurent Reynaud, Yvon 

Gourhant. A global security architecture for operated 

hybrid WLAN mesh networks. Computer Networks, 

54(2),2010: 218-230 

[6] Kuby J.: Immunology. Fifith Edition by Richard A. 

Goldsby et al. 

[7] F.M.Burnet. The Clone Selection Theory of Acquired 

Immunity. Gambridge: Gambridge University Press (1959) 

[8] S A Hofmeyr, and S Forrest. Architecture for an artificial 

immune system. Evolutionary Computation, vol. 8 (2000) 

443-473 

[9] S Forrest, A S Perelson, L Allen, and R Cherukuri. Self- 

Nonself Discrimination in a Computer. Proceedings of 

IEEE Symposium on Re-search in Security and Privacy, 

Oakland, (1994) 

[10] Panigrahi BK, Yadav SR, Agrawal S, et al. A clonal 

algorithm to solve economic load dispatch. Electric Power 

Systems Research. 77(10), (2007):1381-1389 

[11] Tao Li. An immune based dynamic intrusion detection 

model. Chinese Science Bulletin. 50, (2005): 2650-2657


[12] Tao Li. An immunity based network security risk 

estimation. Science in China Ser. F Information Sciences. 

48 (2005):557- 578 


Jin Yang received his M.S. degree 

and the Ph.D. degree in computer 

science from Sichuan University, 

Sichuan, China. He is an Associate 

Professor in Department of Computer 

Science at LeShan normal university. 

His main research interests include 

network security, artificial immune, 

knowledge discovery and expert 

systems. 

Tang Liu received his M.S. degree 

in College of Computer Science, 

Sichuan University, China, in 2009. 

Since 2008, he has been a instructor 

in College of Fundamental 

Education, Sichuan Normal 

University. His research interests are 

in the area of wireless sensor 

networks. 

Lingxi Peng born in 1978. 

Associate professor, Ph.D. and senior 

membership of China. His main 

research interests include network 

security and artificial immune.


Research on Family and Shops Real-time Status 

of 3G Wireless Remote Monitoring System 

Abstract—Aiming at the development situation of 3G 

network, realization scheme is proposed based on 3G 

network B/S mode of embedded family status remote 

real-time monitoring system, avoid the shortage of 

traditional surveillance system which cannot 

networking,require a lot of storage media and query 

forensics more difficult. By constructing the embedded 

server can complete video collection, treatment, storage and 

transmission, and the development of embedded Web server 

software realize the real-time monitoring of the remote 

video. The scheme of the server is based on ARM11 

hardware platform,embedded Windows CE software 

platform,with the aid of Windows CE operating system 

friendly human-machine interface watch history monitoring 

records, you can also phone with 3G web browser in 

real-time monitoring. Experimental results show, System 

has many characteristics about strong real-time property, 

good interaction and lower development costs and so on, 

which can be widely used in remote video monitoring. 

Index Terms—Real-time monitoring, WinCE , 3G , B/S , 

ARM 


Because of the rapid development of 3G network 

technology, the mobile equipments based on the high 

speed of third generation (3G) mobile network 

application are more and more, the mobile phone as video 

monitoring equipment is a new 3G application, and also 

likely will be very popular a kind of application. With the 

features of mobile phone enhanced, we can basically 

inspect 3G mobile phone for an all-purpose terminal [1-2] . 

Because the 3G mobile phone not only can be used as a 

collection of real-time data message, and then transferred 

to the center platform through its processing, is also can 

be used as real-time monitoring equipment for the 

client [3] .To overcome the constraints of the traditional 

network bandwidth’s lack, mobile video surveillance 

system can well meet the requirements of real-time and 

video quality, it not only possess such as video 

monitoring function, basically can also achieve real-time 

and anywhere remotely monitor the real-time control 

application requirements [4-5] . General induction has the 

following advantages: the real-time and anywhere 

advantages, it satisfies no matter in any time, any place, 

through what equipment can undertake video monitoring, 

the cost advantage, because in the today's society, 

basically everyone have the mobile phone, that is used as 

one end of equipment of the video monitoring is a kind of 


doi:10.4304/jsw.6.5.814-818 

Qian Zhao 

School of Communication and Information Engineering 

Xi’an University of Science and Technology 

Xi’an, China 

Email: qianzhaoza@sina.com 

multiple use of resources,it reduces the cost of investment, 

and compared with the ordinary PC ,the mobile phone's 

price is much lower ,in cost investment, therefore, have a 

great advantage. 

II. SYSTEM GENERALLY STATES 

Based on 3G family conditions the remote real-time 

monitoring system includs both the client and the server. 

Client uses the generic 3G mobile phone and mobile 

phone Web browser to the server through the 3G network 

access, and obtain real-time monitoring data information. 

Server-side collects, processes, stores the video 

information, and accepts the browser's visit, transmisses 

the real-time monitoring information to the browser, and 

it can process the monitoring information, when there is 

the abnormal, it can automatically send message to cell 

phone to report [6] . System framework shown in figure 1: 

Figure 1. System framework figure 

III. THE KEY TECHNOLOGY 

In the whole system design, the core technology is the 

models of 3G technology and B/S network. 

A. 3G technology 

3G is a new technology that combines the wireless 

communication technology and the Internet and other 

multimedia technology,the 3G technology's 

improvements are mainly that the speed of transmissing 

such as graphics and sound or video information data has 

been greatly improved, it also can handle images and 

video and other forms of medias, it not only can provide 

all the information service which the second generation


communication (2G) can provide and on this basis have 

improved greatly, such as in transmission speed has 

greatly improved. Restricted and influenced by the 

configuration factors such as the network bandwidth, 

mobile equipment and terminal equipment and so on,in 

the 2G era, Most of the mobile phone users online 

function are just some simple applications, such as 

download some pictures or music, Internet browsing 

simple news page, etc. In the current 3G era, due to the 

widespread greatly improved,mobile terminal can provide 

more more advanced application or services,for instance 

you can via mobile phone watch TV, browse more online 

information, also mobile video call and more. 

The 3G mobile phone has the functions of 

information collection, video browsing and the removable 

characteristic,so in the whole system structure,the 3G 

mobile phone can either as sle equipment, which can be 

used in real-time picture film and video collection or 

information alarm monitoring, etc, but also as a 

monitoring client to browse real-time video, alarm view, 

control result analysis or check functions of sle 

equipment configuration, etc. 

B. B/S model 

In the traditional video monitoring system, basically 

its uses is based on C/S model, this model requires the 

monitoring client install the client software, therefore, it 

needs users spend a lot of time and effort to maintain. 

This system uses is based on B/S model, really realizes 

the user zero maintenance, and can also cross-platform 

operation, it not only saves the manpower and material 

resources also enables users in the environment as long as 

have a network to video monitor, it is really convenient 

for users [7] . 

IV. SERVER-SIDE DESIGN 

In the embedded Linux environment, there are three 

Web server: httpd, thttpd and Boa. Httpd is the simplest 

Web server,whose function is the weakest, and which 

does not support the authentication and CGI.If Web 

server only provides some static page,for example,simple 

on-line help,system introduction etc, it can completely be 

realized by static server httpd. Boa and thttpd support 

authentication, CGI,etc. the functions are more complete. 

If need to improve the security of the system, or need to 

interact with users, such as data query, real-time status 

query etc, it must use dynamic Web technology, can 

choose one of these two kinds of server to realize. The 

system adopts Boa to achieve embedded Web server. 

Between the server-side and each control points we use 

the wired mode, after obtaining video data it submits to 

the background MCU to complete data analysis and other 

processes, video displaying through LCD monitor and 

storing by mass storage unit, transmiting the real-time 

monitoring information to the browser, and can process 

the monitoring information, abnormal ,it automatically 

send informations to the mobile to report, simultaneously, 

but also can locally alarm. 

A. Server-side hardware design 


As the server part need to finish the functions of video 

information collection,display, storage,inquiry and 

judgment alarm.This design selects the Samsung 

company's S3C6410 embedded processor as the 

core,combining 3G network technology, again 

complementary with related devices realize its function 

demand.S3C6410 is based on 16/32 bit RISC kernel's low 

cost,low power consumption,high-performance 

microprocessors solutions, adopts 64/32-bit internal bus 

architecture, interior has integrated many powerful 

hardware accelerators, its frequency can reach 

533MHz.S3C6410 have excellent external memory 

interface ability ,can satisfy the bandwidth requirements 

of the high-end communications services. Memory 

system has DRAM and Flash/ROM two external memory 

ports, which can do parallel access. DRAM port can be 

configured to mobile DDR or standard SDRAM. 

Flash/ROM port supports NAND-Flash, NOR-Flash, 

OneNAND ,CFand ROM type external memory. 

Server part hardware principle diagram includes: 

S3C6410 processor, power, storage unit, Wi-Fi module, 

LCD monitor, buzzer ,mass storage, power supply 

module, clock module and RS232 serial interface. 

Hardware structure is shown as Figure 2. 

S3C6410 processor is responsible for the unit's 

control, computation, processing and other funtions; its 

storage unit is SDRAM and FLASH, Wi-Fi module 

mainly complete data transmission;LCD display the 

received data;when the image is abnormal,buzzer will 

alarm;the mass storage has large enough storage capacity 

to let the receive data stored,played back and 

processed;power module supply DC 5V for CPU and 

other modules as power supply. 

Figure 2. Hardware structure 

B. Server-side software design 

The software framework of this system is mainly 

based on embedded Windows CE operating system to 

realize system functions in EVC graphical interface 

development environments [8] . The server-side's 

application development process uses the program design 

of multithreading modular, makes each function of the 

equipment according to its completing a specific task, 

tasks properties, real-time demand, combining with the 

data flow, detailed differentiate each functional modules. 

In the software design of this system ,there are basically


the following modules: system interface, operating log, 

network communication, database storage, historical data 

(1) WinCE operation system 

Windows CE is having a embedded operating system 

which has the preemptive multi-tasking function, and has 

strong communication ability. It is a newly designed 

operating system products which Microsoft designed 

especially for mobile equipment and consumer electronic 

products, embedded application such as non-PC field. 

Therefore according to its features of application 

environment, Windows CE is designed with highly 

modular, good real-time, strong communication ability, 

supportting multiple CPU embedded operating system. 

Through the platform cutting tool Visual Studio.NET we 

establish the operating system platform according to 

system needs and carry on the corresponding 

configuration. This design uses the Windows CE6.5 

version. 

(2) Device driver development 

WinCE driver is divided into: this machine device 

drivers and stream interface drivers. This machine 

equipment means integrated into the target platform,its 

driver is provided by the original equipment manufacturer 

(OEM) .Stream interface drivers refers to the external 

device drivers connected to the WinCE operation into 

platform,developed by the user.Stream interface drivers 

see the external devices as special file in the file 

system,through document reading function indirectly 

access external devices. 

(3) Network communications introduction 

The network communication module is the main part of 

the system, it contains three data channels: Monitor 

channel, control channel and video data channel.Monitor 

channel is used to transmit command data to the control 

sle equipment;Video data channel is used to transmit 

video data of each group. Three channels use different 

communication port, so each channel transmits data 

independently of each other.Network communication 

module design development is through the network 

programming interface WindowsSocket, abbreviation 

Winsock to fulfill. According to the system 


Figure 3. Software structure 

query and other parts. Software structure is shown as 

Figure 3. 

browser/server's network transmission model, in the 

server-side the SOCKET type of listening SOCKET, 

control SOCKET are established,in the client-side the 

SOCKET type of request SOCKET,control SOCKET are 

established,they all use TCP protocol to encapsulate the 

transfer data. In addition, both ends of the server and 

client use a multicast class (CMulticast), it is the 

packaged class specially for video transmission, derived 

from CObject,which defines SOCKET and group 

SOCKET of the transceiver video data of SOCKET 

type,thus realized using multicast communication mode 

to transmit the video packets encapsulated by the UDP 

protocol. Network data communication process is shown 

in figure 4. 

Figure 4. Network data communication processes 

(4)Application design 

The application layer on server-side of this system is 

developed on the basis of graphics development 

environment of Embedded Visual C++ 4.0.EVC is the 

mainstream development tool on WINCE,it is similar to


VC++ development environment,but provides some 

unique tools and resources to develop the applications on 

WINCE [9] .EVC program support a subset of MFC 

library,can give developers to provide the most powerful 

support. 

For the application development process of the server 

terminal we adopt the modularized program design,make 

each function of the equipment,according to its 

completing a specific task,tasks properties,real-time 

demand,combining with the data flow,detailed 

differentiate each functional modules.Based on the 

system function requirements, embedded server terminal 

has the following several main functional modules:data 

transceiver module,judge alarm module, database storage 

module,real-time/history video data inquiry module, 

system interface module. 

The server-side software uses multithreading to 

realize modular:main program, network communication 

thread (transceiver thread),SMS sending thread,alarm 

thread etc.Main program is first executed,after system is 

initialized, it configurate 3G module,and then creates 

three thread respectively executed follow: receiving data 

storage and display data and judgment alarm, three 

threads are finished, main program exit. The alarm thread 

compares surveillance video data state with alarm 

threshold,if overrun the buzzer is started,and then it waits 

for police clearance to determine whether the exit. When 

abnormal, but no real-time monitoring on the 

client-side,SMS sending thread will send a warning 

message to the client. Software flow chart is shown in 

Figure 5. 

Application software started, first we do the 

initialized work, including initializing the 3G network 

module, recording log books, opening the database etc 

tasks, if the initialization is fails ,the applications cannot 

run normally to exit. After the initialization is successful , 

the server has been monitoring the family status, and 


Figure 5. Main software process of the server 

(a) System interface. The system interface is a 

man-machine interface which the user use to view the 

sampled video, a well-designed system user interface 

helps users conveniently carry on equipment operation, 

data storage and querying historical video data 

locally .Software interface uses hierarchical structures, 

each test is processed separately by the metord of the 

pop-up sub-interface ,so it makes the software for more 

modular management .The whole interface framework of 

the software including the menu bar ,the toolbars ,the 

button ,the edit box, the main window status bar and 

other parts. 

(b) Operation log.The operation log records the user 

name, time, instructions and other detailed informations 

of users' equipment operation. Log information is stored 

in a text files to search easily. 

(c) Data communication module. The data 

communication module realize the data communication 

function through 3G wireless network, receiving the 

control instruction and sending the video data. 

(d) Database stores. The data storage finish the 

database storage of the real-time data ,a good database 

stores scheme not only save system memory ,disk and 

other hardware overhead ,but also can improve the 

efficiency of the database query [10] . 

(e) Historical data query. The historical data query 

module completes the operations of querying the 

historical video data information, by observing the 

historical data,the situation when no one at home can be 

analyzed. The historical data query model including form 

code design and database query code design. 

waits or keeps communications with the client ,when 

closing the equipment ,the real-time monitoring is over, 

otherwise have been circulating. 

V. CONCLUSIONS 

Base on the remote real-time monitoring system of


3G family conditions, the Web server of the embedded 

video network monitoring system directly connect to 3G 

networks, without restrictions of the cable length and 

signal attenuation, simultaneously, the network is a no 

distance concept, which completely abandoned regional 

concepts, expanded the supervised areas. The client is an 

ordinary 3G mobile phone, anywhere and anytime it can 

carry on the real-time monitoring for family safety 

condition. Also, because video compression and Web 

function are focused into a small size equipment inside, it 

directly connects to the LAN or WAN network, 

plug-and-see, the system's real-time, stability and 

reliability are greatly improved, without specially 

assigned management, which is very suitable for the 

environment of non-People on duty. With the rapid 

development of the computer technology and the network 

technology, people's request on the video monitoring 

system will be higher and higher .This system not only in 

the family status remote real-time monitoring system has 

important role, but also in the e-commerce, the video 

conference, the long-distance teaching, remote medical 

treatment ,water conservancy and electricity monitoring 

etc has broad application prospects. 


This work was supported by Industrial Science and 

Technology Department of Shaanxi Province PR project 

“Research on Forward Multi-user MIMO Transmitting 

Technologies for Distributed Wireless Communication 

Systems” NO: 2010K01-074 . 

REFERENCES 

[1] Hu. Huai, Jia. Jing, Chang .Yong, “Ship remote monitoring 

system based on 3G technology,” Ship Science and 

Technology, 2009,31 (02) , pp. 101-103. 

[2] Liu. Bo, Song. Junde, “Research on 3G service platform 

architecture ,” Mobile Communications, 2005, 29 (9) , pp. 

60-63. 

[3] Cheng .Juan, Ping. Xijian, “ Embedded on-board geographic 

information system of integrated GPRS service ,”Computer 

Engineering, 2006,32 (17) , pp.244-245. 

[4] Liu .Wei, Chen. Hexin, “The family's remote monitoring 

system design,” Communication Technology, 2009,42 (01) , 

pp.312-313. 

[5] Sun. Jiangbo, “Video monitoring system design based on 

embedded Linux,” Journal of the Wuhan Polytechnic 

University, 2006,25 (3) , pp.31-36. 

[6] Larry Doolittle, Jon Nelson. BoaWebserver [EB / OL]. 

Http://www.boa.org. 

[7] Zhang. Xihuang, Chai .Zhilei, “The features and realization 

of CGI in embedded Web server,” Mini-Micro Computer 

Systems, 2003,24 (11) , pp.2046-2047. 

[8] Zhou .Yulin, Windows CE. Net kernel’s customization and 

application development , Beijing: Electronic Industry Press, 

2005, pp.17 - 21. 

[9] Wang. Bing, EVC senior programming and application 

development . Beijing: China Water Conservancy and 

Hydropower Press, 2005. 

[10] Chen. Jinjiang, Development and application of the 

embedded database technology in coal mine rig monitoring 

system , Kunming: Kunming University of Technology, 

2007. 


Qian Zhao was born in Shannxi, 

China ,in 1977.He received the M.S. 

degrees in communication and information 

system from Xi’an University of Science 

and Technology, Shannxi ,China , in 2005. 

He is a lecturer in School of 

Communication and Information 

Engineering at Xi’an University of 

Science and Technology. He has published over thirty papers in 

academic journals and international conference. His research 

interests include B3G Mobile communications key technology.


Information Fusion Based Fault Location 

Technology for Distribution Network 

1, 2 

Qingle Pang 

1. School of Information and Electronic Engineering, Shandong Institute of Business and Technology, Yantai, China 

2. College of Computer Science, Liaocheng University, Liaocheng, China 

Email: stefam@163.com 

Abstract—Along with the increasing the level of distribution 

network intelligence and the network complexity, the 

automatic fault location technology for distribution network 

is particularly important. But the traditional fault location 

methods based on one information source is impossible to 

locate the faults accurately because of there are losses or 

faults in the information from the distribution. So, the 

information fusion based fault location method for 

distribution network is proposed. When a fault occurs, one 

information matrix is created based on the action of all the 

protective relays, and the other information matrix is 

created based on the wave data of the current which is 

recorded at the foot node. The above information matrixes 

are combined using D-S evidence theory and the fault 

location is realized. The simulation results show that the 

fault location method for distribution network not only 

realizes accurate fault location, but also possesses stronger 

robustness. 

Index Terms—distributed network, fault location, 

information fusion, D-S evidence theory 


With the increasing complexity of the distribution 

network, feeder automation has a rapid development. 

Feeder automation provides strong support for fault 

location in distribution network. At the same time, it also 

provides strong support for improving the quality and 

reliability of power supply [1-3] . The traditional fault 

location methods based on coordination between recloser 

and sectionalizer exist many shortcomings, such as, 

reclosing to fault point again, slow response rate and 

expanding fault regions easily. So the fault section 

location based on information of protection units and 

feeder terminal units(FTUs) are becoming research 

hotspots. Where the matrix algorithm [1, 2] and the over 

heated arcs searching algorithm [4, 5] have the simple 

principle and well definition. But these methods are based 

on complete and accurate information uploaded by FTUs. 

The intelligent fault location algorithm can realize fault 

section location at the case of exiting uncertain 

information, such as, genetic algorithm based fault 

location [6] , rough set theory based fault location [7, 8] , 

multi-agent based fault location [9] , Bayes probability 

based fault location [10] and Petri-Nets based fault 

location [11] . But these methods can only provide a certain 


doi:10.4304/jsw.6.5.826-833 

diagnosis result and the diagnosis result have the 

following disadvantages: 

• It cannot reflect the fitness of the method. When the 

quality of communication is good, the information based 

on SCADA is reliable. But when the communication 

channel is disturbed, the information isn’t reliable and the 

diagnosis result may be wrong. The difference cannot be 

reflected. 

• It cannot provide the failure indication information. 

• It makes against combination of different fault 

section location methods. 

In one word, there are two problems for the control 

center to resolve the fault location with the information 

uploaded by the feeder automation system: 

• For the vast amount and redundancy of the 

information, it is hard to process the data manually for the 

operators. 

• Under the influence of the work environment of the 

equipment and the communication system, there are 

always some losses or faults in the information from the 

distribution network. 

As a result, it is very important to syncretize the 

information from different information resources and find 

a new fault location method to locate the fault accurately 

even there are losses or faults in the information from the 

distribution. 

Dempster-Shafer(D-S) evidence theory is one of the 

mathematical tools developed in the 70s. D-S evidence 

theory can robustly deal with incomplete data. The D-S 

evidence theory can be a tool for system modeling and 

information fusion. In fault diagnosis, because different 

evidences make different contributions to different faults, 

evidence importance should be considered for specific 

fault diagnosis through multi-resource information 

fusion [12, 13] . 

In this paper, in order to improve the fault location 

accuracy and robustness, through constructing the 

information matrixes based on the action of all protective 

relays switch and the wave data of the current 

respectively, the fault information are combined using the 

D-S evidence theory. 

II. PROBLEM OF FAULT LOCATION BASED ON RECLOSERS 

AND SECTIONALIZERS


A. Principle of Fault Location Based on Reclosers and 

Sectionalizers 

Now, there are three methods to implement the feeder 

automation, that is, mutual coordination of recloser with 

voltage-time type of sectionalizer, recloser with recloser 

and recloser with over-current pulse count type 

sectionalizer. Among them, the mutual coordination of 

recloser with voltage-time type of sectionalizer is the 

most popular one [14] . 

The Figure 1 shows the schematic diagram of fault 

isolation based on mutual coordination of recloser with 

voltage-time type of sectionalizer for typical radial 

distribution networks, where A is a recloser and B, C, D 

are sectionalizers. The first reclosing time of the recloser 

is 15s and the second reclosing time is 5s. The X-time 

limits of the sectionalizers B and D are 7s and the Y-time 

limits are 5s. The X-time limits of the sectionalizers C 

and E are 14s and the Y-time limits are 5s. The operation 

logics of the recloser and senctionalizers are showed as 

Figure 2. 


(a) 

(b) 

(c) 

(d) 

(e) 

(f) 

(g) 

Opening state of senctionalizer 

Closing state of senctionalizer 

Opening state of recloser 

Closing state of recloser 

Figure 1. Schematic diagram of typical radial distribution networks 

A 

B 

C 

D 

E 

15s 5s 

7s 

X 

5s 

Y 

7s 

X 

14s 

X 

5s 

Y 

5s 

Y 

7s 

X 

5s 

Y 

7s 

X 

5s 

Y 

Note: High is opening, low is closing 

14s 

X 

Figure 2. Scheme of operation logic of all switches 

The figure 1 (a)-(g) show the process of the fault 

section location. The figure 1 (a) is the normal operation 

state for the radial distribution networks. The figure 1 (b) 

shows that when a permanent fault happens in the c 

section, the recloser A trips off, and it leads to losing 

voltage of the line and the sectionalizers B, C, D and E 

tripping. The figure 1 (c) shows the recloser A performs 

the first reclosing operation after 15s of the fault tripping. 

The figure 1 (d) shows that after 7s time limits, the 

senctionalizer B recloses automatically and the power 

supply arrives at the b section. The figure 1 (e) shows that 

after another 7s time limits, the senctionalizer D recloses 

automatically and the power supply arrives at the d 

section. The figure 1 (f) shows that after 14s time limits 

5s 

Y


after the senctionalizer B reclosing, the senctionalizer C 

recloses automatically. Because the fault of c section is 

the permanent fault, the recloser A trips off again and 

leads to the line losing voltage and the senctionalizer B, C, 

D and E tripping again. Because the senctionalizer C 

loses voltage before its Y-time limit (5s), the 

senctionalizer is locked. The figure 1 (g) shows that after 

5s time limits after the recloser A trips off again, the 

recloser A recloses secondly and the senctionalizer B, D 

and E s recloses automatically. Because the senctionalizer 

C is locked in the switching-off state, the fault section is 

isolation and the power supply of the non-fault area is 

restored. 

B. Problem of Fault Location Based on Reclosers and 

Sectionalizers 

After fault isolation, the operation logics of switches 

are sent to the control center through the intelligent 

distributed terminal, the control center determines the 

fault location to remove the fault source. But the 

information received by the control center may be wrong 

because of existing strong electromagnetic interference in 

the communications network of the distribution network. 

The control center will make the wrong decision. 

III. FAULT LOCATION METHOD BASED ON INFORMATION 

FUSION 

A. Fault Location Method Based on D-S Evidence 

Theory 

Let Θ be the frame of discernment i.e. the finite set of 

N mutually exclusive and exhaustive hypotheses. 2 Θ is 

the power set of Θ, such that if Θ={1, 2, …, N}, then 

2 Θ ={Φ, {1}, {2}, …, {N}, {1,2}, {1,3}, …, {N-1, N}, 

{1,2,3}, Θ},where Φ denotes the empty set. 

Definition 1 A basic probability assignment is a function 

m: 2 Θ →[0, 1], which satisfies the following conditions: 

⎧ 

⎪∑ 

mA ( ) = 1 

⎨ A⊂Θ 

(1) 

⎪ ⎩ m( φ) 

= 0 

m(A) is called basic probability number. It represents the 

proportion of all relevant and available evidence that 

supports the claim that a particular element of Θ belongs 

to the set A but to no particular subset of A. 

Definition 2 The plausibility function is defined as: 

Θ 

Pl :2 → [0,1] and Pl( A) = ∑ m( B) 

(2) 

B∩A≠φ The belief function Bel(A) measures the total amount 

of probability that must be distributed among the 

elements of A. It reflects inevitability and signifies the 

total degree of belief of A and constitutes a lower limit 

function on the probability of A. On the other hand, the 

plausibility function Pl(A) measures the maximal amount 

in A. It describes the total belief degree related to A and 

constitutes an upper limit function on the probability of 

A. 

Suppose m1 and m2 are two basic probability 

assignment functions formed based on information 

obtained from two different information sources in the 


same frame of discernment Θ. According to Dempster’s 

orthogonal rule of evidence combination, the combination 

of m1 and m2 is as follows: 

∑ 

∑ 

m1( A) m2( B) 

A∩ B= C 

m1⊕ m2( C) 

= 

1 − m ( A) m ( B) 

A∩ B= 

φ 

1 2 

The belief values of the action information of reclosers 

and sectionalizers and the current wave information at 

root node is regarded as m1 and m2 respectively. Then the 

m1 and m2 are combined according to equation (3) and the 

belief value of every protection zone is obtained. 

According to the belief value m, the fault location can be 

determined. The scheme of fault location method based 

on D-S evidence theory is showed as Figure 3. 

Figure 3. Scheme of fault location method based on D-S evidence 

theory 

B. Model of Distribution Network 

The simulation model is a 17-node 24.9kV distribution 

network with feeder automation, which is built in 

PSCAD/EMTDC, showed as Figure 4. R1-R4 indicate the 

reclosers and D1-D9 indicate the sectionalizers [16] . 

Figure 4. Simulation model of distribution network 

C. Belief Assignment of FTU Information 

According to the actions of reclosers and 

sectionalizers, the information matrix of the reclosers and 

sectionalizers is defined as R and Dk,(k=1,2, …, u): 

⎡r11 ⎢ 

r21 R = ⎢ 

⎢� ⎢ 

⎣ru1 r12 r22 � 

ru2 r13 r23 � 

ru3 r14 r24 � 

ru4 r15 

⎤ 

r 

⎥ 

25 ⎥ 

� ⎥ 

⎥ 

ru5⎦ 

(4) 

(3)


⎡dk01 dk02 dk03 dk04 dk05 

⎤ 

⎢ 

dk21 dk22 dk23 dk24 d 

⎥ 

k25 

Dk 

= ⎢ ⎥ 

⎢ � � � � � ⎥ 

⎢ ⎥ 

d d d d d 

(5) 

⎣ kv1 kv2 kv3 kv4 kv5⎦ 

where u is the number of the reclosers in the distributed 

network, v is the number of the sectionalizers in the 

corresponding recloser protection zone. In the 

information matrix, the first column indicates the first 

opening action, the second column indicates the first 

closing action, the third column indicates the second 

opening action, the forth column indicates the second 

closing action and the fifth column indicates the third 

opening action. Each row of the information matrix R 

corresponds to a recolser. The first row of the information 

matrix Dk indicates the switch actions of the 

corresponding recolser and the rows from 2 to u indicate 

the switch actions of the sectionalizers in corresponding 

recloser protection zone. rij=1 or dkij=1 indicates that the 

corresponding action happens and rij=0 or dkij=0 indicates 

that the corresponding action doesn’t happen. 

Suppose the actual information matrix of reclosers 

received by the control center from the distributed 

network is RA, the actual information matrix of 

sectionalizers in different protection zone received by the 

control center from the distributed network is DAk 

respectively. The matrix RA is compared with an m×5 

matrix which elements are all 1, and a statistics matrix 

RAS is obtained by counting the number of same 

elements in every row between the two matrixes. After 

the matrix RAS is normalized, the action probability 

matrix of reclosers PR is obtained. Similarly, the matrix 

DAk is compared with every ideal information matrix 

when faults occur in corresponding sectionalizer 

protection zone, and a statistics matrix DASk is obtained 

by counting the number of same elements between the 

two matrixes. The first row of DASk is set to 0. After the 

matrix DASk is normalized, the action probability matrix 

of sectionalizers PDk is obtained. 

Suppose the permanent fault occurs in the node 856. 

The reclosers R1, R2, R4 and the sectionalizers of their 

protection zone don’t act. Only R3, D4, D5, D6 and D7 

perform automatic switch-off and switch-on action. After 

several automatic switch-off and switch-on actions, the 

fault section is isolated and the power supply of the sound 

area is restored. The action time sequence diagram is 

showed as Figure 5. 

The Figure 5 shows the operation logic of the recloser 

and sectionalizers in R3 protection zone. Suppose the fault 

happens at 0.3s, the recloser R3 acts firstly, the four 

sectionalizers switch off because of losing voltage. After 

1.5s, the recloser recloses automatically. Then, the 

sectinalizers D4, D5, D6 and D7 reclose automatically in 

turn after itself 0.7s time limits, where the sectinalizers 

D4, D5 and D6 reclose successfully. When the 

sectionalizer D7 recloses, the fault is connected to the 

system again because the fault section is located in D7 

section, the recloser R3 switches off again and all the 

sectionalizers switches off because of losing voltage. 

Because the time interval between the two switching-on 


time of D7 is less than the setting time limit 0.5s, the 

sectionalizer is locked. After 0.5s, the recloser recloses 

automatically, and the sectinalizers D4, D5 and D6 

recloses automatically in turn after itself 0.7s time limits, 

where the sectinalizers D4, D5 and D6 recloses 

successfully. The locked sectionalizer D7 don’t reclose, 

the fault section is isolated and the power supply of the 

sound zone is restored. 

R3 

D4 

D5 

D6 

D7 

1 2 3 4 5 6 7 8 

1 2 3 4 5 6 7 8 

1 2 3 4 5 6 7 8 

1 2 3 4 5 6 7 8 

1 2 3 4 5 6 7 8 

Note: High is opening, low is closing 

Figure 5. Scheme of protection action operation logic 

The information matrix of the recloser R3 is followed: 

⎡0 0 0 0 0⎤ 

⎢ 

0 0 0 0 0 

⎥ 

R = ⎢ ⎥ 

⎢1 1 1 1 0⎥ 

⎢ ⎥ 

⎣0 0 0 0 0⎦ 

The information matrixes of the sectionalizers D4, D5, 

D6 and D7 are followed as: 

⎡0 0 0 0 0⎤ 

⎢ 

0 0 0 0 0 

⎥ 

D1 

= ⎢ ⎥ , D 2 = [ 0 0 0 0] 

⎢0 0 0 0 0⎥ 

⎢ ⎥ 

⎣0 0 0 0 0⎦ 

⎡1 1 1 1 0⎤ 

⎢ 

1 1 1 1 0 

⎥ 

⎢ ⎥ 

⎡0 0 0 0⎤ 

D3 

= ⎢1 1 1 1 0⎥, 

D4 

= 

⎢ 

0 0 0 0 

⎥ 

⎢ ⎥ 

⎢ ⎥ 

⎢1 1 1 1 0⎥ 

⎢⎣0 0 0 0⎥⎦ 

⎢ 

⎣1 1 1 0 0⎥ 

⎦ 

The above-mentioned information matrixes can be 

obtained by the control center, if the relosers and 

sectionalizers can act correctly and there are no 

misoperations, moreover, there are no disturbances in the 

communication channel. The location result is true. 

Under the influence of the work environment of the 

equipment and the communication system, there are 

always some losses or faults in the information from the 

distribution network. Suppose the actual information 

matrixes obtained by the control center are followed as:


⎡0 0 0 0 0⎤ 

⎢ 

0 0 0 0 0 

⎥ 

RA = ⎢ ⎥ 

⎢1 1 1 1 1⎥ 

⎢ ⎥ 

⎣0 0 0 0 0⎦ 

DA 2 = [ 0 0 0 0] 

, DA1 

⎡0 0 0 0 0⎤ 

⎢ 

1 0 0 0 1 

⎥ 

= ⎢ ⎥ 

⎢0 0 0 0 0⎥ 

⎢ ⎥ 

⎣0 0 1 0 0⎦ 

⎡1 1 1 1 0⎤ 

⎢ 

1 1 1 1 0 

⎥ 

⎢ ⎥ 

⎡0 0 0 0⎤ 

DA3 

= ⎢0 1 1 1 0⎥, 

DA4 

= 

⎢ 

0 0 0 0 

⎥ 

⎢ ⎥ 

⎢ ⎥ 

⎢1 1 0 1 1⎥ 

⎢⎣0 0 0 1⎥⎦ 

⎢ 

⎣1 1 1 1 1⎥ 

⎦ 

According to these wrong information matrixes, the 

first belief value m1 can be calculated as following. 

Comparing RA with a 4×5 matrix which elements are all 

1, a statistics matrix RAS is obtained by counting the 

number of same elements in every row between the two 

matrixes. After the matrix RAS is normalized, the action 

probability matrix of reclosers PR is obtained as following: 

⎡0⎤ ⎡0⎤ ⎢ 

0 

⎥ ⎢ 

RWS = ⎢ ⎥ 

0 

⎥ 

, PR 

= ⎢ ⎥ 

⎢5⎥ ⎢1⎥ ⎢ ⎥ ⎢ ⎥ 

⎣0⎦ ⎣0⎦ Every element in PR from top to down is the probability 

value of the recloser R1, R2, R3 and R4 respectively. 

Similarly, the matrix DA1, DA2, DA3 and DA4 is 

compared with the corresponding ideal information 

matrix when the fault occurs in corresponding 

sectionalizer protection zone respectively, and the 

statistics matrix DAS1, DAS2, DAS3 and DAS4 is obtained 

respectively by counting the number of same elements 

between the two matrixes. The first row of DASk is set to 

0. After the matrix DASk is normalized, the action 

probability matrix of sectionalizers PDk is obtained. 

These matrixes are as followed as: 

⎡ 0 ⎤ 

⎢ 

0.4348 

⎥ 

m11 = PD1= 

⎢ ⎥, 

m12 = PD2= 

[ 0] 

, 

⎢0.3043⎥ ⎢ ⎥ 

⎣0.2609⎦ ⎡ 0 ⎤ 

⎢ 

0.1935 

⎥ 

⎢ ⎥ 

⎡ 0 ⎤ 

m13 = PD ⎢ 3 = 0.2419⎥, 

m14 = PD2= 

⎢ 

0.6667 

⎥ 

⎢ ⎥ 

⎢ ⎥ 

⎢0.2581⎥ ⎢⎣0.3333⎥⎦ ⎢ 

⎣0.3056⎥ ⎦ 

So, the probability that the fault locates in recloser R1, 

R2, R3 and R4 is 0, 0, 1 and 0 respectively. If the fault 

locates in the recloser R1 protection zone, the locked 

probability of the recloser R1, the sectionalizer D1, D2 and 

D3 is 0, 0.4348, 0.3043 and 0.2609 respectively. There 

are no sectionalizers in the recloser R2 protection zone, so 

only recloser R2 is considered and its locked probability is 

0. If the fault locates in the recloser R3 protection zone, 

the locked probability of the recloser R3, the sectionalizer 

D4, D5, D6 and D7 is 0, 0.1935, 0.2419, 0.2581 and 0.3056 


respectively. If the fault locates in the recloser R4 

protection zone, the locked probability of the recloser R4, 

the sectionalizer D8 and D9 is 0, 0.6667 and 0.3333 

respectively. 

D. Belief Assignment of Fault Recorder Information 

When a permanent fault occurs and the corresponding 

protection acts, the actions of the reclosers will result in 

appearing a current pulse at the root node. So, two section 

fault current corresponding to two protection action at the 

root node can be detected. The diagram comparing 

operation logic of the recloser R3 with the current 

waveform is showed as Figure 6. The diagram shows that 

the time interval between two current pulses is 

corresponds with the time interval between two reclosing, 

and the time interval between two reclosing corresponds 

with the locked sectionalizer. So, through calculating the 

time interval between the two current pulses at the root 

node, the control center can determine the locked 

sectionalizer, and the fault zone can be located. 

I(kA) 

R3 

0.3 

0.2 

0.1 

0 

(a) Node 856 permanent fault 

0 1 2 3 4 5 6 

t/s 

1 

0 1 2 3 4 5 6 

t/s 

(b) Node 830 permanent fault 

Figure 6. Diagram comparing operation logic with the current 

waveform


The time interval T has relationship with the location 

of the locked sectionalizer. It shows as Figure 7. 

(a) Node 856 permanent fault 

(b) Node 830 permanent fault 

Figure 7. Relationship between time interval T and location of the 

locked sectionlizer 

The diagram shows that the function between the time 

interval of two reclosing actions and the setting X-time 

limit of the locked sectionalizer is as follows: 

T ≈ tc1+ tx× d 

(6) 

where T is the setting time interval between two tripping 

actions, tc1 is the setting time of the first reclosing and 

tc1=1.5s, tx is the X-time limit of sectionalizer and tx 

=0.7s, d indicates the position of the locked fault zone, 

for example, when the fault located at node 856, d is set 

at 4; when the fault located at node 830, d is set at 2. 

So, the second belief assignment m2 can be obtained 

through detecting the time interval between two section 

fault current at the root node, the process is followed as: 

Firstly, the time interval between two fault current 

pulses, showed as Figure 8. 


I(kA) 

R3 

Figure 8. Diagram of detection between two fault current pulses 

Secondly, a 1×u probability matrix PCk of the locked 

recloser or sectionalizer corresponding to the protection 

zone k is initialized to 0, where u is the maximal number 

of sectionalizers in the recloser protection zones. The 

locked probability of each recloser or sectionalizer is 

calculated according to the time interval T. The 

probability distributions are showed as Figure 9. 

Figure 9. Diagram of probability distributions of the locked 

element 

Finally, suppose u=5, according to T, the locked 

probability of recloser and sectionalizer is as follows. 

If T0) row of PCk is set to p and the 

element of the d row of PCk is set to 1-p. The probability 

matrixes of other protection zone are composed of before 

m row of the PCk, where m is the number of 

sectionalizers of corresponding to protection zone. 

For example, the detected time interval between two 

fault current pulses is 4.2926s, the belief distribution 

matrix is followed as:


⎡0⎤ ⎢ 

0 

⎥ 

⎢ ⎥ 

m23 = PC3= 

⎢0⎥ ⎢ ⎥ 

⎢0⎥ ⎢ 

⎣1⎥ ⎦ 

The belief matrix shows that the locked probability of 

the fourth sectionalizer is 1, and the locked probabilities 

of the other sectionalizers are 0. 

E. Fault Location Method Based on D-S Evidence 

Theory 

The probability matrix PDk and PCk is regarded as m1k 

and m2k respectively. Then the m1k and m2k are combined 

according to equation (3) and the probability assignment 

Pk of every protection zone is obtained. The element pri 

of the matrix PR multiplies the matrix Pk respectively and 

the final probability assignment matrix P is obtained. 

According to the matrix P, the fault location can be 

determined. 

IV. SIMULATION ANALYSE 

In the simulation model showed as Figure 4, suppose 

the permanent fault occurs in the D3 protection zone. The 

information matrixes received by the control center 

receive from the distribution network are as follows: 

⎡1 1 1 1 0⎤ 

⎡1 1 1 1 0⎤ 

⎢ 

0 0 0 0 0 

⎥ ⎢ 

RA ⎢ ⎥ 

0 1 1 1 1 

⎥ 

= , DA1 

= ⎢ ⎥ 

⎢0 0 0 0 0⎥ 

⎢1 1 1 1 0⎥ 

⎢ ⎥ ⎢ ⎥ 

⎣0 0 0 0 0⎦ 

⎣1 1 0 0 0⎦ 

DA 2 = [ 0 0 0 0 0] 

, 

⎡0 0 0 0 0⎤ 

⎢ 

1 0 0 0 0 

⎥ 

⎢ ⎥ 

⎡0 0 0 0 0⎤ 

DA3 

= ⎢0 0 0 0 0⎥ 

, DA4 

= 

⎢ 

1 0 0 0 0 

⎥ 

⎢ ⎥ 

⎢ ⎥ 

⎢1 0 1 0 1⎥ 

⎢⎣1 0 0 0 0⎥⎦ 

⎢ 

⎣1 0 0 1 1⎥ 

⎦ 

According to above principle, the action probability 

matrixes are as follows: 

⎡1⎤ ⎡ 0 ⎤ 

⎢ 

0 

⎥ ⎢ 

0.2667 

⎥ 

PR = ⎢ ⎥, m11 = PD1 = ⎢ ⎥, 

m12 = PD2 

= [ 0 ] , 

⎢0⎥ ⎢0.3333⎥ ⎢ ⎥ ⎢ ⎥ 

⎣0⎦ ⎣ 0.4 ⎦ 

⎡ 0 ⎤ 

⎢ 

0.3409 

⎥ 

⎢ ⎥ 

⎡ 0 ⎤ 

m13 = PD1= ⎢0.2727⎥, m14 = PD 

⎢ 

1 = 0.5882 

⎥ 

⎢ ⎥ 

⎢ ⎥ 

⎢0.2045⎥ ⎢⎣0.4118⎥⎦ ⎢ 

⎣0.1818⎥ ⎦ 

The current wave detected at root node 800 is showed as 

Figure 4. The calculated time T is 3.6514s. According to 

formula (7) and (8), the probability matrix PCk is 

obtained as follows: 


⎡ 0 ⎤ 

⎢ 

0 

⎥ 

m21 = PC1= ⎢ ⎥, 

m22 = PC2= 

[ 0] 

⎢ 0 ⎥ 

⎢ ⎥ 

⎣0.9977⎦ ⎡ 0 ⎤ 

⎢ 

0 

⎥ 

⎢ ⎥ 

⎡0⎤ m23 = PC3= ⎢ 0 ⎥, 

m24 = PC 

⎢ 

4 = 0 

⎥ 

⎢ ⎥ 

⎢ ⎥ 

⎢0.9977⎥ ⎢⎣0⎥⎦ ⎢ 

⎣0.0023⎥ ⎦ 

According to formula (3), the action probability matrix 

m=m1⊕ m2 is combined. The element pr1- pr4 of the 

matrix PR multiplies the matrix m1- m4 respectively and 

the final probability assignment matrix P is obtained. 

⎡0 0 0 1 0⎤ 

⎢ 

0 0 0 0 0 

⎥ 

P = ⎢ ⎥ 

⎢0 0 0 0 0⎥ 

⎢ ⎥ 

⎣0 0 0 0 0⎦ 

I(kA) 

Figure 10. Current wave at root node 800 

So, the fault position locates the sectionalizer D3 of the 

recloser R1 protection zone. The result of fault location is 

true. 

V. CONCLUSION 

The fault location method based on the information of 

reclosers and sectionalizers will be wrong when the 

electromagnetic interference exits in the communication 

system. Using D-S evidence theory, the information of 

reclosers and sectionalizers and the information of current 

wave at root node can be combined. The combined 

information can locate the fault position accurately. 


This research has been supported by National 

Postdoctoral Science Foundation of P.R. China 

(20090461204), Natural Science Foundation of Shandong 

Province (ZR2010EL030), Postdoctoral Innovative 

Projects of Shandong Province (200903066) and Colleges 

and universities in Shandong Province Science and 

Technology Plan Project (J09LG09). 

REFERENCES 

[1] J. Liu, J. L. N, and Y. Du, “A unified matrix algorithm for 

fault section detection and isolation in distribution system,”


Automation of Electric Power Systems, vol.23, no.1, pp.31- 

33, Jan. 1999. 

[2] F. G. Zhu, D. S. Sun, Y. B. Yao, and et al, “Optimized 

matrix arithmetic of line fault location based on field 

terminal unit,” Automation of Electric Power Systems, 

vol.24, no.8, pp.42-44, Aug. 2000. 

[3] W. H. Chen, C. W. Liu, and M. S. Tsai, "Fast fault section 

estimation in distribution substations using matrix-based 

cause-effect networks," Power Engineering Review, 

vol.21, no.8, pp.61-61, Aug. 2001. 

[4] J. Liu, Z. A. Wang, “A new approach identify faulty 

section in distribution system,” Journal of Xi’an Jiaotong 

University, vol.34, no.2, pp.7-10, Feb. 2000. 

[5] J. Liu, H. L. Cheng, and P. X. Bi, “A simplified model for 

distribution system,” Proceedings of the CSEE, vol.21, 

no.12, pp. 77-82, Dec. 2001. 

[6] Z. Z. Guo, B. Chen, C. P. Liu, and et al, “Fault location of 

distribution network based on genetic algorithm,” Power 

System Technology, vol.31, no.11, pp.88-92, Nov. 2007. 

[7] H. C. Shu, X. F. Sun, dand D. J. Si, “A study fault 

diagnosis in distribution line based on rough set theory,” 

Proceedings of CSEE, vol. 21, no.10, pp.73-77, 82, Oct. 

2001. 

[8] Y. M. Sun, Z. W. Liao, “Assessment of data mining model 

based on the different combination rough set with neural 

network for fault section diagnosis of distribution 

networks,” Automation of Electric Systems, vol.27, no.6, 

pp.31-35, Jun. 2003. 

[9] Q. L. Pang, H. L. Gao, and M. J. Xiang, “Multi-agent 

based fault location algorithm for smart distribution grid,” 

in Pproceedings of 10th IET International Conference on 

Developments in Power System Protection. DPSP 2010. 

The Hilton Deansgate, Manchester, UK, vol.2010, no. 

558CP, pp.55, Apr. 2010. 

[10] Y. Y. Wang, Y. Luo, and G. Y. Ru, “Fault location based 

on Bayes probability likelihood ratio for distribution 


networks,” Automation of Electric Power Systems, vol.29, 

no.19, pp.54-57, Sep. 2005. 

[11] X. Luo, M. Kezunovic, "Implementing fuzzy reasoning 

Petri-Nets for fault section estimation," IEEE Transactions 

on Power Delivery, vol.23, no.2, pp.676-685, Apr. 2008 

[12] G. H. Zhang, M. Y. Duan, J. H. Zhang, and et al, “Power 

system risk assessment based on the evidence theory and 

utility theory,” Automation of Electric Power Systems, 

vol.33, no.23, pp. 1-4, 47, Dec. 2009. 

[13] Y. Song, C. S. Wang, “N-k contingency identification 

method under double failure incident based on evidence 

and functional group decomposition,” Proceeding of the 

CSEE, vol.28, no.28, pp,47-53, Sept. 2008. 

[14] J. Liu, J. L. Ni, and Y. H. Deng, Distribution Automation 

System, Beijing: China Water Power Press, 1998. 

[15] J. Fei, and Y. D. Shan, “Study of automatic fault location 

system in the distribution networks,” Proceeding of the 

CSEE, vol.20, no.9, pp.32-34, 40, Sept. 2000. 

[16] Z. X. Han, Power System Analysis, Hangzhou: Zhejiang 

Uniniversity Press, 2005. 

Qingle Pang was born in Liaocheng, China 

on October 28, 1969. He received the B.Sc. 

degree in electrical technology from 

Shandong University of Technology, China 

in 1994, and his Ph.D. in control theory and 

control engineering from Shandong 

University, China in 2007. He was an 

associate professor in the school of 

Information and Electronic Engineering, Shandong Institute of 

Business and Technology, China. He is now working in 

electrical engineering post-doctoral research station from 

Shandong University. His research interests are smart 

distribution grid, power system protection, fault detection and 

location for distribution network. He is an author and coauthor 

of more than 30 journal papers and conference proceedings.


Study on Remote Aided Diagnosis System of 

Mental Health Base on Export Knowledge Base 

Xiaoyong Wang 

Intelligent Control Institute Zhejiang Wanli University, P.R.China 

wxy0574@126.com 

Abstract—By the developing and maturing of remote 

communications technology, computer network technology 

and multimedia technology , internet applications has 

passed into a new phase. With the fast developing of the 

world economy, morbidity rate of morbidity rate in every 

country shows an upward tendency year by year. According 

to the curiative features of mental disease, application 

principles of expert system and the idea of object oriented 

programming, the paper expatiates the design proposal of 

remote aided diagnosis system base on the Browser-Server 

three-layer architecture. This system takes Internet as 

platform, puts forth the expendable knowledge base 

structure of mental illness professional field that is suitable 

for this system, by the object oriented analytic technique. 

And this structure can be mapped the relational database to 

set up relevant expert knowledge database. Which breaks 

the geographical limitation of mental health test, has good 

extensibility and usability, and will give a lot of assistance to 

the diagnosis of mental health. 

Index Terms—expert knowledge database, remote 

communications technology, computer network technology, 

Aided Diagnosis System 


Nowadays , There's a high incidence of mental 

diseases in every country all around the world, and 

according to statistics, mental diseases occupies eight of 

the ten diseases that deprive of strength or ability the 

most presently. And 70 percent of the depression patients 

cant not get effective treatment for various reasons. The 

suicide rate grows rapidly in the recent decades. 

According to the World Health Organization, there was a 

60 percent increase of suicide rate in the past fifty years. 

The number of suiciders increased the fastest, and a great 

part of the suiciders died of depression. Mind-cure and 

medication are the main two treatments of mental 

diseases, and those diagnoses are still in a traditional 

experience-stage, and mainly depend on various clinical 

diagnostic indices and experimental results. Lack of 

diagnostic experience will undoubtedly affect the 

diagnostic results. General speaking, it will take a 

professional doctor several years to practice to 


doi:10.4304/jsw.6.5.834-841 

Yuefeng Fang 

Zhejiang Wanli University ,P.R.China 

fangyf@zwu.edu.cn 

accumulate certain experience. The research tasks of 

expert system knowledge dam in artificial intelligence are 

to explain and rearrange the expert knowledge of 

professional field, and set up man-machine system 

inferred and developed from these knowledges. It 

provides related knowledges to express the technology, 

inferring control-mechanism and problem solving 

strategy. Meanwhile, the developing of professional 

artificial intelligence Expert System often needs 

professional artificial intelligence development language 

and tool, e.g.PROLOG and LISP, andthe system weak in 

is very complicated openness and flexibility which will 

effect its comprehensive application a lot. How to set up a 

intelligent network application system that takes generalpurpose 

system as running platform, more open and 

flexible in architecture, and easier to infer and reason, is 

the direction deserving of study [1]. In recent years, by 

the development and maturation of remote 

communication technology, computer network 

technology and multimedia technology, these network 

database technology, middleware technology, COM 

technology and computing paradigm base on Internet and 

its Browser and Server provide some basic information to 

develop the investigation. 

II. PRESENTING OF PLAN AND DESIGN 

The design objective of remote aided diagnosis sstem 

of mental health base on export knowledge base is to 

make the best of existing network technology, computer 

technology and modern information technology to set up 

a geniusnet based on the professional knowledge of 

Psychological Medicine, which can realize computerassisted 

instruction while mental health diagnosing and 

explain the relevanl knowledge. The consumers can 

remote access the system to get a new aids of mental 

health diagnosing that is more convenient, economical 

and practical. The main features are intelligence, 

remoteability, usability and extendibility. 

A. Logic and Structured Design of the System 

The system develops the intelligence of the inferring 

aided diagnosis under the help of some information of the 

AL expert System Field, and the logical structure design


bases on the fundamental structure of expert system and 

meets functional requirement, to set up the logical 

structure design model, showed as Fig.1. 

Figure 1. Logical structure of the system 

The model mainly contains the following four parts: 

knowledge base, intelligent heuristic mechanism, 

knowledge base management abbr: KBM, and customer 

interface abbr: UI. 

• Knowledge Base 

The professional field knowledge and its creating 

dynamic information mentioned in the system is kept in 

different types of knowledge base, such as large 

mathematical library, explanatory base, experimental rule 

base, medical symptoms rule base, comprehensive rule 

base and dynamic case base. 

• Intelligent Heuristic Mechanism 

It is used to realize the computer-aided diagnosis , 

equals to inference mechanism. 

• Knowledge Base Management ,abbr. KBM 

It is used to manage and service for the Knowledge 

Base. 

• Customer Interface abbr. UI 

It provides the customer with a convenient system user 

interface. 

B. System Architecture Design 

In order to manage the functional modules better, and 

based on the Logical Structure, a Browser/Server-typed 

three-layer architecture is set up by Component Object 

Model technology, showed as Fig .2. 

Figure 2. Architecture frame of the system and the UAP 


Essentially speaking, the system is to use the remote 

data information dynamically, and do some related logic 

reasoning. That's to say, it is a service system that 

provided by the Internet and for the dynamic use of 

medical data information. So the relative ideal model to 

analyse the system architecture is that the client ask the 

public server for a service, then the server chooses a 

proper services application program automatically and 

feedback the results of execution to the client. Traditional 

Client-Server system has many limitations, such as nontelescopic, 

hard-to-manage, hard-to-upgrade, hard to 

cross the platform, and poor performance, and so on. 

While combining with COM component technology and 

based on Browser-Server, the three-tier architecture the 

weakness of the traditional system. It adds WEB server, a 

new layer to the traditional Client-Server architecture to 

realize a the three-tier architecture which includes 

presentation layer, business logic middle layer and db 

server layer. 

On presentation layer, the Client Browser downloads 

HTML pages from the WEB server to make the interface 

between the AP and consumers.And the client 

components in the page can exchange message with other 

client components or server components. 

On the business logic middle layer, the server-side 

components are packaged up as middle layer to run on 

the WEB. It separates the presentation logic from 

business logic and application logic to provide data 

calculating and accessing, and so on. 

On db server layer, all space that components can 

access form this layer, such as database system, mail 

servers and groupware Server, and so on. Unlike the 

traditional two-layer Client-Server structure, these 

services do not face to the customer service client directly, 

but be carried out by the server components and feedback 

the results of execution to the client. 

In this architecture, there is only formal logic of AP 

in the client that is browser. That figures out the 

disadvantage of maintaining, updating and platformcrossing 

of Client-Server model. The middle layer is to 

reflect and maintain the business logic, such as to send 

the information of presentation layer to db server layer, 

manage the complex data, and nalysis and monitor the 

spontaneous processes. Middle layer does not undertake 

the missions of presentation layer and data layer, but to 

link and coordinate the two. Therefore, the middle layer 

can circulate in different computer from the WEB 

Browser , and manage several requests from the WEB 

Browser user simultaneity. 

This architecture is suit for the design needs of 

remote aided diagnosis system of mental health base on 

export knowledge base, and realize the changing of 

remote information from static publishing to dynamic 

using. Meanwhile, the architecture raises the application 

efficiency of the data base on WEB, component method 

and componentize software realize the layering and layercomponenting 

of business logic, application logic and 

presentation logic. The system constructed on Internet 

and based on the three-layer logic system adopts the 

code-reusing technology to make the best use of service-


independent software so as to enhance the efficiency of 

software development. That will advance the 

maintainability, extendibility and flexibility of the 

software developed. 

C. Systematic Function Module Design 

By design objective, the schematic diagram of basic 

systematic function module is showed as the following 

Fig .3. 

Mental-illness-diagnosing module: It leads the user to 

make a whole diagnosis step by step, that is from 

symptoms observing, history taking, basic experimental 

checking to diagnoses giving, which shows the whole 

processes of diagnosing. 

Knowledge-explaining module: It includes 

classification explaining, medical principles explaining 

and question answering. The module also gives some 

experimental explanations and instructions of the 

professional knowledge of related fields, such as to 

expound terminology, clinical significance and 

interrelation, and so on. And that will give the user a lot 

of guidances and trainings. 

Knowledge management module: It includes the 

compiling of knowledge base and rule bank. This module 

mainly maintains and manages the domain knowledge 

and rules to set an interface for knowledge updating and 

consummating. 

Figure 3. The schematic diagram of basic systematic function module 

Ⅲ. DEVELOPMENT ENVIRONMENT-CHOOSING OF THE 

SYSTEM 

The system is a new three-layer architecture by the 

model technology of COM component object which is 

based on Browser-Server model. How to is choose the 

appropriate technology and tools to use and develop it is 


crucial to advance the running performance and 

developing efficiency of the system. 

The development of WEB's dynamic applications has 

its own characteristics. It runs the program of server and 

promises the customer to exchange with the server and 

access the back-end database and build HTML page 

dynamically. This development involves a large number 

of number of software tools, is much more delicate than 

that of the traditional Client-Server softwares. In dynamic 

WEB-development technology, the ASP technology of 

Microsoft and the component-object-matching dynamic 

WEB development method are more convenient and 

practical. So the system chooses the Visual InterDev of 

Microsoft to be the integrated development environment 

of the dynamic WEB-application development of ASP. 

The system's WEB thematic framework is developed 

under Visual InterDev. 

Visual InterDev is the dominant tools developed 

cooperating with ASP, and can run the mixed 

programming of VBScript, Javascript and HTML. It can 

integrate with many components developed by third-party 

tools, such as VJ++, VB and VC++, and so on, and the 

application programs developed has good compatibility. 

Furthermore, the system integrates the functions of 

accessing the related database and browsing pages, and so 

on, so that the development of application programs can 

be finished under the same environment, and do not need 

to switch back and forth from different tools, which is of 

great convenience. 

Moreover, to be the third-party tools, Visual J++ 

develops the reusable servers and client component 

objects for ASP. Visual J++ is the visual integrated 

development environment Java-programmed under 

Visual Studio by Microsoft. It promises the developers on 

Windows platform either to write the Java code by the 

Pure Java mode of Sun Company, or to write the more 

practical and efficient Java WEB application programs by 

taking the best use of Windows platform. And Microsoft 

also promises the cross call between Java and COM 

components, that is either to write COM components by 

the advantage of Java language, or to improve 

programming efficiency by adding the resources of 

Windows platform into Java application. The custom 

components used in the system is developed by VJ++, 

such as information input component of Client, like BSH 

inputctrl and zhenzhi nputctrl, and inference control 

components of Server, like reasonctrl. 

Ⅳ. PRESENTING OF PLAN AND DESIGN 

According to the structure of the logical component of 

the system in Fig.1, the system logically consists of four 

parts: knowledge base, knowledge base management, 

intelligent thought-provoking mechanisms and user 

interface.Knowledge base is actually a back-end database, 

storing the dynamic use of specialistic knowledge data. 

The knowledge needed in the field of the inference and 

the explanation, and the dynamic information produced in 

facts, rules and the dynamic operation process are stored 

here (ie, integrated database). Meanwhile, the back-end


database provides a basic database management for the 

maintenance and management of such information. 

Smart stimulating mechanism of the system is a logic 

to accomplish some of the specific business, such as the 

diagnostic process guidance. It implements part of the 

core functionality of the system. It needs to access and 

manipulate the knowledge base (ie, back-end database) 

frequently. So it should be put server-side to achieve it. 

One part runs on the server as the server-side script, while 

the other parts which are more complex are compiled into 

COM components by using VJ + +, and they work as 

server-side plug-in type objects. In other words, it 

functions as the connection between middle-tier 

application services and Web network servers. 

User interface is actually the one for the user and the 

system to interact with. This system runs the ASP page 

through the browser. When ASP technology is used and 

integrated scripting language and HTML for Web 

programming are combined with some custom ActiveX 

components, a dynamic interactive remote user interface 

will come into existence. 

Knowledge Base Management is responsible for the 

maintenance and management of datas in Knowledge 

Base. And it actually maintains and manages a back-end 

database, namely, completing some functions of Web 

database system. The system uses integrated ASP 

powerful Web database access features and server-side 

scripting to achieve its functionality. 

The following are the description of the various parts 

of the specific implementation technology. 

A. Establishment and Management of the Knowledge 

Base 

Knowledge is an important component of this 

component in realizing the remote diagnosis of mental 

health. After the exchange of experts in the field, the 

following are my analysis on the uses of the domain 

knowledge in this system: 

From knowledge-level point of view, the goal of the 

system-level knowledge is divided into descriptive 

knowledge (Kd) and process knowledge (Kp). 

Descriptive knowledge provides factual and conceptual 

knowledge in the field, specifically knowledge in the 

field of psychological medicine and concepts, such as: 

mental illness, laboratory examination, medical history or 

the concept note of the symptoms. Procedural knowledge 

is knowledge of reasoning, specifically the knowledge of 

guiding the rules and heuristic of medical diagnostic 

procedure in Medicine. 

From the knowledge representation point of view, the 

structural system of knowledge includes concepts, facts 

and rules, namely, K = C + F + R.(K: items; C: concept; 

F :facts; R :rules.). As the concept is usually included in 

the facts, actually knowledge base contains facts and 

rules. 

The rule is in the form of the following description: 

::=( , 

||...... 

,) 


In the description, the premise and conclusion are 

facts. Therefore, from the logical point of view of 

knowledge, there are only two kinds of predicating rules 

and facts in systemic knowledge base. However, from the 

data point of viewof the relational model of data, two 

kinds of relationships exist in the base, and they can 

create two kinds of tables, rules tables and fact tables. 

All the information needed and produced dynamically 

in the process of system running, including intermediate 

results and diagnostic records, are all stored in the 

comprehensive database (database server), for example, 

the target model structure of a patient. 

Through the analysis of the characteristics of domain 

knowledge and application of object-oriented data 

modeling methods, the database tables and structures in 

knowledge base are established [2]. there are three main 

types of data tables in Knowledge base, as shown in 

Table 1. 

TABLE I. TABLE CLASSIFICATION, FUNCTION AND COMPOSITION 

OF EXPERT KNOWLEDGE BASE 

By the usability, readability and expandability, which 

are needed while setting the knowledge base, the system 

chooses SQL Server to be the db server., which is the 

high-performance relational database management 

system produced by Microsoft. 

B. Data Access Implementation Technology of Expert 

Knowledge Base 

The accession and management to the knowledge base 

in the system are realized by the OLE DB and ADO 

technologies of Microsoft.


ADO technology synthesizes the advantages of 

Remote Data Objects, abbr. RDO, and Data Access 

Objects, abbr. DAO. The object hierarchy of ADO is 

showed as Fig .4. 

Figure 4. The object hierarchy of ADO 

ADO is supported by many programming languages 

and development environments,both of Visual J++ and 

Active Server Pages used in the system support it. And 

ADO can be used in VBScript, so it is suitable for the 

integration of WEB or database server-side very much.To 

realize the application of WEB, the system use the 

VBScript of Active Server Pages to support ADO [3]. 

For example, the linking of Diseases Table and 

displaying its contents by browser can be realized by the 

asp files including the following statement. 

 

 

 

 

 

 

Introduction of mental 

illness 

 

 

 

 

 

 

 

 

Title ID 

 

 

 

Title 

 

 


 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

C. The implementation process of the intelligent 

systemic stimulating mechanism 

The intelligent stimulating mechanism of the system is 

a specific business logic to accomplish. Namely, it is used 

to complete the process of a remote diagnosis, through 

the use of psychological knowledge base stored in the 

medical domain knowledge, heuristic knowledge and 

dynamics of factual information, combined with a certain 

degree of reasoning control strategies. 

With the help of the heuristic knowledge, the System 

combines data-driven and goal-driven to realize the 

intelligent diagnosis inspired by the mechanism. Here are 

two types of knowledge in the rules of the system. One is 

the solution-based rules.In the diagnosis process can 

provide direct guidance of heuristic knowledge rules, 

generally the experience of the psychological knowledge 

of medical experts. It can direct the diagnosis or draw 

conclusions prompt the diagnostic information needed for 

the next step, under the guidance of the existing 

dynamical factual information (patient history, symptoms, 

laboratory test results information). The other one is the 

testing rules. In these rules, the necessary diagnostic 

conditions for a certain kind of conclusion are given, and 

if the existing information is able to meet these conditions, 

the conclusions are drawn. The former is used in datadriven 

strategy, while the latter is used in goal-driven 

strategy. Intelligent instructive mechanisms in Fig.5 are 

put forward with the combination of both strategies [4]. 

From the figure, we can see that in the diagnosis process, 

patients have a a predetermined model, which contains a 

number of items to store and describe the properties of 

the existing dynamic factual information. the solutionbased 

rules are actually the use of rules-based data-driven 

strategy coupled with enlightening information. It makes 

full use of the link between the information of 

predetermined model s and the professional knowledge of 

the system. or directly gives conclusions or rule-based 

model of the scheduled expansion of the items contained 

in the attributes to obtain the diagnostic information 

necessary to promote the efficient and accurate diagnostic


process. When the system model information is not 

intended to provide adequate solution-based rules, we can 

combine goal-driven strategies, and make use of 

empirical rules of the knowledge to identify the 

conditions diagnosed in whole or in part to meet the 

intended model of existing information and diseases of 

the highest degree of confidence as a test target. If an 

exact match is given, the conclusions will be drawn. If it 

is partially matched, we should extend the items of a 

predetermined model to provide diagnostic information 

on the direction of the heuristic. After the extension of a 

predetermined model, we should start from the 

application of the solution-based rules, and repeat the 

process of the combining data-driven and goal-driven, 

and then constantly deepen the end and to complete the 

diagnostic process [5]. 

Figure 5. the map of the process of Intelligent inspired mechanism 

In the running process of the system, this part needs to 

access the knowledge base frequently (ie, back-end 

database) and complete the corresponding logical 

control.Considering the servering efficiency of the system, 

it is realized on the server-side. Some runs on the server 

as scripted server pages, and other complicated hardcore 

are used as plug-in object of the server written as COM 

components by VJ++. 

Here the paper takes reasonctrl, the user-defined 

functional components of the intelligent heuristic 


mechanism for example, to analyse the development 

processes of VJ++ components. 

• To set up ADL File 

When Java develops the COM components, the first 

step is to provide the definitions and interfaces for COM 

-Creatable Classes by Interface Description Language, 

abbr. IDL. 

The interfaces can only use the COM-Classes that can 

be mapped to Java, and OLE is a good choice to be 

compatible types and dual interface automatically. A dual 

interface component promises to be called both by vtable 

and IDispatch. And that promises the server objects 

written by Java to be accessed by other programs in 

different languages completely. 

To set up ADL Files for reasonctrl similar to the 

following form, 

#include 

#include 

[ 

uuid(24BCB100-C7DE-11D4-9423-00E04C67FDDB), 

version(1.0), 

] 

library reasonctrlLib 

{ 

importlib(“stdole2.tlb”); 

[ object, 


dual, 

pointer_default(unique) 

] 

interface Ireasonctrl:IDispatch 

{ 

import”oaidl.idl”; 

HRESULT reason([out] BSTR* pbstrResult ); 

...... 

}; 

[ 


JAVACLASS(“reasonctrl.reasonctrl”), 

PROGID(“reasonctrl.reasonctrl”), 

] 

coclass Creasonctrl 

{ 

[defualt] interface Ireasonctrl; 

...... 

}; 

}; 

• To set up type-library file 

That is to compile the IDL files into specific typelibrary 

files. IDL Compiler of Microsoft can compile the 

reasonctrl.idl files into reasonctrl.tlb ones., and then 

check and see if the contents of type -library meet the 

requirements or not by OLE Object Viewer. 

• To set up the shell-type of Java 

The next thing to do is to set up the shell-type of Java 

for, and type-conversing tools can do this job, such as 

Creasonctrl.class and Ireasonctrl.class. Creasonctrl is the 

shell-type of COM and by which, the reasonctrl samples 

of COM are set up. Ireasonctrl is the interface shell type


Xiaoyong Wang, who is an associate professor, was born in 

Hefei Anhui Province in 1968, got Master of Engineering in 

computer application technology in 2001 from Xidian 

University, Xi'an Shanxi Province. 

She is now doing researches and teaching activities mainly 

in Computer Intelligent-controlling and Application-software 

Development. She participated in WSEAS International 

Conference 2007 in Ningbo, Zhejiang, China, 2007. Main 

contributions include Research on key technology for the global 

goods tracing informational platform, 2008. A Fuzzy Clustering 

Algorithm for Intelligent Mining of Internet Texts , in Computer 

Simulation, Beijing, July 2009 

Yuefeng Fang, who was born in Chaohu Anhui province in 1 

965, is a Master and senior engineer, his main research is on Inf 

ormation Engineering 



Application of Fault Phenomenon Vector 

Distance Discriminance in Woodworking 

Machinery System Fault Diagnosis 

Yun-Jie Xu 

School of Engineering, Zhejiang Agricultural & Forestry University, Lin’an, China 

Email: xyj9000@163.com 

Shu-Dong Xiu 

School of technology, Zhejiang Agricultural & Forestry University, Lin’an, China 

Email: sdxiu@zafu.edu.cn 

Quan-Sheng Men and Liang Fang 

School of technology, Zhejiang Agricultural & Forestry University, Lin’an, China 

Email: {menqs, lfang}@zafu.edu.cn 

Abstract—Aiming at the problem of diagnosis difficulty 

caused by too many factors of woodworking machinery 

system, a kind of diagnosing method based on fault 

phenomenon was presented. The research on woodworking 

machinery system fault phenomenon space arrived at 

conclusion that the emergency of each fault phenomenon 

subject to 0-1 distribution. Therefore, phenomenon vector 

corresponding to each fault formed cluster whose 

accumulation point is expectation of vector. After exclusion 

of abnormal vectors, the distance discrimination was used to 

fault diagnosis to establish expert system based on fault 

phenomenon vector. The confirmed result was return back 

to fault database so that the system achieve self-learning of 

real-time diagnosis experiences. Finally, the example on 

wood-wool working equipment proves that the diagnostic 

method has characteristics of good real-time, simple 

operation and high diagnostic accuracy. 

Index Terms—woodworking machinery system, distance 

discrimination, fault phenomenon vector, fault diagnosis 


The diagnosis of composite fault occurred in 

woodworking machinery is a difficult challenge at 

present. It is hard to diagnose the composite fault exactly 

and comprehensively due to the diversity and influence of 

faults. The purpose of woodworking machinery fault 

diagnosis is to identify whether the technical state is 

normal and determine the nature and site of faults from 

information related to mechanics running. Its essence is 

to find a mapping from fault phenomenon space to fault 

space. In order to accurately find the relationship to 

maximize extend, many scholars research on increasingly 

complex woodworking machinery system and presented 

many fault diagnosis methods based on the idea of expert 




doi:10.4304/jsw.6.5.842-848 

system. The fault diagnosis of woodworking machinery is 

artificial neural network, grey model [1, 2] and Support 

vector machine (SVM). In back propagation artificial 

neural network (BP-ANN), traditional empirical risk 

minimization (ERM) is used on training data set to 

minimize the error. Support vector machine (SVM) based 

on statistical learning theory is used in many applications 

of machine learning because of its high accuracy and 

good generalization capabilities [3, 4].The expert system 

based on neural network and genetic algorithm has 

disadvantages of slow convergence speed of training of 

network or samples, so it is difficult to complete diagnose 

task that has high real-time requirements [5, 6, 7, 8, 9]. 

Although Levenberg-Marquardt (L-M) algorithm can 

overcome the shortcomings, L-M algorithm is a 

combination of gradient method and Gauss-Newton 

method. With t he aid of the approximate second 

derivative, the L-M algorithm is more efficient than t he 

gradient method. Concerned wit h t he t raining process 

and accuracy, the L-M algorithm is superior to vary 

learning rate BP-ANN and SVM [10, 11]. It greatly 

increased complexity of computation and difficulty of 

design. Fault diagnosis expert system based on fuzzy 

theory can describe system fuzzy state, but the key 

reasoning technology is still at the stage of theoretical 

study and far away from the application. In contrast, it is 

simple to design and realize traditional fault diagnosis 

expert system based on rules. However, expert system 

based on rules has two bottlenecks of rule making and 

knowledge acquisition [12,13]To resolve rule making 

problem with complex strategy solving rule is easy to 

return to complex algorithms as neural network and 

genetic algorithm, the problem becomes complex again. 

Machine learning played good effecting solving problem 

of knowledge acquisition, while the machine learning 

strategy is not universal and is prone to induce 

combination explosion. Through the research on


probability distribution of fault phenomenon and 

clustering characteristic of phenomenon group caused by 

faults, the paper applied the idea of distance 

discrimination in diagnose strategy. The diagnosis result 

was feedback to fault database, which provide good 

solution to solving two problems of traditional expert 

system based on rules, so that the complexity of system 

structure and software design difficulty greatly reduced 

and diagnostic efficiency and engineering practicability 

greatly enhanced. Finally, Monte Carlo sampling and 

example of hydraulic excavator proves that the diagnostic 


operation and high diagnostic accuracy. The specific 

arrangement of the paper is as follows: Section 2 builds 

mathematical model of fault phenomenon vector; Section 

3 determines key techniques of fault phenomenon vector 

distance discrimination method; Section 4 performs 

simulation verification of the method taking wood-wool 

working equipment as example; Section 5 concludes our 

work. 

II. FAULT MODEL ESTABLISHMENT 

A. Mathematical Description of Fault Vector 

Phenomenon 

There are many factors led to fault of woodworking 

machinery system, most of which are not major. Under 

the influence of many non-essential factors, the 

phenomenon represented by faults that caused by few 

major factors appeared random. There are following 

facts: a fault may lead to simultaneous multiple 

phenomenons; occurrence of a fault phenomenon may be 

caused by different fault; multiple possible fault 

phenomenon caused by a fault is not certain, but 

statistically law. 

Assume the faults of system are single fault. We can 

know from statistics that there have n types of fault in the 

running history of system S, which forms a fault 

set F = { Fi 

| 1≤ 

i ≤ n} 

, where Fi is the i-th type fault. We 

can also know that there are m types of fault phenomenon 

caused by n types fault, the set of which 

is I = { Ii 

| 1≤ 

i ≤ m} 

. 

Define vector D = ( d1, 

d2, 

�, 

dm) 

, di ∈ B = { 0, 

1} 

, 

( 1≤ 

i ≤ m) 

whose component is a boolean variable to 

represent a fault phenomenon group, which is called as 

fault phenomenon vector. Among them, di=1 says that the 

i-th phenomenon in set i occurs, di=0 says it does not 

occur. As phenomenon caused by a fault constitutes a 

vector, for the fault Fi, the set constituted by all possible 

fault phenomenon vectors is exactly a subspace of m 

dimensional boolean space, which is denoted as Vi. For 

different fault Fi and Fj ( i ≠ j ), the constituted 

subspaces Vi and Vj are not different, but there may be 

common ground. If Vi and Vj are basically same, and the 

spatial distribution of fault phenomenon is probably 

same, the fault Vi and Vj are in a fuzzy set, in other words, 

it is difficult to distinguish the two faults from the 


phenomenon. Each dimension of fault phenomenon is 

subject to 0-1 distribution and respectively has a 

expectation pi, then the fault phenomenon vector has a 

expectation µ . All fault phenomenon vector caused by Fi 

is the vector family around µ in the space. In other 

words, the fault phenomenon vector caused by each fault 

is a natural clustering whose accumulation point is the 

expectation vector. With the above definition, fault 

diagnosis becomes such a problem: given a fault 

phenomenon vector D, to determine i with a method, so 

that D ∈ Vi 

. 

B. Establishment of Expert System Model 

The whole fault diagnosis system is an expert system. 

The process of wood-wool working equipment fault 

diagnosis is shown in Fig. 1. Consists of three main 

stages flow is as follows: 

(a) Retrieving 

According to the current fault phenomenon and 

symptoms of the wood-wool working equipment, retrieve 

the similar case from a database. If the case is suited to 

the current fault phenomenon of the equipment 

Current state of 

equipments 

Fault 

database 

Make a conclusion of 

diagnose 

Retrieving Retrieving Judge Suited 

case 

case 

Revise index 

Store 

Similar 

case 

Reasoning 

machine 

Diagnosis success 

probability order 

table 

Synthesize Transfer 

Determine diagnosis 

success probability 

based on probability 

Fault phenomena 

vector 

Distance 

determinati 

on rule 

database 

Explain 

machine 

Make a conclusion 

of diagnose 

Figure 1. Fault diagnosis expert system model


completely, quote the case directly and make a 

conclusion. 

(b) Modifying 

It takes fault phenomenon vector as input and to reason 

with explained distance discrimination rules and past 

fault data. The reasoning machine issued a diagnostic 

probability order table of diagnosis result according to 

discrimination analysis rules, where the fault with 

maximum probability is the preferred result, and others 

are options by decreasing order of probability. If the case 

is not matched completely, The diagnostic probability 

order table will be available to maintenance personnel for 

reference of further confirm, use the Distance 

determination rule database, parts fault characteristic and 

actions record [16] etc. to Reasoning, adjust, rewrite, 

match and synthesize the case which has been retrieved 

according to the current fault phenomenon of equipment. 

(c) Storing 

Make the corrected case in keeping with the diagnosis 

of the current fault phenomenon, and make a conclusion. 

At the same time, the confirmed result will be fed back to 

fault database for record to prepare for the next diagnostic 

reference. 

The distance discrimination was used to fault diagnosis 

to establish expert system based on fault phenomenon 

vector. The core of fault diagnosis is that it can 

memorize/store the former fault, its environments and the 

process accurately, furthermore, it uses the past diagnosis 

experience, process and methods to complete the current 

diagnosis through analogy and association while 

diagnosing. Therefore, fault diagnosis based on fault 

phenomenon vector is a kind of methods realized through 

analogy [17, 18], and its design mode is to utilize the past 

designed case directly instead of the summary of design 

experience. 

III. KEY TECHNIQUES 

A. Rule-Based Diagnostic Expert Systems 

In the rule-based systems, knowledge is represented in 

the form of production rules. A rule describes the action 

that should be taken if a symptom is observed. The 

empirical association between premises and conclusions 

in the knowledge base is their main characteristic. These 

associations describe cause-effect relationships to 

determine logical event chains that were used to represent 

the propagation of complex phenomena. The general 

architecture of these systems includes domain 

independent components such as the rule representation, 

the inference engine and the explanation system. Basic 

structure of a classical rule-based expert system is shown 

in Fig. 2. 

Expert diagnosis experiences suitably formatted 

consists the basis for the classical expert system approach. 

Fault diagnosis requires domain specific knowledge 

formatted in a suitable knowledge representation scheme 

and an appropriate interface for the human-computer 

dialogue. In this system the possible symptoms of faults 

are presented to the user in a screen where the user can 

click the specific symptom in order to start a searching 


Fault 

database 

Knowledge 

base 

Distance 

determination rule 

database 

User 

interface 

maintenance 

personnel 

Inference 

engine 

Expert 

Figure 2. Basic structure of a rule-based expert system. 

process for the cause of the fault. Additional information 

about checking or measurements is used as input that, in 

combination with stored knowledge in the knowledge 

base guide to a conclusion [19, 20, 21, 23]. 

B. Reasoning Rules Formulation 

The formulation of rules needs to resolve problem of 

fault data table design. Table 1 is the designed fault data 

table of F1, where each line represents a fault 

phenomenon vector. 

Using the above method, we can build fault data table 

for each Fi. Each fault phenomenon obeys standard 0-1 

distribution, the value of which is shown in (1). The 

expectation of each phenomenon is pij, where i represent 

that the phenomenon is caused by the i-th fault; j 

F1 

TABLE I. 

DATA TABLE OF F1 

Phenomenon 

Number I1 I2 I3 I4 … Im 

1 1 0 0 1 … 1 

2 1 0 1 0 … 0 

3 0 0 1 1 … 0 

4 1 0 1 0 … 0 

5 1 0 1 1 … 0 

6 1 0 1 0 … 0 

7 1 0 1 1 … 0 

8 0 0 0 1 … 0 

9 1 0 0 0 … 0 

10 1 1 1 0 … 0 

… … … … … … … 

Total 1000 913 11 946 583 … 50


represents that this phenomenon is the j-th phenomenon 

in the set; N is the sample data amount of this fault; Iijt 

represents that the t-th fault phenomenon vector of fault i 

is caused by the j-th component. 

N 

∑ 

I ijt 

* 

µ i = ( p ij ) j= 

1, 

2, 

� , m 

t= 

1 

= ( 

N 

) j= 

1, 

2, 

�, 

m (1) 

The variance is as (2): 

2 

r = ( S ) 

i 

ij 

1 

= ( 

N −1 

j= 

1, 

2, 

�, 

m 

N 

∑( I ijt − 

t= 

1 

p 

ij 

) 

2 

) 

j= 

1, 

2, 

�, 

m 

Covariance between different phenomenons is as (3): 

∑ 

= [ 

= [ σ ] 

i 

iuv u× 

v= 

m× 

m 

N 

∑I 

iukI 

ivkP( 

Iiuk, 

Iivk 

) ] u× 

v= 

m× 

m 

k= 

1 

( iuk , ivk I I 

(2) 

(3) 

Where, P ) is the joint probability of two 

fault phenomenon, which has only four cases as (4): 

⎧P( 

0, 

0) 

⎪ 

P( 

0, 

1) 

P( I , I ) = ⎨ 

(4) 

iuk ivk 

⎪ P( 

1, 

0) 

⎪ 

⎩ P( 

1, 

1) 

The research on woodworking machinery system fault 

phenomenon space arrived at conclusion that the 

emergency of each fault phenomenon subject to 0-1 

independent and has the same distributions, That denoted 

as Ii1, Ii2 , . . . , Iin. With finite expected value 

* 

2 

µ i = E( 

Iij 

) and finite variance σ i = D( 

Iij 

) . 

Let Sn = Ii1 + Ii2 + … + Iin. 

2 

2 Sn 

σ i 

We know D( Sn 

) = nσ 

i , D( 

) = Also we 

n n 

S ∗ 

know that E( ) = µ 

n 

n 

. 

We know from the large number law,by chebyshev’s 

inequality, then for anyε > 0 , as (5): 

2 

⎛ Sn 

* ⎞ σ i 

P⎜ 

− µ i ≥ ε ⎟ ≤ . (5) 

2 

⎝ n ⎠ nε 

Thus, for fixed ε as (6): 

⎛ Sn 

* ⎞ 

P ⎜ − µ i ⎟ ≥ ε → 0 

(6) 

⎝ n ⎠ 

As n → ∞ . Equivalently as (7): 

⎛ Sn 

* ⎞ 

P ⎜ − µ i ⎟ < ε → 1. 

(7) 

⎝ n ⎠ 

That when the number of sample goes to infinity, the 

expectation limit of samples is equal to that of the overall 

is shown in (5-7) and sample variance is equal to that of 


overall, the covariance of sample is equal to that of 

overall [22]. 

As Table 1 shows, the expectation of I1 caused by F1 is 

equal to p11=913/1000 =0.913. Furthermore, the 

* 

expectation fault phenomenon vector of fault Fi is µ i , 

which is the accumulation point according to probability 

distribution in the space of all fault phenomenon caused 

by Fi. The discrimination analysis idea indicates that 

* 

when perform distance discrimination of all µ i and fault 

phenomenon vector to be diagnosed, then the fault 

phenomenon vector is possible belong to the x-th space. 

That is the probability that it caused by the x-th is the 

largest. In this way, the order result from little to large 

will led to sort of diagnose probability descending. The 

distance here can be Euclidean distance as (8), or be 

Mahalanobis distance as (9): 

* 

i 

Dis = µ − µ 

i 

= 

∑ ∞ 

j= 

1 

( d − p ) 

j 

T 

i 

ij 

∑ −1 

* 

( D − 

i 

i 

* 

Dis ( x, 

G) 

= ( D − ) µ ) 

i 

Where, ∑ −1 

i 

(8) 

µ (9) 

is the inverse matrix of covariance 

matrix. 

The Euclidean distance is intuitive, while the 

Mahalanobis distance needs to compare and discriminate 

the standard overall phenomenon caused by each fault, so 

to as reflect reality. In the practical application, 

Mahalanobis distance needs to know the inverse matrix 

of covariance matrix among all phenomenon, which 

involve inverse operation, so it ie relatively complex. 

C. Design of Learning Strategy 

When the system is built, we should summarize expert 

diagnosis experience and input. The automatic learning in 

system running process can add conformed fault 

phenomenon vector into fault database. The sort 

according to probability from large to little will cause 

misdiagnosis, which means it may be wrong to take the 

fault at the most front as diagnosis result. The 

discrimination may cause mistakes, which is the fact that 

can not mastered by people. If the empirical data is very 

rich, the possibility of misdiagnosis will be very small. 

As to mistakes, the system will be the second diagnosis, 

which is ranked second in the probability of failure as a 

diagnostic output, and so on. 

The storage form affects the problem solving 

efficiency, whereas regulation and evaluation affect the 

problem solving accuracy. The matching degree of the 

fault and fault phenomenon can be expressed as (10): 

s 

n 

∑ 

i= 

1 

2 

D ( c, 

c′ 

) = 1− 

W ( X −Y 

) / n (10) 

i 

Where D s is the matching degree of fault c and fault 

phenomenon c′ ; W i is the weight of characteristic 

parameter i; n is the number of all symptoms; X i and 

i 

i


Y i are respectively the initial character or the conclusion 

credibility of fault c and fault phenomenon c′ . 

If D s = 1 , it indicates that the fault and fault 

phenomenon are most suited, namely matched completely; 

if D s = 0 , the fault and fault phenomenon are 

completely different [19]. 

Experience data is not all valid. According to expert 

experience, the fault phenomenon vector that is obviously 

not corresponding to fault phenomenon correspond to a 

fault, which is identified as abnormal. The abnormal 

should not be discarded directly, but added into database 

after marking. The reason is that if the abnormal after a 

major problem in direct disposal will cause the system to 

continue to drop later, the system will be committing a 

serious error. When conducting distance discrimination, 

these abnormal data should be excluded to avoid affect of 

small probability abnormal on discrimination analysis. If 

this abnormal occurs frequently afterwards, the frequency 

of abnormal will naturally large. According to the 

abnormal determination formula (11), it will not still in 

the scope of abnormal. 

* 

D − µ i 

* 

≥ α%, 

D = ( 1, 

1, 

�, 

1) 

* 

m 

D 

(11) 

Where, α is the abnormal discrimination index that 

can be controlled. 

D. Diagnose Algorithm Design 

Diagnosis algorithm flow is as follows: 

Step 1: Input fault phenomenon to be diagnosed 

Dx = ( d1, 

d 2, 

�, 

d n ) . 

Step 2: For all fault Fi, i = 1, 2, 

�, 

n . 

(a) Compute expectation vector µ i of Fi with (1); 

(b) As to all fault phenomenon vector of Fi, to conduct 

abnormal discrimination with (11); 

(c)Using all abnormal vectors, re-compute expectation 

* 

µ i of Fi. 

Step 3: As far as to be diagnosed vector Dx, compute 

European (or Mahalanobis) distance Disi of each fault 

with (8) or (9). 

Step 4: Order Disi from small to large to obtain 

diagnose probability order table 

Pi k 

⎛ / 

⎜ 

⎝ i 

Step 5: The maintenance personnel confirm faults 

according to probability from small to large. 

Step 6: The confirm result is fed back to fault data 

table. 

k 

⎞ 

⎟ 

⎠ 

The diagnose probability order table 

n 

. 

Pi k 

⎛ / 

⎜ 

⎝ i 

k 

⎞ 

⎟ 

⎠ 

n 

means 

the probability P i of fault whose number is i 

k/ 

k sorted in 

the k-th position of the table. As to the diagnosis result, if 

it is caused by the i1-th fault, it indicates that the first 


diagnosis is successful. If it is caused by the i2-th fault, 

then the first diagnosis is failure and second diagnosis is 

successful. And so on. 

IV. MODEL SIMULATION 

A. Simulation Algorithm Design 

The simulation algorithm is based on the above 

diagnosis algorithm. The standard to measure its 

efficiency is diagnostic success rate DFRk of the k-th 

diagnosis and accumulative success rate DFR k , the 

definition of which is shown in (12) and (13). 

DFRk = nk 

/ M 

(12) 

Where, nk is the frequency of vector to be diagnosed 

after k times diagnosis; M is the total time of diagnosis. 

DFR 

= 

k 

n 

M 

k ∑ i 

i= 

1 (13) 

Where, DFR k is the percentage that fault be 

diagnosed after k times of diagnosis. Obviously, k=1 is 

the fault detection rate. k=2 is the probability that isolate 

fault to two elements, and so on. 

The simulation algorithm is as follows: 

Step 1: As to all faultsi = 1, 2, 

�, 

n , use Monte Carlo 

method to sample according to fault phenomenon vector. 

Each fault generates N groups of sample data. 

Step 2: Extract a fault x using random method and then 

extract a fault phenomenon vector Dx. 

Step 3: To diagnose with diagnosis algorithm and 

⎛ Pi ⎞ k/ 

present probability order table ⎜ 

i ⎟ . 

⎝ k ⎠n 

Step 4: Repeat Step 2 and Step 3 M times. 

Step 5: For k = 1, 2, 

�, 

n , statistical nk. Compute DFRk 

and DFR k , then output. 

B. Simulation Result Analysis 

As to MQ3130-type wood-wool working equipment 

system, there are total 7 typical faults: rolling bearing 

fault、eccentric disk fault、the gear and rack fault、tool 

change Spindle fault、crank-connecting rod mechanical 

fault 、 work piece installation fault and feed drive 

structures fault. That denoted as I1 , I 2, 

� , I7 

. The 

distribution parameter of system fault and corresponding 

phenomenon is shown in Table 2. 

Design fault database with the method of Table 2 and 

conduct simulation, where the distance discrimination use 

Euclidean distance. M=1000, N=1000. The abnormal 

discrimination index α = 30 . Each fault samples to 

generate 1000 vectors and extract 1000 samples for 

simulation. The output result is shown in Table 3 and 

Table 4.


TABLE II. 

DISTRIBUTION PARAMETER OF SYSTEM PARAMETER AND 

CORRESPONDING PHENOMENON 

Results of Table 3 show that the number of abnormal 

of each fault is little, which is consistent with actual 

situation. In Table 4, one time fault detection rate is as 

high as 0.852. The three times accumulative diagnosis 

TABLE III. 

NUMBER OF EXCLUDED ABNORMAL. 

Fault number F1 F2 F3 F4 F5 F6 F7 

Number of 

abnormal 

The k-th 

diagnosis 

F1 F2 F3 F4 F5 F6 F7 

I1 0.5 0.60 0.89 0.88 0.01 0.53 0.01 

I2 0.95 0.80 0.95 0.06 0.01 0.90 0.02 

I3 0.8 0.50 0.96 0.01 0.96 0.9 0.4 

I4 0.1 0.02 0.01 0.96 0.01 0.08 0.98 

I5 0.2 0.02 0.9 0.07 0.89 0.3 0.05 

I6 0.01 0.35 0.01 0.8 0.03 0.00 0.3 

I7 0.9 0.90 0.01 0.02 0.92 0.93 0.02 

I8 0.3 0.40 0.00 0.98 0.01 0.33 0.05 

I9 0.7 0.01 0.01 0.99 0.01 0.68 0.8 

I10 0.3 0.01 0.01 0.02 0.88 0.25 0.01 

1 0 0 25 23 2 1 

TABLE IV. 

DIAGNOSIS SIMULATION RESULT 

Diagnosis success 

probability 

Misdiagnosis 

probability 

Cumulative 

success rate 

1 0.852 0.148 0.852 

2 0.097 0.051 0.949 

3 0.030 0.021 0.979 

4 0.012 0.009 0.991 

5 0.009 0.000 1.000 

6 0.000 0.000 1.000 

7 0.000 0.000 1.000 

success rate is up to 0.979 when k=3, which means the 

probability that isolate fault to three elements can up to 

0.979. The reason is vector distribution parameters of F4 

and F5 is very close. From the above definition we can 

know that these three faults can be regarded as a fuzzy 

group. At the moment, we can regard them as a fault, so 

one time fault detection rate is up to 0.979. The data 


result analysis indicates that this kind of diagnosis 

method based fault phenomenon vector discrimination is 

effective. 

V. CONCLUSION 

The paper presented a kind of woodworking machinery 

system diagnosis method based on fault phenomenon 

vector discrimination analysis. Starting from the 

clustering characteristics of fault phenomenon vector, 

conduct reasoning rule design based on the idea of 

discrimination analysis idea. The expert system model 

was built using determination and exclusion of abnormal. 

Finally, simulation illustration of MQ3130-type woodwool 

working equipment proves that the diagnostic 


operation and high diagnostic accuracy. 

However, the technique is a new branch of artificial 

intelligence, so systemic fruits are still not abundant, 

theories are still not mature, and the research and 

application are still in the exploring stage. If we apply it 

in machinery fault diagnosis system, the techniques of 

fault phenomenon vector and fault, retrieving and 

matching, self-study method, etc. would need further 

improved. With the increasing complication of the 

equipment and systems, fault diagnosis based on fault 

phenomenon vector will become an effective method in 

the fault diagnosis realm. 


The authors wish to thank Shu-Dong Xiu. This work 

was supported in part by a grant from the Science and 

Technology Agency of Zhejiang Province General 

Program Project No. 2007C22080, China; Technology 

Agency of Zhejiang Province R & D Program Plan 

Project No. 2008C02006-1, China. 

REFERENCES 

[1] Li Zhang, Jian-Hua Luo, Su-Ying Yang, “Forecasting box 

office revenue of movies wvith BP neural network”, 

Expert Systems with Applications, vol.36, pp.6580-6587, 

April 2009. 

[2] Wann-Yih Wu, Shuo-Pei Chen, “A prediction method 

using the grey model GMC (1, n) combined with the grey 

relational analysis: a case study on Internet access 

population forecast”, Applied Mathematics and 

Computation, vol.169, pp.198-217, 2005. 

[3] Gavin C. Cawley and Nicola L.C. Talbot, “Fast exact 

leave-one-out cross-validation of sparse least-squares 

support vector machines”, Neural Networks, vol.17, pp. 

1467-1475, December 2004. 

[4] Xue-Cheng Xi, Aun-Neow Poo, Siaw-Kiang Chou, 

“Support vector regression model predictive control on a 

HVAC plant”, Control Engineering Practice, vol.15, 

pp.897-908, 2007. 

[5] Zogg D, Shafai E and Geering H P., “Fault diagnosis for 

heat pump swish parameter identification and clustering,” 

Control Engineering Practice, vol. 14, pp. 1435-1444, 

2006. 

[6] Yun-Jie Xu, Wen-Bin Li, “Forecasting of the total power 

of woodworking machinery based on SVM trained by 

GA,” 2010 The 2nd International Conference on Computer


and Automation Engineering, Vo.01, pp. 358-360, 

February 2010. 

[7] Yun-Jie Xu, Wen-Bin Li, “Fault diagnosis for gearbox 

based on genetic-SVM classifier,” 2010 The 2nd 

International Conference on Computer and Automation 

Engineering, Vo.01, pp. 361-363, February 2010. 

[8] Yun-Jie Xu, Shu-Dong Xiu, “Prediction of wear for wood 

planning tool based on genetic-SVM classifier,” 2010 

International Conference on Electrical and Control 

Engineering, Vol.01, pp.5834-5836, June 2010. 

[9] Yun-Jie Xu, Shu-Dong Xiu, “Accurate diagnosis of rolling 

bearing based on wavelet packet and genetic-support 

vector machine,” 2010 International Conference on 

Electrical and Control Engineering, Vol.01, pp.5589-5591, 

June 2010. 

[10] Guo Kui, Yu Dan, “Spreading L-M method of multiple 

reliability evaluation,” Reliability Engineering, Vol.04, 

pp.157-160, 2003. 

[11] Jie Yu, Yao-Lin Shi , Gui-Xiang Shen , Ya-zhou Jia, 

“Reliability evaluation on CNC lathes based on the 

modified L-M method,” Vol.16, pp.665-668, May 2009. 

[12] Jian-Pei Zhang, Zhong-Wei Li and Jing Yang, “A parallel 

SVM training algorithm on large-scale classification 

problems,” Proceedings of the Fourth International 

Conference on Machine Learning and Cybernetics, vol. 01, 

pp. 1637-1641, August 2005. 

[13] KANGY W., LI J., CAO G. Y., “Dynanue temperature 

model in go fan so fusing Least square support vector 

machines,” Journal of Power sources, vol. 179, pp.683- 

692, 2008. 

[14] Ling-Jun Li, Zhou-Suo Zhang, Zheng-Jia He, “Research of 

mechanical system fault diagnosis based on support vector 

data description,” Journal of Xi'an Jiaotong University, 

vol. 09, pp.910-913, 2003. 

[15] D. Wu, C.-W. Ma and S.-F. Du, “Influences of different 

damaged degrees of ieafminer-infected leaves on the nearinfrared 

spectral reflectance,” Trans. of the CSAE, vol. 23, 

no. 2, pp. 156-159, 2007. 

[16] Choy KL, Lee WB. Design of an intelligent supplier 

relationship management system: a hybrid case based 

neural network approach. Expert Systems with 

Application, vol.24, pp. 225-237, 2003. 

[17] Yang BS, Han T, Kim YS, “Integration of ART-Kohonen 

neural network and case-based reasoning for intelligent 

fault diagnosis,” Expert System with Applications, vol. 26, 

pp. 387-395, 2004. 

[18] Wen-Hong Li, Shao-Wen Sun, Qi Zhang, “Machinery fault 

diagnosis expert system based on case-based reasoning,” 

Journal of Chongqing University: English Edition, vol. 06, 

pp. 273-277, December. 2007. 

[19] Sajja, Akerkar, “Advanced knowledge based systems: 

model, applications & research,” TMRF e-Book, Vol.01, 

pp.50 -73, 2010. 

[20] Su Myat Marlar Soe and May Paing Paing Zaw, “Design 

and implementation of rule-based expert system for fault 

management,” World Academy of Science, Engineering 

and Technology 48, pp. 34-39, 2008. 


[21] Wei Liang, Mechanical fault diagnostics. Bei Jing: China 

coal industry publishing house. 2005. 

[22] Lefebvre, Mario, Applied Probability and Statistics 

[electronic book] by Mario Lefebvre. New York, NY: 

Springer Science +Business Media LLC.2005. 

[23] Wan-Lu Jiang, Shu-Qing Zhang, Yi-Qun Wang, Chaos and 

Wavelet Based Fault Information Diagnosis. Bei Jing: 

China Machine Prees.2005. 

[24] Jun Yang, Intelligent Fault Diagnosis Technology for 

Equipments. Bei Jing: National Defense Industry Press. 

2004. 

Yun-Jie Xu was born in Neimenggu, 

China, in 1976. He received the B.S. 

degree in fluid power transmission and 

control from Dongbei University of 

Mechanical Engineering, Shenyang, 

China, in 1998, and the M.S. degree in 

Mechanical Design and Theory from 

Zhejiang University of Mechanical and 

Energy Engineering, Hangzhou, China, in 

2004. He is currently pursuing the Ph.D. 

degree in forest engineering, University Of Beijing Forestry, 

Beijing, China, in 2009. From April 2004 to Dec. 2010, He is 

serves an lecturer of the School of Engineering, Zhejiang 

Agricultural & Forestry University. His research interests 

include system fault diagnosis and signal propagation in forest. 

Shu-Dong Xiu received B.S. and M.S. degrees in harbin 

institute of technology, Heilongjiang, China, in 1994, and 1988, 

respectively. From April 2004 to Dec. 2010, he was a faculty 

with the School of Engineering, Zhejiang Agricultural & 

Forestry University and was promoted to be an professor in 

2009. His current research interests include forestry machinery 

and woodworking equipment. 

Quan-Sheng Men received B.S. degrees in Zhejiang University 

of Technology, Zhejiang, China, in 1989, respectively. From 

April 2001 to Dec. 2010, he was a faculty with the School of 

Engineering, Zhejiang Agricultural & Forestry University and 

was promoted to be an senior technician in 2010. His current 

research interests include forestry machinery and woodworking 

equipment. 

Liang Fang received M.S. degrees in Jiangsu University, 

Jiangsu, China, in 2007. From September 2007 to Dec. 2010, he 

was a faculty with the School of technology, Zhejiang 

Agricultural & Forestry University and was promoted to be an 

Laboratory Technician in 2007. His current research interests 

include mechatronic control, signal detection and processing.



Chen Yongqiang 

School of Computer Science, Wuhan Textile University, Wuhan 430073, China 

Email:chenyqwh@gmail.com 

Zhang Yanqing 

Department of Computer Science, Georgia State University, Atlanta 30303, USA 

Email: yzhang@zmail.cs.gsu.edu 

Hu Hanping 

Institute for Pattern Recognition and Artificial Intelligence, Huazhong University of Science and Technology, Wuhan 

430074, China 

Email: hphu@mail.hust.edu.cn 

Ling Hefei 

College of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan 430074, 

China 

Email: lhf@mail.hust.edu.cn 

Abstract—An effective and integrated image watermarking 

scheme mainly includes watermark generation, watermark 

embedding, watermark identification, and watermark 

attack. In this paper, a novel discrete wavelet transform 

domain image watermark scheme is proposed to meet the 

watermarking properties: security, imperceptibility and 

robustness. Here watermark comes from a meaningful 

binary image encrypted by two-dimensional chaotic stream 

encryption, which has more security. In the procedure of 

watermark embedding, the watermark is embedded into 

host image through selecting and modifying the wavelet 

coefficients using genetic algorithms with a simple fitness 

function to improve the imperceptibility of watermarked 

image. In order to identify the owner of extracted 

watermark, synergetic neural networks are used in the 

watermarking identification to overcome the limitation of 

correlation analysis or the human sense organ after some 

attacks. The results of our scheme realization and robust 

experiments show that this scheme has preferable 

performance. 

Index Terms—image watermark, genetic algorithm, 

synergetic neural networks, discrete wavelet transform 


Digital watermark is a kind of technology that embeds 

copyright information into multimedia data[1]. Unlike 

encryption, which is useful for transmission but does not 

provide a way to examine the original data in its 

protected form, the watermark remains in the content in 

its original form and does not prevent a user from 

Manuscript received November 15, 2010; revised January 21, 2011; 


Corresponding author: Chen Yongqiang 

Project number: National 863 Hi-Tech Grant 2009AA01Z411 


doi:10.4304/jsw.6.5.849-856 

listening to, viewing, examining, or manipulating the 

content. Digital watermarking technology opens a new 

door to authors, producers, publishers, and service 

providers for protecting their rights and interests in 

multimedia documents. 

An effective image watermarking scheme mainly 

includes watermark generation, watermark embedding, 

watermark identification, and watermark attack. 

Watermark generation refers to what content and form of 

data a watermarking scheme adopts as watermark. The 

data may be original or encrypted from copyright 

information of number, letter, image, and so on. Some of 

copyright information is meaningful or meaningless. 

Meaningful information could be easily authenticated and 

usually needs to be encrypted in practice to strengthen 

watermarking security [2]. 

Watermark embedding is the most important part in a 

watermarking scheme and must meet the two most 

fundamental requirements under the condition of fixed 

watermark size, imperceptibility and robustness. The two 

requirements are in conflict with each other and need to 

reach a trade-off. Watermark embedding can be done in 

either spatial domain or frequency domain. The spatial 

domain watermark embedding manipulates host image 

pixels, especially on least significant bits that have less 

perceptual effect on the image[3]. Although the spatial 

domain watermark embedding is simple and easy to 

implement, it is less robust than frequency domain 

watermark embedding to various attacks and noise, which 

is made on the frequency coefficients of the host image. 

The existing frequency transformation methods for 

watermark embedding include discrete Fourier transform 

(DFT) [4], discrete cosine transform (DCT)[5], and 

discrete wavelet transform (DWT)[6]. Considering 

watermarking imperceptibility, we need to select an


appropriate embedding algorithm to embed the 

watermarking bits into certain frequency coefficients so 

that the quality of watermarked image does not obviously 

decline compared by original host image. But many 

traditional embedding algorithms in the literature can not 

optimize the embedding process by experiential method. 

In recent years, watermarking techniques have been 

improved using optimization algorithms such as genetic 

algorithm (GA) which is a popular evolutionary 

optimization technique invented by Holland [7]. In the 

field of watermarking, GA is mainly used in the 

embedding procedure to search for locations to embed the 

watermark [8-11]. 

The act of watermark detection can be named as 

watermark verification that a watermarking receiver must 

do a yes or no judgment whether a watermark does exist 

in the received image. In general, the normalized 

cross-correlation (NC) value between the original and 

extracted watermark is used in watermark detection [11]. 

Defined a threshold T, a yes judgment can be given if NC 

≥T, or a contrary result will be gotten. After getting a yes 

judgment in watermark detection, especially to the 

meaningful watermark, people may do more things to 

judge the owner of extracted watermark because of 

possibility of spurious watermark. Based on the 

watermark detection, watermark identification is to 

farther judge what degree extracted watermark is similar 

to original watermark and whose extracted watermark 

belongs to. Although bit correct rate (BCR), NC and 

human eyes can be used in watermark identification, they 

all depend upon experimental results and human 

experiences. For distinguishing extracted watermark 

more clearly, meaningful watermark may be recovered 

partly or entirely from watermarked image by neural 

networks [12-14]. The introduction of neural networks 

helps to pave the way for the further development of 

watermark identification techniques. 

The goal of watermark attack is to test the robustness 

of a watermarking system. To simulate the 

communication conditions and deliberate or unintentional 

processing, some attacks, including adding noise, 

filtering, compression and geometrical distortion, need to 

be used in the watermarked image. For copyright 

protection, we use robust watermark in the condition that 

the watermark can partially be recognized and the 

copyright can be preserved after attacked by some means. 

But based on the applied purpose of robust watermark, a 

watermarking scheme need not withstand all kinds of 

attack. 

This paper presents a novel DWT domain gray image 

watermarking scheme. The watermarking data comes 

from a meaningful binary image encrypted by 

two-dimensional chaotic stream encryption. In the 

procedure of watermark embedding, GA is used to select 

the most fit wavelet coefficients to embed watermarking 

bits into the host gray image. After some kinds of attack, 

the extracted watermark can be identified expediently 

through the synergetic neural networks(SNN). The 

experimental results have shown that this scheme has 


preferable performance of security, imperceptibility and 

robustness. 

II. WATERMARK GENERATION 

The paper [2] employed the two-dimensional chaotic 

Logistic map to encrypt the meaningful gray image and 

gave the digital image stream encryption algorithm in 

detail. Here, we simplify the algorithm to encrypt a 

binary image. 

The two-dimensional Logistic map system with simple 

coupled term is defined in (1). 

⎧ x = 4 µ x (1 − x ) + γ y 

⎨ 

⎩y 

= 4 µ y (1 − y ) + γ x 

n+ 1 1 n n n 

n+ 1 2 n n n 

The dynamical behavior of this map system is 

controlled by control parameters of µ 1, 

µ 2 andγ . When 

µ 1 = µ 2 = µ ≥ 0.89 and γ = 0.1, 

the system is chaotic 

and can be used in digital image encryption. 

Let B = [ b s, t] M× M represent a meaningful binary 

image with size M × M , 0≤ s ≤ M − 1 , 0≤t ≤ M − 1, 

bst , ∈ {0,1} . x P and y P are values obtained after the 

map system is iterated P times. Using iterative values 

x i and y i , encryption algorithm is described as follows. 

1)Transform the decimal fraction of x i into binary 

sequence and choose the first M bits to be represented 

as xi,0 xi,1 � xi, M−1. 

Like x i , the decimal fraction of y i is 

represented as yi,0 yi,1 � yi, M−1. 

2)According to the row order s = 0,1, 2, � , M −1and 

i = P+ s, 

do the XOR operation cs, t = bs, t ⊕ xi, 

j, 

j = t . 

3)According to the column order t = 0,1, 2, � , M −1 

and j t 

w = c ⊕ y , 

= , do the XOR operation s, t s, t i, j 

i = P+ s. 

The watermark W = [ w s, t] M× M can be gotten by 

completing above three steps. 

Let � � W = [ w st , ] M × M represent the extracted 

watermark from watermarked image, the decryption 

procedure is described as follows. 

1)The decimal fraction of x i and y i are denoted 

by xi,0 xi,1 � xi, M−1and 

yi,0 yi,1 � yi, M−1respectively. 

2)According to the column order, do XOR 

operation c� = w�⊕ y . 

st , st , i, j 

(1) 

3)According to the row order, do XOR 

b� = c� ⊕x 

and get decrypted image 

operation st , st , i, j 

B� = [ b � 

st , ] M × M . 

From the above encryption and decryption algorithms, 

it can be concluded that if the received image has not 

been processed, the equations � W = W and � B = B are 

satisfied after the decryption.


III. WATERMARK EMBEDDING 

Wavelet transform is a time-frequency analyzing 

method to localize spacial and frequency domain. Using 

wavelet transform to code and compress image can 

acquire good effect having high compress ratio and no 

avail of block and midge noise. In our scheme, the 

encrypted watermark is embedded into host image 

through selecting and modifying the wavelet coefficients 

using GA. 

A. Genetic Algorithm 

GA is a kind of best searching algorithm that simulates 

biological evolution to produce a similar optimal solution 

and widely used in various fields such as pattern 

recognition, decision support and the nearest optimization 

problem. In organism evolution, organisms with defective 

genes are weeded out so that a species of organisms 

preserves its beneficial genes for its descendants. 

Generally, better chromosomes will be produced for 

propagation after crossover or mutation. 

The GA can be briefly depicted as follows. 

1) Code. In GA-based optimization, any possible 

solution in problem field is represented as an individual 

in colony and encoded by a finite-length binary string, 

called the chromosome. The elements in the binary string, 

or the genes, are adjusted to minimize or maximize the 

fitness value. 

2) Original colony. Some individuals or chromosomes 

are selected in random to form original colony as the first 

generation that can reproduce new generation. 

3) Fitness evaluation. The fitness function is defined 

by algorithm designers, with the goal of optimizing the 

outcome for the specific application. For every generation, 

a pre-determined number of chromosomes will 

correspondingly produce fitness values. The fitness 

values decide the probability of the chromosomes' 

survival or removal during the competition. 

Chromosomes with higher fitness values have higher 

probability to contribute more offspring in the next 

generation. 

4) Genetic operation. Three GA operators, selection, 

crossover and mutation, the core components for GA, are 

applied to the chromosomes repeatedly. 

Selection: A large portion of the chromosomes with 

low fitness values is discarded through this natural 

selection step. The selection rate P s defines the portion 

of chromosomes with high fitness values that can be 

survived into the next generation. 

Crossover: Pairs of chromosomes among the survived 

chromosomes are chosen from the current generation to 

produce two new off-springs. A crossover point is 

selected, and the fractions of each chromosome after the 

crossover point are exchanged, and two new 

chromosomes are produced. 

Mutation: Mutation is the occasional random 

alternation of the value in some positions of 

chromosomes. It introduces traits not in the original 

individuals and keeps GA from converging too fast. Most 

mutations deteriorate the individual fitness values. 

However, the occasional improvement of the fitness adds 


diversity and strengthens the individual. Generally 

speaking, the probability P m for mutation is supposed 

to be low. 

These operators are used repeatedly to obtain 

successive generations of chromosomes. Within a 

generation, only the chromosomes with the higher fitness 

values can survive. They will be passed as parent 

chromosomes to the next generation. 

5) Terminating rule. The terminating rule can be 

selected as one of conditions that the generation number 

is more than a defined terminating number or the fitness 

values of chromosomes is unchanged after some 

generations. 

After a number of generations, the chromosomes are 

optimized. We can obtain the near-optimal solution of the 

modeled problem. 

B. Embedding algorithm using GA 

Embedding algorithms using GA in DWT domain 

image watermark have been researched in some papers. 

GA is applied to improve the quality of the watermarked 

image and the robustness of the watermark. But, the main 

drawback in these algorithms lies in the fitness function 

which is developed based on the combination of 

imperceptibility and robustness. The objective functions 

used to measure these properties vary significantly by 

numerical values. The varieties of attacks make difficult 

for equal contribution of imperceptibility and robustness 

in fitness function even if robustness measure is scaled by 

a factor. In our embedding algorithm based on GA, a 

simple fitness function may be developed only 

considering imperceptibility rather than robustness dealt 

with in the procedure of watermark identification. 

Let I = [ I( i, j)] 

(1 ≤ i, j ≤ N ) represent the host 

gray image with size N× N and I� be the optimal 

watermarked image. The embedding algorithm is outlined 

below in detail. 

1) Divide the host image into ordinal un-overlapped 

N N 

2M2M × 

× sub-images. There are 2M2M sub-images represented by 

N 

I i, j ( 1 ≤i, j ≤ 2M 

). 

2) Perform discrete wavelet transform independently to 

every sub-image I i, j and get the sub-image low subband 

LL i, j= [ LLi, 

j, s, t] 

, high subband HH i, j, 

two middle 

subbands HL i, j= [ HLi, 

j, s, t] 

and LH i, j= [ LHi, 

j, s, t] 

. 

Because the texture and edge information are mainly 

represented in the biggish wavelet coefficients of HH, HL 

and LH subbands, the watermark will be embedded into 

the low or middle subband. 

N N 

3) There are 2M × 2M × 3 subband positions in the 

host image so the chromosome is encoded to 

N N log 2( 2M × 2M 

× 3) bits . Each chromosome represents 

a position to embed the watermark. 

4) For one chromosome, modify the corresponding 

coefficients as (2) and (3).


' 

LL 

CF 

i, j, s, t 

' 

i, j, s, t 

M M 

⎧ α1 

⎪LLi, 

jst , , + 2 ∑∑LLi, 

jst , , if wst 

, = 1 

⎪ M s= 1 t= 

1 

= ⎨ M M 

⎪ α1 

LLi, jst , , − LL 2 

i, jst , , if wst 

, = 0 

⎪ ∑∑ 

⎩ M s= 1 t= 

1 

(2) 

M M 

⎧ α2 

⎪CFi, 

jst , , + 2 ∑∑ CFi, jst , , if wst 

, = 1 

⎪ M s= 1 t= 

1 

= ⎨ M M 

⎪ α2 

CFi, jst , , − CF 2 

i, jst , , if wst 

, = 0 

⎪⎩ 

∑∑ M s= 1 t= 

1 

(3) 

The parameter α 1 and α 2 are the embedding 

intensities and CF maybe one of HL and LH . Do 

inverse discrete wavelet transform after modifying the 

' 

' 

wavelet coefficients and get I i, j. 

All of I i, j are united 

' ' 

to a watermarked image I = [ I ( i, j)] 

(1 ≤i, j ≤ N ). 

5) Define the fitness function using peak 

signal-to-noise ratio (PSNR) between I = [ I( i, j)] 

and 

' ' 

I = [ I ( i, j)] 

. 

2 2 

N × max( I ( i, j)) 

N N 

' 2 

PSNR = 10× log 10( 

) 

( Iij ( , ) − I( i, j)) 

∑∑ 

i= 1 j= 

1 

6) Create some random chromosomes into an original 

colony and give the values of s P and P m . Evaluate the 

fitness values of chromosomes and do the genetic 

operation until the process of GA stops and the optimal 

watermarked image I� is gotten. 

The final chromosome of GA and parameters 

( µ 1, µ 2, γ, x0, y0, P) 

of the two-dimensional chaotic 

Logistic map system can be looked upon the key of this 

watermarking scheme used in watermark extraction. 

IV. WATERMARK IDENTIFICATION 

To an encrypted meaningful watermark extracted from 

watermarked image, people maybe not distinguish its 

decrypted form through technical indexes, such as BCR, 

NC, and eyes, because of some interferential causations 

to watermarked image in the communication and usage. 

The SNN can effectively identify the extracted watermark 

in our former research[15] so that it is used in this scheme 

too. 

A. Synergetic Neural Networks 

The SNN model is a top-down network constructed by 

synergetic different from traditional network constructed 

by the method researched in single neuron’s characteristic, 

configuration and connection[16]. 


(4) 

Dynamical system can be described by state vector in 

Synergetic. Let a state vector be q= ( q1, q2, � , q ' ) . 

M 

A synergetic associative pattern recognition system can 

be described by dynamical evolutionary process, in which 

the system evolves by neural network learning to fill 

incomplete data set and form pattern. Furthermore, let 

' 

prototype pattern number be M and prototype pattern 

' 

' ' 

vector’s dimension be N , where satisfies M ≤ N . 

A dynamical equation can be described by (5). 

. 

∑ ∑ (5) 

q= λ v ( v q) −B ( v q) ( v qv ) −Cq 

( qq ) 

+ + 2 + + 

k 

k k k 

' 

k≠k ' 

k k k 

where q as recognizable pattern vector with original 

input value q0 = q(0) 

can be decomposed into 

prototype v k and remnant vector w , having 

' 

M 

∑ ξk 

k = 1 

k and 

+ 

w = 0 

q= v + w 

vk . Attention parameter 

λ k is positive. B and C are appointed coefficients 

and must be more than zero. Prototype pattern vector k v 

is expressed as vk = ( vk1, vk2, � v ' ) ′ and 

kN 

+ 

v k is an 

adjoint vector of v k , which satisfies an orthogonal 

⎧ ′ 

+ 

1, 

k = k 

condition vk vk 

= δ kk′ 

= ⎨ . All v k will be 

⎩0, 

k ≠ k′ 

normalized as 

' 

N 

∑ vkl 

= 0 and centered as 

l = 1 

' 

N 

2 1/2 

k = ( ) 1 

2 ∑ kl = 

l= 

1 

v v 

. 

Order parameter ξ k is defined as k k vq ξ 

+ 

= . The 

dynamical equation can be rewritten by order parameter. 

k = k k − B∑ 2 

k′ ' 

M 

k − C ( ∑ 

2 

k′ ) k 

k′ ≠ k k′ 

= 1 

ξ� λξ ξ ξ ξ ξ (6) 

D= ( B+ C) ∑ ξ , (6) is simplified to (7). 

k 

2 

Used ' 

' 

k 

� (7) 

ξ ξ λ ξ 

2 

k = k( − D+ B k ) 

So the SNN model is constructed with three layers. 

The top layer is the input layer. All order parameter 

neurons form the middle layer. The down layer is the 

output layer. 

B. Watermark Extracting 

The watermark extracting is the contrary producer of 

watermark embedding. The DWT transforms of the 

received watermarked image I � and host image I 

could be done according to the rule of watermark 

embedding and the final chromosome of GA need be


gotten from the key. The extracted encrypted watermark 

�W can be extracted by (8). 

' 

⎪ 

⎧CF − CF 

, , , , , , 0 , 1 

i j u v i j u v> ⇒ ws 

t= 

⎨ ' 

⎪⎩ 

CF − CF 

, , , , , , 0 , 0 

i j u v i j u v< ⇒ ws 

t= 

The CF presents one of LL , HL and LH . 

Then using the parameters ( µ 1, µ 2, γ , x0, y0, P) 

of 

the two-dimensional chaotic Logistic map system from 

the key and decryption algorithm, we can decrypt the 

extracted encrypted watermark � W and get the 

decrypted watermark � B . 

C. Decrypted Watermark Identification using SNN 

The decrypted watermark identification may be taken 

for the process that a special existing watermark is 

formed and recognized in a mass of watermark patterns 

so that the pattern recognition method may be used in the 

watermark algorithm. 

The original meaningful watermark image and some 

binary images having the same size and similar content as 

the watermark image are select to makeup a prototype 

' 

pattern set including M components. All of 

two-dimensional binary images are transferred to 

one-dimensional sequences and the vectors 

[ 1, 2, 

, ' ] T 

' ' 2 

vk = vk vk � v ( k = 1, 2, � , M , N = M ) 

kN 

can be gotten. A prototype pattern set may be composed 

' 

' ' 

of these M pattern vectors only if M ≤ N . 

According to the SNN model, the learning algorithm of 

networks is the training process that adjoins vectors are 

calculated through prototype pattern vector. 

1) Compute prototype pattern vector v k satisfying 

normal and center condition. 

2) Compute the according adjoint vector 

+ 

v k of 

prototype pattern vector v k . 

Used the SNN method, the watermark detection and 

identification may be accomplished at one time through 

recognized pattern. Now the constants of synergetic 

dynamic equation are given to B = C = 1 and λ k = 1 . 

Thus the recognition process of SNN is showed as 

following. 

1) Compute the test pattern vector q(0) = { qi} 

, 

' 

i = 1, 2, � , N , which satisfying normal and center 

condition, too. 

2) Achieve the according order parameter ξ k ( 0) 


prototype patterns. According to synergetic slaving 

principle, the pattern having the most value of order 

parameters will prevail in the synergetic evolution, and 

thus the watermark embedded in the host image carrier 

may be detected firstly. 

3) Evolve by (7) until the neural networks becomes 

stabilized to specific prototype pattern. 


(8) 

The decrypted watermark sequence represented by the 

special prototype pattern can revert to the original 

meaningful watermark so that the decrypted watermark is 

identified. 

V. SCHEME REALIZATION AND ROBUST EXPERIMENTS 

In order to test the validity of the proposed scheme, we 

select Peppers image as host gray image with size 

N = 512 and a binary face image as original 

meaningful image with size M = 64 treated from The 

Database of Faces [17]. In the mean time, other four 

binary face images are selected to compose a prototype 

pattern vector set including five one-dimensional vector 

components with the original watermark. Therefore there 

' 

are M = 5 components in the prototype pattern set. 

The images used in our scheme are all listed in Fig.1. 

(a) (b) 

(c) (d) 

(e) 

Figure 1. Images in the scheme 

After selected control parameters µ 1 = µ 2 = µ = 0.9 , 

γ = 0.1 and initial values x 0 = 0.1 , y 0 = 0.11 , 

P = 500 , the binary face image is encrypted to 

watermark used the encryption algorithm. The face image 

and watermark are showed in Fig.1(b) and Fig.1(d) 

respectively. 

The embedding intensities α 1 = 0.02 and 

α 2 = 1.5 are set firstly. Used GADS Toolbox in the 

Matlab7.0, the watermark is embedded into the peppers


image. The host Peppers image and watermarked Peppers 

image are showed in Fig.1(a) and Fig.1(c). In the GA, the 

best position LL 2,1 and the most value of PSNR 

53.1417 are gotten through about 15 generations. 

To image watermark, the possible attacks in the usage 

are adding noise, range change, linear filter, and lossy 

compression, which used to review the image 

watermark’s robustness. In the Matlab7.0 software 

environment, we accomplished attack experiments to the 

watermarked Peppers image: adding gaussian noise with 

zero mean and 0.0005 variance, adding salt-pepper noise 

which zero mean and 0.0005 variance, strengthening 

contrast from [0.1 0.9] to [0 1], weakening contrast from 

[0 1] to [0.1 0.9], doing 3×3 median filter and wiener 

filter, and JPEG compression with quality 50%. After 

these attacks, the extracted watermark can be identified 

rightly by SNN in the 20-50 steps of evolution and some 

robust experimental results are showed in Figs. 2-8. In 

these figures, (a) are extracted watermarks, (b) are 

encrypted images of extracted watermarks, and (c) are 

evolution lines of SNN. 

(a) (b) (c) 

Figure 2. Evolution of robust experiments for Gaussian noise 

(a) (b) (c) 

Figure 3. Evolution of robust experiments for salt-pepper noise 

(a) (b) (c) 

Figure 4. Evolution of robust experiments for strengthen 

(a) (b) (c) 

Figure 5. Evolution of robust experiments for weaken 


(a) (b) (c) 

Figure 6. Evolution of robust experiments for median filter 

(a) (b) (c) 

Figure 7. Evolution of robust experiments for wiener filter 

(a) (b) (c) 

Figure 8. Evolution of robust experiments for JPEG compression 

From the Figs. 2-8, we can see that some extracted 

watermarks in Fig.3, Fig.5 and Fig.8 can be directly 

identified by our eyes or correlation analysis, but others 

in Fig.2, Fig.4 and Fig.6 can’t be. Used the SNN, the 

SNN evolution results of watermarks tends to 1 and the 

watermarks embedded in the host image could be easily 

identified. 


An effective digital watermark scheme must meet three 

main properties: security, imperceptibility and robustness. 

In our scheme the two-dimensional chaotic stream 

encryption is used to encrypt a meaningful image to 

generate a watermark. The watermark encrypted from a 

meaningful image can not be fabricated so that there is 

very strong watermarking security. GA is adopted to find 

the best position to embed watermark to wavelet 

coefficients of host image in order to guarantee the 

quality of watermarked image. This kind of evolutionary 

optimization technique can improve watermarking 

imperceptibility and robustness. In the procedure of 

watermarking identification, SNN has the ability to 

recognize the original watermark quickly and accurately 

after attacks. 

In our scheme realization and robust experiments, the 

results prove the feasibility and validity of our proposed 

scheme. But in the watermarking embedding, one of the 

limitations is that the embedding intensity is given by 

experience in this scheme. In the next step, we will use 

GA to find the best value of embedding intensity to 

improve embedding performance. It is better that GA


could be used to select the embedding position and 

intensity synchronously. 


This work was supported in part by grants from the 

National 863 Hi-Tech Grant 2009AA01Z411 and the 

2008 importance project of Wuhan Textile University. 

REFERENCES 

[1] I. J. Cox, M. L. Miller, and J. A, “Bloom, Digital 

watermarking,” San Francisco: Morgan Kaufmann 

Publishers, 2002 

[2] H. P. Hu and Y. Q. Chen, “Image watermarking technique 

based on two-dimensional chaotic stream encryption,” The 

First International Workshop on Security in Ubiquitous 

Computing Systems, LNCS 3823, pp.817-824,2005 

[3] I. Nasir, Y. Weng, J. M. Jiang, and S. Ipson, “Multiple 

spatial watermarking technique in color images,” Signal, 

Image and Video Processing, vol.4, pp.145-154, 2009 

[4] M. David, S. R. Jordi, and F. Mehdi, “Efficient self- 

synchronised blind audio watermarking system based on 

time domain and FFT amplitude modification,” Signal 

Processing. Vol.90, pp.3078-3092, 2010 

[5] W. Liu and C. H. Zhao, “Digital watermarking for volume 

data based on 3D-DWT and 3D-DCT,” The 2nd 

International Conference on Interaction Sciences: 

Information Technology, Culture and Human, pp.352- 

357, 2009 

[6] B. Deepayan and A. Charith, “Video watermarking using 

motion compensated 2D+t+2D filtering,” The 12th ACM 

workshop on Multimedia and security, pp.127-136, 2010 

[7] J. Holland, “Adaptation in natural and artificial systems,” 

University of Michigan Press, Ann Arbor, MI ,1975 

[8] P. Kumsawat, K. Attakitmongcol, and A. Srikaew, “A new 

approach for optimization in image watermarking by using 

genetic algorithms,” IEEE Transactions on Signal 

Processing, vol.53, pp.4707-4719, 2005 

[9] Y. T. Wu and F. Y. Shih, “Genetic algorithm based 

methodology for breaking the steganalytic systems,” IEEE 

Transactions on Systems, Man, and Cybernetics, vol.36, 

pp.24-31,2006 

[10] H. C. Huang, J. S. Pan, Y. H. Huang, F. H. Wang, and K. 

C. Huang, “Progressive watermarking techniques using 

genetic algorithms,” Circuits Systems Signal Processing, 

vol.26, pp.671-687, 2007 

[11] S. C. Chu, H. C. Huang, Y. Shi, S. Y. Wu, and C. S. Shieh, 

“Genetic watermarking for zerotree-based applications,” 

Circuits Systems Signal Process, vol.27, pp.171-182, 2008 

[12] Z. F. Wang, N. C. Wang, and B. C. Shi, “A novel blind 

watermarking scheme based on neural network in wavelet 

domain,” The 6th World Congress on Intelligent Control 

and Automation, vol. 1, pp.3024-3027, 2006 

[13] S. Huang, W. Zhang, W. Feng, and H.Q. Yang, “Blind 

watermarking scheme based on neural network,” The 7th 

World Congress on Intelligent Control and Automation, 

vol.1, pp.5985- 5989, 2008 

[14] C. Y. Chang and S. J. Su, “The application of a full counter 

propagation neural network to image watermarking,” 

Proceedings of IEEE on Networking, Sensing and Control, 

pp.993-998, 2005 

[15] Y. Q. Chen, H. P. Hu, and X. T. Li, “Extracted watermark 

identification using synergetic pattern recognition,” The 

4th International Symposium on Multispectral Image 

Processing and Pattern Recognition, Vol.6043, pp.256-264, 

2005 


[16] H. Hanken, “Synergetic computers and cognition-a top- 

down approach to neural nets,” Berlin: Springer-Verlag, 

1991 

[17] AT&T Laboratories Cambridge. The Database of Faces. 

http://www.cl.cam.ac.uk/research/dtg/attarchive/facedataba 

se.html. 2010-3-24 

Chen Yongqiang Wuhan China, July 

1967. He received the B.S. degree in 

Fluid Drive and Control, M.S. degree in 

Mechanical Design and Theory and Ph.D. 

degree in Pattern Recognition and 

Intelligence System from Huazhong 

University of Science and Technology, 

China, in 1989, 2001 and 2005, 

respectively. 

Dr. Chen is currently an Associate Professor of School of 

Computer Science at Wuhan Textile University, Wuhan, China. 

His research areas include computer graphics, digital image 

processing, multimedia technology, artificial intelligence and 

digital watermarking. 

Dr. Chen is a member of ACM, China Computer Federation 

and China Society of Image and Graphics. He has been awarded 

the third prize of “Prize of Scientific Progress” of Hubei 

Province and the third prize of “Prize of Scientific Progress” of 

Wuhan City. He published one book and about 40 journal or 

conference papers. 

Zhang Yanqing He received the B.S. 

and M.S. degrees in computer science 

from Tianjin University, China, in 1983 

and 1986, respectively, and the Ph.D. 

degree in computer science from the 

University of South Florida, Tampa, in 

1997. 

Dr. Zhang is currently an Associated 

Professor of the Computer Science 

Department at Georgia State University, Atlanta, USA. His 

research interests include computational intelligence, data 

mining, bioinformatics, web intelligence, and intelligent 

parallel/distributed computing. 

Dr. Zhang is a member of the Bioinformatics and 

Bioengineering Technical Committee, and the Data Mining 

Technical Committee of the IEEE Computational Intelligence 

Society. He has co-authored two books, co-edited two books 

and four conference proceedings. He published 15 book 

chapters, 65 journal papers and over 130 conference/workshop 

papers. 

Hu Hanping He received the M.S. and Ph.D. degrees in 

Pattern Recognition and Intelligence System from Huazhong 

University of Science and Technology, China, in 1995 and 1998, 

respectively. 

Dr. Hu is currently a Professor of the Institute for Pattern 

Recognition and Artificial Intelligence at Huazhong University 

of Science and Technology, Wuhan, China. His research areas 

include information security, computer networks, digital image 

processing, artificial intelligence and information hiding. 

Dr. Hu is a committeeman of the intelligent Automation 

Committee in the Chinese Association of Automation. He 

published about 70 journal or conference papers.


Ling Hefei He received the B.S. degree in Energy and Power 

Engineering and Ph.D. degree in Computer Science and 

Technology from Huazhong University of Science and 

Technology, China, in 1999 and 2005, respectively. 

Dr. Ling is currently an Associated Professor of the School 

of Computer Science and Technology at Huazhong University 

of Science and Technology, Wuhan, China. His research areas 

include multimedia security, digital watermarking, copy 

detection, digital media forensics and intelligent video 

processing. 

Dr. Ling is a senior member of ACM, IEEE and China 

Computer Federation. He published about 40 journal or 

conference papers. 



An Efficient Method for Improving Query 

Efficiency in Data Warehouse 

Zhiwei Ni 1,2 

1)School of Management, Hefei University of Technology, Hefei , Anhui, China 

2)Key Lab. of Process Optimization and Intelligent Decision-making, Ministry of Education, Hefei , Anhui, China 

zhwnelson@163.com 

Junfeng Guo 1,2 , Li Wang 1,2 and Yazhuo Gao 1,2 

1)School of Management, Hefei University of Technology, Hefei , Anhui, China 

2)Key Lab. of Process Optimization and Intelligent Decision-making, Ministry of Education, Hefei , Anhui, China 

alloy1129@yahoo.com.cn, wl820609@163.com, yazhuogao@163.com 

Abstract—There are lots of performance bottlenecks for 

real-time queries in mass data. Many methods can only 

improve the efficiency for frequently used queries, but it is 

not advisable to neglect the non-frequently used queries. 

This paper proposes a new integrated index model called 

BBI and illustrates the application of this model. Based on 

the feature of data warehouse and OLAP queries, this index 

model is built with inverted index, aggregation table, bitmap 

index and b-tree. It greatly promotes not only the efficiency 

of frequently used queries, but also the performance of 

other queries. The analytical and experimental results 

demonstrate the utility of BBI. 

Index Terms—Aggregation Table, Inverted Index, Bitmap 

Index, B-Tree Index 


Data warehouse (DW) is defined as a subject-oriented, 

integrated, steady and time varying data set which 

supports enterprises or organizations to make decisions. 

As the decision maker needs to query several values from 

one subject for real-time analysis processing, the 

multidimensional model of DW is usually implemented as 

star schemes to meet the requirements. This kind of 

hierarchical model is highly unnormalized and queryoriented. 

There are two kinds of table in star schemes. 

One is fact table which contains basic quantitative 

measurements of a business subject, the other is 

dimension table that describes the facts. If there are more 

than one fact table in a DW, it can be called galaxy model 

which is actually constituted of several star schemes. 

Complex queries are always requested in DW. When 

users need to process multi-dimensional analysis, multitable 

joins may be involved. Although data can be stored 

in a multi-dimensional database, DW usually stores data 

in the form of relational database. As the number of 

dimension and the overall size of data sets increase, the 

size of DW often grows to gigabytes or terabytes. When 

the complex queries are implemented on mass multidimensional 

data , the query efficiency is far beyond 

satisfaction. Fact tables in data warehouses which store 

business measures usually have millions of records or 


doi:10.4304/jsw.6.5.857-865 

more. Such tables usually have more than 10 attribute 

dimensions. For example, to select records of which time 

between 01/01/2008 and 09/10/2009, it needs to join the 

tables and compare the time value of millions of each 

record with the requested value. So it is necessary to 

retrieve the data more efficiently. Some methods and 

technologies were proposed to improve queries efficiency, 

such as materialized view[1], feature selection[5], index 

technology[3], etc. 

Materialized view is a kind of pre-computed structure, 

it materializes the calculated results ahead of using. The 

pre-computed values are often mean, sum, average, etc. 

Queries on materialized views are fast responded, because 

no join needs to be made on successive requests, and the 

records in views are less than the original tables. 

Materialized views can be applied to OLAP, but due to the 

limit of storage space, it is infeasible to store results of all 

queries. Some heuristic algorithms have been used to find 

an approximate optimal solution. For example, greedy and 

genetic algorithms that based on requirement and 

probability are applied to generate views. But once queries 

are made on the records which are not materialized, the 

efficiency can not be improved, and it is unacceptable for 

any delay when users need the results urgently. So there is 

limitation of materialized method. 

Feature selection is a procedure to select a subset 

from the original feature set by eliminating redundancy 

and less informative features so that the subset contains 

only the most discriminative features [4]. Applied to 

dimension reduction, a set of attributes that best represents 

the overall data set is found out by feature selection. But 

feature selection has the same problem with materialized 

view that when the queries involve the dimensions which 

are not selected, the efficiency of this method deceases. 

Index technology greatly cuts down the load of I/O, 

which is highly effective in real application. Compared 

with materialized view, the space cost of index is reduced. 

Many indexes are classified into data-partitioning indexes, 

such as B+ tree and R-tree family and other tree indexes. 

B+ tree indexes are often adopted in databases to retrieve 

rows of a table with specified values involving one or


more columns. Data are placed in some partitions by the 

sequence of key values which also need to be precomputed. 

Using these indexes to answer a query, system 

should find the partitions which contain the related data by 

comparing the key value from the root node to the leaf 

nodes. Those all search paths that may be potentially 

matched must be explored. Data-partitioning indexes are 

effective for single keyword queries. As the dimension 

quantity of both the indexes and the queries increases, the 

efficiency deceases. 

OLAP query includes point query and cube query. It 

takes less time to process point query whose result is a 

single value or one record. Cube query returns a list of 

values or multiple rows of data which are aggregated from 

the data set. It needs to traversal the whole data table to 

get the data set and is relatively long in the duration of 

query implementation. For example, if we need the total 

revenue which is sorted only by time and location, it is a 

cube query that needs to traversal the whole sale fact table 

to aggregate the data. As for OLAP query,the common 

index technology can not meet the requirements of query 

efficiency. Data cube plays an essential role in fast OLAP 

query, but high dimensional data cube requires massive 

memory and disk space, and the current algorithms are 

unable to materialize the full cube under such conditions. 

It is called curse of dimensionality. Besides, it is hard to 

build data cubes based on relational databases, so how to 

promote the query efficiency becomes the key problem. 

As for OLAP application,paper [17] used inverted index 

which is common technology of search engine to build 

shell cube. Though the new cubing approach reduces the 

space cost of data cube and promotes the efficiency of 

queries, many queries require to be computed at run time. 

Introducing inverted-index in search engine 

technology, integrating join-index and materialized view, 

we propose a new index model in this paper based on user 

interest and query statistics, called BBI. The new index 

can not only greatly promote the efficiency of frequently 

used queries, but also improve the performance of other 

queries. With a good performance in storage size, it is 

suitable to be applied in OLAP and other complex queries. 

In Section 2, the relevant knowledge of existing 

approaches and technology are presented; our approach is 

elaborated and the efficiency of the theory is analyzed in 

Section 3. BBI is further discussed in Section 4. Section 5 

describes the performance evaluation of our approach. 

Section 6 is the conclusion. 

II. RELATED WORK 

There are some feasible methods with acceptable 

performance for the queries that require multi-table joins 

in high dimension data warehouses when the queries are 

based on the dimensions that are important or commonly 

used, but not all dimensions can be included with most 

methods while the space expense grows rapidly as the 

dimension grows. The common technologies are view 

materialization, feature selection, index technology, etc., 

they are not independent with each other, and many 

researches concentrate on integrating various technologies 

to improve the efficiency of DW. 


A. Materialized View 

Dynamic materialized view [7] selectively materializes 

only a subset of rows which are the most frequently 

accessed. Compared to conventional materialized view 

which maintains all rows of a view, the set of dynamic 

materialized views can be changed dynamically and the 

storage space is reduced. A method in [8] materializes the 

views in a data warehouse to reduce the query response 

time. Aiming at the insufficient consideration of the 

dynamic update, the method can wash out materialized 

views and add new materialized views in the set of current 

materialized views on the basis of the greedy algorithm. 

Paper [9] proposes a constrained evolutionary algorithm 

for materialized views. Constraints are incorporated into 

the algorithm through a stochastic ranking procedure [18]. 

The basic principle of this technology has been well 

described in [10] and [11]. But [7], [8], [9], [10], [11] all 

avoid the question of generality. 

An algorithm is presented in [1] for building 

materialized sample views for database approximation. 

The core technique is called ACE (Appendability, 

Combinability, and Exponentiality) Tree, improved from 

B+-tree, that is suitable for organizing and indexing a 

sample view, but the algorithm is not useful when 

integrated views are needed. 

B. Feature selection 

Feature selection techniques are targeted at finding a 

set of attributes that best represent the overall data. [2], [4] 

and [19] are traditional techniques of feature selection 

which focus on maximizing data energy or classification 

accuracy for dimension reduction [19]. The algorithm in 

[4] groups the features into different clusters based on 

feature similarity and selects a representative feature from 

each cluster, so that the feature redundancy is reduced. As 

a result, selected features may have no overlap with 

queried attributes. In this case, to neglect any attribute 

may bring troubles when queries are based on the 

attributes that have not been selected. 

C. Index 

Patrick O’Neil and Dallan Quass presented a review of 

indexing technology in Paper [6] which included join 

index, bitmap index and B+-tree. Two indexing structures 

called Bit-Sliced indexes and Projection indexes were 

introduced as well. The core of the method was the 

combination of B+-tree and bitmap index, which was not 

complex but effective. 

At present, B+-tree and bitmap index are two of the 

most important indexes in most database software. In fact, 

the new edition of ORACLE database has involved the 

improved indexing technology. The key indexes of 

ORACLE are shown in Fig. 1 and Fig. 2. Fig. 1 shows the 

application of B-tree index. Based on the rules of B-tree, 

relational table is divided by a numerical attribute, all 

rows of the table are distributed in the leaf nodes of a tree. 

In leaf nodes, the values of indexed attribute are 

considered as labels. Each label corresponds to a value 

called ‘rowid’ which is used to uniquely identify one row 

of data in ORACLE database and can be seen as physical 

addresses. Leaf nodes are connected by indicator so that it


is convenient to use range query. When queries are made 

on keywords, it is rapidly to find the ‘rowid’ in leaf nodes 

by b-tree rule and localize the rows of data in database by 

‘rowid’. B-tree index in ORACLE is suitable for highcardinality 

columns and it is inexpensive to update on 

keys relatively. But it is inefficient for queries using ‘or’ 

predicates. 

Fig. 1. B-tree index in ORACLE 

Fig. 1. shows the structure of bitmap index which is 

built for gender in ORACLE. The word ‘start’ and ’end’ 

represent a segment of storage space. The ‘rowid’ of 

‘8.0.1’ is the beginning address and ’12.0.1’ is end. 

Bitmap shows the distributing of all the gender keywords. 

‘1’ means the keyword appears in the row, ‘0’ means the 

contrary. Bitmap index in ORACLE is suitable for lowcardinality 

columns and it is efficient for queries using OR 

predicates. But it is expensive to update key column. At 

the same time, each keyword needs a bitmap whose length 

is the same as the fact table. When the row number is 

large, it is difficult to use and store bitmap. ORACLE 

often uses the subsection structures to solve this problem 

which also shows in Fig. 2,but as the number of segment 

increases, performance decreases quickly, because it 

needs to checkall the bitmap segments to find answer. 

Fig. 2. Bitmap index in ORACLE 

Several papers proposed designs for index 

recommendations based on optimization rules [14], [15], 

[16].Since the effectiveness of these indexes degrades 

when the query patterns change, Michael Gibas and his 

collaborators [5] introduced a technique to recommend 

indexes based on index types that were frequently used for 

high-dimensional data sets and to dynamically adjust 

indexes as the underlying query workload changes. 

Paper [3] evaluated experimentally the dimension-join 

indexes using the TPC-H benchmark and showed a new 

index structure using bitmap and B+ tree that can 

dramatically improve the performance for some queries. 


Papers [12] proposed Hybrid Index which is suitable for 

given data by building with B+ tree and hash table, but it 

is not universality. Paper [13] introduced a new kind of 

multi-table join index and proved it was more effective 

than only using multi-table join by experiment. But this 

technique also only has the index on the frequently using 

data to advance the query performance; it upgrades 

nothing when queries are made on the data which has not 

been paid attention. 

III. APPROACH 

A. Inverted Index 

Inverted index comes from the functional needs of 

attribute retrieval in practice. Each item in the index table 

has an attribute value and an ID which corresponds to the 

value. Inverted index tables identify ID by attribute, which 

is different from the common tables that identify attribute 

by ID, that is why it is called inverted index. Table 1 is a 

fact table which has three dimension columns and one 

numerical value column. Table 2 is the first order inverted 

index of table 1. The keyword ‘male’ appears in line 1, 2, 

5 in table 1; so there is one line in table 2 which records 

‘male’ and ‘1,2,5’. The rest can be deduced accordingly. It 

is easy to find that the storage space of inverted index 

smaller than that of the original table because it does not 

include the numerical value column. 

Table 1. Fact Table 

TID sex age specialty score 

1 male 20 computer 90 


3 female 19 computer 83 




Table 2. First-order Inverted Index 

Word TID 

male 1,2,5 

female 3,4,6 

19 3,5 

20 1,2,4,6 

computer 1,2,3,4,5,6 

Second-order and high-order inverted index can also 

be built. Table 3 shows the second-order inverted index of 

sex and age. The TID of male and 19 is 5, it tells that male 

and 19 both appear in line 5. The advantage of inverted 

index is fast keyword retrieval. 

Table 3. Second-order Inverted Index 

Word TID 

male ,19 5 

male ,20 1,2 

female,19 3 

female,20 4,6


Table 2 can be translated into Table 4 by using bitmap, 

which can further decrease the storage space and increase 

the query efficiency. 

Table 4. Bitmap-based Inverted Index 

Word TID 

male 110010 

female 001101 

19 001010 

20 110101 

computer 111111 

B. Definitions 

Definition 1 High-frequency Join 

Queries which relate to the same dimensions are 

classified into the same category. If the ratio of the 

number of queries in the same category to the total 

number of all queries is over HFJ, the category is called 

high-frequency Join. HFJ is the threshold of highfrequency 

which is between 0 and 1. The lower the HFJ is, 

the more the high-frequency Join there are, which leads to 

the large space occupation. The setup of HFJ depends on 

the real cases. 

Definition 2 Fact table can be formalized as Table 

(Dimx1, Dimx2,…,Dimxn, M1,…,Mm). Aggregation 

tables are derived from the original fact table, indicating to 

AgTable (Dimx1, Dimx2,…,Dimxi, f(M1),…,f(Mm)). 

Dimx1,Dimx2,…,Dimxi indicates the remaining 

dimensions after the aggregation. f ( ) is the aggregation 

function. There are (n-i) dimensions in AgTable, that is 

less than original fact table. 

C. Approach Overview 

The abstract of our approach is as follows: In 

condition of the limited storage space, inverted index is 

built for fact table in database while aggregation table is 

established for frequently used queries, aggregation table 

is applied for frequently used queries to get quick 

response and inverted index is adopted for other queries to 

reduce response time. So it improves the efficiency of 

both frequently used queries and the other queries. 

This paper establishes aggregation based on the 

frequency of queries from users and adopts bitmap to 

optimize inverted index. A solution is proposed for the 

inefficient performance of bitmap when the number of 

records becomes too large. The approach has the 

advantages as quick response for frequently used queries, 

performance promotion for other queries and reasonable 

space cost. 

D. Application of Our Approach 

The establishing process of our approach can be 

mainly divided into three steps: getting high-frequency 

Joins; building inverted indexes; building aggregation 

tables and join indexes. If better space performance is 

required, then only the aggregation table for those highfrequency 

joins meet the thresholdα will be built (α is 

HFJ which has been introduced before). The function 

ofα is to reduce the inconsequence of building too many 

aggregation tables. For some high-frequency joins, the 


number of involved dimensions is slightly smaller than the 

total number, there is only tiny advancement by applying 

aggregation table for them. So this kind of high-frequency 

joins will be ignored while building aggregation tables. 

Algorithm 1 Creating Index and AgTable 

Input: Fact Table, Query Set 

Output: Inverted Index, AgTable(Dimx1,Dimx2 …Dimxi 

, f(M1),…,f(Mm)) 

begin 

(1)Scan Query Set → Get High Frequency Dim and 

High Frequency Join 

(2)Scan Database → Create First-order Inverted 

Index 

IF space is surplus 

For each High Frequency Dim 

Create High-order Inverted Index(X≥2) 

End 

(3)For each High Frequency Join 

IF i


Theorem 1 The ratio of aggregation table to fact table 

is getting smaller rapidly with the increase of rows of any 

dimension table or the number of dimensions. 

Theorem 2 Fact table and the structure of our 

approach have the same the space complexity. 

Time complexity analysis is as follows: When there is 

no index, Cube queries of OLAP need to join the fact 

table with dimension table to build a temporary view. A 

summary query needs a traversal of the whole table, so the 

time complexity is O (Mn). When our approach is 

adopted, there are two situations: 

(1)If the aggregation table can be used, it costs time to 

search for aggregation and query on it. As there are join 

index in RAM, it can quickly find out the aggregation, so 

the time is mainly spent on the second step. Query by 

using aggregation, the time complexity is O (Mn-i), ‘i’ is 

(1-α )*n which stands for the difference of dimension 

number of fact table and aggregation table. O (Mn-i) is 

much smaller than O (Mn), the more the dimensions in the 

dimension table are and the bigger the fact table is, the 

more efficient our approach is. 

(2) If no aggregation table can be used, then inverted 

index can be adopted by OLAP query. The total time can 

be divided into three parts: the time of reading out the 

result id set from inverted index, the time of getting the 

intersection of result id set, the time of querying on the 

fact table with id. If inverted index is expressed by bitmap, 

time for intersection can be omitted, so the total time 

complexity is O (M*n). Otherwise, the intersection time 

should be considered. In fact, the time for calculating is 

much less than reading time. But for convenience, we 

suppose they are in same level. If we use dichotomy to 

sort one set, the time complexity is O (M ㏒ M), so the 

time for sorting all sets is O (nM ㏒ M). The time for 

getting common elements on N ordered set is O (M*n). At 

last the total time complexity is O (nM ㏒ M+ M*n), 

equal to O (nM(1+㏒ M)). It is still much smaller than O 

(Mn). 

IV. BBI 

The application of bitmap in expressing inverted index 

decreases the space cost, as well as promotes the 

efficiency, but there are two problems for using bitmap on 

huge database. Firstly, every keyword needs a long bitmap 

with the same length of the whole table. If there are many 

records in a table, it is hard to store and use the bitmap. 

Suppose there are 10000000 records in the table, however 

one keyword needs 125000 bytes for bitmap, getting the 

intersection of two 125000 bytes bitmap is also difficult. 

Secondly, high-cardinality columns like time dimension 

have many keywords and it is common that most 

keywords appear rarely, so it wastes much space by using 

bitmap inverted index to express each keyword. For the 

above two problems, we made an improvement on our 

approach. We use bitmap, b-tree and inverted index and 

apply two different methods to high-cardinality columns 

and low-cardinality columns, called BBI. 


A. BBI index for low-cardinality columns 

For low-cardinality columns’ keywords, the number is 

not large and the frequency is high. For these columns we 

use the BBI index structure as shown in Fig. 3. It divides 

the fact table into some segments which still can be 

divided into many blocks. Index blocks are established on 

each table block by using inverted index and bitmap, 

containing the initial and ending address of the block. One 

bit of the bitmap refers to one row and stands for whether 

a keyword appears in that row. After establishing all the 

index blocks, it establishes index segment for the index 

blocks. Index segments has the similar structure as index 

block, but one bit of the bitmap refers to one block and 

stands for whether a keyword appears in that block. It 

establishes chief index on the total index segments and 

also uses a bit to stand for whether a keyword appears in 

that segment. BBI is an open index, using the same 

method, it can build higher order index on the table, but 

not only third order. 

Such as Fig. 3, it establishes index block for each 256 

rows, and builds index segment for each 64 index blocks; 

at last built a chief index for 64 index segments. 

Fig. 3. BBI index for low-cardinality columns 

The main space cost is the storage space of index 

blocks, because the space of index segments and chief 

index is only a small portion of the index blocks’ . We can 

see from Fig. 3 that establishing BBI for one column costs 

the same space as building bitmap index of ORACLE. 

Suppose there are N rows in the fact table, the total 

number of keywords of all the low-cardinality columns is 

M, K bytes bitmap is used for one keyword in each block, 

so it comes that: 

N 

NM 

Space = * M * K = 

8K 

8 

(1) 

It can be derived from (1) that the space of BBI is only 

related to the number of rows and the number of 

keywords. Suppose there are one million rows and one 

thousand keywords, the space cost of BBI for all the 

keywords is 125M. Compared to the space of table with 

one million rows, it is much smaller. 

B. BBI index for high-cardinality columns 

For high-cardinality columns’ keywords, the number 

is large but the frequency is low. For these columns we 

use the BBI index structure like Fig. 2. It also divides the 

fact table into many blocks. Index is established by using


B-tree, inverted index and bitmap. Index is organized by 

b-tree and the inverted indexes are stored in the leaf 

nodes. According to the property of high-cardinality 

columns, it will waste too much space to establish whole 

inverted index for all the keywords in all the rows, so it 

only stores the bitmap for the block in which the 

keywords appear, containing the tab of block. All the leaf 

nodes are linked by indicator for range query. Such as Fig. 

4, there are 3000 keywords in one column, it only find out 

the id of the keywords in dimension table. The keyword 

’1’ exists in some rows of block 5 and block 121, so there 

are two bitmaps in leaf node 1 for keyword ‘1’. Each leaf 

node stores the bitmaps of 10 keywords at most, sorting 

by ID. 

Fig. 4. BBI index for high-cardinality columns 

The main space cost is the storage space of index leaf 

nodes. Suppose there are N rows in the fact table, the total 

number of keywords of all the low-cardinality columns is 

M, each keyword may appear in L blocks on average (it 

means each keyword needs L rows of bitmap in the leaf 

node on average), it uses K bytes for each row of bitmap, 

so it comes that: 

Space = M * L * K 

(2) 

It can be seen from (2) the space of BBI for highcardinality 

columns is related to size of one block, the 

distribution of keywords and the number of keywords. 

Suppose there are 100000 keywords, each keyword may 

exist in 10 blocks on average and each bitmap is 

expressed by 32 bytes, the space cost of BBI is 30M. 

Two extreme conditions are considered. The first 

condition is that every keyword exists in all blocks. It is 

obvious that the space cost can be calculated by (1). The 

second extreme condition is that for one column, the 

keyword in each row is different from each other. So the 

number of keywords is the same as row number. If the 

row number is N and the dimension number is D, we can 

derive that: 

Space = N * D * K 

(3) 

Suppose there are one million rows and one hundred 

high-cardinality columns with the second extreme 

condition, it uses 32 bytes for one bitmap; the BBI space 

cost of these a hundred million keywords is 3.2G. 

Comparing to the space of table with one hundred highcardinality 

columns, it is much smaller. 


V. PERFORMANCE EVALUATION 

In this section the performance evaluation of BBI is 

illustrated. In experiments, we used both real data and 

simulated data to verify the space and time performance 

with the variation of row number and dimension number. 

The experiments are performed on fedora core 10.0 with 

Inter(R) Core(TM)2 2.83G CPU and 2GB RAM, data is 

stored on a local disk which is 7,200-rpm. The space of 

simulated data is 1622.8M which contains 5 million rows 

and 22 dimensions, including 11 high-cardinality columns 

and 11 low-cardinality columns, it is randomly generated. 

The real data set contains 7.79 million rows and 12 

columns, and the space is 1080M. 

A. IO performance of BBI 

For retrieval speed, one of the most important 

influences is I/O cost. On the one hand, if it is not 

consecutively stored in physical address, the efficiency of 

disk I/O will decline. On the other hand, it costs huge I/O 

for querying on huge data, so using BBI to decrease I/O 

cost can increase the retrieval efficiency. 

Fig. 5 shows the change of I/O cost when establishing 

BBI in stages. As the number of columns on which BBI 

established grows, 100 queries are repeatedly executed on 

the simulated data. 

Fig.5. I/O Cost 

In fact, if all the data is read into RAM at a time, it 

doesn’t need disk I/O any more if the following queries 

are executed on the same data. But it is hard to read all the 

huge data to RAM, so only the essential data is read at a 

time in the experiment, and released after query. When 

there is no BBI, the I/O cost is close to 40000M. When 

there are 4 columns with BBI, the I/O cost decreased to 

24100M. When there were 20 columns with BBI, the I/O 

cost is only 5400M. 

B. Space size of aggregation table 

We use the simulated data and simulated queries in 

this experiment. We vary HFJ to see the change of highfrequency 

join number. Fig. 6 shows change tendency of 

ratio of high-frequency joins to all joins when HFJ ranges 

from 1% to 5% and the columns number are 6,10,14,18,22 

respectively. There are 10 high-frequency joins when 

there are 22 columns and HFJ equals to 0.5%, but there 

are only 3 when HFJ equals to 2%. No high-frequency 

joins exist when HJF increases to 4%. It shows the


common discipline by using simulated data and simulated 

queries which are all randomly generated. 

Fig. 7 shows the storage size of aggregation table, 

α =2/3. The storage size of aggregation table decreases 

rapidly with the increase of HFJ, but the tendency is not 

regular. Through analysis we can see that high-cardinality 

columns often affect more than low-cardinality columns, 

because high-cardinality columns contain more keywords. 

So if a high-cardinality column is deducted, the storage 

size of aggregation will drastically decrease. In the 

condition of 22 columns, when the HFJ increases from 

1.5% to 2%, the number of high-frequency joins decreases 

1, but the aggregation table size decreases 200M. At the 

same time, when HFJ increases from 2.5% to 3.0%, the 

number of high-frequency joins also decreases 1, but the 

aggregation table size only decreases 7.1M. 

Fig.6. Number of High-frequency Joins 

Fig.7. Storage Size of Aggregation table 

C. Performance of our approach 

In this section, the overall test is illustrated. Our 

approach has two functions, one is only to establish BBI 

as common index, and apply it to promote the efficiency 

for querying on huge data, the other is to use BBI and 

aggregation table to promote the efficiency for OLAP 

query. Based on real data, we first establish BBI, B-Tree 

index, traditional bitmap index and compare their 

efficiency for kinds of queries, then verify the OLAP 

performance of our approach. 

There are 7.79 million rows in the fact table. We 

choose 6 low-cardinality columns to establish BBI, B- 

Tree and traditional bitmap index, then 6 high-cardinality 

columns. Fig.8 shows the comparison of space costs. First 

two low-cardinality columns only have 9 keywords, both 

storage size of BBI and traditional bitmap are 12M, but B- 


Tree is over 100M. When two high- cardinality columns 

are added, keyword number is over 31000, the traditional 

bitmap already can not be built, its storage size is 

predicted to be over 38000M. At this time, though BBI 

also used bitmap, its size only increases to 550M. At last, 

the number of all keywords is over 8.2 million, the 

traditional bitmap is strongly unsuitable, but the storage 

size of BBI and B-Tree are 767M and 593M, both of 

which show the good space performance for huge number 

of keywords. We can see that BBI has good space 

performance no matter how many the columns are. 

Fig.8. Comparison of Space Costs 

Fig.9. Comparison of Query Performance 

As proved above, BBI has good query performance. 

Suppose there are three conditions, the queries involve 

1,6,12 columns. We simulate queries for each condition. 

Fig. 9 shows the comparison of query performance of BBI 

and B-Tree. BBI always surpasses B-tree in the three 

conditions especially when the columns increase. BBI 

inherits the advantages of bitmap and b-tree, it could 

easily answer reply the range query, multi keywords 

query, ‘or’ query and etc. 

At last, we verify the comprehensive performance 

(containing cube query for OLAP) of our approach. We 

establish the aggregation table on real data with the 

condition thatα =2/3, HFJ=3%. For comparison, we also 

use greedy algorithm to establish materialized views 

which has the same space cost of our approach 

(aggregation table and BBI). Fig.10 shows the comparison 

of comprehensive performance. As cube queries increase, 

it turns up that some queries could not use materialized 

views and aggregation tables, but BBI is suitable for all 

the queries which could use inverted index to established 

data cube for OLAP (paper [17]).


Fig.10. Comparison of Comprehensive performance 


A new approach is introduced in this paper, which can 

be used to decrease the time cost for frequently used 

queries, including cube queries. The core of the approach 

is BBI which is an integrative index, inheriting the 

advantage of bitmap index, b-tree index and inverted 

index. Inverted index represented by bitmap is adopted to 

get result quickly by intersection operations. Meanwhile, 

tree structure is used to accelerate speed for range queries. 

BBI can be built on both high-cardinality and lowcardinality 

columns and it is suitable to all data types. BBI 

established on different columns can cooperate with each 

other to promote the efficiency. Combining with 

aggregation which is based on user-driven, the approach 

can promote the efficiency for of cube queries as well. It 

does not only greatly promote the efficiency of frequently 

used queries, but also improve the performance of other 

queries. This paper discusses the space and time 

complexity of BBI, and the experiment results show good 

performance on space and time of this approach. Future 

work is planned to be focused on live update strategy. 


This work was supported by the National Natural 

Science Foundation of China under Grant NO. 70871033 

and the National High-Tech Research and Development 

Plan of China under Grant NO. 2007AA04Z116. 

REFERENCES 

[1] Joshi.S, Jermaine.C, “Materialized Sample Views for 

Database,” [J] IEEE Transactions on Knowledge and Data 

Engineering, Volume 20, Issue 3, pp: 337 – 351, March 

2008 

[2] Guangrong.Li, Xiaohua.Hu and etc, “A Novel 

Unsupervised Feature Selection Method for Bioinformatics 

Data Sets through Feature Clustering” Proc. Granular 

Computing, 2008. IEEE International Conference on 26-28 

Aug. 2008(GrC 2008.), pp: 41 - 47 

[3] Dimension-Join: A New Index for Data Warehouses 

http://www4.wiwiss.fu-berlin.de/dblp/resource/record/ 

conf/sbbd/BizarroM01 

[4] M.E Morita, R. Sabourin, F. Bortolozzi, and C.Y. Suen, 

“Unsupervised Feature Selection Using Multi-Objective 

Genetic Algorithm for Handwritten Word Recognition”, in 

the 7th International Conference on Document Analysis 

and Recognition, Edinburgh, Scotland, 2003, pp.666-670. 


[5] Gibas.M, Canahuate.G, Ferhatosmanoglu.H, “On row 

Index Recommendations for High-Dimensional Databases 

Using Query Workloads” IEEE Transactions on 

Knowledge and Data Engineering, Volume 20, Issue 

2, Feb. 2008 Page(s):246 - 260 

[6] P·O’Neil, D·Quass. Improved query performance with 

variant indexes [EB/OL]. http://www.cs.duke.edu/~ 

junyang /courses/cps216-2003-spring/papers/oneil-quass- 

1997.pdf,1997-05. 

[7] Jingren.Zho, Larson.P.A, Goldstein.J, Luping.Ding, 

“Dynamic Materialized Views”, Data Engineering, 2007. 

ICDE 2007. IEEE 23rd International Conference on 15-20 

April 2007 Page(s):526 - 535 

[8] Yin.GS ,Yu.X, Lin.LD, ” Strategy of Selecting 

Materialized Views Based on Cache updating”, IEEE 

International Conference on Integration Technology 

Shenzhen, CHINA, MAR 20-24, 2007 pp:789-792 

[9] Jeffrey.Xu.Yu, Xin.Yao, ChiHon.Choi, Gang.Gou. 

“Materialized View Selection as Constrained Evolutionary 

Optimization”, IEEE Transactions on Systems, Man, and 

Cybernetics, Part C: Applications and Reviews, Volume 

33, Issue 4, Nov. 2003 Page(s):458 - 467 

[10] H. Gupta and I. S. Mumick, “Selection of views to 

materialize under a maintenance cost constraint,” in Proc. 

7th Int. Conf. Database Theory,1999, pp. 453–470. 

[11] A. Shukla, P. Deshpande, and J. F. Naughton, 

Materialized view selection for multidimensional 

datasets,in Proc. 24th Int. Conf. Very Large Data Bases, 

1998, pp. 488–499. 

[12] Byeong-Seob You, Dong-Wook Lee, et al. Hybrid Index 

for Spatio-temporal OLAP Operations[A] //International 

Conference on Advances in Information 

Systems(ADVIS 2006). Germany:Springer,2006:110-118. 

[13] Wen Juan, Xue Yongshen, et al. An Efficient Method for 

Multi-Table Joining in Data Warehouse[J]. Journal of 

Computer Research and Development, 2005, 44(11): 

2010~2017(in Chinese). 

[14] M. Frank, E. Omiecinski, and S. Navathe, “Adaptive and 

Automated Index Selection in RDBMS,” Proc. Third Int’l 

Conf. Extending Database Technology (EDBT ’92), 1992. 

[15] S. Choenni, H. Blanken, and T. Chang, “On the Selection 

of Secondary Indexes in Relational Databases,” Data and 

Knowledge Eng., 1993. 

[16] A. Capara, M. Fischetti, and D. Maio, “Exact and 

Approximate Algorithms for the Index Selection Problem 

in Physical Database Design”, Knowledge and Data 

Engineering, IEEE Transactions on Volume 7, Issue 

6, Dec. 1995 Page(s):955 - 967 

[17] Xiaolei Li, Jiawei Han, Hector Gonzalez.Highdimensional 

OLAP:a minimal cubing approach[A]. 

NASCIMENTO M A,OZSU M T,KOSSMANN D,et al. 

International Conference on Very Large Data Bases(VLDB 

2004).San Fransisco:Morgan Kaufmann,2004:528-539. 

[18] Yan.Jun; Liu, Ning; Yan, Shuicheng; Yang, Qiang; Chen, 

Zheng;, “Synthesizing Novel Dimension Reduction 

Algorithms in Matrix Trace Oriented Optimization 

Framework”, Data Mining, 2009. ICDM '09. Ninth IEEE 

International Conference on 6-9 Dec. 2009 Page(s):598 - 

606 

[19] Smalter.A, Huan.Jun, Lushington.G, “Feature Selection in 

the Tensor Product Feature Space”, in the ICDM '09. 2009 

Page(s):1004-1009.


Zhiwei Ni (Tongcheng City, Anhui 

Province,1963), Professor, Doctoral 

supervisor. 

He received his master degree from the 

Department of Computer Science and 

Engineering, Anhui University, Hefei, 

China, 1991 and a PhD degree from the 

Department of Computer Science and 

Technology, Hefei, China, in 2002, all in computer science. 

He is currently a full Professor in the School of Management 

and also the Director for the Institute of Intelligent Management 

in Hefei University of Technology, Hefei, China. His major 

research interests include Artificial Intelligence, Machine 

Learning, Intelligent Management and Intelligent Decision 

Technique. 

Junfeng Guo (Hefei City, Anhui Province, 1983), graduate 

student. He focuses on the research of business intelligence, 

data warehouse and OLAP (online analytical processing). 

Li Wang (Fuyang City, Anhui Province, 1982), PhD student. 

His research field includes business intelligence, data mining 

and cloud computing. 

Yazhuo Gao (Zibo City, Shandong Province, 1984), PhD 

student. Her research field includes business intelligence, data 

mining and OLAP. 



Co-simulation Study of Vehicle ESP System 

Based on ADAMS and MATLAB 

Shengqin Li 

State Key Laboratory of Automotive Safety and Energy, Tsinghua University, Beijing 100084, China 

Email: lishengqin@126.com 

Le He 

State Key Laboratory of Automotive Safety and Energy, Tsinghua University, Beijing 100084, China 

Email: luckhero@263.net 

Abstract—ESP is an active safety system for road vehicles to 

control the dynamic vehicle motion in emergency, the 

composition and working principle of ESP were introduced 

in the paper, and the control technology of ESP was studied 

too. A virtual prototype model of a vehicle model was built 

in ADAMS/Car, and the yaw fuzzy control co-simulation 

model of vehicle was established in Matlab/Simulink, to 

study the stability of vehicle with ESP disabled and enabled 

under sine with dwell. Results showed that, the vehicle 

electronic stability program can make the handling 

performance under big steering wheel angle, and improve 

the vehicle stability. 

Index Terms—co-simulation; virtual prototype; ESP; 

stability 


Electronic stability program (ESP) is an evolution of 

antilock brake technology designed to help drivers 

maintain handling control of their vehicles in high-speed 

or sudden maneuvers and on slippery roads. (Refs. [1], 

[2])Antilock brakes (ABS) have wheel speed sensors and 

the ability to apply brake pressure to individual wheels. 

ESP has additional sensors that monitor how well the 

vehicle is responding to a driver’s steering input. If the 

sensors determine that the vehicle is straying from the 

chosen path, brake pressure will be automatically applied 

as necessary at individual wheels to bring the vehicle 

back to the direction that the driver is steering. In 

addition, in many cases engine power is reduced by 

means of an electronic throttle, thus slowing the vehicle 

down even more. 

ESP is a vehicle control system comprising sensors, 

brakes, engine control modules, and a microcomputer that 

continuously monitors how well the vehicle responds to 

the driver’s steering input. (Refs. [3])The computer 

compares a driver’s commands to the actual behavior of 

the vehicle. In general, when the sensors indicate the 

vehicle is leaving the intended line of travel, ESP applies 

This work was supported by the National high technology program 

Foundation China (Grant No. 2006AA110101) 


doi:10.4304/jsw.6.5.866-872 

the brake pressure needed at each individual wheel to 

bring the vehicle back on track. In some cases ESP also 

reduces the force exerted by the engine. The way ESP 

systems are programmed to respond to the information 

from the sensors varies among vehicle models. Some 

systems intervene sooner and take away more driver 

control of speed than others. 

ESP first appeared in Europe in the 1995 model year 

and in the U.S. market a few years later (Memmer, 2001). 

As is typical of new technologies, ESP initially was 

available as optional equipment on luxury cars. However, 

by model year 2001 it was standard on a number of highselling 

vehicles and available as an option in many more. 

For the 2004 model year, ESP was on all cars and light 

trucks manufactured by Audi, BMW, and Mercedes, and 

on some models produced by just about every other 

automaker. The marketing names of ESP systems vary. 

For example, BMW refers to its system as Dynamic 

Stability Control (DSC), Mercedes calls it Electronic 

Stability Program (ESP), Toyota calls it Vehicle Stability 

Control (VSC), Ford calls it AdvanceTrac, and General 

Motors uses the names StabiliTrak, Active Handling, and 

Precision Control. 

NHTSA estimates that the installation of ESP will 

reduce single vehicle crashes of passenger cars by 34 

percent and single vehicle crashes of sport utility vehicles 

(SUVs) by 59 percent, with a much greater reduction of 

rollover crashes. (Refs. [4], [5])NHTSA estimates that 

ESP has the potential to prevent 71 percent of the 

passenger car rollovers and 84 percent of the SUV 

rollovers that would otherwise occur in single vehicle 

crashes. 

Manufacturers first began equipping vehicles with 

ESP, introduced under many different names, in the mid- 

1990s in Europe, and the technology appeared in other 

markets several years later. As with many new 

technologies, ESP first appeared as an option on more 

expensive luxury vehicles but within a few years was 

being offered as standard equipment on these and other 

less expensive models. Although Europe and Japan 

initially led the way, ESP is now standard on many 

vehicles in the United States. In Europe, 5 million ESP 

are expected to produce annually by 2004, and in the U.S


ESP is just catching on. On the other hand, in South 

Korea the demand of ESP is slowly rising. 

A co-simulation model is established in the paper, 

based on yaw fuzzy control technology, the simulation 

results are studied, and used to study the vehicle handling 

and stability test method and rules. 

II. SYSTEM COMPOSITION AND WORKING 

PRICINPLE OF ESP 

The ESP system aims at helping the driver to maintain 

vehicle stability, its main design philosophy is that the 

system should help the driver to keep the vehicle 

controllable i.e. to avoid excessive vehicle side slip 

angles. (Refs. [6], [7], [8])This is achieved by using 

individual wheel brakes to control the vehicles yaw 

motion. A typical ESP system include traditional brake 

system, sensors (such as, wheel speed sensors, wheel 

steering angle sensor, later acceleration sensor, yaw 

sensor, and rake master cylinder pressure sensor), 

hydraulic modulator, stability electronic control unit 

(ECU), and other support system. 

At the present, the control methods of most ESP 

system are differential braking control. Figure 1 shows 

the action of ESP using single wheel braking to correct 

the onset of oversteering or understeering, if the vehicle 

has entered a left curve that is extreme for the speed it is 

traveling. The rear of the vehicle begins to slide which 

would lead to a vehicle without ESP turning sideways 

unless the driver expertly countersteers. (Refs. [9]~[11]) 

In a vehicle equipped with ESP, the system immediately 

detects that the vehicle’s heading is changing more 

quickly than appropriate for the driver’s intended path, it 

momentarily applies the right front brake to turn the 

heading of the vehicle back to the correct path. (Refs. 

[12])In the situation of understeering, the ESP system 

rapidly detects that the vehicle’s heading is changing less 

quickly than appropriate for the driver’s intended path, it 

momentarily applies the left rear brake to turn the 

heading of the vehicle back to the correct path. 

Figure 1 Note how the caption is centered in the column. 

The agency proposes to adopt the ESP definition based 

on the Society of Automotive Engineers (SAE) Surface 

Vehicle Information Report J2564 (revised June 2004). 

The ESP is defined as a system that has all of the 

following attributes: 

(a) Augments vehicle directional stability by applying 

and adjusting the vehicle brakes individually to induce 

correcting yaw torques to the vehicle. 


(b) Is computer-controlled, which uses a close-loop 

algorithm to limit understeer and oversteer of the vehicle 

when appropriate. 

(c) Has a means to determine vehicle yaw rate and to 

estimate its sideslip or the time derivative of sideslip. 

(d) Has a means to monitor driver steering input. 

(e) Has an algorithm to determine the need, and a 

means to modify engine torque, as necessary, to assist the 

driver in maintaining control of vehicle. 

(f) Is operational over the full speed range of the 

vehicle (except below a low –speed threshold where loss 

of control is unlikely). 

III. VEHICE MODEL 

A. The ADAMS Model 

An ADAMS model is established to study the ESP 

control system, based on the appropriate simplification of 

prototype vehicle. 

The vehicle model is created in Adams, using the 

graphical user interface of Adams Car. (Refs. [13], [14]) 

The modeling process is structured to facilitate ease of 

modification later in the design, starting with creating 

hard points denoting the various key locations of the 

suspension system. This is followed by creating links 

using those hard points, and finally adding joints and 

constraints between the links to complete the geometry. 

The mass and inertia properties are then added to the 

components of the suspension system. 

The suspension parts are created using the cylinder 

member in the Adams Car template tool box. The hard 

points already created mark the end points of each of the 

suspension links, these points are used to create the 

suspension geometry. Each suspension element is 

modeled as a separate part, connected to other parts 

through joints. The prototype vehicle employs a High 

Place Double A-arm suspension on front suspension. For 

each wheel a lower control arm, upper control arm and 

knuckle are modeled as parts. The upper and lower 

control arms are connected to the chassis with bushings 

and to the knuckles with spherical joints. The bushings 

are used to introduce some steering compliance and they 

are modeled to be extremely stiff in the vertical and 

longitudinal directions compared to the lateral direction. 

The prototype vehicle uses a rack and pinion steering 

system. The rack is connected to the chassis with a 

translational joint. 

Figure 2 shows the front High Place Double A-arm 

suspension model in Adams Car, and figure 3 shows the 

rear Multi-links suspension model in Adams Car. The 

springs are modeled as nonlinear single component forces 

(S force in Adams) acting between two points which are 

the mounting points of the strut. The force is defined by a 

2D curve with spring deflection on the x axis and force 

on the y axis. The shock absorbers are modeled as 

nonlinear single component forces acting between the 

same points as the strut. The force is defined by a 2D 

curve with deformation velocity along the x axis and 

force on the y axis.


Once the parts are created, their mass and inertia 

properties are defined. The data for the mass, inertia and 

joint locations were gained by the company. The graphics 

of the vehicle body are modeled, and the mass and inertia 

properties of the body are incorporated in the model. 

Figure 2 The front suspension model. 

Figure 3 The rear Multi-links suspension model. 

The tires and road are modeled using the Adams Tire 

module available in Adams Car. One of the default flat 

road profiles available in the Adams database is used for 

the road. The Pac2002 tire property file suitable for 

P205/55R16 tires was used. (Refs. [15])Road model 

documents are established by the road builder in the 

ADAMS/Car, and the parameters are setup. Figure 4 

shows the vehicle ADAMS model. 

Figure 4 The ADAMS model of vehicle. 


The vehicle model is run only in coast mode for testing 

the ESP systems, and for this reason a drive train is not 

incorporated in the model. Instead, forced motions are 

applied at the wheels to control the speed. The motions 

are then switched off using scripted controls to let the 

vehicle coast while the maneuver is performed. (Refs. 

[16]) Since the scripted controls cannot be used during 

co-simulation with Matlab, torqueses are used instead of 

the forced motions to control the vehicle speed when cosimulating 

with Matlab. 

In any computer model, the accuracy of the simulation 

relies on the accuracy of the model and the vehicle 

parameters used to build the model. Hence the mode is 

checked against experimental data to ensure that the data 

matches to confirm the accuracy of the model. 

In order to validate the ADAMS model, double lane 

change experiment and simulation are carried out, and the 

results are contrasted, shown in figure 5. Figure 6 shows 

the contrast of experiment and simulation under the Sine 

with Dwell maneuver, it can be seen that, the curves of 

simulation and experiment are in good agreement, so it is 

considered that, the ADAMS model can reflect the basic 

characteristics of vehicle, and can be used to simulate and 

analyze the vehicle stability and ESP control system. 

yaw velocity / deg/s 

20 

15 

10 

5 

0 

-5 

-10 

-15 

-20 

experiment 

simulation 

-25 

0 2 4 6 8 

time / s 

10 12 14 16 

Figure 5 Contrast of simulation and experiment under Double Lane 

Change 

yaw velocity/deg/s 

40 

30 

20 

10 

0 

-10 

-20 

-30 

-40 

-50 

experiment 

simulation 

-60 

0 0.5 1 1.5 

t /s 

2 2.5 3 

Figure 6 Contrast of simulation and experiment under Sine with Dwell 

B. Co-simulation Control Model 

In order to simulate the ESP control system, the 

ADAMS model should be translated into S-function in 

Matlab/Simulink. (Refs. [17], [18])Co-simulation is the 

process of simulating a system where two or more 

separate simulation programs are simultaneously used to 

model various aspects of the system and these simulation 

programs communicate during run-time, to simulate the


whole system, thus affecting each other’s output. In this 

case the vehicle is modeled in Adams-Car whereas the 

brake system is modeled in Simulink and a co-simulation 

is setup to run the vehicle model in Adams using the 

brake model in Simulink. 

The various steps involved in setting up a cosimulation 

between Adams and Simulink are: 

a. Loading Adams/Controls 

b. Defining Input and Output Variables 

c. Referencing Input Variables in the Adams Model 

d. Exporting the Adams Block 

e. Connecting the Adams Block and the CES Block in 

Simulink 

f. Running the Co-simulation 

g. Things to Remember 

Figure 6 is the co-simulation principle of ESP, the 

yaw velocity and side slid angle of vehicle body are 

obtained from ADAMS model, and are contrasted with 

the same parameters which are calculated from the 

reference model, thus the stable state of vehicle can be 

estimated and intend to brake the wheel. 

Figure 6 The co-simulation model of ESP system 

The adams_sub module in figure 6 is the S-function 

obtained from ADAMS model, which includes the whole 

vehicle information. The input and output state variables 

have been defined while the sub-system being 

established, such as, the brake pressure of each wheel was 

defined as input state variable while the brake model 

being established, the later velocity, longitude velocity 

and yaw velocity were defined output state variables 

while vehicle body being established, and the steering 

wheel angle was defined as output state variable while 

steering system being established. Desired module is the 

linear two degrees of freedom vehicle model, used to 

calculate the desired state parameters of vehicle. ESP 

module is the core of co-simulation model, which can 

complete the estimation of vehicle stable state and active 

yaw control. The fuzzy control principle is adopted based 

on the yaw velocity in the paper, which is shown in figure 

7. 


Figure 7 The fuzzy control principle based on yaw velocity 

Once the system is setup, the simulation time is 

entered in the box on top of the screen and the play 

button is clicked to run the simulation. (Refs. [19]) 

Simulink invokes Adams and runs the model in 

Adams/Car while the damper forces are calculated in 

Simulink and fed into Adams while the simulation is 

running. Thus co-simulation is achieved. 

IV. SIMULATION RESULTS 

NHTSA has proposed a new Federal motor vehicle 

safety standard (FMVSS). FMVSS No. 126, Electronic 

Stability Control Systems, would require ESP systems on 

passenger cars, multipurpose passenger vehicles, trucks 

and buses with a gross vehicle weight rating of 4,536 Kg 

(10,000 pounds) or less. 

As shown in Figure 8, the Sine with Dwell maneuver 

was based on a single cycle of sinusoidal steering input. 

A single cycle input is performed at a frequency of 0.7 

Hz, with a 500 ms pause between completion of the third 

quarter cycle and initiation of the fourth quarter cycle. 

Figure 8 The Sine with Dwell maneuver 

To begin the maneuver, the driver accelerates the 

vehicle to a speed of approximately 52 mph, at which 

point the throttle is released and a programmable steering 

controller is engaged. Since the maneuver entrance speed 

is always 50 mph, increasing the magnitude of the 

steering wheel angles is used to increase maneuver 

severity. This is accomplished by multiplying the steering 

wheel angle capable of producing a lateral acceleration of 

0.3g during Slowly Increasing Steer testing (δ0.3g) by a 

series of scalars. The steering wheel angles nominally 

begin at 1.5*δ0.3g, and are increased in increments of 

0.5*δ0.3g until the steering wheel angle of 6.5*δ0.3g or 

270 degrees is used (whichever was greater). Sine with


Dwell tests are performed with left-right and right-left 

steering. 

The proposed criterion performance limit establishes 

the displacement threshold to ensure that the ESP 

intervention used to achieve acceptable lateral ability 

dose not compromise the ability of the vehicle to 

response to the driver’s input. The proposal would require 

that an ESP-equipped vehicle a gross vehicle weight 

rating of 3500Kg or less would have a lateral 

displacement of at least 1.83 meters at 1.07 seconds after 

the initiation of steering, and the ESP-equipped vehicle a 

gross vehicle weight rating of 3500 Kg above would have 

a lateral displacement of at least 1.52 meters at 1.07 

seconds after the initiation of steering. 

Based on consideration of all available test data, 

NHTSA ultimately decided a metric based on the YRR 

1.0 seconds after completion of steer would meet the two 

requirements and effectively augment the later value, as 

indicated in Figure 9. Specifically, the yaw rate ratio of 

the vehicle are measured at 1.5 to 1.75 seconds after 

completion of steer, where the yaw rates of the vehicles 

equipped with fully enabled ESP systems had decayed to 

approximately zero while those associated with the fully 

disabled tests remained quite high. 

Figure 9 Steering wheel position and yaw rate information used to 

assess lateral stability 

A formal definition of the lateral stability performance 

criteria is provided below. 

⎧ rT 

0+ 

1 

⎪ × 100% 

≤ 35% 

⎪ rpeak 

⎨ 

r 

⎪ T0 

+ 1. 

75 

× 100% 

≤ 20% 

⎪ 

⎩ rpeak 

In both criterion, rpeak is the first local yaw rate peak 

produced after the second steering reversal, 

r T0+ xyaw 

rate 

at x (x=1, 1.75) seconds after completion of a maneuver’s 

dynamic steering inputs. 

The sine with dwell maneuver simulation is carried out 

on the road with high attachment coefficient, according to 


FMVSS 126 sine with dwell standard, the results of 

vehicle with ESP disabled and enabled are shown in 

figure 10 and figure 11. As shown in Figure 10, when the 

steering amplitude is 80 and 120 degrees, following a 

smooth 0.7 Hz sinusoidal pattern, the absolute value of 

the yaw velocity increases with the absolute value of the 

steering angle, and then the vehicle changes to clockwise 

yaw velocity in response to right steering. At two seconds 

after the beginning of steering, the steering wheel has 

been turned back to straight ahead, and the yaw rate 

returns to zero after a fraction of a second response time. 

At that point, the vehicle is being steered straight ahead, 

and it is going straight ahead without any yaw rotation. 

The vehicle is responding closely to the steering input, 

and the driver is in control. However, when the steering 

amplitude is increased to 169 degrees, the vehicle spins 

out, exhibiting oversteer loss of control. This condition is 

identified in the yaw rate trace. When the steering is 

straight ahead at time = 2 seconds, the yaw rate for this 

run is still about 35 deg/sec. However, there is a time lag 

past the instant of steering to straight ahead even for the 

previous runs where there was no loss of control. What is 

different is that the yaw rate does not swiftly decline to 

zero as it does with a vehicle under control. At time = 3 

seconds, the yaw rate is still the same, and it has actually 

increased at time =4 seconds in this example. The 

physical interpretation of this graph is that the driver has 

turned the wheels straight ahead and wants the vehicle to 

go straight, but the vehicle is spinning clockwise about a 

vertical axis through its center of gravity. It is out of 

control in a spinout. The driver’s steering input is not 

causing the vehicle to take the desired path and heading, 

and the vehicle would depart the road surface sideways or 

even backward. 

Figure 10 Sine with dell maneuver test of a vehicle without ESP 

Figure 11 shows another series of tests of the same 

vehicle but with ESP enabled. The first two runs were at 

80 and 120 degrees of steering angle, and the vehicle’s 

yaw rate declined to zero in a fraction of a second after


the steering command. This is the same good response to 

steering exhibited by the vehicle with ESP disabled in the 

previous figure. The third run was conducted at 180 

degrees of steering angle. This is greater than the 169 

degrees that caused a severe loss of control without ESP, 

but the yaw rate returned to zero with the steering angle 

just as quickly as in the runs with less steering. The final 

set of curves in Figure 11 represents a run conducted with 

279 degrees of steering angle. This would be the left-right 

portion of the performance test proposed for the ESP 

system of this vehicle since 279 degrees is 6.5 times the 

steering angle that produces 0.3g steady state lateral 

acceleration for this example vehicle. In this case, the 

yaw rate did not return to zero nearly instantaneously as it 

had at lower steering angle. Instead, it steadily declined 

after the steering was turned to straight ahead, and the 

vehicle was completely stable and going straight in about 

1.75 seconds. Clearly, the vehicle remained in control 

compared to its behavior without ESP (see Figure 10) in 

which turning the steering to straight ahead had no effect 

on the vehicle’s heading. However, the ESP system 

required some time to cause the vehicle to stop turning in 

response to the driver’s straight ahead steering command. 

It can be concluded that, ESP can make the handling and 

stability performance on big lateral acceleration and slip 

angle improved, and make the driver drive the vehicle 

normally. 

Figure 11 Sine with dell maneuver test of a vehicle with ESP 


(1) The virtual prototype model of vehicle and cosimulation 

model of ESP control system based on the 

fuzzy control principle were established in ADAMS and 

Matlab, in order to simulate study the performance of 

ESP. From the simulation, the model was shown to give 

accurate results for the purposes of this study; 

(2) The ESP model was developed in Matlab/Simulink 

and a co-simulation was set up to integrate the ESP 

model with the vehicle model. ESP fuzzy controller based 

on the yaw velocity can improve the controlling stability 


of the automotives by initiatively finishing the 

implementation of the wheel braking, thus the driver can 

help the driver to keep the vehicle controllable; 

(3) The design cost can be decreased, and the 

development cycle can be shortened, by means of virtual 

prototype and co-simulation technology. 


The authors wish to give their sincere thanks to the 

editor and the anonymous referees, for their valuable 

suggestions and helpful comments which improved the 

presentation of the paper. 

REFRENCE 

[1] CHARLES M. FARMER. “Effect of Electronic Stability 

Control on Automobile Crash Risk,” Traffic Injury 

Prevention, pp. 317–325, May, 2004. 

[2] Ma Chun-Hui, Wu Zhi-Lin, Wang Liang-Mo, Li Song- 

Yan. “Modeling and controlling method for vehicle ESP 

system,” Journal of Nanjing University of Science and 

Technology, Vol.34, pp.108-112, February 2010. (In 

Chinese) 

[3] Y.Shibahata, K.Shimada and T.Tomari. “Improvement of 

Vehicle Maneuverability by Direct Yaw Moment Control,” 

Vehicle System Dynamics, Vol.22, pp.465-481, 1993. 

[4] WANG Xia, HANG Li. “The dynamics simulation of 

automotive controlling stability,” Journal of Shenyang 

Institute of Aeronautical Engineering, Vol.26, No.4, pp.24- 

26, 2009 (In Chinese) 

[5] “Vehicle Modeling and ADAMS-SIMULINK CO- 

SIMULATION with integrated continuously controlled 

Electronic Suspension (CES) and Electronic Stability 

Control (ESC) Models,” [D]: Sughosh Jagannatha Rao, 

B.S.M.E, Univ. of Ohio State University 

[6] MSC Software Corp., “Getting Started Using 

Adams/Controls Introducing and Starting the Tutorials”, 

Mechanical Dynamics Inc., 2002 

[7] Chen, B.-C., Peng, H., “Differential braking based rollover 

prevention for Sport Utility Vehicles with human-in-theloop 

evaluations”. Vehicle System Dynamics, Vol. 36, No. 

4-5, pp. 359-389, 2001. 

[8] U.S. Department of Transportation. “National Highway 

Traffic Safety Administration, Laboratory Test Procedure 

for FMVSS 126, Electronic Stability Control Systems, ” 

TP-126-01, April 10, 2008 

[9] FAN Xiao-bin; XIA Qun-sheng. “Vehicle Stability Direct 

Yaw Moment Fuzzy Control Based on Virtual 

Prototyping,” Tractor & Farm Transporter, Vol.37, pp.47- 

49, 2010 (In Chinese) 

[10] Kinjawadekar, T., Dixit, N. Heydinger, G.J., Guenther, 

D.A., and Salaani, M.K., “Vehicle Dynamics Modeling 

and Validation the 2003 Ford Expedition with ESC using 

CarSim”, SAE Paper 2009-01-0452, April 2008 

[11] Federal Register; Friday, April 6, 2007; Part II, 

Department of Transportation; National Highway Traffic 

Safety Administration; 49 CFR Parts 571 and 585, Federal 

Motor Vehicle Safety Standards; Electronic Stability 

Control Systems; Controls and Displays; Final Rule 

[12] Dang, Jennifer N., “Preliminary Results Analyzing The 

Effectiveness of Electronic Stability Control (ESC) 

Systems” September 2004, DOT HS809790 

[13] Pan, W. and Papelis, Y.E., “Real-Time Dynamic 

Simulation Of Vehicles With Electronic Stability Control:


Modeling And Validation,” Vehicle Systems Modeling and 

Testing, Vol. 1, Nos. 1/2/3, 2005 

[14] Tingvall C, Krafft M, Kullgren A, Lie A. “The 

effectiveness of ESP (electronic stability program) in 

reducing real-life accidents,” Traffic Injury Prevention, 

Vol. 5, pp. 37–41, 2004 

[15] Najm W, Sen B, Smith J, Campbell B. “Analysis of light 

vehicle crashes and pre-crash scenarios based on the 2000 

General Estimates System,” Report no. DOT-HS-809-573. 

U.S. Department of Transportation, Washington, DC, 2003 

[16] Brown, T. A. et al., “Rollover Stability Control for an 

Automotive Vehicle, ” US patent No. 6,263,261 B1 

[17] Van Zanten, A. T., “Bosch ESP systems: 5 years of 

experience, ” SAE 2000-01-1633 

[18] Branicky, M.S., “Studies in Hybrid Systems: Modeling, 

Analysis, and Control,” PhD Thesis, Laboratory for 

Information and Decision Systems, MIT, Cambridge, 

USA, 1995 


[19] Fennel, Ding, ‘A Model-Based Failsafe System for the 

Continental Teves Electronic Stability Program (ESP)’, 

SAE 2000-01-1635 

Li Shengqin graduated from 

Northeast Forest University (NEFU), 

Harbin, Heilongjiang, P.R. China, in 208, 

and received her Doctor Degree in 

Vehicle Application Engineering. 

She is currently an associate 

professor in the Traffic College, 

Northeast Forest University, China, and 

engages in the postdoctoral research in 

the State Key Laboratory of Automotive Safety and Energy, 

Tsinghua University, Beijing, China. Her research interest 

covers vehicle dynamics and control.


An Improved Fuzzy C-means Clustering Algorithm 

based on PSO 

Abstract—To deal with the problem of premature 

convergence of the fuzzy c-means clustering algorithm 

based on particle swarm optimization, which is sensitive to 

noise and less effective when handling the data set that 

dimensions greater than the number of samples, a novel 

fuzzy c-means clustering method based on the enhanced 

Particle Swarm Optimization algorithm is presented. Firstly, 

this approach distributes the memberships on the basis of 

the distance between the sample and cluster centers, making 

memberships meet the constraints of FCM. Then, 

optimization strategy is presented that the optimal particle 

can be guided to close the group effectively. The 

experimental results show the proposed method significantly 

improves the clustering effect of the PSO-based FCM that 

encoded in membership. 

Index Terms—clustering, particle swarm algorithm, fuzzy C 

means, membership, constraint strategy 


Generally speaking, clustering is in accordance with 

certain requirements and rules to distinguish between 

things, and classification of the process. Clustering 

algorithm is a set of classification of the data that 

distribution is unknown, the aim is to find the structure 

hidden in data, and as much as possible to make the data 

that have same nature attributed to the same class 

according to some measure of similarity degree. 

Clustering is a form of unsupervised learning whereby 

objects that similar to each other are put into the same 

cluster. It is the first stage of knowledge acquisition 

concerning a group of objects that is obtaining knowledge 

of classes. 

Fuzzy clustering methods that based on the objective 

function is the most studied in the literature and the most 

widely used in practice, such algorithm takes the 

clustering problem as a constrained optimization problem, 

by solving the optimization problem to determine the 

fuzzy partition and the clustering results in data set. Such 

algorithms are characterized by simple and easy to apply 


doi:10.4304/jsw.6.5.873-879 

Qiang Niu 

School of Computer Science &Technology 

China University of Mining & Technology 

Xuzhou Jiangsu, China 

niuqiang@vip.163.com 

Xinjian Huang 

School of Computer Science &Technology 

China University of Mining & Technology 

Xuzhou Jiangsu, China 

xinjian1020@yahoo.cn 

and clustering performance is good, can take use of the 

classical optimization theory as its theoretical support, 

and easy for the programming. 

Fuzzy c-means clustering algorithm (FCM) [1-2] is an 

effective algorithm and is one of the most used clustering 

methods. But when the data set has a higher dimension, 

the clustering effect of FCM is poor, and it is difficult to 

find the global optimum [3-4]. 

Particle Swarm Optimization (PSO) [11] is one of the 

modern heuristic algorithms under the evolutionary 

algorithms, and has proved to be very effective for 

solving global optimization, and gained lots of attention 

in various engineering applications. It is not only a 

recently invented high-performance optimizer that is easy 

to understand and implement, but it also requires little 

computational bookkeeping and generally only a few 

lines of code. It is a stochastic search technique with 

reduced memory requirement, computationally effective 

and easier to implement compared to other evolutionary 

algorithms. 

Clustering problems can be attributed to optimization 

problems under certain conditions, Particle Swarm 

Optimization is an optimization algorithm based on the 

theory of swarm intelligence, which could be 

implemented and applied easily to solve various function 

optimization problems, or the problems that can be 

transformed to function optimization problems. PSO is 

easy to describe and implement, it also has a strong 

global search capability and a faster convergence [12]. 

Many PSO-based Fuzzy Clustering Algorithms are 

proposed [5-9]. However, in most of these algorithms the 

particle is encoded by cluster centers, less of these 

algorithms use the method that the particle is encoded by 

membership. 

If a data set has n samples and c clusters, each sample 

has d dimensions. While encoded by membership, a 

particle is an one-dimensional row vector with n×c rows. 

While encoded by cluster centers, a particle is an onedimensional 

row vector with c×d rows. In a data set, d is


usually less than n, then c×d is less than n×c, so most 

of PSO-based FCM is encoded by cluster centers. But 

when the particle is encoded by cluster centers, the range 

of particle is difficult to confirm for different clusters 

have different centers. When the particle is encoded by 

membership, the rang of particle is [0,1], and the PSObased 

FCM is better than FCM on processing the data 

that d is more than n [9]. When the particle is encoded by 

membership, the sum of the membership between a 

sample and all cluters should be one, this is the constraint. 

Thomas A. Runkler et al. put forward a method for Fuzzy 

clustering constraints when the particle is encoded by 

membership in [9]. When the sum of membership 

between a sample and all cluters is not one, the method in 

[9] increases or decreases the insufficient or extra parts 

evenly. And their method is sensitive to noise, and less 

effective when handling the data set that dimensions less 

than the number of samples [9]. 

In order to solve the above problems, this paper 

proposes an improved method for the distribution of 

membership, having a better effect on handling the data 

containing noise, and also on low dimensional and high 

dimensional data sets. At the same time, optimization 

strategy is presented that the optimal particle can be 

guided to close the group effectively. 

The rest of the paper is organized as follows. After the 

introduction, Section II gives a description of the 

generalized FCM clustering. In Section III, PSO is briefly 

described, and the improved method is introduced. 

Section IV provides the experiments conducted over three 

different data sets and discusses the results. Finally, 

Section V concludes the paper. 

II. FUZZY C-MEANS CLUSTRING ALGORITHM 

Different clustering criteria can produce different 

clustering methods. Clustering algorithm can be put into 

traditional hard clustering and fuzzy clustering algorithm 

if in accordance with the range of membership. 

Traditional hard clustering division is “either-or” type of 

a division, namely the membership for a sample to a 

clustering is either 0 or 1. Since fuzzy clustering 

algorithm extends the range of membership, so that it can 

take any value within 0 to 1, which has better clustering 

effect and data expression, has become a hot research in 

this field. Fuzzy C-means algorithm theory is 

substantially complete, applications are relatively wide. 

The following is a brief introduction on fuzzy C means 

clustering algorithm. 

The data set � has n samples and c 

n 

clusters, and . The objective function of FCM is: 

X X X X � , , � 1 2 � 

p 

X � R 

m 

i 

n 

c 

�� 

m 

2 

J ( U, 

E) 

� ( � ) || x � e || (2.1) 

k�1i�1 

m is constant, and m>1. Cluster i is expressed as 

e . The membership between sample k and 

i ( i � 1, 

2, 

� � �, 

c) 

cluster i is expressed as � ( i � 1, 

2, 

�� 

�, 

c, 

k � 1, 

2, 

�� 

�, 

n) 


ik 

ik 

k 

i 

�ik 

�{ 0, 

1}, 

�i, 

k; 

� �ik 

� 1, 

�k 

i�1 

c 

(2.2) 

The method to optimize the FCM model is alternating 

optimization through the necessary conditions: 

� 

ik � 

c 

� 

j �1 

n 

� 

k �1 

ei � n 

1 

|| xk 

� ei 

|| 

( ) 

|| x � e || 

k 

j 

2 

m�1 

(2.3) 

m 

( � ik ) xk 

(2.4) 

, 1 � i � c 

m 

( � ) 

� 

k �1 

ik 

The final aim is minimizing (2.1) [10]. 

The specific steps for FCM: 

Step 1: set the number of clusters c (2≤c1), initializing the matrix of membership (or 

initializing cluster centers), set the maximum iterations n. 

Step 2: calculate various cluster centers (or the matrix 

of membership). 

Step 3: calculate the matrix of membership (or various 

cluster centers). 

Step 4: repeat step 2 and step 3, until the completion of 

the maximum number of iterations. It can also set a 

convergence precision as the condition for a loop 

terminates. 

III. IMPROVED FCM CLUSTERING ALGORITHM BASED ON 

PSO 

A. Fuzzy c means algorithm based on PSO 

Particle swarm optimization (PSO) [11] is an 

optimization algorithm based on the theory of swarm 

intelligence, the cooperation and competition among 

particles produced swarm intelligence to guide the 

optimization search. PSO is easy to describe and 

implement, it also has a strong global search capability 

and fast convergence. But PSO also has defects, as in 

convergence condition, all particles are in the direction of 

the optimal, if optimal particle is not good enough, it can 

easily fall into local optimum [12]. 

In the D-dimensional space, the number of the particle 

is m. [ , , , ] 

1 2 

D i 

i i i X � X X � X is the position of particle i, 

its velocity is , ] D i 

Vi � [ Vi , V , V 

1 i � , its best position is 

2 

Pi � [ Pi 

, P , , ] 

1 i � P . 

2 i 

, ] 

D 

i , , � P 

1 i2 

iD 

P [ Pg � P 

is the best 

position of all particles. Each particle velocity is updated 

by (3.1), then each particle position is updated by (3.2). 

n 

n 

�P � X �� c r �P X � 

V � 

n�1 

n 

i � Vi 

� c r 

(3.1) 

1 1 i i 2 2 g i 

�1 

X � X � V 

(3.2) 

n 

i 

n 

i 

In (3.1) and (3.2), i=1,2, … ,m expresses different 

particles, c1 and 2 are acceleration constants; and 

are random real numbers drawn from [0,1]; n denotes 

evolutionary epochs [13]. From a sociological point of 

c 1 r 2 r 

n 

i


view, the first part of (3.1) as “Memory” entry, is the 

previous velocity, that the current velocity is by the 

impact of previous velocity; the second part is 

“cognitive” entry, represents particles itself thinking; the 

third part is the “social” entry, reflects the collaboration 

between the particles and information sharing, which 

guide the particles toward the optimal position in the 

entire group. 

PSO algorithm steps: 

Step 1: initialize the particle swarm, including 

population size, initial position and velocity of particles, 

etc. 

Step 2: calculate fitness for each particle, storage each 

particle best position P and its fitness, and choose the 

best 

particle that has the best fitness as G ; best 

Step 3: update the velocity and the position of each 

particle according to (3.1) and (3.2); 

Step 4: calculate the fitness of each particle after 

update the position, compare the fitness of each particle 

with its best previous fitness P , if better than it, then 

best 

set the current position as P ; best 

Step 5: compare the fitness of each particle with the 

group best previous fitness, if better than it, then set the 

current position as G ; best 

Step 6: search algorithm to determine whether the 

results meet the conditions set by the end of (usually 

good enough to adapt to a preset value or the maximum 

number of iterations), if preconditions not met, then 

return to Step 3; if preconditions are met, then stop 

iteration, output the optimal solution. 

While PSO is used in FCM, particle can be encoded by 

membership or by cluster centers. If the data set has n 

samples and c clusters, each sample has d dimensions. 

When particle is encoded by membership, a particle is an 

one-dimensional row vector with n × c rows, that is 

�x 11, 

x12, 

�, x1c 

, �, 

xn1, 

xn2, 

�, 

x � , x 

nc ij means the 

membership between sample i and cluster j. When 

particle is encoded by cluster centers, a particle is an onedimensional 

row vector with c × d rows, that is 

�x 11, 

x12� 

x1d 

, �, 

xc1, 

xc2, 

�, 

xcd� 

, xij 

means the value 

of cluster i in dimension j. If n>d, encoded by cluster 

centers is simple, and could better handle data sets that 

n>d. If n


The distance between sample k and cluster i is , the 

membership between sample k and cluster i is u , 

uik 

�[ 

0, 

1] 

. If � 1, 

the change of is determined 

u 

l 

by . 

i 

c 

� 

i�1 

c 

�uik ik 

i�1 

If u � 1 , use (3.5), smaller distance, more plus; 

ik 

greater distance, less plus. 

c 

� 

i�1 

c 

u ik � uik 

� ( 1� 

�uik 

) � ( 1� 

li 

c ) (3.5) 

i�1 

l 

� 

i 

i�1 

If u � 1 , use (3.6), greater distance, more 

ik 

reduction; smaller distance, less reduction. 

l 

( (3.6) 

c 

i 

ik � uik 

� 1� 

�uik 

) � c 

i�1 

�li 

i�1 

u 

Secondly, for optimal particle in PSO has an important 

role in guiding the group, if can get better optimal particle 

in each iteration, then can speed up the convergence and 

optimize cluster results, so this paper puts forward a new 

optimization method for optimal particle. The new 

method optimizes optimal particle through optimizing the 

worst sample in the optimal particle. 

Data set X � { x , ,..., } has n samples and 

1 x2 

xn 

c( 2 � c � n) 

clusters, m>1 is constant, � means the 

ik 

membership between sample k and cluster i, ei 

means 

cluster i. Using (3.7) get the fitness of sample k, that is 

L(k) 

, to fix the worst sample in the optimal particle. 

Greater fitness, the worse the sample. 

c 

� 

i�1 

m 

2 

L( 

k) 

� ( � ) x � e (3.7) 

� means the membership between the worst sample 

n 

and cluster n, d means the distance between the worst 

n 

sample and cluster n. The irrational distribution of 

membership cause the worst sample, the most reasonable 

distribution is that � and are proportional. 

n n d 

Let � d �� d � � �� 

nd and 1 1 2 2 

n �1 �� 2 �� 

�� 

n � 1 , 

t 

gets 

k � � k � 

, tk � d d d �� dk 

dk 

�� 

� d , so 

1 2 3 �1 �1 

n 

t1 

� t2 

�� 

� �tn 

� and d are proportional. 

n n 

This improved method firstly considers the distance 

between sample and different clusters. For the clusters 

near the sample, if the sum less than one, plus more, if the 

sum more than one, decreased less; for the clusters away 

from the sample, if the sum less than one, plus less, if the 

sum more than one, decreased more. Improved constraint 


ik 

k 

i 

l 

i 

ik 

method makes the sample anear its closer clusters and 

away from clusters that far from it in each iteration. Then, 

optimizing the worst sample in the optimal particle, to 

ensure the membership in near clusters big and the 

membership in far clusters small, optimization of the 

worst sample means optimizing the optimal particle at the 

same time, serve to the purpose that speeding up the 

convergence and optimizing cluster results. Compare to 

method in [9], the improved method adds the process of 

computing the distance, but gives up the process of using 

the sigmoid function (i.e. (3.3)), so the computed amount 

between the improved method and method in [9] is not 

obviously. 

IV. EXPERIMENT TESTING AND COMPARATIVE ANALYSIS 

Hardware environment of experiment is PC with 

Intel(R) Core(TM)2 Duo, CPU E7400 2.80GHz, 2GB 

RAM. Operating system is Windows XP Professional, 

program code is achieved in platform of Visual Studio 

2005 using C#. 

Test data sets are: (1) Single outlier data set: [- 

1.2,0.5,0.6,0.7,1.5,1.6,1.7], 7 group with 1dimension data, 

one with the point [0.5,0.6,0.7], one with the point 

[1.5,1.6,1.7], a single outlier at -1.2. (2) Iris data set: 150 

vectors with 4 features, has 3 clusters, each cluster has 50 

samples. (3) Lung cancer data set: 32 vectors with 56 

features (for containing 5 unknown data, only use 32 

vectors with 54 features), has 3 clusters, the first contains 

9 samples, the second contains 13 samples, the third 

contains 10 samples. 

When the number of particle is 10 and with 100 

iterations, table 1 to table 3 show the index of average 

clustering effect after running various methods. DIC 

stands for average distance inside clusters, DBC stands 

for average distance between clusters, OFV stands for 

objective function value, SCR stands for successful 

classification rate. 

TABLE I. 

SINGLE OUTLIER DATA SET 

FCM Paper [9] Improved method 

DIC 0.28 0.56±0.05 0.26±0.01 

DBC 2.28 0.44±0.24 2.29±0.03 

OFV 1.54 2.79±0.11 1.53±0.07 

SCR (%) 85.71 82.86±11.43 85.71 

TABLE II. 

IRIS DATA SET 


DIC 0.65 1.93±0.03 0.67±0.02 

DBC 3.30 0.13±0.12 2.98±0.04 

OFV 160.51±0.01 227.36±0.03 155.92±8.52 

SCR (%) 89.33 46.60±2.07 90.47±2.33 

TABLE III. 

LUNG CANCER DATA SET 


DIC 4.24±0.37 4.32±0.04 4.07±0.10 

DBC 0.04±0.01 0.23±0.06 2.07±0.47 

OFV 204.32 200.49±0.03 123.01±8.72 

SCR (%) 42.81±7.19 61.25±7.50 69.38±3.75


Table I shows the improved method is better than the 

method in [9] in various performance, and better than 

FCM in DIC and DBC. From table I, the improved 

method is better than FCM in various performances 

except in SCR. The improved method is obviously better 

than method in [9] in various performances too. Also, the 

method in [9] is worse than FCM in all performances. So, 

we can know the improved method has the best effect in 

data sets that has noise, and we can know the difference 

between the improved method and FCM when handling 

small data sets is not obvious. 

Table II shows the improved method is better than the 

method in [9] obviously, and also better than FCM in the 

iris data set. Comparing FCM with method in [9], we can 

know FCM is better than method in [9], and the 

difference is obviously, because when PSO based FCM is 

encoded by membership, the algorithm can only better 

handle data sets that dimension greater than the number 

of sample, do not has a better performance in handling 

data sets that dimension less than the number of sample, 

like iris data set. However, table II shows the improved 

method makes the PSO based FCM that encoded by 

membership can also have a better effect in handling data 

sets that dimension less than the number of sample, and 

the advantages are obvious. 

Table III shows the improved method is better than the 

method in [9] and FCM obvious in lung cancer data sets. 

From the comparation between FCM and the method in 

[9], we can see FCM is not suitable for data sets that 

dimension better than the number of sample, PSO based 

FCM that encoded by membership can better handle them, 

and the improved method is obviously better than the 

method in [9]. 

Table I to table III show the improved method is better 

than FCM and method in [9] not only in data sets that 

have noise, but also in data sets that have a low 

dimension or a high dimension, when the iteration is 

taken at 100. 

Take the number of particle at 10, increasing iteration 

from 1 to 100, Fig. 1 to Fig. 6 show the change of 

successful classification rate and objective function value 

by using different methods. OFV means objective 

function value, SCR means successful classification rate. 

Fig. 1 shows the method in [9] is worse than the other 

two methods in successful classification rate, the 

difference between FCM and the improved method is not 

obvious too when the iteration is change. So we can see 

the method in [9] is more sensitive to noise than the other 

two methods. 

From Fig. 2, we can see the improved method has a 

faster convergence than the method in [9], and finally 

better than FCM. The change of objective function value 

in the improved method has a clear trend that is the value 

of objective function in the improved method decrease 

with the increase of iterations. 


Figure 1. Comparation of SCR in single outlier data set. 

Figure 2. Comparation of OFV in single outlier data set. 

Figure 3. Comparation of SCR in iris data set.


Figure 4. Comparation of OFV in iris data set. 

From Fig. 3, knows the improved method is better than 

the method in [9] obviously, successful classification rate 

in the improved method and the method in [9] is not at 

the same level, and the improved method has a faster 

convergence with the increase of iterations, then finally 

better than FCM. When the iteration is 32, the improved 

method nearly has a convergence, compare to the method 

in [9] with a convergence number at 56. While the 

number of iteration increases to 70, the successful 

classification rate of the improved method is better than 

FCM. 

From Fig. 4, knows objective function value in the 

improved method and the method in [9] is not at the same 

level too, and the improved method has a fast 

convergence with the increase of iterations, using less 

than 10 iterations. The overall level of the improved 

method is significantly lower than FCM, but the overall 

trend of the improved method in objective function value 

is not stable enough. The trend that PSO based FCM 

encoded by membership can also has a good effect on 

data sets that dimension less than the number of sample is 

obviously. 

Figure 5. Comparation of SCR in lung cancer data set. 

Fig. 5 shows the performance of FCM is worse, the 

improved method is better than the method in [9] 

obviously, and having a faster convergence. From Fig. 5, 

knows the curve of FCM is the worst one, having the 


lowest successful classification rate, using 78 iterations to 

come to a convergence. The method in [9] has a middle 

level successful classification rate, but the rate of coming 

to convergence is the biggest one, using 80 iterations. The 

improved method only uses 12 iterations to come to a 

convergence, the successful classification rate is 

significantly high than others, is the best one. 

Figure 6. Comparation of OFV in lung cancer data set. 

Fig. 6 displays the performance of FCM is the worst 

one, the method in [9] is little better than FCM, but worse 

than the improved method obviously. The performance of 

FCM and the method in [9] is stable enough in objective 

function value, but they are so worse. The improved 

method in objective function value has a fast convergence 

that only uses 18 iterations, and with the increase of 

iteration, the volatility of the curve decreased. 

To sum up, the experiment using single outlier data set 

show the method in [9] is sensitive to noise, the improved 

method and FCM can handle the noise better; the 

experiment using iris data set show the new method 

improves the clustering effect of PSO based FCM better 

than FCM in data sets that dimensions less than the 

number of samples; the experiment using lung cancer 

data set show FCM is not suitable for high-dimensional 

data sets, the improved method gets the best clustering 

effect. So the improved method is better than the method 

in [9] obviously, having a faster convergence, and better 

than FCM in various data sets. 


For the problems when PSO-based fuzzy clustering 

algorithm is encoded by membership, this paper improves 

the method of achieving constraint and puts forward an 

optimization method for optimal particle. In the previous, 

PSO-based FCM that encoded by membership can only 

better handle data sets that dimensions greater than the 

number of samples, but not suitable for data sets that 

dimensions less than the number of samples. 

Three typical data sets are used to verify different 

algorithms. Experiments show that the improved method 

can handle the noise better than previous methods, further 

improves clustering effect in data sets that dimensions 

greater than the number of samples, and gets better effect 

than FCM in data sets that dimensions less than the


number of samples at the same time, making PSO based 

FCM encoded by membership can better handle data sets 

that dimensions less than the number of samples too. The 

desired effect is achieved 

REFERENCES 

[1] Maria Halkidi, Yannis Batistakis and Michalis 

Vazirgiannisv. On Clustering Validation Techniques. 

Journal of Intelligent Information Systems. 2001, (17): 

107-145. 

[2] J. C. Bezdek.Pattern Recognition with fuzzy objective 

function algorithms. Plenum Press, New York, 1981. 

[3] Cai WL, Chen SC, Zhang DQ. Fast and robust fuzzy cmeans 

clustering algorithms incorporating local 

information for image segmentation. Pattern Recognition, 

2007, 40(3): 825-833. 

[4] Jiayin Kang, Lequan Min, Qingxian Luan. Novel modified 

fuzzy c-means algorithm with applications. Digital Signal 

Processing, 2009, 19: 309-319. 

[5] WEN ZW, LI RJ. Fuzzy c-means clustering algorithm 

based on improved PSO. Application Research of 

Computers, 2010, 27(7): 2520-2522. 

温重伟, 李荣钧. 改进的粒子群优化模糊 C 均值聚类算 

法[J]. 计算机应用研究,2010 ,27 (7): 2520-2522. 

[6] LI LL, LI M, LIU XY. Image segmentation algorithm 

based on particle swarm optimization fuzzy c-means 

clustering. Computer Engineering and Applications, 2009, 

45(31): 158-160. 

李丽丽, 李明, 刘希玉. 基于粒子群模糊 C-均值聚类的 

图像分割算法[J]. 计算机工程与应用,2009 ,45 (31): 

158-160. 

[7] PU PB, WANG G, LIU TA. Research of improved fuzzy 

c-means algorithm based on particle swarm optimization. 

Computer Engineering and Design, 2008, 29(16): 4277- 

4279. 

蒲蓬勃, 王鸽, 刘太安. 基于粒子群优化的模糊 C-均值 

聚类改进算法[J]. 计算机工程与设计,2008 ,29 (16): 

4277-4279. 

[8] YANG GQ, ZHU CM. Particle swarm optimization 

algorithm based fuzzy kernel clustering method. Journal of 

Shanghai Jiao Tong University, 2009, 43(6): 935-939. 


杨广全, 朱昌明. 基于粒子群优化的模糊核聚类方法[J]. 

上海交通大学学报,2009 ,43 (6): 935-939. 

[9] Thomas A. Runkler, Christina Katz. Fuzzy clustering by 

particle swarm optimization [C]. 2006 IEEE International 

Conference on Fuzzy Systems. 2006: 601-608. 

[10] Jiayin Kang, Lequan Min, Qingxian Luan, Xiao Li and 

Jinzhu Liu. Novel modified fuzzy c-means algorithm with 

applications. Digital Signal Processing. 2009, (19): 309- 

319. 

[11] Riccardo Poli, James Kennedy and Tim Blackwell. Particle 

Swarm Optimization. Swarm Intell. 2007, (1): 33-57. 

[12] WANG JW, LI HN. Summary of particle swarm 

optimization algorithm. Modern Computer, 2009, 301: 22- 

27. 

王杰文, 李赫男. 粒子群优化算法综述[J]. 现代计算机, 

2009,301:22-27. 

[13] Gerhard Venter and Jaroslaw Sobieszczanski-Sobieski. 

Particle Swarm Optimization. AIAA JOURNAL. 2003, 

41(8): 1583-1589. 

Qiang Niu born in 1974, doctor, 

associate professor in China University 

of Mining and Technology, received 

B.S. in Northeastern University in 1997, 

received Master in 2004 and PHD in 

2010 in China University of Mining 

and Technology. His main research 

interests are intelligent optimization 

algorithms and data mining. 

Xinjian Huang born in 1986, master 

graduate student, received the B.S. 

degree in China University of Mining 

and Technology in 2010. His main 

research interests are particle swarm 

optimization and its application in time 

series.


Classification of Bio-potential Surface Electrode 

based on FKCM and SVM 

Hao Liu 

School of Textile, Tianjin Polytechnic University, Tianjin, China 

Key Laboratory for Advanced Textile Composite of Ministry of Education, Tianjin Polytechnic University, Tianjin, 

China 

liuhao_0760@yahoo.com.cn 

Xiaoming Tao 

Institute of Textiles and Clothing, Hong Kong Polytechnic University, Hong Kong, China 

tctaoxm@polyu.edu.hk 

Pengjun Xu 

School of Textile, Donghua University, Shanghai, China 

tcxupj@inet.polyu.edu.hk 

Guanxiong Qiu 

Key Laboratory for Advanced Textile Composite of Ministry of Education, Tianjin Polytechnic University, Tianjin, 

300160, China 

qiuguanxiong@tjpu.edu.cn 

Abstract—In this paper, a method which is used for 

evaluating the performance of bio-potential surface 

electrode (BSE) with multi-index is presented. The Fuzzy 

kernel C-means (FKCM) algorithm and KF statistic are 

employed for classifying the BSE samples and searching an 

optimal classification amount respectively. Subsequently, a 

discriminant function is constructed by support vector 

machines (SVM) for recognizing the new measured samples. 

Experimental result shows classification correction ratios of 

improved FKCM algorithm are 96.3% and 85% on the IRIS 

and BSE dataset according a priori knowledge, 

furthermore, the recognition correction ratios of SVM 

algorithm are 96.3% and 90% on the IRIS and BSE dataset. 

Index Terms—FKCM, SVM, classification, recognition, biopotential 

surface electrode 


Bio-potential surface electrode (BSE) is an importance 

unit in some health monitoring devices, especially in 

some wearable bio-potential monitoring garments [1-2]. 

In [3], the measurement is performed in vivo and the biopotential 

is utilized for evaluating the performance of 

BSE. In [4-7], the impedance spectra are utilized for 

evaluating the performance of textile-base BSE on some 

device by more objective methods. However, efficient 

evaluation methods for analyzing the performance of 

Manuscript received December 24, 2010; revised February 1, 2010; 

accepted February 14, 2010. 

Research Grants Council of the Hong Kong SAR Government 

(Grant No. PolyU5277/07E). 


doi:10.4304/jsw.6.5.880-886 

BSE with multi-indexes are absence so far. Improved 

FKCM and SVM which are the algorithms based on 

kernel method have better capability for solving the 

nonlinear problems than FCM and Fisher linear 

discriminant analysis method [8-11]. FKCM which is a 

generalization of the conventional fuzzy C-means 

clustering algorithm (FCM) is presented, and the concrete 

algorithm is shown in [12, 13]. The theory of SVM is 

based on the idea of structural risk minimization (SRM) 

[14-17], and the good generalization ability of SVM is 

obtained by finding a large margin between two classes. 

In this paper, the FKCM and SVM are employed for 

classifying and recognizing the BSE samples with multiindexes, 

however the analysis result can provide help for 

design and improvement of BSE. 

II. EXPERIMENTAL MATERIAL AND MEASUREMENT 

INDEXES 

A. Mearsurement indexes 

Five types of BSE were used in this study. A pair of 

99.99% gold button electrodes (G), 4 types of fabric 

electrodes which are terry or plain coating silver/silver 

chloride (TC or PC) electrode and terry or plain coating 

silver (TS or PS) electrode are selected and each type of 

electrode with 6 pair of specimens are measured. The 

measurement indexes are composed of static open circuit 

potential (OCP) differences (z1 and z2), low frequency 

impedance at 0.01, 1, and 10Hz (z3, z4 and z5), dynamic 

OCP variation (z6). Furthermore, low frequency 

impedance represents the measured resistance in low 

frequency domain, magnitude of static OCP reflects the


symmetry of two electrodes, and static OCP variation 

represents the stable of electrodes in measurement 

process. Dynamic OCP variation represents the noise of 

interface of electrode/electrolyte. However, the less are 

magnitude of these indexes, the better are performances 

of BSE. Experiments are performed at 20 degree 

temperature and 65% relative humidity. 

B. Measurement data 

Data of 26 samples are acquired by using 

electrochemical station which is composed of a computer, 

Electrochemical Interface 1252A (Solartron, UK) and 

Frequency Analyzer Response 1287 (Solartron, UK), a 

code 

G1 

G2 

TC1 

TC2 

TC3 

TC4 

TC5 

TC6 

TS1 

TS2 

TS3 

TS4 

TS5 

TS6 

PC1 

PC2 

PC3 

PC4 

PC5 

PC6 

PS1 

PS2 

PS3 

PS4 

PS5 

PS6 

z1 

(mV) 

TABLE I. 

INITIAL MEASUREMENT DATA OF 26 BSES 

z2 

(mV) 

Measurement Indexes 

z3 

(ohm) 

z4 

(ohm) 

z5 

(ohm) 

series of signal conditioning devices and sensors. 

z6 

(mV) 

1.4 5.4 792320 22521 3235.8 1.34 

2.5 180.7 652490 23365 3788.3 1.002 

0.1 7.6 3526 309.6 125 0.2 

1.2 23.3 1521 283.5 117.7 1.11 

0.5 2.1 1953 235.3 102.5 0.13 

1.2 12.5 871 278.9 111 0.512 

1.2 2.1 613 195.6 77.0 0.483 

0.3 2.5 685 206.2 85.4 0.494 

5.7 28.6 28007 1469.7 342.9 0.71 

0.5 12.0 55407 2707.6 597.4 0.79 

4.3 62.3 44980 2176.8 493.3 0.98 

1.1 4.3 62134 3539.1 805.7 0.829 

0.8 4.0 65408 4015.9 878.5 0.939 

4 67.3 66192 3072.3 671.8 1.059 

0 11.0 5471 626.3 259.7 0.32 

1.5 8.3 6991 679.9 265.3 0.15 

10 161 6701 641.7 228.4 0.41 

0.5 2.6 517 153.8 69.1 0.553 

0.5 3.4 720 275.7 108.6 0.38 

0.2 2.1 619 260.4 99.7 0.204 

0.2 7.3 63082 2923.3 637.8 1.92 

0.3 10.5 131000 9958 2142.7 1.94 

0.1 12.4 73751 3618.5 769.3 2.21 

1.4 0.5 114000 6358.2 1350.7 1.902 

6.0 4.3 132000 6973.8 1492.1 0.976 

0.7 7.2 127000 7307.4 1576.5 2.208 


III. CLASSIFICATION AND RECOGNITION ALGORITHMS OF 

BSE 

A. Data pre-processing 

To enhance the efficiency and accuracy of data 

analysis, it is often necessary to utilize pre-processing on 

the dataset before applying the improved FKCM 

clustering analysis algorithm. The data preprocess 

procedure consist of data normalization, correlative 

analysis of indexes, calculation of index weight, and 

calculation of initial clustering centers. 

Let x be vector that is composed of measurement 

indexes of a sample, Dataset X consists of all samples 

X= { x1, x2, �, 

xn} 

and can be represented as , 

xi = ( xi1, xi2, �, 

xip) 

where , ( 1,2, , ) 

i n = � . Vector x 

p 

is called as pattern of input space Ω also. The set X 

p 

with n patterns is a subset of input space Ω . The 

measured indexes can be represented by 

Z= { z1, z2, � , zp} 

. 

To eliminate effect of indexes’ quantity difference to 

final evaluation, initial measurement data should be 

normalized to compress the data in [0,1]. Data 

normalization function is 

x′ - x′ 

i min 

x i = 

x′ max -x'min 

Where i x and i ′ x denote the normalized vector and 

measurement vector of the ith sample respectively, 

x′ max and x ′ min denote the maximum and minimum of 

measurement vectors in initial data set respectively. 

The redundancy of indexes on data set exists, so 

correlation analysis of indexes is often necessary for 

reducing the dimension of data. 

cov( 

zi, z j) 

ρ ( zi, z j) 

= , i, j = 1,2, � , p (2) 

Dz Dz 

i j 

Where ρ is the correlative coefficient between indexes, 

z i and z j denote the random index in index set, cov is 

the covariance function. 

The contribution of indexes of samples in classification 

is different, and the clustering result and actual 

classification have the better consistent after the index 

weights of samples are introduced in classification. 

However, the methods of index weight acquisition have 

subjective method and objective method. In this paper, 

the objective method is employed for acquiring the index 

weight of all indexes. 

(1)


w 

= 

n n 

∑∑ 

i= 1 k= 

1 

j p n n 

∑∑∑ 

j= 1 i= 1 k= 

1 

| z − z | 

ij kj 

| z − z | 

ij kj 

where wj denotes weight of the jth index, 

n 

zij= xij ∑ x 

i= 

1 ij ( i = 1,2, ⋅⋅⋅ , n; j = 1,2, ⋅⋅⋅ , p) 

To enhance the stability of clustering result, a 

determinate initial clustering center is desired prior to 

performing clustering algorithm. In this paper, the 

relation matrix R is constructed by employing the 

included angle cosine formula (4), and the samples are 

classified by threshold partition method. Though the 

classifying effect is not well, the classification result can 

provide a decided initial clustering center for improved 

FKCM clustering algorithm. 

r 

ij 

= 

p 

∑ 

| x x | 

k = 1 

ik jk 

p p 

2 2 

( ∑x )( ) 

ik ∑x 

jk 

k= 1 k= 

1 

(3) 

(, i j = 1,2, � , n) 

(4) 

where r ij denotes the similarity between the ith and the 

jth sample. 

B. FKCM algorithm 

However, the FCM clustering analysis is somewhat 

limited in real world problems and nonlinear clustering 

analysis would be highly desirable. An efficient method 

of obtaining the nonlinear cluster algorithm is to first map 

p 

the patterns of input space Ω into some higher 

q 

dimensional feature space Ω using a kernel 

function φ() ⋅ , the FCM can be performed in this feature 

space. When the kernel function is chosen, the Euclid 

distance between i x and x j in feature space 

1/2 

is dˆ ij( xi, x j) = [ K( xi, xi) − 2 K( xi, x j) + K( 

xj, x j)] 

, 

i, j = 1,2, � , n. 

Let V be clustering center matrix in input space, 

V= ( v , v , , v ) v = ( v , v , �, v ),( i = 1,2, � , c) 

. 

1 2 � c i i1 i2 ip 

Let Û be membership matrix in feature, 

Uˆ= ( uˆ ˆ ˆ 

1, u2, � , un) 

, 

uˆ ˆ ˆ ˆ 

i = ( ui1, ui2, �, uic),( i = 1,2, � , n) 

. Hence, Objective 

function of FKCM clustering algorithm in feature space 

is 

m( ; , 

c n 

) = ∑∑ 

j= 1 i= 

1 

m ˆ 2 

ji ij 

Jˆ XUD ˆ ˆ uˆd , 2 < c < n (5) 

New clustering center vectors in feature space are 


j 

j 

n 

ukj m 

i 

n 

ukj 

m 

k= 1 i= 

1 

vˆ φ( v ) ( ˆ ) φ( 

x )/ ( ˆ ) , j = 1, 2, � , c (6) 

= =∑ ∑ 

The φ( xi ) is dropped in (6). Unfortunately, the mapping 

function φ() ⋅ may not be known explicitly and if the 

dimension of the feature space 

q 

Ω is very high or 

infinite, it is difficult to solve for objective function by 

(6). To get around this difficulty, the problem is 

reformulated to involve only the dot product of the 

patterns x i ( i= 1, 2, � , n) 

in the feature space. 

K 

n 

m 

∑( 

uˆkj ) K( 

xk, xi) 

( , ˆ i j) = φ( i) ⋅ φ ( k = 1 

j) = n 

m 

∑ ( uˆ 

kj ) 

k = 1 

n n 

∑∑ uˆkj m 

uˆtj m 

K xkxt j j = φ j ⋅ φ j 

k= 1 t= 

1 = 

n 

2 

⎛ m ⎞ 

⎜∑( uˆ 

kj ) ⎟ 

k = 1 

x v x v (7) 

K ( vˆ , vˆ ) ( v ) ( v ) 

uˆ 

= 

= 

ij c 

j= 

1 

(1 / dˆ 

( x , vˆ 

)) 

j= 

1 

2 1/( m−1) 

ij i j 

(1 / dˆ 

( x , vˆ 

)) 

2 1/( m−1) 

ij i j 

( ) ( ) ( , ) 

⎝ ⎠ 

(1 / ( K( x , x ) − 2 K( x , vˆ ) + K( 

vˆ , vˆ 

))) 

c 

∑ 

∑ 

i i i j j j 

(1 / ( K( x , x ) − 2 K( x , vˆ ) + K( 

vˆ , vˆ 

))) 

i i i j j j 

1/( m−1) 

1/( m−1) 

When ˆ 2 

d ( , ˆ ij xiv j ) = 0, 

uˆ 1, ˆ ij = uit = 0, ( t∈[1, j) ∪ ( j, c]) 

. 

ˆ 1/2 

d ( , ˆ ) [ ( , ) 2 ( , ˆ ) ( ˆ , ˆ 

ij xiv j = K xixi− K xiv j + K v j v j )] , 

i = 1, 2, �, n; j = 1, 2, �, 

c 

(10) 

To acquire the optimization membership degree matrix 

* 

Û and corresponding distance matrix * ˆD , the equation of 

ˆ() l ˆ( 

l 1) 

| J J | 

− 

− must be convergent, that is, equation of 

lim Jˆ 

( ; ˆ, ˆ 

m XUD) comes into existence [18]. Hence, the 

l→∞ 

variance ε can be set at a random small value, the initial 

distance matrix and the initial membership matrix are 

known, then the iterative algorithm can be performed by 

() ( 1) 

(5), (7), (8), (9), and (10), if ˆ l 

| ˆ l− 

J − J | < ε , the 

iterative algorithm ceases, and the constraint optimization 

membership matrix () ˆ l 

U and the constraint optimization 

() 

distance matrix ˆ l 

D can be acquired, finally, samples is 

classified in terms of maximum membership principle. 

When dot product operation is performed between pattern 

and indexes weight vector in kernel function of the 

FKCM clustering algorithm, the improved FKCM 

clustering algorithm can be obtained. Furthermore, the 

statistic F in (11) can be utilized for acquiring the optimal 

classification amount. Statistic F is a conventional index 

for evaluating the clustering validity and the nonlinear 

factors of data set are not considered, hence the KF index 

is constructed by using kernel function and shown in (12) 

for evaluating more efficiently the clustering validity . 

(8) 

(9)


( n−c) ni 

|| xi −x|| 

SSA ( c −1) 

i= 

1 

F = = 

c ni 

SSE ( n − c) 

( c −1) || x −x 

|| 

c 

∑ 

∑∑ 

i= 1 t= 

1 

it i 

c 

2 

( n−c) ˆ ∑ nd i ( xi, x) 

i= 

1 

c− c ni 

2 

∑∑d 

xit xi 

i= 1 t= 

1 

SSA ( c −1) 

KF = = 

SSE ( n − c) 

( 1) ˆ ( , ) 

2 

2 

(11) 

(12) 

Where: KF is statistic of samples vectors in feature 

space, SSA is between class variance, SSE is inner-class 

variance, xi is the mean vector of the ith class samples 

vectors, x is the mean vector of the whole samples 

vectors, it x is the ith class and the tth sample vector, i n 

is amount of samples of the ith class. 

C. SVM Algorithm 

According to the above classification result, we can 

acquire a training set S = {( x1, y1),( x2, y2), � ,( xN, 

yN)} 

p 

of N data points, where i ∈Ω x is the ith input pattern 

and yi ∈ R is the ith output pattern. In most cases, the 

searching of a suitable hyperplane in an input space is too 

restrictive to be of practical use. Hence, suppose 

m 

{ ϕ j ( x)} j= 

1 represents the nonlinear transfer set from 

the input space to feature space, where m is the dimension 

of feature space. Hence, a decision hyperplane in feature 

can be defined as 

y i 

T 

i[ w ϕ ( x ) + b] 

≥ 1, 

i 1, �, 

N, 

= (13) 

Where ϕ (⋅) 

is a kernel function which maps the input 

space into a higher dimensional space, m and n are, 

respectively, the dimensions of the input space and 

feature space. However, this function is not explicitly 

constructed. In order to have the possibility to violate 

(15), in case a separating hyperplane in this higher 

dimensional space does not exist, slack variables ξ k are 

introduced such that 

T ⎧y 

i[ 

w ϕ( 

xi 

) + b] 

≥ 1− 

ξi 

, i = 1, 

�, 

N 

⎨ 

⎩ξi 

≥ 0, 

i = 1, 

�, 

N 

(14) 

Subsequently, according to the structural risk 

minimization principle, the risk bound is minimized by 

considering the optimization problem 

1 

w 

T 

Minimize ⋅ w + C∑ 

ξ i (15) 

2 

= 1 

subject to (14). Where C is a constant and can be 

regarded as a regularization parameter. Tuning this 

parameter can obtain a balance between margin 

maximization and classification violation. In order to 


i 

N 

solve the constraint optimal problem, one constructs the 

Lagrangian and transformed into the dual 

⎧Maximize 

⎪ 

N N 

⎪ 

1 

i= 1, �, 

N 

⎪W( 

α) = ∑αi − ∑αα 

i iyiyK i ( xi, xj) 

⎨ 

i= 1 2 i, j= 

1 

⎪ 

N 

⎪ Subject to ∑ yiαi = 0, 0 ≤αi ≤ C, i = 1, �, 

N 

⎪⎩ 

i= 

1 

(16) 

Searching the optimal hyperplane in (15) is a quadratic 

programming (QP) problem, according to the Kuhn- 

Tucker theorem, the solution of the optimal problem must 

satisfies the equality 

⎧αi( 

yi( xw i i + b) 

− 1 + ξi) 

= 0 

⎨ 

⎩( 

C − αi) ξi 

= 0 

, i = 1, �, 

N (17) 

In (12), α i is zero for most of samples, when the non- 

zero values α i are satisfied with the equality sign in 

(14), the pattern x i corresponding with α i is called 

support vector. 

* 

If α is the optimal solution in (16), then 

w 

* 

= 

N 

* 

∑ α i yi 

xi 

(18) 

i= 

1 

That is the weight coefficient vector of optimal 

classification hyperplane is linear combination of training 

pattern vectors. After solving the above problems, the 

optimal classification function can be acquired as 

N 

⎛ * 

* ⎞ 

f ( x) 

= sign⎜∑ 

yiα 

i K( 

xi 

⋅ x) 

+ b ⎟ (19) 

⎝ i= 

1 

⎠ 

IV. RESULET AND DISCUSSION 

In order to evaluate the performance of BSE and verify 

effect of these algorithms, the improved FKCM and SVM 

were performed on IRIS dataset and measurement dataset 

respectively. Matlab7.01 software was utilized for data 

processing and analyzing. 

A. Pre-processing of measurement dataset 

After data normalization, procedures of correlation 

analysis, weight computation, and initial clustering center 

were performed on IRIS dataset and measurement dataset 

in turn. Correlation coefficients of IRIS dataset were 

shown in Tab. II and Tab. III by using (2). However, 

some indexes with strong correlation can be eliminated 

by Pearson Correlation coefficient analysis method. 

Because of strong correlation in indexes, 4 z , 5 z and 4 z′ can 

be deleted in data analysis process. Furthermore, index 

weight coefficient of measurement dataset and IRIS 

dataset was, respectively, WM = (0.2399 0.2858 0.2967 

0.1776) and WIRIS = (0.2374 0.1785 0.2785 0.30) by 

using (3). 

FKCM and SVM Performed on IRIS Dataset


TABLE II. 

PEARSON CORRELATION COEFFICIENT OF ALL MEASURED INDEXES 

IN IRIS DATASET 

TABLE IV. 

COMPARISON OF WRONG CLASSIFICATION AMOUNT OF SAMPLES BY 

USING THREE CLUSTERING METHODS 

Clustering Amount in 

Algorithm 1 st Amount in 

Class 2 nd Amount in 

Class 3 rd Total Correction 

Class amount Ratio (%) 

FCM 0 12 4 16 89.33 

FKCM 0 9 3 12 92.00 

Improved 

FKCM 

code 

z′ 1 

2 z′ 

0 4 3 7 95.33 

TABLE III. 

STATISTIC F AND KF OF IRIS DATASET USING IMPROVED FKCM 

Preset Class Amount 2 3 4 5 6 7 

Actual Class Amount 2 3 4 5 6 7 

F validity index 339 342 284 233 197 192 

KF validity index 430 485 409 346 289 273 

IRIS dataset contains 150 examples with 4 dimensions 

and 3 classes. One class is linearly separable from the two 

other; the latter are not linearly separable from each 

other. Iterative accurate degree and class number are, 

respectively, 10e-5 and 3, and the kernel function of 

improved FKCM clustering algorithm chooses 

p 

2 

k( xy , ) = exp( ∑ ( w *( )) ) 

i 1 i xi − yi σ 

= 

, 

whereσ 

= 0.8 . 

Tab. IV showed the F and KF value were all maximal 

when samples in IRIS were classified into three classes, 

and the result was consistent with the actual classification 

of samples. Furthermore, wrong classification amount of 

samples and classification correction ratios using FCM, 

FKCM, improved FKCM in IRIS dataset were shown in 

Tab. V, samples in the 1st class was classified correctly 

which shows three clustering algorithms can solve 

efficiently the linear classification problems, however, 

the improved FKCM is obviously better than FCM and 

KFCM for solving the nonlinear problems. 

In order to compare the performance of four 

discriminant methods: the SVM, Kernel Fisher 

Discriminant Analysis (KFDA), back promulgation 

neural network (BPNN), and radial basis function neural 

z′ 3 

z′ 4 

z′ 1 1 -0.12 0.87 0.82 

z′ 2 -0.12 1 -0.42 0.37 

z′ 3 0.87 -0.42 1 0.96 

z′ 4 0.82 -0.37 0.96 1 


TABLE V. 

PEARSON CORRELATION COEFFICIENT OF ALL MEASURED INDEXES IN 

MEASUREMENT DATASET 

Index 

code 

z 1 

2 z 

network (RBFNN), We performed these algorithms on 

IRIS dataset, the training samples were utilized for 

training the networks, then employing the trained 

networks to recognize the testing samples. The 

parameters of BPNN were set as follow: the neurons of 

input layer were the amount of the training samples, the 

neurons of hidden layer were obtained according the 

equation n 1 = i + o + a where 1 n , i and o were the 

neurons amount of hidden layer, input layer and output 

layer respectively, parameter a were taken in the integer 

domain [1,10]. The transfer function of hidden layer was 

set for ”tansig”, the transfer function of output layer was 

set for ”purelin”, training algorithm selected “Levenberg- 

Marquardt” method, the output tolerance was 0.05 , 

training error was set for 0.001,the transfer coefficient 

of RBFNN was 1.0. In KFDA algorithm, Gauss kernel 

2 

function 

k( 

x, y) 

= exp( − || x − y || / σ ) 

was selected 

and parameterσ was 0.7. In SVM algorithm, the kernel 

function was ‘gaussian’ and parameter of ‘kerneloption’ 

was 2, and bound on the lagrangian multipliers was 1000. 

Then above four methods were performed on IRIS 

dataset, and the effects of every method were evaluated 

by the correction recognition ratios to testing samples. 

The four fifth of each class of IRIS were using as 

training samples, and the other samples were using as 

testing samples, the discriminant functions were 

constructed by above described four methods. And 

experimental results showed the correction recognition 

ratios of BPNN, RBFNN, KFDA, and SVM methods in 

Tab. VI. Obviously, the effects of SVM method and 

KFDA were better than other two methods. Therefore, 

the SVM method was chosen for constructing the 

discriminant function of measurement dataset. 

B. SVM and FKCM Performed on Measured Dataset 

26 samples were measured on electrochemical 

station and 6 indexes are obtained. G, TS and PS are 

polarization electrodes but TC and PC are nonpolarization 

electrodes, the priori knowledge showed the 

non-polarization electrodes had the better performance 

than polarization electrodes [19]. Improved KFCM was 

performed on measurement dataset. Tab. VII showed KF 

value was maximal when samples were classified into 

z 3 

z 4 

5 z 

z 1 0.63 0.03 0.03 0.04 -0.12 

1 

z 2 0.63 1 0.35 0.36 0.35 -0.05 

z 3 0.03 0.35 1 0.97 0.92 0.31 

z 4 0.03 0.36 0.97 1 0.99 0.43 

z 5 0.04 0.35 0.92 0.99 1 0.51 

z 6 -0.12 -0.05 0.31 0.43 0.51 1 

z 6


TABLE VI. 

COMPARISON OF CORRECTION RECOGNITION RATIOS USING ABOVE 

FOUR METHODS ON IRIS DATASET 

Discriminant Methods BPNN RBFNN KFDA SVM 

Correction Recognition 

Ratio (%) 

two classes. Obviously, two classes was best 

classification amount, and the class amount of 

classification result was consistent with the priori 

knowledge. 

Tab. VIII showed the first class has 14 samples and 

the second class has 12 samples by using improved 

FKCM algorithm. The four samples of TS2, TS4, TS5 

and PC3 are disagreement with the priori knowledge, and 

approximate 85% samples agree with the priori 

knowledge. Subsequently, the front eight of each class 

measured samples were selected as training samples, and 

the other measured samples were selected as testing 

samples. According to the previous describe, SVM was 

performed on training samples and acquires a 

discriminant function and some support vectors, and 90% 

recognition correction ratio can be acquired by applying 

discriminant function on the testing samples. 

TABLE VII. 

STATISTIC KF OF MEASUREMENT DATASET USING IMPROVED FKCM 

Preset Class Amount 2 3 4 

Actual Class Amount 2 3 - 

KF validity index 15 10 - 

TABLE VIII. 

CLASSIFICATION NUMBER AND QUALITY GRADE OF SAMPLES 

code Membership 

degree 

Number 

of class 

Membership Number 

grade code 

degree of class grade 

G1 0.53 2 2 G2 0.54 2 2 

TC1 0.69 1 1 PC1 0.71 1 1 

TC2 0.65 1 1 PC2 0.69 1 1 

TC3 0.68 1 1 PC3* 0.54 2 2 

TC4 0.77 1 1 PC4 0.75 1 1 

TC5 0.74 1 1 PC5 0.72 1 1 

TC6 0.73 1 1 PC6 0.69 1 1 

TS1 0.54 2 2 PS1 0.51 2 2 

TS2* 0.82 1 1 PS2 0.54 2 2 

TS3 0.62 2 2 PS3 0.53 2 2 

TS4* 0.82 1 1 PS4 0.55 2 2 

TS5* 0.77 1 1 PS5 0.55 2 2 

TS6 0.63 2 2 PS6 0.55 2 2 

a.* express class number of the sample is different according the priori knowledge and improved 

FKCM algorithm 


93.33 86.67 96.67 96.67 

V. CONCLUSION 

In this paper, a generalization evaluation method is 

proposed and employed in quality evaluation of BSE. By 

classifying the IRIS dataset using FCM, FKCM and 

improved FKCM clustering algorithm, experimental 

result shows improved FKCM is more efficient than other 

two algorithms for solving the nonlinear problem. The 

constructed KF statistic can help find the optimal 

classification amount of measured samples, and the 

classification result is consistent with a priori knowledge. 

Subsequently, we construct the discriminant function for 

recognizing the new measured samples by SVM 

algorithm, and a well recognition effect can be acquired. 

In all, the improved FKCM and SVM algorithms can 

compose a complete evaluation method for the 

performance of BSE. 


The authors would like to thank Research Grants Council 

of the Hong Kong SAR Government (Grant No. 

PolyU5277/07E). This work was supported in part by a 

grant from The Hong Kong Polytechnic University. 

REFERENCES 

[1] Jerald Yoo, Long Yan, Seulki Lee, Hyejung Kim, and Hoi- 

Jun Yoo, “A wearable ECG acquisition system with 

compact planar-Fashionable circuit board-based shirt,” 

IEEE Trans. Inform. Tech. Biomed., vol. 13, pp. 897–902, 


[2] Tae-Ho Kang, Carey R. Merritt, Edward Grant, Behnam 

Pourdeyhimi, and H. Troy Nagle, “Nonwoven fabric active 

electrodes for biopotential measurement during normal 

daily activity,” IEEE Trans. Biomed. Eng., vol. 55, pp. 

188–195, January 2008. 

[3] Enzo Pasquale Scilingo, Gemignani A, Paradiso R, Taccini 

N, Ghelarducci B, and De Rossi D, “Performance 

evaluation of sensing fabrics for monitoring physiological 

and biomechanical variables,” IEEE Trans. Inform. Tech. 

Biomed., vol. 9, pp. 345–352, September 2005. 

[4] Rattfalt L, Linden M, Hult P, Berglin L and Ask P, 

“Electrical characteristics of conductive yarns and textile 

electrodes for medical applications,” Med. Biol. Eng. 

Comput., vol. 12, pp. 1251–1257, December 2007. 

[5] Westbroek P, Priniotakis G, Palovuori E, De Clerck K, 

Van Langenhove L and Kiekens P, “Quality control of 

textile electrodes by electrochemical impedance 

spectroscopy,” Text. Res. J., vol. 2, pp. 152–159, February 

2006. 

[6] Priniotakis G., Westbroek P., Van Langenhove L. and 

Kiekens P., “An experimental simulation of human body 

behaviour during sweat production measured at textile 

electrodes,” Int. J. Cloth. Sci. Technol., vol. 3–4, pp. 232- 

241, October 2005. 

[7] L Beckmann, C Neuhaus, G Medrano, N Jungbecker, M 

Walter, T Gries, et, al., “Characterization of textile 

electrodes and conductors using standardized measurement 

setups,” Physiol. Meas., vol. 31, pp. 233–247, February 

2010. 

[8] GROLAMI M., “Mercer kernel based clustering in feature 

space,” IEEE Trans. Neural Netw., vol.13, pp.780–784, 

May 2002. 

[9] Bernhard Scholkopf, Sebastian Mika, Chris J. C. Burges, 

Philipp Knirsch, Klaus-Robert M¨uller, Gunnar Ratsch, et,


al., “Input space versus feature space in kernel-based 

methods,” IEEE Trans. Neural Netw., vol.10, pp. 1000– 

1017, September 1999. 

[10] Chua, K.S., “Efficient computations for large least square 

support vector machine classifiers,” Pattern Recogn. Lett., 

vol. 24, pp. 75–80, January 2003. 

[11] Vapnik, V., Statistical Learning Theory, Wiley, New York, 

1998. 

[12] ZHANG Li, ZHOU Weida and JIAO Licheng, “Kernel 

clustering algorithm,” Chinese Journal of Computers, vol. 

25, pp. 587–590, June 2002. 

[13] ZHAO Yugao and PENG Hong, “An improved fuzzy 

kernel clustering algorithm,” Journal of Xihua University: 

Natural Science Edition, vol. 26, pp. 48–50,69, March 

2007. 

[14] Shawe-Taylor, J. and Bartlett, P.L., “structural risk 

minimization over data-dependent hierarchies,” IEEE 

Trans. Inform. Theor., vol. 44, pp. 1926–1940, September 

1998. 

[15] Chih-Wei Hsu and Chih-Jen Lin, “A comparison of 

methods for multi-class support vector machines,” IEEE 

Trans. Neural Netw., vol. 13, pp. 415–425, March 2002. 

[16] Qin Yuping, Wang Xiukun, Li Xiangna and Wang Chunli, 

“Study on multi-class text classification algorithm based 

on hyper-sphere support vector machines,” Computer 

Engineering and Applications, vol.44, pp. 166–168, July 

2008. 

[17] K. Crammer and Y. Singer, “On the learnability and design 

of output codes for multiclass problems,” Mach. Learn., 

vol. 47, pp. 201–233, February/March 2002. 

[18] BEZDEK J C., “Convergence theory for fuzzy c-Means: 

counterexamples and repaires,” IEEE Trans on SMC, vol. 

17, pp. 873–877, September/October 1987. 

[19] John G. Webster, Medical instrumentation : application 

and design, 3 rd ed., New York : Wiley, 1998, pp. 183-232. 


Hao Liu received the B.Eng. and M.Eng. degrees in textile 

material and design from Tianjin Polytechnic University, 

Tianjin, China, in 2001 and 2004. He is currently working 

towards the PhD degree in textile engineering at Tianjin 

Polytechnic University, Tianjin, China. His research interests 

are data analysis and image processing in textile and clothing 

Engineering, smart textiles and apparel and wearable electronics. 

Xiaoming Tao is Chair Professor and Head of Institute of 

Textiles and Clothing. Prof. Tao graduated with a B.Eng. in 

Textile Engineering and a first class prize for undergraduate 

students from East China Institute of Textile Science and 

Technology in 1982. She gained her PhD in Textile Physics 

from University of New South Wales, Australia in 1987.Her 

research interests are fibrous materials, textile Science and 

technology, smart textiles and apparel and wearable electronics 

and photonics. 

Pengjun Xu is currently working towards the PhD degree in 

textile engineering at Donghua University, Shanghai, China. His 

research interests are fibrous materials, smart textiles and 

apparel. 

Guanxiong Qiu is professor of Tianjin Polytechnic University. 

His research interests are textile material and knitted 

engineering.


Consonant Recognition of Dysarthria Based on 

Wavelet Transform and Fuzzy Support Vector 

Machines 

Zhuo-ming Chen 1 

1 The First Affiliated Hospital, Jinan University, Guangzhou, China 

Email: tchzm@21cn.com 

Wei-xin Ling 2 and Jian-hui Zhao 3, * 

School of Science, South China University of Technology, Guangzhou, China 

Email: 2 lingweixin@21cn.com, 3 awordman@163.com 

Tao-tao Yao 4 

4 The First Affiliated Hospital, Jinan University, Guangzhou, China 

Email: selena.567.com@163.com 

Abstract—Consonant(in Chinese) recognition had important 

clinical significance in the assessment of dysarthria, while 

the consonants were so short and unstable that the 

recognition results of traditional methods were ineffective. 

The algorithm described in this paper extracted a new 

feature(DWTMFC-CT) of the consonants employing 

wavelet transformation, and the difference of similar 

consonants can be described more accurately by the feature. 

Then the algorithm classified consonants using multi-class 

fuzzy support vector machines(FSVM). In order to reduce 

the computation complexity caused by using the standard 

fuzzy support vector machines for multi-class classification, 

this paper proposed a algorithm based on two stages. 

Experimental results shown that the proposed algorithm 

could get better classification results while reducing the 

training time greatly. 

Index Terms—wavelets transform, fuzzy theory, support 

vector machines, consonant recognition 


There are a large number of pronunciation-impaired 

patients in China. It’s very important to assess the patient 

with dysarthria accurately. The methods of traditional 

dysarthria assessment, including Franchy Dysarthria 

Assessment and the dysarthria assessment method made 

by Zhongkang, with great subjectivity, often lead to 

diagnose inaccurately and incorrectly. Speech analysis is 

an effective assessment tool. With non-invasive and 

objective, this method can test a large number of patients 

in a short time. Speech analysis is often based on some 

long vowels [1-2], but practice shows that the clinical 

significance of consonant is more important than the long 

vowels’, and there is almost no objective assessment in 

dysarthria now. Automatic and accurate identification of 

the 21 categories of consonant in the Mandarin Language 

*Corresponding author 


doi:10.4304/jsw.6.5.887-893 

is the key to objective assessment of consonant. 

Because of consonant’s instability, strong dynamic and 

short duration, so to identify them is difficult. There are 

two ways to improve the correct recognition rate: 1) To 

extract better feature parameters of consonant; 2) To 

choice a suitable recognition method. Feature parameters 

of speech are used widely, including Linear Prediction 

Cepstrum Coefficient (LPCC) and Mel Cepstrum 

Coefficient (MFCC), etc. They are assumed that speech 

signal is short-time stationarity, but the consonants are 

very unstable signal, so these parameter models have 

poor effect in consonant recognition [3]. 

It’s a good choice to extract the consonant features by 

employing wavelet transformation, because wavelet 

transformation has good localize characteristics of time 

domain and frequency domain. The time-frequency 

window can adjust according to the signal’s shape and 

multi-resolution analysis, so it can describe non-stable 

signal more precisely. 

HMM technology [4], Gaussian mixture model [5] and 

neural network [6] are used widely in the field of speech 

recognition, but all of them have some defects, which are 

hard to make up. HMM is poor in classification decisionmaking, 

and need to priori statistical knowledge first; 

Gaussian mixture model is also based on statistical 

theory, and need to a large number of training samples to 

get good recognition effect; neural network’s problems 

are hard to determine the network structure, local 

optimization and easy to over learning. Support vector 

machine is the important theory based on VCdimensional 

theory of statistical learning theory and 

structural risk minimizes principle. It seeks the best 

compromise between the Modal complexity and learning 

ability to obtain the best extension according to the 

limited sample information. It can solve small sample, 

nonlinear, high dimension and local minimal problems 

[7].


This research combines the advantages of wavelet 

transformation and support vector machine, and puts 

forward a new two-stage consonant recognition method. 

The experiments show that this method can make higher 

accuracy for consonant identification with small 

samples.. 

II. FUZZY SUPPORT VECTOR MACHINE 

A. Two-class fuzzy support vector machine 

Standard support vector machine (SVM) is a twoclass 

classifier, if there are k training samples{(xi, yi), 

i=1,2,…,k}. in which xi∈R N , yi∈{+1,-1} is the 

corresponding class label. If the training samples are 

linearly separable, finding separating hyper plane w • xi, 

+ b = 0 to make each sample has: yi [w • xi, + b] -1 ≥ 0. 

Solving the optimal hyper plane is equivalent to 

make‖w‖ 2 /2 minimize, that is solving the solution of 

optimal problem: 

k 

⎧ 1 2 

⎪min 

w + C∑ξi 

⎨ 2 

i= 

1 

(1) 

⎪ 

⎩st 

.. yi( w⋅ xi + b) ≥1 − ξi, 

i= 1, �, 

k 

In which: C is penalty factor, which controls the penalty 

degree of misclassification; ξi is slack variable, which 

compensates some samples that cannot be correctly 

classified by hyper plane. 

The principle between FSVM and standard SVM is 

similar, the difference is that FSVM weights 

classification error ξi caused by each input point through 

the fuzzy factor q i , qiξ i means each input point 

corresponding classification error. The smaller fuzzy 

factor i q , leading to smaller classification error qiξ i , 

which reduces the importance of the wrongly classified 

samples [8]. The objective function as follows: 

k 

⎧ 1 2 

⎪min 

w + C∑qiξi ⎨ 2 

i= 

1 

(2) 

⎪ 

⎩st 

.. yi( w⋅ xi + b) ≥1 − ξi, 

i= 1, �, 

k 

To solve the above objective function, it can be 

transformed into the corresponding dual form. we can get 

the two classification decisions function: 

k 

f ( x) = sign( ∑ aiyixix+ b) 

(3) 

i= 

1 

It transformed the sample points by nonlinear and 

mapped to high dimensional feature space for the 

nonlinear problems. It could be realized classification by 

linear classifier in high-dimension. By the method of 

introducing the kernel function, it needn’t to know the 

exact mapping function, and could calculate the inner 

product between the samples. The decision function was: 

k 

f ( x) = sign( ∑ aiyK i ( xi⋅ x) + b) 

(4) 

i= 

1 

The kernel function mainly used in this research was dorder 

polynomial kernel function: 


( , ) ( 1) d 

K xi xj = xi⋅ xj 

+ (5) 

The choice of penalty factor C and order d of the 

kernel function can impact the FSVM greatly. There are 

mainly two methods, including the experience 

deterministic method and the grid-search method, applied 

in the current practice. The grid-search method was used 

in this paper, setting the C value space is {1, 10, 100, 

1000, 10000}and d is {1, 1.5, 2, 2.5, 3}, to make the 

space of C and d divided into grids, which trialed one by 

one to determine the optimal parameters in each grid 

point. 

B. multi-class fuzzy support vector machines 

FSVM algorithm was originally designed for twoclass 

classification problems. It needs to be extended to 

multi-class classifier when dealing with multi-class 

problems. There are two ways used widely at present: 

one-to-many and one-to-one. The method of one-to-many 

makes some class samples classified as one-class in turn 

when it trains, remaining samples classified as the other 

one, so k-class samples are constructed to k FSVM. 

Unknown samples are classified to the class that has the 

largest classification function value. The practice of oneto-one 

method is to design a FSVM between any twoclass 

samples, so k-class samples need to design k (k-1) / 

2 FSVM. When classifying an unknown sample, the 

category of the sample is the class that gets most votes by 

last. 

Experiment showed that the one-to-one method could 

get better classification effect than the one-to-many 

method [9], however, its time complexity is O (k2). The 

performance of the algorithm would drop dramatically 

with the number of categories increases. These FSVM got 

by one-to-one training method, organized into a directed 

graph (DirectedGraph, DG) structure with unique root 

node, then got the fuzzy support vector machine 

(FDGSVM), shown in Fig. 1. For the k classification 

problem, FDGSVM required only k-1 dimensions [10], 

which reduced the time complexity effectively. 

3 

4 

2 

3 

4 

1 

2 

3 

4 

1,4 

2,4 1,3 

1 

2 

3 

2 

3,4 3 2,3 1,2 

4 3 2 1 

Figure 1. The flow chart of FDGSVM (K=4). 

The flow of FDGSVM was as follows: A list was 

formed by all samples. A test sample was input from the 

root. First, it was determine that whether it belonged to 

the first category or the last category in the list. After 

1 

2


removing the category which was not selected, went to 

the next level node in DG. The new list was made the 

same treatment until the node was leaf node. 

III. WAVELET TRANSFORMATION 

2 

Ψ() t ∈ L ( R) 

, the Fourier transform ˆ ( ω) 

2 

If 

Ψ meets 

the permit conditions C = Ψ ˆ ( ω) ω dω 

=1) layer, then it 

expands as following: 

j−1j j 

A2 [ f( t)] = D2 [ f( t)] + A2 [ f( t)] 

(8) 

In which, 2 j 

D is the detail coefficient, representing 

the high-frequency component of the j layer, 2 j 

A is the 

approximate coefficient, representing the low-frequency 

component of the j layer, when j = 1, 

0 

f () t = A2[ f()] t . 

This research chose db4 wavelet as the mother 

wavelet, because the db wavelet is the compact 

orthogonal wavelets, and has a good expansibility. So it 

could weigh the border problem that brought by 

increasing the support set length flexibly (to increase the 

concentrative degree of energy) [11]. 

IV. ALGORITHM DESCRIPTION 

A. The two-stage recognition algorithm 

To reduce the algorithm’s time complexity and ensure 

its accuracy, the whole algorithm was divided into two 

stages, shown in Fig. 2. 

The first stage is the rough classification stage of 

consonant. Among a numerous acoustic parameters of 

consonants, we extracted features such as: length, 

periodicity, relative energy and zero-crossing rate for the 

consonant classification [12]. Using FDGSVM, the 21 

consonants were divided into 7 rough categories: 

C1(b,d,g), C2(1,m,n,r), C3(z,zh,j), C4(p,t,k), C5(c,ch,q), 

C6(f,h) and C7(s,sh,x). The second stage is the fine 

classification stage of consonants. Because of high 

similarity between the different consonants in the same 

rough category, it could be described more accurately by 


∫ 

R 

using the wavelet transformation to extract distinguishing 

features of consonants. And more detailed delineations of 

7 categories were done by using FDGSVM again (using a 

separate fine classifier for each rough category). Finally 

the purpose of identifying each consonant was achieved. 

Rough stage 

Fine stage 

C1 

FDGSVM1 

b d g 

21 consonants 

FDGSVM0 

…… 

…… 

RC 

Figure 2. The two-stage recognition algorithm flow (RC=rough 

classifier, FCG= fine classifier group).. 

FCG 

The detail steps of the algorithm were as follows: 

Step 1 Set the training sample set was S={S1, S2, …, 

S7}, in which Si(i=1,…,7) was sample set of the ith rough 

category. The training samples were extracted 

consonant’s length(L), periodicity(P), relative energy(E) 

and zero-crossing rate (including the mean zero-crossing 

rate(MZ), the last zero-crossing rate(LZ) and phonetic 

rhyme transition zero-crossing rate(TZ)) to form the 

characteristic parameters(F = (L, P, E, MZ, LZ, TZ)) of 

rough category. 

Step 2 Set a sample point SPi,j∈Si (j=1,…, Ni, Ni is 

the number of samples in Si). Rough category feature 

with different units and orders, were normalized first, and 

then calculated the average of each feature, composed the 

center of feature vector F of Si: 

Fi= ( LPEMZLZTZ 

i, i, i, i, i, i) 

(in which Li = ∑ Li, j Ni 

, 

other features were treated similarly). 

The calculation method of each sample fuzzy factor 

qi,j was: qi, j= 1−di, j max { di,1 , � , di, 

Ni} 

. In which, 

di,j I is Euclidean distance between the sample point SPi,j 

feature vector and the feature vector center F i . 

Step 3 The algorithm trained 21 FSVM by using the 

training sample S, then organized these trained FSVM 

into a fuzzy directed graph support vector machines: 

FDGSVM0 for rough classification. 

Step 4 New consonant feature vectors (DWTMFC- 

CT) were extracted based on discrete wavelet transform 

(see section B), as the fine feature vectors in the second 

phase. 

Step 5 Based on parameters DWTMFC-CT, similar 

to the method of step 2, the algorithm then calculated the 

fuzzy factor for each sample (7 rough classification 

sample sets were calculated independently), The 

C7 

FDGSVM7 

s sh x


algorithm trained the Ki*(Ki-1)/2 FSVM used for 

subdividing rough category Ci by using the training 

sample Si, (Ki is the number of class of consonants in Ci), 

then organized them into fine classifier group: FDGSVMi 

(i = 1, ..., 7). 

Step 6 The algorithm extracted test samples feature 

parameters F and DWTMFC-CT, and input feature 

parameter F to the rough classifier FDGSVM0, the most 

appropriate fine classifier would be selected from the fine 

classifier group, and then determined test samples belong 

to different categories according the parameters 

DWTMFC-CT. 

B. feature parameters DWTMFC-CT extraction 

Some speech features had be extracted using wavelet 

transform by some researchers, such as MWBC [13], 

DWT-MFC [14]. The method of DWT-MFC was 

improved appropriately in this paper as it could adapt the 

features of consonant recognition. The extraction method 

of the new consonant feature vector DWTMFC-CT was 

as follows: 

Step 1 Enflame: According to the different length of 

consonant, the consonant signal was split up into several 

frames on average, as shown in TABLE I. 

Step 2 Wavelet transformation: Using db4 as mother 

wavelet, the consonant signals were decomposed to 3 

TABLE I. 

THE NUMBER OF FRAMES 

rough category C1 C2 C3 C4 C5 C6 C7 

number 1 2 2 3 3 4 4 

layers, then 3 groups of detail coefficients: 

1 

D 2 , 

3 

D 2 and a group of similar coefficients: 3 

2 

2 

D 2 , 

A were 

extracted. 

Step 3 Spectrum combined: The 4 groups of wavelet 

coefficients were done by Fast Fourier Transform (FFT), 

the wavelet coefficients spectrum were translated from 

the time domain into the frequency domain, then all of 

the wavelet coefficients will combined into a full 

spectrum. 

Step 4 Calculation cepstrum coefficient: The full 

spectrum above were done by M-order Mel filter, and got 

Mel spectrum, furthermore, translated by discrete cosine 

transformation to obtain cepstral vector coefficient dh= 

(dh,1, dh,2,…,dh,M ) of the hth frame, in which h=1,…, 

Nci, NCi is the number of the frames. 

Step 5 Cepstrum coefficient splice: Each frame was 

processed by the method above, and cepstrum 

coefficients of all frames were spliced, then could get 

NCi*M-dimension feature vectors DWTMFC-CT=(d1, 

d2,…,dNCi). 

V. EXPERIMENTS 

Speech samples were collected in a quiet 

environment, recorded by Cool Edit software. The 

number of recording channel was one, its frequency was 

16000Hz and precision was 16bit. The distance from 

microphone to people’s mouth is form 10 cm to 20 cm. 

10 females and 10 males (healthy and speech Mandarin) 


Figure 3. A graph describing the result of FSVM(C1-C2). 

Figure 4. A graph describing the result of FSVM(C3-C4). 

Figure 5. A graph describing the result of FSVM(C5-C6).


were called to do experiments. Each consonant was read 

5 times, then all of the samples were front-end processed 

and segmented the vowel parts and consonant parts of 

consonants by usual. Finally each consonant had 100 

samples. 

The aim of the first experiment was to determine 

whether the feature parameter vector F (including length 

(L), periodicity (P), relative energy (E) and zero-crossing 

rate(Z)) was useful in the first stage of consonant 

recognition. 100 samples in each rough category set were 

selected randomly. Two-class fuzzy support vector 

machine was used as classifier. For example, three 

pictures of results were shown in three rough category 

pairs (C1-C2 (Fig. 3), C3-C4 (Fig. 4), and C5-C6 (Fig. 

5)). We can see that, there were significant differences 

among the rough categories with the feature parameter 

vector F. 

In the second stage of consonant recognition, in order 

to test the effectiveness of the 7 fine classifiers, 10 

experiments were done. In each experiment, 60 samples 

in each consonant sample set were selected randomly as 

the training samples, the remaining 40 samples as the test 

samples. That is, the ith fine classification had 60*Ki 

category title 

Accuracy(%) 

TABLE III. 

THE RECOGNITION RESULTS OF 7 CLASSES 

training samples and 40*Ki test samples. The Mel 

cepstrum coefficients (MFCC) and DWTMFC-CT of 

consonant samples were extracted (the number of FFT 

points was min {512,the number of sample points}, Mel 

filter order was 24). SVM and FSVM were also used for 

comparative experiments. The recognition results of 7 

classes were shown in TABLE II and TABLE III. The 

results of TABLE II and TABLE III were the average of 

10 experiments. 

We can see that: compared with the speech feature 

classification by using the MFCC directly, the 

DWTMFC-CT based on wavelet transformation could 

reduce a large number of support vectors and improve the 

correct rate, particularly in rough categories with similar 

pronunciation: C3 (z similar to zh), C5 (c similar to ch), 

C7 (s similar to sh), and the effectiveness of C1 (with the 

strongest instability and the shortest length) was also 

significant. It showed that unstable consonants can be 

described more accurately by the multi-scale wavelet 

analysis. When chosen the same DWTMFC-CT feature 

vectors, compared with SVM, the recognition correct 

rates of FSVM were better, while the difference of 

support vectors number was small. It showed that FSVM 

SVM(MFCC) SVM(DWTMFCC-CT) FSVM(DWTMFCC-CT) 

number of support 

vectors 

Accuracy(%) 


vectors 

Accuracy(%) 


vectors 

C1 [b,d,g] 84.33 63.4 89.83 47.2 91.58 46.8 

C2 [l,m,n,r] 93.75 88.8 97.13 47.5 98.38 50.6 

C3 [z,zh,j] 85.75 61 90.92 49.5 92.17 54.4 

C4 [p,t,k] 85.67 68.7 89.33 51.5 91.67 50.2 

C5 [c,ch,q] 82.58 68.7 88.17 41.8 90.83 44.2 

C6 [f,h] 95.50 10.1 98.25 7.3 99.25 8.4 

C7 [s,sh,x] 86.42 75.9 95.50 23.5 96.58 24.8 

Average 87.71 61.34 92.73 38.33 94.35 39.91 

category title 

TABLE II. 

COMPARISON AMONG SVM(DWTMFCC), SVM(MFCC) AND FSVM(DWTMFCC-CT) 

SVM(MFCC) VS 

SVM(DWTMFCC-CT) 

Difference of 

Accuracy(%) 

Difference of the 


vectors(%) 

SVM(DWTMFCC-CT) VS 

FSVM(DWTMFCC-CT) 


Accuracy(%) 



vectors(%) 

SVM(MFCC) VS 

FSVM(DWTMFCC-CT) 


Accuracy(%) 



vectors(%) 

C1 [b,d,g] 6.52 -25.55 1.95 0.85 8.60 -26.18 

C2 [l,m,n,r] 3.60 -46.51 1.29 -6.53 4.93 -43.02 

C3 [z,zh,j] 6.03 -18.85 1.37 -9.90 7.48 -10.82 

C4 [p,t,k] 4.28 -25.04 2.61 2.52 7.01 -26.93 

C5 [c,ch,q] 6.76 -32.03 3.02 -5.74 9.99 -28.13 

C6 [f,h] 2.88 -27.72 1.02 -15.07 3.93 -16.83 

C7 [s,sh,x] 10.51 -69.04 1.13 -5.53 11.76 -67.33 

Average 5.80 -60.04 1.77 -3.97 7.67 -53.69 



Figure 6. The Standard Deviation of accuracy of each category. 

Figure 7. The Standard Deviation of the support vector number of 

each category. 

with introducing fuzzy membership could describe the 

importance of each training sample to classification 

result. So it could optimize SVM classification face and 

improve recognition accuracy (Fig. 6, Fig. 7). 

At last the efficiency of algorithm (CFDCT) 

mentioned in this paper was further validated by 

comparing with the BP neural network (three structures, 

hidden nodes is 49, the maximum training number is 500 

and error goal


algorithm. Control and Decision, vol. 17, No. 1, pp. 65-68, 

2002. 

[7] ZHAI Yong-jie; HAN Pu; WANG Dong-feng; WANG 

Guo-peng. Sisk function based sum algorithm and its 

application to a slight malfunction diagnosis. Proceedings 

of the Csee, vol. 23, No. 9, pp. 198-203, 2003. 

[8] QI Li; LIU Yu-shu. Fuzzy Support Vector Machine Based 

on Two Stage Clustering. Computer Engineering, vol. 34, 

No. 1, pp. 4-6, 2008. 

[9] Hsu Chihwei, Lin Chihjen. A comparison of methods for 

multi-class support vector machines. IEEE Trans on Neural 

Networks, 

[10] Zhang Jun; Zhang De-yun; Fu Peng. Objective Speech 

Quality Evaluation Based on Fuzzy Multi-Class Support 

Vector Machine. Journal of Xi an Jiaotong University, vol. 

40, No. 2, pp. 199-202, 2006. 

[11] DONG Chang-hong, MTLAB toolbox of theory and 

application of wavelet analysis. Beijing: National Defence 

Industry Press, 2004. 

[12] Xu Bing-zheng, Qiu Wei. Classification and Recognition 

of Chinese (Putonghua) Consonants.Journal of Chinese 

Information Processing, vol. 1, No. 7, pp. 33-39, 1993. 

[13] MO Jia-ling; HU Wei-ping. Speech Features Extraction 

Based on Invariant Sets Multi-wavelet. Audio Engineering, 

vol. 33, No. 7, pp. 63-67, 2009. 

[14] LIU Ming, DAI Bei-qian, LI Hui, LI Xiao-han, LU Wei. A 

New Speech Feature Extracted by Wavelet Analysis & 

Mel-Frequancy Filtering. Journal of Circuits and Systems, 

vol. 5, No. 1, pp. 21-25, 2000. 

Zhuo-ming Chen received the B.S. 

degree in clinical medicine and M.S. 

degree in neurology from Jinan 

University, Guangzhou, China in 198* 

and 199*, respectively, and the Ph.D. 

degree in Neuropsychology From South 

China Normal University, Guangzhou, 

China in 2007.He is the chief of 

Rehabilitation Department and Center of 

Language Disorder Diagnosis and 

treatment at the first affiliate hospital of Jinan University. He is 

the moonlighting researcher in the Psychological Applying 

Center at the South China Normal University. He is the main 

researcher in the “Language disorder diagnosis apparatus” and 

“Cognitive disorder diagnosis apparatus”. He is proficient in 

neuro-rehabilitation and diagnosis and treatment of language 

disorder. Dr. Chen had been the subeditor of the , editorial board member of the , and so on. 

Wei-xing Ling received the B.S., M.S., and Ph.D. degree in 

computer science from South China University of Technology, 

China. She is associate professor at School of Sciences. Her 

main researches are neural networks and fuzzy theory. 

Jian-hui Zhao received the B.S. degree in mathematics from 

South China University of Technology, China. He is on reading 

postgraduate. His main researches are speech recognition and 

image recognition. . 

Tao-tao Yao is on reading postgraduate. Her main researches 

are neuro-rehabilitation and diagnosis and treatment of language 

disorder. 



ELECTRE I Decision Model of Reliability 

Design Scheme for Computer Numerical Control 

Machine 

Jihong Pang 

College of Mechanical Engineering, Chongqing University, Chongqing 400044, China 

Department of Management, Guangxi University of Technology, Liuzhou 545006, China 

E-mail: pangjihong@163.com 

Genbao Zhang and Guohua Chen 

College of Mechanical Engineering, Chongqing University, Chongqing 400044, China 

E-mail: genbaozhang@163.com; 59782071@163.com 

Abstract—The ELECTRE I is one of the most extensively 

used methods to solve multiple criteria decision making 

(MCDM) problems. In this paper, we propose a novel AHPbased 

ELECTRE I method of reliability design scheme 

decision for computer numerical control (CNC) machine. 

Based on the AHP method combined with ELECTRE I, the 

decision model is built to select the optimal design scheme. 

The AHP method is applied to determinate the weights of 

reliability design factors through the decision model. 

ELECTRE I method is then designed to rank reliability 

design scheme in order of decision maker’s preference. To 

evaluate performance of the developed algorithm, an 

illustrative example of CNC machine is given. The 

computational results show that the proposed approach is 

reliable and performs well. 

Index Terms—reliability design, scheme decision, multiple 

criteria decision making, electre, analytic hierarchy process 


Reliability design of CNC machine has been widely 

applied during the past decades. High reliability proves 

not only successful experience in manufacturing field, 

bust also strategic need for manufacturing enterprises 

improving market competence. Strictly speaking, 

performance is less important than reliability in a CNC 

machine. It is the key for quality of the products to realize 

the value of reliability design. Furthermore, the 

optimization reliability design scheme has played an 

important role to ensure the reliability and rationality of 

the product development design [1]. More and more 

enterprises attach great importance to vouch for the 

reliability of the mechanical products in the development 

and application of reliability design [2]. 

On the other hand, customers require high quality CNC 

machine with high performance, high reliability and 

security. Therefore, at the product development and 

Manuscript received June 1, 2010; revised November 1, 2010; 


Corresponding Author: Jihong Pang 


doi:10.4304/jsw.6.5.894-900 

design stage, adequate decision method that select the 

optimal reliability design scheme for the CNC machine 

are essential. However, reliability design is also a 

complex task because of the large number of reliability 

factors that have to be taken into consideration in the 

product design process [3]. The enormous complexity of 

reliability design makes product designers hard to select 

an optimum scheme from many design schemes. 

Much research has been done on reliability-based 

design optimization. Youn et al. [4] presented the 

conjugate mean value (CMV) method for the concave 

performance function in the performance measure 

approach (PMA) of reliability-based design optimization. 

Du and Chen [5] developed the sequential optimization 

and reliability assessment method for probabilistic 

optimization. Using a single-loop strategy with 

deterministic optimization and reliability assessment, 

their application results demonstrated the effectiveness of 

reliability-based design method. Gea and Oza [6] 

proposed a two-level approximation method to solve the 

reliability-based design optimization problem. Chwail 

and Choi [7] presented an improved method to solve 

reliability-based design optimization problem. To 

estimate the effect of the response surface error, the 

developed method used the prediction interval to obtain 

an optimum reliability design. 

Because of CNC machines with their millions of 

components, reliability design evaluation and 

optimization is becoming more and more complex and 

difficult. This decision and optimization model is often 

called MCDM problem. In a recent paper [8], the authors 

have provided a reliability assessment method to improve 

the efficiency for solving problem of probabilistic 

optimization with changing variance. In order to improve 

the accuracy of nonlinear and multi-dimensional 

performance functions, Lee et al. [9] proposed an inverse 

reliability analysis method was applied to improve the 

accurate probability of failure calculation for reliability 

design optimization. Zhang et al. [10] provided 

probabilistic perturbation method multi-objective 

optimization problem of reliability optimization design.


They put forward Ant Colony Algorithm to improve the 

road header operational reliability. Injoong [11] proposed 

a system design-for-reliability method and reliability 

object model tree for reliability design of complex 

systems. Bhattacharjee et al. [12] established reliability 

design optimization formulation based on response 

surface method under uncertainty environment. And then 

the structural reliability was evaluated by the Advanced 

First-Order Second-Moment Method. 

This paper develops a decision method of reliability 

design scheme for CNC machine using AHP assessment 

model and elimination and choice translating reality 

(ELECTRE) method. The AHP method is applied to 

determinate the weight factors through the selecting and 

decision-making model. Then, the ELECTRE I method is 

used to select the alternatives combining AHP method. 

The objective is to select the optimal reliability design 

scheme, satisfying customers in the aspects of quality and 

reliability needs to the most degree. 

The paper is organized as follows. Section 2 describes 

the AHP and ELECTRE I method. The framework of the 

proposed AHP-based ELECTRE I algorithm are 

demonstrated. In Section 3, the hierarchic architecture 

model for the reliability design scheme is established 

based on AHP method. It then proposes a novel method 

for the scheme decision during the CNC machine 

reliability design process. An illustrative example of 

CNC machine is provided in order to assess the 

contribution of the proposed approach. The final diction 

offers concluding remarks. 

II. DECISION OF THE PROPOSED MODEL FOR RELIABILTY 

DESIGN SCHEME 

The decision of reliability design scheme for CNC 

machine is a MCDM problem in engineering worlds. The 

decision of reliability design scheme is a choice made 

from two or more reliability design schemes. The 

selection of reliability design scheme is very critical for 

product development staff because the optimum design 

adds vital value on the product quality and reliability. 

When a new product is under study, product development 

teams should make a major strategic decision of 

reliability design scheme. 

In the decision making process considered in this 

paper, it is very important to find a suitable method to 

solve the alternatives selection problem. The best 

decision of product development team is pursuing high 

reliability and quality for the good design product. 

Therefore, the AHP-based ELECTRE I method is 

developed to make the decision of reliability design 

scheme. 

The decision making process of reliability design 

scheme for CNC machine is shown in Fig.1. Since the 

decision of reliability design scheme for CNC machine is 

a quite complicated process, AHP method is first applied 

to build the decision model so as to aid decision support. 

When the weights of reliability design indicators are 

confirmed by using AHP approach, ELECTRE I method 

must be taken to determine the rank of reliability design 


scheme. The following sections describe the decision 

process of reliability design scheme for CNC machine. 

AHP 

ELECTRE 

Reliability design scheme 

decision for CNC machine 

Decision model of reliability 

design scheme 

Confirmation of reliability design 

indicators weights using AHP 

Determining reliability design 

scheme rank using ELECTRE 

Figure 1. Decision making process of reliability design scheme 

A. Application of AHP in weighting design indicators 

The analytic hierarchy process (AHP) method was first 

proposed by Saaty [13]. The AHP is widely used as one 

of the popular methods in solving all kinds of problems 

of MCDM and calculating weighting vector method [14- 

15]. The primary advantage of the AHP approach is to 

incorporate judgments on qualitative and quantitative 

data [16]. First, AHP breaks down a complex MCDM 

problem into a hierarchy of interrelated decision 

indicators and alternatives. Then, the indicators and 

alternatives are compared in pair-wise comparison within 

each level. The standardized comparison scale of 9 levels 

is used to compare the importance of all indicators, such 

as “3” means “moderately more important”. 

Once the weights of reliability design indicators are 

calculated by AHP method, the ELECTRE I approach 

will be used to obtain the four ranking scheme scores of 

the CNC machine. 

B. Decision of reliability design scheme by using 

ELECTREⅠ 

To rank a set of alternatives, the ELECTRE method as 

outranking relation theory was used to analyze the data of 

a decision matrix. The Elimination and Choice 

Translating Reality (ELECTRE) method was first 

introduced in [17]. It is one of the most extensively used 

outranking methods reflecting the decision maker’s 

preferences in many fields. The ELECTRE I approach 

was then developed by a number of variants [18]. 

Teixeira [19] utilized the ELECTRE I method in a multi 

criteria decision model supports decision makers. Shanian 

and Savadogo [20] provided ELECTRE I method to 

select the material of bipolar plates based on multiple 

conflicting objectives. The transport sustainability was 

firstly evaluated by ELECTRE method in [21], then the 

modification of ELECTRE I was used to reduce the 

subjectivity of decision makers. 

ELECTRE method reflect the dominance of relations 

among alternatives by outranking relations [22]. It is 

possible that the alternatives can be compared by these


outranking relations built in the way. Different 

ELECTRE method, concordance and discordance indexes 

are two types of indices pair-wise comparison between 

alternatives in ELECTRE I. With a simple analysis of the 

concordance reliability index, ELECTRE I method was 

applied to select the optimal reliability design scheme in 

this paper. 

We assume that A1, A2,…, Am are m possible 

alternatives for optimum reliability design scheme of 

CNC machine, C1, C2,…, Cn are criteria that used to 

describe the alternative characters, after the assignment, 

defined as xij for the degree of alternative Ai with respect 

to criteria Cj. Let Wn be the weight for importance of Cn, 

which is determined by AHP method. The computation 

flow process of ELECTRE I method is stated in the 

following paragraphs. 

Step 1. Normalization of matrix and weighted matrix 

Considering concepts on the interval numbers of 

decision matrix, the normalized matrix of R ij = [ rij 

] is 

calculated by (1): 

x ij 

rij = , i = 1, 2, …, n j = 1, 2, …, 

m (1) 

m 

x 

2 

∑ ij 

i= 

1 

Thus, the weighted matrix depends on normalized 

matrix assigned to it is given by: 

r ⋅w r ⋅w … r ⋅w 

r21 ⋅w1 Vij = R× W = 

� 

r22⋅w2 � 

… 

� 

r2n⋅w 

� 

r ⋅w r ⋅w � r ⋅w 

11 1 12 2 1n⋅n 

m1 1 m2 2 

mn n 

Where 0≤w1,w2,…,wn≤1. The weights of the attributes 

are expressed by these constants. Besides, the correlation 

coefficients of normalized interval numbers are between 

0 and 1. 

Step 2. Ascertainment of concordance and discordance 

interval sets 

Considering that reliability design scheme decision is a 

multi-attribute decision with preference information, the 

decision rules are reasoned by the concordance and 

discordance interval sets, and then the attribute sets are 

obtained through these decision rules. Let A = { abc , , , � } 

denote a finite set of alternatives, in the following 

formulation we divide the attribute sets into two different 

sets of concordance interval set (Cab) and discordance 

interval set (Dab). The concordance interval set is applied 

to describe the dominance query if the following 

condition is satisfied: 

{ jx x } 

Cab aj bj 

n 

(2) 

= ≥ (3) 

On complementation of Cab, we obtain the discordance 

interval set (Dab) using (4): 

{ aj bj} 

D = jx < x = J− 

C 

(4) 

ab ab 


Step 3. Calculation of the concordance interval matrix 

According to the deciders’ preference for alternatives, 

the concordance interval index (Cab) between Aa and Ab 

can be obtained using (5): 

Cab 

= ∑ w 

(5) 

j∈Cab The concordance index indicates the preference of the 

assertion “A outranks B”. The concordance interval 

matrix can be formulated as follows: 

− c(1, 2) … c(1, m) 

C = 

c(2,1) 

� 

− 

� 

… 

� 

c(2, m) 

� 

c( m,1) c( m,2) 

� − 

Step 4. Calculation of the discordance interval matrix 

First, we consider the discordance index of d(a,b) , 

which can be viewed as the preference of discontent in 

decision of scheme a rather than scheme b. More 

specifically, we define: 

d(a,b) = 

max v − v 

j∈Dab max 

j∈J, m, n∈I j 

aj bj 

v − v 

mj nj 

Here scheme m, n is used to calculate the weighted 

normalized value among all scheme target attributes. 

Then, using discordance interval index sets, we can 

obtain discordance interval matrix as: 

− d(1,2) … d(1, m) 

D = 

d(2,1) 

� 

− 

� 

… 

� 

d(2, m) 

� 

d( m,1) d( m,2) 

� − 

Step 5. Determine the concordance index matrix 

The concordance index matrix for satisfaction 

measurement problem can be written as follows: 

m 

m 

a= 1 b 

(6) 

(7) 

(8) 

c = ∑∑ cab ( , )/ mm ( −1) 

(9) 

Here c is the critical value, which can be determined by 

average dominance index. Thus, a Boolean matrix (E) is 

given by: 

⎧eab 

( , ) = 1 if cab ( , ) ≥c 

⎨ 

⎩eab 

( , ) = 0 if cab ( , ) < c 

(10) 

Step 6. Determine the discordance index matrix 

On the contrary, the preference of dissatisfaction can 

be measured by discordance index:


d = 

m 

m 

∑∑ 

a= 1 b 

dab ( , ) 

mm ( −1) 

(11) 

Based on the discordance index mentioned above, the 

discordance index matrix (F) is given by: 

⎧ ⎪ f ( ab , ) = 1 if dab ( , ) ≤ d 

⎨ 

⎪⎩ f ( ab , ) = 0 if dab ( , ) > d 

Step 7. Calculate the net superior and inferior value 

(12) 

Let c a and d a be the net superior and net inferior 

value respectively. c a sums together the number of 

competitive superiority for all alternatives, and the more 

and bigger, the better. The c a is given by: 

c 

n n 

∑ ∑ (13) 

= c − c 

a ( a, b) ( b, a) 

b= 1 b= 

1 

On the contrary, d a is used to determine the number of 

inferiority ranking the alternatives: 

d 

n n 

∑ ∑ (14) 

= d − d 

a ( a, b) ( b, a) 

b= 1 b= 

1 

Smaller is better. This is the biggest reason that smaller 

net inferior value gets better dominant then larger net 

inferior value by sequence order. 

III. ILLUSTRATIVE EXAMPLE 

As an illustration of the use of the proposed method for 

reliability design scheme decision for CNC machine, a 

numerical example is presented in this study. To examine 

the potential applications of the AHP-based ELECTRE I, 

we taken into account the design standards data obtained 

from CNC machine. 

A. Confirmation of reliability design indicators weights 

using AHP 

The AHP method was utilized to calculate the 

indicators weights of the reliability design scheme (RDS). 

Based on a basic reliability design, CNC machine was 

taken as one of the references to estimate the design 

schemes. The all reliability design indicators were 

selected: Mean Time To First Failure (MTTFF, hour), 

Mean Time Between Failures (MTBF, hour), Mean Time 

To Repair (MTTR, hour), Annual Maintenance Charge 

Rate (AMCR, %), Inherent Reliability (IR, %) and 

Failure Rate (FR). The structure of decision hierarchy is 

shown in Fig.2. 

Then, the task of the experts in the expert team is to 

create individual pair-wise comparison matrix for all 

design indicators. The matrices of these values are given 

in Table I. 

By applying the AHP method, the importance weights 

of the all reliability design indicators with respect to the 

main objective were obtained, the details of the calculated 

results are shown in Table II. 


Decision of reliability design scheme 

MTTFF MTBF MTTR AMCR IR FR 

RDS 1 RDS 2 … RDS n 

Figure 2. Decision model of reliability design scheme 

TABLE I. 

PAIR-WISE COMPARISON MATRIX FOR RELIABILITY INDICATORS 

TABLE II. 

THE IMPORTANCE WEIGHTS BY AHP METHOD 


W 0.2336 0.1652 0.3355 0.1021 0.0424 0.1212 

Rank 2 3 1 5 6 4 

λmax 6.5162 

CI 0.1032 

RI 1.24 

CR 0.0833 


MTTFF 1 2 1 1 4 3 

MTBF 1/2 1 1/2 2 3 2 

MTTR 1 2 1 3 7 5 

AMCR 1 1/2 1/3 1 2 1/2 

IR 1/4 1/3 1/7 1/2 1 1/7 

FR 1/3 1/2 1/5 2 7 1 

The importance weights of reliability design indicators 

were accepted because the associated CR were smaller 

than 0.1, as is shown in Table II. Therefore, the decision 

matrix of the proposed hierarchical structure for decision 

model of reliability design scheme is consistent. The 

results indicate that the calculation and analysis are 

accurate and rational. 

B. Determining the scheme rank using ELECTRE 

a. Confirming normalized and weighted matrix 

In this case study, four reliability design schemes for 

CNC machine are compared with respect to six reliability 

design indicators (see Fig.2). There is a close relation 

between reliability design scheme and reliability design 

indicators measured by quantitative index. Thus, the 

system of decision index to estimate the reliability must 

be established with the quantitative data. In addition, 

since it takes much decisive data to select an optimal 

scheme by using the proposed approach, the values of


qualitative research on indicators for reliability design 

scheme are presented in Table III. 

The indicators of MTTFF, MTBF and IR are the 

“bigger-the-better” type of indicators, and others are the 

“smaller-the-better” type. According to normalization 

method, the normalized matrix can be determined by 

using (1): 

⎡0.4441 0.5059 0.4642 0.3500 0.4972 0.3116⎤ 

⎢ 

0.5527 0.4513 0.5261 0.5088 0.5068 0.4708 

⎥ 

R = ⎢ ⎥ 

⎢0.5132 0.5333 0.4023 0.6594 0.4712 0.6578⎥ 

⎢ ⎥ 

⎣0.4836 0.5059 0.5880 0.4287 0.5233 0.4985⎦ 

Based on the importance weights (see Table Ⅱ) and 

(2), the weighted matrix is calculated as follows: 

⎡0.1037 0.0836 0.1558 0.0357 0.0211 0.0378⎤ 

⎢ 

0.1291 0.0745 0.1765 0.0519 0.0215 0.0571 

⎥ 

V= ⎢ ⎥ 

⎢0.1199 0.0881 0.1350 0.0673 0.0200 0.0797⎥ 

⎢ ⎥ 

⎣0.1130 0.0836 0.1973 0.0438 0.0222 0.0604⎦ 

b. Computing process by using ELECTRE 

In this work, we are interested in making decision for 

the best alternatives. As a result, the computing process 

was proposed to rank four reliability design schemes of 

CNC machine by using ELECTRE I methods. With 

respect to (3), the concordance interval sets can be 

ascertained as follows: 

C 

C 

C 

C 

C 

C 

12 

14 

23 

31 

34 

42 

= {2,3,4,6}, 

= {2, 

3,4,6}, 

= {1,4,5,6}, 

= {1,2,3}, 

= {1,2, 

3}, 

= {2,4,5}, 

C 

C 

C 

C 

C 

C 

13 

21 

24 

32 

41 

43 

= {4,5,6}, 

= {1,5}, 

= {1, 

3,6}, 

= {2,3}, 

= {1,2,5}, 

= {4,5,6}. 

Accordingly, based on the concept of discordance 

interval set, we have the discordance interval sets using 

(4) as follows: 

D 

D 

D 

D 

D 

D 

12 

14 

23 

31 

34 

42 

TABLE III. 

THE VALUES OF RELIABILITY DESIGN SCHEME 


RDS1 1350 1850 7.5 2.58 93.5 0.045 

RDS2 1680 1650 8.5 3.75 95.3 0.068 

RDS3 1560 1950 6.5 4.86 88.6 0.095 

RDS4 1470 1850 9.5 3.16 98.4 0.072 

= {1,5}, 

= {1, 

5}, 

= {2,3}, 

= {4,5,6}, 

= {4,5, 

6}, 

= {1,3,6}, 

D 

D 

D 

D 

D 

D 

13 

21 

24 

32 

41 

43 


= {1,2,3}, 

= {2,3,4,6}, 

= {2, 

4,5}, 

= {1,4,5,6}, 

= {3,4,6}, 

= {1,2,3}. 

Using (5), the concordance interval index can be 

obtained. For example, the concordance interval index 

of c ( 1, 

2) 

and c( 1, 

3) 

can be calculated as follows: 

c( 

1, 

2) 

= ∑ w j 

j∈c12 

= 0. 

1652 + 0. 

3355 + 0. 

1021+ 

c( 

1, 

3) 

= ∑ w j 

j∈c12 

= 0. 

1021+ 

0. 

0424 + 0. 

1212 = 

0. 

1212 

0. 

2657 

= 

0. 

7240 

Similarly, the same procedure is applied to calculate 

the other concordance interval indexes. After all 

concordance interval indexes had been calculated, the 

concordance interval matrix is given as: 

⎡ − 

C = 

⎢0. 

2760 

⎢0. 

7343 

⎢⎣ 

0. 

4412 

0. 

7240 

− 

0. 

5007 

0. 

3097 

0. 

2657 

0. 

4993 

− 

0. 

2657 

0. 

7240⎤ 

0. 

6903⎥ 

0. 

7343⎥ 

− ⎥⎦ 

Furthermore, the concordance index can be determined 

by (9), which is expressed as follows: 

c 

4 4 

= ∑ ∑ 

a= 1 b= 

1 

c( 

a, 

b) 

= 

4× 

( 4 −1) 

6. 

1652 

12 

= 

0. 

5138 

Once the concordance index was calculated, according 

to (10), the concordance index matrix is given as: 

⎡− 

E = 

⎢0 

⎢1 

⎢⎣ 

0 

1 

− 

0 

0 

0 

0 

− 

0 

1⎤ 

1⎥ 

1⎥ 

−⎥⎦ 

Therefore, the net superior values for each scheme are 

obtained by (13): 

c 

c 

c 

c 

1 

3 

4 

2 

4 

4 

1 

1 1 

( 0. 

7240 0. 

2657 0. 

7240 ) 

( 0. 

2760 0. 

7343 0. 

4412 ) 0. 

2622 

1 = ∑ c − ∑ 

b c b 

b = b = 

= + + 

− + + = 

4 

4 

2 

1 1 

( 0. 

2760 0. 

4993 0. 

6903 ) 

( 0. 

7240 0. 

5007 0. 

3097 ) 0. 

0688 

2 = ∑ c − ∑ 

b cb 

b = b = 

= + + 

− + + = − 

4 

4 

3 

1 1 

( 0. 

7343 0. 

5007 0. 

7343 ) 

( 0. 

2657 0. 

4993 0. 

2657 ) 0. 

9386 

3 = ∑ c − ∑ 

b cb 

b = b = 

= + + 

− + + = 

4 

4 

= ∑ c4 

− ∑ 

b cb 

4 

b = 1 b = 1 

= ( 0. 

4412 + 0. 

3097 + 0. 

2657 ) 

− ( 0. 

7240 + 0. 

6903 + 0. 

7343 ) = −1. 

1320 

Similarly, the discordance index can be obtained by 

using (7) using the same count. For example, the 

discordance index of d ( 1, 

2) 

and d( 1, 

3) 

can be calculated 

as follows:


max( 

0. 

1037 − 0. 

1291, 

0. 

0211− 

0. 

0215) 

d( 

1, 

2) 

= 

0. 

0254 

0. 

0254 

= = 1. 

0000 

0. 

0254 

max( 

0. 

1037 − 0. 

1199, 

0. 

0836− 

0. 

0881, 

0. 

1558− 

0. 

1350) 

d( 

1, 

3) 

= 

= 

0. 

0208 

0. 

0420 

= 

0. 

0420 

0. 

4949 

Using the same counting method, the remaining 

discordance interval indexes are computed. After all 

discordance interval indexes had been determined by the 

similar computational process, the discordance interval 

matrix is given as: 

⎡ − 

D = 

⎢0. 

8189 

⎢1. 

0000 

⎢⎣ 

1. 

0000 

1. 

0000 

− 

0. 

5455 

1. 

0000 

0. 

4949 

1. 

000 

− 

1. 

0000 

0. 

2220⎤ 

0. 

4351⎥ 

0. 

3780⎥ 

− ⎥⎦ 

Furthermore, using the discordance interval matrix 

described above, the discordance index can be 

determined by (11): 

d 

4 4 

= ∑ ∑ 

a= 1 b= 

1 

d ( a, 

b) 

= 

4 × ( 4 − 1) 

8. 

8945 

12 

= 

0. 

7412 

Based on the discordance index calculated above, the 

discordance index matrix (F) is obtained by using (12) as 

follows: 

⎡− 

F = 

⎢1 

⎢1 

⎢⎣ 

1 

1 

− 

0 

1 

0 

1 

− 

1 

0⎤ 

0⎥ 

0⎥ 

−⎥⎦ 

Finally, based on the concept of net inferior ranking 

the alternatives, the net inferior values for each scheme 

are obtained by (14): 

d 

d 

d 

d 

1 

2 

3 

4 

4 

4 

= ∑ d1 

− ∑ 

b d b1 

b = 1 b = 1 

= ( 1. 

0000 + 0. 

4949 + 0. 

2220 ) 

− ( 0. 

8189 + 1. 

0000 + 1. 

0000 ) = −1 

. 1020 

4 

4 

= ∑ d 2 − ∑ 

b d b 2 

b = 1 b = 1 

= ( 0 . 8189 + 1 . 0000 + 0 . 3936 ) 

− ( 1. 

000 + 0. 

5455 + 1 . 000 ) = − 0. 

3330 

4 

4 

= ∑ d 3 − ∑ 

b d b 3 

b = 1 b = 1 

= ( 1. 

000 + 0. 

5455 + 0. 

3780 ) 

− ( 0. 

4949 + 1. 

000 + 1. 

000 ) = −0 

. 5714 

4 

4 

= ∑ d 4 − ∑ 

b d b 4 

b = 1 b = 1 

= ( 1.. 

000 + 1. 

000 + 1. 

000 ) 

− ( 0. 

2220 + 0. 

3936 + 0. 

3780 ) = 


2. 

0064 

After all the net superior values and net inferior values 

for each scheme are calculated, reliability design scheme 

can be sorted by the calculations. 

c. Discussion of the sorting results of reliability design 

scheme 

According to computing the net superior and net 

inferior values for each scheme, the sorting results are 

shown in Table IV. 

TABLE IV. 

SORTING RESULTS OF RELIABILITY DESIGN SCHEME 

Net superior 

values 

Net inferior 

values 

Ranking of Net 

superior values 

Table IV compares the performances of each design 

scheme with the net superior and net inferior values. The 

computation results of the net superior values show that 

RDS3 have the max value, which is the best scheme (see 

Table Ⅳ). On the other hand, sorting the reliability 

design scheme based on the net inferior values, RDS1 

finished top while RDS4 ranked last. According to the 

theory of ELECTRE I, excluding RDS2 and RDS4, the 

optimal schemes of reliability design for CNC machine 

include RDS1 and RDS3. 

IV. CONCLUSION 

Ranking of Net 

inferior values 

RDS1 0.2622 -1.1020 2 1 

RDS2 -0.0688 -0.3330 3 3 

RDS3 0.9386 -0.5714 1 2 

RDS4 -1.1320 2.0064 4 4 

The conclusion of this study is that the optimal 

reliability design scheme can be selected accurately by 

using AHP and ELECTRE I method. Firstly, we adopt 

AHP method to calculate the weights of reliability design 

indicators. Then, AHP-based ELECTRE I methodology 

were utilized synthetically to rank the design schemes. 

The approach proposed in this paper presents diverse 

choices for product designers select the best alternatives. 

Finally, the implementation of the novel method is 

demonstrated by the illustrative example of CNC 

machine. The results of computational experiments 

indicated that the proposed algorithms possess good 

application prospect. 

As mentioned above, this research was motivated by a 

selecting problem of reliability design scheme. In 

practice, reliability design scheme decision for CNC 

machine usually consists of multi-objective optimization 

model. Thus, various effective factors of optimized model 

need to be considered in the decision process. The focus 

of future studies will concentrate on other ELECTRE 

methods such as ELECTRE II and III. We will also 

research other methods to select reliability design scheme 

for CNC machine.



Project supported by the National High-Tech. R&D 

Program, China (No. 2009AA04Z119), the National 

Natural Science Foundation, China (No. 50835008), the 

National Major Scientific and Technological Special 

Project for “High-grade CNC and Basic Manufacturing 

Equipment”, China (No.2009ZX04014-016 ; 

2009ZX04001-013;2009ZX04001-023; 2010ZX04014- 

015), and supported by Open Research Foundation of 

State Key Lab. of Digital Manufacturing Equipment & 

Technology in Huazhong University of Science & 

Technology. 

REFERENCES 

[1] L. Jinghong, Z. P. Mourelatos, and T. Jian, "A single-loop 

method for reliability-based design optimisation," 

International Journal of Product Development, pp. 76-92, 

2008. 

[2] H. Shimizu, Y. Otsuka, and H. Noguchi, "Design review 

based on failure mode to visualise reliability problems in 

the development stage of mechanical products," 

International Journal of Vehicle Design, vol. 53, pp. 149- 

165, 2010. 

[3] A. Mohsine, G. Kharmanda, and A. El-Hami, "Improved 

hybrid method as a robust tool for reliability-based design 

optimization," Structural and Multidisciplinary 

Optimization, vol. 32, pp. 203-213, 2006. 

[4] B. D. Youn, K. K. Choi, and Y. H. Park, "Hybrid analysis 

method for reliability-based design optimization," 

Transactions of the ASME. Journal of Mechanical Design, 

vol. 125, pp. 221-32, 2003. 

[5] X. P. Du and W. Chen, "Sequential optimization and 

reliability assessment method for efficient probabilistic 

design," Journal of Mechanical Design, vol. 126, pp. 225- 

233, 2004. 

[6] H. C. Gea and K. Oza, "Two-level approximation method 

for reliability-based design optimisation," International 

Journal of Materials & Product Technology, vol. 25, pp. 

99-111, 2006. 

[7] [7] K. Chwail and K. K. Choi, "Reliability-based design 

optimization using response surface method with 

prediction interval estimation," Journal of Mechanical 

Design, p. 121401 (12 pp.), 2008. 

[8] X. L. Yin and W. Chen, "Enhanced Sequential 

Optimization and Reliability Assessment method for 

probabilistic optimization with varying design variance," 

Structure and Infrastructure Engineering, vol. 2, pp. 261- 

275, 2006. 

[9] I. Lee, K. K. Choi, L. Du, and D. Gorsich, "Inverse 

analysis method using MPP-based dimension reduction for 

reliability-based design optimization of nonlinear and 

multi-dimensional systems," Computer Methods In 

Applied Mechanics and Engineering, vol. 198, pp. 14-27, 

2008. 

[10] Z. Qiang, L. Shouju, and T. Ying, "Multi-objective 

Incomplete Probability Information Optimization 

Reliability Design Based on Ant Colony Algorithm," 

Journal of Software Engineering and Applications, pp. 

350-3, 2009. 

[11] K. Injoong, "Reliability object model tree (ROM-Tree): a 

system design-for-reliability method," Microelectronics 

Reliability, pp. 438-46, 2010. 


[12] P. Bhattacharjee, K. Ramesh Kumar, and T. A. Janardhan 

Reddy, "Reliability design evaluation and optimization of a 

nitrogen gas bottle using response surface method," 

International Journal of Reliability, Quality and Safety 

Engineering, vol. 17, pp. 119-132, 2010. 

[13] T. L. Saaty, "The Analytic Hierarchy Process," New York: 

McGraw Hill, 1980. 

[14] N. Y. Secme, A. Bayrakdaroglu, and C. Kahraman, "Fuzzy 

performance evaluation in Turkish Banking Sector using 

Analytic Hierarchy Process and TOPSIS," Expert Systems 

with Applications, vol. 36, pp. 11699-11709, 2009. 

[15] S. Parthasarathy and N. Anbazhagan, "Evaluating ERP 

implementation choices using AHP," International Journal 

of Enterprise Information Systems, pp. 52-65, 2007. 

[16] T. Soota, H. Singh, and R. Mishra, "Defining 

characteristics for product development using quality 

function deployment: a case study on Indian bikes," 

Quality Engineering, pp. 195-208, 2008. 

[17] R. Benayoun, B. Roy, and N. Sussman, "Manual de 

reference du programme electre," Note De Synthese et 

Formaton, vol. 25, 1966. 

[18] B. Roy, "Classement et choix en presence de points de vue 

multiples: La methode ELECTRE," R.I.R.O, vol. 8, pp. 57- 

75, 1968. 

[19] A. Teixeira De Almeida, "Multicriteria modelling for a 

repair contract problem based on utility and the ELECTRE 

I method," IMA Journal of Management Mathematics, vol. 

13, pp. 29-37, 2002. 

[20] A. Shanian and O. Savadogo, "ELECTRE I decision 

support model for material selection of bipolar plates for 

Polymer Electrolyte Fuel Cells applications," Journal of 

New Materials for Electrochemical Systems, vol. 9, pp. 

191-199, 2006. 

[21] [N. Bojkovic, I. Anic, and S. Pejcic-Tarle, "One solution 

for cross-country transport-sustainability evaluation using 

a modified ELECTRE method," Ecological Economics, 

vol. 69, pp. 1176-1186, 2010. 

[22] A. T. de Almeida, "Multicriteria decision model for 

outsourcing contracts selection based on utility function 

and ELECTRE method," Computers & Operations 

Research, vol. 34, pp. 3569-3574, 2007. 

Jihong Pang was born in Beihai, 

Guangxi Zhuangzu Autonomous Region, 

China, on January 10, 1978. He received 

his master's degree in Management 

Science and Engineering from Tianjin 

University in 2006. 

Currently, he is a PH.D candidate 

with Mechanical Engineering at the 

College of Mechanical Engineering, 

Chongqing University in China since 2008. His main research 

interest is quality and reliability engineering, industrial 

engineering (IE), Enterprise Resource Planning (ERP). 

Genbao Zhang is a professor at Chongqing University. His 

main research interest is quality and reliability engineering, 

enterprise informatization, advanced manufacturing technology. 

Guohua Chen is a PH.D candidate with Mechanical 

Engineering at Chongqing University. His main research 

interest is quality engineering, supply chain management.


Fractional Modeling Method Research on 

Education Evaluation 

Chunna Zhao 

Information Engineering College, Capital Normal University, Beijing, China 

Email: chunnazhao@163.com 

Yu Zhao 

Yunnan Technician College, Kunming, China 

Email: yuzhao66@163.com 

Liming Luo 

Information Engineering College, Capital Normal University, Beijing, China 

Email: luolm@mail.cnu.edu.cn 

Yingshun Li 

Engineering College, Shenyang University of Technology, Liaoyang, China 

Email:liyingshunok@163.com 

Abstract—Education assessment is one of important 

measures that will be to ensure and continuously improve 

the educational level and educational quality. A fractional 

order model method for education evaluation is proposed in 

this paper. Improved fractional Basset force model is 

referred to constitute the complex education evaluation 

model. The detailed descriptions are displayed through 

building block diagram. An algorithm for linear fractional 

order systems is described. The fractional evaluation model 

is composed of fractional order and common coefficient. 

Model parameters can be determined by a large number of 

actual data and mathematical statistics method. The 

proposed model was applied to actual course evaluation 

work of Capital Normal University. The practicability and 

effectiveness of the method have been validated. 

Index Terms—fractional order, model, evaluation 


Post-Secondary Education Accreditation Council is set 

up in the United States in 1975. And this is the first 

institution of higher education assessment. In China 

educational assessment of colleges and universities began 

in the twentieth century, the eighties. Educational 

assessment is the efficient means that higher education 

institution realizes the higher education self-perfecting, 

self-regulation and self-improvement. Higher education 

assessment is one of important measures that will be to 

ensure and continuously improve the educational level 

and educational quality of China's institutions of higher 

Manuscript received January 10, 2010; revised June 15, 2010; 

accepted July 16, 2010. 

This work is supported by Beijing Education Committee 

Science(KM201010028021) and Technology Foundation and Liaoning 

Nature Science Foundation (20082044, 20060624) 


doi:10.4304/jsw.6.5.901-907 

learning. It aims to improve the quality of instruction and 

promote teaching reform. There is more and more 

research on education evaluation recently. Reference [1] 

applies a growth modeling approach to the stability of 

teaching effectiveness. Reference [2] contributes to the 

conceptual and empirical distinction between appraisals 

of teaching behavior and self-reported competence 

acquirement within academic education evaluation. 

Reference [3] examined the effects of embedding special 

education instruction into pre-service general education 

assessment courses. Some methods had not considered 

the existence of objective weight, so that the result is too 

subjective. Because education evaluation is a complex 

nonlinear process that affected by many factors, 

traditional integer order calculus model is unable to 

accurately describe its action. Fractional order system is 

established on the idea of fractional order calculus and 

theory of fractional order differential equations, which is 

an extension to the conventional calculus problems. 

It is well known that fractional order systems itself is 

an infinite dimensional filter due to the fractional order in 

the differentiator or integrator while the integer-order 

systems are with limited memory(finite dimensional). 

There has been a surge of interest in the possible 

engineering application of fractional order differentiation. 

Examples may be found in [4] and [5]. Some applications 

including automatic control are surveyed in [6]. The 

significance of fractional order theory is that it is a 

generalization of classical integral order theory, which 

could lead to more adequate modeling and more robust 

control performance. Fractional order systems could 

model various real materials more adequately than integer 

order ones and thus provide an excellent modeling tool in 

describing many actual dynamical processes[7]. 

Fractional model provides the scientific basis for 

prevention and treatment of satellite monitoring


absorption rate [8]. The nematode movement can be 

simulated through fractional model [9]. It may be used 

for building “love” models using fractional-order system 

[10]. [11] modeled iron meteorites crystallization by 

fractional theory. And there are some people pay close 

attention to unemployment rates by means of fractional 

calculus [12]. 

Fractional model is a mathematical modeling approach 

based on fractional calculus, and it provides a powerful 

decision support and scientific basis for education 

evaluation. Fractional education evaluation model is 

proposed in this paper. Model parameters can be obtained 

by the corresponding actual data. It aims to reduce the 

complexity while improving the scientific validity of the 

assessment results based on the fractional model. 

The remaining part of this paper is organized as 

follows. In Section II, mathematical foundation of 

fractional calculus is briefly introduced; in Section III, a 

fractional model method is presented for education 

evaluation; in Section IV, some practical examples are 

presented to verify the feasibility. Finally, conclusions are 

drawn in Section V. 

II. BRIEF INTRODUCTION OF FRACTIONAL CALCULUS 

Although the fractional order calculus is a 300-years 

old topic, the theory of fractional order derivative was 

developed mainly in the 19th century. References [13-18] 

provide a good source of references on fractional calculus. 

Fractional calculus is a generalization of integration 

and differentiation to a fractional, or non-integer order 

fundamental operator aDt � , where a and t are the 

lower/upper bounds of integration and � the order of the 

operation. 

� 

� d 

� R( 

�) 

� 0 

� 

�� 

dt 

� 

D ��1 R( 

�) 

� 

� t 

( �� 

) 

� ( d�) R( 

�) 

� 0 

a 

�� 

0 (1) 

a t 

which R( � ) is the real part of � . Moreover, the 

fractional order can be a complex number as discussed in 

[19]. In this paper, we focus on the case where the 

fractional order is a real number. 

Caputo’s fractional-order differentiation is defined by 

( m�1) 

1 t 

� 

f ( � ) 

D f() t � d� 

(1 ) a 

� 

� �� (2) 

( t ��) 

a t 

where � �m� � , m is an integer, and 0�� 

1 . 

Similarly, by Caputo’s definition, the integral is described 

by 

� 1 f ( � ) 

D f() t d 

( �) ( t �) � 

� 

�� 

a t 


t 

� � (3) 

a 1 � 

As in the case for conventional calculus, fractionalorder 

derivatives and integrals have the following 

qualities: 

(1) If f () t is an analytic function of the variable t , the 

� 

derivative Dt f() t is an analytic function of t and � . 

(2) The operation Dt � and the usual derivative of order 

� 

n�Z , � � n give the same result; The operation Dt � 

� 

and the usual n-fold integral with n�Z , � �n 

give the 

0 

same result, and Dt f() t � f() t . 

(3) The operator should be linear: 

� � 

� 

D [ af ( t) �bg( t)] �aD f ( t) �bDg( 

t) 

t t t 

(4) For the fractional-order integrals of arbitrary order, it 

holds the additive law of exponents (semi-group 

property): 

(4) 

� � �� D [ D f( t)] � D f( t) 

(5) 

t t t 

Linear fractional-order differential equations are the 

fundamental governing equations. The linear fractionalorder 

differential equation is defined as 

aD yt � a D yt �� 

(6) 

� aD yt �aDyt�b n 

�n t () n�1 �n�1 

t () 

1 

�1 

t () 0 

�0 

t () 

Substituting fractional-order differentiation definition 

into the above equation, one may find that 

� � 

�� b(7) 

[( t�a)/ h] [( t�a)/ h] 

0 

( �i) m 

( �n) 

j yt jh j y 

� � � � t jh 

0 

� � � � 

n 

h j�0 h j�0 

where the binomial coefficients 

evaluated recursively with 

� � 1, 

� 

( �i 

) 

0 

� � �1 

� 

� � 

� j � 

( �i) 

i ( �i) 

j � 1� 

� j�1 

� can still be 

( �i 

) 

j 

, j �1, 2, � (8) 

By slight rearrangement of the terms, the closed-form 

solution of the fractional-order differential equation can 

be obtained as 

� � 

� 

y �b yt�jh� 

� � 

n [( t�a)/ h] 

1 i 

( �i 

) 

t � � 

� 

n 

� 

j 

i 

� � � 

i i�0 h j�1 

� �i 

i�0 

h 

Fractional evaluation model will be proposed based on 

the above theory in this paper. Model parameters can be 

obtained by the corresponding actual data. It aims to 

reduce the complexity while improving the scientific 

validity of the assessment results based on the fractional 

model. 

III. FRACTIONAL MODEL 

Here fractional model of education evaluation will be 

established based on the above fractional order systems. 

Educational evaluation is to assess the course system. 

(9)


Firstly, each course is evaluated by the proposed method. 

Then course system evaluation result can be gotten 

depending on the course evaluation results. And then the 

curriculum and the construction could be analyzed. You 

can also compare the differences between courses. 

Previous courses are assessed by the expert ratings. 

The human factors have an important effect, and the 

result is not so objective. In order to overcome the 

drawbacks, fractional modeling approach is adopted. 

Based on some obtained actual data, the effect of course 

evaluation can be taken through fractional model. It is the 

relative objective, realistic, and highly persuasive. 

A. Each Course Evaluation 

Fractional model of education evaluation can be 

modeled by improved Basset force and fractional model 

in this paper. Basset force mainly describes the process 

that the ball moves in a straight line. And when the ball 

sinks into the viscous fluid, two-phase flow in the actual 

ball movement is not linear motion. The force is impacted 

by other particles movement. The force should be 

connected with particle size, particle and fluid density 

ratio and fluid pulsation frequency so on. The process is 

multi-factor process. Refer to the model of fractional 

Basset force, based on the character of education 

evaluation and a lot of relevant data, fractional education 

evaluation model for one course can be modeled. 

� 

� 9 � 

� � 

�1�2�� 

D x() t �bD x() t �aDx() t � x() t � 1(10) 

where x() t expresses the final evaluation result of the 

course. Fractional coefficient � shows characteristics of 

the teachers. Whether teachers have the experience, 

passion, quality and evaluation of previous classes so on? 

And it also includes the teachers teaching methods and 

means. Fractional factor � shows the overall 

characteristics of the students. It includes the students 

understanding of the course, interest and the past 

accumulation of knowledge so on. Other coefficients are 

determined by teachers and school students. They express 

that students are main body and teacher is the organizer. 

And it fully reflects that the masters of classroom are 

students. The primary role of teachers is to guide and 

answer questions in the classroom. And it should not be 

spoon-fed the traditional speaking. Students’ active 

learning should be mainly part. Coefficient � shows the 

evaluation of students for the course. It is mainly 

confirmed by recognition degree and benefit from the 

course of students. And it fully reflects student initiative. 

The coefficient a expresses the evaluation of course 

teachers. It is mainly determined based on the teaching 

content, teaching hours and teaching arrangement aspect. 

Coefficient b expresses students and teachers for the 

objective evaluation of teaching and learning 

environment. It is decided by the credits set, the college 

supporting degree for the program and class conditions so 

on. 

The fractional simulation block diagram is used for the 

proposed fractional model[20]. Through building 


Simulink model, the numerical solution of fractional 

order nonlinear calculus equation can be obtained directly. 

A fractional calculus module has been mainly adopted. In 

fractional calculus model, a modified approximation 

method is introduced[21]. Based on series expansion and 

recurrence, the continuous rational transfer function is 

2 N 

' 

� � ds �b�hs� 1 �s/ 

�k 

Gs () � ( �� b h) 

� 2 

� 

�d(1 ��) s �b�hs�d� �k��N1 � s / �k 

� 

(11) 

where, 

2 k �� 1 1 

N�k� (1 ��) N�k� (1 ��) 

2N1 2 2 

' �b� � 

2N�1 2N�1 k h b 

� � � � � � 

�d� 2 k �� 1 1 

N�k� (1 ��) N�k� (1 ��) 

2N1 2 2 

�b� � 

2N�1 2N�1 k h b 

� � � � � � 

�d� (12) 

(13) 

where 2N� 1 is the order of approximation, and b , d 

are improvement factor. Here b �10 , d � 9 , N � 3 , 

and the pre-specified frequency range is �b � 0.001 , 

� � 1000 . 

h 

B. Course System Evaluation 

Linear fractional order system is adopted for each 

course assessment result. Output is the final result of the 

evaluation system that we expect. The fractional order is 

important. The different course evaluation results are 

introduced as orders of fractional order systems. Other 

parameters are determined by the characteristics of 

curriculum. 

� � 

aD yt () �aD yt () �� 

1 2 

1 2 

�n 

� aD yt ( ) �ayt ( ) �1 

n 

0 

(14) 

One coefficient ai 

shows the characteristic parameters 

of the course. And � i is the above evaluation result. 

Through a comprehensive study for college courses, we 

can select 

a � , a �10 ��, i�1,2, � , n (15) 

0 1 

i i 

Then the above fractional model can be expressed as: 

� � 

10 � D y( t) �10 � D y( t) 

�� 

1 2 

1 2 

�n 

�10 � D y( t) � y( t) 

�1 

n 

(16) 

Based on the above algorithm, the numerical solutions 

to the linear fractional-order differential equation can be 

obtained with a MATLAB function. Results of curve can 

be achieved by the established fractional model. 

IV. COMPARATIVE SIMULATIONS 

Taken the recent course evaluation as an example, 

there are three courses in software engineering system.


They are Software Testing Technology, Software Testing 

Practice and Software Engineering Practice respectively. 

Firstly each course is analyzed. 

A. Software Testing Technology Evaluation 

Conventional course evaluation score of Software 

Testing Technology is 78 out of a hundred. And then the 

course is evaluated by the proposed fractional model in 

this paper. There are two fractional coefficients in the 

above fractional model, which 

ranges between 0 and 1. 

The coefficient � shows characteristics of the 

teachers. The larger the value of fractional coefficients 

shows that teachers have the higher overall quality of, the 

better characteristics and ways. And it is converse if the 

value is smaller. The fractional coefficient can be 

determined grounded on the above factors, expert ratings 

and data analysis so on. And there is � � 0.8 . The 

fractional order coefficient � is decided due to the past 

content knowledge and overall level of students to 

understand. The coefficient is relatively large if the 

student has accumulated a strong knowledge of related 

courses with active learning. It is converse if the value is 

smaller. According to the previous data analysis of the 

students selected the course, sorting and statistics, 

mathematical statistics and other conv entional methods, 

the 

fractional coefficient can be gotten � � 0.6 . 

The rest factors should be confirmed by teaching 

teachers and school students all together. The 

scope of 

these three factors is 1 to 10. Coefficient � is evaluated 

from the perspective of the students for course. It is 

mainly reflected on students’ satisfaction for the 

curriculum and interest for course content. The value is 

large shows that students are satisfied with the curriculum 

and interested in course content. And they believe that 

teachers have very good teaching ways and means. It is 

converse if the value is smaller. The value is 6 through 

investigating the overall students in the classroom. The 

coefficients a is mainly determined by teachers and 

curriculum. The larger coefficient indicates that teachers 

think that the course is enough for students and the 

content is appropriate for students, whereas the smaller 

the value. The coefficient value is 9 in this course. 

Coefficient b is confirmed from the perspective of 

students and teachers for teaching objective configuration. 

In other words, it reflects that support degree of external 

factors for the course. And it includes the institutions 

support and the school classroom environment. The 

larger coefficient indicates that the course has been 

attached importance to and various external factors are 

favorable for the course. While it is converse if the value 

is smaller. The coefficient is taken as 8 under the joint 

participation 

of all students and Teaching Committee. 

Based on the above data and related methods, the 

fractional program evaluation model can be obtained: 

� 9 � 

� � 

�1�2�6� 0.8 

0.8 

D x t 

0.6 

D x t Dx t x t 


() �8 () �9 () � () � 1 

(17) 

Simulation model can be built by using the above 

algorithm, as shown in Fig. 1. 

Figure 1. Simulink model 

According to the algorithm proposed in this paper, the 

corresponding course curve can be obtained with 

MATLAB, as shown in Fig. 2. 

Figure 2. Output curve of Software Testing Technology 

The result obtained by the proposed method is close to 

0.72 in this figure. It is equivalent to 72 percentile points. 

And it is slightly inferior to the results of traditional 

assessment methods. Through a lot of surveys and 

interviews, it can be found that this method is more 

objective for course evaluation and avoids a lot of human 

factors. 

B. Software Testing Technology Evaluation 

In Software Testing Technology, the coefficient � 

shows characteristics of th e teachers. And it can be 


and data analysis so on. And there is also � � 0.8 . The 



understand. According to the previous data analysis of the 


mathematical statistics 

and other conventional methods,


the fractional coefficient can be gotten � � 0.7 . 

Coefficient � is mainly 

reflected on students’ 

satisfaction for the curriculum and interest for course 

content. Here is � � 7 . 

The coefficients a is 8 in this course. Coefficient b 

reflects that support degree of external factors for the 

course. And it includes the institutions support and the 

school classroom environment. The coefficient 

is taken as 

8 under the joint participation of all students and 

Teaching Committee. 

Then fractional model 

is as follows: 

� 9 � 

� � 

�1�2�7� 

0.8 

0.8 

D x() t 

� � 

0.7 

8 D x() t 8 Dx() t 

� x() t � 1 

(18) 

Simulatio n model 

can be built by using the above 

algorithm, as shown 

in Fig. 3. 


The final result curve 

of the evaluation can 

be gotten 

by building Simulation 

model 

block, as shown in Fig. 4. 

The result is 0.825. 

Figure 4. Output curve of Software Testing Practice 


C. Software Engineering Practice Evaluation 

Course parameters of Software Engineering Practice 

can also be obtained. The coefficient � can be 


and data analysis so on. And there is � � 0.9 . The 



understand. According to the previous data analysis of the 


mathematical statistics and other conventional methods, 

the fractional coefficient can be gotten � � 0.8 , and it is 

high. Coefficient � is evaluated from the perspective of 

the students for course. It is mainly reflected on students’ 

satisfaction for the curriculum and interest for course 

content. The value is 8 through investigating the overall 

students in the classroom. The coefficients a is mainly 

determined by teachers and curriculum. And it is 7 in this 

course. Coefficient b is confirmed from the perspective of 

students and teachers for teaching objective configuration. 

And it includes the institutions support and the school 

classroom environment. The larger coefficient indicates 

that the course has been attached importance to and 

various external factors are favorable for the course. It is 

9 in this course. 

So the fractional model is: 

� 9 � 

� � 

�1�2�8� 0.9 

0.9 

D x t � 

0.8 

D x t � Dx t � x t 

() 9 () 7 () () � 1 


(19) 

Simulation model can be built as shown in Fig. 5. 

Through building Simulation block, the final result curve 

is as shown in Fig. 6. The result is 0.905.


Figure 6. Output curve of Software Engineering Practice 

Analysis: Teachers are not the most important for 

curriculum and student evaluation in education evaluation. 

In contrast the overall quality characteristic of the 

students is the most important for the course evaluation. 

From the view of above three courses, the course 

evaluation result is high when the teacher did not think 

highly of their own but the overall quality of students is 

high and student evaluation is good, such as Software 

Engineering Practice. While the final result of the 

evaluation is not good when the teacher evaluation is high 

whereas the evaluation of the students is low. It can be 

shown in this evaluation model system that external 

factors and students play an important role and the 

proportion of teachers is not large. The actual data and 

results could be more convincing. It is fully consistent 

with the actual teaching activities. And it is shown that 

the fractional order model is valid. 

D. Course System Evaluation 

The fractional order linear model is taken based on 

each course evaluation result. Output is the evaluation of 

the course system. 

�1 �2 

�3 

10 �1D y( t) �10 �2D y( t) �10 �3D 

y( t) � y( t) 

�1 

(20) 

There is a fractional order system founded on the 

above three course evaluation results 

0.91 0.83 0.72 

9.1 D y( t) �8.3 D y( t) �7.2 D y( t) � y( t) 

�1 

(21) 

The result is the output yt ( ) � 0.78 , as shown in Fig. 7. 


Figure 7. Output curve of course system evaluation 

Analysis: It is obviously that the poor evaluation result 

has a great influence on the overall evaluation program in 

the course evaluation system. And it is one of the 

characteristics of the proposed model. It should be taken 

into account the effect of each course not the average. It 

aimed at improving the overall quality and not just 

pursuing the optimal assessment of individual course. 

And it will improve the overall standard of teaching 

rather than creating one or two quality courses. So 

students will get the most benefit. 

The results are consistent with the actual situation, and 

there is more obvious difference in evaluation results. It 

is clear that the proposed method is practical and 

effective. 

V. CONCLUSION 

Education evaluation is a complex multi-factorial 

process, and it is not modeled by integer order model 

accurately. While fractional order system can model the 

complex process. And the corresponding curves can be 

shown in the model block. A fractional modeling method 

of course evaluation is proposed in this paper. The 

fractional model is based on the linear fractional order 

systems. The coefficients of the model are ascertained by 

a large number of related data. Results of the assessment 

can be obtained through MATLAB program. At last, the 

method was applied to actual course assessment instance 

of Capital Normal University Information Engineering 

College. And result indicates that this method is highly 

efficient for solving real-world problems. 

REFERENCES 

[1] W. M. Herbert, “Do university teachers become more 

effective with experience: a multilevel growth model of 

students' evaluations of teaching over 13 years,” Journal of 

Educational Psychology, vol. 99, no.4, pp. 775-790, 2007. 

[2] B. Edith and L. Bernhard, “Academic course evaluation: 

theoretical and empirical distinctions between self-rated


gain in competences and satisfaction with teaching 

behavior,” European Psychologist, vol.14, no.4, pp.297- 

306, 2009. 

[3] S. B. Kimberlee, A. W. Lesley, H. H. Kristin, et al, “The 

efficacy of embedding special education instruction in 

teacher preparation programs in the United States,” 

Teaching and Teacher Education, vol. 24 no.8, pp. 2087- 

2094, 2008. 

[4] C. N. Zhao, F. Pan, and D. Y. Xue, “H∞ controller design 

for fractional order system,” Journal of Northeastern 

University, vol. 27, pp. 1189-1192, 2006. 

[5] D. Y. Xue and C. N. Zhao, “Fractional order PID 

controller design for fractional order system,” Control 

Theory and Applications, vol. 24, pp. 771-776, 2007. 

[6] C. N. Zhao and X. D. Zhang, “The application of 

fractional order PID controller to position 

servomechanism,” IEEE WCICA, p. 3380-3383, 2008. 

[7] D. Y. Xue, C. N. Zhao, and Y. Q. Chen, “Fractional Order 

PID Control of A DC-Motor with Elastic Shaft: A Case 

Study,” American Control Conference, p. 3182-3187, 2006. 

[8] O. Pontus and E. Lars, “Estimation of absorbed PAR 

across Scandinavia from satellite measurements. Part II: 

Modeling and evaluating the fractional absorption,” 

Remote Sensing of Environment, vol. 110, pp. 240-251, 

2007. 

[9] S. Hapca, J. W. Crawford, K. MacMillan, et al, “Modelling 

nematode movement using time-fractional dynamics,” 

Journal of Theoretical Biology, vol. 248, pp. 212-224, 

2007. 

[10] W. M. Ahmad and R. El-Khazali, “Fractional-order 

dynamical models of love,” Chaos, Solitons & Fractals, 

vol. 33, pp. 1367-1375, 2007. 

[11] J. W. Richard, F. M. William, H. Jenise, et al, “Modeling 

fractional crystallization of group IVB iron meteorites,” 

Geochimica et Cosmochimica Acta, vol.72, p. 2198-2216, 

2008. 

[12] M. C. Guglielmo and A. G. Luis, “Modelling the US, UK 

and Japanese unemployment rates: Fractional integration 

and structural breaks,” Computational Statistics & Data 

[18] S. E. Hamamci, M. Koksal, “Calculation of all stabilizing 

fractional-order PD controllers for integrating time delay 

systems,” Computers & Mathematics with Applications, 

Vol. 59, no. 5, pp. 1621-1629, 2010. 

[19] A. Oustaloup, F. Levron, B. Mathieu, and F.M. Nanot, 

“Frequency band complex noninteger differentiator: 

characterization and synthesis,” IEEE Trans. Circuits Syst. 

I, vol. 47, pp. 25-39, 2000. 

[20] C. N. Zhao, Y. Zhao, Y. Liu, Y. S. Li, L. M. Luo, 

“Fractional Personnel Losing Modeling Approach and 

Application,” IEEE CISE, p. 14-17, 2009. 

[21] D. Y. Xue, C. N. Zhao, and Y. Q. Chen, “Modified 

Approximation Method of Fractional Order System,” IEEE 

ICMA, p. 1043-1048, 2006. 

Chunna Zhao was born in Liaoning 

Province, China, in 1978. She received 

her B.S. degree in Liaoning Normal 

University, Dalian, in 2001. She received 

her M.S. and Ph.D. degree in 

Northeastern University, Shenyang, in 

2004 and 2006 respectively. She studied 

in Northeastern University postdoctoral 

from 2006 to 2008. Her current research 

interests include evaluation, data analysis, fractional systems 

and mathematics modeling. Her address is with Capital Normal 

University, Information Engineering College, Beijing, 100048, 

China. (Corresponding author to provide phone: 86-10- 

68901370-213 & 15011489020; e-mail: 

chunnazhao@163.com.) 

Yu Zhao was born in Liaoning Province, China, in 1977. He 

received his B.S. degree in Shenyang University, Shenyang, in 

2006. She received her M.S. degree in yunnan university of 

finance and economics, Kunming, in 2008. His current research 

interests include evaluation, data analysis, logistics management. 

Analysis, vol.52, pp. 4998-5013, 2008. His address is with Yunnan Technician College, Kunming, 

[13] D. Y. Xue, C. N. Zhao, and F. Pan, “Simulation model 

method and application of fractional order nonlinear 

system,” Journal of System Simulation, pp. 2405-2408, 

China. Email: yuzhao66@163.com. 

2006. 

Liming Luo was born in Beijing, China, in 1965. He received 

[14] C. N. Zhao, X. D. Zhang, and Y. R. Sun, “Simulation of his M.S. degree in Capital Normal University, Beijing, in 2008. 

commensurate fractional order systems,” Journal of System His current research interests include evaluation, data analysis. 

Simulation, pp. 3948-3950, 2008. 

His address is with Capital Normal University, Information 

[15] A. Dzielinski, D. Sierociuk, “Ultracapacitor Modelling and Engineering College, Beijing, 100048, China. Email: 

Control Using Discrete Fractional Order State-Space 

Model,” Acta Montanistica Slovaca, vol. 13, no.1, pp. 

208–216, 2008. 

luolm@mail.cnu.edu.cn. 

[16] S. Y.n Xing, J. G. Lu, “Robust stability and stabilization of Yingshun Li was born in Liaoning Province, China, in 1971. 

fractional-order linear systems with nonlinear uncertain She received her B.S. degree in 1993. She received her M.S. in 

parameters: An LMI approach,” Chaos, Solitons & Dalian University of Technology in 2000. And she received her 

Fractals, Vol. 42, no. 2, pp. 1163-1169, 2009. 

Ph.D. degree in Northeastern University, Shenyang, in 2006. 

[17] L. M. Richard, J. R. Thomas, “Fractional-order elastic Her research interest covers intelligent system, pattern 

models of cartilage: A multi-scale approach,” recognition and chemical automation. Her address is with 

Communications in Nonlinear Science and Numerical Engineering College, Shenyang University of Technology, 

Simulation, vol. 15, no. 3, pp. 657-664, 2010. 

Liaoyang, China. Email:liyingshunok@163.com. 



Immune Genetic Evolutionary Algorithm of 

Wavelet Neural Network to Predict the 

Performance in the Centrifugal Compressor and 

Research 


Department of Mathematics and Computer Science, Liuzhou Teachers College 

Liuzhou, Guangxi, 545004, China 

gxhsz@126.com 

Abstract—Prediction of the performance of centrifugal 

compressors, the traditional methods using BP neural 

network. This single neural network for forecasting problem 

is not high enough precision, slow convergence and easy to 

fall into local optimal solution. In order to more accurately 

predict the performance of centrifugal compressors, the 

implicit commit identify problems early. Are the immune 

algorithm, genetic algorithm, wavelet theory, the 

combination of neural networks, established immune 

genetic algorithm optimization of wavelet neural network 

model (IGA-WNN). Realized to predict the performance of 

centrifugal compressor, and the predicted results with the 

BP neural network model prediction results and the wavelet 

neural network model prediction results were compared. 

Simulation results show that: the prediction model, can 

achieve the centrifugal compressor performance prediction 

and monitoring. Which, IGA-WNN optimal prediction 

results: with a simple algorithm, structural stability, the 

convergence speed and generalization ability of the 

advantages of prediction accuracy of 99% over traditional 

methods of prediction accuracy of 15%, with a certain 

Theoretical study and practical value. 

Index Terms—immune algorithm, genetic algorithm, 

wavelet theory, centrifugal compressor, performance 

prediction. 


As centrifugal compressor owns the features of high 

rotating speed, small size, small occupancy area, etc, 

therefore, it is widely used in the petrochemical industry. 

The following stages in the process of new product 

development for centrifugal compressor are needed: 

theoretical design, modeling production, performance test 

and prototype production. When designing, we must test 

on centrifugal compressor’s performance, which can help 

us know whether the centrifugal compressor meet the 

requirements or not. If the result is not good, then a 

further optimization is needed. However, this approach 

will inevitably need large input of manpower, material 

and financial resources, so applying a new method can 

predict the performance of centrifugal compressor before 

designed to provide theoretical guidance for the optimal 

design. Artificial neural networks can train the existing 


doi:10.4304/jsw.6.5.908-914 

samples in order to predict the conditions of centrifugal 

compressor’s best performance, shorten the testing time, 

increase the efficiency and cut down the cost. In the past, 

the BP neural network approach is often considered as the 

main network approach, however, for complex 

centrifugal compressors, in order to obtain a better 

prediction, the further optimization is a must. When 

conventional neural network recognizes centrifugal 

compressor, there exists the defects with low 

identification accuracy, poor convergence speed, and 

easily falling into local optimal solution. Therefore, it’s 

necessary to seek for an optimal identification model with 

high identification accuracy, fast convergence speed and 

strong global searching optimization, which is significant 

for the research and understanding of centrifugal 

compressor’s fault and performance. 

To solve these problems, the immune genetic 

algorithm is proposed wavelet neural network model 

(IGA-WNN). Advantages of this model are: 

(1) structure determination, avoiding the blindness in 

the structure design of BP network; 

(2) the linear distribution of network weight factors 

and the convexity of study objective function 

fundamentally prevent issues such as the local 

optimization in network training process; 

(3)simple concept of algorithm, faster convergence 

speed; 

(4) strong function learning ability, able to approach 

any nonlinear function in high precision. 

The results show that: the immune genetic algorithm 

for wavelet neural network model (IGA-WNN) 

prediction accuracy 99%. 

II. IMMUNE GENETIC ALGORITHM OPTIMIZATION OF 

WAVELET NEURAL NETWORK MODEL (IGA-WNN) 

Wavelet neural network WNN is based on the wavelet 

theory, put forward artificial neural network as a 

feedforward network. It is based on the wavelet function 

as the neuron activation function, wavelet scaling, 

translation factors, as well as connection weights, the 

error energy function in the optimization process to be 

adaptive. Use of models such as (1) type shown.


M 

y( x) 

= f ( ∑ω 

jϕ 

a, 

b[ 

∑v 

ji xi 

]) 

(1) 

j= 

1 

N 

i= 

1 

Type in xi (i = 1,2,3, ... N) is the i-input; y (x) for the 

network output; ωj first j for the hidden layer nodes to 

output layer weights; vji for the first i- input to the j-node 

hidden layer weights; f (•) for the hidden layer to output 

layer activation function; ϕ a, 

b ( •) 

for the wavelet 

function, a, b are scaling and wavelet function for the 

translation factor; N input number; M is the number of 

hidden layer neurons 

Error energy function is: 

1 ∑= 

K 

2 

E = [ T − f ( xk 

)] 

(2) 

2 k 1 

Type in: x K as the sample; K for the sample number. 

Wavelet theory, Zhang and Benveniste first proposed 

wavelet neural network (WNN), its forward neural 

networks as a new form of training with back propagation 

algorithm to approximate WNN arbitrary nonlinear 

function. Wavelet neural network (WNN)’s structure 

shown in Figure 1. 

Figure1. Schematic diagram of wavelet neural network 

Wavelet neural network is basically divided into three 

layers: input layer, wavelet function and output layers. 

The output expression is: 

( _ 

y 

m 

∑ 

j 

j 

_ 

= w jψ x) 

(3) 

Type in: x = x1 

..., x i ,..., xm 

, ψ j That the wavelet 

function layer node function, w j equivalent to the 

wavelet coefficients after the wavelet decomposition.For 

a multi-variable nonlinear coupled system, need to build 

multi-dimensional wavelet function as the wavelet layer 

WNN node, where more than scalar wavelet function 

using the product to build multidimensional wavelet 

function.Therefore, WNN output expression to read: 

n m xi 

− bij 

y = ∑ w j∏ 

( ) 

a 

j 

ψ (4) 

i ij 

Among them, the scalar function ψ, multi-dimensional 

wavelet function in wavelet layer node number n, 

coefficient of Wj, scale parameters a ij and b ij are the 

translation parameters need to be identified and 


estimated.Here choose ψ (x) =- xexp (-1 / 2x2). For the 

parameters n, a ij and b ij , using immune algorithm is 

estimated that in the course of the iterative algorithm for 

updating coefficients Wj. 

Immune algorithm is a combination of deterministic 

and stochastic selection heuristic random search 

algorithm, which is considered to be adaptive immune 

response in a simple simulation, this study of antibodies 

and antigen response process to complete. Immune 

algorithm includes initialization antibody group, clonal 

selection, antibody clones, affinity mutation, clone 

inhibition, immune selection, new members of the other 

steps. each step corresponds to an evolutionary 

mechanism for the immune system. The evolution 

process of antibody optimization problem as a candidate 

solution, antigen as the optimal solution, process of 

finding the optimal solution can be achieved by the 

immune algorithm. 

For a set of training data St={( xk , yk ), k = 1,…, nd} , 

Objective function obtained here: 

( a ) ( b ) , n) 

nd 

1 

∧ ⎛ ⎞ 

J ij , = ∑ ⎜ − ⎟ × ij 

y 

m n m× 

n 

k y (5) 

k 

n ⎝ ⎠ 

Where, k 

y∧ is the y k estimate. 

Calculated using the following formula between the 

antibody and antigen affinity: 

f 

aff 

( a ) , ( b ) , n) 

ij m× 

n 

= 

1+ 

exp 

( η J ( a ) , ( b ) , n ) 

aff 

ij m× 

n 

1 

ij m× 

n 

d 

k 

ij m× 

n 

Where, 0


Step 4: According to the affinity of antibodies size, 

degree of similarity and incentive on the antibody clone, 

mutation, suppression and selection; 

Step 5: Obtained from step 4, remove the new antibody 

population in one of the largest affinity antibodies to 

determine the antibody corresponding objective function 

value is less than the set value, if less than, the end of the 

algorithm; Otherwise, return to step 2. 

Genetic Algorithm (GA) is a global optimization 

algorithm based on natural selection and natural heredity, 

and it uses the selection, crossover and mutation of the 

three genetic operators abstracted from the mechanism of 

natural selection to operate on the parameters encoding 

string, which features with overall importance, rapidity, 

good adaptability and strong robustness, etc and is able to 

achieve global search in complex, multimodal, nonlinear 

and non-differentiable space. The general steps of using it 

to solve problems are: 

Step1:conduct chromosome bit string encoding on the 

parameters that need to be optimized; 

Step2:generate initial population; 

Step3: evaluate the group and solute the fitness value 

of each individual; 

Step4: apply genetic operators on groups and generate 

new population, executing circularly. Specific genetic 

manipulation includes selection, crossover and mutation. 

When carrying out network training, the author applies 

genetic algorithm to optimize the network parameters of 

wavelet neural network predictive model and determine 

hidden layer nodes---the number of wavelet in order to 

simplify network structure and improve the accuracy, 

adaptability and robustness of anode effect prediction. 

The author uses genetic algorithm optimization 

approach to identify the number of hidden nodes is the 

number of wavelet basis, while the second part is the 

training method of training: first with the genetic 

algorithm to train the network's scale factor, translation 

factor and weights the same time determine the wavelet 

Number, and then use gradient descent on the scale, 

translation factors and weights, the threshold for 

secondary training. The specific steps: 

1. encode 

For chromosomes to mixed binary and real coded form 

of the binary code to each network that the effectiveness 

of Cain layer unit, randomly generated N (determined 

based on experience) one structure, each individual 

corresponds to a structure, the wavelet network The 

number of input and output nodes to reflect the anode 

effect by the number of parameters determined by the 

actual situation, when encoding only the number of 

hidden layer wavelet encoding, 0 for invalid connections, 

1 connection is valid;On the network weights, scale 

factor, translation factor to real coding, the wavelet 

corresponding to each individual network, weight, scale 

factor, translation factor compiled a code string sequence 

as a gene. Code string obtained by mixing the form of 

coding 

h h h h h h h h h h ο ο ο 

φ w ⋅⋅wa 

b ⋅⋅φHw 

⋅⋅ 

11 11 1 1 H1 

wH 

1aHb 

Hφ 

1 

w ⋅⋅ 

Ο 11 wH 

1 

1…0…1 (10) 


Where: h is hidden, O for the output layer. 

2. Swarm initialization 

For the initialization for weights, threshold, scale 

factor and shift factor, the initialization interval is [-1, 1]. 

Set the population size to S, crossover probability to Pc, 

mutation probability to Pm, and the maximum number of 

genetic iterations to Gmax. 

3. Calculation of individual fitness 

Evaluate the training results according to the fitness 

function, and the evaluation evidence is: 

f 

1 

= 

1+ 

E 

(11) 

4. determine whether the fitness value meets the 

overall requirements, if it satisfies, refer to 6, otherwise 

execute 5. 

5. Selection, crossover and mutation 

Adopt the roulette selection method to get a large 

number of individuals with strong fitness, and then use 

them to directly copy the next generation. At the same 

time, carry out crossover and mutation on the individuals 

which need to be conducted these operations according to 

the crossover probability Pc and mutation probability Pm 

in order to generate the next generation, then refer to 3, to 

evaluate its fitness value. 

6. if it meets the pre-conditions or maximum 

iterations, end the loop to obtain the optimal chromosome 

and decode it into corresponding number of hidden layer 

wavelet basis, weight, scale factor and shift factor. 

7. Secondary Training 

After determining the number of hidden layer network 

nodes, use gradient descent algorithm to conduct 

secondary optimization training on parameters such as 

weight, threshold, scale factor, scale factor and shift 

factor. The parameters obtained after two optimizations 

will be used as the final parameters of genetic wavelet 

network prediction model, and furthermore used to 

predict the performance of centrifugal compressors. 

III. THE EXTRACTION OF MODEL IDENTIFICATION 

PARAMETER 

The factors affecting the performance of centrifugal 

compressor own the following performance parameters: 

flow(Q), blade incidence(βA) , rotational speed(n) , 

number of leaves(Z), number of guide vanes(Z1), 

impeller diameter(D), ratio of impeller outer and inner 

diameter(D1/D2), the impeller outlet width(b2), pressure 

ratio(Pd/Ps), etc. Simply calculating, select 4 main 

performance parameters: flow(Q), installation angle of 

the impeller outlet (β2A), pressure ratio(ε=Pd/Ps)and 

efficiency(η)to establish predictive model. Research on 

the centrifugal compressor with the same rotate speed, 

and take the pressure ration corresponding to 4 different 

installation angles of the impeller outlet(β2A) and the 

performance curve of efficiency changing with the flow. 

Utilize partial data of these known performance curve to 

conduct training for BP neural network, and then make 

use of other data to test the network prediction ability that 

has been trained. According to the efficiency and pressure 

ratio of 16 flow points of each blade angle obtained


through 4 performance testing data, divide the data into 

groups, as shown in table 1. These data constitute 64 

pairs of sample points. The input mode of each sample is 

constituted by 1 angle value and 1 flow value, while the 

output mode is constituted by 1 efficiency or pressure 

β2A° 

30 

30 

Q 

m 3 /s 

1.52 

1.51 

1.50 

1.49 

1.48 

1.47 

1.46 

1.45 

1.44 

1.43 

1.52 

1.51 

1.50 

1.49 

1.48 

1.47 

ε 

1.21 

1.31 

1.40 

1.50 

1.66 

1.71 

1.80 

1.91 

2.00 

2.10 

1.21 

1.32 

1.34 

1.47 

1.58 

1.66 

η 

% β2A° 

71.00 

73.00 

75.00 

77.00 

79.00 

81.00 

83.00 

84.00 

86.00 

89.00 

71.00 

73.00 

75.00 

77.00 

79.00 

81.00 

32 

32 

Table 1 

Sample Centrifugal Compressor Performance Prediction 

Q 

m 3 /s 

1.67 

1.66 

1.65 

1.64 

1.63 

1.62 

1.61 

1.60 

1.59 

1.58 

1.67 

1.66 

1.65 

1.64 

1.63 

1.62 

IV. SIMULATION RESULTS 

ε 

1.31 

1.41 

1.60 

1.70 

1.81 

1.90 

2.00 

2.11 

2.30 

2.40 

1.32 

1.41 

1.65 

1.73 

1.82 

1.96 

η 

% 

74.00 

76.00 

77.00 

78.00 

80.00 

81.00 

82.00 

83.00 

84.00 

86.00 

74.00 

76.00 

77.00 

78.00 

80.00 

81.00 

Immune genetic algorithm to optimize the initial 

antibody population or the population size S = 50, the 

maximum number of iterations Gmax = 10000, training 

for gradient descent learning rate η = 0.85, momentum 

factor α = 0.921, learning error to set ε = 0.001 , The 

maximum number of learning steps epoch = 10, 

crossover probability Pc = 0.25, mutation probability Pm 

= 0.01, the maximum number of learning steps epoch = 

10. For the learning rate η, used in the training process 

adaptive method for faster convergence rate and improve 

the prediction of real-time. 

To paragraph (1) group of 40 samples of data input, 

after several training comparison, the number of hidden 

layer neurons is taken as 10, to build the prediction 

model. To subsection (2) group of 24 samples to be 

predicted forecast data generation model identification, in 

the iteration stops after 4800 time iterations, the correct 

identification rate of 99%, At this point, pressure 

ratio( ε ) and efficiency ( η )errors are 1.0406 × 10 -6 , 

1.0386×10 -6 ; Recognition rate WNN model is greater 

than 90%; BP network model is over 85% recognition 

accuracy. Figure 2 and Figure 3 is IGA-WNN, WNN and 

BP training error comparison chart, we can see from the 

figure: IGA-WNN's convergence speed. 


ratio. Divide the samples into 2 groups: there are 40 

samples in group (1), which are mainly used in the 

network training to establish predictive model; there are 

24 samples in group (2) used to test the predictive model. 

β2A° 

34 

34 

Q 

m 3 /s 

1.84 

1.83 

1.82 

1.81 

1.80 

1.79 

1.78 

1.77 

1.76 

1.75 

2.01 

2.00 

1.99 

1.98 

1.97 

1.86 

ε 

1.42 

1.52 

1.71 

1.90 

2.00 

2.10 

2.20 

2.31 

2.40 

2.61 

1.35 

1.53 

1.74 

1.95 

2.02 

2.14 

η 

% 

78.00 

80.00 

81.00 

82.00 

83.00 

85.00 

86.00 

87.00 

88.00 

89.00 

82.00 

83.00 

84.00 

86.00 

87.00 

88.00 

β2A° 

36 

36 

Q 

m 3 /s 

2.01 

2.00 

1.99 

1.98 

1.97 

1.96 

1.95 

1.94 

1.93 

1.92 

2.06 

2.04 

1.99 

1.96 

1.93 

1.88 

ε 

1.61 

1.72 

1.80 

1.90 

2.12 

2.33 

2.43 

2.52 

2.61 

2.70 

1.34 

1.56 

1.71 

1.85 

2.10 

2.16 

Figure 2.ε training error comparison chart 

Figure 3. η training error comparison chart 

η 

% 

82.00 

83.00 

84.00 

86.00 

87.00 

88.00 

89.00 

90.00 

91.00 

92.00 

85.00 

86.00 

87.00 

88.00 

89.00 

90.00 

In order to better understand and explain the immune 

genetic algorithm neural network model to predict the 

performance advantages of centrifugal compressors. Of


the study, also using BP neural network and wavelet 

neural network WNN data were processed using the same 

training samples, and establish a three-layer neural 

network model, the number of hidden neurons is set to 

1O, and the immune genetic Algorithm optimization of 

wavelet neural network model has the same structure of 

the network, the training sample to predict recognition. 

The simulation results from Matlab, you can obviously 

see, IGA-WNN and BP in the 4800 iteration of the 

changes within the training error, IGA-WNN method of 

convergence speed and accuracy of better than WNN 

method; WNN method of convergence Speed and 

accuracy is obviously better than the BP method. Table 2 

Table 2 

Pressure ratio (ε ) and Efficiency (η ) prediction table 

shows the centrifugal compressor pressure ratio (ε ) and 

efficiency ( η ) predictions. 

In order to better observe the centrifugal compressor 

pressure ratio (ε ) and efficiency (η ) of the predicted 

results, the measured value and the IGA-WNN, WNN, 

BP predicted values with the scatter plot to display. It is 

clear that: IGA-WNN best prediction. Figure 4, Figure 5, 

Figure 6, Figure 7, Figure 8 and Figure 9. 

β2A° ε BP WNN IGA-WNN Q(m 3 /s) η (%) BP WNN IGA-WNN 

30 

32 

34 

36 

1.21 

1.32 

1.34 

1.47 

1.58 

1.66 

1.32 

1.35 

1.34 

1.73 

1.82 

1.96 

1.35 

1.53 

1.74 

1.95 

2.02 

2.14 

1.34 

1.56 

1.73 

1.85 

2.10 

2.16 

1.01 

1.30 

1.40 

1.50 

1.61 

1.68 

1.25 

1.40 

1.41 

1.71 

1.80 

1.81 

1.40 

1.46 

1.71 

1.91 

1.92 

2.10 

1.42 

1.53 

1.71 

1.81 

2.07 

1.91 

1.16 

1.33 

1.38 

1.50 

1.61 

1.69 

1.32 

1.39 

1.39 

1.74 

1.83 

1.93 

1.39 

1.53 

1.75 

1.96 

2.01 

2.14 

1.39 

1.57 

1.74 

1.85 

2.11 

2.10 

1.21 

1.32 

1.35 

1.47 

1.58 

1.66 

1.32 

1.36 

1.35 

1.73 

1.82 

1.95 

1.36 

1.53 

1.74 

1.95 

2.02 

2.13 

1.34 

1.56 

1.73 

1.84 

2.10 

2.16 

Figure4. BP to centrifugal compressor's pressure ratio forecast contrast 

chart 


1.52 

1.51 

1.50 

1.49 

1.48 

1.47 

2.01 

2.00 

1.99 

1.98 

1.97 

1.86 

1.67 

1.66 

1.65 

1.64 

1.63 

1.62 

2.06 

2.04 

1.99 

1.96 

1.93 

1.91 

71.00 

73.00 

75.00 

77.00 

79.00 

81.00 

82.00 

83.00 

84.00 

86.00 

87.00 

88.00 

74.00 

76.00 

77.00 

78.00 

80.00 

81.00 

85.00 

86.00 

87.00 

88.00 

89.00 

90.00 

71.30 

73.10 

75.40 

77.30 

79.60 

81.20 

82.40 

83.20 

84.50 

86.40 

87.10 

88.30 

74.20 

76.50 

77.40 

78.50 

80.20 

81.50 

84.40 

85.20 

86.50 

87.40 

88.50 

89.30 

71.12 

73.06 

75.16 

77.12 

79.23 

81.09 

82.15 

83.09 

84.19 

86.15 

87.05 

88.12 

74.09 

76.18 

77.16 

78.19 

80.09 

81.19 

84.82 

85.75 

86.86 

87.82 

88.85 

89.79 

71.01 

73.01 

75.02 

77.00 

79.02 

81.00 

82.00 

83.01 

84.02 

86.00 

87.00 

87.99 

74.00 

75.99 

77.01 

78.01 

80.00 

81.02 

84.99 

85.99 

87.01 

88.00 

89.00 

90.01 

Figure5. BP to centrifugal compressor's potency forecast contrast chart


Figure6. WNN to centrifugal compressor's pressure ratio forecast 

contrast chart 

Figure7. WNN to centrifugal compressor's potency forecast contrast 

chart 

Figure8. IGA-WNN to centrifugal compressor's pressure ratio forecast 



Figure9. IGA-WNN to centrifugal compressor's potency forecast 



Simulation Description: immune genetic algorithm 

neural network model(IGA-WNN) with structural 

stability, simple algorithm, global search capability and 

convergence speed, generalization ability, etc., can reflect 

the nonlinear centrifugal compressor performance 

Problem. The measured performance of the centrifugal 

compressor 40 samples for training, 24 samples to 

predict, predict the correct rate of 99%. Predicted effect, 

centrifugal compressor performance monitoring presents 

a fast convergence and high accuracy, low cost 

performance prediction model developed in the 

experiment, the centrifugal compressor to improve 

efficiency, reduce costs and has practical value to the 

similar Problem with some guidance. 

REFERENCES 

[1] Yue bangguo, “Talking about bedpan’s flushing filth 

function and saving water,” Chinaware, 2001(5).(in 

Chinese) 

[2] Wu ziniu, The basic elements of computing 

hydromechanics. Beijing: the Science Press. 2001. 1-8. (in 

Chinese) 

[3] Hong fangwen, “Kinetic object round free surface flowing 

field value simulation and experimentation research,” 

Doctor Degree thesis. Wuxi, Jiangsu, China: Chinese 

watercraft science research center. 2001.4. (in Chinese) 

[4] C.W. Hirt, B.D. Nichols. Volume of fluid (VOF) method 

for the dynamics of free boundaries. Journal of 

Computational Physics, 1981, 39:201-225. 

[5] Lin huzong, Phantasmagoric flowing science-multiphase 

hydromechanics. Beijing: the book concern of Qinghua 

University, Guangzhou: the book concern of Jinan 

University. (in Chinese). 

[6] Wang Xi Feng, Gao Ling, Zhang twins. based on wavelet 

technology for network traffic analysis and forecasting [J]. 

Computer Applications and Software, 2008,25 (8) :70-72. 

(in Chinese)


[7] Liu Yuan, Dai Yue, Jian-Hua Cao. based on wavelet neural 

network traffic prediction of chaotic time series [J]. 

Computer Engineering, 2008, 34 (16): 105 - 107. (in 

Chinese). 

[8] SOUTHEAST UNIVERSITY Hou Jun, Jiang Shuyan and 

so on. based on wavelet analysis and hierarchical decisionmaking 

method for analog circuit fault diagnosis [J]. 

Computer Applications, 2010,27 (6) :2057-2060. (in 

Chinese). 

[9] Li Li, Li Hongqi, Xieshao Long and so on. clonal selection 

based on immune particle optimization algorithm [J]. 

Computer Science, 2008, 35 (10): 253 - 278. (in Chinese) 

[10] Chenzhen Wei, Zheng Guo crisis. Wavelet neural network 

predictive model Simulation [J]. Computer Simulation, 

2008, 25 (6): 147-150(in Chinese). 

[11] Zhourong Yi, Li Shuqing, cows will never , Wavelet 

Neural Network on Safety Management Evaluation[J], 

Coal Science and Technology,2006(5):67-70(in Chinese). 

[12] Lidi Wei, Shao-hua, SHEN Yuan-tong. PSO based wavelet 

neural network [J]. Engineering Geophysics, 2007, 4 (6): 

529-532(in Chinese). 

[13] F Calise, et al. Simulation and exergy analysis of a hybrid 

SolidOxide Fuel Cell (SOFC)- Gas Turbine System [ J ]. 

Energy,2006, 31 (15) : 3278 - 3299. 

[14] V M Janardhanan, V Heuveline, O Deutschmann. 

Performance analysis of a SOFC under direct internal 

reforming conditions[J].Journal of Power 

Sources,2007,172(1):296-307. 

[15] Ko Mei Jing, Hu Liming. wavelet neural network based on 

genetic pressure sensor nonlinear correction [J]. Actuators, 

2007,20 (4) :816-819(in Chinese). 

[16] Cao there. an immune particle swarm optimization 

algorithm and the wavelet neural network learning 

application [J]. Computer Applications and Software, 

2009,26 (6) :189-191(in Chinese). 

[17] Zhao Guorong, Wang Xibin, Gao Qing Wei , PSO neural 

network transfer of INS Alignment[J], System 

Simulation,2010,22(3):670-673 (in Chinese). 


Author: Shengzhong Huang, 1957, 

male, Guangxi Hezhou, associate professor. 

Research Interests: Computer Intelligence 

algorithms and digital image processing. In 

ten years teaching computer applications, 

published more than 20. 

Funding Project: New Project For 

Guangxi Higher Education Reform 

(2010JGB135).


Development of Optimization Design Software 

for Bevel Gear Based on Integer Serial Number 


Xiaoqin Zhang 

Mechanical and Electrical Engineering College 

Hebei Normal University of Science & Technology, Qinhuangdao, China 

Email: zxqwlc@163.com 

Yu Rong, Jingjing Yu, Liling Zhang and Lina Cui 

Mechanical and Electrical Engineering College 

Hebei Normal University of Science & Technology, Qinhuangdao, China 

Email: lixiangcg@126.com 

Abstract—Bevel gear drive is widely used, quality of which 

not only affects its own transmission performance, size and 

weight, but also has some impact on the machine's 

performance. This paper introduces optimization design 

software for bevel gear, in which automatic optimization 

design is realized. In the paper mathematical model, 

programming of design data and realization of optimization 

design based on genetic algorithm are described in detail. 

The paper proposed integer serial number encoding genetic 

algorithm, which effectively deals with continuous and 

discrete variable optimization problem and reduces the code 

length of the string to improve the encoding and decoding 

efficiency, no invalid solution or duplicate solutions. 

Index Terms—bevel gear, optimization design, genetic 

algorithm, integer serial number encoding, augmented 

penalty function 


Bevel gear drive, characterized by changing direction, 

high coincidence and smooth transmission, etc., is widely 

used in the aerospace, automotive and large mechanical 

transmission system. So its design quality not only affect 

its own transmission performance, size and weight but 

also have some impact on the machine's performance. In 

practical engineering design, involving many parameters, 

consuming much calculation time, prone to error, and 

repeated calculation, query and drawing are needed for 

series of product design, resulting in substantial 

duplication of effort, so the development of bevel gear 

design software finding the optimal design is of great 

significance [1] [2] [3]. 

Optimization design of bear gear includes the 

continuous variables and discrete variables. The 

traditional method is to round the optimal design to the 

adjacent discrete points. Thus, design point might run out 

Manuscript received January 1, 2010; revised December 10, 2010; 

accepted January 8, 2011 

Corresponding author: Zhang Xiaoqin 


doi:10.4304/jsw.6.5.915-922 

of the feasible region, besides traditional optimization 

method mostly bases on gradient algorithm, which is 

likely trapped into local minimum search. Many studies 

indicate that the genetic algorithm has strong ability of 

general optimization, which is very effective in treating 

optimization problem containing continuous and discrete 

variables[4][5]. 

This paper proposes the optimization design software 

for straight tooth bevel gear based on genetic algorithm 

encoding with integer serial number, no invalid solution 

or duplicate solutions, the code length of the string 

reduced, the encoding and decoding efficiency improved. 

II. SOFTWARE FUNCTION 

The software can accomplish strength-calculation and 

optimal design of straight tooth bevel gear based on GB/T 

Calculation Methods of Load Capacity for Bevel Gear. 

Moreover friendly user interface is developed using VB 

language shown in Fig. 1. Input the known data, click on 

the button "optimization calculation", the user can get the 

optimization results of straight tooth bevel gear. Click on 

the button "gear materials and fatigue limit query", the 

user can get the corresponding fatigue limit of gear 

materials, shown in Fig. 2. Click on the button 

“coefficient calculation”, various coefficients can be 

calculated, shown in Fig. 3. For example, select Yfsa and 

click on the button “calculate coefficient”, then the 

interface for calculating Yfsa appear as shown in Fig. 4. 

Input the known data and click on the calculating button, 

the results of recombination tooth coefficient for pinion 

and gear are gained. 

III. CRITICAL TECHNOLOGY 

A. Establishing Mathematical Model 

Establishing mathematical model is prime step for gear 

design, in which the design variables, objective function


Figure 1. User interface 

Figure 2. Gear material 

and constraints are determined through studying gear 

design theory. 

B. Programming of A Large Number of Graph and Table 

In gear design process, involving a large number of 

graph and table data, how to invert artificial seeking into 

automatically calculation is a basic problem in the 

process of CAD operations. 

C. Genetic Algorithm Realizing 

The augmented penalty function and integer serial 

number encoding is used in genetic algorithm. 


Figure 3. Coefficient calculation 

Figure 4. Yfsa calculation 

Software design flow chart is shown as Fig. 5 

IV. MATHEMATICAL MODEL FOR OPTIMIZATION DESIGN 

Optimization objects for straight tooth bevel gear are 

various. If user requirements needed, the consumption of 

material is the measurement of design. Therefore the 

optimization object is the volume of frustum of cone of 

the bevel gear pair [2][6][7]. 

A. Object Function 

Bevel Gear volume are the sum of pinion volume and 

gear volume, while each bevel gear volume is similar to 

the volume of frustum of cone between the big end and 

small end of the pitch circle. Therefore, according to the 

volume formula of frustum of cone, volume calculation 

formula of straight tooth bevel gear pair can be expressed 

as: 

V = V1+ 

V2 

π ⎡ mz1 

2 mz1 

R −b 

mz1 

R −b 

mz1 

2⎤ 

= bcosδ1⎢( 

) + ( × ) + ( × ) ⎥ + 

3 ⎣ 2 2 R 2 R 2 ⎦ 

π ⎡ mz2 

2 mz2 

R −b 

mz2 

R −b 

mz2 

2⎤ 

bcosδ2 

⎢( 

) + ( × ) + ( × ) 

3 

⎥ 

⎣ 2 2 R 2 R 2 ⎦ 

(1) 

Where, b is the face width, δ1, δ2 are respectively cone 

angle of pinion and gear, m is modulus, z1 is tooth 

number of pinion, R is cone pitch, z2 is tooth number of 

gear. 

Figure 5. Software design flow chart


B. Design Variable 

The independent design parameters of volume of 

straight bevel gear drive include big end module m, tooth 

number of pinion z1, face width coefficient φR, so the 

design variables are. 

X = [m,z1,φR] T = [x1,x2,x3] T (2) 

C. Constraint Conditions 

1) Contact strength conditions 

The bending stress σH and contact stress σF of gears 

should be not more than the allowable value, i.e. 

4. 

7KT1 

σ H = Z E Z H Z ε ≤ [ σ ] 

2 3 3 H 

ϕ R ( 1− 

0. 

5ϕ 

R ) m z1 

u 

(3) 

Where, ZE is the elastic coefficient; ZH is regional 

coefficient of pitch point, for straight gear ZH = 2.5; Zε is 

the coincidence coefficient; K is load coefficient; T1 is 

input torque; [σH] is the allowable contact stress. 

2) Tooth root bending strength conditions 

σ 

4. 

7KT1Y 

= 

2 

ϕ ( 1− 

0. 

5ϕ 

) z 

≤ [ σ ] (4) 

FSa1 

ε 

F1 

R 

R 

2 3 

1 m 

2 

u + 1 

F1 

YFSa2 

σ F 2 = σ F1 

≤ [ σ F1] 

(5) 

Y 

FSa1 

σF1,σF2 are respectively tooth root bending stress of 

small and large bevel gear; [σ]F1, [σ]F2 are respectively 

allowable bending stress of small and large bevel gear; 

YFSa1, YFSa2 are respectively tooth recombination 

coefficient of small and large bevel gear; Yε is the 

contact ratio coefficient. 

3) The maximum peripheral speed conditions. 

For straight bevel gear, the peripheral speed of the 

average diameter vm should meet: 

vm


kv=polyval(p,x) 

Figure 6. Dynamic load coefficient 

C. Programming of Gear Recombination Coefficients 

In gear design, the tooth shape coefficient YFa and 

stress correction coefficient YSa are the most complicated 

to determine, affected by the gear module, helix angle, 

equivalent number of teeth, variable coefficient, pressure 

angle, tooth root fillet radius and other factors. Ref. [6] 

gives the calculating formula in detail, in which a 

transcendental equation is needed to solve. 

1) Solving of transcendental equation. 

When calculating YFa or YSa, it’s needed to solve 

equation: 

θ=2.*G.*tan(θ)./zv+H (9) 

Where x is needed to calculate, G and H are middle 

variables, and zv is equivalent tooth number. 

This equation can be solved by MATLAB function 

fsolve( ). For example, 

x = fsolve(fun,x0) (10) 

Equation (10) starts at x0 and tries to solve the 

equations described in fun. 

The program of solving (9) is as followings: 

sita0=pi./6.*ones(1,length(zv)) %starting point 

sitax=@(x)bevelsita(x,G,H,zv) % function handle 

sita=fsolve(sitax,sita0) 

Where bevelsita is a M-file function, the content of 

which is as followings: 

function s=bevelyfasita(x,G,H,zv) 

s=x-2.*G.*tan(x)./zv+H; 

2) Recombination gear tooth shape coefficient 

In the paper, the tooth shape coefficient YFa and stress 

correction coefficient YSa are integrated into one 

parameter, that is, YFa* YSa, which are called 

recombination gear tooth shape coefficient, and 

programmed by one M-file function, BevelGearYfsa.m, 

the content of which is as followings: 

function [yfs1,yfs2]=bevelgearYfsa(m,z1,u,fr) 

%generating method, m: module; z1: tooth number 

of %pinion; u: rotating speed ratio; fr: tooth 

face %coefficient 

if app==1 % applied to industry; 

d1=m*z1 %big end diameter 

… 

else app==2 % applied to automobile 


d2=m*z1 % big end diameter 

… 

end 

… 

sf1=(zv1*sin(pi/3-sita1)+sqrt(3)*(G/cos(sita1)pa0/mm))*mmsf2=(zv2*sin(pi/3-sita2)+sqrt(3)*(G/cos(sita2)pa0/mm))*mm 

pf1=(pa0/mm+2*G^2/(cos(sita1)*(zv1*cos(sita1)^2- 

2*G)))*mm 

pf2=(pa0/mm+2*G^2/(cos(sita2)*(zv2*cos(sita2)^2- 

2*G)))*mm 

hfa1=(0.5*zv1*(cos(alf)/cos(alffa1)-cos(pi/3sita1))+0.5*(pa0/mn-G/cos(sita1)))*mnhfa2=(0.5*zv2*(cos(alf)/cos(alffa2)-cos(pi/3sita2))+0.5*(pa0/mn-G/cos(sita2)))*mn 

%calculating tooth shape factor of pinion and gear 

yfa1=(6*hfa1/mm)/((sf1/mm)^2)*cos(alffa1)/cos(alf) 

yfa2=(6*hfa2/mm)/((sf2/mm)^2)*cos(alffa2)/cos(alf) 

La1=sf1./hfa1 

La2=sf2./hfa2 

qs1=sf1./2./pf1 

qs2=sf2./2./pf2 

%calculating stress correction coefficients of 

pinion %and gear 

ysa1=(1.2+0.13.*La1).*qs1.^(1./(1.21+2.3./La1)) 

ysa2=(1.2+0.13.*La2).*qs2.^(1./(1.21+2.3./La2)) 

% recombination gear tooth shape coefficients 

of %pinion and gear 

yfs1=yfa1.*ysa1 

yfs2=yfa2.*ysa2 

VI. REALIZATION OF OPTIMIZAITION DESIGN 

A. Principle of Genetic Algorithm 

Genetic Algorithm (GA) is referred to as a search 

method of optimal solution to simulating Darwin's 

genetic selection and biological evolution process. 

Genetic algorithm is a series of random iterations and 

evolutionary computations simulating the process of 

selection, crossover and mutation occurred in natural 

selection and population genetic, in according to the 

survival of the fittest, through crossover and mutation, 

good quality gradually maintained and combined, while 

continually producing better individuals and out of bad 

individuals. Through the generational produce and 

optimizing the individual, the whole group evolves 

forward and constantly approaches to the optimal 

solution. 

Genetic algorithm, not requiring gradient information 

and continuous function, optimization results being 

global, applied to mechanical design optimization 

problems, can effectively avoid local optimal solutions, 

and get the global optimal solution [2] [7]. 

B. Outline of the Genetic Algorithm 

The following outline summarizes how the genetic 

algorithm works [8]: 

1) The algorithm begins by creating a random initial 

population.


2) The algorithm then creates a sequence of new 

populations. At each step, the algorithm uses the 

individuals in the current generation to create the next 

population. To create the new population, the algorithm 

performs the following steps: 

a) Scores each member of the current population 

by computing its fitness value. 

b) Scales the raw fitness scores to convert them 

into a more usable range of values. 

c) Selects members, called parents, based on their 

fitness. 

d) Some of the individuals in the current population 

that have lower fitness are chosen as elite. These elite 

individuals are passed to the next population. 

e) Produces children from the parents. 

Children are produced either by making random 

changes to a single parent - mutation - or by combining 

the vector entries of a pair of parents - crossover. 

f) Replaces the current population with the 

children to form the next generation. 

3) The algorithm stops when one of the stopping 

criteria is met. 

C. Integer Serial Number Encoding 

In the paper, integer serial number encoding method is 

used, in which each chromosome represents the number 

of a variable value in discrete collection. The advantages 

of integer serial number encoding relative to the binary 

are smaller seek space and simple, intuitive operation. 

For example, a design variable has 10 optional discrete 

values, 4-bit binary code is needed (16 kinds of solutions 

can be expressed), then will result in 6 invalid or 

duplicate combination solutions. However, integer serial 

number encoding can fully express combination solutions 

with only numbers from 0 to 9, so the genetic algorithm 

does not produce an invalid solution or repeated solution, 

but also reduces the code string length and improves the 

encoding and decoding efficiency [9]. 

In the standard straight bevel gear design, the 

independent design variables are three: m, z1, fR, where 

m and z1 are discrete, and fR is continuous. If fR is kept 

two fractions, it is changed into discrete variable. 

Consistent with modulus constrains, if mmax=50, the 

standard modulus number is 36, m = [1.5 1.75 2 2.25 2.5 

2.75 … 45 50], consistent with tooth number constrains, 

the number of z1 is 28, z = [13 14 15 16 17 … 38 39 

40], consistent with tooth width coefficient constraints, 

the number of fR is 6, fR = [0.25 0.26 0.27 0.28 0.29 0.3]. 

So code ranges of x(1), x(2) and x(3) are respectively 1- 

36, 1-28 and 1-6 [6][7]. 

Decoding ways of the three parameters are separately 

as follows: 

Modulus m: 

switch x(1) 

case 1 

m=1.5 

case 2 

m=1.75 


… 

case 35 

m=45 

case 36 

m=50 

otherwise 

disp ('modulus is not in the range') 

end 

Tooth number z: 

z=12+ x(2) 

Face width coefficient fR: 

fR=0.24+x(3)/100 

D. Initial Population 

The initial population is produced by following 

program: 

A=rand(PopulationSize,3) 

X1=round(1+(19-1).*A(1:PopulationSize,1)) 



InitialPopulation =[X1 X2 X3] 

Where, rand() is random function, round() is rounding 

function and PopulationSize(Population Size) is the 

number of individuals. 

Population size is an important parameter in genetic 

algorithm, which specifies how many individuals there 

are in each generation. With a large population size, the 

genetic algorithm searches the solution space more 

thoroughly, thereby reducing the chance that the 

algorithm will return a local minimum that is not a global 

minimum. However, a large population size also causes 

the algorithm to run more slowly. So it can not be too 

large or too small. The range of population size is 10-160. 

It is determined to be 50 by trying. 

The program is as follows: 

options = gaoptimset(' Populationsize ',50) 

E. Fitness Function 

Gear optimization problem is nonlinear constrained 

optimization problems, commonly used penalty function 

method. Penalty function method is simple in principle, 

easy in algorithm, wide application and widely used. 

However, many problems exist in penalty, for example, 

only when the penalty factor r → ∞ (outside point 

method) or r → 0 (interior point method), converge the 

algorithm is, the iterative process is of slow convergence. 

In addition, when the initial value of the penalty factor r0 

is inappropriate, the penalty function may become sick, 

making optimization difficult [5]. 

In this paper, augmented penalty function combined 

Lagrange multiplier method with penalty function 

method for solving straight bevel gear optimization 

problems. The form of augmented penalty function is as 

follows: 

m 1 

2 2 

M( 

x, 

λ, 

r) 

= f ( x) 

+ ∑{ 

max[ 0, 

λ1 

j + rg j ( x)] 

− λ1 

j} 

+ 

2r 

l 

∑ 

r 

λ hp 

( x) 

+ 

2 

2 p 

p= 

1 

∑ 

j= 

1 

2 

[ h ( x)] 

p


Where, f(x) is the objective function; gj(x) is the 

inequality constraints, m is the number of inequality 

constraints; hp(x) is the equality constraints; l is the 

number of equality constraints; r is penalty factor; λ1j the 

multiplier vector for inequality constraint functions; λ2p 

multipliers for the equality constraints vector [10]. 

F. Solving through MATLAB GA Toolbox 

MATLAB is advanced mathematics software launched 

by the MathWorks Company since the mid 1980s, which 

faces to science and engineering. 

MATLAB genetic toolbox is customized toolbox for 

genetic algorithm, with which various problems to 

optimize using genetic algorithm can be easily figured 

out. 

Below is the detailed process to realizing optimization 

design of straight bevel gear using MATLAB 7.1 genetic 

toolbox [8]. 

1) Convert the optimization mathematical model of 

IV into the following forms applied to MATLAB: 

a) Design variable: 

X=[x 1,x2,x3] T 

b) Object function: 

F(X)=π*u* (u+1) *x3 (x1*x2) 3 *(1-x3+x3 2 /3)/8→min 

c) Constraint conditions: 

Linear inequality constraints: 

A*X


fitnessFunction = @bevelobj; 

%Set optimization options: 

options = gaoptimset; %fault options 

% change optimization options 

options = gaoptimset(options,'PopInitRange' ,[1 1 1 ; 

36 28 6 ]); 

options = gaoptimset(options,'MigrationDirection' , 

'both'); 

options = gaoptimset(options,'TolFun' ,0.001); 

options = gaoptimset(options,'TolCon' ,0.001); 

options = gaoptimset(options,'SelectionFcn', 

{@selectiontournament 4 }); 

options = gaoptimset(options,'MutationFcn', 

@mutationadaptfeasible); 

options = gaoptimset(options,'HybridFcn' ,{ @fmincon 

[ ]}) 

%Call ga( ) function: 

[X,FVAL]=ga(@FitnessFcn,nvars,A,b,Aeq,beq,LB, 

UB, @bevelnonlcon,options) 

VII. INTERFACE BETWEEN VISUAL BASIC 

LANGUAGE AND MATLAB SOFTWARE 

Interface between Visual Basic language and 

MATLAB software can be realized by three methods, 

DLL(dynamic linked library), DDE(dynamic data 

exchange) or ActiveX technology. In the paper, the third 

method is adopted. Partial programs are as follows: 

Private Sub OptiCommand_Click() 

‘Announce ActiveX loading MATLAB 

Dim matlab As Object 

‘Loading MATLAB application program 

Set matlab = CreateObject("Matlab.Application") 

‘Define variables for input parameters 

Dim p As Double, n1 As Double, u As Double 

Dim z1 As Double 

Dim fR As Double 

Dim mi As Double, mj As Double 'gear material 

… 

‘Define variables for output results 

Dim outm As Variant 

Dim outz1 As Variant, outz2 As Variant 

… 

‘The known data in UI are transmitted to input 

parameters 

p = Val(Textp.Text) 

u = Val(Textu.Text) 

n1 = Val(Textn1.Text) 

z1 = Val(Textz1.Text) 

fr= Val(TextfR.Text) 

… 

‘Visual Basic variables are transmitted to MATLAB 

workspace 

Call MATLAB.PutWorkspaceData("orgm", "base", 

orgm) ‘original machine working conditions 

Call MATLAB.PutWorkspaceData("wkm", "base", 

wkm) ‘work machine working conditions 

Call MATLAB.PutWorkspaceData("pregrd", "base", 

accgrd) ‘precision grade 


Call MATLAB.PutWorkspaceData("p", "base", p) 

Call MATLAB.PutWorkspaceData("n1", "base", n1) 

Call MATLAB.PutWorkspaceData("u", "base", u) 

Call MATLAB.PutWorkspaceData("z1", "base", z1) 

Call MATLAB.PutWorkspaceData("m", "base", min) 

Call MATLAB.PutWorkspaceData("fr", "base", fr) 

Call MATLAB.PutWorkspaceData("supp", "base", 

supp) 

‘Executing the MATLAB program of bevel gear 

optimization design 

matlab.execute ("bevelgearopt") 

‘Results from MATLAB are transmitted to Visual 

Basic 

Call MATLAB.GetWorkspaceData("M", "base", mout) 

Call MATLAB.GetWorkspaceData("Z1","base", outz1) 

… 

‘Optimization results are transmitted to UI 

Textoutm.Text = mout 

Textoutz1.Text =outz1 

… 

VIII. APPLICATION EXAMPLE 

It is known that input power P=5kw, small gear speed 

n1=960r/min, gear ratio u=4.8; load stability, expected 

life is 15 years, 300 days a year, accounting for 30% of 

working time. Work condition is steady, pinion is 

cantilever, big gear is two-support. 

Result compared with the results of traditional 

optimization design and conventional design is shown in 

Table II. 

TABLE II. 

RESULTS COMPARED WITH THE TRADITIONAL OPTIMIZATION DESIGN 

AND CONVENTIONAL DESIGN 

Design 

method 

Genetic 

algrithm 

Traditional 

optimization 

Conventional 

design 

Module 

Tooth 

number 

Face 

width 

coefficient 

Volume 

2.75 24 0.3 6.8834e+005 

3 33 0.27 2.1604e+006 

3 34 0.3 2.5408e+006 

IX. CONCLUSION 

Optimization design software for straight bevel gear is 

developed in the paper to achieve the automatic 

optimization by using VB and MATLAB, in which the 

genetic algorithm is selected and the augmented penalty 

function and integer serial number encoding is used, so 

the global optimal solution can be obtained, design 

efficiency greatly improved, weight of bevel gear 

reduced, the production cycle shortened. 

REFERENCES 

[1] Zeng Qiang, “The R & D of Gear CAD system based on 

artificial intelligence,” Xi'an Architecture and Technology 

university, 2003


[2] Liang Shangming, Yin Guofu, “Modern mechanical 

optimization Methods,” Beijing,: Chemical Industry Press, 

2005, pp.231-234 

[3] Ang Xueye, Ding Jianmei, “3-Dimensional parametrical 

modeling method of straight bevel gear,” engineering 

graphics transaction, vol. 6, 2007, pp.22~25 

[4] Bi Changchun, Ding Yuzhan, “Application of real number 

encoding genetic algorithm in helical gear drive 

optimization design,” mechanical science and technology, 

vol. 11, 2000, pp.82-84 

[5] Chen Xiuning, “Mechanical optimization Design,” 

Zhejiang University Press, Hangzhou, 1997, pp. 252-260 

[6] Cheng Daxian, Mechanical Design Handbook. Machine 

drive, Beijing: Chemical Industry Press, 2004. 

[7] Zhu Xiaolu, Gear drive design manual, Beijing: Chemical 

Industry Press, 2005. 

[8] The MathWorks, Inc, http://www.mathsworks.com 

[9] Laumanns M, Thiele L, Zitzler E, Welzl E, Deb K, 

“Running time analysis of multi-objective evolutionary 

algorithms on a simple discrete optimization problem,” 

Parallel Problem Solving from Nature—PPSN vol. VII, 

Spain: Granada, 2002, pp.44-53 

[10] Schwefel H. Evolution and optimum seeking, New York: 

John Wiely &Sons, 1995. 


Xiaoqin Zhang: born in Tangshan, 

China in March, 1971. A PHD Candidate, 

Mechanical Engineering College, 

Yanshan University, Qinhuangdao, China, 

Major in CAD/CAM.


Study on Operating Mechanisms and Dynamics 

Behavior of Agile Supply Chain 

Guohua Chen 

College of Mechanical Engineering Chongqing University ,Chongqing400044, China 

Institute of Mechanical & Auto Engineering Xiangfan University, Xiangyang414053, China 

E-mail:59782071@163.com 

Genbao Zhang and Jihong Pang 

College of Mechanical Engineering Chongqing University ,Chongqing400044, China 

E-mail:genbaozhang@163.com; pangjihong@163.com 

Abstract —To study deeply the operating mechanisms and 

system underlying behavior of agile supply chains, a new 

method—systems dynamics (SD) is introduced into the 

analysis of agile supply chain’s behaviors.According to the 

characteristics of agile supply chain, its operating 

mechanisms was analysed and the dynamics model of it was 

established .Then the simulation analysis of systemic 

behavior of agile supply chain was conducted under the 

circumstances of disturbance of market. At the same time, 

the simulation of ordering cycle and target inventory’s 

influence to the behavior of agile supply chain was run. The 

results indicate that: for agile supply chain, the delivery 

ratio of it can be increased not only through adjusting order 

cycle time, but also through changing target inventory, 

which all can increase delivery ratio of the whole supply 

chain. 

Index Terms—agile supply chain, operating mechanisms, 

dynamics behavior characteristics, simulation 


Agile supply chain is defined as the dynamic network 

of supply and demand, which composed of a number of 

supply-side and demand-side entities can do the rapid 

response to environmental changes, demand-side entities 

in the competitive, cooperative and dynamic market 

environment [1-3]. Agile supply chain emphasizes the 

importance of supply chain’s rapid response capability to 

market change and customer demand. It requires large 

enterprise groups, complex production process, even 

specific products, each employee to have agility, which is 

distinctly different from lean supply chain [4]. 

Nowadays, the study on behavior characteristics of 

agile supply chain is by the method of qualitative and 

static analysis. However, supply chain is a dynamic and 

balanceable system. The static analysis method can not 

show the whole supply chain’s operational discipline, and 

the qualitative method can only obtain some perceptual 

knowledge, not achieve quantitative acquaintance. 

Manuscript received June 10, 2010; revised November 10, 2010; 


Corresponding Author: Guohua Chen 


doi:10.4304/jsw.6.5.923-929 

System dynamics provides a qualitative and quantitative, 

semi-quantitative analysis of the problem. It characterized 

as a precursor to quantitative support [5]. Since Forrester 

published “industrial dynamics” in 1961, the system 

dynamics method has been applied in a variety of 

industrial policy-making and strategic issues [6,7]. 

Although the system dynamics model of supply chain 

constructed by Forrester is called in question in a very 

long time .since system dynamics in the early 20th 

century was introduced to China, thousands of people , 

including Wang Qifan [8,9], Su MaoKang [10], Hu 

Yukui [11] and other scholars, involved in the application 

of system dynamics research work in China,but in the 

field of supply chain management, the literature about 

applied research is relatively rare. 

As time goes by, the role in supply chain management 

research using system dynamics is increasingly 

recognized [12], and its application recently is more 

widely. The application of system dynamics method to 

study supply chain’s issues is hot recently. Now research 

about supply chain is related inventory [13,14], retailer’s 

behavior [13], logistics financial balance [15], the 

stability of supply chain [16] as well as the ability to replan 

[17] etc. The paper [13] established a new model of 

supply chain by exploiting computer simulation software 

provided by system dynamics to simulate ordering 

behavior of retailers. An analysis was made on the 

various changing indexes under two strategies: (1) 

ordering amount in terms of sale amount and(2) ordering 

amount in terms of sale amount and inventory. The paper 

[14] focuses on the analysis of simulated impact of the 

radio frequency identification (RFID) system on 

thenventory replenishment of the thin film transistor 

liquid crystal display (TFT-LCD) supply chain in Taiwan. 

A global operations and logistics case of a well-known 

LCD monitor manufacturer in Taiwan has been studied. 

The pull-base dulti-agents supply chain was accordingly 

modeled and simulated with AnyLogic. An automatic 

inventory replenishment unction adopting the (s,S) 

policyis enabled with RFID or not. The studies of paper 

[15] were made on keeping the balance of supplydemand 

of funds in the supply chain system based on 

system dynamic theory. System dynamic logic model and


mathematical equations showed that when the logistic 

system experienced perturbation, a supply-demand 

balance could be achieved by the self-organization of the 

logistic system; when the perturbation accumulated into 

macro fluctuation, balance could be achieved by macro 

evolution of the system self-organization. The paper [16] 

proposed a system-dynamics-based criterion for stability 

judgment. With simulation, the criterion could be used to 

describe the nonlinearities of supply chain system with 

1st order exponential lag and Pure Time Delay (PTD). 

The criterion could be used to judge the influences 

exerted on supply chain stability by decision 

behavior.The paper [17] analyzed the behavior of the 

generic system under study through a simulation model 

based on the principles of the system dynamics 

methodology. The simulation model provides an 

experimental tool, which can be used to evaluate 

alternative long-term capacity planning policies(“what-if” 

analysis) using total supply chain profit as measure of 

policy effectiveness.In all these papers no document 

studies behavior of agile supply chain. So, through the 

simulation of agile supply chain on different conditions, 

some important characteristics can be obtained, which 

can supply the reference for supply chain operation. 

II. The OPERATING MECHANISMS AND 

DYNAMICS MODEL OF AGILE SUPPLY CHAIN 

OPERATION 

Agile supply chain runs in the form of market demandpull 

from the downstream supply chain close to market 

customers to the upstream supply chain close to supplier. 

Agile supply chain can be understood as the mode that 

the enterprises in the downstream supply chain send order 

message to the upstream enterprises according to demand 

conditions on necessary time, and the upstream 

enterprises organize production according to product 

storage to meet the needs of the upstream enterprises.In 

this paper, related study is conducted on the basis of 

taking three-tier supply chain (which includes supplier, 

manufacturer and distributor ) as object. the operating 

mechanisms and dynamics Model of agile supply chain 

can be seen from Fig.1. 

In Fig. 1 the relation among these variables is: 

average demand=SMOOTH(market demand, smooth 

time); 

order 2= supplier's target inventory - supplier's 

inventory; order ratio 1= order 1/ supplier's order cycle 

time; 

supplier's production= order ratio 1-reject ratio 1inspection 

ratio 1; 

inspection ratio1=supplier's production*percent of pass 

1/inspection time 1; 

reject ratio 1=supplier's production*(1 - percent of 

pass 1)/inspection time 1; 

supplier's inventory=inspection ratio 1-output ratio 1; 

output ratio 1=order ratio 2; 

delivery ratio 1=IF THEN ELSE(supplier's inventory < 

output ratio 1 - inspection ratio 1:AND: supplier's 


inventory + inspection ratio 1>0, (supplier's inventory + 

inspection ratio 1)/ output ratio 1 ; 

delivery ratio 1=IF THEN ELSE(supplier's inventory + 

inspection ratio 10, (ditributor's 

inventory+inspection ratio 3)/sales rate 3 , IF THEN 

ELSE(ditributor's inventory


order ratio1 

supplier's order 

cycle time 

order 1 

delivery ratio 2 

distributor's order 

cycle time 

ouput ratio 2 

manufacturer's 

target inventory 

order ratio 3 

distributor's target 

inventory 

supplier's 

production 

supplier's target 

inventory1 

percent of pass 1 

reject ratio1 inspection time 1 

supplier's inventory 

inspection ratio 1 ouput ratio1 


inspection time 2 


inventory inspection ratio 2 

distributor's 

production 

order 2 

reject ratio2 


production order ratio 2 

manufacturer's order 

cycle time 

distributor's 

inventory 

inspection ratio 3 sales ratio 

reject ratio 3 


delivery ratio 1 

average demand 

inspection time 3 delivery ratio 3 market demand 

smooth time 

order 3 

Figure 1. Operating mechanisms and Dynamics Model of agile 

supply chain 

III. SIMULATION ANALYSIS 

Simulation software used in this paper is vensim. 

vensim is a software developed and used in recent years, 

which can assist to complete the system modeling and 

flow-chart drawing and can further present the simulation 

results [11,18]. 

A. Systemic Behavior Under The Circumstance Of 

Market Disturbance 

Suppose market demand obeys normal distribution 

function RANDOM NORMAL ( 0 ,10 , 5 , 5 , 0 ). 

Other state variables’ initial values are set as follows: 

supplier's order cycle time= manufacturer's order cycle 

time = ditributor's order cycle time=2; 

inspection time 1= inspection time 2=inspection time 

3=smooth time=2; 

percent of pass 1= percent of pass 2= percent of pass 

3=0.95; 

supplier's target inventory= manufacturer's target 

inventory= ditributor's target inventory=20; 

Time step of simulation =0.125; 

Initial time=0,Final time=100; 

Fig. 2 shows the changes of simulation, including 

supplier's inventory, manufacturer's inventory, 

distributor’s inventory and delivery ratio 1, delivery ratio 

2, delivery ratio 3 before market demand's disturbance. 

Now, add a disturbing function to market demand, then 

the distribution function of market demand is RANDOM 

NORMAL (0,10,5,5,0) +PULSE(50,100)*5. Systemic 

behavior changes can be seen from Fig. 3 after market 

demand's disturbance. 

From the change between Fig. 2 and Fig. 3, we can see 

the curves of supplier’s inventory ,manufacturer’s 

inventory and distributor’s inventory shift down to 

horizontal axis after market demand ’s disturbance and 

fluctuates slightly , which explains that for agile supply 

chain their delivery is still stable when market demand is 


stable or not ,but market demand’s disturbance has 

influence to supplier’s inventory, manufacturer’s 

inventory and distributor’s inventory and their devery 

ratio. From table I, we can see clearly that with market 

demand’s increasing suddenly the values of delivery ratio 

1 and delivery ratio 2 reduces relatively large, but the 

value of delivery ratio 3 changes from 0.9930 to 0.9164, 

reducing relatively small, which shows that market 

demand ’s disturbance has a smaller effect on delivery 

ratio 3 than delivery ratio 1 and delivery ratio 2. 

market demand 

10 

7.5 

5 1 1 

2.5 

0 

1 

1 

1 

1 

1 

1 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

market demand : 1 1 1 1 1 1 1 1 1 1 1 1 1 1 

40 

20 

0 

-20 

-40 

2 1 

1 

3 

1 

3 

3 3 1 2 

3 2 3 1 3 3 3 

3 3 3 3 3 

2 2 

2 2 2 

2 

1 

2 

1 

1 1 

2 2 2 2 

3 

1 

1 1 1 

1 

1 

1 

2 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

supplier's inventory : 1 1 1 1 1 1 1 1 1 1 1 1 

manufacturer's inventory : 2 2 2 2 2 2 2 2 2 2 

distributor's inventory : 3 3 3 3 3 3 3 3 3 3 3 

1 

0.75 

0.5 

0.25 

0 

1 2 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

1 

(a)Market demand obey normal distribution 

(b)Inventory changes before market demand's disturbance 

3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 

delivery ratio 1 : 1 1 1 1 1 1 1 1 1 1 1 1 



(c)Delivery ratio changes before market demand's disturbance 

Figure 2. Systemic behavior before market demand's disturbance 

1 

1 

1 

1 

1 

1 

1


20 

15 

10 

5 

1 

0 

1 1 

1 

1 

1 

market demand 

1 

1 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

market demand : 1 1 1 1 1 1 1 1 1 1 1 1 1 1 

40 

20 2 1 

0 1 

3 

1 

3 

3 3 1 2 

3 

3 3 3 

2 2 

2 1 

1 2 1 2 1 

1 

2 

3 

3 1 2 

3 3 3 2 3 3 

2 1 2 1 2 1 1 

2 1 

-20 

2 

1 

-40 

1 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 




1 

0.75 

0.5 

0.25 

(a)Market demand after disturburbance 

(b)Systemic behavior after disturbance 

3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 

0 

1 2 

0 10 20 30 40 50 

1 

60 70 80 90 100 

Time (Week) 

1 

2 

3 

1 

1 

1 

3 1 

2 

3 

2 

2 

2 1 

3 1 

3 

3 2 

1 

1 

1 

1 2 




(c)delivery ratio changes after market demand's disturbance 

Figure 3. Systemic behavior after disturbance 

B. Order cycle time’s influence to systemic behavior 

Generally speaking, increasing order cycle time can 

influence the delivery ratio of the whole supply chain. 

However, for agile supply chain, the fact is not as so. 

When the value of supplier’s order cycle time is only 

changed (but the manufacturer’s order cycle time and the 

distributor’s order cycle time is unchanged), the curve 

line of supplier’s inventory changes accordingly, but the 

curve liners of the manufacturer’s inventory and the 

distributor’s inventory don’t change,which explains 

supplier’s order cycle time has only inluence on 

supplier’delivery,not on other delivery. 


1 

3 

TABLE I. 

STATISTICS OF AVERAGE DELIVERY RATIO UNDER DISTURBANCE 

strategies 

Delivery ratio 

1 


2 


3 

Before disturbance 0.9424 0.9624 0.9930 

After disturbance 0.7683 0.8347 0.9164 

Fig. 4 (a1) and (a2) is a systemic behavior chart when 

distributor’s order cycle time is changed as 10 ,and 

supplier’s order cycle time and manufacturer’s order 

cycle time is unchanged (they are still 2). From Fig. 4 

(a1) and (a2),we can see that with increasing distributor’s 

order cycle time to 10,the curve of distributor’s inventory 

lies down axis, which states distributor’s delivery ( it is 

also can be seen from delivery ratio 3) can not meet 

market demand. However , the curves of supplier’s 

inventory and manufacturer’s inventory are still above 

axis, almost unchanged, which states supplier’s and 

manufacturer’s delivery ( it is also can be seen from 

delivery ratio 1 and delivery ratio 2 ) can still meet 

downstream demand. 

Fig. 4(b1) and (b2) is a systemic behavior when the 

value of supplier’s order cycle time and manufacturer’s 

order cycle time is changed as 10,and distributor’s order 

cycle time is unchanged (still 2).From Fig. 4(b1) and 

(b2),we can see that when change supplier’s order cycle 

time and manufacturer’s order cycle time as 10,the two 

curves of supplier’s inventory and manufacturer’s 

inventory lies down horizontal axis, which states 

supplier’s delivery and manufacturer’s delivery ( it is also 

can be seen from delivery ratio 1 and delivery ratio 2) can 

not meet downstream demand. However, the curve of 

distributor’s inventory is almost unchanged, still above 

horizontal axis , which states distributor’s delivery ( it is 

also can be seen from delivery ratio 3) can still meet 

market demand. 

All these explain that for supplier, manufacturer and 

distributor, changing whose order cycle time only whose 

delivery, that is to say, all delivery ratio of the whole 

supply chain can be increased through changing 

supplier’s order cycle time, manufacturer’s order cycle 

time and distributor’s order cycle time simultaneously. 

To deeply illustrate the relationship between systemic 

behavior and order cycle time, we analyzed statistics 

about delivery raio,which can be seen from Table II. 

Related explanation: order cycle time (10,10,2) 

expresses supplier’s order cycle time,manufacturer’s 

order cycle time are 10, distributor’s order cycle time is 

2; order cycle time (2,2,10) expresses supplier’s order 

cycle time and manufacturer’s order cycle time are 2 

respectively, and distributor’s order cycle time is 10. 

All of these results demonstrate that for agile supply 

chain, the strategy of changing order cycle time only 

improves the delivery ratio of local supply chain, not of 

the whole supply chain. The delivery ratio of the whole 

supply chain can be increased through reducing all order 

cycle times.


40 

20 

0 

-20 

-40 

1 

2 

3 

1 

2 

3 

1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 

3 3 3 3 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

supplier's inventory : 1 1 1 1 1 1 1 1 

manufacturer's inventory : 2 2 2 2 2 2 

distributor's inventory : 3 3 3 3 3 3 3 

1 

0.75 

0.5 

0.25 

2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 

0 

1 

0 

3 3 

10 

3 

20 

3 3 

30 

3 

40 

3 3 

50 

3 

60 

3 3 

70 

3 

80 

3 3 

90 

3 

100 

Time (Week) 




30 

0 

-30 

-60 

1 

3 

1 

2 2 

3 3 3 3 3 3 3 3 3 3 

1 

2 

3 

2 

2 

1 1 1 2 1 2 2 

1 1 2 

1 2 1 2 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

supplier's inventory : 1 1 1 1 1 1 1 1 

manufacturer's inventory : 2 2 2 2 2 2 

distributor's inventory : 3 3 3 3 3 3 3 

1 

0.75 

0.5 

0.25 

0 

(a1) Systemic behavior when only change distributor’s order 

cycle time as 10 

(a2) delivery raio changes when only change distributor’s order cycle time 

as 10 

60 

(b1) Systemic behavior when change supplier’s and manufacturer’s 

3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 

1 2 

0 

1 2 1 2 1 2 

10 20 

1 2 1 2 1 2 

30 40 

1 2 1 2 1 2 

50 60 

1 2 1 2 1 2 

70 80 

1 2 1 2 

90 100 

Time (Week) 




(b2) delivery raio changes when change supplier’s and manufacturer’s order 

cycle time as 10 respectively 

Figure 4. Influence of order cycle time to systemic behavior 

C. Target inventory’s influence to systemic behavior 

There are 3 kinds of target inventories, which are of 

supplier, manufacturer and distributor. Target inventory’s 


3 

3 

3 

3 

TABLE II. 

STATISTICS OF AVERAGE DELIVERY RATIO UNDER 2 STRATEGIES 

strategies 

order cycle time 

(10,10,2) 

order cycle time 

(2,2,10) 

Delivery 

ratio 1 

Delivery 

ratio 2 

Delivery 

ratio 3 

0.0000 0.0000 0.9930 

0.9626 0.9972 0.0000 

influence to systemic behavior is analyzed through 

changing 3 strategies of target inventories. 

When the value of supplier’s target inventory is 

changed, the curve line of supplier’s inventory moves 

accordingly (when increase its value, the curve of 

supplier’s inventory moves up, and lower its value, the 

curve of it moves down),but the curve liners of 

manufacturer’s inventory and distributor’s inventory 

almost unchanged including its shape, which means 

supplier’s target inventory has influence to the delivery of 

supplier, not to the delivery of manufacturer and 

distributor. All these are similar to the influence of 

manufacturer’s target inventory and distributor’s target 

inventory to their inventories. 

Fig. 5(a1) and (a2) is systemic behavior when the 

value of supplier’s target inventory is changed as 10,and 

the values of manufacturer’s inventory and distributor’s 

inventory are still 20 ,which shows the changes of 

simulation,including supplier's inventory, manufacturer's 

inventory, ditributor's inventory and delivery ratio 1, 

delivery ratio 2, delivery ratio 3. From Fig. 5(a1) we can 

see that when reduce supplier’s target inventory , the 

curve of supplier’s inventory moves down slightly, but 

the two curves of manufacturer’s inventory and 

distributor’s inventory move hardly. At the same time, 

the curve of delivery ratio 1 changes clearly,and the 

curves of delivery ratio 2 and delivery raio 3 show little 

change. 

Fig. 5(b1) and (b2) is systemic behavior when the 

values of manufacturer’s target inventory and 

distributor’s target inventory are changed as 10, and 

supplier’s target inventory is still 10. It can be seen from 

Fig. 5(b1),when we reduce manufacturer’s target 

inventory and distributor’s target inventory, the two 

curves of manufacturer’s inventory and distributor’s 

inventory accordingly move down ,and their delivery 

ratio ( delivery ratio 2 and dlivery raio 3) become 

lower,not to meet delivery requirements. 

To explain concretely the phenomena, we calculated 

the values of delivery ratio including supplier’s target 

inventory changed and unchanged, which can be seen 

from Table III. 

Related explanation: Target-inventories (20,10,10) 

expresses supplier’s target inventory is 20,manufacturer’s 

target inventory and distributor’s target inventory are 10 

respectively; Target-inventories(10,20,20) expresses 

supplier’s target inventory is changed as 10, and 

manufacturer’s target inventory and distributor’s target 

inventory are still 20 respectively.


40 

20 2 

3 

3 

1 2 3 2 

3 2 

3 2 3 2 3 3 

2 2 3 

3 

2 

2 

3 3 

2 2 

3 2 3 

2 

3 

0 1 

1 

1 

1 

1 

1 

1 

1 

1 

1 1 1 1 

1 

-20 

2 1 

-40 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 




1 

0.75 

0.5 

0.25 

3 2 3 1 2 3 2 3 1 2 3 2 3 1 2 3 2 3 1 2 3 2 3 1 2 3 2 3 2 3 2 3 2 3 

1 

1 

1 

1 

0 

1 2 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 




20 

10 

0 

-10 

-20 

1 

2 

1 

1 

1 

1 

1 1 

1 

1 1 

1 

1 

1 

2 

3 

2 2 3 2 

2 

3 3 3 

3 

3 

2 3 2 

3 

2 3 

2 

3 3 

2 

2 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

supplier's inventory : 1 1 1 1 1 1 1 1 1 

manufacturer's inventory : 2 2 2 2 2 2 2 

distributor's inventory : 3 3 3 3 3 3 3 3 

1 

0.75 

0.5 

0.25 

(a1) Systemic behavior when only change supplier’s target inventory 

as 10 

(a2) delivery ratio changes when only change supplier’s target 

inventory as 10 

(b1) Systemic behavior when change manufacturer’s and 

distributor’s target inventory as 10 respectively 

3 1 2 3 1 1 3 1 2 1 2 3 1 1 3 1 1 3 1 2 1 1 3 1 2 1 

3 

3 

2 

3 

2 

3 

2 2 

0 

1 2 

0 10 20 30 40 50 60 70 80 90 100 

Time (Week) 

1 

3 

2 




(b2) delivery ratio changes when only change supplier’s target 

Figure 5. Influence of target inventory to systemic behavior 

From table 3, we can see when the value of supplier’s 

target inventory is changed at 10, delivery ratio 1 has 


2 

1 

3 

1 

2 

3 

1 

2 

1 

3 

2 

TABLE III. 

STATISTICS OF AVERAGE DELIVERY RATIO UNDER 2 STRATEGIES 

strategies 

Target-inventories 

(20,10,10) 

Target-inventories 

(10,20,20) 

Delivery 

ratio 1 

been changed ,but delivery ratio 2 and delivery ratio 3 

unchanged. All of these results demonstrate, the delivery 

ratio of the whole supply chain can be improved through 

increased different target inventories. 

IV. CONCLUSIONS 

From this study, there are a number of conclusions can 

be drawn. They are as follows: 

(1) The delivery of Agile supply chain is stable, 

whether it is in market demand’s disturbance or not. 

(2) Traditionally, increasing order cycle time can 

promote the delivery ratio of the whole supply chain 

.However ,for agile supply chain, order cycle time only 

affect on local supply chain, not on the whole supply 

chain. The delivery ratio of the whole supply chain can be 

increased through reducing all order cycle times.And 

there is a appropriate value of order cycle time which can 

make the whole inventory in supply chain reach 

minimum but the delivery can be meet. 

(3) For agile supply chain, the strategy of changing 

target inventory increases only the delivery of local 

supply chain, not of the whole supply chain. The delivery 

ratio of the whole supply chain can be increased through 

changing different target inventory. 

ACKNOWLEDGMENTS 

Project supported by the National High-Tech. R&D 

Program, China (No. 2009AA04Z119), the National 

Natural Science Foundation, China (No. 50835008), the 

National Major Scientific and Technological Special 

Project for “High-grade CNC and Basic Manufacturing 

Equipment”, China (No.2009ZX04014-016 ; 

2009ZX04001-013;2009ZX04001-023; 2010ZX04014- 

015), and supported by Open Research Foundation of 

State Key Lab. of Digital Manufacturing Equipment & 

Technology in Huazhong University of Science & 

Technology and the Scientific and Technological Projects 

of Xiangfan City(NO.2010GG3A44). 

REFERENCES 

Delivery 

ratio 2 

Delivery 

ratio 3 

0.9628 0.7244 0.8339 

0.6255 0.9624 0.9930 

[1] S.H. Ma. and Y. Lin., Supply chain, Beijing: China 

machine press, pp.50-52,May 2005. 

[2] H.X. Zhao, R. Du, C.D. Qin and L. Zhou.Study on Value 

of Technological Innovation in Agile Supply Chain Based 

on Inter-Organizational Knowledge Sharing.Chinese 

Journal of Management Science, pp.426-429,October 

2008. 

[3] M.M. Han, X.Q. XU and S.S. Zhang .XML and Agility of 

Agile Supply Chain Management System. China 

Mechanical Engineering, July 2002. 

[4] F.T. ZENG, H.Y. Li and J. Sun. Primary Analysis on


Agile SC and Lean SC. Logistics Technology.January 

2004. 

[5] Q.F. Wu.. Systems dynamics, Shanghai: Shanghai 

University of Finance Press, 2009. 

[6] Y.,Barlas System dynamics: system feedback modeling 

for policy analysis in knowledge for sustainable 

development-an insight into the encyclopedia of life 

support systems, Paris, France, Oxford, UK: UNESCO 

publishing-Eolss Publishers, June 2002. 

[7] J. D. Sterman, Business dynamics:system thinking and 

modeling for a complex world, New York:McGraw-Hill, 

January 2000. 

[8] Q.F. Wang.Comprehensive and Dynamic Analysis and 

Model Set of Large Complex System. JOURNAL OF 

MANEGEMENT SCIENCES IN CHINA,pp. 15~ 19, 

February 1999. 

[9] Y.Y. Cai, Q.F.Wang and J.G. Jia .Review: From System 

Dynamics to Organization Learning .Chinese Journal of 

Management Science, pp.237~ 247, August 2000. 

[10] M.K.Su. Principle and Application of System 

Dynamics[M].Shanghai:Shanghai Jiaotong University 

Press. June 1988. 

[11] Y.K. Hu, Y.G. Han and Z.Y. Cao. Evolution of System 

Dynamics Model. SYSTEMS ENGINEERING - 

THEORY & PRACTICE, pp.132~ 136,October,1997. 

[12] Tow ill D R. Time compression and supply chain 

management - a guided tour [J]. Supply Chain M 

anagement, pp.15~ 27,January 1996. 

[13] G.Z. Jia. and X.L.Wang. Research on retailer performance 

in supply chain based on system dynamic, Industrial 

Engineering Journal, pp.56-59,December 2009. 

[14] W. Shujen, L. Shih-Fei, and W. Weiling, The simulated 

impact of RFID-enabaled supply chain on pull-based 

inventory replenishiment in TFT-LCD industry, 

International Journal of Production Economics, Vol.112, 

pp.570-586,2008. 

[15] Y. Wang and H.Y. Hao., Blance of supply and demand of 

logistic funds in supply chain system based on system 

dynamics, Industrial Engineering Journal, Vol.12, pp.46- 

50,2009. 


[16] C.Luo., S.L.Jia. and H.W. Wang., Stability criterion of 

supply chain based on system dynamics, Computer 

Integrated Manufacturing Systems, Vol.13,pp.1762- 

1767 ,2007. 

[17] D. Vlachos, P. Georgiadis, and E. Lakovou, A system 

dynamics model for dynamic capacity planning of 

remanufacturing in closed-loop supply chains, Computers 

& Operations Research,Vol. 34, pp.367-394,2007. 

[18] L.B.Zhang, Y.Q. Han , J. Chen , Z. Yu and Y.Z. MA. A 

Review:the Application of System Dynamics in Supply 

Chain Management[J] Systems Engineering, Vol. 23,pp. 

8-15,2005. 

Guohua Chen was born in Huanggang 

City, Hubei Province, China, on 

December 10, 1976. He received his 

master's degree in Mining Engineering 

from Henan Polytechnic University in 

2004,6. Currently, he is a PH.D 

candidate with Mechanical Engineering 

at the College of Mechanical 

Engineering, Chongqing University in 

China since 2008. His main research interest is quality 

engineering, industrial engineering (IE), supply chain 

management, etc. 

Genbao Zhang is a professor at Chongqing University. 

His main research interest is quality and reliability engineering, 

enterprise informatization, advanced manufacturing technology. 

Jihong Pang is a PH.D candidate with Mechanical 

Engineering at Chongqing University. His main research 

interest is uality and reliability engineering , industrial 

engineering (IE), Enterprise Resource Planning (ERP).


Unified Service Platform for Accessing Grid 

Resources 

Abstract�Web Services Resource Framework (WSRF) 

redefines Grid Services standards and extends Web Services 

by adding stateful resources. Using GT4 to develop WSRF 

Grid Services is a taxing work, and it is difficult to build and 

deploy these services dynamically. Addressing these issues, 

this paper proposes a unified service platform which can 

provide a series of unified service interfaces for accessing 

kinds of different Grid resources. On the platform, Grid 

resources are independent of the service interfaces. The 

platform provides unified service interfaces to access different 

Grid resources on server, so Grid services developers only pay 

attentions to realizing the native methods of Grid resources 

and configuring necessary resource database. The remainder 

work of composing typical Grid Services such as mapping the 

resources into the service interfaces would be automatically 

finished by the platform. It means that Grid Services 

development becomes native application development. What is 

more, there are no needs to restart the service container when 

deploying/undeploying Grid resources, so it does not affect 

other resources. The platform provides service-users with two 

types of clients, one is for directly invoking the unified 

interfaces and the other is a proxy client associating to the 

specific Grid resource. Finally, the test shows that the service 

development and deployment is much easier on this platform 

and the service performs well. 

Index Terms�WSRF, GT4, Grid resources, Web Services 


Web Services Resource Framework (WSRF) specifies 

stateful services by adding stateful resources to stateless 

Web Services. The Globus Toolkit (GT) is a software toolkit 

can use to program Grid-based applications. The Globus 

Toolkit 4(GT4), in fact, includes a complete implementation 

of the WSRF specification. GT4 Java WS Core is the 

common runtime component provides a set of Java libraries 

and tools which are needed to build both WS and non-WS 

services. This paper discusses the issue that using GT4 to 

build WSRF Grid Services [1] [2]. 

Developer may encounter some problems using GT4 Java 

WS Core to developing Grid Services and Grid-based 

applications. 

1) Heavy coding workload to generate a WSRF service. 

Generating Grid Services Using GT4 is different to 


doi:10.4304/jsw.6.5.930-936 

Shaochong Feng 

Dept. 2, Mechanical Engineering College, Shijiazhuang, China 

Email: fscsat@gmail.com 

Yuanchang Zhu and Yanqiang Di 

Dept. 2, Mechanical Engineering College, Shijiazhuang, China 

Email: {YuanchangZ, YanDi}@gmail.com 

developing native applications. Besides the kernel code of 

the services and resources, many other configuration files, 

build files and scripts must be finished manually. As Fig.1 

shows, writing a simple stateful web service that uses 

WSRF to keep state information needs at least 4 steps: 

� Define the service's interface. This is done with 

WSDL. 

� Implement the service. This is done with Java. 

� Define the deployment parameters. This is done 

with WSDD and JNDI. 

� Compile everything and generate a GAR file. This 

is done with Ant. 

Figure 1. Generating a WSRF service GAR file 

It is a great block for the newer, and is a load for the 

trained programmer.


2) Difficult to dynamically deploy and undeploy. The 

final work of deploy a WSRF Grid Service is placing the 

generated GAR files to services container. Although GT4.2 

supports dynamic deployment in standalone container, the 

third party Web services containers such as Tomcat are not 

supported. If developers want to deploy or undeploy Grid 

Services, they have to stop the services container first, 

executing the deploying or undeploying, and then restart the 

services container. The other active stateful services and 

resources running in the container would be consequentially 

affected. What is more, services modifying and transferring 

cannot be achieved dynamically. 

Addressing these issues, this paper proposes a Unified 

Service Platform (USP) for accessing Grid resources based 

on GT4 Java WS Core. Under USP, the main workload of 

developing Grid applications is realizing the native methods 

of the Grid resources, just like developing native application. 

The Grid resources can be dynamically deployed and 

undeployed under USP regardless what kind of Web 

services container is used. 

The remainder of the paper is organized as follows: 

Section II briefly summarizes the related researches and 

works. Section III analyzes the modes of accessing the Grid 

resources. Section IV describes the architecture of USP. 

Section V presents the experiment on USP. Conclusions and 

future work are presented in Section VI. 

II. RELATED WORKS 

The research group form University of Marburg proposed 

GDT (Grid Development Tools) which is part of the 

Marburg Ad-hoc Grid Environment (MAGE). GDT is a 

bundle of Eclipse Plugins useful for Service and Application 

Development, Workflow Creation, Grid Management and 

others in the Eclipse Integrated Development Environment 

[5]. GDT greatly reduced the workload under Eclipse to 

develop Grid applications. In [7], they modified the Axis 

web service engine utilized by GT4 to allow dynamic 

loading and unloading of Grid services. Hot Deployment 

Service (HDS) was constructed to provide applications with 

the capability to remotely deploy, undeploy and redeploy 

services onto a running node. In [8] an approach to discover 

Grid resources and to deploy Grid services based on 

peer-to-peer technologies was presented. 

Reference [9] and reference [10] specially researched the 

dynamic services deployment. Eun-Kyu Byun developed 

Universal Factory Service (UFS) that provided a dynamic 

Grid service deployment mechanism and a resource broker 

called Door service [9]. Jon B. Weissman proposed a 

dynamic service architecture consists of several core 

services and components, the kernel was the Adaptive Grid 

Service (AGS) [10]. However, these researches are all based 

on OGSI/GT3, and with the development of Grid 

technology, OGSI/GT3 has been replaced by WSRF/GT4. 

The OGSI services don�t have states information, so the 


proposed approaches in [9] and [10] are not compatible with 

WSRF. 

FuQiang Li presented an approach to deploy visualization 

services dynamically in [11]. The Uniform Visualization 

Service (UVS) is developed in this paper to provide 

dynamic visualization services deployment on the Grid, and 

a service named Agency Service, is used as a resource 

broker/dispatcher and service states communicator. Service 

users send requests to Agency Service. The Agency Service 

find available or applicable services from service center, 

retrieve the available service codes from the service codes 

center, and find available resources from MDS. The Agency 

Service can transfer the service codes to the available 

resources. The UVS on the resources creates UVS instance 

to create the service instance with the service codes. The 

service instance has its own service resources, which can be 

interacted and collaborated with the service users. In this 

architecture, the service providers only concern on 

developing the visualization service codes but not finding 

available resources. 

Addressing the issue that services container can only 

accommodate language-special services, Pu Liu and 

Michael J. Lewis described the implementation of an 

approach that allows Grid services developers to write their 

code in one language, and have the services running in 

different service containers, namely GT4, WSRF.NET, and 

gSOAP and ACE [12]. 

Li Qi and Doc. Hai Jin proposed a highly available 

dynamic deployment infrastructure (HAND) in [13], which 

addressed dynamic service deployment at both the container 

level and the service level. 

These researches are significative for this paper, but all of 

them concerned the service itself, the service interfaces are 

tightly bound to the resources. USP proposed in this paper 

keeps the services interfaces and the resources completely 

separated, users can access the Grid resources through the 

same service interface. USP provides the unified service 

interface and the mechanism of managing different 

resources on server, so the Grid Service developing is 

predigested to developing resources and configurating 

database. 

III. MODES OF ACCESSING GRID RESOURCES 

WSRF adds stateful resources to stateless Web Services. 

So the WSRF service interfaces usually bind the resources. 

The kernel idea of this paper is thoroughly disparting the 

service interfaces and Grid resources. USP publishes the 

unified WSRF service interfaces which access Grid 

�� terms are 

listed as follows: 

1) Proxy Resource (PR): PR is a Java class, and it is the 

resource directly related to the published service interfaces 

by USP. PR implements the defined functions in USP�s 

WSDL file, and provides access to the real resource.


2) Real Resource (RR): RR also is a Java class, and it is 

the resource with the final functions. RR is reserved in the 

resource database. 

3) Instance of Proxy Resource (IPR): IPR is an object of 

class PR. 

4) Instance of Real Resource (IRR): IRR is an object of 

class RR. 

USP takes the user-defined Grid resources as RR, and 

publishes unified service interfaces for accessing the RRs. 

The published service interfaces only retrieve PR, and they 

know nothing about the RR. USP automatically maps the 

access from client to �� 

methods. 

When client invokes the service published by USP, the 

PR and RR on server may interact with each other in three 

modes: 

A. Singleness IPR mode 

As Fig.2 shows, in the singleness IPR mode, there is only 

one instance of proxy resource. All clients access IRRs 

through the Singleness IPR, so the invoked service interface 

should include the parameters to indicate the invoking RR, 

the specified IRR and the exact method with all of the 

�� 

B. IPR-RR mode 

As Fig.3 shows, in the IPR-RR mode, USP maintains an 

instance of proxy resource for each real resource class, and 

one IPR may contain several instances of real resource. The 

clients access the different instances of the same RR class 

through the same IPR. It means the invoked RR can be 

judged by the IPR, so the invoked service interface should 

include the parameters to indicate the specified IRR and the 

exact method with all of the �� 

C. IPR-IRR mode 

As Fig.4 shows, in this mode, instance of proxy resource 

and instance of real resource are peer to peer. The clients 

access the IPR-IRR pairs. So the invoked service interface 

�� 

parameters. 

Client1 

Client2 

Client3 

Client4 

Service 

Interface 


Instance of 

Proxy 

Resource 

Figure 2. Singleness IPR mode 

Instan 

ce1_A 

Instan 

ce2_A 

Instan 

ce1_B 

Instan 

ce2_B 

Real Resource A 

Real Resource B 

Client1 

Client2 

Client3 

Client4 

Client1 

Client2 

Client3 

Client4 

Service 

Interface 

Service 

Interface 

Instance1 of 

Proxy 

Resource 

Instance2 of 

Proxy 

Resource 

Figure 3. IPR-RR mode 

Instance1 of Proxy 

Resource 


Resource 


Resource 


Resource 

Figure 4. IPR-IRR mode 

Instan 

ce1_A 

Instan 

ce2_A 

Instan 

ce1_B 

Instan 

ce2_B 

Instan 

ce1_A 

Instan 

ce2_A 

Instan 

ce1_B 

Instan 

ce2_B 

Obviously, in the first two modes, it is probable that 

client access different IRRs through the same IPR, and the 

fault in one IRR accessing may fail several of the other 

IRRs. What is more, in view of GT4, the IPR and IRR could 

not be created together with these two modes, extra 

attentions must be paid to manage the IRR, and it may 

increase the complexity of the system. Anyway, this paper 

adopts the IPR-IRR mode. 

IV. IMPLEMENT OF THE MAIN COMPONENTS 

UIS IPR IRR 

RQDS 

SLQS 

Resource 

manager 

Resource database 

Server load 

monitor 

UIS: Unified invoking Service 

RQDS: Resource Query&Deploy Service 

SLQS: Server Load Query Service 

Figure 5. Structure of USP server 

Besides IPR and IRR, USP server is composed of three 

WSRF services and several backup modules like Fig.5 

shows. 


Real Resource B 


Real Resource B


A. Resource database 

Resource database maintains the RR class information, 

such as the full class name (including class package), public 

class methods and so on. The information of the active IRR 

is also recorded here. 

B. Resource manager 

This module takes responsibility to map IPR to IRR. The 

functions are listed as following: 

� Queries and configures the resource database. 

� Obtains the RR class according to the queried RR�� 

name utilizing Java reflection/introspection 

� 

mechanism. Farther, all of the RR class native 

methods can be achieved. 

Maps the access to IPR to corresponding IRR 

according to the necessary parameter provided by 

the UIS. 

C. Server load monitor 

The performance parameters such as CPU type, CPU 

utilization, memory size, memory utilization and bandwidth 

are gained by server load monitor module. 

D. Unified Invoking Service(UIS) 

As the kernel part of USP, UIS is a WSRF service, and it 

provides clients with the unified invocation interfaces to 

requesting the RR on server. IPR-IRR mode described in 

Fig.3 is adopted here. The main interfaces published by UIS 

are listed as follows: 

1) Service interface for resource instantiation: The service 

interface CreateResource(Sting strRRName)can create an 

IPR-IRR pair. The strRRName parameter specified the 

identity of the accessed RR. Receiving the invoking on this 

interface, server creates an instance of PR, then queries the 

resource database, obtains the RR by strRRName, and 

creates the object of it. So an IPR-IRR pair is maintained on 

server for each client. 

2) Service interface for accessing the RR: Because the 

native methods are different between different RRs, the 

service interface for accessing �� 

designed to be unique, and it is WSRFFunc(String strParam, 

String strFuncName). The strFuncName parameter specifies 

�� 

included in the strParam parameter. Receiving the request 

on this interface, server invokes IPR, and then maps this 

invoking to the IRR through resource manager module. 

3) Interface for the notification: in GT4, resource exposes 

a resource property (RP) as a topic for client to subscribing, 

a notification would be triggered each time the value of the 

RP changes, and then client subscribing the topic receives 

the notification. The topic published by RR are also 

uncertain, so the USP implies a unified methods 

WSRFNoty(String str)in UIS. PR publishes one topic for 

client subscribing, and all the message of the notification 

such as the real topic of RR and the value of RP is encoded 

in the str parameter. 


E. Resource Query&Deploy Service( RQDS) 

RQDS provides WSRF services facilitating users to 

program under USP. 

1) Service interface for querying RR information: the 

service interface QueryAllClass() queries all of the RRs, and 

generates a text file for client. The text file lists all the class 

names of RRs. 

2) Service interface for querying the pointed RR 

information: the service interface QueryClass(String 

strRRName) queries detailed information of pointed RR 

specified by the strRRNamer parameter, and generates three 

file for client. The first file lists all the public methods with 

parameters of the class. The second file shows an example 

of invoking the unified service interfaces published by UIS. 

The third file is a Java file, it is the proxy client, and 

sometimes a proxy notifier is also generated if the pointed 

RR can notify the client. UIS publishes only one service 

interface WSRFFunc �� 

Although client may invoke this interface follow the 

example in first file, the client program would be very 

complex, because clients have to make sure the 

strFuncName �� and 

encode the strParam parameter according to the achieved 

second text file generated by QueryClass interface. So, the 

QueryClass interface can build new Java classes based on 

the queried RR class and the client stubs, and they are the 

proxy client and proxy notifier. The proxy client 

encapsulates the client stubs and provides same public 

methods with the pointed remote RR; analogously the proxy 

notifier provides the same interfaces with the pointed remote 

RR. So if client instantiates the proxy client, and invokes its 

methods, the proxy client would map the invoking to the 

WSRFFunc request, and finally the request would be sent to 

��corresponding method on server. It greatly 

advantages the USP users. 

3) Service interface for deploying Grid resource: the 

service interface DeployingRR(String strConfgFile, String 

strRR) is for users to remotely deploy the RR and configure 

the resource database. The strConfgFile parameter specifies 

a local text file, which should define the following elements: 

RR full class name, location on the server, and dependence 

relationship. The other strRR parameter refers the native 

Java class file to be deployed as RR. The file transmission 

from client to server is based on GridFTP. 

All of the interfaces of the RQDS are related to the 

resource manager module which reads and writes the 

resource database. 

F. Server Load Query Service(SLQS) 

�� 

information such as the CPU type, CPU utilization, memory 

size, memory utilization and bandwidth. 

G. Process of accessing the resources 

Under USP, the process of accessing the Grid resources is 

described in Fig.5 (it is supposed that client has invoked


RQDS, generated the proxy client, and instantiated the 

IPR-IRR pair by calling CreateResource). 

1) Client invokes the native method Func(String str) of 

the proxy client. 

2) Proxy client transforms Func(String str) to 

WSRFFunc(String strParam, String strFuncName), and 

invoke the client stub. The parameter of Func is encoded to 

be the strParam parameter, and proxy client also specifies 

the invoking method by strFuncName parameter. 

3) Client stub sends the request WSRFFunc to UIS. 

V. PERFORMANCE TEST 

The proposed unified service platform is based on GT4 

Java WS Core, it provides unified WSRF service interfaces 

to accessing the resources on server through the IPR-IRR 

pair, and reduces the workload in service development and 

configuration. USP adds middleware to the interaction 

between service interfaces and the real resources, so the 

service performance would be affected consequentially. 

The performance test is executed to examine the 

performance lost and development efficiency increase. The 

test mainly inspected the response time to finish a service 

invoking. What is more, we compared the workload by 

measuring time consumption for a trained GT4 programmer 

to developing and deploying new Grid Service. 

A. Response time test 

We took a simple RR class ClassA with native method int 

Func(String str) for example. The Func directly returns an 

integer value, and does nothing else. 

Using GT4, the ClassA is instantiated in hard code by 

resource factory when Creating IPR, and service interface 

directly calls ��Func method and get the returned value. 


Figure 6. Consequent diagram of accessing Grid resources based on USP 

4) UIS finds the corresponding IPR of the client. 

5) USP invokes ��WSRFFunc method. 

6) 7) Resource manager queries resource database and 

��Func method according to the strFuncName 

parameter. 

8) UIS decodes the strParam parameter, and call the 

Func method. 

Finally, client receives the responds from server. 

The implement of the notification is a converse process. 

Under USP, the ClassA was recorded in resource database, 

USP queries the database, gets ClassA and its Func method 

utilizing Java reflection/introspection mechanism, and then 

invokes its method. It is necessary to code and decode the 

parameters in the interaction between IPR and instance of 

ClassA. 

The testing environment is described in Tab.1. We 

separately carried out experiments in Windows and Linux, 

and all the tests were executed in LAN connected by a 

switch. 

References [14] and [15] have mentioned that in 

Windows environment, the service response time seemed to 

be affiliated with the length of the parameters. If the data 

length is about between 200 bytes and 3000 bytes, the 

latency is relatively small, as long as out of this range, 

longer or shorter, the latency would be much heavier. So in 

Windows test, the length of str parameter was set to 500 

bytes. 

In Windows, we separately executed 50 tests under USP 

and GT4. In each test, we continually invoked the service 

interface for 100 times, measured the invoking response 

time and in the end of each test, the mean of all the time 

consumption values were calculated as the result of this test.


The test scheme in Linux was similar except that the 

length of str parameter was 5 bytes. 

server 

client 

TABLE I. TEST ENVIRONMENT 

CPU Pentium(R)4 2.6G 

memory 512M 

Operation system 

WinXP pro sp2 

Ubuntu8.04 

Grid toolkit GT4.0.8 

JDK Java-1.5.0 

Service container Tomcat5.5.29 

Ethernet 100M 

CPU Pentium(R)4 2.6G 

memory 512M 

Operation system 

WinXP pro sp2 

Ubuntu 8.04 

JDK Java-1.5.0 

Service container Tomcat5.5.29 

Ethernet 100M 

other switch DLink 100M 

Fig.7 shows the test data in Windows, while Fig.8 

indicates the test in Linux. Tab.2 analyses the data. From 

these two figures, we can see that the values of service 

response time under USP and GT4 are approximate. Tab.2 

shows that under USP, the mean service time in Windows 

rises from 19.27ms to 20.34ms, and the one in Linux rises 

from 37.84ms to 38.56ms comparing to the service directly 

utilizing GT4. The service performance loss is no more than 

6%. 

Figure 7. Service time consumption in Windows 


Figure 8. Service time consumption in Linux 

TABLE II. DATA IN WINDOWS 

Windows(ms) Linux(ms) 

max min mean max min mean 

GT4 17.2 13.77 19.27 41.72 35.34 37.84 

USP 29.09 14.82 20.34 43.26 36.32 38.56 

In fact, the performance of GT4 service is not stable for 

some reasons, just like what Fig.7 and Fig.8 have showed, 

the service response time always floats in a certain range. So 

a little service time rising is not meaningful to Grid 

Services. 

So, the negative effect on service performance taken by 

USP is not evident. 

B. Workload test 

The performance can be ensured, how about the 

workload? We separately measured the time consumption 

for a trained programmer to finish new Grid Service 

utilizing GT4, GDT and USP. The function of Grid Service 

is same to the one used in response time test. The comparing 

of time consumption is listed in Tab.3. We can see that the 

developing efficiency increases a lot. 

TABLE III. TIME CONSUMING COMPARE 

Tools USP GDT GT4 

Time 4.2 min 11.4min 37.8 min 

C. Test conclusion 

From the test, we can easily come to the conclusion that 

Grid Service development under USP is efficient and the 

outcome service performs almost as well as the service 

directly generated by GT4. 


A unified service platform for accessing Grid recourses is 

proposed in the paper. USP separates the service 

development and resource development, and provides


unified service interfaces to access different resources. The 

main workload for a USP user willing to deploy Grid 

resources is realizing the native methods and configuring the 

resource database. The resources under USP can be 

deloyed/undeployed dynamically. Besides the common 

client stub which can directly invoke the unified service 

published by USP, users can also achieve the proxy client 

through resource query&deploy service. The proxy client 

and proxy notifier could greatly facilitate the service users. 

The final experiment proofs that the service performance 

under USP is acceptable for Grid applications and 

development workload reduce a lot. 

This paper only proposes a prototype. The future works 

include management of the third-party resources used by 

RR, GUI for USP users, high QoS of USP, and so on. 


The authors wish to thank Dr. Xianguo Meng from 

Dept.2 of Mechanical Engineering College. 

REFERENCE 

[1] http://www.globus.org/toolkit/.(2010-10-11) 

[2] �� 

http://gdp.globus.org/gt4-tutorial/.(2010-2-5) 

[3] http://tomcat.apache.org. (2010-2-5) 

[4] Ken Arnold, Jams Gosling, David Holmes. The Java TM 

Programming Language, Fourth Edition. Pearson Education, Inc. 

2006 

[5] http://mage.uni-marburg.de/trac/gdt/wiki. (2009-12-3) 

[6] T. Friese, M. Smith, and B. Freisleben. GDT:A Toolkit for Grid 

Service Development. In Proc. of the 3rd International Conference 

on Grid Service Engineering and Management, pages 131-148, 2006 

[7] M. Smith, T. Friese, and B. Freisleben. Intra-Engine Service Security 

for Grids Based on WSRF. In Proceedings of the 2005 IEEE 

International Symposium on Cluster Computing and Grid 

��s 644-653, 2005. 

[8] Kay Dornemann and Bernd Freisleben. Discovering Grid Resources 

and Deploying Grid Services Using Peer-to-Peer Technologies. In 

Proceedings of the 2009 International Conference on Advanced 

Information Networking and Applications Workshops, pages 

292-297, 2009 

[9] Eun-Kyu Byunt, Jae-Wan Jangt, Wook Jungt et al. A Dynamic Grid 

Services Deployment Mechanism for On-Demand Resource 


Provisioning. In Proceedings of the 2005 IEEE International 

Symposium on Cluster Computing and the Grid, pages 863-870, 

2005 

[10] Jon B. Weissman, Seonho Kim, and Darin England, Supporting the 

Grid Service Dynamic Lifecycle. In Proceedings of the 2005 IEEE 

International on Cluster Symposium and Computing the Grid, pages 

808- 815,2005 

[11] Fu Qiang Li, Bin Gong, Cheng Xing. Dynamic Visualization Service 

Deployment in Grid Scientific Workflow. In Proceedings of the 2008 

Seventh International Conference on Grid and Cooperative 

Computing, pages 201-205, 2008 

[12] Pu Liu, Michael J. Lewis. Unified Dynamic Deployment of Web and 

Grid Services. In the Proceedings of 2007 IEEE Conference on Web 

Service, pages 26-34, 2007 

[13] L. Qi, H. Jin, I. Foster, and J. Gawor. HAND: Highly Available 

Dynamic Deployment Infrastructure for Globus Toolkit 4. In the 

Proceedings of 15th EUROMICRO International Conference on 

Parallel, Distributed and Network-Based Processing, pages155-162, 

2007 

[14] Shaochong Feng, Yanqiang Di, Yuanchang Zhu, et al. Developing 

WSRF-Based Web Service RTI Using GT4. In Proceedings of 2009 

First International Workshop on Education Technology and 

Computer Science, pages 1066-1069, 2009 

[15] Feriese,Thomas.��-�� 

[Doctoral dissertation], Fachbereich Mathmatik und Informatik 

Universit Marburg ,2006 

Shaochong Feng was born in Hebei, China. He receives his 

Master�s degree in Guidance, Navigation and Control from 

Mechanical Engineering College in 2007. His technical 

interests include distributed modeling and simulation, Grid 

computing and Cloud computing. 

Yuanchang Zhu was born in Heilongjiang, China. He 

received the B.A., M.A., and Ph.D in 1982, 1988 and 2005. 

He is a professor at Mechanical Engineering College. His 

study interests include M&S, Cloud Computing and so on. 

Yanqiang Di was born in Hebei, China. He received the 

B.A., M.A., and Ph.D in 1995, 1998 and 2009. He is a 

teacher at Mechanical Engineering College. His study 

interests include M&S, Grid Computing and database 

system.


Research on an Improved Terrain Aided 

Positioning Model 

Li Shidan 

Dept. Electronic Engineering, Tsinghua University, Beijing, China 

Email: lisd06@mails.tsinghua.edu.cn 

Sun Liguo, Li Xin, Wang Desheng 

Dept. Electronic Engineering, Tsinghua University, Beijing, China 

Email: {slg00, xinli05, wangdsh_ee}@mails.tsinghua.edu.cn 

Abstract—Terrain aided positioning (TAP) is a kind of 

positioning method which acquires position information 

from the terrain elevation datum underneath the vehicle. 

This method has the characteristics of autonomy, allweather, 

anti-interference, strong stealthiness and high 

accuracy. It is widely used in the navigation system for 

various aircrafts, cruise missiles and underwater vehicles. 

The fundamentals of TAP is that it firstly measures the 

terrain elevation underneath the vehicle using relevant 

sensors, then compares these datum with the referenced 

Digital Elevation Map (DEM) and acquires the position 

information through matching algorithm. The system model 

for TAP currently used totally depends on the referenced 

DEM and the position acquired is the position referenced to 

the map rather than the true position. Due to the DEM 

error which is introduced during production procedure, the 

position on the map is not the real position. In order to 

overcome the problem, the paper proposes an improved 

TAP model which introduces the map error into the system 

model and gets the recursive solution based on the Bayesian 

framework which is numerically solved by RPF particle 

filter. From the simulation results, the new model has 

extraordinary performance for handling the error of DEM 

and the algorithm can estimate the map error and acquire 

the accurate position. 

Index Terms—Terrain Aided Positioning, non-linear 

estimation, Bayesian iteration, Particle Filter, RPF 


Positioning and navigation system plays an important 

role on any vehicle no mater aerial, surface or underwater. 

The positioning system widely used nowadays can be 

divided into two categories: non-autonomy system and 

autonomy system. 

Non-autonomy system, represented by the satellite 

based systems, such as GPS and GNSS can give accurate 

position information by receiving signals from external 

devices. These satellite based systems can easily be 

jammed by electromagnetic interference and even the 




doi:10.4304/jsw.6.5.937-943 

satellite itself may be attacked during war time. 

Meanwhile, due to the rapid attenuation of 

electromagnetic wave through sea water, such satellite 

based systems are not suitable for underwater vehicles. 

Autonomy system, include inertial navigation system 

(INS) and terrain aided navigation system (TAN), does 

not need external devices for positioning and has the 

characteristics of all-weather, anti-interference, strong 

stealthiness and so on. They are widely used in many 

kinds of military vehicles for main or backup positioning 

system. Although INS can give high accurate positioning 

information during short time period, it has time 

accumulated errors which should be corrected by other 

systems, such as TAN or GPS, during work time, while 

the positioning error of TAN system mainly depends on 

the complexity and similarity of the terrain. TAN can also 

give high accurate position information from rugged 

terrain. For example, the TERPROM system widely used 

on NATO’s military aircrafts has horizontal position 

accuracy of 10~25 meters [1]. 

From the introduction above, it can be inferred that the 

autonomy poisoning system, especially TAN system, 

plays an important role on military usage during war time 

and may be the only choice for underwater vehicles. 

However the terrain aided positioning system currently 

used severely depends on the accuracy of the referenced 

Digital Elevation Map (DEM) which can be introduced 

with errors during producing [2] and finally leads to large 

differences between true position and estimated position. 

The paper just focuses on such problem and proposes an 

improved system model which introduces the map error 

components to the system model. The paper gives 

feasibility proof for the new model based on Bayesian 

theory and uses RPF particle filter for numerically getting 

the result. The simulation results confirm that the new 

model can estimate the map error and give more accurate 

position information compared to the basic model. 

II. SYSTEM MODEL 

The paper mainly discusses the terrain aided 

positioning system for aircrafts, but it can be easily 

extended to the underwater systems.


The airborne sensors for measuring terrain underneath 

the aircrafts are composed of barometric altimeter and 

radar altimeter. Fig. 1 shows the measurement procedure. 

The barometric altimeter measure the aircraft’s altitude 

above sea level while the radar altimeter gets the 

elevation above the ground, and the difference between 

the two makes the terrain elevation underneath the 

vehicle. During the measuring, the pressure fluctuation 

and atmospheric turbulence may affect the barometric 

altimeter while the terrain roughness and the ground 

vegetation may influence the radar altimeter. There are 

several papers discussed such error aspects [3, 4, 5]. This 

paper concentrates on the influence caused by the DEM 

error and does not introduce the sensor error components 

into the system model. 

Figure 1. Terrain elevation measurement. 

The basic model for terrain aided positioning is 

⎧ k + = k + k 

⎪ 

⎨ k = k + k 

⎪ 

⎩y 

k = h ( k ) + v 

* * 

* * 

e 1 e w 

* * 

x x e 

x 

where * 

e k is the position error of INS at time k which is 

modeled as time accumulated error. x k is the horizontal 

position from INS which is the true position 

k 

(1) 

* 

x k plus INS 

error * 

e k . y k is the terrain elevation measurement which 

is the true terrain elevation ( ) 

* * 

h x plus the 

measurement noise v k . The process noise w k and 

measurement noise vk are independent with each other 

and also independent with the state * 

e k . They have 

Gaussian distribution w k ~ N( 

0, 

Qk 

) , v k ~ N( 

0, 

Rk 

) . 

* 

Due to the strong non-linearity of the terrain h ( ⋅) 

, 

this model is classified into the non-linear model and the 

terrain aided positioning belongs to the typical non-linear 

estimation problem. 

* 

In the basic model above, h ( ⋅) 

represents the terrain 

elevation map. Because the map component is denoted as 


k 

( ) 

* * 

h x k in the measurement equation, the estimated 

* 

position x ˆ k which is corrected by the * 

e ˆ k through x k is 

* 

the position referenced to the map h ( ⋅) 

. That is to say, 

* * 

if h ( ⋅) 

is the real map then x ˆ k is the true position 

* * 

whereas if h ( ⋅) 

is the map with errors then x ˆ k is just 

* 

the position on h ( ⋅) 

rather than the true position. Since 

the real map can not be acquired, the basic model can 

only estimate the position referenced to the map rather 

the true position. Fig. 2 depicts the relationship of the 

values in the procedure. From the correction by the TAN, 

the system can give the relative position between the 

aircraft and the mountain and the error of that position 

compared to the true position is just the map error. 

Usually the relative position is enough for anti-collision 

usages. However the true position is also preferred in 

many occasions. So a more complicated model should be 

built to estimate the map error from time to time. Hence, 

we need to introduce the map error into the system model. 

Figure 2. Relationship of the values in TAN. 

To make the discussion simple, we assume that the 

map errors have the regional stability, thus they can be 

treated as constant parameters in a certain area. 

Let Δ Hk 

, Δ Vk 

be the horizontal and vertical errors 

of terrain elevation map respectively where Δ H k 

composes of two components of x and y directions. Then 

the TAN referenced elevation map can be expressed as 

* 

h( ) = h ( x + ΔH 

) + ΔV 

x (2) 

* 

k 

* 

k 

* 

where h ( ⋅) 

represents the real terrain elevation map, 

* 

x k is the real position. 

Let x = xk 

+ ΔHk 

* 

* 

, substitute to equation (2), we get 

h ( x ) = h( 

x − ΔHk 

) − ΔVk 

(3) 

Substitute (3) to the basic model (1), we get the new 

measurement equation using h (⋅) 

as 

y k = h( 

k − Δ k ) − ΔVk 

+ vk 

* 

x H 

(4) 

Then we can get the new system model containing the 

map error components: 

k 

k


⎧ k + = k + k 

⎪ 

⎪ 

⎪Δ 

k + = Δ k 

⎨ΔVk 

+ = ΔVk 

⎪ 

= + 

⎪ k k k 

⎪⎩ 

yk 

= h( 

k − Δ k ) − ΔVk 

+ vk 

* 

* * 

e 1 e w 

H 1 H 

1 

* * 

x x e 

x H 

(5) 

where equations Δ H k +1 = ΔH 

k and Δ Vk +1 = ΔVk 

are 

the reflections of the assumption above, meaning that the 

error components are constant in the area concerned. 

The compact form of (5) is 

⎧ k + = k + k 

⎪ 

⎨ k = k + ⋅ k 

⎪ 

⎩y 

k = h( 

k − Δ k ) − ΔVk 

+ vk 

* 

' 

e 1 e w 

* 

x x F e 

x H 

(6) 

* ⎡ e ⎤ k 

⎢ ⎥ ⎡1 

where ek 

= ⎢ΔHk 

⎥ , F = ⎢ 

⎢ ⎥ 

⎣ΔV 

⎣0 

k ⎦ 

0 

1 

0 

0 

0 

0 

⎡wk 

⎤ 

0⎤ 

' ⎥ 

0 

⎥, w 

⎢ 

k = 

⎢ 

0 . 

⎥ 

⎦ 

⎢⎣ 

0 ⎥⎦ 

III. RECURSIVE BAYESIAN ESTIMATION 

According to Bayesian theory, a Bayesian estimation 

problem is defined by the joint density of the parameters 

and the observations, p ( x, 

y) 

= p( 

y | x) 

p( 

x) 

. The 

estimator under the minimum mean square error criterion 

is the posterior mean x ˆ MS = ∫ xp( 

x | y) 

dx [6]. So if 

n 

R 

there exists some relationship between the observation 

and the parameter, then the parameter can be estimated 

by the observation. From the information theory’s point 

of view, the observation contains the information of the 

parameter when there exists stochastic relationship 

between the two. So we can make use of the observation 

to estimate the parameter. 

Let Y k be the augmented measurement vector 

consisting of all the measurements up to time step k. 

From Bayesian formula [6] and the new system model (6), 

we have the posterior probability density function update: 

p( 

yk 

| ek, 

Yk 

−1) 

⋅ p( 

ek 

| Yk 

−1) 

p( 

ek 

| Yk 

) = 

p( 

y | Y ) 

(7) 

k 

k−1 

−1 

* 

= αk 

⋅ pv 

( yk 

−h( 

xk 

−ek 

−ΔH 

k) 

+ ΔVk 

) ⋅ p( 

ek 

| Yk 

−1) 

k 

where 

* 

α k = ∫ pv ( yk 

− h( 

x 

N k − ek 

− ΔHk 

) + ΔVk 

) ⋅ p( 

ek 

| Yk 

⋅ dek 

R k − 1) 

. 

The priori probability density function update is 

p( 

ek 

+ 1 | Yk 

) = ∫ p( 

ek 

ek 

Yk 

p ek 

Yk 

de 

N + 1 | , ) ⋅ ( | ) ⋅ k 

R 

. (8) 

* * 

= p ( e − e ) ⋅ p( 

e | Y ) ⋅ de 

∫ 

N w 

R k 

k+ 

1 

Given the initial prior density p ( e0 | Y− 

1) 

= p( 

e0) 

, 

we can recursively generate the posterior probability 

density through equation (7) and (8). And the state 


k 

k 

k 

k 

estimate is e ˆ k = E[ ek 

| Yk 

] with covariance matrix 

ˆ 

T 

ˆ ) ( ˆ 

k E[( k k k k ) | Yk 

] e e e e P − ⋅ − = . 

The recursive Bayesian equations above are the 

theoretical solution for model (6) and are intractable due 

to the complexities of the posterior and priori probability 

density function in the non-linear model. Usually the 

numerical methods should be used for calculating the 

result, such as point mass filter (PMF) or particle filter 

(PF). This paper uses the PF to solve the model. 

IV. PARTICAL FILTER 

The fundamental of particle filter is to use particles 

with weights for representing the probability density 

function (PDF) and uses the recursion of the particle set 

to replace the recursion of the posterior density function. 

When the complex probability density function is 

represented by the particle set the integration in α k and 

(8) can be easily calculated according to Monte Carlo 

integration theory. 

i i M 

Let { e k , wk} i= 

1 be particle set for the posterior PDF 

M 

∑ 

i= 

1 

i 

p( e k | Yk 

) , where w = 1 , then 

k 

k 

k 

M 

∑ 

i= 

1 

i 

p( 

| Y ) ≈ w ⋅ ( e − e ) 

e δ (9) 

where δ (⋅) 

is Dirac-Delta function. The estimator and its 

covariance are 

= 

= 

k 

∑ 

eˆ 

k 

= E[ 

e 

= 

= 

∫ ek 

⋅ N ∑ 

R 

M 

∑ 

i= 

1 

k 

| Y ] 

i 

k 

k 

M 

i= 

1 

w ⋅e 

i 

k 

w 

i 

k 

T 

∫ ( ek 

−eˆ 

k) 

⋅( 

ek 

−eˆ 

k) 

⋅ 

N ∑ 

R 

M 

i= 

1 

⋅ 

i= 

1 

i i 

i T 

w ⋅( 

e −eˆ 

) ⋅( 

e −eˆ 

) 

k 

k 

k 

k 

k 

k 

k 

k 

k 

M 

k 

( e 

T 

Pˆ 

= E[( 

e −eˆ 

) ⋅( 

e −eˆ 

) | Y ] 

k 

k 

i 

k 

i 

− e ) ⋅ de 

δ (10) 

k 

k 

i 

i 

w ⋅δ 

( e −e 

) ⋅de 

(11) 

The recursion of the particle set is simply explained 

below: 

i i M 

Let { e k−1 , wk−1} i= 

1 be the particle set at time k-1 

which represents the posterior PDF p( e k −1 | Yk 

−1) 

. At 

time k, we first draw sample i 

e k from an easy sampling 

i 

distribution q( e k | ek 

−1, 

yk 

) and then update its weight 

using 

w 

i 

k 

∝ w 

i 

k −1 

i i i 

p( 

yk 

| ek 

) ⋅ p( 

ek 

| ek 

⋅ 

i i 

q( 

e | e , y ) 

k 

k 

k −1 

k 

k 

k 

−1 

) 

k 

k 

(12)


i i M 

Then the new particle set { e k , wk} i= 

1 represents the 

posterior PDF p( e k | Yk 

) . Thereby this simple particle 

set recursion takes place of the intractable recursion of 

the posterior PDF. 

i 

The distribution q( e k | ek 

−1, 

yk 

) is called importance 

sampling density function which can be selected 

according to requirements. Usually for convenient usage, 

i 

the state transition PDF p( e k | ek 

−1) 

is chosen [7]. 

i 

Substitute p( e k | ek 

−1) 

into (12) yields 

i i 

i 

wk ∝ wk 

− 1 ⋅ p( 

yk 

| ek 

) 

(13) 

Because the last three components of the state vector in 

the new model are constant parameters, we need special 

treatment for random components and constant 

components respectively during particle recursion. We 

use important sampling density function for random 

components and leave constant components unchanged 

during particle transition: 

* i * * 

⎡e 

⎤ 

k ~ p( 

ek 

| ek 

−1) 

i ⎢ i i ⎥ 

e k = ⎢ ΔHk 

= ΔHk 

−1 

⎥ (14) 

⎢ i i ⎥ 

⎣ ΔVk 

= ΔVk 

−1 

⎦ 

* * 

* 

p( ek 

| ek− 

1) 

~ N( 

ek−1, 

Qk 

) (15) 

i 

* i i i 

p( 

y | e ) = p ( y −h( 

x −e 

−ΔH 

) + ΔV 

) (16) 

k 

k 

vk 

k 

Equations (13) to (16) finally accomplish the particle 

filter recursion for map error model (6). 

In order to decrease the impacts on PF performance 

caused by the phenomenon of particle degeneracy and 

particle collapse, we need some resampling scheme for 

effective representing the PDF. 

Since our new model (6) consists constant parameter 

which can be seen as random variable with extremely 

small process noise, common particle filters such as SIS, 

ASIR are not suitable for handling the model with small 

process noise states which can lead to severe particle 

degeneracy phenomenon due to the lose of particle 

diversity [7]. In this paper, we chose Regularized Particle 

Filter (RPF) [8] for resampling which can maintain the 

particle diversity to the maximum extent. 

During resampling, we actually resample from the 

discrete distribution 

k 

k 

M 

∑ 

i= 

1 

i 

p( 

| Y ) ≈ w ⋅ ( e − e ) 

k 

k 

k 

e δ (17) 

which makes that the new samples cannot get rid of the 

old particle set and leads to singular composition after 

several iterations. The main idea of RPF is to make the 

PDF continuous by introducing kernel function and let 

the particle evolve in the continuous space. The 

resampling distribution function for RPF is 

M 

k 

k 

i 

k 

i 

p( x | Y ) w ⋅ K ( x − x 

k 

k 

≈ ∑ 

i= 

1 


k 

h 

k 

i 

k 

) 

k 

(18) 

1 

Kh 

x) 

= nx 

h 

⎛ x ⎞ 

K⎜ 

⎟ 

⎝ h ⎠ 

where x 

> 

of kernel function K (⋅) 

. (⋅) 

( (19) 

n is the dimension of x, h 0 is the bandwidth 

K can be seen as a 

symmetric probability density function on 

K (⋅) 

is called the rescaled kernel. 

h 

x n 

R and 

The kernel and the bandwidth are chosen so as to 

minimize the mean integrated square error between the 

true posterior density and the corresponding regularized 

weighted empirical measure in (18). In a special case of 

equally weighted sample, the optimal choice of the kernel 

is the Epanechnikov kernel [8]. To reduce computing cost, 

we use Gaussian kernel instead and the corresponding 

optimal bandwidth is [9] 

1 

− 

n x + 4 

h = A ⋅ N 

(20) 

opt 

1 

x + 4 

with = ( 4/( 

+ 2)) 

n 

A nx 

. 

For implement, the new particle set can be generated 

by 

i* 

i 

i 

x k = xk 

+ hoptDkε (21) 

where D k is the square root of the empirical covariance 

i i M i 

matrix of the samples { x k , wk} i= 

1 . ε is the sample 

drawn from the kernel function. 

The resampling procedure of RPF is 

--------------------------------------------------------------------- 

� Calculate the effective number of particles N eff [10] 

N < N 

� IF eff thr 

� Calculate the empirical covariance matrix S k 

i i M 

for particle set { x k , wk} i= 

1 

� Calculate the square root Dk of S k 

� Resample the particle set using Systematic 

Resampling [11] method, and get the new set 

{ x , , −} 

= 

i i M 

k wk i 1 

� FOR i=1:M 

i 

� Draw sample ε from kernel 

i* 

i 

i 

� Update particle x k = xk 

+ hoptDkε � END FOR 

� END IF 

--------------------------------------------------------------------- 

V. SIMULATION RESULTS 

The reference DEM data for the simulation is from 

ASTER GDEM produced by METI and NASA, which 

has grid size of 30 meters [12]. The area we use is 

between 40 and 41 degrees latitude north and 105 and 

106 degrees longitude west.


In our simulation, we use the DEM from ASTER 

* 

GDEM to be the real terrain h ( ⋅) 

and add some map 

errors to get the reference DEM used for filtering. We use 

bilinear interpolation method to draw elevation data from 

the grid map. 

The simulation procedure is as follows: 

--------------------------------------------------------------------- 

Step 1, chose a trajectory on the real map to be the 

* N 

reference trajectory and get { x k} k = 1. 

N 

Step 2, get the output of the INS { x k} k = 1 by equation 

* * ⎧ek 

+ 1 = ek 

+ w k 

⎨ 

and the initial INS error 

* * 

⎩xk 

= xk 

+ ek 

* 

e 0 . 

Step 3, given the initial distribution of state p e ) , 

we can get estimate 

and acquire 

{ ˆ = 

* N 

k} k 1 

{ ˆ } = 

k 

N 

k 1 

( 0 

e through RPF particle filter 

x using equation 

xˆ − k 

* * ˆ 

k xk 

e 

= . 

Step 4, repeat Step 2 and Step 3 to do Monte Carlo 

simulation several times and get average estimation error 

to evaluate the system performance. 

--------------------------------------------------------------------- 

The map errors used in our simulation are 

T 

ΔH = [ 200, 

150] 

, ΔV = 30 . 

The DEM and the flight path used are depicted in Fig. 

3. This area has a mountain from north to south and the 

flight path we chose is a uniform speed trajectory with a 

turn round in the middle of the path. Fig. 3 shows an 

estimated result for map error model. 

Y (m) 

3 

2.5 

2 

1.5 

1 

0.5 

x 10 4 

Start 

End 

Estimated Path 

Real Path 

0 1 2 3 4 5 6 

X (m) 

x 10 4 

0 

Figure 3. Flight path and estimation result for map error model. 

Fig. 4 is the average root mean square error (RMSE) 

curve for horizontal position estimation which is 

generated by using 100 Monte Carlo simulations. The 

solid line is the RMSE of the map error model while the 

dash line is for the basic model. Because the map error 

model actually estimates the map errors and corrects the 

position with them, so it can get more accurate position 

information. From the figure, the horizontal RMSE of the 

new model is about 200 meters smaller than the basic 

model which is close to the horizontal map error we set, 


thus confirms the new model’s capability of map error 

estimation. However, the new model has much slower 

convergence speed than the basic model which may 

caused by the high dimensionality of the state vector in 

the new model. 

RMSE (m) 

10 5 

10 4 

10 3 

Map Error Model 

Basic Model 

10 

0 50 100 150 200 250 300 

2 

Time (s) 

Figure 4. Horizontal RMSE with MC=100. 

Fig. 5 shows the terrain elevation estimation error for 

both map error model and the basic model. The terrain 

elevation estimation for map error model is acquired by 

* * * 

h ( x k ) ≈ h( 

xˆ 

k − ΔHˆ 

k ) − ΔVˆ 

k (22) 

* 

where x = x 

* 

− eˆ 

, while for basic model 

ˆ k k k 

* * * 

h ( x ) ≈ h( 

xˆ 

) . (23) 

k 

Apparently, the new model takes map errors (both 

horizontal and vertical error) into consideration and uses 

the estimated error to correct the measurement while the 

basic model just depends on the map itself. So from the 

simulation results shown in Fig. 5, the error of basic 

model is about 20 meters higher than the new model 

which is close to the vertical error of the map we set. 

RMSE(m) 

10 3 

10 2 

10 1 

10 

0 50 100 150 200 250 300 

0 

Time(s) 

k 

Map Error Model 

Basic Model 

Figure 5. Terrain Elevation Estimation Error with MC=100. 

Fig. 6 is the average map error estimate with 100 

Monte Carlo simulations. The dash line is the true error 

while the solid curve is the estimated parameter. From the 

figure, the filter is best for the vertical error estimation 

which has rapid convergence, high accuracy and good 

stability. That because the vertical error component has a


linear structure in the system model that can be even 

extracted from the system model and use Kalman filter 

for solving, such as Rao-Blackwellize method [13]. So 

when using particle filter, such component with simple 

structure can be easily estimated. For horizontal map 

errors, the convergence of Y direction is worse than the X 

direction, that because the terrain in the second half path 

varies more on the X direction than Y direction which 

benefits the component estimation on X direction. 

Meanwhile, the convergent speed for horizontal errors is 

very slow which may caused by the constant 

characteristic of these states. 

220 

200 

180 

Horizontal X direction 

0 50 100 150 200 250 300 

Horizontal Y direction 

200 

150 

100 

0 50 100 150 

Vertical 

200 250 300 

200 

100 

0 

0 50 100 150 200 250 300 

Figure 6. Map error estimation. 

Fig. 7 and Fig. 8 show the particle evolution process 

for SIS and RPF respectively for comparison of the 

effectiveness of these two different algorithms. The 

figures show the histogram of one component of the state 

in the particle set at different time step which can be seen 

as the distribution of that component. The component we 

chose to show is the X direction error of the map error 

component which is a constant parameter in the state 

vector. 

Fig. 7 is for SIS. As mentioned above, the commonly 

used particle filter is not suitable for solving the model 

with constant parameters. For these parameters the initial 

particle set at k=1 contains the whole data values for the 

evolution that there will be no new values generated in 

later time step since they have no process noise. So the 

initial distribution must cover the true value we estimated 

otherwise the filter can not give that value. From Fig. 7, 

after several iterations the distribution is concentrated to 

some distinct values and the state can hardly move to 

other values. 

In Fig. 7, when k=1, the initial distribute is a Gaussian 

distribution with mean equal to 180 and can cover the 

true value of 200 which is the map error we set. After 

some iterations, the amount of effective values decreases 

and when k=16 there are only two bars in the histogram 

which do not contain the true value thus after that the 

filter can not give the accurate value of 200. 

In SIS, since the parameter components in the particle 

do not change during state transition step, the degeneracy 

phenomenon of these components will affect the 

evolution of the other state components and lead to slow 


convergent speed, low accuracy and even divergent when 

the particles fall into bad bars. 

x 104 

k=1 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

x 104 

k=6 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

x 104 

k=9 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

x 104 

k=16 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

Figure 7. particle evolution for SIS. 

Fig. 8 is for RPF. RPF uses an effective resampling 

scheme which draws new particles from a continuous 

distribution constructed from the discrete one. It can 

make the singular bar extent to a region according to 

some distribution which can generate new values during 

evolution. From Fig. 8, the distribution of the parameter 

moves towards the true value with time and concentrate 

to the true value when convergent. 

x 104 

k=1 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

x 104 

k=7 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

x 104 

k=29 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

x 104 

k=37 

4 

2 

0 

80 100 120 140 160 180 200 220 240 260 280 

Figure 8. particle evolution for RPF. 

In Fig. 8, the initial distribution is also Gaussian. When 

k=7, the mean of the parameter move away from 180 and 

the right side of the distribution expands. The mean 

becomes 190 when k=29 and finally gets to 200 which is 

the true value when k=37. During each time step, RPF


can maintain the particle diversity to the maximum extent 

and the particle set can move towards the right direction 

as long as the true value is covered by the distribution. 

However, this moving procedure is much slower than SIS. 

According to the comparison above, the common SIS 

particle filter can hardly handle the constant parameter 

estimation in our model and the particle degenerate 

phenomenon always occurs after several iteration steps. 

When using the resampling scheme of RPF, the particle 

diversity can be maintained and the particle set can move 

to the true value. Meanwhile, we found that the constant 

parameter incorporated in the complex non-linear system 

is hard to estimate, the filter is slow on convergence 

speed and sensitive to the initial value. 

V. CONCLUSION 

In this paper, the system model of terrain aided 

positioning system is studied and an improved system 

model is proposed which overcomes the disadvantage of 

the dependency on the accuracy of the map that exists in 

the basic model. The new model can estimate map error 

and correct the position to acquire more accurate position 

information. The paper selects particle filter for this 

nonlinear model and compares the performance of SIS 

and RPF particle filters. From our simulation, the RPF 

has much better performance for our new model which 

contains constant parameter in the state vector. With RPF, 

our simulation results confirm the better performance of 

the new model than the basic model that the accuracy of 

the horizontal position estimation is improved by around 

200 meters which is close to the map error we set. 

With this map error model, our terrain aided 

positioning system can give actual position information 

other than just the position on the map. This progress will 

make the system more preferable for use. 


We should thank Jia Ke for his DEM data acquisition 

and preprocessing work. And we also appreciate the 

institutions of METI and NASA for their opening support 

of ASTER GDEM datum. 

REFERENCES 

[1] M. Cowie, and N. Wilkinson, “Latest development of the 

TERPROM (R) Digital Terrain System (DTS) ,” 2008 

Ieee/Ion Position, Location and Navigation Symposium, 

Vols 1-3: 658-668, 2008. 

[2] K. J. Markham, and W. A. Morris, “Digital terrain 

elevation models produced using radar altimetry and 

GPS data,” Igarss 2002: Ieee International Geoscience 

and Remote Sensing Symposium and 24th Canadian 

Symposium on Remote Sensing, Vols I-Vi, Proceedings: 

2723-2725, 2002 

[3] K. B. Anonsen, and O. Hallingstad, “Terrain aided 

underwater navigation using point mass and particle 

filters,” 2006 IEEE/ION Position, Location and 

Navigation Symposium, Vols 1-3: 1027-1035, 2006. 

[4] P. J. Nordlund, and F. Gustafsson, “Recursive Estimation 

of 3-Dimensional Aircraft Position Using Terrain-Aided 

Positioning,” Linkoping, 2001. 


[5] P. Frykman, “Applied particle filters in integrated 

aircraft navigation,” Electrical Engineering, Linkoping 

University, 2003. 

[6] J. M. Bernardo, and A. F. M. Smith, Bayesian Theory, 

2nd ed., New York: Wiley, 1998. 

[7] B. Ristic, S. Arulampalam, and N. Gordon, Beyond the 

Kalman Filter, Artech House, Boston, London, 2004. 

[8] B. W. Silverman, “Density Estimation for Statistics and 

Data Analysis,” Vol. 26 of Monographs on Statistics and 

Applied Probability, Chapman & Hall, London, 1986 

[9] C. Musso, N. Oudjane, and F. LeGland, “Improving 

regularised particle filters,” in Sequential Monte Carlo 

Methods in Practice, New York: Springer, 2001. 

[10] A. Kong, J. S. Liu, and W. H. Wong, “Sequential 

imputations and Bayesian missing data problems,” 

Journal of the American Statistical Association, vol. 89, 

no. 425, pp. 278-288, 1994. 

[11] G. Kitagawa, “Monte Carlo filter and smoother for non- 

Gaussian non-linear state space models,” Journal of 

Computational and Graphical Statistics, vol.5, no. 1, pp. 

1-25, 1996. 

[12] “http://asterweb.jpl.nasa.gov/,” January, 2011. 

[13] A. Doucet, N. d. Freitas, K. Murphy, and S. Russell, 

“Rao-Blackwellised particle filtering for dynamic 

bayesian networks,” Proceedings UAI2000: 176-183, 

2000. 

Li Shidan, born in 1983, Ph. D. 

candidate in Department of Electronic 

Engineering, Tsinghua University. His 

research interests mainly focus on 

particle filters, navigation technique and 

radar signal processing. 

He is currently doing research at the 

High-speed Signal Processing and 

Network Transmission Institute in 

Tsinghua University. He has 

participated in the terrain aided navigation project and the 

marine navigation radar system project. 

Sun Liguo, born in 1982, Ph. D. candidate in Department of 

Electronic Engineering, Tsinghua University. His research areas 

include navigation and positioning technique, signal processing 

on Geographic Information System. 

Li Xin, born in 1967, Ph. D. candidate in Department of 

Electronic Engineering, Tsinghua University. Her interests 

focus on the target tracking technique and aircraft’s navigation 

system. 

Wang Desheng, born in 1946, Prof. in Department of 

Electronic Engineering, Tsinghua University. He is doing 

research on signal processing and framework techniques on 

radar display and control terminal. He is one of the first for 

developing raster scan radar terminal and the frequency jumper 

radar system.


Research on Integrated Information Platform of 

Agricultural Supply Chain Management Based on 


Yan-e Duan 

Beijing University of Agriculture Computer and Information Engineering College, Beijing, China 

Email: duanyane@126.com 

Abstract—In recent years, accidents in food quality & safety 

frequently occur, and more and more people have begun to 

think highly of food quality & safety and encourage food 

producers to be able to trace the origin of ingredients and 

process of production along the supply chain. With the 

development of IT, more and more practices have shown 

that the supply chain of agricultural products should rely on 

IT. The using of IT directly decides the degree of 

agricultural informatization and efficiency of agricultural 

supply chain management. In this paper, on the basis of 

introducing the meanings and characteristics of supply 

chain management and agricultural supply chain 

management, it also analyzes the information flow's 

attributes throughout the process of agricultural supply and 

the technological attributes of Internet of Things, finally, the 

designing method and architecture of integrated 

information platform of agricultural supply chain 

management based on internet of things was discussed in 

detail. 

Index Terms—supply chain management, agricultural 

product, Internet of Things, RFID 


With 9.5% of world arable land and freshwater 

resource that is only 31% of world average; China has 

successfully fed 22% of world population From 1978 

until 2006, the Engel Coefficients of Chinese urban and 

rural residents have declined respectively from 57.5% 

and 67.7% to 35.8% and 43%, and the food of the 

Chinese has also turned from shortage to abundance. The 

above development shows that China has contributed 

greatly to world food supply. However, the 

developmental process of food supply has the 

characteristics of “pollution first and then elimination”, 

which means the process is inevitably accompanied with 

relatively serious problems of pollution and food safety [1] . 

Specially, in recent years, the frequency and severity of 

product recalls has been increasing over the past decade. 

Notable examples of recent ‘global’ product recalls 

include the 2008 pork dioxin recall in Ireland, the 2008 

melamine tainted milk recall in China, and especially the 

global recalls caused by Bovine Spongiform 

Encephalopathy (BSE). More and more people have 

begun to think highly of food quality & safety and 

encourage food producers to be able to trace the origin of 

ingredients and process of production along the supply 

chain [2] 

. Agricultural supply chain management is 

becoming one of the hot issues in the researches on 


doi:10.4304/jsw.6.5.944-950 

supply chain management and causes wide concern of 

scholars at home and abroad. The research of agricultural 

supply chain management would optimize the deploying 

of agricultural resource, lower the risk of agriculture, 

improve the efficiency of agricultural production and 

advance the sustainable development of agriculture. 

Following the social infomationization, the Internet is 

often viewed as the means to reduce income disparities 

between the urban and rural populations in the 

developing world. Specially, with the development of 

information technology (IT), numerous innovative of IT 

based applications have emerged around the world that 

promise to bridge the digital divide. In the field of 

agriculture, technological innovation and competition 

have led to improvements in supply chain management 

for food products. More and more practices have shown 

that the supply chain management of agricultural 

products should rely on network information technology 

to improve its level of infomationization, internetization 

and intelligentization. 

II. CONCEPT OF SUPPLY CHAIN MANAGEMENT 

A supply chain is a system of organizations, people, 

technology, activities, information and resources involved 

in moving a product or service from supplier to customer. 

The Council of Supply Chain Management Professionals 

(CSCMP) defines Supply Chain Management (SCM) as 

follows: "Supply Chain Management encompasses the 

planning and management of all activities involved in 

sourcing and procurement, conversion, and all logistics 

management activities. Importantly, it also includes 

coordination and collaboration with channel partners, 

which can be suppliers, intermediaries, third-party service 

providers, and customers. In essence, supply chain 

management integrates supply and demand management 

within and across companies. Supply Chain Management 

is an integrating function with primary responsibility for 

linking major business functions and business processes 

within and across companies into a cohesive and highperforming 

business model" [3] . The advantages of supply 

chain management are numerous, like the reduction of 

product losses, increase in sales, reduction of transaction 

costs, a better control of product quality and safety and 

the dissemination of technology, capital and knowledge 

among the chain partners. A successful SCM 

implementation is expected to enhance the relationship 

between upstream suppliers and downstream customers.


Its objective is to produce and distribute the commodity 

in the right quantity, to the right place, and at the right 

time to minimize overall cost while maintaining customer 

satisfaction and firm performance [4] 

(Fig.1). 

Farmer 

Farmer 

Internal supply chain 

Supplier Purchaser Manufacturer Distributor Customer 

Information flow 

Logistics 

Capital flow 

Figure 1. Flow of Supply Chain 

Over the past two decades, SCM, emphasizing the 

interdependence of buyer and supplier firms working 

collaboratively to improve the performance of the entire 

supply, has generated extensive interest in both academic 

and practitioner communities (Shin et al., 2000; 

Narasimhan and Kim, 2007). A range of new supply 

chain management tools have been developed over the 

past decade. For example, ‘Efficient consumer response’ 

(ECR) has been developed to increase the consumer 

orientation and cost-effectiveness of supply chains (Kurt 

Salmon Associates, 1993). New management systems 

have been implemented to improve logistics, increase the 

use of information and communications technologies and 

boost quality management (Lambert and Cooper, 2000). 

Material supply 

Plan of Plant 

Material supply 

Plan of Plant 

Farmland 

Job 

… … … 

Farmland 

Job 

Raw 

Agricultural 

Products 

At present, in many developed countries such as 

America, Holland, Japan and European Union, the 

research of ASCM has gotten a very high level. For 

example, in America, the production of agriculture has 

been entrepreneurial; the developing mode of agricultural 

enterprises is “industry + industry + industry”. The 

process of agricultural enterprises’ operation has been 

basically finished. In Europe, people tend to have shorter 

food supply chain, i.e. sending products from the farms 

directly to the families so as to ensure the freshness and 

safety, and avoid uncertainty and information asymmetry. 

Holland summarizes the successful organization of 

agricultural chain to: research, service of information and 

education, high quality and stability of agricultural 

product supply, and global vision of market. 

In China, the research on ASCM begins in 1999 and is 

currently still in its infancy. Compared with developed 


New generation cooperatives are emerging, strengthening 

the position of farmers’ groups (Cook et al., 2001) and 

strategic partnering and vertical alliances are cementing 

sustainable partnerships throughout the supply chain 

(Zylbersztajn & Farina, 1999). In the side of models, 

there are a variety of supply chain models which address 

both the upstream and downstream sides. Such as: 

Supply-Chain Operations Reference-model (SCOR), 

American Productivity & Quality Center's (APQC) 

Process Classification Framework and the Supply Chain 

Best Practices Framework etc. All of these models 

address both the upstream and downstream sides, and 

need use IT to collect, analyze, transmit, manage and 

visualize their information and results. 

Shipping 

III. AGRICULTURAL SUPPLY CHAIN MANAGEMENT 

Products 

Processing 

Shipping 

Figure 2. Flow Frame of Agricultural Supply Chain 

We speak of a ‘supply chain’ when different actors are 

linked from ‘farm to fork’ to achieve a more effective and 

consumer-oriented flow of products. The ASCM is a 

branch of SCM and means the research and application of 

SCM in the field of agriculture; it is a kind of tracking 

management of “from farm to table”. ASCM involves the 

whole products’ flow throughout agricultural 

production’s phase from grow seedlings to customers [5] 

. 

In this whole process, the actors of ASCM may include 

growers, pickers, packers, processors, storage and 

transport facilitators, marketers, exporters, importers, 

distributors, wholesalers, and retailers (Fig.2). 

Shipping 

Dispatch Sale 

Shipping 

Customer 

countries, Chinese agricultural product supply chains 

consist of the millions of small scale farmers, who are not 

well structured and organized in the supply chain. The 

status quo of food supply chain in China can be 

summarized as following: A long and unsustainable 

supply chain, inadequate policy support, limited 

infrastructure for storage, inefficient information and 

knowledge flows, the lower level of internetization and 

intelligentization etc. With the development of 

Information Technology, specially the development of 

internet of things technology (such as RFID, Sensor etc.) 

provides new opportunity for research on ASCM, 

infomationization, internetization and intelligentization of 

ASCM would be the new researching content of 

agricultural modernization and trend of global agriculture 

development.


IV. ANALYSIS OF INTERNET OF THINGS 

In 1995, in the book of “the road ahead”, BILL 

GATES firstly brought up the concept about connection 

of “thing to thing”. In 1999, EPCglobal united more than 

100 enterprises and created the International 

Telecommunication Union and formally brought up the 

concept of Internet of Things (IoT). In 2005, ITU 

published “ITU Internet Reports 2005: Internet of Things 

(IoT)” introduced that we are standing on the brink of a 

new computing and communication era, one that will 

radically transform our corporate, community, and 

personal spheres. 

A. Concept of Internet of Things 

According to the definition of ITU, The IoT describes 

a worldwide network of billions or trillions of objects that 

can be collected from the worldwide physical 

environment, propagated via the Internet, and transmitted 

to end-users. Services are available for users to interact 

with these smart objects over the Internet, query their 

states, as well as their associated information, and even 

control their actions [6] . The purpose of IoT is to create a 

huge network through the combination of different 

smarter devices (such as RFID, GPS, RS) and networks 

to realize the information sharing of global things from 

any place, and any time [7] 

(Fig.3). 

Gather up 

any thing 

Process any 

information 

Connect 

any place 

IoT 

Serve any 

object 

Figure 3. ttributes Frame of IoT 

B. The Technology System of IOT 

In the field of IT, the IoT is a technology revolution 

that represents the future of computing and 

communications. It refers to a network of objects, and is 

often a self-configuring wireless network. With 

continuing developments in miniaturization and declining 

costs, IoT is becoming not only technologically possible 

but also economically feasible to make everyday objects 

smarter, and to connect the world of people with the 

world of thing. Embedded intelligence in things 

themselves will distribute processing power to the edges 

of the network, offering greater possibilities for data 

processing and increasing the resilience of the network. 

As a whole, in this technology system, the coral 

technologies mainly include: RFID, sensor, “3S”, WSN, 

and cloud computing etc. (Fig.4). 


Communicate 

any network 

Collect any 

data 

Smart Card 

Barcode 

RFID 

GPS 

Sensor 

… 

Webcam 

High-speed 

Access 

Network 

Cloud 

Computing 

SPSS 

Internet 

Center 

Figure 4. Technology System of IOT 

… 

DB 

GIS 

The radio frequency identification (RFID) is a 

technology that can be used to tag physical objects, 

allowing them to be detected and identified automatically. 

RFID has been perceived as a critical technology for 

improving efficiency and effectiveness in production and 

operations of manufacturing and service organizations 

and for improving SCM in various types of organizations 

RFID technology is classified as a wireless automatic 

identification and data capture technology which includes 

bar coding, optical recognition, biometrics, card etc. 

Basic RFID system consists of three components: antenna, 

RF tag and reader. The purpose of an RFID system is to 

enable data to be transmitted by a portable device, called 

a tag, which is read by an RFID reader and processed 

according to the needs of a particular application. Using 

RFID, the IoT can comprise millions of networked 

embedded devices also called smart items [8] 

. 

The sensor is another important technology for 

acquisition of data, and is mainly used to capture part of 

the existing data, correlate and synchronize these data, 

analyze them, and finally, carries out a reactive activity 

without user intervention. The components of a (remote) 

sensing node include the following: sensing and actuation 

unit (single element or array), processing unit, 

communication unit, power unit and other applicationdependent 

units Sensors can be simple point elements or 

can be multipoint detection arrays, and it has the 

capability of large scale deployment, low maintenance, 

[9] 

scalability, adaptability for different scenarios etc . 

The technology of “3S” includes RS (Remote Sense), 

GIS (Geography Information System) and GPS (Global 

Position System ) and is mainly used to provide the 

location of a particular tagged object, acquire the 

information of an object or phenomenon by multiple 

satellites, aircraft, etc, and finally store, analyze, manage, 

and present data that are linked to location(s) by GIS 

[10] 

software . 

For the transmission of information, wireless sensor 

network (WSN) and GPRS/GSM are the mainly used 

network technology in the field of agriculture. WSN 

mainly consists of spatially distributed autonomous 

sensors to cooperatively monitor physical or 

environmental conditions, such as temperature, sound, 

pressure, motion or pollutants. There are four basic


components in a WSN: an assembly of distributed or 

localized sensors; an interconnecting network (usually, 

but not always, wireless-based), a central point of 

information clustering; and a set of computing resources 

at the central point (or beyond) to handle data correlation, 

event trending, status querying, and data mining. 

For the process of information, cloud computing is 

Web-based processing and is location independent 

computing, whereby shared servers provide resources, 

software, and data to computers and other devices on 

demand, as with the electricity grid. Cloud computing 

describes a new supplement, consumption, and delivery 

model for IT services based on the Internet, and it 

typically involves over-the-Internet provision of 

dynamically scalable and often virtualized resources [11] 

. 

In addition, in the side of integration of information and 

service, service-oriented architecture (SOA) is a flexible 

set of design principles used during the phases of systems 

development and integration in computing. SOA defines 

how to integrate widely disparate applications for a Webbased 

environment and uses multiple implementation 

platforms. It generally provides a way for consumers of 

services, such as web-based applications, to be aware of 

available SOA-based services. 

V. ANALYSIS ABOUT INFORMATION PLATFORM OF ASCM 

In order to react effectively and quick to consumer’s 

demand, supply chain management is consumer-oriented. 

It aims at coordination of production processes. In ASCM, 

if all relevant information is accessible to any relevant 

company; every company in the supply chain has the 

possibility to and can seek to help optimizing the entire 

supply chain rather than sub optimize based on a local 

interest. This will lead to better planned overall 

production and distribution which can cut costs and give 

a more attractive final product leading to better sales and 

better overall results for the companies involved. So, the 

agricultural data are the vital basis of ASCM, every 

activity in this chain involves the creation, processing and 

communication of information. As an important subpart, 

the integrated information platform of ASCM is to realize 

integration and seamless access of multi-source 

information from any place at any time. The terminal 

Material 

Supply 

Seeds 

Technique 

Device 

Pesticides 

Fertilizer 


Farmland 

Production 

Water Soil 

Farmland 

Climate … 

Product 

Process 

Manufacture 

and 

Processing 

Enterprise 

target is to improve the level of infomationization, 

internetization and intelligentization of ASCM, and 

realize just in time delivery, supply base reduction, 

supplier integration, efficient information transmission, 

and collaborative relationships, help firms to trace 

products along food chains. 

A. Analysis of Agricultural Information 

Agricultural supply chain is a very complicated 

process, and involves many different phases and different 

actors. Every phase involves many kinds of operation, 

and every operation involves many kinds of factors, from 

environment to humane, from ecology to economic, from 

geography to society etc. First, agricultural production is 

closely related with spatial factors, every farmland has its 

own geographic location and boundaries. Second, any 

agricultural system has many factors; each factor also 

contains many sub-factors. For example, in the crops of 

biological factor, there are wheat, paddy rice, corn, cotton 

and other factors etc [12] . Third, agricultural data come 

from multiple sources, such as on-board sensors, soil 

sampling, remote sensing, satellite, web-cam and history 

material etc, the category of data includes text, number, 

image, sound and video etc. Finally, all of this 

information is always changing following the time and 

space. So, agriculture production is a very complicated 

ecology system and has the attributes of areadecentralization, 

object-diversify, data-mass creaturemutation 

and factors-uncertain. So, as a whole, ASCM is 

closely related with multi-source, dynamic and enormous 

information, and agricultural supply chain process has the 

attributes of dispersed collection points, long average 

collection period, low speed, enormous data, bad 

conditions of field etc [13] . In addition, during the process 

of agricultural supply, the flow of information is not 

single line, every phase has many relationship with other 

parts or industries, and is carried out by different 

operators (manufacturers, distributors, service suppliers, 

consumers) [5] 

Product 

Shipping 

Transportation 

Company 

Figure 5. Agri-food Flow of ASCM 

(Fig.5). All of these factors increase the 

degree of difficulty of agricultural data acquisition and 

transmission. 

Distribution 

Supermarket 

Shop 

Fair 

Restaurant 

… 

Customer


B. Designing of Information Platform 

The data are the core and basis of ASCM. The 

potential of using these data will reach its full extent 

when suitable collecting and transmitting technology and 

method are developed. So, the main function of 

information platform of ASCM is to improve the speed 

and precise of data collecting, ensure the reliable and 

seamless transmission, enhance the central process ability 

and advance intelligent service level of ASCM. 

According to the characteristic of agricultural information 

flow, from the side of technology, because of the 

attributes of overall sense, reliable transfer and intelligent 

process, IoT begins to become the main method of data 

acquisition and transmission and would become an 

important technology across the supply chain to collect, 

analyze, transmit and manage the whole data. IoT can 

comprise millions of networked embedded devices also 

called smart items; these devices are capable of collecting 

information about themselves, their environment, and 

Basic Information 

Of Pre-production 

Communication Interface 

Farmland Information Acquisition 

Sensors Weather 

Station 

RS 

GPS 

Object 

with 

RFID 

WSN 

1) Fast Acquisition of Data 

The collection of data still proves a demanding task 

and directly affects the efficiency and quality of ASCM. 

Because of the multi-source of data, we should select 

different collecting technologies to acquire the data of 

different phases (Fig.7). 

GPS 

Spatia 

l Data 

RS 

image 

Data 

Multi-source Data of Agriculture 

Video 


Crop 

Figure 7. Acquisition of Multi-source Data 


Environment 

Sensing 

Data 

associated devices and communicate this information to 

other devices and systems via the all-connecting internet. 

It can monitor vulnerable environments and prevent or 

limit natural disasters. In recent years, researchers have 

begun studying how to use IoT in agriculture. For 

example, many kinds of sensors have been produced for 

sensing agricultural objects, such as crops, animals, at the 

same time, by wireless network; the sensed data can be 

transmitted to Internet [12] 

. 

The information platform of ASCM based on IoT 

would be an integrated system that integrates all kinds of 

information which includes the agricultural production, 

purchase, warehousing, shipment, delivery and retail and 

realizes the information interchange between different 

phases. The core function is fast acquisition, seamless 

connection, reliable transmission and in-time search and 

trackback of information (Fig.6). 

Information 


Processing 

Restaurant 

Supermarket 

Fair 

Customer 

Farmer… 

Object 

with 

RFID 

Figure 6. Information Flow of ASCM Based on IoT 

RFID 

Tag 

Record 

History 

Data 

and 

Graph 

Transportation 

Information 

Information Center 

of Agriculture 

Internet 

Object 

with 

RFID 

For example, during the stage of farmland production, 

for collecting the data of environmental factor, such as 

factors of soil (such as PH, CEC, humidity, electrical 

conductivity etc.), factors of climate (such as temperature, 

illumination, wind speed etc.), we can deploy different 

kinds of sensors (such as Temperature Sensor, Moisture 

Sensor) according to the farmland’ attribute of geography 

and crop’s attribute of growth. During the deploying of 

sensors, we should think of that the in-time dynamic 

changing of farmland environment would interfere with 

the transmission of radio signal and diminish this 

affection as far as possible; for acquiring the environment 

data of crop’s growth, we can deploy web-cam for 

capturing the image or video of crop’s growth, and 

deploy weather station for getting the environmental 

climate data of farmland; for acquiring the geographic


information of farmland, we can use GPS to measure 

geographic location such as, provide data on location in 

terms of latitude, longitude and altitude of farmland. In 

addition, airborne data collection systems through RS 

technologies, such as aerial photographs and satellite 

remote sensing provide periodic land use, land cover and 

other thematic information [10] 

. 

During the stage of processing and transportation of 

products, for tracking products status and identify every 

product, we can use RFID technology. RFID is a generic 

term that is used to describe a system that transmits the 

identity (in the form of a unique serial number) of an 

object or person wirelessly, using radio waves. RFID will 

start linking up the supply chain from the farm gate to the 

restaurant plate – and every point in between. The RFID 

tag is dimensioned to approximate a size of an individual 

product and multiple such tags are deposited in a 

container of the product at the harvesting stage. At each 

product handling stage the tags are programmed with the 

time and location of the event as well as any other 

attributes relevant to the handling process. Therefore, the 

entire history of crop handling is stored in the tags and 

can be detected and identified automatically at any time. 

2) Reliable Transmission of Data 

In the ASCM, the reliable flow of information is very 

important for track, manage, collaborate and plan the 

production of agricultural products. In the part of data 

process and transmission, the main function is to ensure 

that the information from different intelligent devices can 

be reliably transmitted to different users through network 

infrastructure, such as mobile communications network 

(such as GSM, GPRS, TD-SCDMA), wireless sensor 

network (such as ZigBee), and satellite communication 

network etc. 

Multi-source 

Data 

Of 

Agriculture 

Spatial Data 

Process System 

WSN/ GPRS/GSM 

Non-spatial Data 

Process System 

Internet 

Figure 8. Transmission of Multi-source Data 

In the technological system of data transmission, WSN 

is one of the most suitable technologies for capturing real 

world data. WSN is an infrastructure comprised of 

sensing (measuring), computing, and communication 

elements that gives an agricultural administrator the 

ability to instrument observe, and react to events and 

phenomena in a specified environment. In addition, 

connecting WSN to the Internet can standardize 

contextual data and make them can be shared with other 

entities, and can analyze these data, take decisions in 


Internet 

Information 

Pool 

remote premises, and finally implement these decisions 

back in the real world through sensors [14] 

. 

3) Constructing the Information Center of 

Agriculture 

The purpose of modern ASCM is to improve the level 

of agricultural information process and enhance the 

intelligent management and decision of agricultural 

production. For improving the using efficiency of 

information, we should use different technology and 

method to process, analysis, merge, classify and statistic 

multi-data and create public database and special 

database, spatial database and relationship database etc. 

for different users. For example, we can use GIS to store, 

manage and analyze geographical reference data, and use 

relationship database to analysis non-spatial data, and use 

data mining to analyze data from different perspectives 

and summarizing it into useful information, and use SPSS 

to realize different statistics of data. 

Finally, for further improving the efficiency of on-line 

analysis and process, the technology of cloud computing 

has been used. Using the intelligent cloud computer 

platform can ensure that enormous information of internet 

are real-time analyzed, processed, managed and 

controlled and create an efficient and reliable decision 

service system for the high level management and largescale 

industry application. In addition, following the 

application of service-oriented architecture (SOA) in 

different industrial infomationization and internetization, 

many experts begin to research how to realize the 

application of SOA in agricultural field. Specially, for 

ASCM, it is consumer-oriented and aims at coordination 

of different production sub-processes, a system based on 

a SOA will package functionality as a suite of 

interoperable services that can be used within multiple 

separate systems from several business domains. So using 

SOA can not only improve the level of infomationization, 

internetization and intelligentization of ASCM, but also 

enhance the level of service of ASCM. 

SUMMARY 

Technological innovation and competition have led to 

improvements in supply chain management for 

agricultural products. Yet, global market standards are 

stringent. Consumers demand safe and nutritional food, 

excellent quality and just-in-time delivery. So, 

collaboration between trade partners has become 

increasingly important for the success of cross-border 

trade in the competitive market. Agricultural supply chain 

management is a powerful tool to achieve this 

collaboration. Through supply chains, producers in 

developing countries and emerging economies can access 

market information and knowledge to hone their valueadded 

activities. However, developing cross-border 

supply chains is very complex, and requires a lot of 

information and expertise about how to build chains, as 

well as communication and commitment from all the 

chain partners. Specially, for the ASCM, the research of 

integrated information platform is the key of improving 

level of ASCM. The developing of information 

acquisition and process technology mainly depends on


the development of modern Information Technology, 

such as computer technology, electronics, satellite 

navigation technology, RS, sensor technology and 

network technology etc. In the developed countries, 

agriculture automation has been applied widely, and the 

degree of agriculture informationization and networking 

has attained a very high level. But in China, as a 

developing country, the research of agriculture 

informationization still belongs to the beginning level. 

However, following the occur of IoT, it will advance the 

developing of digital agriculture, and the research of 

intelligent agriculture based on IoT will create a huge 

intelligent agricultural production and supply network 

and connect the entire farm, farm villages and trade 

market of a big city such as Beijing together. It can not 

only improve the quality of agricultural product, the 

efficiency of agricultural production, the level of 

agricultural product supply, but also efficiently solve the 

emergent dispatch of food in special situation. 


This work is elaborated within the project of “Research 

on Growth Evaluation and Physiological-Ecological 

Simulation of Peach”. No. KM201010020011 funded by 

the Beijing Municipal Commission of Education. 

REFERENCES 

[1] Xuewen TAN. The Sustainability of Food Supply Chain in 

China: Review and Comment. Published by Rural 

Development Institute of Chinese Academy of Social 

Sciences, 2007.9. 

[2] Athanasios S. Voulodimos, Charalampos Z. Patrikakis, 

Alexander B. Sideridis, Vasileios A. Ntafis, Eftychia M. 

Xylouri. A complete farm management system based on 


animal identification using RFID technology. Computers 

and Electronics in Agriculture 70 (2010) 380–388. 

[3] http://en.wikipedia.org/wiki/Supply_chain_management. 

2011.1 

[4] Chin S. Ou, Fang C. Liu, Yu C. Hung, David C. Yen. A 

structural model of supply chain management on firm 

performance. International Journal of Operations & 

Production Management. Vol. 30 No. 5, 2010 pp. 526-545. 

[5] Maurizio Mariani. Sustainable agri-food supply chains and 

systems, Forum of China-Europe, 2007.06. 

[6] Lei CHEN, Mitchell TSENG, Xiang LIAN. Development 

of foundation models for Internet of Things. Front. Comput. 

Sci. China 2010, 4(3): 376–385. 

[7] Su-bin SHEN, Qu-li FAN, Ping ZONG. Study on the 

Architecture and Associated Technologies for Internet of 

Things. Journal of Nanjing University of Posts and 

Telecommunications (Natural Science).2009, 29(6):1-11. 

[8] ITU. ITU Internet Reports 2005: The Internet of Things, 

ITU (2005). 

[9] Juan Ignacio Vazquez, Aitor Almeida, Iker Doamo. Flexeo: 

Architecture for Integrating Wireless Sensor Networks into 

the Internet of Things. UCAMI 2008, ASC 51, p219-218 

Technical Review 2009, 26(5). 

[10] http://en.wikipedia.org/wiki/Remote_sensing. 2011.1 

[11] http://en.wikipedia.org/wiki/Cloud_computing. 2011.1 

[12] Yaqin WANG, Hua GAO, Huaxue TAO, Research on the 

Framework System and Technical Platform of Digital 

Agriculture . 5th FIG Regional Conference Accra, Ghana, 

2006:1-13. 

[13] Christos Goumopoulos* and Achilles D. Kameas, Alan 

Cassells. An Ontology-driven System Architecture for 

Precision Agriculture Applications. Int. J. Metadata, 

Semantics and Ontologies, Vol. 4, Nos. 1/2, 2009:72-84. 

[14] KAZEM SOHRABY, DANIEL MINOLI, TAIEB ZNATI. 

WIRELESS SENSOR NETWORKS--Technology, 

Protocols and Applications. A JOIN WILEY &SONS, 

INC.2007.

Aims and Scope. 

Call for Papers and Special Issues 

Journal of Software (JSW, ISSN 1796-217X) is a scholarly peer-reviewed international scientific journal focusing on theories, methods, and 

applications in software. It provide a high profile, leading edge forum for academic researchers, industrial professionals, engineers, consultants, 

managers, educators and policy makers working in the field to contribute and disseminate innovative new work on software. 

We are interested in well-defined theoretical results and empirical studies that have potential impact on the construction, analysis, or management 

of software. The scope of this Journal ranges from the mechanisms through the development of principles to the application of those principles to 

specific environments. JSW invites original, previously unpublished, research, survey and tutorial papers, plus case studies and short research notes, 

on both applied and theoretical aspects of software. Topics of interest include, but are not restricted to: 

• Software Requirements Engineering, Architectures and Design, Development and Maintenance, Project Management, 

• Software Testing, Diagnosis, and Validation, Software Analysis, Assessment, and Evaluation, Theory and Formal Methods 

• Design and Analysis of Algorithms, Human-Computer Interaction, Software Processes and Workflows 

• Reverse Engineering and Software Maintenance, Aspect-Orientation and Feature Interaction, Object-Oriented Technology 

• Component-Based Software Engineering, Computer-Supported Cooperative Work, Agent-Based Software Systems, Middleware Techniques 

• AI and Knowledge Based Software Engineering, Empirical Software Engineering and Metrics 

• Software Security, Safety and Reliability, Distribution and Parallelism, Databases 

• Software Economics, Policy and Ethics, Tools and Development Environments, Programming Languages and Software Engineering 

• Mobile and Ubiquitous Computing, Embedded and Real-time Software, Database, Data Mining, and Data Warehousing 

• Internet and Information Systems Development, Web-Based Tools, Systems, and Environments, State-Of-The-Art Survey 

Special Issue Guidelines 

Special issues feature specifically aimed and targeted topics of interest contributed by authors responding to a particular Call for Papers or by 

invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are of interest to the Journal. 

Preference will be given to proposals that cover some unique aspect of the technology and ones that include subjects that are timely and useful to the 

readers of the Journal. A Special Issue is typically made of 10 to 15 papers, with each paper 8 to 12 pages of length. 

The following information should be included as part of the proposal: 

• Proposed title for the Special Issue 

• Description of the topic area to be focused upon and justification 

• Review process for the selection and rejection of papers. 

• Name, contact, position, affiliation, and biography of the Guest Editor(s) 

• List of potential reviewers 

• Potential authors to the issue 

• Tentative time-table for the call for papers and reviews 

If a proposal is accepted, the guest editor will be responsible for: 

• Preparing the “Call for Papers” to be included on the Journal’s Web site. 

• Distribution of the Call for Papers broadly to various mailing lists and sites. 

• Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be 

informed the Instructions for Authors. 

• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact 

information. 

• Writing a one- or two-page introductory editorial to be published in the Special Issue. 

Special Issue for a Conference/Workshop 

A special issue for a Conference/Workshop is usually released in association with the committee members of the Conference/Workshop like 

general chairs and/or program chairs who are appointed as the Guest Editors of the Special Issue. Special Issue for a Conference/Workshop is 

typically made of 10 to 15 papers, with each paper 8 to 12 pages of length. 

Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop: 

• Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers of XYZ Conference”. 

• Sending us a formal “Letter of Intent” for the Special Issue. 

• Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees. 

Information about the Journal and Academy Publisher can be included in the Call for Papers. 

• Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus 

the evaluation from the Session Chairs and the feedback from the Conference attendees. 

• Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. 

Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced. 

• Providing us the completed and approved final versions of the papers formatted in the Journal’s style, together with all authors’ contact 

information. 

• Writing a one- or two-page introductory editorial to be published in the Special Issue. 

More information is available on the web site at http://www.academypublisher.com/jsw/.

(Contents Continued from Back Cover) 


Yongqiang Chen, Yanqing Zhang, Hanping Hu, and Hefei Ling 

An Efficient Method for Improving Query Efficiency in Data Warehouse 

Zhiwei Ni, Junfeng Guo, Li Wang, and Yazhuo Gao 

Co-simulation Study of Vehicle ESP System Based on ADAMS and MATLAB 

Shengqin Li and Le He 

An Improved Fuzzy C-means Clustering Algorithm based on PSO 

Qiang Niu and Xinjian Huang 

Classification of Bio-potential Surface Electrode based on FKCM and SVM 

Hao Liu, Xiaoming Tao, Pengjun Xu, and Guanxiong Qiu 

Consonant Recognition of Dysarthria Based on Wavelet Transform and Fuzzy Support Vector 

Machines 

Zhuo-ming Chen, Wei-xin Ling, Jian-hui Zhao, and Tao-tao Yao 

ELECTRE I Decision Model of Reliability Design Scheme for Computer Numerical Control Machine 

Jihong Pang, Genbao Zhang, and Guohua Chen 

Fractional Modeling Method Research on Education Evaluation 

Chunna Zhao, Yu Zhao, Liming Luo, and Yingshun Li 

Immune Genetic Evolutionary Algorithm of Wavelet Neural Network to Predict the Performance in 

the Centrifugal Compressor and Research 


Development of Optimization Design Software for Bevel Gear Based on Integer Serial Number 


Xiaoqin Zhang, Yu Rong, Jingjing Yu, Liling Zhang, and Lina Cui 

Study on Operating Mechanisms and Dynamics Behavior of Agile Supply Chain 

Guohua Chen, Genbao Zhang, and Jihong Pang 

Unified Service Platform for Accessing Grid Resources 

Shaochong Feng, Yuanchang Zhu, and Yanqiang Di 

Research on an Improved Terrain Aided Positioning Model 

Shidan Li, Liguo Sun, Xin Li, and Desheng Wang 

Research on Integrated Information Platform of Agricultural Supply Chain Management Based on 


Yan-e Duan 

849 

857 

866 

873 

880 

887 

894 

901 

908 

915 

923 

930 

937 

944

Journal of Software - Academy Publisher

Create successful ePaper yourself

Delete template?

Save as template?