Journal of Networks - Academy Publisher
Journal of Networks - Academy Publisher
Journal of Networks - Academy Publisher
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
<strong>Journal</strong> <strong>of</strong> <strong>Networks</strong><br />
ISSN 1796-2056<br />
Volume 6, Number 7, July 2011<br />
Contents<br />
Special Issue: Selected Best Papers <strong>of</strong> the International Workshop on Computer Science for<br />
Environmental Engineering and EcoInformatics (CSEEE 2011)<br />
Guest Editors: Tianlong Gu and Shenghui Liu<br />
Guest Editorial<br />
Tianlong Gu and Shenghui Liu<br />
SPECIAL ISSUE PAPERS<br />
Multi-Constrained Routing Algorithm for Multimedia Communications in Wireless Sensor <strong>Networks</strong><br />
Xin Yan, Layuan Li, and F. J. An<br />
Reputation-aware Service Selection based on QoS Similarity<br />
Shenghui Zhao, Guoxin Wu, Guilin Chen, and Haibao Chen<br />
Cost Aggregation Strategy with Bilateral Filter Based on Multi-scale Nonlinear Structure Tensor<br />
Li Li and Hua Yan<br />
A Collaborative Nonlocal-Means Super-resolution Algorithm Using Zernike Monments<br />
Lin Guo and Qinghu Chen<br />
Mathematical Model and Hybrid Scatter Search for Cost Driven Job-shop Scheduling Problem<br />
Jie Bai, Kai Sun, and Gen Ke Yang<br />
Multi-objective Genetic Algorithm for System Identification and Controller Optimization <strong>of</strong><br />
Automated Guided Vehicle<br />
Xing Wu, Peihuang Lou, and Dunbing Tang<br />
WebVR—Web Virtual Reality Engine Based on P2P Network<br />
Zhihan Lv, Tengfei Yin, Yong Han, Yong Chen, and Ge Chen<br />
An Energy-Efficient Communication Protocol for Wireless Sensor <strong>Networks</strong><br />
Fengjun Shang<br />
Robust Cross-layer Design <strong>of</strong> Wireless Multimedia Sensor <strong>Networks</strong> with Correlation and<br />
Uncertainty<br />
Lei You and Chungui Liu<br />
The E-Commerce Model <strong>of</strong> Health Websites: An Integration <strong>of</strong> Web Quality, Perceived Interactivity,<br />
and Web Outcomes<br />
Chung-Hung Tsai<br />
937<br />
939<br />
950<br />
958<br />
966<br />
974<br />
982<br />
990<br />
999<br />
1009<br />
1017
A New Method <strong>of</strong> Time-frequency Synthesis <strong>of</strong> Harmonic Signal Extraction from Chaotic<br />
Background<br />
Erfu Wang, Zhifang Wang, Jing Ma, and Qun Ding<br />
Provable Data Possession <strong>of</strong> Resource-constrained Mobile Devices in Cloud Computing<br />
Jian Yang, Haihang Wang, Jian Wang, Chengxiang Tan, and Dingguo Yu<br />
Image Compression Based on Improved FFT Algorithm<br />
Juanli Hu, Jiabin Deng, and Juebo Wu<br />
Correlative Peak Interval Prediction and Analysis <strong>of</strong> Chaotic Sequences<br />
Qun Ding, Lu Wang, and Guanrong Chen<br />
REGULAR PAPERS<br />
An Energy Efficient Dynamic Clustering Protocol Based on Weight in Wireless Sensor <strong>Networks</strong><br />
Ming Zhang and Suoping Wang<br />
Performance <strong>of</strong> UWB Systems with Direct-Sequence Bipolar Pulse Amplitude Modulation and<br />
RAKE Reception over IEEE 802.15.3a Channel<br />
Jingjing Wang and Hao Zhang<br />
Data Accuracy Estimation for Spatially Correlated Data in Wireless Sensor <strong>Networks</strong> under<br />
Distributed Clustering<br />
Jyotirmoy Karjee and H.S Jamadagni<br />
Networking as a Service: a Cloud-based Network Architecture<br />
Tao Feng, Jun Bi, Hongyu Hu, and Hui Cao<br />
1025<br />
1033<br />
1041<br />
1049<br />
1057<br />
1065<br />
1072<br />
1084
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 937<br />
Special Issue on Selected Best Papers <strong>of</strong> the International Workshop on Computer Science for<br />
Environmental Engineering and EcoInformatics (CSEEE 2011)<br />
Guest Editorial<br />
This special issue comprises <strong>of</strong> 14 selected papers from the International Workshop on Computer Science for<br />
Environmental Engineering and EcoInformatics (CSEEE 2011). The conferences received 860 paper submissions from<br />
15 countries and regions, <strong>of</strong> which 450 were selected for presentation after a rigorous review process. From these 450<br />
research papers, through two rounds <strong>of</strong> reviewing, the guest editors selected 14 as the best papers on the Networking<br />
Technologies and Information track <strong>of</strong> the Conference. The candidates <strong>of</strong> the Special Issue are all the authors, whose<br />
papers have been accepted and presented at the CSEEE 2011, with the contents not been published elsewhere before.<br />
2011 International Workshop on Computer Science for Environmental Engineering and EcoInformatics will continue<br />
the excellent tradition <strong>of</strong> gathering world-class scientists, engineers and educators engaged in the fields <strong>of</strong> Computer<br />
Science and Environmental Biotechnology to meet and present their latest activities. CSEEE 2011 held on July 29-31,<br />
2011, Kunming, China. This conference is sponsored by International Association for Scientific and High Technology,<br />
and is in cooperation with Yunnan University, and it is technical co-sponsored by Kunming University <strong>of</strong> Science and<br />
Technology.<br />
“Multi-Constrained Routing Algorithm for Multimedia Communications in Wireless Sensor <strong>Networks</strong>”, by Xin Yan,<br />
Layuan Li and F. J. An, proposes a novel routing model that can depict multiple service requirements in multimedia<br />
sensor networks, and designs a new multi-constrained routing algorithm MCRA for multimedia communications.<br />
Theoretical analysis and simulation experiments are provided to validate their claims.<br />
“Reputation-aware Service Selection Based on QoS Similarity”, by Shenghui Zhao, Guoxin Wu, Guilin Chen and<br />
Haibao Chen, proposes a reputation evaluation method for Web Services which can gradually adjusting the reputations<br />
based on eliminating the collusive behaviors <strong>of</strong> consumers step by step. The experimental results show that the model<br />
can identify the conclusive consumers and improve the exact rate <strong>of</strong> reputation evaluation and success rate <strong>of</strong> service<br />
selection.<br />
“Cost Aggregation Strategy with Bilateral Filter Based on Multi-scale Nonlinear Structure Tensor”, by Li Li and Hua<br />
Yan, proposes a novel cost aggregation method for stereo matching with modified bilateral filter. By constructing the<br />
multi-scale nonlinear structure tensor and adding the new corresponding weight in cost aggregation, more pixels similar<br />
with central pixel are aggregated in a support window and the final disparity map are more accurate..<br />
“A Collaborative Nonlocal-Means Super-resolution Algorithm Using Zernike Monments”, by Lin Guo and Qinghu<br />
Chen, proposes an efficient improved algorithm by introducing Zernike moments as representation <strong>of</strong> image invariant<br />
features into similarity measure. Experimental results indicate the proposed method can handle real video sequences<br />
with general motion pattern, and performances better than the comparing methods.<br />
“Mathematical Model and Hybrid Scatter Search for Cost Driven Job-shop Scheduling Problem”, by Jie Bai, Kai Sun<br />
and Gen Ke Yang, proposes a cost driven model <strong>of</strong> the job-shop scheduling problem in which the solutions are driven<br />
by business inputs, such as the cost <strong>of</strong> the product transitions, revenue loss due to the machine idle time and<br />
earliness/tardiness penalty.<br />
“Multi-objective Genetic Algorithm for System Identification and Controller Optimization <strong>of</strong> Automated Guided<br />
Vehicle”, by Xing Wu, Peihuang Luo and Dunbing Tang, proposes a multi-objective genetic algorithm (MOGA) with<br />
Pareto optimality and elitist tactics for the control system design <strong>of</strong> automated guided vehicle (AGV).<br />
“WebVR—Web Virtual Reality Engine Based on P2P network”, by Zhihan Lv, Tengfei Yin, Yong Han, Yong Chen<br />
and Ge Chen, introduces a multi-user online virtual reality engine -- WebVR. The core model innovation <strong>of</strong> WebVR<br />
engine is mapping the geographical space and virtual space to the P2P overlay network space, and build quad-tree index<br />
for the three spaces, and they identify the geocoding based on Hash value, which is used to index the user list, terrain<br />
data, and the model object data. The model greatly improves the hit rate <strong>of</strong> 3D geographic data search under P2P<br />
overlay network.<br />
“An Energy-Efficient Communication Protocol for Wireless Sensor <strong>Networks</strong>”, by Fengjun Shang, proposes an<br />
energy-efficient Single-Hop Active Clustering (SHAC) algorithm for wireless sensor networks. Through both<br />
theoretical analysis and numerical results, it is shown that SHAC prolongs the network lifetime significantly against the<br />
other clustering protocols such as LEACH-C and EECS.<br />
“Robust Cross-layer Design <strong>of</strong> Wireless Multimedia Sensor <strong>Networks</strong> with Correlation and Uncertainty”, by Lei You<br />
and Chungui Liu, uses a cross-layer method to deal with the robust lifetime optimization <strong>of</strong> wireless multimedia sensor<br />
network (WMSN) with energy consumption uncertainty and proposes to model the uncertainty as a polyhedral set.<br />
“The E-Commerce Model <strong>of</strong> Health Websites: An Integration <strong>of</strong> Web Quality, Perceived Interactivity, and Web<br />
Outcomes”, by Chung-Hung Tsai, integrates web quality (system quality, information quality, and service quality),<br />
perceived interactivity (human-message, human-human), and web outcomes (web usage, web satisfaction, and web<br />
loyalty) to explore the e-commerce model <strong>of</strong> health websites. A survey <strong>of</strong> 1076 users <strong>of</strong> health websites was conducted<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.937-938
938 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
to validate the proposed model. The findings show that web quality has significantly positive effect on perceived<br />
interactivity, web usage, and web satisfaction separately, which in turn influence web loyalty.<br />
“A New Method <strong>of</strong> Time-frequency Synthesis <strong>of</strong> Harmonic Signal Extraction from Chaotic Background”, by Erfu<br />
Wang, Zhifang Wang, Jing Ma and Qun Ding, proposes a new synthesis about wavelet threshold and empirical mode<br />
decomposition (EMD) complementary <strong>of</strong> new harmonic signal extraction by experimental simulation.<br />
“Provable Data Possession <strong>of</strong> Resource-constrained Mobile Devices in Cloud Computing”, by Jian Yang, Haihang<br />
Wang, Jian Wang, Chengxiang Tan and Dingguo Yu, proposes a novel PDP scheme, in which a trusted third-party<br />
agent (TPA) takes over most <strong>of</strong> the calculations from the mobile end-users. By using bilinear signature and Merkle hash<br />
tree (MHT), the scheme reduces communication and storage burden, and is fit for mobile devices.<br />
“Image Compression based on improved FFT Algorithm”, by Juanli Hu, Jiabin Deng and Juebo Wu, adopts Radix-4<br />
Fast Fourier transform (Radix-4 FFT) to realize the limit distortion for image coding, and to discuss the feasibility and<br />
the advantage <strong>of</strong> Fourier transform for image compression. It aims to deal with the existing complex and<br />
time-consuming <strong>of</strong> Fourier transform, according to the symmetric conjugate <strong>of</strong> the image by Fourier transform to<br />
reduce data storage and computing complexity.<br />
“Correlative peak interval prediction and analysis <strong>of</strong> chaotic sequences”, by Qun Ding, Lu Wang and Guanrong Chen,<br />
proposes a digital circuit design for the logistic-map module used in chaotic stream ciphers, analyzes the factors that<br />
may affect the output <strong>of</strong> the sequences, and develops a calculation method for estimating the output sequential<br />
correlative peak interval.<br />
We wish to thank the Kunming University <strong>of</strong> Science and Technology for providing the venue to host the conference.<br />
We would like to take this opportunity to thank the authors for the efforts they put in the preparation <strong>of</strong> the manuscripts<br />
and for their valuable contributions. We wish to express our deepest gratitude to the program committee members for<br />
their help in selecting papers for this issue and especially the referees <strong>of</strong> the extended versions <strong>of</strong> the selected papers for<br />
their thorough reviews under a tight time schedule. Last, but not least, our thanks go to the Editorial Board <strong>of</strong> the<br />
<strong>Journal</strong> <strong>of</strong> <strong>Networks</strong> for the exceptional effort they did throughout this process.<br />
In closing, we sincerely hope that you will enjoy reading this special issue.<br />
Guest Editors:<br />
Tianlong Gu, Guilin University <strong>of</strong> Electronic Technology, P.R. China<br />
Shenghui Liu, Harbin University <strong>of</strong> Science & Technology, P.R. China<br />
© 2011 ACADEMY PUBLISHER<br />
Tianlong Gu was born in Shanxi, China, on 1st October 1964. He received the Bachelor Degree from Taiyuan<br />
University <strong>of</strong> Technology in 1984, received Master Degree from Xidian University in 1986 and received his<br />
Ph.D. degree from Zhejiang University in 1996. From 1998 to 2002, he was a postdoctoral research fellow and<br />
visiting pr<strong>of</strong>essor within Murdoch University and Curtin University <strong>of</strong> Technology, Australia. He has published<br />
more than 130 papers, and authored 6 books. His main research interests include formal method, knowledge<br />
engineering and mobile computing.He is a full pr<strong>of</strong>essor in school <strong>of</strong> computer science & engineering at Guilin<br />
University <strong>of</strong> Electronic Technology, Guilin, China, and Ph.D. supervisor in school <strong>of</strong> computer science &<br />
technology at Xidian University, Xian, China.<br />
Shenghui Liu was born in Heilongjiang, China, on July 24, 1961. Bachelor <strong>of</strong> Automatic Control(6/1982).<br />
Master <strong>of</strong> Computer Science (3/1985). Doctor <strong>of</strong> Management Engineering(6/2009). Harbin University <strong>of</strong><br />
Science & Technology, Harbin, Heilongjiang, P.R.China. Now he is a pr<strong>of</strong>essor and the dean <strong>of</strong> S<strong>of</strong>tware School<br />
in Harbin University <strong>of</strong> Science & Technology.He has wide research interests, mainly information technology.<br />
He has published above 50 papers in journals or conference proceedings and some <strong>of</strong> the papers are indexed by<br />
SCI, EI. He has won various awards in the past. He served as many workshop chair, advisory committee or<br />
program committee member <strong>of</strong> various international conferences.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 939<br />
Multi-Constrained Routing Algorithm for<br />
Multimedia Communications in Wireless Sensor<br />
<strong>Networks</strong><br />
Xin Yan and Layuan Li<br />
Department <strong>of</strong> Computer Science, Wuhan University <strong>of</strong> Technology, Wuhan 430063, P.R. China<br />
Email: yanxin@whut.edu.cn, jwtu@public.wh.hb.cn<br />
F. J. An<br />
Faculty <strong>of</strong> EEMCS, Delft University <strong>of</strong> Technology, 2600 GA Delft, The Netherlands<br />
Email: anfengju@hotmail.com<br />
Abstract — The existing routing protocols designed for<br />
real-time or multimedia applications in sensor networks<br />
usually adopt relatively simple routing models where fewer<br />
service metrics are considered, which is not sufficient for<br />
real-time or multimedia data transportations. Furthermore,<br />
for the sake <strong>of</strong> route discovery or the acquisition <strong>of</strong> a target<br />
location, they usually need extra localization equipments or<br />
beacon exchanges to obtain the geographic location <strong>of</strong> each<br />
sensor node or construct a coordinate system for sensor<br />
nodes, which imports extra costs to routing algorithms. In<br />
this paper, firstly we propose a novel system model that can<br />
comprehensively depict the service requirements <strong>of</strong><br />
multimedia applications, and on the basis <strong>of</strong> this system<br />
model, we design a new multi-constrained routing algorithm,<br />
MCRA, for multimedia communications in sensor networks.<br />
MCRA not only can provide end-to-end delay guarantee<br />
and packet loss ratio guarantee for multimedia<br />
communications, but also can balance and improve the<br />
energy consumption in sensor nodes. Besides, MCRA adopts<br />
several effective policies to suppress message flooding and<br />
lessen data redundancy. In MCRA, neither the acquisition<br />
<strong>of</strong> target location nor the route discovery process requires<br />
any extra measurement equipment or coordinate system<br />
based on location message exchange, however, the target<br />
location we concern can be easily figured out by a<br />
localization scheme without message exchange. In addition,<br />
we may optionally apply MAC multicast and differentiation<br />
service in MCRA so as to further lower its control message<br />
overhead and differentiate forwarding priority levels for<br />
real-time data and best-effort traffic in MAC layer.<br />
Theoretical analysis and simulation experiments are<br />
provided to validate our claims.<br />
Index Terms — routing algorithm, multimedia applications,<br />
sensor networks, QoS, message suppression, localization<br />
Ⅰ. INTRODUCTION<br />
A wireless sensor network (WSN) is comprised <strong>of</strong> small,<br />
low powered, self-organizing sensor nodes, densely<br />
deployed in the area to be monitored. These networks can<br />
support a wide rang <strong>of</strong> applications, such as earthquake<br />
response, health monitoring, battlefield surveillance etc.<br />
Some <strong>of</strong> these applications may be augmented due to the<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.939-949<br />
use <strong>of</strong> real-time or multimedia data. Real-time or<br />
multimedia applications have stringent requirements <strong>of</strong><br />
quality <strong>of</strong> service (QoS), for instance, end-to-end delay,<br />
packet loss ratio etc, during the data transmissions.<br />
Several main MAC layer protocols have been<br />
developed for multimedia communications in sensor<br />
networks. IEEE 802.11e scheme has provisions for<br />
service differentiation at MAC layer in sensor networks,<br />
though it was proposed for ad hoc networks initially [1].<br />
In this scheme, the service differentiation is obtained by<br />
changing the duration <strong>of</strong> the Inter-Frame Spacing (IFS)<br />
and the Contention Window (CW) size based on the<br />
priority <strong>of</strong> the packet. The scheme can provide QoS<br />
services for multimedia communications from two<br />
aspects: timeliness and reliability, thanks to its broadcast<br />
and multicast functions [2]. In addition, IEEE<br />
802.15.4/ZigBee specification designed for the low data<br />
rate, low power consumption, and low cost networks can<br />
provide a Guaranteed Time Slot (GTS) mechanism to<br />
allocate a specific duration within a super frame structure<br />
for real-time transmissions [3,4].<br />
Although each layer in sensor network stack may<br />
provide QoS services for multimedia communications,<br />
routing protocol in network layer is always playing the<br />
most important role. Routing protocol can provide not<br />
only QoS guarantees but also network load-balance and<br />
congestion management for multimedia data streams.<br />
As we know, majority <strong>of</strong> routing protocols in sensor<br />
networks are oriented to various applications. For<br />
multimedia applications, routing protocols should aim at<br />
providing timeliness and reliability services for them, and<br />
manage to balance the energy consumption in sensor<br />
nodes [5]. Additionally, routing protocols designed for<br />
multimedia communications are supposed to have better<br />
capacities <strong>of</strong> message suppression and data aggregation<br />
than common applications, because <strong>of</strong> more data<br />
redundancies in multimedia applications [6]. Therefore,<br />
routing algorithms working in multimedia sensor<br />
networks should be able to not only provide QoS services<br />
for multimedia communications but also suffice for other<br />
communication requirements in WSN (e.g., optimal
940 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
energy consumption, localization, message suppression<br />
and data aggregation etc). Meanwhile, they also ought to<br />
be resilient and adapt to the dynamics and scalability <strong>of</strong><br />
multimedia sensor networks.<br />
One the other hand, for most multimedia applications<br />
in sensor networks, location is more important than a<br />
specific node ID. Target data without position<br />
information retrieved from sensor nodes is usually<br />
unmeaning. It requires sensor nodes in WSN are<br />
location-aware, which needs to resort to GPS etc<br />
measurement equipments or some localization algorithms<br />
to locate these nodes. However, GPS etc measurement<br />
equipments are impractical even do not work completely<br />
in some especial environments [7]. To overcome this<br />
weakness, some virtual or logical coordinate systems are<br />
proposed [8,9,10], which construct a coordinate space for<br />
route discovery and target localization by node-location<br />
message exchange. Whereas, the construction <strong>of</strong> virtual<br />
or logical coordinate system needs to consume drastically<br />
the limited resources in sensor nodes. Worse for them,<br />
both geographic information based routing algorithms<br />
and virtual (or logical) coordinate system based routing<br />
algorithms, the inaccurate node-location information<br />
sometimes emerging could result in the failure <strong>of</strong> route<br />
discovery process. Hence, we should evade the risk <strong>of</strong><br />
routing failure caused by the imprecise location<br />
information as possible as we can, when designing a<br />
routing algorithm for multimedia sensor networks.<br />
Ⅱ. RELATED WORK<br />
For the sake <strong>of</strong> QoS provision and adaptation to the<br />
communication characteristics (energy constrained,<br />
limited computation capacity, and less memory<br />
availability) in sensor networks, T. He et al proposed a<br />
routing protocol named SPEED to provide s<strong>of</strong>t real-time<br />
guarantees for real-time communications in sensor<br />
networks [11]. In SPEED, end-to-end s<strong>of</strong>t real-time<br />
communication is achieved by maintaining a desired<br />
delivery speed across the sensor network through a<br />
combination <strong>of</strong> feedback control and non-deterministic<br />
geographic forwarding. SPEED is a stateless and<br />
geographic position based routing protocol without<br />
end-to-end path set-up before packets forwarding.<br />
However, SPEED neither takes another important QoS<br />
metric (packet loss ratio) into account, nor balances well<br />
the energy consumption in sensor nodes. Moreover, before<br />
computing data routes, SPEED must assume each sensor<br />
node is location-aware.<br />
In [12], E. Felemban et al designed a routing protocol<br />
called MMSPEED, which presents not only the service<br />
timeliness by guaranteeing multiple packet delivery speed<br />
options, but also the service reliability in a probabilistic<br />
multi-path forwarding manner. The QoS provision is<br />
realized in a localized way without global network<br />
information by employing localized geographic packet<br />
forwarding with dynamic compensation to <strong>of</strong>fset the local<br />
decision inaccuracies. Although MMSPEED adds<br />
reliability <strong>of</strong> route discovery in SPEED, and is more<br />
suitable for large-scale dynamic sensor networks, it still<br />
has some disadvantages similar to SPEED, for instance,<br />
© 2011 ACADEMY PUBLISHER<br />
geographic forwarding based and lack <strong>of</strong> important QoS<br />
constraints.<br />
L. Shu et al presented a two phase geographic greedy<br />
forwarding (TPGF) routing algorithm for multimedia<br />
sensor networks, which supports multipath transmission<br />
and hole-bypassing, as well as shortest path transmissions<br />
[13]. TPGF consists <strong>of</strong> two phases: geographic<br />
forwarding and path optimization, wherein geographic<br />
forwarding is responsible for exploring a delivery<br />
guaranteed route while bypassing the holes in WSN, path<br />
optimization is responsible for optimizing the found path<br />
with the least number <strong>of</strong> nodes by a method <strong>of</strong> label<br />
based optimization. Nevertheless, TPGF and MMSPEED<br />
fall into the same category essentially.<br />
In order to differentiate video and audio applications<br />
which both employ TPGF as their routing protocol, on<br />
the basis <strong>of</strong> TPGF, a multi-priority multi-path selection<br />
scheme (MPMPS) is proposed for transport layer in WSN<br />
[14]. MPMPS supports multiple transmission priorities<br />
and chooses the maximum number <strong>of</strong> paths for<br />
maximizing throughput <strong>of</strong> streaming data transmission<br />
and guaranteeing the end-to-end transmission delay.<br />
K. Akkaya et al proposed an energy-aware QoS<br />
routing protocol for sensor networks [15]. The protocol<br />
finds a least-cost, delay-constrained path for real-time<br />
data in terms <strong>of</strong> link cost that captures nodes’ energy<br />
reserve, transmission energy, error rate and other<br />
communication parameters. Moreover, the throughput for<br />
non-real-time data is maximized by adjusting the service<br />
rate for both real-time and non-real-time data at the<br />
sensor nodes. The main problem <strong>of</strong> this protocol is that it<br />
requires complete knowledge <strong>of</strong> network topology at each<br />
node in order to compute a route, so that it is unsuitable<br />
for the large-scale sensor networks. In addition, like<br />
SPEED etc, this protocol does not take packet loss ratio<br />
into account, and must resort to extra localization<br />
equipments or algorithms to locate sensor nodes before<br />
route discovery.<br />
Seen from the mentioned above, like GPSR [16],<br />
SPEED, MMSPEED, TPGF, MPMPS etc are all<br />
geographic greedy forwarding based routing algorithms,<br />
which need the geographic coordinate <strong>of</strong> each node in<br />
sensor networks to compute the routing path. They also<br />
leave out some important QoS constraints in their routing<br />
models. Besides, all <strong>of</strong> the routing algorithms are based<br />
on data-driven delivery mode, which mode is unsuitable<br />
for the multimedia applications with periodic data.<br />
On the basis <strong>of</strong> the discussed above, we manage to<br />
design a routing algorithm for multimedia<br />
communications with periodic data in sensor networks,<br />
which is supposed to satisfy the following design goals: 1)<br />
end-to-end delay guarantee; 2) end-to-end packet loss<br />
ratio guarantee; 3) optimal node energy consumption; 4)<br />
minimum MAC layer support (i.e., it does not need<br />
special QoS-aware MAC support); 5) optional MAC<br />
layer services (MAC multicast and differentiation service<br />
are available); 6) extra position measurement equipment<br />
or location message exchanges unnecessary; 7) message<br />
suppression and data aggregation; 8) resilience and<br />
reliability.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 941<br />
Ⅲ. SYSTEM MODEL<br />
A WSN can be represented as a weighted, connected<br />
graph G=(V,E), where V is the set <strong>of</strong> nodes and E denotes<br />
the set <strong>of</strong> wireless communication links connecting the<br />
nodes.<br />
A. Network Service Model<br />
It is believed that in WSN the communications from<br />
multiple source nodes to one sink node construct a<br />
reversed multicast tree. Suppose that s(s∈V) is a source<br />
node <strong>of</strong> the multicast tree, and M( M ⊆{ V − { d}}<br />
) is the<br />
set <strong>of</strong> source nodes <strong>of</strong> the multicast tree, where d(d∈V) is<br />
the sink node. Let us use T(M,d) to denote the multicast<br />
tree. In order to describe the service requirements from<br />
multimedia applications in sensor networks (i.e., the<br />
timeliness and reliability <strong>of</strong> communications, as well as<br />
the optimal energy consumption in sensor nodes), we<br />
define the network service model as follows, which is<br />
able to comprehensively depict these requirements.<br />
⎧e2e<br />
_ delay( p( s, d)) ≤ D<br />
⎪ priority _ levels( f ( T )) ≥ 2<br />
⎨<br />
⎪ arrival _ probability( p( s, d )) ≥ P<br />
⎩<br />
⎪ energy _ cos t( T ( M , d )) = min[...]<br />
Where e2e_delay, priority_levels, arrival_probability and<br />
energy_cost denote the end-to-end delay, the number <strong>of</strong><br />
traffic priority levels, the packet arrival probability, and<br />
the energy consumption respectively; p(s,d) is the path<br />
from the source node s( ∀ s∈M) to the sink node d in the<br />
reversed multicast tree T(M,d); and f(T) refers to the<br />
traffic type in T(M,d); D and P denote the constraints to<br />
the end-to-end delay and the packet arrival probability<br />
respectively.<br />
B. Routing Model<br />
Aiming at sufficing for QoS requirements <strong>of</strong><br />
multimedia applications, formula (1) represents the<br />
network services that should be provided by WSN. The<br />
network services mainly consist <strong>of</strong> the service from MAC<br />
layer and the service from network layer. However, with<br />
respect to the network layer, the routing service model<br />
should be defined as follows:<br />
⎧ residual _ energy( T ( M , d )) ≥ E<br />
⎪<br />
⎪e2e<br />
_ delay( p( s, d)) ≤ D<br />
⎨<br />
⎪ packet _ loss( p( s, d )) ≤ R<br />
⎩<br />
⎪ hopcount( p( s, d )) = min[...]<br />
Where residual_energy, e2e_delay, packet_loss and<br />
hopcount denote the residual energy ratio in sensor node,<br />
end-to-end delay, end-to-end packet drop ratio and<br />
hop-count respectively; E, D and R denote the constraints<br />
to the residual energy ratio, end-to-end delay and packet<br />
drop ratio respectively.<br />
In (2), there exist the following relationships:<br />
© 2011 ACADEMY PUBLISHER<br />
(1)<br />
(2)<br />
⎧residual<br />
_ energy( T ( M , d )) =<br />
⎪<br />
⎪<br />
min{ residual _ energy( n), n ∈ T ( M , d )}<br />
⎪ e2e _ delay( p( s, d))<br />
=<br />
⎪<br />
⎨ ∑ delay( e) + ∑ delay( n)<br />
⎪ e∈p( sd , ) n∈p( sd , )<br />
⎪ packet _ loss( p( s, d))<br />
=<br />
⎪<br />
⎪1<br />
− ∏ (1− packet _ loss( n))<br />
⎩ n∈p( s, d )<br />
Where n∈V, e∈E; delay(n) and delay(e) are the delay<br />
functions <strong>of</strong> sensor nodes and wireless links respectively.<br />
Note that hereby we use the hop-count <strong>of</strong> a path to<br />
represent the accumulated node energy consumption<br />
along the path.<br />
Ⅳ . PROPOSED ALGORITHM<br />
In this section, we present a routing algorithm named<br />
MCRA (Multi-Constrained Routing Algorithm), which is<br />
based on query-flooding and query-driven data delivery<br />
mode, because this mode has its intrinsic resilience and<br />
reliability [17,18]. In advance suppose that the delay<br />
metric <strong>of</strong> duplex bi-directional wireless links in sensor<br />
networks has symmetric property.<br />
A. Routing Procedure<br />
The message used to query an event occurred within a<br />
surveillance area is usually called interest. In MCRA, the<br />
format <strong>of</strong> interest message is defined as Fig. 1, where each<br />
item between a pair <strong>of</strong> parentheses is the comment to the<br />
corresponding field name.<br />
interest.type (query event type)<br />
interest.nodes (visited node set)<br />
interest.hopcount (trip hop count)<br />
interest.e2e_delay (trip time record)<br />
interest.packet_loss (accumulated packet drop ratio)<br />
interest.D (end-to-end delay constraint)<br />
interest.R (constraint to packet drop ratio)<br />
interest.E (constraint to energy consumption)<br />
interest.neighbors (temporary neighbor table)<br />
interest.TTL (time to live)<br />
Figure 1. The format <strong>of</strong> interest message<br />
1) Starting from a sink node, interest messages (copies<br />
<strong>of</strong> one interest) are flooded to all <strong>of</strong> neighbors <strong>of</strong> the sink<br />
node. When some intermediate node in the network (e.g.,<br />
node k) receives an interest, the node k begins to measure<br />
its residual energy and the packet drop ratio in it (the<br />
packet drop ratio is a statistical value during a period <strong>of</strong><br />
time, which has been stored in this node), as well as the<br />
current system time from the synchronous clock in node k.<br />
Afterwards, node k uses the detected information to<br />
calculate and rewrite the two fields interest.e2e_delay and<br />
interest.packet_loss in this interest, as illustrated in Fig. 1.<br />
Of course here k ∉ interest. nodes , that is to say,<br />
interests will never visit the nodes which they have already<br />
visited.<br />
(3)
942 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
2) The intermediate node k starts to check QoS<br />
constraints, as illustrated in Fig. 2. If the result <strong>of</strong> the<br />
following expression is true:<br />
residual _ energy( k) ≥ E ∧<br />
interest. packet _ loss ≤ R ∧<br />
(4)<br />
interest. e2e_ delay ≤ D<br />
Then interest.hopcount=interest.hopcount+1, and the ID<br />
<strong>of</strong> node k is added into the list interest.nodes in the interest.<br />
Otherwise, the interest will be discarded by node k.<br />
Monitored<br />
Area s<br />
m<br />
i<br />
Data Interest<br />
Figure 2. The routing process <strong>of</strong> MCRA<br />
k<br />
n<br />
j<br />
sink<br />
3) Delaying for a period <strong>of</strong> time, if node k receives<br />
multiple interests that have the same query event type (i.e.<br />
interest.type) but traveled along different paths, node k<br />
will select the best interest (i.e., that one with the<br />
minimum value <strong>of</strong> interest.hopcount, or<br />
interest.e2e_delay, or interest.packet_loss) in them, and<br />
drops the others (details in subsection B.2). Afterwards,<br />
node k forwards the interest to its neighbors excluding the<br />
nodes visited by the interest, in terms <strong>of</strong> a restraining<br />
forwarding scheme (details in subsection B.1).<br />
4) The above steps are repeated until the interest arrives<br />
at a node s that matches the content <strong>of</strong> field interest.type in<br />
this interest, i.e., source node, eventually.<br />
5) Node s performs operations in a similar way to other<br />
nodes except for not forwarding the interest. It begins to<br />
read the list value from field interest.nodes in this interest,<br />
and sends data towards the sink node by using the node list<br />
value as the travel path <strong>of</strong> its packets. The node list<br />
information will guide the data forwarding by means <strong>of</strong><br />
piggybacking, i.e., the path information is carried by the<br />
packets, as shown in Fig. 2. Note that the logical<br />
coordinate vector (see subsection D) <strong>of</strong> node s is also sent<br />
towards the sink by means <strong>of</strong> piggybacking.<br />
6) When the sink node receives the data sent by source<br />
node s, it also receives the logical coordinate vector <strong>of</strong><br />
node s (i.e., the hop-count information from multiple sink<br />
nodes to the source node s) carried back by the packets.<br />
Now the sink node may adopt the coordinate system<br />
proposed in subsection D to calculate the coordinate<br />
value <strong>of</strong> source node s by its logical coordinate vector.<br />
The differentiation service in MAC layer may be<br />
applied to MCRA, so that real-time traffic and best-effort<br />
traffic in networks are classified into different priority<br />
levels to forward, which provides differentiation service<br />
for sensor networks [19]. Meanwhile, in order to avoid<br />
the collisions in routing process as soon as possible, here<br />
we may optionally apply a MAC protocol supporting<br />
MAC multicast/broadcast, e.g., CAPWAP protocol (RFC<br />
© 2011 ACADEMY PUBLISHER<br />
5416), though MAC protocols without<br />
multicast/broadcast are also able to accomplish the<br />
routing mission in MCRA [20].<br />
It should be mentioned here that MCRA does not need<br />
the support <strong>of</strong> any special QoS-aware protocol in MAC<br />
layer. The common MAC protocols also can cooperate<br />
with the MCRA well.<br />
B. Message Suppression<br />
During the routing forwarding process mentioned<br />
above, we adopt some policies <strong>of</strong> message suppression in<br />
order to reduce the message redundancies and the<br />
re-transmission probability caused by collisions. Message<br />
suppression focuses on reducing the number <strong>of</strong> interests<br />
occurred during the routing process, so that it can help not<br />
only save the energy consumption in sensor nodes but also<br />
reduce the time <strong>of</strong> convergence <strong>of</strong> route discovery. In<br />
MCRA, the policies <strong>of</strong> message suppression consist <strong>of</strong> two<br />
aspects: restraining forwarding and deferring forwarding.<br />
1. Restraining Forwarding<br />
First <strong>of</strong> all, similar to other routing algorithms, each <strong>of</strong><br />
nodes in MCRA periodically broadcasts a beacon packet<br />
(HELLO message) to its neighbors, so that every node can<br />
keep a neighbor table to store the information passed by<br />
beacons. Each entry inside the table has the following<br />
fields: (NeighborID, ExpireTime), where field ExpireTime<br />
is used to timeout this entry. If a neighbor entry is not<br />
refreshed after certain timeout, it will be removed from the<br />
neighbor table. Since geographic position information is<br />
not necessary in our routing process, no position<br />
information inside our neighbor tables.<br />
The main ideal <strong>of</strong> restraining forwarding policy is how<br />
to lessen the amount <strong>of</strong> interest messages by restraining<br />
some <strong>of</strong> nodes from forwarding interests. As shown in Fig.<br />
3, after an interest from node n1 is sent to node n2, node n2<br />
floods this interest to its neighbors, node n3, n4, n5, and n6<br />
excluding the node n1, in manner <strong>of</strong> multicast (because<br />
interest cannot be sent back to its visited nodes). In<br />
addition, node n2 stores its neighbor information into field<br />
interest.neighbors in this message, so did node n1. Note<br />
that the neighbor information <strong>of</strong> node n2 replaces the<br />
existing neighbor information <strong>of</strong> node n1 in this interest<br />
field interest.neighbors.<br />
ID 246<br />
ID<br />
2<br />
3<br />
5<br />
...<br />
n3<br />
n6<br />
n4<br />
n5<br />
n2<br />
ID<br />
1<br />
3<br />
4<br />
5<br />
6<br />
n1<br />
Unrestrained Node<br />
Restrained Node<br />
Unknown Node<br />
ID<br />
2<br />
Figure 3. The restraining forwarding policy<br />
r<br />
Interest<br />
When node n3 receives the interest, it begins to<br />
compare each element in the list interest.neighbors with
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 943<br />
each entry in its neighbor table and calculate the value <strong>of</strong><br />
following expression:<br />
ΦN⊆Φ INTN<br />
(5)<br />
Wherein N Φ denotes the set <strong>of</strong> its neighbors, and Φ INTN<br />
is the set <strong>of</strong> the elements in the interest field<br />
interest.neighbors. If the value <strong>of</strong> (5) is true, this node is a<br />
restrained node. That means it discards this interest<br />
immediately, e.g. node n3 and n6 in Fig. 3. Otherwise, this<br />
node is an unrestrained node, which forwards interest to its<br />
neighbors excluding the visited nodes, moreover rewrites<br />
field interest.neighbors in the interest, e.g. node n4. So the<br />
amount <strong>of</strong> interest messages in network can be lessened<br />
drastically.<br />
2. Deferring Forwarding<br />
The main purpose <strong>of</strong> deferring forwarding policy is to<br />
lower the amount <strong>of</strong> interests by dynamically deferring the<br />
forwarding actions on nodes. Deferring forwarding policy<br />
is able to make these nodes have enough time to collect<br />
and merge the interests from their neighbors as many as<br />
possible.<br />
Let the forwarding delay on some node be ∆ τ . As<br />
shown in Fig. 2, during a period <strong>of</strong> time ∆ τ , node k<br />
receives multiple interests with the same interest.type<br />
value from its neighbors, and all <strong>of</strong> which suffice for QoS<br />
constraints, i.e., the result <strong>of</strong> expression (4) is true. Node k<br />
will list these interests in some order, e.g., in<br />
interest.e2e_delay, or interest.packet_loss, or<br />
interest.hopcount order. And then, node k selects the best<br />
element from the ordered list and forwards it, drops the<br />
others. After forwarding the best interest, if the node k<br />
again receives some interests with the same interest.type<br />
and satisfying the constraint criteria (4), it also discards<br />
them.<br />
Considering the factors that influence forwarding delay,<br />
according to our results <strong>of</strong> extensive simulation<br />
experiments, we give an empirical formula that calculates<br />
the value <strong>of</strong> ∆ τ . For the forwarding delay on sensor node<br />
k, the formula is as follows:<br />
D D−T N<br />
V D V<br />
trip k<br />
∆ τk = ρk<br />
× × (6)<br />
Wherein D is the end-to-end delay constraint, |V| is the<br />
total number <strong>of</strong> sensor nodes, T denotes the trip time <strong>of</strong><br />
trip<br />
an interest, N is the number <strong>of</strong> neighbors <strong>of</strong> sensor node<br />
k<br />
k, and ρ is the instantaneous queue size (in bytes) in<br />
k<br />
node k. From (6), we can see that an arbitrary sensor node<br />
(<strong>of</strong> course, apart from sink nodes) is able to easily<br />
calculate its forwarding delay ∆ τ by reading the<br />
information in interest and its neighbor table.<br />
C. Data Aggregation<br />
In multimedia applications, since sensor nodes in a<br />
monitored area might generate significant redundant data,<br />
and that duplicate or similar data packets from multiple<br />
sensors need to be aggregated, so that the amount <strong>of</strong><br />
transmissions would be reduced.<br />
© 2011 ACADEMY PUBLISHER<br />
Suppose that node s and node i in a same monitored area<br />
send the data they detected respectively to their<br />
downstream node k, as shown in Fig. 4. If there exists<br />
redundancy in the data, node k will perform the<br />
aggregation computation by using some aggregation<br />
functions, such as suppression (eliminating duplicates),<br />
min, max and average etc. Some <strong>of</strong> these functions can be<br />
performed either partially or fully on each node in WSN.<br />
Monitored Area<br />
s<br />
m<br />
i<br />
Data Interest<br />
k<br />
Figure 4. The process <strong>of</strong> data aggregation<br />
n<br />
j<br />
sink<br />
In Fig. 4, if node i selects node j as its downstream node<br />
rather than node k, the aggregation computation will be<br />
performed on the sink node.<br />
Recognizing that computation consumes much less<br />
energy than communication, substantial energy savings<br />
can be obtained through the above process <strong>of</strong> data<br />
aggregation.<br />
D. Localization Scheme<br />
Borrowed from the idea <strong>of</strong> logical coordinate system<br />
[10], we design a new localization approach based on<br />
hop-count information for MCRA. The key difference is<br />
that instead <strong>of</strong> constructing a coordinate space for routing,<br />
we only use the hop-count information acquired from<br />
routing process to calculate the location <strong>of</strong> target. It not<br />
only avoids the message overhead induced by the<br />
coordinate space construction, but also eliminates the<br />
negative impact upon routing process due to the imprecise<br />
node coordinate information.<br />
1<br />
0<br />
Sink Common Node Source<br />
m<br />
s(6,7,4,3)<br />
Figure 5. The localization example<br />
Suppose that network nodes are placed on a plane <strong>of</strong><br />
rectangle m units in length and n units in width (or other<br />
given size shapes, e.g., triangle, ellipse etc), where there<br />
are 4 perimeter nodes (or called landmarks), as shown in<br />
Fig. 5. Here we may take these landmarks as sink nodes<br />
(multiple sinks are used to improve the positioning<br />
precision <strong>of</strong> source nodes). We also assume that these sink<br />
nodes have known their respective hop-counts to some<br />
2<br />
n<br />
3
944 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
source node, e.g., the node s(6,7,4,3), where the numbers<br />
between the pair <strong>of</strong> parentheses form the logical<br />
coordinate vector <strong>of</strong> node s. Each <strong>of</strong> elements in this<br />
vector in order represents the respective hop-count from<br />
each sink to node s. The serial number <strong>of</strong> each element is<br />
the ID <strong>of</strong> its corresponding landmark (sink node). So the<br />
actual coordinate value <strong>of</strong> source node relative to these<br />
sinks can be computed in the following manner:<br />
∑<br />
∑<br />
L<br />
L<br />
−<br />
−<br />
i<br />
j<br />
x= m , y= n<br />
L + L L + L<br />
∑ −<br />
i ∑ + − +<br />
i ∑ j ∑ j<br />
(7)<br />
Where L ur denotes the logical coordinate vector <strong>of</strong> target;<br />
i − , i + , j − , and j + are the IDs <strong>of</strong> left, right, down, and<br />
up perimeter nodes in this coordinate plane, respectively.<br />
E. Network Dynamics<br />
Although most <strong>of</strong> sensor network architectures assume<br />
that sensor nodes are stationary, it is sometimes deemed<br />
necessary to support the mobility <strong>of</strong> sink nodes, and that<br />
the communications usually fail due to the energy<br />
exhaustion in sensor nodes or other causes. Besides the<br />
dynamics <strong>of</strong> network topology, the dynamics caused by<br />
the imprecision <strong>of</strong> network state information also demands<br />
that routing protocols are able to adapt to these variations<br />
<strong>of</strong> network state.<br />
In MCRA, there are two policies used to implement this<br />
function. The first policy is notification update, as shown<br />
in Fig. 6, during the data delivery, once the node k detects<br />
the communication failure from node s to k, it sends a<br />
notification message towards the sink node, immediately,<br />
which informs the sink node restarts a new routing process.<br />
Meanwhile, we also may use hold-down timer in sink<br />
node, if necessary. The period <strong>of</strong> hold-down timer is<br />
usually set to 3ω , where ω is the average time<br />
interval <strong>of</strong> the query occurrences on sink nodes.<br />
Monitored<br />
Area s<br />
Data Interest<br />
i<br />
m<br />
k<br />
n<br />
j<br />
Notification<br />
Figure 6. The adaptation to the network dynamics<br />
sink<br />
The second policy is periodic update, which is that sink<br />
node restarts periodically a new routing process in certain<br />
time interval ∆ t . Compared to cable networks, here ∆ t<br />
should have a larger value. We usually set ∆t≥ 30ω<br />
,<br />
where ω has the same meaning as above.<br />
Ⅴ. DISCUSSION<br />
A. Correctness Pro<strong>of</strong><br />
The correctness and feasibility <strong>of</strong> MCRA can be<br />
approved by the following two non-formalized theorems.<br />
© 2011 ACADEMY PUBLISHER<br />
Theorem 1 If a feasible path that suffices for QoS<br />
constraints exists, MCRA is able to find it.<br />
Pro<strong>of</strong>: In the routing process <strong>of</strong> MCRA, interest<br />
messages are diffused in a restricted flood manner to seek<br />
all <strong>of</strong> the feasible paths. Hence, only if the paths that<br />
satisfy QoS constraints exist, MCRA must be able to find<br />
them. The theorem holds.<br />
Theorem 2 The paths found by MCRA form a<br />
loop-free reversed multicast tree with optimal energy<br />
consumption, which suffices for the end-to-end delay and<br />
packet drop ratio requirements.<br />
Pro<strong>of</strong>: As mentioned before, in MCRA, interest<br />
message neither visits those paths that do not suffice for<br />
QoS constraints, nor visits those paths visited by it. In<br />
addition, each <strong>of</strong> paths from each source node to the sink<br />
node is the path with minimal energy consumption (i.e.,<br />
minimal hop-count), because the hop-count <strong>of</strong> a path<br />
represents the accumulated energy consumption along the<br />
path (mentioned in section II). Besides, MCRA balances<br />
the energy consumption in networks by using the<br />
constraint to the residual energy ratio in sensor nodes.<br />
Thus, we may deem that the tree constructed by these<br />
optimal paths is a loop-free and reversed multicast tree that<br />
has the optimal energy consumption and satisfies the<br />
end-to-end delay and packet drop ratio requirements. The<br />
theorem holds.<br />
B. Complexity Analysis<br />
In MCRA, the overhead <strong>of</strong> message transmissions<br />
determines its complexity, since not only the total energy<br />
consumption but also the computation complexity in<br />
networks are proportional to the number <strong>of</strong> transmissions.<br />
Suppose that there exist |Q| queries in G=(V,E). We<br />
also may assume no collisions or very few collisions<br />
when messages transmit among nodes if the value <strong>of</strong> |Q|<br />
is not large, so each node in G performs 2|V|<br />
transmissions (|V| HELLO messages plus |V| interest<br />
messages) at most, where |V| denotes the number <strong>of</strong> nodes.<br />
Hence, we can draw that the complexity <strong>of</strong> message<br />
overhead is O(|V||Q|) theoretically.<br />
C. Simplified MCRA<br />
For the sake <strong>of</strong> suppression messages, we adopt<br />
restraining forwarding and deferring forwarding policies<br />
in MCRA. However, the restraining forwarding policy<br />
also imports extra control overhead (HELLO messages)<br />
so as to establish neighbor tables. We design a simplified<br />
MCRA named MCRA-S, which does not apply the<br />
restraining forwarding policy during its routing process.<br />
According to the complexity analysis above, the<br />
complexity <strong>of</strong> MCRA-S should be O(|V||Q|) theoretically,<br />
which is similar to MCRA, maybe even slightly better<br />
than MCRA. But this case only happens when the<br />
network has relatively low node density. As the sensor<br />
node density increases, MCRA would be getting better<br />
than MCRA-S because <strong>of</strong> the increasing collisions in<br />
MCRA-S. Our simulation experiments will prove this<br />
conclusion.<br />
In MCRA-S, despite the lack <strong>of</strong> neighbor table on<br />
sensor node, we may calculate the N value in (6) by<br />
k
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 945<br />
the following equation used to compute the neighbor<br />
number <strong>of</strong> a sensor node in [21,22].<br />
V<br />
N r<br />
S π<br />
2<br />
k = (8)<br />
Where |V| is the total number <strong>of</strong> sensor nodes, S denotes<br />
the total surface area covered by all sensor nodes, and r is<br />
the communication radius <strong>of</strong> node k.<br />
Ⅵ . SIMULATION<br />
In our evaluation, we compare the performance <strong>of</strong> four<br />
different routing algorithms: MCRA, SPEED, Directed<br />
Diffusion (DD) [23,24] and MCRA-S. Directed Diffusion<br />
is a typical data-centric algorithm based on query-driven<br />
data delivery mode, which optimizes single objective (e.g.<br />
energy savings) by selecting empirical good paths and by<br />
caching and processing data in network. SPEED is a<br />
representative algorithm that can guarantee the timeliness<br />
<strong>of</strong> multimedia communications by a combination <strong>of</strong><br />
feedback control and non-deterministic geographic<br />
forwarding. In addition, both Directed Diffusion and<br />
SPEED are correlative with our algorithm MCRA.<br />
We simulate MCRA on NS2 (ns-2.33), because this<br />
version has implemented many MAC layer protocols<br />
applied in WSN and Directed Diffusion algorithm, as<br />
well as XCP (explicit congestion control protocol) that is<br />
similar to SPEED algorithm [25]. Table I describes the<br />
main setting parameters and scenarios for our<br />
simulations.<br />
TABLE I<br />
THE SIMULATION SCENARIO SETTINGS<br />
Configuration Options Setting Values<br />
Routing MCRA, SPEED, DD, MCRA-S<br />
llType (link layer type) LL (delay_: 0.25ms, bandwidth_:<br />
not used)<br />
macType (MAC layer protocol) Mac/802_11<br />
function)<br />
(without multicast<br />
ifqType (interface queue type) Queue/DropTail/PriQueue<br />
ifqLen (interface queue length) 50<br />
antType (antenna type) Antenna/OmniAntenna<br />
propType (propagation model) Propagation/TwoRayGround<br />
phyType (network interface type) Phy/WirelessPhy<br />
channel (channel type) Channel/WirelessChannel<br />
energyModel (energy model) EnergyModel<br />
Terrain (300m, 300m)<br />
Node placement Uniform distribution<br />
Node number and (Variables)<br />
Communication range<br />
We present the following metrics to evaluate and<br />
compare the performance <strong>of</strong> the four routing algorithms:<br />
1) end-to-end delay under different node number; 2)<br />
end-to-end packet loss ratio under different node number;<br />
3) control message overhead under different node number<br />
and different communication range (radio radius)<br />
respectively; 4) packet delivery ratio<br />
N sec (Nsec and<br />
N req<br />
Nreq are the successful connection number and sum <strong>of</strong><br />
request connections respectively) under different node<br />
number; 5) average energy consumption J<br />
(J and n are<br />
n<br />
© 2011 ACADEMY PUBLISHER<br />
the total energy consumption and the node number<br />
respectively) under different node number.<br />
In our simulations, sink nodes and source sensor nodes<br />
are selected randomly in the scenarios, and the flows<br />
between them (4 flows in each simulation scenario where<br />
there are 4 queries and 4 sink nodes for DD, MCRA and<br />
MCRA-S) are CBR traffic pattern with a rate <strong>of</strong> 50<br />
packets/second. End-to-end delay constraint D, packet<br />
drop ratio constraint R, and residual energy ratio constraint<br />
E are uniformly distributed in [100ms, 250ms], [10%,<br />
30%], and [5%, 10%], respectively. The communication<br />
ranges in the scenarios with varying node number are set<br />
to 30m, on the other hand, the node number in the<br />
scenarios with varying communication range is 100. Note<br />
that our simulations do not count in the extra costs<br />
imported by measurement equipments or location<br />
message exchanges in SPEED and DD. In our<br />
experiments, each <strong>of</strong> the reported values is the average<br />
result over 100 runs with different random seeds and<br />
different random node topologies.<br />
A. End-to-end Delay<br />
End-to-end delay measures the network delay<br />
performance <strong>of</strong> these algorithms. Fig. 7 plots the<br />
end-to-end delay for the four different routing algorithms.<br />
At each point, we average the e2e delays <strong>of</strong> all then<br />
packets from the 24 flows (100 runs with 4 flows each).<br />
Delay (ms)<br />
300<br />
260<br />
220<br />
180<br />
140<br />
100<br />
60<br />
MCRA<br />
SPEED<br />
DD<br />
MCRA-S<br />
Node Number<br />
20<br />
10 40 70 100 130 160 190 220<br />
Figure 7. End-to-end delay vs. node number<br />
Seen from Fig. 7, SPEED has the best e2e delay<br />
performance as its optimized objective, in particular, in<br />
the route acquisition phase, because it is a<br />
non-deterministic geographic routing with less initial<br />
delay cost. Obviously, Directed Diffusion that only<br />
optimizes the energy savings by flooding has the worst<br />
e2e delay performance. Both MCRA and MCRA-S<br />
perform much better than DD, because the e2e delay is<br />
considered as an important element in their routing model.<br />
However, when the network has low density (
946 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
in Fig. 8 are the summary <strong>of</strong> 100 randomized runs. From<br />
the reported data, we can clearly see MCRA and<br />
MCRA-S are better than SPEED and DD. The main<br />
cause is that the e2e packet loss ratio in MCRA is an<br />
important performance constraint during the route<br />
discovery. In comparison to MCRA and MCRA-S, when<br />
the network has fewer sensor nodes (
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 947<br />
Energy Consumption<br />
(mWhr)<br />
40<br />
35<br />
30<br />
25<br />
20<br />
15<br />
10<br />
5<br />
MCRA<br />
SPEED<br />
DD<br />
MCRA-S<br />
Node Number<br />
0<br />
10 40 70 100 130 160 190 220<br />
Figure 12. Average energy consumption vs. node number<br />
F. Localization Error<br />
Unlike SPEED, GPSR, and BVR [9] etc., MCRA is<br />
not a routing protocol based on location information, i.e.,<br />
the route discovery does not need the location<br />
information <strong>of</strong> sensor nodes. So the localization error in<br />
MCRA only affects the positioning precision <strong>of</strong> targets.<br />
Let δ = δδ denote the position error, where x y<br />
δ and x<br />
δ are the position errors in horizontal and vertical<br />
y<br />
direction respectively. By extensive experiments, we plot<br />
the position error with respect to different sensor node<br />
number and different sink (i.e., perimeter node) number,<br />
as shown in Fig. 13. From it, we can find the relationships<br />
between the localization error and the node density as<br />
well as the perimeter node number.<br />
Position Error (%)<br />
50<br />
40<br />
30<br />
20<br />
10<br />
4 Perimeter Nodes<br />
6 Perimeter Nodes<br />
8 Perimeter Nodes<br />
10 Perimeter Nodes<br />
Node Number<br />
0<br />
10 40 70 100 130 160 190 220<br />
Figure 13. Position error vs. sensor node and sink number<br />
Ⅶ . CONCLUSION<br />
This paper presents a multi-constrained routing<br />
algorithm MCRA based on query-flooding and<br />
query-driven data delivery mode for multimedia<br />
applications with periodic data in sensor networks.<br />
MCRA can not only provide end-to-end delay guarantee<br />
and packet loss ratio guarantee for multimedia<br />
communications, but also improve and balance the<br />
energy consumption in sensor nodes. Besides, MCRA<br />
adopts efficient policies to suppress message flooding and<br />
lessen data redundancy. In MCRA, extra position<br />
measurement equipment or location message exchanges<br />
are unnecessary, and that routing computation does not<br />
require the geographic or logical coordinate information<br />
<strong>of</strong> sensor nodes, however, target locations we concern<br />
still can be figured out on sink nodes by using hop-count<br />
information. In addition, we may optionally adopt MAC<br />
© 2011 ACADEMY PUBLISHER<br />
multicast in MCRA in order to further lessen its control<br />
message overhead. Meanwhile, MAC differentiation<br />
service can be applied to MCRA, so that real-time traffic<br />
and best-effort traffic in WSN can be classified into<br />
different forwarding priority levels. Theoretical analysis<br />
and extensive simulations not only demonstrate the<br />
correctness <strong>of</strong> MCRA, but also show that it has a good<br />
overall performance, thanks to the low end-to-end delay<br />
and loss ratio <strong>of</strong> data delivery, the low average energy<br />
consumption, the high packet delivery ratio, and the<br />
moderate control message overhead. Our further work is<br />
to investigate the performance <strong>of</strong> MCRA when the<br />
number <strong>of</strong> queries in WSN increases, and to manage to<br />
improve the localization precision <strong>of</strong> sensor nodes.<br />
APPENDIX A: THE METACODE MCRA<br />
//1. The definition <strong>of</strong> interest packet header<br />
struct interest_header {<br />
char *type; char **nodes; long hopcount; double<br />
e2e_delay; double packet_loss; double D; double R;<br />
double E; char **neighbors; double TTL; }<br />
//2. The main metacode MCRA<br />
MCRA (V, M, D, R, E, TTL) {<br />
Initialization (V);<br />
while (M≠0) {<br />
for each sin k∈ M {<br />
interest.nodes = AddList(null, sink.ID);<br />
sink.broadcast_interests(interest);<br />
for each k∈{ V − sin k}<br />
{<br />
if ( k ∉ {int erest. nodes}<br />
) {<br />
interest.e2e_delay += k T ∆ ; /* Here k T ∆ is<br />
the trip time from the upstream <strong>of</strong> node k to it. */<br />
if (interest.e2e_deay >= TTL) {<br />
k.drop_interest(interest); }<br />
interest.packet_loss *= packet _ loss ; k<br />
if ( kresidual . _energy≥<br />
E &&<br />
interest. packet _loss<br />
≤ R &&<br />
interest.2 e e _delay<br />
≤ D ) {<br />
interest.hopcount += 1;<br />
interest.nodes = AddList(interest.nodes,<br />
k.ID);<br />
} else { k.drop_interest(interest); }<br />
/* Check if the node is source node. */<br />
if ( k = interest.type) { s = k ;<br />
s.merge_interests( ∆ τ ); /* Here s ∆ τ is s<br />
the delay <strong>of</strong> source node s. */<br />
s.coordinate_vector =<br />
AddList(interest.hopcount);<br />
s.send_data(interest.nodes,<br />
s.coordinate_vector, sink);<br />
} else {<br />
if (k.RFF(interest) == “Restrained Node”)<br />
{k.drop_interest(interest);<br />
} else { k.merge_interests( ∆ τ ); k
948 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
N = DeleteList(k.neighbors,<br />
interset.nodes); /* Here N is the set<br />
<strong>of</strong> node k’ neighbors except for the visited nodes. */<br />
k.forward_interest(N); } }<br />
}<br />
}<br />
sink.TLS(s.coordinate_vector); /* Calculate the actual<br />
coordinate values <strong>of</strong> each monitored target node by Target<br />
Localization Scheme (TLS) */<br />
M=M-sink;<br />
}<br />
if (M=0) end;}}<br />
//3. The meta-code <strong>of</strong> initialization subroutine<br />
Initialization (V) {<br />
for each k∈ V {<br />
k.exchange_beacons(HELLO);<br />
k.get_neighbors(NeighborID, ExpireTime);<br />
}<br />
}<br />
//4. The meta-code Restraining Forwarding Function<br />
(RFF)<br />
RFF (INT) {/* Here INT denotes interest. */<br />
if (this.neighbors ⊆ INT.neigbors) {<br />
this = “Restrained Node”;<br />
} else {this = “Unrestrained Node”;<br />
INT.neighbors = this.neighbors;<br />
}<br />
return (“Restrained Node” || “Restrained Node”);<br />
}<br />
//5. The meta-code Target Localization Scheme (TLS)<br />
TLS ( L ur ) {/* Here L ur is the logical coordinate vector <strong>of</strong><br />
target. */<br />
if ( L ur >= 4) {<br />
∑<br />
∑ ∑ );<br />
return ( x= m<br />
L −<br />
i<br />
L − + i L +<br />
i<br />
return (<br />
y = n<br />
∑ L −<br />
j<br />
L + L<br />
);<br />
∑ ∑<br />
− +<br />
j j<br />
/* Here i − , i + , j − , and j + are the IDs <strong>of</strong> left, right,<br />
down, and up landmarks in this coordinate plane,<br />
respectively, which are known; m and n are also known<br />
planar shape parameters. */<br />
}<br />
}<br />
ACKNOWLEDGMENT<br />
This work is supported by the Ph.D. Program<br />
Foundation <strong>of</strong> Ministry <strong>of</strong> Education <strong>of</strong> China under Grant<br />
No. 200804971030, the Natural Science Foundation <strong>of</strong><br />
Hubei Province <strong>of</strong> China under Grant No. 2008CDB347,<br />
and the Fundamental Research Funds for the Central<br />
Universities <strong>of</strong> China under Grant No. 2010-Ia-049.<br />
© 2011 ACADEMY PUBLISHER<br />
REFERENCES<br />
[1] J. R. Gallardo, P. Medina, and W. Zhuang, “QoS<br />
mechanisms for the MAC protocol <strong>of</strong> IEEE 802.11<br />
WLANs,” Wireless <strong>Networks</strong>, vol. 13, no. 3, pp. 335-349,<br />
June 2007.<br />
[2] I. H. Hou and P. R. Kumar, “Admission control and<br />
scheduling for QoS guarantees for variable-bit-rate<br />
applications on wireless channels,” in Proc. <strong>of</strong> the 10th<br />
ACM International Symposium on Mobile Ad Hoc<br />
Networking and Computing, New Orleans, LA, USA, pp.<br />
175-184, May 2009.<br />
[3] J. K. Song, J. D. Ryoo, and S. C. Kim et al, “A dynamic<br />
GTS allocation algorithm in IEEE 802.15.4 for QoS<br />
guaranteed real-time applications,” in Proc. <strong>of</strong> the IEEE<br />
International Symposium on Consumer Electronics, pp. 1-6,<br />
June 2007.<br />
[4] J. J. Garcia and T. Falck, “Quality <strong>of</strong> service for IEEE<br />
802.15.4-based wireless body sensor networks,” in Proc. <strong>of</strong><br />
the 3rd International Conference on Pervasive Computing<br />
Technologies for Healthcare, pp. 1-6, April 2009.<br />
[5] I. F. Akyildiz, T. Melodia, and K. R. Chowdhury, “A survey<br />
on wireless multimedia sensor networks,” Computer<br />
<strong>Networks</strong>: the International <strong>Journal</strong> <strong>of</strong> Computer and<br />
Telecommunications Networking, vol. 51, no. 4, pp.<br />
921-960, 2007.<br />
[6] N. Dimokas, D. Katsaros, and Y. Manolopoulos,<br />
“Cooperative caching in wireless multimedia sensor<br />
networks,” Mobile <strong>Networks</strong> and Applications, vol. 13, no.<br />
3-4, pp. 337-356, August 2008.<br />
[7] F. Ingelrest and S. R. David, “Localized broadcast<br />
incremental power protocol for wireless ad hoc networks,”<br />
Wireless <strong>Networks</strong>, vol. 14, no. 3, pp. 309-319, June 2008.<br />
[8] A. Jadbabaie, “On geographic routing without location<br />
information,” in Proc. <strong>of</strong> the 43rd IEEE Conference on<br />
Decision and Control, vol. 5, pp. 4764-4769, Dec. 2004.<br />
[9] R. Fonseca, S. Ratnasamy, and J. Zhao et al, “Beacon vector<br />
routing: scalable point-to-point routing in wireless<br />
sensor-nets,” in Proc. <strong>of</strong> the 2nd conference on Symposium<br />
on Networked Systems Design & Implementation, vol. 2,<br />
pp. 329-342, 2005.<br />
[10] Q. Cao and T. Abdelzaher, “A scalable logical coordinates<br />
framework for routing in wireless sensor networks,” in<br />
Proc. <strong>of</strong> the 25th IEEE International Real-Time Systems,<br />
pp. 349-358, Dec. 2004.<br />
[11] T. He, J. Stankovic, and L. Chenyang et al, “SPEED: a<br />
stateless protocol for real-time communication in sensor<br />
networks,” in Proc. <strong>of</strong> the 23rd International Conference<br />
on Distributed Computing Systems, pp. 46-55, May 2003.<br />
[12] E. Felemban, C. G. Lee, and E. Ekici et al, “MMSPEED:<br />
multipath multi-SPEED protocol for QoS guarantee <strong>of</strong><br />
reliability and timeliness in wireless sensor networks,”<br />
IEEE Transactions on Mobile Computing, vol. 5, no. 6, pp.<br />
738-754, 2006.<br />
[13] L. Shu, Y. Zhang, and L. T. Yang et al, “Geographic<br />
routing in wireless multimedia sensor networks,” in Proc.<br />
<strong>of</strong> the Second International Conference on Future<br />
Generation Communication and Networking, vol. 1, pp.<br />
68-73, Dec. 2008.<br />
[14] L. Zhang, M. Hauswirth, and L. Shu et al, “Multi-priority<br />
multi-path selection for video streaming in wireless<br />
multimedia sensor networks,” in Proc. <strong>of</strong> the 5th<br />
International Conference on Ubiquitous Intelligence and<br />
Computing, vol. 5061, pp. 439-452, June 2008.<br />
[15] K. Akkaya and M. Younis, “Energy and QoS aware routing<br />
in wireless sensor networks,” Springer Cluster Computing<br />
<strong>Journal</strong>, vol. 8, no. 2-3, pp. 179-188, 2005.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 949<br />
[16] J. Chen, Y. Guan, and U. Pooch, “Customizing GPSR for<br />
wireless sensor networks,” in Proc. <strong>of</strong> the IEEE<br />
International Conference on Mobile Ad-hoc and Sensor<br />
Systems, pp. 549-551, Oct. 2004.<br />
[17] F. Tang, M. Guo, and M. Li et al, “Secure routing for<br />
wireless mesh sensor networks in pervasive<br />
environments,” International <strong>Journal</strong> <strong>of</strong> Intelligent Control<br />
and Systems, vol. 12, no. 4, pp. 293-306, 2007.<br />
[18] M. Busse, T. Haenselmann, and W. Effelsberg,<br />
“Energy-efficient forwarding in wireless sensor networks,”<br />
Pervasive and Mobile Computing, vol. 4, no. 1, pp.<br />
3-32, 2008.<br />
[19] I. S. Hwang and J. H. Wu, “Performance assessment <strong>of</strong><br />
service differentiation in IEEE 802.11e wireless LANs,”<br />
International <strong>Journal</strong> <strong>of</strong> Ad Hoc and Ubiquitous<br />
Computing, vol. 3, no.1, pp. 21-32, 2008.<br />
[20] R. Fantacci, G. Vannuccini, and G. Vestri, “Performance<br />
analysis <strong>of</strong> a multiple access protocol for voice and data<br />
support in multiuser broadband wireless LANs,” Wireless<br />
<strong>Networks</strong>, vol. 14, no. 1, pp. 17-28, Jan. 2008.<br />
[21] R. Nagpal, H. Shrobe, and J. Bachrach, “Organizing a<br />
global coordinate system from local information on an ad<br />
hoc sensor network,” in Proc. <strong>of</strong> the 2nd International<br />
Workshop on Information Processing in Sensor <strong>Networks</strong><br />
(IPSN '03), Palo Alto, CA, US, pp. 333-348, April 2003.<br />
[22] J. Bachrach, R. Nagpal, and M. Salib et al, “Experimental<br />
results and theoretical analysis <strong>of</strong> a self-organizing global<br />
coordinate system for ad hoc sensor networks,”<br />
Telecommunication Systems, vol. 26, no. 2-4, pp. 213-233,<br />
2004.<br />
[23] C. Intanagonwiwat, R. Govindan, and D. Estrin, “Directed<br />
diffusion: a scalable and robust communication paradigm<br />
for sensor networks,” in Proc. <strong>of</strong> the 6th Annual<br />
international Conference on Mobile Computing and<br />
Networking (MobiCom'00), Boston, Massachusetts, US, pp.<br />
56-67, Aug. 2000.<br />
[24] L. Khelladi and N. Badache, “On the performance <strong>of</strong><br />
directed diffusion in dense sensor networks,” in Proc. <strong>of</strong><br />
4th International Conference on Innovations in<br />
Information Technology, Dubai, pp. 113-117, Nov. 2007.<br />
[25] I. A. Qazi and T. Znati, “On the design <strong>of</strong> load factor based<br />
congestion control protocols for next-generation<br />
networks,” in Proc. <strong>of</strong> IEEE INFOCOM 2008 - The 27th<br />
Conference on Computer Communications, pp. 96-100,<br />
April 2008.<br />
Xin Yan: received his M.Sc. degree in<br />
electrical engineering from the Hubei<br />
University <strong>of</strong> Technology, China, in 1997,<br />
and his Ph.D. degree in computer science<br />
from the Wuhan University <strong>of</strong><br />
Technology, China, in 2006.<br />
He is an associate pr<strong>of</strong>essor at the<br />
Department <strong>of</strong> Computer Science, Wuhan<br />
University <strong>of</strong> Technology, China. He was a postdoctoral<br />
researcher at the Network Architectures and Services Group,<br />
Delft University <strong>of</strong> Technology, The Netherlands, from Jan.<br />
2009 to Jan. 2010. His main research interests lie in new<br />
Internet-like network architectures, and the modeling and<br />
performance analysis <strong>of</strong> network behavior and complex<br />
infrastructures.<br />
F. J. An: received her B.S. degree in electrical engineering from<br />
the Delft University <strong>of</strong> Technology, The Netherlands, in 2010.<br />
© 2011 ACADEMY PUBLISHER<br />
She is currently pursuing the M.Sc. degree in telecommunication<br />
electrical engineering at the Delft University <strong>of</strong> Technology, The<br />
Netherlands. Her research interests lie in mobile computing,<br />
wireless sensor networks, and ad hoc wireless networks.<br />
Layuan Li: received his B.S. degree from the Harbin Institute <strong>of</strong><br />
Military Engineering, China, in 1970 and his M.Sc. degree in<br />
communication and electrical systems from the Huazhong<br />
University <strong>of</strong> Science and Technology, China in 1982.<br />
He is a pr<strong>of</strong>essor and Ph.D. supervisor at the Department <strong>of</strong><br />
Computer Science, Wuhan University <strong>of</strong> Technology, China, and<br />
the Editor-in-Chief <strong>of</strong> the <strong>Journal</strong> <strong>of</strong> Wuhan University <strong>of</strong><br />
Technology. His research interests include high speed computer<br />
networks and protocol engineering. He received the National<br />
Special Prize by the Chinese government in 1993.
950 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Reputation-aware Service Selection based on<br />
QoS Similarity<br />
Shenghui Zhao 1,2 , Guoxin Wu 2<br />
1 Department <strong>of</strong> Computer Science and Technology, Chuzhou University, Chuzhou, China<br />
2 School <strong>of</strong> Computer Science & Engineering, Southeast University, Nanjing, China<br />
Email:zsh@chzu.edu.cn, gwu@seu.edu.cn<br />
Guilin Chen, Haibao Chen<br />
Department <strong>of</strong> Computer Science and Technology, Chuzhou University, Chuzhou, China<br />
Email:glchen@chzu.edu.cn, chb@chzu.edu.cn<br />
Abstract— For the up-and-coming computing models like as<br />
cloud computing, service is the standard package for<br />
meeting all kinds <strong>of</strong> consumers' requirements. Web Services<br />
are the concrete implement <strong>of</strong> the service. When users<br />
request and consume Web Services, services' reputations<br />
will play a vital role in users' selection. A gradually<br />
adjusting reputation evaluation method <strong>of</strong> Web Services is<br />
proposed based on eliminating the collusive behaviors <strong>of</strong><br />
consumers step by step, and a reputation-aware model for<br />
service selection is designed. In order to adjust reputations,<br />
QoS similarity is computed firstly according to the<br />
differences between advertised QoS from service providers<br />
and delivered QoS from service consumers' evaluation, next,<br />
current reputation is attained; then the consumers are<br />
sorted based on reputation using clustering algorithm and<br />
the potential collusive consumers are mined using<br />
association rules algorithm; finally, the updated reputation<br />
is recalculated and saved in the reputation center included<br />
in the model. The experimental results show that the model<br />
can identify the malicious consumers and improve the exact<br />
rate <strong>of</strong> reputation evaluation and success rate <strong>of</strong> service<br />
selection.<br />
Index Terms—Web Service, quality <strong>of</strong> service (QoS),<br />
reputation update, clustering algorithm, collusive<br />
consumers<br />
I. INTRODUCTION<br />
With the widespread <strong>of</strong> SOA, Web Services has<br />
become the main computing paradigm across Internet,<br />
new computing patterns are springing up such as cloud<br />
computing and CPS (Cyber Physical Systems) etc. A Web<br />
Service is a self-described and self-contained application<br />
that uses standard Internet technologies to interact with<br />
other Web Services, which can be published and accessed<br />
through the web. At present, many corporations and<br />
organizations have implemented their core application<br />
through buying the Web Services on Internet. For example,<br />
salesfore.com provides ERP service for users. Along with<br />
the maturation <strong>of</strong> service market, more and more service<br />
providers can provide the same or similar service, how to<br />
rationally select satisfied service has been turned into one<br />
<strong>of</strong> the key problems in Web Services research fields.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.950-957<br />
When service requestors select required services<br />
among many services with similar functionality, services'<br />
non-functional properties is an important considerable<br />
criterion, such as QoS (Quality <strong>of</strong> Service), reputation,<br />
etc. Generally speaking, QoS <strong>of</strong> Web Service is described<br />
by response time, reliability, availability, security and<br />
execution cost and so on. In the early service transactions,<br />
QoS information was published by service providers, but<br />
it was not always exact and up-to-date. For the interest <strong>of</strong><br />
ensuring the veracity <strong>of</strong> QoS properties, it should be a<br />
direct and valid method to appraise the QoS by requestors<br />
after invoking the Web Service. These values can be<br />
acted as the references for subsequent consumers to select<br />
the service. Many researches on service selection adopt<br />
this scheme.<br />
However, in the practical transactions, some feedbacks<br />
about QoS are falsity information due to the vicious<br />
estimation aiming at service providers. Thus, relying only<br />
on feedback estimation <strong>of</strong> QoS can not provide accurate<br />
methods for service selection. Reputation based service<br />
selection methods were proposed later, most <strong>of</strong> which<br />
were reputation evaluation on the basis <strong>of</strong> appreciable<br />
QoS after invoking a service, and then computed<br />
predicted reputation integrating multi historical values<br />
and current value. Above methods can wipe <strong>of</strong>f influence<br />
<strong>of</strong> little vicious users at a certain extent and improve the<br />
success rate <strong>of</strong> service selection. But, the community<br />
collusions may be occurred among consumers or among<br />
consumers and providers, services' reputation may be<br />
either lower or rose up which leads to distortion <strong>of</strong><br />
reputation.<br />
This paper discusses service selection based on<br />
reputation, in which distinguishes and filters out the<br />
collusive consumers through collusive behavior analysis<br />
methods. Then these collusive consumers' ratings are<br />
ignored, and decrease the influence <strong>of</strong> the malicious<br />
consumers, which can improve the veracity <strong>of</strong> reputation<br />
and the success rate <strong>of</strong> service selection. The rest <strong>of</strong> the<br />
paper is organized as follows. In section 2, we introduce<br />
related work. Section 3 proposes a method for Web<br />
Service reputation evaluation. A model for service
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 951<br />
selection is set up in Section 4. Next is experimental<br />
analysis. At last, we conclude the paper.<br />
II. RELATED WORK<br />
It is essential to acquire the QoS information when<br />
service selection is depending on QoS. [1] presented that<br />
both service selection and composition were QoS-aware;<br />
the QoS was measured by monitoring system according to<br />
service operations. An approach for measuring quality <strong>of</strong><br />
Web Services based on the superposition <strong>of</strong> uncertain<br />
factors was proposed, and a judging method for<br />
determining priorities among Web Services, which can<br />
help users select satisfied service[2]. A dynamic and QoSdriven<br />
model for service selection was proposed in [3],<br />
and the dynamic QoS data were computed according to<br />
users' feedback. In [4], the QoS attributes’ data obtained<br />
from service providers was revised, and feedback<br />
similarity came from service consumers was used to<br />
weight QoS data' trustworthiness, that strengthened the<br />
accuracy <strong>of</strong> the service selection. In [5], service-level<br />
agreements were discussed in order to set the penalties<br />
over the lack <strong>of</strong> QoS for web services. It ensured that the<br />
trustworthiness <strong>of</strong> a service-oriented environment relies<br />
on reliable QoS monitoring in certain sense.<br />
Although QoS based service selection is essential, due<br />
to the services' marketability, it is hard to avoid the<br />
dishonest service providers. So, service selection must be<br />
trust or reputation based [6], that can assure service's<br />
trustworthiness. In [7], Yao Wang et al. reviewed and<br />
concluded the service selection' criteria, and presented that<br />
it was necessary to implement service selection depending<br />
on trust and reputation. The authors in [8] suggested a<br />
framework <strong>of</strong> service selection based on reputation in a<br />
semantic network. The reputation was computed by<br />
different service consumers. In [9], Malik et al. had<br />
proposed a model to compute the reputation <strong>of</strong> a web<br />
service in accordance with the personal evaluation <strong>of</strong> the<br />
previous users. The characteristic <strong>of</strong> this method was the<br />
credibility <strong>of</strong> the users <strong>of</strong> evaluating services has been<br />
taken into account. If the rater tried to provide a fake<br />
rating, then its credibility would be decreased and the<br />
rating <strong>of</strong> this user would become less important in the<br />
reputation <strong>of</strong> the web service.<br />
Obviously, QoS can help consumers select the service<br />
with high quality, and reputation has been used to make<br />
consumers select the service providers which honestly<br />
<strong>of</strong>fer the service with advertised QoS. Making use <strong>of</strong><br />
reputation, consumers can find or select secure, reliable<br />
and trusted Web Services. So, the service quality's<br />
reputation is vital important to select the genuine service<br />
required by the consumers. In [10], Maximilien and Singh<br />
designed a multi-agent framework based on ontology for<br />
QoS. The users’ ratings which depended on the different<br />
qualities satisfied varied consumers' trust requirement<br />
used for computing the reputation <strong>of</strong> the web service, and<br />
it would be the selection criterion. That was dynamic<br />
selection.<br />
Ping Wang et al.[11] expressed an idea that aggregating<br />
previous assessment records (bodies <strong>of</strong> evidence) via<br />
consumers' feedbacks and witness <strong>of</strong> network referrals to<br />
© 2011 ACADEMY PUBLISHER<br />
derive a more objective reputation score on the specific<br />
service. Then two factors was defined, confidence degree<br />
and support degree based on evidence theory, to enhance<br />
the discrimination <strong>of</strong> the quality <strong>of</strong> existing evidence to<br />
help providers avoid malicious assessment. In [12],<br />
service providers' reputations were figured out through<br />
applying the current reputation and historical data with<br />
various weights. According to providers' reputation and<br />
services' reputation, a method for measuring service<br />
providers' trust was proposed. By ranking the trust value,<br />
consumers can select more trusted service. In [13], the<br />
authors developed a framework aiming to select Web<br />
Services based on the trust policy expressed by the users.<br />
The framework allowed the users to select a web service<br />
matching their needs and expectations.<br />
In above-mentioned literatures about service selection,<br />
some only proposed the models <strong>of</strong> applying reputation,<br />
and some gave the rating methods on reputation, but they<br />
lacked <strong>of</strong> controlling the situations <strong>of</strong> providing falsity<br />
reputation information. That is, there are only few<br />
researches on consumers' collusion and deception.<br />
Although some researches [14][15] have considered<br />
collusion among consumers, their analysis objects and<br />
methods are different from our paper. Moreover, most<br />
literatures didn't give solutions on integrality <strong>of</strong><br />
reputation data.<br />
For the sake <strong>of</strong> solving above problems, in our<br />
reputation-aware service selection model, the reputation is<br />
computed based on the similarity <strong>of</strong> service quality's<br />
attributes, and a method for collusion behavior analysis is<br />
proposed. Besides, while constructing a system reputation<br />
model, we take into account reputation storage and<br />
security, which can provide a relative secure and trusted<br />
service selection scheme for consumers.<br />
III. A METHOD FOR RATING WEB SERVICE<br />
The evaluation method is based on following<br />
suppositions:<br />
1) One service provider <strong>of</strong>fers one service only, and the<br />
provider's reputation can be apperceived by its' service<br />
reputation.<br />
2) The reputation published to UDDI by service<br />
providers is authentic.<br />
3) The reputation center is trustable. It can be acted as a<br />
broker for service consumers and providers' behaviors.<br />
4) If one service consumer selects a service, it must<br />
trust in the service provider.<br />
5) If the consumer is honesty, its evaluation is honesty<br />
too.<br />
A. Computation <strong>of</strong> Web Service's Reputation.<br />
Supposing that a Web Service s j has m attributes <strong>of</strong><br />
QoS, what is expressed as ( q1, q2,..., q m)<br />
. For any user<br />
u i , its invocation to s j 's attributes is represented as<br />
ij ij ij<br />
( q1 , q2,..., q m)<br />
. The advertised QoS <strong>of</strong> s j is shown as<br />
j j j<br />
( Ad_ q , Ad_ q ,..., Ad_ q ) . After s j being invoked by<br />
1 2<br />
m
952 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
u i , its feedback rate on s j 's QoS is denoted as<br />
( Eval _ q , Eval _ q ,..., Eval _ q ) .<br />
ij ij ij<br />
1 2<br />
m<br />
Definition 1 Similarity <strong>of</strong> service s j 's quality<br />
j<br />
Sim q<br />
After current user i u invoking service j s , u i will give<br />
a new value on QoS' attributes. The similarity degree<br />
Sim can be figured out in (1).<br />
j<br />
q<br />
Sim<br />
= 1−<br />
m<br />
∑<br />
j k = 1<br />
q<br />
ij j<br />
( Eval _q−Ad_ q )<br />
k k<br />
m is the number <strong>of</strong> QoS attributes.<br />
Definition 2 Service s j ' s current reputation<br />
m<br />
2<br />
rep<br />
j<br />
cur<br />
(1)<br />
s j 's current reputation is the newest reputation after<br />
j<br />
this invoking. repcur is computed using (2) which is<br />
figured as Fig. 1.<br />
j j<br />
rep = 1−sinh(1 − Sim ) . (2)<br />
cur q<br />
Thereinto,sinh( x) = (exp( x) −exp( − x))<br />
/ 2 .<br />
Figure 1. Result <strong>of</strong> Function sinh(x)<br />
It can be seen from Fig. 1, when advertised QoS is<br />
closer to received QoS, the higher reputation is gained.<br />
j<br />
j<br />
Especially, if Simq is equal to 1, repcur can achieve<br />
j<br />
j<br />
highest value 1. When Sim is less than 0.2, rep q<br />
cur is<br />
nearly 0. This is very similar to the actual application<br />
occasions. So we adopt the function in formula (2).<br />
But current reputation had something to do with rater's<br />
credibility degree. If rater is trustful user, its reputation is<br />
trustworthy and can be contained in the global reputation.<br />
Otherwise, its reputation is not trustful and will be ignored<br />
in the global reputation.<br />
B. Reputation Update and Adjustment<br />
When j s is invoked at each time, s j 's current<br />
j<br />
reputation repcur can be computed using formula (2).<br />
However, for other users, they not only use the last result,<br />
but also consult historical data. A reputation center being<br />
involved in the service selection model is designed in<br />
© 2011 ACADEMY PUBLISHER<br />
Section 4. Reputation center plays the role <strong>of</strong> computing<br />
and keeping each service's global reputation. Hence, after<br />
each time invoking s j , it should update the stored s j 's<br />
global reputation.<br />
Assuming parameter δ is the difference between<br />
received QoS and advertised QoS, that is,<br />
m<br />
∑<br />
ij j<br />
δ = ( Eval _q−Ad_ q )/ m .If δ is larger than<br />
i=<br />
1<br />
k k<br />
0, it can be explained that received QoS is prior to<br />
advertised QoS. Then the reputation will be increased<br />
apparently; otherwise it will be decreased.<br />
j<br />
Definition 3 Service s j 's global reputation rep glo .<br />
j<br />
s j 's global reputation repglo is associated with all<br />
historical reputations and current reputation. After each<br />
invoking,<br />
rep and historical data are updated and<br />
j<br />
cur<br />
j<br />
rep glo is adjusted using formula (3).<br />
j j<br />
rep = rep × (0.9 + 0.1× exp( δ )) (3)<br />
glo upd<br />
upd<br />
n<br />
∑ i cur<br />
n<br />
i / ∑ i , i<br />
i= 1 i=<br />
1<br />
j j di<br />
In (3), rep = rep _ × γ γ γ = λ<br />
j<br />
rep is the ith current reputation <strong>of</strong> service<br />
i_cur s j computed by formula (2). γ is the aging factor for the<br />
i<br />
ith service reputation, λ ∈ [ 0,1]<br />
. A smaller λ means only<br />
recent reputations are included and a larger λ means more<br />
reputations are included. d i is the time interval <strong>of</strong><br />
between last rating time and current time. For instance,<br />
when using current reputation, d i almost equals to 0, so<br />
γ is 1. That is, the current reputation has not aged.<br />
i<br />
C. Reputation Storage<br />
In general, there are three places to store reputation<br />
information, rater (service consumer), ratee(service<br />
providers) and a third party(reputation center). This paper<br />
stores the information in ratee and a third party,<br />
respectively. The advantage <strong>of</strong> saving the global<br />
reputation in ratees is that consumers can find the satisfied<br />
service's reputation from providers at the time <strong>of</strong><br />
reputation center collapsing. Due to computing global<br />
reputation in reputation center, each service's rating<br />
information should be saved. A database is created in<br />
reputation center to save rating information. The storage<br />
format includes five items, listed in Table I.<br />
Rater ID<br />
(UID)<br />
TABLE I. STORAGE FORMAT OF REPUTATION<br />
Service ID<br />
(SID)<br />
Reputation<br />
Value with<br />
encrypted<br />
Time<br />
Stamp<br />
Transaction<br />
Number<br />
(TID)<br />
Rate0001 Serice001 Repkey TM1 Tran0001<br />
When a consumer finds the satisfied service providers,<br />
he will query the providers' reputation in either reputation
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 953<br />
center or providers selves. In order to prevent the<br />
providers to tamper with the saved reputation, the<br />
reputations will be handled in particular. The method is<br />
adopting the RcertX[16] idea to save the reputation<br />
information, where each service's reputation is put in one<br />
certificate, so all reputations are linked as a body to avoid<br />
the provider juggling with the reputation. After each<br />
global reputation is updated, it will be signed digital and<br />
stored in reputation center and sent to ratee. When users<br />
take out the reputations from the providers, they will use<br />
public key published by reputation center to decrypt the<br />
messages. Of course, public key has been already<br />
delivered to all consumers by broadcast, and it can be<br />
transmitted among consumers too.<br />
D. Algorithm for collusion behavior analysis<br />
After service transactions, honest users will give<br />
fairness and objectivity valuation. But the malicious<br />
consumers will give unpractical estimation, like as rising<br />
up or playing down the reputation intentionally, as well as<br />
collusive users. If the proportion <strong>of</strong> collusive users is<br />
higher, it will affect the service providers' reputation more<br />
greatly. Therefore, reputation model should be capable <strong>of</strong><br />
identifying the collusive users and reduce the negative<br />
influence.<br />
For the sake <strong>of</strong> finding the collusive consumers, we<br />
design an algorithm named as CBA (Collusive Behavior<br />
Analysis) which has two main operations: classification<br />
and mining. We will use k-means cluster algorithm [17] to<br />
create multi clusters and find the correlations among<br />
consumers adopting the association rule mining algorithm.<br />
The CBA algorithm steps are as follows:<br />
Step 1 Take out all records <strong>of</strong> service s j from the<br />
database in reputation center, and sort the reputation<br />
values into three clusters (marked as 0,1,2) using k-means<br />
cluster algorithm. The 0th cluster denotes the class with<br />
the lowest reputation, and the 2th cluster represents the<br />
class with the highest reputation. Average the reputation<br />
distributed in the 1th cluster, and makes the mean as the<br />
references to analyze the collusion.<br />
Step 2 Owing to the reputations in the class <strong>of</strong> 0th<br />
clusters are all lower generally, which express that all<br />
assessment are on the low side to service s j . The reason<br />
has two sides, on the one hand, the service s j is bad<br />
service originally; on the other hand, it may be the<br />
collusive consumers purposely debasing s j ' s reputation.<br />
For differentiating them, taking out some others service<br />
from all services except for s j , the extraction proportion<br />
is 33%. Then, each service records construct their data<br />
collection which will be analyzed applying the first step,<br />
respectively. Moreover, take out users included in 0th<br />
cluster deriving from each service as a list. This user list is<br />
merged into a new list with service s j ' s 0th cluster user<br />
list.<br />
Step 3 Applying the association rules algorithm, on<br />
given support degree, the largest frequent items can be<br />
© 2011 ACADEMY PUBLISHER<br />
mined, in which there are different users.<br />
Step 4 The different users in the largest frequent items<br />
may be looked as collusive community. So, if the user<br />
exists in the community, it will be marked as dishonest<br />
user and set hi=0; otherwise, set hi=1.<br />
j<br />
After detecting the collusive users, recalculate rep , cur<br />
j j<br />
viz. repcur = hi × rep . So, when hi=0, the current<br />
cur<br />
reputation is zero, too.<br />
For example, there are two services, s1 and s2. Their 30<br />
transaction records are listed as Table II.<br />
TABLE II. TRANSACTION RECORD<br />
TID UID SID RepValue TID UID SID RepValue<br />
1 u1 s1 0.8 16 u6 s1 0.66<br />
2 u2 s2 0.5 17 u7 s2 0.83<br />
3 u2 s1 0.6 18 u1 s1 0.89<br />
4 u2 s1 0.65 19 u8 s2 0.83<br />
5 u3 s2 0.55 20 u9 s1 0.91<br />
6 u1 s2 0.8 21 u6 s2 0.95<br />
7 u2 s2 0.85 22 u5 s1 0.78<br />
8 u1 s1 0.86 23 u4 s2 0.68<br />
9 u3 s1 0.2 24 u5 s1 0.88<br />
0<br />
1<br />
u4 s1 0.6 25 u6 s2 0.95<br />
1<br />
1<br />
u5 s1 0.75 26 u6 s2 0.75<br />
2<br />
1<br />
u4 s2 0.7 27 u1 s1 0.9<br />
3<br />
1<br />
u5 s2 0.55 28 u4 s2 0.85<br />
4<br />
1<br />
u5 s2 0.6 29 u3 s1 0.58<br />
5<br />
1<br />
u4 s2 0.7 30 u1 s2 0.79<br />
Based on the above algorithm and using k-means<br />
cluster algorithm, we extract all s1 records and s2 records<br />
and divide them into three clusters, respectively. The<br />
results are shown in Table III and Table IV.<br />
TABLE III. DIVIDE S1 RECORDS INTO THREE CLUSTERS.<br />
---------Cluster0---------<br />
00029[0.58,u3,s1]<br />
00010[0.6,u4,s1]<br />
00003[0.6,u2,s1]<br />
00009[0.2,u3,s1]<br />
---------Cluster1---------<br />
00001[0.8,u1,s1]<br />
00022[0.78,u5,s1]<br />
00011[0.75,u5,s1]<br />
00016[0.66,u6,s1]<br />
00004[0.65,u2,s1]<br />
-----------Cluster2---------<br />
00018[0.89,u1,s1]<br />
00024[0.88,u5,s1]<br />
00027[0.9,u1,s1]<br />
00008[0.86,u1,s1]<br />
00020[0.91,u9,s1]
954 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
TABLE IV. DIVIDE S2 RECORDS INTO THREE CLUSTERS<br />
---------Cluster0---------<br />
00002[0.5,u2,s2]<br />
00005[0.55,u3,s2]<br />
00014[0.6,u5,s2]<br />
00013[0.55,u5,s2]<br />
-----------Cluster1---------<br />
00030[0.79,u1,s2]<br />
00015[0.7,u4,s2]<br />
00006[0.8,u1,s2]<br />
00026[0.75,u6,s2]<br />
00023[0.68,u4,s2]<br />
00012[0.7,u4,s2]<br />
-----------Cluster2---------<br />
00007[0.85,u2,s2]<br />
00019[0.83,u8,s2]<br />
00025[0.95,u6,s2]<br />
00028[0.85,u4,s2]<br />
00017[0.83,u7,s2]<br />
From Table III and Table IV., we can see, users' list<br />
(u3,u4,u2) and (u2,u3,u5) derived from cluster0 are made<br />
up <strong>of</strong> two lines. If there are more services, they will<br />
3, 4, 2<br />
formed as ⎧u u u ⎫<br />
⎪ ⎪ in which collusive consumers could<br />
⎨u2, u3, u5⎬<br />
⎪...... ⎪<br />
⎩ ⎭<br />
be mined using association rule mining algorithm.<br />
IV. SERVICE SELECTION OF REPUTATION- AWAREE<br />
Service selection model based on reputation is designed<br />
and shown as Fig.2. It includes two roles: service<br />
providers, service requestors, and a data center: UDDI, as<br />
well as three agents: discovery agent, selection agent and<br />
rating agent. Discovery agent helps service consumers to<br />
find services meeting requirements, and selection agent<br />
selects the service with highest reputation for consumers<br />
from the satisfied services. The purpose <strong>of</strong> rating agent is<br />
to evaluate the newest reputation <strong>of</strong> service providers and<br />
update the global reputation.<br />
Figure 2. Model <strong>of</strong> Service Selection<br />
Satisfied<br />
Services<br />
The process <strong>of</strong> service selection is as follows.<br />
1) If a service provider joins the system, it will publish<br />
its advertised QoS to UDDI acquiescently.<br />
© 2011 ACADEMY PUBLISHER<br />
2) Service requestors (viz. consumers) send their<br />
requirements which contain functional and nonfunctional<br />
descriptions to discovery agent.<br />
3) Discovery agent queries the UDDI. If UDDI has the<br />
satisfying services, it returns the results to discovery agent.<br />
4) Discovery agent submits the results to selection<br />
agent.<br />
5) Selection agent will query the reputations <strong>of</strong> all<br />
matched services to reputation center.<br />
6) Selection agent sorts the reputations <strong>of</strong> all satisfied<br />
services, and returns service with the highest reputation to<br />
requestor.<br />
7) Service provider and requestor carry out their<br />
transaction.<br />
8) After using service, consumer evaluates the aware<br />
quality <strong>of</strong> service, and sends information back to<br />
reputation center.<br />
9) Rating agent sends request to UDDI for promised<br />
QoS <strong>of</strong> service provider. Then updates the service' global<br />
reputation according to formula (2) and (3).<br />
10) Reputation center makes digital signature for the<br />
reputation and feed it back to the provider.<br />
CBA algorithm will be executed by reputation center<br />
after a period <strong>of</strong> time. The results can be used in 9).<br />
V. EXPERIMENTATION<br />
In order to validate our reputation model, we develop a<br />
simulation program written in Java. It simulates multi<br />
service provider and consumers' transaction behavior. The<br />
transaction results are saved like as the format listed in<br />
Table I. For the purpose <strong>of</strong> finding the collusive users,<br />
CBA algorithm is employed to make clustering and mine<br />
data for the transaction records at regular intervals. As<br />
time goes on, collusive consumers will gradually be<br />
steady. So the diminished reputations also tend towards<br />
stability. In experiments, the records being used are in a<br />
fixed time <strong>of</strong> transactions.<br />
A. Experimental Environment<br />
In our simulation environment, the number <strong>of</strong> service<br />
providers is 100, and the number <strong>of</strong> service consumers is<br />
60. For the convenience <strong>of</strong> experiment, the kind <strong>of</strong><br />
services <strong>of</strong>fered by providers is 1. Each service's QoS<br />
includes availability, response time, reliability, security<br />
and cost. Each service's attributes has four levels: bad,<br />
ordinary, good, and excellent. We assumed that there are<br />
10% excellent, 20% bad, 30% good and 40% ordinary<br />
among the 100 providers.<br />
However the service level is, the collusive consumers<br />
will give bad evaluation, and honest consumers will give<br />
authentic estimation. The appraisal is like as Table V.<br />
service<br />
level<br />
TABLE V. ESTIMATION RANGE FROM CONSUMERS<br />
proportion<br />
<strong>of</strong> all<br />
providers<br />
default<br />
<strong>of</strong> the<br />
attributes<br />
Estimation<br />
Range from<br />
Collusive<br />
Consumers<br />
Estimation<br />
Range from<br />
Honest<br />
Consumers<br />
Excellent 10% 1 0.3~0.5 0.9~1<br />
Good 30% 0.85 0.3~0.5 0.7~0.9<br />
Ordinary 40% 0.7 0.3~0.5 0.5~0.7<br />
Bad 20% 0.5 0.3~0.5 0.3~0.5
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 955<br />
At the beginning <strong>of</strong> experiment, assumed all users are<br />
honest and all services' initial reputation are 0.6.The value<br />
<strong>of</strong> execution transaction is 0.6, meaning that if reputation<br />
is greater than or equal to 0.6, a transaction will be<br />
executed. At the same time, the successful transaction<br />
refers to the reputation greater than a threshold. Given<br />
threshold 0.7, viz. if the estimation value is no less than<br />
0.7, the transaction is successful. We will take success rate<br />
<strong>of</strong> service as a validity measurement <strong>of</strong> reputation model.<br />
Success rate <strong>of</strong> service is defined as the ratio <strong>of</strong> the<br />
number <strong>of</strong> times to the number <strong>of</strong> total transactions. When<br />
finding maximum frequent items, we set the support<br />
degree 0.9.<br />
B. Experimental Data<br />
In order to examine the validity and veracity <strong>of</strong><br />
reputation computing in service selection model, we use<br />
rate <strong>of</strong> service success and exact rate <strong>of</strong> reputation to<br />
measure, respectively. Comparison is implemented in the<br />
two situations <strong>of</strong> considering collusion and no considering<br />
it.<br />
(1) Comparison <strong>of</strong> Success Rate<br />
At first, we study the success rate when the ratio <strong>of</strong><br />
collusion (RoC) varies from 10% to 40% increased by<br />
10%. The results <strong>of</strong> considering collusion and no<br />
considering collusion are shown in Fig.3 and Fig.4.<br />
Figure 3. Success Rate <strong>of</strong> Considering Collusion under different RoC<br />
Figure 4. Success Rate <strong>of</strong> No Considering Collusion under different<br />
RoC<br />
From Fig.3 and Fig.4, the success rate reduces<br />
evidently as the RoC increasing, it can be explained that<br />
© 2011 ACADEMY PUBLISHER<br />
along with the number <strong>of</strong> round grew higher, success rate<br />
goes to steady. When the RoC is 10%, the both success<br />
rates are approaching to 90%, what can be explained the<br />
effect <strong>of</strong> collusion is not distinctness. But, when the RoC<br />
arrives at 40%, the success rate <strong>of</strong> considering collusion is<br />
higher than no considering collusion. Obviously,<br />
considering collusion has better effect for improving<br />
success rate. If the ratio <strong>of</strong> collusive users is bigger, it will<br />
have a larger effect on the service providers' reputation.<br />
TABLE VI. PROPORTION OF DIFFERENT SERVICE LEVEL<br />
service level Proportion <strong>of</strong> service providers<br />
Excellent 10% 20% 30% 40% 50%<br />
Good 20% 20% 20% 20% 20%<br />
Ordinary 50% 40% 30% 20% 10%<br />
Bad 20% 20% 20% 20% 20%<br />
Secondly, we compare the success rate <strong>of</strong> different ratio<br />
<strong>of</strong> excellent services under the circumstances <strong>of</strong> 20%<br />
RoC. The ratio <strong>of</strong> excellent services (RoE) is listed in<br />
Table VI. When the ratio <strong>of</strong> excellent service changes<br />
from 10% to 50%, the rates <strong>of</strong> success are displayed in<br />
Fig.5.<br />
Figure 5. Success Rate <strong>of</strong> Different Ratio <strong>of</strong> Excellent Services<br />
From Fig.5, we can see, no matter how change the RoE<br />
is, success rate has no great effect in certain collusive<br />
consumers. The phenomenon is consistent with the fact,<br />
because <strong>of</strong> the success rate only be affected on the number<br />
<strong>of</strong> collusive consumers, as well as the ratio <strong>of</strong> collusion is<br />
fixed.<br />
(2) Exact Rate <strong>of</strong> Reputation<br />
A service s j ' exact rate is represented as:<br />
j j<br />
exactRate = 1 − repcur / rep . glo<br />
The experimental results are shown in Fig.6. Each<br />
value is the average <strong>of</strong> exact rates in each round. With the<br />
increase <strong>of</strong> rounds, about 500 rounds later, the exact rate<br />
goes to stable. The exact rates <strong>of</strong> considering collusion<br />
and no considering collusion are about 92% and 72%,<br />
respectively. Above result illustrates our reputation model<br />
considering collusive behavior can eliminate the influence<br />
<strong>of</strong> collusion effectively.
956 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Figure 6. Exact Rate <strong>of</strong> Reputation <strong>of</strong> Considering and no Considering<br />
collusion<br />
(3) Only considering QoS<br />
In order to differentiate reputation-aware selection<br />
from QoS-aware selection, we set the experiment <strong>of</strong> the<br />
QoS-aware service selection only relying on the<br />
evaluated QoS. Table V is hold good. The threshold <strong>of</strong><br />
QoS is set to 0.6, that is, if the evaluated QoS is 0.6, the<br />
transaction succeeds. The colluding consumers are filtered<br />
out using our CBA algorithm and total QoS value but not<br />
global reputation is computed. Total QoS value is the<br />
average <strong>of</strong> received QoS. The service selection is based on<br />
the ranking QoS, obviously, the service with highest QoS<br />
will be selected.<br />
Figure 7. Service Selection only based on QoS<br />
The result is shown as Fig.7. Compared with Fig.3, the<br />
rate <strong>of</strong> success is lower than 15% in general. That<br />
illustrates service selection <strong>of</strong> reputation-aware is superior<br />
to QoS-aware'.<br />
VI. CONCLUSION<br />
This paper introduces a reputation model considering<br />
collusive consumers. We get the current reputation by<br />
utilizing the similarity between advertised QoS from<br />
service providers and delivered QoS from consumer’s<br />
evaluation, then update the global reputation and save<br />
them into reputation center and service providers. At the<br />
same time, in order to prevent providers tampering with<br />
reputation, we use the digital signature. Experimental<br />
results show that the success rate <strong>of</strong> transaction<br />
© 2011 ACADEMY PUBLISHER<br />
considering collusion is higher than no considering<br />
collusion.<br />
In order to find collusive consumers, we make use <strong>of</strong><br />
k-means cluster algorithm to classify the consumers, and<br />
use association rule algorithm to mine collusive<br />
consumers, then adjust the service reputation through<br />
eliminating the collusive consumers gradually. In fact,<br />
collusion exists not only in consumers, but also exists<br />
between consumers and providers. Because <strong>of</strong> pr<strong>of</strong>its,<br />
they have enthusiasm to make collusion. How to reduce<br />
the second collusion is another challenge.<br />
ACKNOWLEDGMENT<br />
This research was partially supported by Anhui<br />
Educational Department Natural Sciences Project<br />
(KJ2010A251).<br />
REFERENCES<br />
[1] Liangzhao Zeng, Hui Lei, and Henry Chang. "Monitoring<br />
the QoS for Web Services," Proceedings <strong>of</strong> the 5th<br />
international conference on Service-Oriented<br />
Computing(ICSOC 2007). 2007, LNCS 4749, pp.132-144.<br />
[2] Liu Yue, Weiyi Liu, Xiaoling Wang,Jin Li. "An Approach<br />
for Measuring Quality <strong>of</strong> Web Services Based on the<br />
Superposition <strong>of</strong> Uncertain Factors," <strong>Journal</strong> <strong>of</strong> Computer<br />
Research and Develeopment. vol.46, no.5, 2009, pp.841-<br />
849.<br />
[3] Liu Y, Ngu A, Zengl Z. "QoS computation and policing in<br />
dynamic Web service selection," Proceedings <strong>of</strong> t he 13th<br />
International World Wide Web Conference. New York ,<br />
USA : ACM Press , 2004, pp.66-73.<br />
[4] Yan Li,Minghui Zhou, Ruichao Li, Donggang Cao,Hong<br />
Mei. "Service Selection Approach Considering the<br />
Trustworthiness <strong>of</strong> QoS Data," <strong>Journal</strong> <strong>of</strong><br />
S<strong>of</strong>twate.Vol.19,no.10, 2008, pp. 2620-2627.<br />
[5] R. Jurca, B. Faltings, and W. Binder. "Reliable QoS<br />
monitoring based on client feedback," Proceedings <strong>of</strong> the<br />
16th International World Wide Web Conference<br />
(WWW07), 2007, pp. 1003-1011.<br />
[6] T. Sobh and K. Elleithy. "Service Selection Should be<br />
Trust and Reputation-Based," Advances in Systems,<br />
Computing Sciences and S<strong>of</strong>tware Engineering, 2006,<br />
pp.359-364.<br />
[7] Yao Wang, Julita Vassileva ."Toward Trust and Reputation<br />
Based Web Service Selection: A Survey," In International<br />
Transactions on Systems Science and Applications, 2007,<br />
pp.118-132.<br />
[8] Ali Shaikh Ali, Shalil Majithia, Omer F. Rana and et al.<br />
"Reputation-based semantic service discovery,"<br />
Concurrency and Computation: Practice & Experience -<br />
First International Workshop on Emerging Technologies<br />
for Next-generation GRID (ETNGRID 2004) . vol.18,<br />
issue 8, 2006, pp.817-826.<br />
[9] Zaki Malik, Athman Bouguettaya. "RATEWeb: Reputation<br />
Assessment for Trust Establishment among Web Services,"<br />
The VLDB <strong>Journal</strong>. vol.18,no.4, 2009, pp.885-911,doi:<br />
10.1007/s00778-009-0138-1.<br />
[10] E. M. Maximilien and M. P. Singh. "Multiagent System for<br />
Dynamic Web Services Selection," Proceeding <strong>of</strong> first<br />
Workshop on Service-Oriented Computing and Agent-<br />
Based Engineering (SOCABE at AAMAS), 2005, pp.25-<br />
29.<br />
[11] Ping Wang, Kuo-Ming Chao, Chi-Chun Lo, Ray Farmer,<br />
Pu-Tsun Kuo. "A Reputation-based Service Selection
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 957<br />
Scheme," 2009 IEEE International Conference on e-<br />
Business Engineering. IEEE Computer Society,2009,<br />
pp.501-506.<br />
[12] Keting Yin, Bo Zhou, Shuai Zhang, Yixi Chen,et al. "An<br />
effective approach to select trustable web services,"<br />
WiCOM '08. 4th International Conference on Wireless<br />
Communications, Networking and Mobile Computing,<br />
2008, pp.1-6.<br />
[13] A.S. Ali, S.A. Ludwig, and O.F. Rana. "A cognitive trustbased<br />
approach for web service discovery and selection,"<br />
Proceeding <strong>of</strong> the third European Conferenc. on Web<br />
Services,2005, pp. 38-40.<br />
[14] Wanita Sherchan, Seng W. Loke and Shonali<br />
Krishnaswamy. "Explanation-aware service selection:<br />
rationale and reputation.Service," Oriented Computing and<br />
Applications, vol.2, no.4, 2008,pp.203–218,doi:<br />
10.1007/s11761-008-0032-5.<br />
[15] Babak Khosravifar, Jamal Bentahar, Philippe Thiran,<br />
Ahmad Moazin, and Adrien Guiot."An Approach to<br />
Incentive-based Reputation for Communities <strong>of</strong> Web<br />
Services," 2009 IEEE International Conference on Web<br />
Services, 2009, pp.303-310.<br />
[16] Beng Chin Ooi, Chu Yee Liau and Kian-Lee Tan.<br />
"Managing trust in peer-to-peer systems using reputationbased<br />
techniques," LNCS 2003, Volume 2762/2003, pp.2-<br />
12, doi: 10.1007/978-3-540-45160-0_2.<br />
© 2011 ACADEMY PUBLISHER<br />
[17] MacQueen, J.B. "Some methods for classification and<br />
analysis <strong>of</strong> multivariate observations," Proceedings <strong>of</strong> 5th<br />
Berkeley Symposium on Mathematical Statistics and<br />
Probability, University <strong>of</strong> California Press, Berkeley, 1967,<br />
pp. 281-297.<br />
Shenghui Zhao Ph.D candidate <strong>of</strong> Southeast University,<br />
associate Pr<strong>of</strong>essor <strong>of</strong> Chuzhou University. Her major research<br />
fields include network security, Web Services and distributed<br />
computing.<br />
Guoxin Wu Pr<strong>of</strong>essor <strong>of</strong> Southeast University, doctoral<br />
supervisor. His major research fields include trust network,<br />
distributed computing and network security.<br />
Guilin Chen Pr<strong>of</strong>essor <strong>of</strong> Chuzhou University. His research<br />
interests include distributed computing, pervasive computing<br />
and virtualization.<br />
Haibao Chen Teacher <strong>of</strong> Chuzhou University. His major<br />
research fields include trust and Cloud computing.
958 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Cost Aggregation Strategy with Bilateral Filter<br />
Based on Multi-scale Nonlinear Structure Tensor<br />
Li Li<br />
Shandong Provincial Key Laboratory <strong>of</strong> Digital Media Technology, Shandong Economic University, Jinan, China<br />
School <strong>of</strong> Computer Science and Technology, Shandong University, Jinan, China<br />
Email :lily_jn @ sina.com<br />
Hua Yan<br />
Shandong Provincial Key Laboratory <strong>of</strong> Digital Media Technology, Shandong Economic University, Jinan, China<br />
Email: yhzhjg @ sdu.edu.cn<br />
Abstract—This paper proposed a novel cost aggregation<br />
method for stereo matching with modified bilateral filter. In<br />
original bilateral filter, only spatial and range weights are<br />
used to compute the similarity <strong>of</strong> two considering pixels and<br />
a new weight based on structure tensor is added in our<br />
method. By smoothing each element <strong>of</strong> the structure tensor<br />
considering both the spatial and gradient distances <strong>of</strong><br />
neighboring pixels, the nonlinear structure tensor for each<br />
pixel is constructed. We adopt the Log-Euclidean calculus as<br />
tensor dissimilarity function to compute the structure tensor<br />
distance <strong>of</strong> two considering pixels. Then the multi-scale<br />
value is computed by summing <strong>of</strong> the tensor distances in<br />
each scale. So a new weight based on multi-scale structure<br />
tensor distance is set up and included in bilateral filter for<br />
cost aggregation. By constructing the multi-scale nonlinear<br />
structure tensor and adding the new corresponding weight<br />
in cost aggregation, more pixels similar with central pixel<br />
could be aggregated in a support window and the final<br />
disparity map could be more accurate. Experimental results<br />
have confirmed the effectiveness <strong>of</strong> our proposed method.<br />
Index Terms—stereo matching, cost aggregation, multi-scale<br />
nonlinear structure tensor, Log-Euclidean tensor distance,<br />
bilateral filter<br />
I. INTRODUCTION<br />
Stereo matching is a key problem in computer vision.<br />
According to authors [1] stereo matching algorithms<br />
usually perform four steps: cost computation, cost<br />
aggregation, disparity computation or optimization and<br />
disparity refinement. Cost aggregation is mandatory for<br />
local stereo matching algorithms to improve signal noise<br />
rate (SNR) and <strong>of</strong>ten adopted by global ones. The cost<br />
aggregation step is to aggregate initial matching costs in a<br />
support window. An ideal support window should be<br />
adjusted according to image content to include only the<br />
pixels with the same disparity. Many cost aggregation<br />
methods have been presented while this behavior is far<br />
from ideal. This paper proposed a novel cost aggregation<br />
strategy with modified bilateral filter based on multi-scale<br />
nonlinear structure tensors.<br />
Many adaptive window methods have been presented<br />
to include more pixels having the same disparity values<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.958-965<br />
with the central pixel by varying the size, the shape and<br />
the position <strong>of</strong> the support window [2-4]. Different from<br />
the adaptive window methods, the adaptive weight<br />
method (AW) [5] adopts a fixed window and assigns a<br />
weight to each pixel in the support window according to<br />
the spatial and color similarity with the central pixel and<br />
gains a high performance. The AW method is based on<br />
bilateral filter which is a non-iterative feature-preserving<br />
image smoothing technique. Bilateral filter assigns a<br />
geometric (spatial filter) and a color proximity (range<br />
filter) constraint independently and a higher weight is<br />
assigned to the pixel with both smaller spatial and color<br />
distances to the central pixel. In AW method, the weight<br />
<strong>of</strong> a pixel within the support window is obtained by<br />
applying two independent bilateral filters in the<br />
neighborhood <strong>of</strong> potential correspondence. To further<br />
improve the disparity accuracy and decrease the<br />
computational time, many modified approaches against<br />
the AW method have been presented in recent years. The<br />
segment-based support method (SS) adds the segment<br />
information [6] in cost aggregation step. By using<br />
segment information and removing the spatial weight, the<br />
SS method can further improve the accuracy <strong>of</strong> disparity<br />
maps. But the computational time is almost double <strong>of</strong> that<br />
<strong>of</strong> the AW method. To decrease the execution time, a<br />
simplified asymmetrical strategy was proposed in [7].<br />
The bilateral filter is enforced on the reference image<br />
only and weights are computed by means <strong>of</strong> a two pass<br />
approach. These simplifications yield a real-time<br />
implementation and worse but reasonable results<br />
compared with the AW method. A fast bilateral stereo<br />
method (FBS) combines traditional local approach with a<br />
symmetric adaptive weight strategy based on two<br />
independent bilateral filters applied on a regular block<br />
basis [8]. Disparity maps yielded by the FBS method are,<br />
in general, less noisy compared with the AW method and<br />
one can trade accuracy for speed and vice versa by<br />
modifying the block size.<br />
Danny Barash pointed out that the nature <strong>of</strong> bilateral<br />
filter resembles that <strong>of</strong> anisotropic diffusion [9]. So<br />
recently many stereo matching methods based on partial<br />
differential equations (PDE) have also been presented
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 959<br />
[10-12]. Usually these methods have solutions <strong>of</strong><br />
disparity maps by constructing partial differential<br />
equations which are derived from minimizing energy<br />
functions and include different diffusion equations<br />
modeling the smoothness assumption. The diffusion<br />
equations as regularization terms in PDE model are all<br />
constructed basing on structure tensors <strong>of</strong> images. The<br />
structure tensor can be used as a local geometry indicator<br />
to analyze the geometric structure <strong>of</strong> image and widely<br />
used in image segmentation, corner detection and object<br />
tracking areas [13-15]. However they have not been used<br />
in cost aggregation method directly until now.<br />
To fill the void in existing stereo approaches, this<br />
paper presented a novel cost aggregation method against<br />
the AW method. Our approach adds a new weight in the<br />
original bilateral filter based on structure tensor proximity<br />
<strong>of</strong> corresponding pixels. By using structure tensor<br />
information, the final weight can take into account the<br />
local geometric structure <strong>of</strong> image and then can more<br />
accurately detect the similarity <strong>of</strong> two pixels. Usually a<br />
Gaussian kernel is used to smooth each element <strong>of</strong> the<br />
structure tensor over a local window to remove noise.<br />
While the Gaussian kernel is isotropic, some weak<br />
features <strong>of</strong> the image will be smoothed out. So we<br />
construct a nonlinear structure tensor based on bilateral<br />
filter which assigns a weight to a pixel according to both<br />
spatial and gradient distances with the given one. The<br />
neighboring pixels that have shorter spatial and gradient<br />
distances to the central one should have higher weights in<br />
averaging process. In this way, the nonlinear structure<br />
tensor, which is adaptive to the image local structures,<br />
could be constructed and hence similarity <strong>of</strong><br />
corresponding pixels could be better distinguished. Then<br />
the Log-Euclidean function is selected to estimate<br />
structure tensor proximity (that is structure tensor<br />
distance) <strong>of</strong> corresponding two pixels for easy<br />
implementation and efficient computation. To deal with<br />
the scale difference <strong>of</strong> structure tensor <strong>of</strong> the image, we<br />
compute the multi-scale value by summing <strong>of</strong> structure<br />
tensor distances in each scale. So a new weight based on<br />
the multi-scale structure tensor distance could be set up<br />
and used in cost aggregation function. After cost<br />
aggregation the final disparity map could be obtained by<br />
the winner-takes-all (WTA) strategy without any post<br />
processing step. The flowchart <strong>of</strong> the algorithm is shown<br />
in Fig. 1. The experimental results on Middlebury test set<br />
indicate that the performance <strong>of</strong> our proposed method is<br />
competitive with the other state-<strong>of</strong>-the-art cost<br />
aggregation strategies.<br />
The rest <strong>of</strong> paper is organized as follows. Section II<br />
briefly introduces the adaptive weight method. Section III<br />
presents our proposed cost aggregation algorithm in detail.<br />
Experimental results are given in section IV. Section V<br />
gives conclusions and an outlook to possible future work.<br />
II. THE ADAPTIVE WEIGHT METHOD<br />
In this section, we briefly introduce the adaptive<br />
weight method which bases on the original bilateral filter<br />
and will be compared with our method in experimental<br />
test section.<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 1. The flowchart <strong>of</strong> the algorithm<br />
In bilateral filter, given an image I , the value <strong>of</strong> a pixel<br />
p in the filtered image I ~ is a weighted average value<br />
described as follows<br />
~<br />
I ( p)<br />
=<br />
∑<br />
C<br />
q∈S<br />
( p)<br />
W ( I(<br />
p)<br />
− I(<br />
q))<br />
W ( p − q)<br />
I(<br />
q)<br />
∑<br />
C<br />
q∈S<br />
( p)<br />
S<br />
W ( I(<br />
p)<br />
− I(<br />
q))<br />
WS<br />
( p − q)<br />
. (1)<br />
where S( p)<br />
is a support window centered in pixel p, WS<br />
and W are weighting functions related to spatial distance<br />
C<br />
and color distance between p and q respectively. The<br />
higher weight should be assigned to the pixels with both<br />
smaller spatial and color distances to the central pixel.<br />
The adaptive weight method uses the two independent<br />
bilateral filters to execute cost aggregation. Given a pixel<br />
p in the reference image and the potential<br />
l<br />
I L<br />
corresponding pixel p in the matching image with<br />
r<br />
I R<br />
disparity d , the aggregated cost<br />
~<br />
C( pl<br />
, pr<br />
, d)<br />
is computed<br />
as follows<br />
~<br />
C(<br />
p , p , d)<br />
=<br />
l<br />
r<br />
∑<br />
W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) C(<br />
q , q , d)<br />
C<br />
ql∈S<br />
( pl<br />
)<br />
qr<br />
∈S<br />
( pr<br />
)<br />
∑<br />
C<br />
W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q ))<br />
C<br />
qr<br />
∈S<br />
( pr<br />
)<br />
ql∈S<br />
( pl<br />
)<br />
l<br />
C<br />
l<br />
l<br />
S<br />
l<br />
S<br />
S<br />
l<br />
S<br />
l<br />
l<br />
C<br />
l<br />
C<br />
C<br />
r<br />
C<br />
r<br />
r<br />
S<br />
r<br />
S<br />
S<br />
r<br />
S<br />
r<br />
r<br />
r<br />
l<br />
r<br />
. (2)<br />
where the initial matching cost C( q is single pixel<br />
l , qr<br />
, d)<br />
truncated absolute differences (TAD) score between<br />
corresponding pixels q and assuming the disparity<br />
l<br />
r<br />
value is d , the spatial distance and range distance<br />
both are Euclidean, the weighting functions and<br />
q<br />
DS<br />
W<br />
DC S
960 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
W both are Gaussian. The adaptive weight method<br />
C<br />
provides excellent results in a WTA framework which are<br />
comparable to some global methods without any complex<br />
reasoning. To more accurately distinguish the similarity<br />
<strong>of</strong> neighboring pixels, in our approach a new weight<br />
based on structure tensor proximity is added in cost<br />
aggregation function against the AW method considering<br />
the high performance <strong>of</strong> structure tensor to detect the<br />
local geometric features <strong>of</strong> the image.<br />
A. Initial Matching Cost<br />
III. OUR PROPOSED METHOD<br />
Given a reference image I and a matching image L<br />
I , R<br />
the matching cost indicates similarity <strong>of</strong> the two images<br />
with a disparity map. Many stereo matching algorithms<br />
adopt a criteria based on the color difference between two<br />
corresponding pixels. To improve the robustness to noise<br />
or distortions, we adopt a truncated L1 norm as initial<br />
matching cost function given by<br />
C ∇<br />
2<br />
2<br />
( pl<br />
, pr<br />
, d)<br />
= || I L ( pl<br />
) − I R ( pr<br />
) || + λM<br />
|| ∇I<br />
L ( pl<br />
) − I R ( pr<br />
C 0 ( p l , p r , d ) = − log[ δ M + ( 1−<br />
δ M ) exp( −C(<br />
p l , p r , d ) / σ M )]<br />
. (3)<br />
where p , are corresponding two pixels<br />
l ( x,<br />
y)<br />
pr ( x − d,<br />
y)<br />
in the reference image I and the matching image<br />
L<br />
I R<br />
T<br />
respectively assuming the disparity is d , ∇ = ( ∂x<br />
∂y)<br />
is<br />
the gradient operator, λ , M δ and M σ are predefined<br />
M<br />
parameters and C is the initial matching cost<br />
0 ( pl<br />
, pr<br />
, d)<br />
which will be used in the next cost aggregation step.<br />
B. Nonlinear Structure Tensor<br />
In the cost aggregation step, the initial matching cost<br />
will be aggregated using the bilateral filter process<br />
expressed by (2). A pixel in the support window will be<br />
assigned a weight based on the spatial and color distances<br />
with the central pixel as described before. To more<br />
accurately detect the similarity <strong>of</strong> considering two pixels,<br />
a new weight is added into (2) based on the proximity <strong>of</strong><br />
structure tensors. Firstly the nonlinear structure tensor <strong>of</strong><br />
the reference image is computed.<br />
The classic differential geometry theory [16] provides<br />
a method to analyze the local geometric structure <strong>of</strong> an<br />
image. Let us consider a multi-valued reference image:<br />
n<br />
2<br />
I ( p ) : Ω → R defined on a domain Ω ∈ R where<br />
L<br />
l<br />
+<br />
n ∈ N is the number <strong>of</strong> the image channel, p is a<br />
l ( x,<br />
y)<br />
pixel in the domain. The local variations <strong>of</strong> the vector<br />
norm || dI || can be given by<br />
L<br />
|| dI<br />
L<br />
2<br />
||<br />
) ||<br />
n<br />
T<br />
T ⎛dx⎞<br />
= dX GdX G = ∑ ∇I<br />
i∇I<br />
i dΧ<br />
= ⎜ ⎟<br />
i=<br />
1<br />
⎝dy⎠<br />
. (4)<br />
where G is symmetric as well as semi-positive-definite<br />
and its coefficients are<br />
g<br />
11<br />
=<br />
© 2011 ACADEMY PUBLISHER<br />
n<br />
n<br />
n<br />
2<br />
2<br />
∑ I i g = = ∑ =<br />
x 12 g 21 I i I x i g y 22 ∑ I iy<br />
i=<br />
1<br />
i=<br />
1<br />
i=<br />
1<br />
2<br />
. (5)<br />
We call G a structure tensor because it indicates the<br />
local geometry <strong>of</strong> the image. In fact, the eigenvalues λ , +<br />
2<br />
λ <strong>of</strong> G are the maximum and minimum <strong>of</strong><br />
−<br />
|| dI L ||<br />
while the orthogonal eigenvectors θ , + θ <strong>of</strong> G are<br />
−<br />
corresponding variation directions. Structure tensor has<br />
been used in many image applications to present the local<br />
geometric feature <strong>of</strong> the image. However, the computing<br />
derivatives is sensitive to noise, it needs to smooth the<br />
derivatives for noise reduction. Usually an isotropic<br />
Gaussian kernel is used to smooth each <strong>of</strong> the four<br />
elements in the 2× 2 structure tensor in a local window.<br />
As we all know, such a smoothing operation will smooth<br />
out some weak features and the results will not be<br />
accurate. So we construct a nonlinear structure tensor<br />
using bilateral filter same as [14]. During the smoothing<br />
the structure tensor, we consider both the spatial distance<br />
and gradient distance in the averaging weight assignment.<br />
Here the gradient distance for a neighboring pixel<br />
ql ∈ S(<br />
pl<br />
) <strong>of</strong> the central pixel p is given by<br />
l<br />
D −<br />
2<br />
G ( pl<br />
, ql<br />
) = ( I x ( pl<br />
) − I x ( ql<br />
)) + ( I y ( pl<br />
) I y ( ql<br />
))<br />
2<br />
. (6)<br />
where I , and , are the first<br />
x ( pl<br />
) I y ( pl<br />
) I x ( ql<br />
) I y ( ql<br />
)<br />
order partial derivatives <strong>of</strong> the reference image I along<br />
L<br />
the horizontal and vertical directions at pixel p and l ql<br />
respectively. For the multi-valued image, we first<br />
transformed the image into the intensity image before<br />
computing the gradient distance. The spatial distance is<br />
still Euclidean expressed by<br />
D −<br />
2<br />
S ( pl<br />
, ql<br />
) = ( x p − x ) (<br />
l q + y<br />
l p y l ql<br />
)<br />
2<br />
. (7)<br />
Then by considering both the spatial and gradient<br />
distances, we define a bilateral weighting function <strong>of</strong><br />
smoothing the structure tensor for each pixel p in the<br />
l<br />
reference image I as follows<br />
gˆ<br />
i,<br />
j<br />
( p ) =<br />
l<br />
L<br />
∑<br />
G<br />
ql∈S<br />
( pl<br />
)<br />
W ( D ( p , q )) W ( D ( p , q )) g ( q )<br />
∑<br />
G<br />
G<br />
ql∈S<br />
( pl<br />
)<br />
l<br />
G<br />
l<br />
l<br />
l<br />
S<br />
S<br />
S<br />
l<br />
S<br />
l<br />
l<br />
i,<br />
j<br />
W ( D ( p , q )) W ( D ( p , q ))<br />
where is the element <strong>of</strong> structure tensor for each<br />
g i,<br />
j<br />
l<br />
l<br />
.(8)<br />
neighboring pixel <strong>of</strong> the central pixel p in the support<br />
l<br />
window, the range <strong>of</strong> indices is , j = 1,<br />
2 , ˆ is the<br />
i g i,<br />
j<br />
filtered element value <strong>of</strong> structure tensor for the central<br />
pixel p , l WG = exp( −DG<br />
/ 2σ<br />
G ) and WS = exp( −DS<br />
/ 2σ<br />
S )<br />
both are Gaussian functions based on the spatial distance<br />
D and the gradient distance respectively. Compared<br />
G<br />
DS<br />
with the Gaussian kernel, the weighting function is<br />
anisotropic and adaptive to the image local structure by<br />
using bilateral filter.<br />
C. Multi-scale Nonlinear Structure Tensor Distance<br />
Multi-scale structure tensor was firstly defined in [17]<br />
and named as multi-scale fundamental forms. It has been<br />
widely used in multi-valued images fusion or merging,
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 961<br />
noise filtering and segmentation etc. By considering the<br />
scale differences <strong>of</strong> the multi-valued image, we construct<br />
a multi-scale structure tensor by using tensor information<br />
at each scale. The images at different scales can be<br />
obtained by smoothing the original image I with a<br />
L<br />
series <strong>of</strong> Gaussian kernels k with different standard<br />
deviations ξ . By increasing the values <strong>of</strong> ξ , a fine to<br />
coarse scale space can be formed. Then a multi-scale<br />
structure tensor can be expressed by<br />
G<br />
m<br />
=<br />
n<br />
∑<br />
i=<br />
1<br />
ξ<br />
∇(<br />
I ∗ Fξ<br />
) ∇(<br />
I ∗ Fξ<br />
)<br />
L<br />
m<br />
i<br />
L<br />
T<br />
m i<br />
. (9)<br />
where m ∈ L is the current scale level, L is the total<br />
number <strong>of</strong> scales, ∗ is the convolution operator and F<br />
is the Gaussian function with the deviation ξ . Finally<br />
m<br />
the computed multi-scale structure tensor is bilateral<br />
filtered for different scales as described before. We<br />
should notice that each scale is nonlinearly filtered<br />
separately and each scale has three different channels.<br />
One key factor in the tensor space analysis is a proper<br />
choice <strong>of</strong> the tensor distance norm to measure the<br />
similarity between tensors. For simplicity, we use the<br />
recently proposed Log-Euclidean calculus [18] as the<br />
measure function <strong>of</strong> tensor similarities. The Log-<br />
Euclidean distance between two tensors and G for<br />
pixels and q respectively is given by<br />
pl l<br />
G pl<br />
l q<br />
ξm<br />
ˆ<br />
ˆ 2<br />
DT ( pl<br />
, ql<br />
) = tr((log<br />
m(<br />
G p ) − log m(<br />
G )) )<br />
l<br />
ql<br />
. (10)<br />
where tr (.) is the trace operator <strong>of</strong> a matrix, log m(.)<br />
is<br />
the logarithm operator <strong>of</strong> a matrix, the “hat” denotes that<br />
the structure tensor has been bilateral filtered as described<br />
before. The logarithm <strong>of</strong> a matrix is computed by<br />
decomposing the matrix to its eigenvalues and<br />
eigenvectors, taking the logarithms <strong>of</strong> the eigenvalues,<br />
and constructing a matrix from the eigenvectors and the<br />
newly computed eigenvalues. Thus the matrix logarithm<br />
operation is particularly easy to compute when structure<br />
tensor G is given in terms <strong>of</strong> its eigenvalues and<br />
eigenvectors.<br />
ˆ<br />
The tensor distance for multi-scale structure tensor can<br />
be defined as the square root <strong>of</strong> the sum <strong>of</strong> the Log-<br />
Euclidean distances for all scales and can be rewritten as<br />
∑ − L 1<br />
~<br />
=<br />
ˆ m<br />
2<br />
) − log ( ˆ m<br />
DT<br />
( pl<br />
, ql<br />
) tr((log<br />
m(<br />
G p m Gq<br />
)) )<br />
l<br />
l<br />
m=<br />
0<br />
. (11)<br />
~<br />
Then the new weight WT = exp( −DT<br />
/ 2σ<br />
T ) between<br />
two considering pixels p and is computed which is<br />
l q l<br />
based on the similarity <strong>of</strong> their tensors.<br />
D. Cost Aggregation and Disparity Computation<br />
By the above procedures, the final cost aggregation<br />
equation can be expressed as follows<br />
© 2011 ACADEMY PUBLISHER<br />
~<br />
C(<br />
p , p , d)<br />
=<br />
l<br />
r<br />
∑<br />
~<br />
W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q )) C ( q , q , d)<br />
C<br />
ql∈S<br />
( pl<br />
)<br />
qr<br />
∈S<br />
( pr<br />
)<br />
C<br />
∑<br />
~<br />
W ( D ( p , q )) W ( D ( p , q )) W ( D ( p , q ))<br />
C<br />
ql∈S<br />
( pl<br />
)<br />
qr<br />
∈S<br />
( pr<br />
)<br />
l<br />
C<br />
l<br />
l<br />
S<br />
l<br />
S<br />
S<br />
l<br />
S<br />
l<br />
l<br />
T<br />
l<br />
T<br />
T<br />
l<br />
T<br />
l<br />
l<br />
0<br />
l<br />
l<br />
r<br />
. (12)<br />
where W and are spatial weight and range weight<br />
S WC<br />
respectively same as the expression in (2) used by the<br />
AW method. Compared with (2), a new weight is<br />
included which reflects the similarity <strong>of</strong> two considering<br />
pixels based on their structure tensors in our approach. A<br />
neighboring pixel is assigned a high weight to the central<br />
pixel not only if their spatial and range distances are<br />
small but also if they have similar local geometric<br />
structures. The final weight could more accurately<br />
distinguish the similarity <strong>of</strong> two relevant pixels. It is<br />
worth noting that, in the above equation, we simply<br />
execute the filter on the reference image only different<br />
from the adaptive weight method to decrease the<br />
execution time.<br />
Finally the disparity map is obtained in a WTA<br />
framework and expressed as below<br />
~<br />
d ( pl<br />
) = arg min( C(<br />
pl<br />
, pr<br />
, d ))<br />
d∈D<br />
. (13)<br />
where D represents the set <strong>of</strong> all allowed disparities.<br />
IV. EXPERIMENTAL RESULTS<br />
In this section, we aim at assessing the performance <strong>of</strong><br />
our proposed cost aggregation approach based on the<br />
modified bilateral filter with multi-scale nonlinear<br />
structure tensor. We used the Middlebury test bed<br />
provided by authors <strong>of</strong> [1] to evaluate our approach<br />
performance compared with the cost aggregation method<br />
based on the original bilateral filter and the other state-<strong>of</strong>-<br />
the-art methods.<br />
Firstly we compared the results <strong>of</strong> our method with<br />
that <strong>of</strong> the method using the original bilateral filter (OBF).<br />
Both methods adopt asymmetric strategy which executes<br />
filter on the reference image only for simplicity and is<br />
different from that <strong>of</strong> the adaptive weight method. For<br />
comparison our method and the OBF method both select<br />
the L1 norm as initial cost function which is also different<br />
from the AW method. A constant set <strong>of</strong> parameters are<br />
run for all test images. The initial cost function<br />
7<br />
parameters are λ M = 2 , 10 and −<br />
δ M = σ M = 2 , the<br />
bilateral structure tensor parameters are σ G = 20 ,<br />
σ S = 2.<br />
5 and the window size = 5× 5 . For multi-scale<br />
structure tensor computation, the total number <strong>of</strong> scales<br />
L is 3, the Gaussian kernelξ 0 = 0.<br />
6 , 1 1.<br />
0 = ξ and 2 1.<br />
4 = ξ .<br />
The cost aggregation parameters are σ C = 15 , σ S = 10.<br />
5 ,<br />
σ T = 5 and the window size = 21× 21.<br />
The corresponding<br />
disparity maps are plotted in Fig. 2. From the figures, we<br />
can see that our proposed method has better results than<br />
the OBF method using the original bilateral filter. This<br />
manifests that including new weight in cost aggregation<br />
function can actually improve the disparity accuracy,
962 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Figure 2. The disparity maps <strong>of</strong> two methods. The top is input reference images <strong>of</strong> “Tsukuba”, “Venus”, “Teddy” and “Cones” images. The second<br />
row is the truth disparity maps. The third row is the results <strong>of</strong> the OBF method. The last row is our results.<br />
especially in discontinuity area, because <strong>of</strong> structure<br />
tensor reflecting the local geometric feature <strong>of</strong> the image.<br />
On the other hand, the nonlinear filter and the<br />
construction multi-scale values <strong>of</strong> structure tensors can<br />
further remove the noise and improve the accuracy <strong>of</strong> the<br />
discontinuity location fitting for real images which<br />
usually have different light sources or distortions. To<br />
manifest this, we have the same comparison experiments<br />
on other two datasets which are available at Middlebury<br />
test bed. Each dataset has 9 different images that exhibit<br />
3 different exposure and 3 different lighting variations.<br />
Fig. 3 shows the both exposure and lighting variations <strong>of</strong><br />
Figure 3. The left image <strong>of</strong> the Art dataset with three different<br />
exposures and under three different light conditions<br />
© 2011 ACADEMY PUBLISHER<br />
left image <strong>of</strong> the Art dataset. Quantitative comparative<br />
results are given in Table I. The experimental parameters<br />
are all the same with those in first the experiment. The<br />
focus here is evaluation <strong>of</strong> the raw cost aggregation<br />
methods which don’t deal explicitly with occlusions and<br />
TABLE I.<br />
QUANTITATIVE COMPARATIVE RESULTS OF OUR METHOD WITH THE<br />
OBF METHOD FOR REAL TEST IMAGES.<br />
Art<br />
Dataset<br />
Book<br />
Dataset<br />
OBF method Our method<br />
Vis. Dis. Vis. Dis.<br />
Art1-1 15.03 23.59 14.57 21.47<br />
Art1-2 12.39 19.13 12.40 17.08<br />
Art1-3 14.26 22.01 14.08 19.50<br />
Art2-1 13.91 21.74 13.48 19.54<br />
Art2-2 11.07 18.16 10.74 16.05<br />
Art2-3 15.88 22.48 15.62 20.15<br />
Art3-1 13.24 18.06 12.88 15.48<br />
Art3-2 13.67 16.66 13.12 14.58<br />
Art3-3 26.80 24.65 26.61 21.40<br />
Book1-1 12.57 25.39 12.29 24.63<br />
Book1-2 13.21 25.47 13.13 25.17<br />
Book1-3 23.02 36.26 23.45 36.11<br />
Book2-1 13.85 28.78 13.68 28.55<br />
Book2-2 15.19 27.55 15.35 27.30<br />
Book2-3 14.46 28.31 14.65 28.33<br />
Book3-1 12.39 25.26 12.01 24.54<br />
Book3-2 12.61 25.26 12.61 24.70<br />
Book3-3 17.81 33.36 17.95 32.75
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 963<br />
we only report the percentage <strong>of</strong> bad pixels (i.e. pixel<br />
whose absolute disparity error is greater than 1) in nonoccluded<br />
region (Vis.) and in near depth discontinuity<br />
( Dis.) described by [1]. From the table we can see that<br />
our proposed method can outperform the OBF method<br />
for real images having different exposure and light<br />
source changes due to including multi-scale nonlinear<br />
structure tensor information. Our results have been<br />
plotted in Fig. 4.<br />
In the third experiment, to evaluate the effectiveness<br />
<strong>of</strong> the multi-scale nonlinear structure tensor, our method<br />
is compared with the method using simple structure<br />
tensor information for clean and real test images. The<br />
same parameters are adopted by both methods for<br />
fairness. The quantitative comparative results are given<br />
in Table II. From the table, we find that our method<br />
using the multi-scale nonlinear structure tensor is more<br />
effective than the method using simple structure tensor<br />
for real images while not for clean images, mainly<br />
because real images have more noise or distortions and<br />
multi-scale nonlinear structure tensor can be used to<br />
effectively remove noise which is approved by the last<br />
experiment. We should notice that the computational<br />
time mainly depends on the bilateral filter process not<br />
multi-scale nonlinear structure tensor computation, so<br />
execution time <strong>of</strong> our approach increases a little when<br />
compared with the method using simple structure tensor.<br />
Lastly we compared our proposed method with the<br />
state-<strong>of</strong>-the-art cost aggregation strategies. For our<br />
method, we adopt the optimal parameters minimizing the<br />
Vis.+Dis. error on the whole test images: window<br />
size = 31× 31 , σ S = 15.<br />
5 and the other parameters are<br />
same as those in the first experiment. The results <strong>of</strong> the<br />
other four top methods used here were reported in [7]. It<br />
is worth noting that these results reported by [7] are<br />
obtained using the original cost function proposed by the<br />
authors <strong>of</strong> each paper and the results for AW and SS<br />
available on the Middlebury evaluation sites including<br />
the post processing steps are not used. We have reported<br />
quantitative comparative results in Table III. From the<br />
table, we can see that our proposed method have results<br />
comparable to the best performing cost aggregation<br />
strategies. The most similar method with our approach is<br />
the AW method based on the original bilateral filter. The<br />
AW method outperforming our results is mainly due to<br />
its symmetric strategy while our approach adopts the<br />
asymmetric strategy. However our method runs faster<br />
than the AW method and the AW run is 3226 seconds<br />
according to [7] while our method takes 210 seconds<br />
without any accelerating techniques on the Teddy image.<br />
V. CONCLUSIONS<br />
In this paper, we proposed a novel cost aggregation<br />
algorithm based on modified bilateral filter with multiscale<br />
nonlinear structure tensor for local stereo matching.<br />
A new weight is included in cost aggregation equation<br />
based on the structure tensor distance. For reducing noise<br />
influence, in structure tensor computation, each element<br />
<strong>of</strong> the structure tensor is smoothed using a bilateral filter<br />
and the nonlinear structure tensor is constructed. Then a<br />
multi-scale tensor is computed at each scale for<br />
considering the scale difference <strong>of</strong> the multi-valued<br />
image. The nonlinear structure tensor can be used in the<br />
Log-Euclidean measure function as tensor similarities.<br />
Figure 4. The disparity maps <strong>of</strong> our method for the Art and Book datasets. The first image <strong>of</strong> the two datasets is truth disparity maps respectively for<br />
the Art and Book datasets.<br />
© 2011 ACADEMY PUBLISHER
964 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
TABLE II.<br />
DISPARITY MAPS OF OUR METHOD AND THE METHOD USING SIMPLE STRUCTURE TENSOR.<br />
The method using simple<br />
structure tensor<br />
Our method<br />
Vis. Dis. Vis. Dis.<br />
Tsukuba 3.01 9.76 3.11 10.44<br />
Venus 6.97 12.49 6.68 13.15<br />
Teddy 10.86 19.67 10.86 19.70<br />
Cones 6.45 12.76 6.41 12.77<br />
Art1-1 14.57 21.74 14.57 21.47<br />
Art1-2 12.25 17.82 12.40 17.08<br />
Art1-3 13.77 20.12 14.08 19.50<br />
Art<br />
Dataset<br />
Art2-1<br />
Art2-2<br />
Art2-3<br />
13.42<br />
10.74<br />
15.57<br />
19.98<br />
16.72<br />
20.70<br />
13.48<br />
10.74<br />
15.62<br />
19.54<br />
16.05<br />
20.15<br />
Art3-1 12.92 16.09 12.88 15.48<br />
Art3-2 13.20 15.08 13.12 14.58<br />
Art3-3 26.22 22.14 26.61 21.40<br />
Book1-1 12.48 24.76 12.29 24.63<br />
Book1-2 13.17 25.08 13.13 25.17<br />
Book1-3 24.01 36.25 23.45 36.11<br />
Book<br />
Dataset<br />
Book2-1<br />
Book2-2<br />
Book2-3<br />
13.97<br />
15.52<br />
15.00<br />
28.51<br />
27.56<br />
28.77<br />
13.68<br />
15.35<br />
14.65<br />
28.55<br />
27.30<br />
28.33<br />
Book3-1 12.05 24.41 12.01 24.54<br />
Book3-2 12.80 24.91 12.61 24.70<br />
Book3-3 18.44 32.86 17.95 32.75<br />
TABLE III.<br />
QUANTITATIVE COMPARATIVE RESULTS OF OUR METHOD WITH THE OTHER FOUR TOP ALGORITHMS.<br />
Tsukuba Venus Teddy Cones<br />
Vis. Dis. Vis. Dis. Vis. Dis. Vis. Dis.<br />
SS[4] 2.19 7.22 1.38 6.27 10.50 21.20 5.83 11.80<br />
AW[3] 3.33 8.87 2.02 9.32 10.52 20.84 3.72 9.37<br />
FBS[6] 2.95 8.69 1.29 7.62 10.71 20.82 5.23 11.34<br />
VW[2] 3.12 12.40 2.42 13.30 17.70 25.50 21.20 27.30<br />
Our method 2.70 10.72 4.93 11.83 10.86 19.70 6.41 12.77<br />
Finally the multi-scale tensor distance is set up which is<br />
the square root <strong>of</strong> sum <strong>of</strong> tensor distances at each scale.<br />
So the new weight can be computed and related to the<br />
multi-scale tensor distance. The proposed new algorithm<br />
not only considers the spatial and range distances <strong>of</strong> two<br />
pixels same as the original bilateral filter, but also the<br />
local geometric feature distance <strong>of</strong> them. Our new weight<br />
can more accurately reflect the similarity <strong>of</strong> two pixels<br />
which is important for cost aggregation approach. The<br />
experimental results confirm the effectiveness <strong>of</strong> our<br />
approach compared with the OBF method using clean<br />
and real test sets and the other state-<strong>of</strong>-the-art strategies.<br />
In the future, we plan to devise new cost aggregation<br />
methods based on the structure tensor to further improve<br />
the accuracy <strong>of</strong> disparity maps and decrease the<br />
computational time. We are also interested in using the<br />
diffusion equation constructed by structure tensor in the<br />
variational framework to devise the global stereo<br />
matching method.<br />
ACKNOWLEDGMENT<br />
The authors would like to thank financial supports<br />
from National Natural Science Foundation <strong>of</strong> China<br />
© 2011 ACADEMY PUBLISHER<br />
under Grant Nos. 60970048, Natural Science Foundation<br />
<strong>of</strong> Shandong Province Grant Nos. 2009ZRB019SF.<br />
REFERENCES<br />
[1] D. Scharstein and R. Szeliski, “A taxonomy and<br />
evaluation <strong>of</strong> dense two-frame stereo correspondence<br />
algorithms”, International <strong>Journal</strong> <strong>of</strong> Computer Vision,<br />
pp.7-42, 2002.<br />
[2] O. Veksler, “Fast variable window for stereo<br />
correspondence using integral images”, In Proc. Of Conf.<br />
on CVPR 2003, pp.556-561, 2003.<br />
[3] H. Hirschmuller, P. Innocent, and J. Garibaldi, “Real-time<br />
correlation-based stereo vision with reduced border<br />
errors”, International <strong>Journal</strong> <strong>of</strong> Computer Vision, 47:1–3,<br />
2002.<br />
[4] J. Kim, K. Lee, B. Choi, and S. Lee, “A dense stereo<br />
matching using two-pass dynamic programming with<br />
generalized ground control points”, In Proc .<strong>of</strong> Conf. on<br />
Computer Vision and Pattern Recognition 2005, pp.<br />
1075–1082, 2005.<br />
[5] K. J. Yoon and I. S. Kweon, “Adaptive support-weight<br />
approach for correspondence search”, IEEE TPAMI, 28(4),<br />
pp.650-656, 2005.<br />
[6] F. Tombari, S. Mattoccia, and L. Di Stefano,<br />
“Segmentation-based adaptive support for accurate stereo<br />
correspondence”, In Proc. <strong>of</strong> Conf. on PSIVT, 2007.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 965<br />
[7] M. Gong, R.G. Yang, W. Liang, and M.W. Gong, “A<br />
performance study on different cost aggregation<br />
approaches used in real-time stereo matching”,<br />
International <strong>Journal</strong> Computer Vision, 75(2), pp.283-296,<br />
2007.<br />
[8] S. Mattoccia, S. Giardino, and A. Gambini, “Accurate and<br />
efficient cost aggregation strategy for stereo<br />
correspondence based on approximated joint bilateral<br />
filtering”, In Proc. <strong>of</strong> Conf. on ACCV, 2009.<br />
[9] Danny Barash, “A Fundamental relationship between<br />
bilateral filtering, adaptive smoothing and the nonlinear<br />
diffusion equation”, IEEE TPAMI, Vol. (24), No.6, June,<br />
2002.<br />
[10] D. Scharstein and R. Szeliski, “Stereo matching with<br />
nonlinear diffusion”, International <strong>Journal</strong> <strong>of</strong> Computer<br />
Vision, 28(2), pp.155-174, 1998.<br />
[11] R. Ben-Ari and N. Sochen, “A geometric approach for<br />
regularization <strong>of</strong> the data term in stereo-vision”,<br />
International <strong>Journal</strong> <strong>of</strong> Math Imaging, vol.31, pp.17-33,<br />
2008.<br />
[12] R.B. Ari and N. Sochen, “Variational stereo vision with<br />
sharp discontinuities and occlusion handling”, In Proc. <strong>of</strong><br />
Conf. on ICCV, Rio de Janeiro, Brazil, pp.1-7, 2007.<br />
[13] S. Han, W. Tao, D. Wang, X.Ch. Tai and X.L. Wu,<br />
“Image segmentation based on grabcut framework<br />
integrating multi-scale nonlinear structure tensor”, IEEE<br />
Transactions on Image Processing, 18(10), pp. 289-302,<br />
June 2009.<br />
[14] L. Zhang, L. Zhang and D. Zhang, “A multi-scale bilateral<br />
structure tensor based corner detector”, In Proc. <strong>of</strong> Conf.<br />
on ACCV, 2009.<br />
© 2011 ACADEMY PUBLISHER<br />
[15] M. Donoser, S. Kluckner and H. Bisch<strong>of</strong>, “Object tracking<br />
by structure tensor analysis”, In Proc. <strong>of</strong> International<br />
Conference on Pattern Recognition, 2010.<br />
[16] M. D. Carmo, “Differential geometry <strong>of</strong> curves and<br />
surfaces”, Prentice Hall, 1976.<br />
[17] J. Weickert, B. Romeny and M. A. Viergever, “Efficient<br />
and reliable schemes for nonlinear diffusion filtering”,<br />
IEEE Transactions on Image Processing, vol.7, pp.398-<br />
410, 1998.<br />
[18] V. Arsigny, P. Fillard, X. Pennec, and N. Ayache, “Log-<br />
Euclidean metrics for fast and simple calculus on diffusion<br />
tensors”, Magnetic Resonance in Medicine, vol. 56(2), pp.<br />
411-421, 2006.<br />
Li Li works as an instructor in the School <strong>of</strong> Computer Science<br />
and Technology at the Shandong Economic University. She<br />
received the B.E. degree in Motor Engineering from Shandong<br />
Technology University in 1998 and M.E. degree from<br />
Shandong University in 2000, and now is currently a Ph.D.<br />
student in the School <strong>of</strong> Computer Science and Technology at<br />
the Shandong University. Her research interests lie in computer<br />
vision, especially in 3D reconstruction and stereo matching.<br />
Hua Yan works as an associate pr<strong>of</strong>essor in the School <strong>of</strong><br />
Computer Science and Technology <strong>of</strong> Shandong Economic<br />
University. She received the B.S. degree in Physics in 1997,<br />
M.E. degree in Communication and Information System in<br />
2004 and D Sc Tech degree in Communication and Information<br />
System in 2007 from Shandong University. Dr. Yan’s research<br />
interests include image and video processing, multimedia data<br />
retrieval and super resolutions.
966 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
A Collaborative Nonlocal-Means Superresolution<br />
Algorithm Using Zernike Monments<br />
Lin Guo 1,2 * , Qinghu Chen 1<br />
1 School <strong>of</strong> Electronic and Information, Wuhan University, Wuhan, 430074, China<br />
2 School <strong>of</strong> Physics and Electronics, Hubei University, Wuhan, 430062, China<br />
Email: gunglin@gmail.com, cqh@eis.whu.edu.cn<br />
Abstract—Super-resolution (SR) with probabilistic motion<br />
estimation is a successful algorithm to circumvent the<br />
limitation <strong>of</strong> motion estimation upon conventional superresolution<br />
methods. However, the algorithm can’t match<br />
similar patches with rotation or scale. This paper presents<br />
an efficient improved algorithm by introducing Zernike<br />
moments as representation <strong>of</strong> image invariant features into<br />
similarity measure. A collaborative strategy is proposed<br />
combining the moment based proximity and the bilateral<br />
proximity <strong>of</strong> nonlocal means (NL-means) algorithm for joint<br />
determination <strong>of</strong> weights. For the invariant property <strong>of</strong><br />
Zernike moments, structure-similar pixels with rotation or<br />
scale can also be matched for computation <strong>of</strong> weights.<br />
Furthermore, the collaborative mechanism ensures higher<br />
accuracy <strong>of</strong> weights for a better estimation <strong>of</strong> each pixel in<br />
SR images. Experimental results indicate the proposed<br />
method is able to handle general video sequences with<br />
superior performance in SR reconstruction to the compared<br />
algorithms.<br />
Index Terms—super resolution, Zernike moments,<br />
probabilistic motion estimation, nonlocal means,<br />
collaborative<br />
I. INTRODUCTION<br />
Super-Resolution (SR) technique is the fusion <strong>of</strong> a<br />
sequence <strong>of</strong> low-resolution noisy, blurred images to<br />
produce a higher resolution image or sequence<br />
overcoming the inherent resolution limitation <strong>of</strong> LR<br />
imaging systems. Since Huang and Tsai first proposed the<br />
concept <strong>of</strong> SR in 1984, the SR technique has attracted a<br />
lot <strong>of</strong> attention in the image processing community due to<br />
its wide variety <strong>of</strong> application in image enhancement,<br />
medical imaging, high definition televisions and<br />
computer vision. A great deal <strong>of</strong> literature about SR can<br />
be found, and the representatives are referred in [1-4].<br />
SR reconstruction is an ill-posed inverse problem. A<br />
wisdly used model for this problem is described as<br />
follows:<br />
y � DHM z � n . (1)<br />
t t , s s<br />
Where the measurements yt, t = 1, 2, …, T, are results <strong>of</strong><br />
different motion, noise, blur, and decimation parameters<br />
* Corresponding author: Lin Guo, gunglin@gmail.com<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.966-973<br />
from an original high resolution reference image zs. The<br />
matrix Mt,s indicates the geometric warp <strong>of</strong> yt relative to<br />
the high resolution image zs. And H is the blur matrix.<br />
Both <strong>of</strong> them are assumed for simplicity to be linear<br />
space and time invariant. Similarly, D denotes the fixed<br />
spatial resolution decimation. Gaussian random noise n is<br />
assumed to be added to the measurements.<br />
To recover zs from yt, Mt,s and H must be known or can<br />
be reliably estimated from inputs. Most <strong>of</strong> the existing SR<br />
methods are roughly based on an estimation <strong>of</strong> the motion<br />
between frames followed by the super-resolution fusion<br />
<strong>of</strong> inputs according to these motion vectors. As it is well<br />
know, motion estimation <strong>of</strong> sub-pixel precise between<br />
frames is indispensable and commonly regarded very<br />
critical for successful SR reconstruction.<br />
However, it is a challenging task to obtain highly<br />
accurate motion estimation with an affordable<br />
computation load. In fact, it’s almost impossible to<br />
handle actual scenes with complex motion patterns or<br />
very low quality. Inaccurately registration <strong>of</strong>ten leads to<br />
deteriorated reconstruction results even compared to a<br />
simple interpolated version. So motion estimation has<br />
become the bottleneck for the conventional SR methods<br />
to get excellent performance.<br />
In order to overcome the above problem, several recent<br />
articles [4-7] attempted to deliver SR methods avoiding<br />
explicit motion estimation apart from above conventional<br />
methods. The algorithm in [5] relies on extending their<br />
previous steerable kernel regression method to multiframe<br />
super-resolution. The approach in [6] is based on<br />
the sparse 3D transform-domain collaborative filtering<br />
and iterative projection on the observation constrained<br />
subspace. The method in [7] develops the notion <strong>of</strong><br />
probabilistic motion estimation into the classical SR<br />
formulation, which is regarded as a generalization <strong>of</strong> the<br />
very successful nonlocal means (NL-means) denoising<br />
method [8] to serve the super-resolution task [4]. The<br />
main idea <strong>of</strong> the NL-means is that the pixel is estimated<br />
as a weighted average <strong>of</strong> similar pixels in its nonlocal<br />
neighborhood, and the weights are computed according to<br />
the similarity between two pixels. It shows simple and<br />
robust to noise. However, the similarity measure only in<br />
intensity is crude to ensure the accuracy <strong>of</strong> weights for no<br />
any information about the underlying image features are<br />
considered. For example, structure-similar patches with<br />
rotation or various scales are unable to be matched. As a<br />
result, unsuitable weights are calculated and assigned to
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 967<br />
pixels, and hence lead to the estimation value to deviate<br />
from the true one.<br />
Several recent papers have tried to improve the NLmeans<br />
algorithm in image denoising. Ref. [9] and [10]<br />
renders a similar approach by employing affine gray scale<br />
transformations to find patches at the same or different<br />
scales. Ref. [11] uses cross-scale (i.e., downsampled)<br />
neighborhoods in the NL-means filter [12]. Ref. [13]<br />
introduces SIFT as rigid invariant features to compute the<br />
similarity between different patches. SIFT features as<br />
local descriptors are suitable for image retrieval, affine<br />
registration etc., but they are not for denoising and SR.<br />
Ref. [14] develops Hu moment as rotationally invariant<br />
feature into similarity measure strategy. Hu moment is<br />
the simplest moment, but it’s not efficient in many cases.<br />
Ref. [15] replaces the geometrical Hu moment in [14]<br />
with Zernike moment, and shows a competitive<br />
performance in image denoising application. Zernike<br />
moment is proved to be superior to geometrical moment<br />
for better capabilities <strong>of</strong> invariant feature representation<br />
because <strong>of</strong> its orthogonal property. This motivates us to<br />
introduce Zernike moments as invariant descriptors <strong>of</strong><br />
image shape features into the similarity measure to<br />
improve the super-resolution results. Here, we need to<br />
point out that the intuitive approach through independent<br />
interpolation <strong>of</strong> each frame followed by the Zernike<br />
moment based denoising processing is unable to provide<br />
super-resolution results [4].<br />
Our super-resolution algorithm using the Zernike<br />
moments is based on the SR framework with probabilistic<br />
motion estimation [7]. Similarity for two pixels is<br />
computed on two small local patches around them in the<br />
Zernike moment images. For the invariant property <strong>of</strong><br />
Zernike moment, the algorithm is enabled to match more<br />
similar patches not only with translation but also with<br />
rotation or scale. However, the images <strong>of</strong> super-resolution<br />
generally may contain complex degradation involving<br />
downsampling, aliasing as well as noise, especially for<br />
higher order moments that are more sensitive to the noise.<br />
So Zernike moments <strong>of</strong> SR images are usually unreliable<br />
to be the sole basis for computation <strong>of</strong> weights. To tackle<br />
the problem, a collaborative algorithm is designed<br />
combining the Zernike moment based proximity and<br />
intensity based bilateral proximity <strong>of</strong> NL-means<br />
algorithm for joint determination <strong>of</strong> weights.<br />
The algorithm proposed in this paper has the following<br />
major features. Firstly, besides retaining the advantage <strong>of</strong><br />
avoiding explicit motion estimation, our algorithm<br />
extends the notion <strong>of</strong> probabilistic motion estimation in<br />
[7] to include not only intensity-similar patches with<br />
translation but also structure-similar patches with rotation<br />
or scale. Secondly, the collaborative similarity measure<br />
strategy balances the influence <strong>of</strong> gray-level based<br />
proximity and invariant moment based proximity. Then<br />
weights with higher accuracy are computed for better<br />
estimation <strong>of</strong> a pixel.<br />
The remainder <strong>of</strong> the paper is as follows. Section II<br />
presents the super-resolution framework with<br />
probabilistic motion estimation on which our method is<br />
based. Section III describes the proposed collaborative<br />
© 2011 ACADEMY PUBLISHER<br />
super-resolution algorithm using Zernike moments in<br />
details. A simplified numerical algorithm in iterative<br />
form is given at last in this section. Section IV shows<br />
experimental results on several general video sequences,<br />
followed by conclusion and discussion in Section V.<br />
II. THE SR FRAMEWORK WITH PROBABILISTIC MOTION<br />
ESTIMATION<br />
According the observation model (1), the Maximum-<br />
Likelihood (ML) estimation <strong>of</strong> high resolution image is<br />
expressed as<br />
T 1<br />
2<br />
zˆ s � arg min � DHMt , szs � y t . (2)<br />
2<br />
2 t�1<br />
The matrix Mt,s in classical SR methods denotes a oneto-one<br />
mapping between pixels in the s-th and the t-th<br />
image. And as such, it introduces sensitivity to errors.<br />
According to the idea <strong>of</strong> probabilistic motion estimation<br />
[7], the one-to-one mapping between pixels in classical<br />
SR methods is substituted to a probabilistic movement<br />
domain. That means every estimated pixel in the<br />
reference image with many possible correspondences in<br />
all the frames <strong>of</strong> the sequences (including itself). Each<br />
pixel inside the domain is assigned a value <strong>of</strong> weight to<br />
denote the probability <strong>of</strong> being correct. The movement<br />
domain is a spatial and temporal neighborhood centered<br />
at the estimated pixel in the reference image with radius<br />
R among all the sequences. For given s and t, the<br />
displacement between the estimated pixel and every pixel<br />
inside the domain is written as [dx(n), dy(n)], n = 1, … ,<br />
N, N = (2R+1) 2 . The location relationship is described by<br />
a matrix Mn with size <strong>of</strong> S1N1S2N2×S1N1S2N2 and value<br />
<strong>of</strong> 1 in one position and 0 for others. S1 and S2 are<br />
sampling factors respectively in horizontal and vertical<br />
direction. For the pixel whose displacement is indicated<br />
by the 1 in Mn, the corresponding weight is denoted by<br />
Wn;t,s , a diagonal matrix with the same size as Mn. Thus,<br />
we get the following equation:<br />
N<br />
M z � �W<br />
M z . (3)<br />
t, s s n; 1t<br />
, s n s<br />
n�1<br />
According to (2) and (3), the probabilistic ML<br />
estimation <strong>of</strong> the high resolution image is formulated as<br />
follows:<br />
1<br />
zˆ � arg min DHM z � y , (4)<br />
s PML 2<br />
N T<br />
2<br />
�� n s t L<br />
Wn<br />
; t , s<br />
n�1 t�1<br />
L<br />
where W , with size <strong>of</strong> N1N2 × N1N2, is the<br />
n; 1t<br />
, s<br />
corresponding weight matrix in low resolution space by<br />
being downsampled from Wn;t,s.<br />
Since both H and Mn are space-invariant, they can be<br />
exchanged in position. Thus, defining x = Hz, the task <strong>of</strong><br />
SR is turned to be a two-step process: first estimation <strong>of</strong><br />
the “blurry” high resolution image x according to (5) and<br />
the subsequent acquirement <strong>of</strong> z from x by using existing<br />
deblurring algorithms.
968 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
1<br />
x DM x y . (5)<br />
N T<br />
2<br />
ˆ s PML � arg min �� n s � t Wn<br />
; t , s<br />
2 n�1 t�1<br />
Minimization <strong>of</strong> (5) leads to a solution represented in<br />
pixel-wise as<br />
( k , l) �N ( i, j) t�1<br />
n; t, s t<br />
s i j � T<br />
xˆ<br />
( , )<br />
T<br />
� �<br />
� �<br />
( k , l ) �N ( i, j) t�1<br />
W ( k, l, i, j) y ( k, l)<br />
,<br />
W ( k, l, i, j)<br />
n; t, s<br />
where (i, j) is an arbitrary coordinate on high resolution<br />
grid. And (k, l)�N(i, j) denotes the (k, l)-th pixel within<br />
the movement domain for pixel (i, j), but is located on<br />
low resolution grid. That is, (k, l) s.t. (S1k, S2l)� N(i, j),<br />
which ensures that the center pixel <strong>of</strong> the patch is on the<br />
decimation grid. The weight Wn;t,s is computed based on<br />
the bilateral proximity strategy according to<br />
W<br />
� 2<br />
� �<br />
� Ri, jDM n xs � Rk , l y �<br />
t<br />
2<br />
n; t, s ( k, l, i, j)<br />
exp�<br />
�<br />
�<br />
�<br />
2 �<br />
� � �<br />
g<br />
� � �<br />
�<br />
�� �<br />
��<br />
2 2 2<br />
� ( ( )) ( ( )) ( ) �<br />
(6)<br />
f dx n � dy n � s � t , (7)<br />
where the function f may be an arbitrary monotonically<br />
non-increasing function, such as Gaussian or box form.<br />
The parameter σg controls the effect <strong>of</strong> the gray-level<br />
difference between two patches. Ri,j is an operator that<br />
extracts a patch <strong>of</strong> a fixed size centred at the (i, j)-th pixel<br />
from an image. The square differences <strong>of</strong> all the pixels <strong>of</strong><br />
two patches are accumulated. Both gray-level proximity<br />
and geometric proximity are considered for similarity<br />
measurement, which helps to enhance the effectiveness <strong>of</strong><br />
the algorithm and robustness to noise. However, the<br />
similar patches in case <strong>of</strong> rotation and various scales can<br />
not be matched in this algorithm because the invariance<br />
property <strong>of</strong> a patch is not taken into account.<br />
III. THE PROPOSED METHOD<br />
A. Zernike Moment Based Image Representation<br />
Moment based image feature representation has a very<br />
wide range <strong>of</strong> applications in the field <strong>of</strong> image<br />
processing and pattern recognition. Zernike moments are<br />
proved to be superior to other moments in noise<br />
sensitivity, redundancy and expression efficiency for the<br />
property <strong>of</strong> orthogonality and invariance. Zernike<br />
Moments with orthogonal basis functions can be used to<br />
represent image features by a set <strong>of</strong> mutually independent<br />
descriptors, with a near zero value <strong>of</strong> information<br />
redundancy [16].<br />
The kernel <strong>of</strong> Zernike Moments is the set <strong>of</strong> orthogonal<br />
Zernike polynomials defined over a unit disk in the polar<br />
coordinate space. The Zernike basis function for order n<br />
and repetition m is<br />
V ( x, y) � V ( r, �) � R ( r)exp( jm�<br />
) , (8)<br />
nm nm nm<br />
© 2011 ACADEMY PUBLISHER<br />
where n is a positive integer or zero, and m is an integer<br />
subject to the following constraints: n- |m| = even and<br />
|m| ≤ n. In addition, θ and r is, respectively, the phase in<br />
polar coordinate space and the distance from point (x, y)<br />
to the origin. And j = � 1 .<br />
The radial polynomial Rnm is defined as<br />
( n�| m|)<br />
/ 2<br />
s<br />
( �1) ( n � s)!<br />
Rnm ( r) � �<br />
r<br />
s�<br />
0 � n� | m | � � n� | m | �<br />
s! � � s �! � � s �!<br />
� 2 � � 2 �<br />
n�2 s<br />
. (9)<br />
Given that f is a complex-valued function on the unit<br />
disk, the Zernike moment for f <strong>of</strong> order n and repetition m<br />
is<br />
n �1<br />
Z � �� f ( x, y) V ( x, y) dxdy , (10)<br />
nm<br />
� 2 2<br />
x � y �1<br />
*<br />
nm<br />
*<br />
where V nm is the complex conjugate <strong>of</strong> V nm .<br />
When f is a digital image, the Zernike moment means<br />
the projection <strong>of</strong> image f(x, y) on above orthogonal bases.<br />
Then (10) becomes<br />
n �1<br />
Z � �� f x y V x y . (11)<br />
*<br />
nm ( , ) nm ( , )<br />
� x y<br />
To reckon the Zernike moments for f (x, y), the image<br />
(or a patch) is first mapped to the unit disk <strong>of</strong> polar<br />
coordinates, moving the origin <strong>of</strong> the unit disc to the<br />
centre <strong>of</strong> the image. In this paper, a square-to-circular<br />
mapping transformation [16] is used so that the<br />
polynomials Rnm(r) need be computed only once for all<br />
pixels mapped to the same circle. Furthermore, fast<br />
computation for Rnm(r) in [17] is adopted to speed up the<br />
calculation.<br />
The magnitude <strong>of</strong> Zernike moment is rotation invariant<br />
as reflected in the mapping to the unit disc. The scale<br />
invariance can be achieved through normalizing the<br />
moments by the zero-order geometric moment [18].<br />
In terms <strong>of</strong> (11), Zernike moments <strong>of</strong> different orders<br />
are calculated with varying n; accordingly, for given n<br />
each moment <strong>of</strong> order n is computed with varying m. And<br />
moments <strong>of</strong> different orders correspond to independent<br />
characteristics <strong>of</strong> the image, which constitutes a multilevel<br />
representation for describing various shape features<br />
<strong>of</strong> the image. The magnitudes <strong>of</strong> these moments can be<br />
presented as images [15]. Fig. 1 shows an image and its<br />
Zernike moment images up to the third order. Fig. 1 (a) is<br />
the image “lena”. Fig. (b)-(g) are the moment images <strong>of</strong><br />
Z00, Z 11, Z 20, Z 22, Z 31 and Z 33, respectively. It can be<br />
seen that the lowest order moment Z 00 displays the main<br />
content <strong>of</strong> the image, the same as the average filtering<br />
result. And the higher moments deliver more detailed<br />
shape characteristics, but are also more sensitive to the<br />
noise. So Zernike moments <strong>of</strong> only up to third order are<br />
used in this paper.<br />
B. Collaborative SR Algorithm<br />
In this section, Zernike moments are first introduced as<br />
invariant features into the similarity measure strategy.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 969<br />
Figure1. The image with noise and its Zernike moment images <strong>of</strong> various orders.<br />
Figure 2. The original image (the left column) and the comparison <strong>of</strong> the weight distribution decided by three different methods.<br />
The Zernike moment based similarity measurement for<br />
SR reconstruction is proposed in this paper as<br />
w<br />
Z<br />
2<br />
�<br />
i, j nZ k ( s ) � k, l Zk ( Yt<br />
) �<br />
� � R M x R<br />
2 �<br />
k<br />
� exp �� 2<br />
� , (12)<br />
� � Z<br />
�<br />
� �<br />
where k ( s ) x Z and k ( t ) Y Z mean the k-th moment image<br />
for high resolution image xs and Yt. And Yt is the<br />
interpolation result <strong>of</strong> the measurement yt. σZ is the<br />
controlling parameter similar with σg. They can be<br />
roughly decided by estimation <strong>of</strong> noise from inputs. So<br />
when the weights are calculated according to (12), a<br />
Zernike moment based SR algorithm can be executed<br />
through (6).<br />
However, in the practical SR task, images may<br />
undergo complex motion and degradation, so their<br />
Zernike moments are usually not accurate enough to be<br />
the only basis for computation <strong>of</strong> weights, especially for<br />
the higher order moments that are more sensitive to the<br />
noise. Moreover, very inadequate information is rendered<br />
for the weak textures in the Zernike moment images,<br />
while the gray-levels <strong>of</strong> images contain the rich<br />
underlying details, and they are also more loyal to the<br />
original images. Thus, a collaborative algorithm is<br />
developed to combine Zernike moment features and the<br />
gray-level based proximity in (7) into the computation <strong>of</strong><br />
weights.<br />
© 2011 ACADEMY PUBLISHER<br />
The computation formula <strong>of</strong> the final weight for a<br />
searching pixel is designed as<br />
1 1<br />
w � wB � wZ � ( wB � wZ<br />
) , (13)<br />
2 4<br />
where B w is an abbreviation for Wn;t,s in (7). B w and Z w<br />
is calculated, respectively, according to (7) and (12). In<br />
the collaborative algorithm, the final weight is jointly<br />
determined by the moment images and bilateral proximity<br />
<strong>of</strong> gray-levels, which leads to a more accuracy calculation<br />
<strong>of</strong> weights in our method. This can be seen in Fig.2,<br />
where a comparison <strong>of</strong> the weight distribution with<br />
different algorithm is given. The 1st column in Fig. 2 is<br />
the original image without noise, the 2nd to the 4th<br />
columns are the weights distribution <strong>of</strong> the NLM<br />
algorithm, Zernike moment based algorithm and the<br />
collaborative algorithm. Differences between the right<br />
three columns in Fig. 2 show that the NLM algorithm<br />
strictly matches similar pixels only with translation. The<br />
moment based algorithm finds more pixels with similarity<br />
both in translation and rotation, but little difference is<br />
reflected for pixels with different similarity. The<br />
collaborative algorithm also can match similar pixels both<br />
with translation and rotation, but greater weights are<br />
assigned to pixels that are more similar. So the weights<br />
are more precise in comparison in the collaborative<br />
algorithm.
970 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
In addition, in order to improve the super-resolution<br />
algorithm, the following points are considered. Firstly, in<br />
our method whenever a pixel is SR estimated, the old<br />
value is replaced for the new one, which is closer to the<br />
true value than the old value. Hence, more accurate<br />
information is provided for computation <strong>of</strong> weights. That<br />
practically helps not only acquire more precise weight but<br />
also speed up the SR process. Secondly, in order to<br />
reinforce the reliability <strong>of</strong> Zernike moments in presence<br />
<strong>of</strong> noise, Yt in (12) is processed by a NLM denoising<br />
before computation <strong>of</strong> moments. Finally, the image <strong>of</strong><br />
moment Z00 viewed just as an average version <strong>of</strong> an<br />
image. Since gray-level information has been combined<br />
into the collaborative mechanism, the moment Z00 is not<br />
necessary any more. Hence, we set Z = {Zk | k = 1, 2, 3, 4,<br />
5} = {Z11, Z20, Z22, Z31, Z33} in our experiments to<br />
decrease the computation.<br />
Summarily, the proposed collaborative SR<br />
reconstruction for video sequences can be expressed in a<br />
simplified numerical algorithm. The iterative form <strong>of</strong> this<br />
numerical algorithm is represented as:<br />
T n<br />
1<br />
( , ) ( , ) 1 s, t y<br />
n�<br />
k l �N i j t�<br />
t<br />
s ( , ) � T n<br />
� ( k , l ) �N ( i, j) �t<br />
�1<br />
s, t<br />
X i j<br />
� �<br />
n n<br />
2<br />
R<br />
1 i, j X s � RS1k<br />
, S2l X<br />
n�<br />
t<br />
2<br />
B; s, t ( , , , ) exp<br />
2<br />
� g<br />
w k l i j<br />
w ( i, j, k, l) ( k, l)<br />
, (14)<br />
w ( i, j, k, l)<br />
� �<br />
� �<br />
� �� ��<br />
�<br />
�<br />
�<br />
�<br />
| i � S k | � | j � S l |<br />
1 2<br />
exp{ �<br />
}<br />
max(| i � S k | � | j � S l |)<br />
1 2<br />
, (15)<br />
5 � n n<br />
2 �<br />
� � Ri , jZ k (V s ) � RS1k<br />
, S (V )<br />
2lZ<br />
k t<br />
2<br />
n�1 �<br />
�<br />
k �1<br />
�<br />
Z; s, t ( , , , ) � exp �� 2<br />
�<br />
� Z<br />
w k l i j<br />
n<br />
where s, t<br />
� �<br />
�� ��<br />
,(16)<br />
w in (14) is computed according to (13).<br />
n T<br />
Especially, when n = 0, { X s } s�<br />
1 in (15) are obtained<br />
by the bilinear interpolation <strong>of</strong> { } 1<br />
T<br />
n T<br />
y s s� . And { Vt } t � 1 in<br />
n T<br />
(16) are the denoised results <strong>of</strong> { X t } t � 1 by NLM<br />
n n<br />
algorithm [8]. Otherwise, when n > 0, V t = X t . Both<br />
n<br />
n<br />
V t and X t are updated after each iteration.<br />
IV. EXPERIMENTS<br />
In this section, the performance <strong>of</strong> the proposed<br />
algorithm is validated. The obtained results <strong>of</strong> processing<br />
several real video sequences with a general motion<br />
pattern are presented. The comparison is provided with<br />
several methods: the bilinear and bicubic interpolation <strong>of</strong><br />
single image as well as the state-<strong>of</strong>-the-art SR algorithm<br />
[4]. The results are evaluated from both the visual effects<br />
and the objective quality measure (PSNR =<br />
© 2011 ACADEMY PUBLISHER<br />
10log10(<br />
255<br />
2<br />
N<br />
Xˆ � X<br />
2<br />
2<br />
) dB, where N is the number <strong>of</strong> pixels<br />
in the true image ˆ X or the constructed image X ).<br />
All the tests in this section were prepared in the<br />
following degradation: The input sequences were blurred<br />
using a 3×3 uniform kernel, down-sampled with a factor<br />
<strong>of</strong> 1:3 in each axis, and then added by additive white<br />
zero-mean Gaussian noise with std = 3. All images were<br />
in the input range [0,255]. In processing all the sequences,<br />
all 30 frames took part in the iterative reconstruction <strong>of</strong><br />
each image.<br />
First, sequences “Miss America”, “Trevor” and<br />
“Foreman” are tested for evaluation <strong>of</strong> PSNR. Table 1<br />
gives the average PSNR for each <strong>of</strong> the three test<br />
sequences, where two iterations were run for our method<br />
and the compared method in [4] with no additional<br />
deblurring followed. Fig. 3 illustrates the PSNR values<br />
frame by frame for every test sequence.<br />
In the experiments, the parameter σg and σZ was set<br />
manually to 2.5 and 2.4 for all the test sequences. The<br />
size <strong>of</strong> the patch used for calculating the weights wB and<br />
wZ was equally selected as 7 × 7 pixels (high resolution<br />
grid) for all sequences. The searching range for<br />
movement domain is 7 × 7 pixels (high resolution grid)<br />
for sequence “Miss America” and “Trevor”, and 19 ×19<br />
pixels (high resolution grid) for sequence “Foreman”,<br />
which has greater displacements between frames.<br />
The table shows that both the method in [4] and our<br />
method can handle sequences <strong>of</strong> arbitrary motion patterns<br />
and achieve effects <strong>of</strong> the state-<strong>of</strong>-the-art compared to the<br />
single image interpolation. And the proposed algorithm<br />
yields superior performance to all compared methods in<br />
PSNR.<br />
Then, an example on sequence “Mobile” with 30<br />
frames <strong>of</strong> 330 × 264 pixels is to reveal the visual effects<br />
for the proposed method and the compared methods. Fig.<br />
4 represents the selected two frames from the<br />
reconstructed results for bilinear interpolation, the<br />
method in [4] and the proposed method. The details are<br />
unfolded by their enlarged parts in Fig. 5. It can be seen<br />
that some numbers in the images, such as “15”, “16”,<br />
“18”, “19”, are obviously clearer in our results than<br />
others, due to that the invariant moments <strong>of</strong> Zernike are<br />
introduced, which improves the NLM algorithm.<br />
V. CONCLUSION AND DISCUSSION<br />
This paper proposes a SR algorithm using Zernike<br />
moments. The algorithm is based on the framework <strong>of</strong><br />
probabilistic motion estimation and needs no explicit<br />
motion estimation. A collaborative similarity measure<br />
strategy is developed in our algorithm to combine the<br />
Zernike moment based proximity and the bilateral<br />
proximity <strong>of</strong> NL-means algorithm. As representation <strong>of</strong><br />
image local invariant features, Zernike moments enable<br />
the algorithm to match more similar pixels not only with<br />
translation but also with rotation or scale. The<br />
collaborative mechanism ensures more suitable weights<br />
assigned to similar patches for better estimation
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 971<br />
© 2011 ACADEMY PUBLISHER<br />
TABLE I. MEAN-PSNR RESULTS FOR THREE TEST SEQUENCES WITH DIFFERENT METHODS.<br />
Sequence Bilinear Bicubic<br />
Protter et al.<br />
[4]<br />
Our method<br />
(1st Iteration)<br />
Our method<br />
(2nd Iteration)<br />
Miss America 33.91 34.31 35.31 35.65 35.87<br />
Trevor 29.42 29.79 30.39 30.58 30.76<br />
Foreman 28.38 28.89 29.57 29.65 29.84<br />
(a)Miss America<br />
(b)Trevor (c) Foreman<br />
Figure 3. The PSNR values <strong>of</strong> each frame reconstructed by different methods<br />
(a) (b) (c)
972 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
(d) (e) (f)<br />
Figure 4. Results for the 9th (the top row) and 14th (the bottom row) frames from “Mobile” sequence. From left to right: bilinear interpolation; the<br />
method in [4]; the proposed method.<br />
<strong>of</strong> images. Experimental results demonstrate that the<br />
proposed algorithm is able to process real video<br />
sequences with general motion patterns with<br />
improvements both in PSNR and the visual effects<br />
compared to the stare-<strong>of</strong>-the-art algorithm.<br />
Several aspects <strong>of</strong> the proposed method may be further<br />
studied to improve the algorithm. Firstly, parameters σB<br />
and σZ reflecting the size <strong>of</strong> the noise are constant during<br />
the iterations. Since noise becomes smaller in the later<br />
reconstruction, it should be reasonable to decrease<br />
parameters σB and σZ appropriately with the increased<br />
iteration. Secondly, Zernike moment used in our method<br />
may be replaced by the Pseudo-Zernike moment, which is<br />
proved able to represent image details better with lower<br />
orders. And fast computation has been proposed to<br />
directly calculate arbitrary Pseudo-Zernike moment with<br />
high order without lower order moments first computed.<br />
Thus, to use Pseudo-Zernike moment instead may help to<br />
improve the performance without computation increased.<br />
Finally, several important parameters in our algorithm are<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 5. Enlarged parts <strong>of</strong> images in Fig. 4.<br />
manually selected in the experiments, such as σB, σZ, and<br />
the searching size for the motion domain. An adaptive<br />
selection strategy may be developed to improve the<br />
algorithm in the future work.<br />
REFERENCES<br />
[1] S. Park, M. Park, M. G. Kang, “Super-resolution image<br />
reconstruction: a technical review,” IEEE Signal<br />
Processing Magazine, vol. 20, pp. 21–36, May 2003.<br />
[2] S. Farsiu, D. Robinson, M. Elad, and P. Milanfar,<br />
“Advances and challenges in superresolution,” Int. J. Imag.<br />
Syst. Technol., vol. 14, pp. 47–57, August 2004.<br />
[3] W. Shao, Z. Hui, “Edge-and-corner preserving<br />
regularization for image interpolation and reconstruction,”<br />
Image and Vision Computing, vol. 26, pp. 1591–1606,<br />
2008.<br />
[4] M. Protter, M. Elad, H. Takeda, et al., “Generalizing the<br />
nonlocal-means to super-resolution reconstruction,” IEEE<br />
Trans. Image Process., vol. 18, pp. 36–51, January 2009.<br />
[5] H. Takeda, P. Milanfar, M. Protter, et al., “Superresolution<br />
without explicit subpixel motion estimation,”
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 973<br />
IEEE Trans. Image Process., vol. 18, pp. 1958-1975,<br />
September 2009.<br />
[6] A. Danielyan, A. Foi, V. Katkovnik, et al., “Image and<br />
video super-resolution via spatially adaptive blockmatching<br />
filtering,” in Proc. Int. Workshop Local and Non-<br />
Local Approx. Image Process., Switzerland, August 2008.<br />
[7] M. Protter, M. Elad, “Super resolution with probabilistic<br />
motion estimation,” IEEE Tran. Image Process., vol. 18,<br />
pp. 1899–1904, August 2009.<br />
[8] A. Buades, M. Morel, “A review <strong>of</strong> image denoising<br />
algorithms with a new one,” Mutiscale Model Simul. vol. 4,<br />
pp. 490–530, February 2005.<br />
[9] S. Alexander, E. Vrscay, and S. Tsurumi, “A simple,<br />
general model for the affine selfsimilarity <strong>of</strong> images,” in<br />
Proc. Int. Conf. on Image Analysis and Recognition,<br />
Lecture Notes in Comput. Sci. 5112, Springer-Verlag,<br />
Berlin, 2008, pp. 192–203.<br />
[10] G. Peyre, “Sparse modeling <strong>of</strong> textures,” J. Math. Imaging<br />
Vision, vol. 34, pp. 17–31, 2009.<br />
[11] M. Ebrahimi and E. Vrscay, “Examining the role <strong>of</strong> scale<br />
in the context <strong>of</strong> the non-local-means filter,” in Image<br />
Analysis and Recognition, Lecture Notes in Comput. Sci.<br />
4633, Springer-Verlag, Berlin, 2008, pp. 170–181.<br />
[12] A. Buades, B. Coll, J. Morel, “Image denoising methods a<br />
new nonlocal principle,” SIAM Review, vol. 52, pp. 113–<br />
147, Jan. 2010.<br />
[13] Y. Lou, P. Favaro, S. Soatto, “Nonlocal similarity image<br />
filtering,” in Reports CAM (8–26), 2008.<br />
[14] S. Zimmer, S. Didas and J. Weickert, “A rotationally<br />
invariant block matching strategy improving image<br />
denoising with non-local means,” in Proc. <strong>of</strong> the Int.<br />
© 2011 ACADEMY PUBLISHER<br />
Workshop on Local and Non-local Approximation in<br />
Image Processing, pp. 135–142, 2008.<br />
[15] Z. Ji, Q. Chen, Q. Sun and D. Xia, “A moment-based<br />
nonlocal-means algorithm for image denoising,”<br />
Information Processing Letters, vol. 109, pp. 1238–1244.<br />
September 2009.<br />
[16] R. Mukundan and K. Ramakrishnan, “Fast computation <strong>of</strong><br />
legendre and zernike moments,” Pattern Recognition, vol.<br />
28, pp. 1433–1442, September 1995.<br />
[17] C. Chong, P. Raveendran and R. Mukundan, “A<br />
comparative analysis <strong>of</strong> algorithm for fast computation <strong>of</strong><br />
Zernike moments,” Pattern Recognition, vol. 36, pp. 731–<br />
742, 2003.<br />
[18] B. Ye and J. Peng, “Invariance analysis <strong>of</strong> improved<br />
Zernike moments,” <strong>Journal</strong> <strong>of</strong> Optics A: Pure and Applied<br />
Optics. vol. 4, pp. 606–614, 2002.<br />
Lin Guo was born in Hubei, China, in 1978. She is currently<br />
pursuing the Ph.D. degree in Wuhan University, Wuhan, China.<br />
Her research interests include super-resolution reconstruction,<br />
image processing and computer vision.<br />
Qinghu Chen was born in Hubei, China, in 1957. He<br />
currently is a pr<strong>of</strong>essor in School <strong>of</strong> Electronic Information at<br />
Wuhan University, Wuhan, China. His main research interests<br />
include image processing and intelligent recognition.
974 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Mathematical Model and Hybrid Scatter Search<br />
for Cost Driven Job-shop Scheduling Problem<br />
Bai Jie<br />
Department <strong>of</strong> Automation, Shanghai Jiao Tong University, Shanghai, China<br />
Email: baijie@sjtu.edu.cn<br />
Sun Kai*<br />
Shandong Provincial Key Laboratory <strong>of</strong> AM&MC Technology for Light Industry Equipment,<br />
Shandong Institute <strong>of</strong> Light Industry, Shandong, China<br />
Email: sunkai79@gmail.com<br />
Yang Gen Ke<br />
Department <strong>of</strong> Automation, Shanghai Jiao Tong University, Shanghai, China<br />
Email: gkyang@sjtu.edu.cn<br />
Abstract— Job-shop scheduling problem (JSP) is one <strong>of</strong> the<br />
most well-known machine scheduling problems and one <strong>of</strong><br />
the strongly NP-hard combinatorial optimization problems.<br />
Cost optimization is an attractive and critical research and<br />
development area for both academic and industrial societies.<br />
This paper presents a cost driven model <strong>of</strong> the job-shop<br />
scheduling problem in which the solutions are driven by<br />
business inputs, such as the cost <strong>of</strong> the product transitions,<br />
revenue loss due to the machine idle time and<br />
earliness/tardiness penalty. And then, a new hybrid scatter<br />
search algorithm is proposed to solve the cost driven jobshop<br />
scheduling problem by introducing the simulated<br />
annealing (SA) into the improvement method <strong>of</strong> scatter<br />
search (SS). In order to illustrate the effectiveness <strong>of</strong> the<br />
hybrid method, some test problems are generated, and the<br />
performance <strong>of</strong> the proposed method is compared with<br />
other evolutionary algorithms such as genetic algorithm and<br />
simulated annealing. The experimental simulation tests<br />
show that the hybrid method is quite effective at solving the<br />
cost driven job-shop scheduling problem.<br />
Index Terms—cost optimization, job-shop scheduling<br />
problem, scatter search, simulated annealing<br />
I. INTRODUCTION<br />
The job-shop scheduling problem (JSP) is one <strong>of</strong> the<br />
most well-known machine scheduling problems and one<br />
<strong>of</strong> the strongly NP-hard combinatorial optimization<br />
problems [1]. Historically, JSP was primarily solved by<br />
the branch-and-bound method and some heuristic<br />
procedures based on priority rules [2]. During the past<br />
decade, researches on meta-heuristic methods to solve the<br />
JSP have been widely studied, such as genetic algorithm<br />
[3],[4], simulated annealing [5], tabu search [6] and<br />
Manuscript received Sep 5, 2010; revised Nov 6, 2010; accepted Dec<br />
15, 2010.<br />
The project is supported by Shandong Provincial Natural Science<br />
Foundation, China (No.ZR2010FQ009) and the National Nature<br />
Science Foundation <strong>of</strong> China (No. 61074150)<br />
*Corresponding author<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.974-981<br />
particle swarm optimization [7]. The majority <strong>of</strong> studies<br />
on JSP, however, are driven by production criteria, such<br />
as total flowtime, maximum complete time (makespan),<br />
maximum tardiness and number <strong>of</strong> tardy jobs, etc.<br />
In nowdays, the critical challenge to manufacturing<br />
enterprise is to become more flexible and pr<strong>of</strong>itable [8].<br />
The “goal” <strong>of</strong> a manufacturing enterprise is not only to<br />
apply advanced technology, but consistently to make<br />
money (i.e., pr<strong>of</strong>its), as discussed in the book The Goal<br />
[9]. The development and application <strong>of</strong> an appropriate<br />
scheduling solution with seamless integration <strong>of</strong> business<br />
and manufacturing play a critical role in any modern<br />
manufacturer achieving this goal [10].<br />
Jiang et al. [11] present a cost driven objective<br />
function for job-shop scheduling problem and solve it by<br />
using genetic algorithm, and the experimental results<br />
demonstrate the effectiveness <strong>of</strong> the algorithm. However,<br />
the study didn’t provide precise mathematical model for<br />
the cost driven job-shop scheduling problem. The major<br />
contributions <strong>of</strong> this paper are summarized as follows:<br />
(1) Based on the characteristics <strong>of</strong> general JSP, a<br />
mathematical model <strong>of</strong> cost driven JSP is presented. The<br />
cost <strong>of</strong> operational transitions between products, the<br />
revenue loss due to machine idle time during the phase <strong>of</strong><br />
product transitions, and the penalty due to missing the<br />
required on-time delivery date, and so on are included in<br />
the model.<br />
(2) A new hybrid evolutionary algorithm, which<br />
combines the strong global search ability <strong>of</strong> scatter search<br />
(SS) with the strong local search ability <strong>of</strong> simulated<br />
annealing (SA), is developed to solve the cost driven JSP.<br />
The organization <strong>of</strong> remain contents is as follows. In<br />
section 2, the formulation and model <strong>of</strong> the cost driven<br />
JSP is presented. Section 3 presents the conceptual<br />
introduction to SS and SA and proposes the hybrid scatter<br />
search (HSS) algorithm to solve the cost driven JSP.<br />
Section 4 provides experimental results and performance<br />
analyses. Section 5 <strong>of</strong>fers concluding remarks.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 975<br />
II. PROBLEM STATEMENT<br />
The JSP may be formulated as follows. There are n<br />
jobs that plan to process on m machines. Each job<br />
involves a set <strong>of</strong> operations, which are performed on the<br />
machines in a pre-specified order. Each machine can<br />
process only one job at a time, and it cannot be<br />
interrupted. Furthermore, the processing time is fixed and<br />
known. The objective widely used is to find a sequence <strong>of</strong><br />
jobs to minimize makespan or total weighted complete<br />
time[1-7]. In the study, we focus on developing a cost<br />
driven model for JSP in order to make manufacturing<br />
enterprises more pr<strong>of</strong>itable.<br />
The multiple job sequences are defined as decision (or<br />
manipulate) variables. The illustration <strong>of</strong> the costs<br />
generated from a scheduling scenario is schematically<br />
shown in Figure 1.<br />
The key ideas behind the developed scheduling system<br />
are constrained cost driven optimization solution, namely,<br />
the cost is defined as an objective function subject to the<br />
relevant constraints. In the paper, the cost <strong>of</strong> the product<br />
transitions, revenue loss due to the machine idle time and<br />
earliness/tardiness penalty are included in the cost model<br />
for optimizing production scheduling solutions.<br />
Figure 1. Cost description <strong>of</strong> job-shop production line<br />
In order to formulate the cost driven job-shop<br />
scheduling problem, we introduce a null job (job 0)<br />
whose processing time and transition cost with other job<br />
are zero and all sequences on each machine are started<br />
from job 0 and ended at job 0. And then we define the<br />
parameter and decision variables as follows.<br />
Paremeters:<br />
i, j : index <strong>of</strong> jobs, and i , j � 0 ~ n<br />
k : index <strong>of</strong> machines, and k � 1 ~ m<br />
t ijk : sequence-dependent transition cost from job i<br />
to job j on machine k<br />
Ci : flowtime <strong>of</strong> job i<br />
p : processing time <strong>of</strong> job i on machine k<br />
ik<br />
C ik : complete time <strong>of</strong> job i on machine k<br />
W k : machine waiting cost <strong>of</strong> machine k per unit<br />
time during job transfer<br />
ei: earliness due date <strong>of</strong> job i<br />
di: tardiness due date time <strong>of</strong> job i<br />
αi: unit time earliness penalties <strong>of</strong> job i<br />
βi: unit time tardiness penalties <strong>of</strong> job i<br />
T (S)<br />
: total transition cost <strong>of</strong> a schedule scenario S<br />
R (S ) : total revenue loss <strong>of</strong> a schedule scenario S<br />
© 2011 ACADEMY PUBLISHER<br />
E (S ) : total earliness/tardiness penalty <strong>of</strong> a schedule<br />
scenario S<br />
Decision variable:<br />
�1<br />
if job j is preceded by job i on machine k<br />
X ijk � �<br />
(1)<br />
�0<br />
else<br />
The precedence coefficient Y ijk is defined as:<br />
�1<br />
if machine k process job i right after machine h<br />
Yihk � �<br />
(2)<br />
�0<br />
else<br />
A. Transition cost<br />
Transition includes work to prepare the machine,<br />
process, or bench for product parts or the cycle. This<br />
includes obtaining tools, positioning work in process<br />
materials, return tooling, cleanup, setting the required jigs<br />
and fixtures, adjusting tools, and inspecting material [12].<br />
Sule and Huang [13] described the activities typically<br />
associated with sequence-dependent and sequenceindependent<br />
operations in machine shop environments.<br />
Transition cost is comprised <strong>of</strong> labor cost for setup<br />
operation and pr<strong>of</strong>it loss during machine idle time.<br />
Let T (S)<br />
denote the total transition cost <strong>of</strong> a schedule<br />
scenario S, the total transition costs <strong>of</strong> the solution S can<br />
be calculated by the following formula:<br />
B. Revenue loss<br />
m<br />
n<br />
n<br />
���<br />
�<br />
ijk � ijk X t<br />
S T(<br />
)<br />
(3)<br />
k�1<br />
i�1<br />
j�1<br />
During the work process <strong>of</strong> JSP, each <strong>of</strong> the jobs is to<br />
be sequentially processed on machine 1~m. If a machine<br />
has finished processing a job, while the next job is being<br />
processed on the previous machine, the machine will be<br />
idle and waiting for the next job. The idle time <strong>of</strong> the<br />
machine will cause revenue loss <strong>of</strong> manpower and fixed<br />
assets for the manufacturer to maintain the ready state <strong>of</strong><br />
the machine. Less machine idle time means higher<br />
utilization <strong>of</strong> manpower and machines. The revenue loss<br />
can be calculated by the following formula:<br />
m<br />
n<br />
n<br />
���<br />
R(<br />
S)<br />
� ( C jk � Cik<br />
� pik<br />
) �Wk<br />
� X ijk (4)<br />
k �1 i�1<br />
j�1<br />
C. Earliness/tardiness penalty<br />
In the proposed cost driven JSP, every job has a duedate<br />
window. Any job completed prior to ei will incur an<br />
earliness penalty, on the other hand, any job completed<br />
after di will incur a tardiness penalty. No penalty,<br />
however, will be incurred if any job can be completed<br />
within the time window [ e i , i d ]. Let αi and βi, which can<br />
be determined by the inventory carrying cost and<br />
tardiness compensation, be the unit time earliness and<br />
tardiness penalties for job i. The earliness/tardiness<br />
penalty E can be calculated as:
976 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
n<br />
�<br />
i �1<br />
E(<br />
S)<br />
� [ � � max( 0,<br />
e � C ) � � � max( 0,<br />
C � d )] (5)<br />
i<br />
D. Mathematical model <strong>of</strong> cost driven JSP<br />
i<br />
Based on the above discussion <strong>of</strong> the problem, the<br />
mathematical model <strong>of</strong> cost driven JSP can be formulated<br />
as:<br />
m n n<br />
m n n<br />
Min{<br />
��� tijk<br />
� Xijk<br />
����<br />
( C jk � Cik<br />
� pik<br />
) �Wk<br />
� Xijk<br />
k �1 i�1<br />
j�1<br />
k �1 i�1<br />
j�1<br />
Subject to:<br />
jk<br />
n<br />
��[<br />
�i<br />
� max( 0,<br />
ei<br />
� Ci)<br />
� �i<br />
� max( 0,<br />
Ci<br />
� di<br />
) ]}<br />
i�1<br />
n<br />
� X ijk<br />
j�0<br />
n<br />
� X ijk<br />
i�0<br />
jk<br />
i<br />
i<br />
i<br />
i<br />
(6)<br />
� 1 i � j , i � 1,...<br />
n , k � 1,...,<br />
m (7)<br />
� 1 j � i , j � 1,...<br />
n , k � 1,...,<br />
m (8)<br />
C � p � ( 1�<br />
X ) M � C n<br />
Cik � pik<br />
� ( 1�<br />
Yihk<br />
) M � Cih<br />
i 1,...<br />
n<br />
ijk<br />
ik i , j � 1,...<br />
, k 1,...,<br />
m<br />
� (9)<br />
� , h , k � 1,...,<br />
m (10)<br />
X ijk �{<br />
0,<br />
1}<br />
i , j � 0,<br />
1,...,<br />
n , i � j , k � 1,...,<br />
m (11)<br />
Constraint (7) defines that a job should be right before<br />
another job on each machine. Constraint (8) denotes that<br />
a job should be right after another job on each machine.<br />
Constraint (9), where M is a very large positive number,<br />
shows that each machine can process at most one job at<br />
any time. Constraint (10) shows that each job can be<br />
processed on at most one machine at any time. Constraint<br />
(11) ensures that the variable only takes the integer 0 or 1.<br />
Obviously, the proposed JSP is a hard, constrained,<br />
m<br />
combinatorial optimization problem with ( n ! ) possible<br />
solutions for n-job m-machine problems. Even for a<br />
simple case with ten jobs and ten machines, the<br />
computation time for complete enumeration <strong>of</strong> all<br />
possible solutions is quite large [1]. If we have a large<br />
number <strong>of</strong> jobs and machines, e.g., thirty jobs and fifty<br />
machines, the complete enumeration <strong>of</strong> all possible<br />
solutions is computationally prohibitive, i.e., no exact<br />
algorithm is capable <strong>of</strong> solving the optimization problem<br />
in a reasonable computation time. Frequently,<br />
evolutionary algorithms as promising approximate<br />
techniques, such as genetic algorithm [3],[4] and<br />
simulated annealing [5], are employed to solve the<br />
scheduling problem <strong>of</strong> finding a desirable, although not<br />
necessarily, optimal solution.<br />
III. HYBRID SCATTER SEARCH FOR JSP<br />
© 2011 ACADEMY PUBLISHER<br />
Scatter search (SS) is an evolutionary method that has<br />
been successfully applied to hard optimization problems<br />
[14-17]. Unlike genetic algorithm, a scatter search<br />
operates on a small set <strong>of</strong> solutions (called reference set)<br />
and makes only limited use <strong>of</strong> randomization as a proxy<br />
for diversification when searching for a globally optimal<br />
solution. Glover [14] proposed a template to serve as the<br />
guideline <strong>of</strong> implementing SS to solve combinational<br />
optimization problems. The template consists <strong>of</strong> five<br />
components which are diversification generation method,<br />
improvement method, reference set update method,<br />
subset generation method and solution combination<br />
method. A template <strong>of</strong> the standard SS can be shown as<br />
follows:<br />
Step 1: Initialization<br />
1.1 Use the diversification generation method to<br />
generate diverse trial solutions.<br />
1.2 Apply the improvement method to enhanced<br />
trial solutions.<br />
1.3 Apply the reference update method to build<br />
the initial reference set (RefSet) from the<br />
enhanced trial solutions.<br />
Step 2: Computation<br />
Do<br />
2.1 Generate subsets <strong>of</strong> RefSet with the subset<br />
generation method.<br />
2.2 Use solution combination method to combines<br />
these subsets and obtain new solutions.<br />
2.3 Use the improvement method to enhance these<br />
new trial solutions.<br />
2.4 Apply the reference update method to update<br />
the RefSet.<br />
While (the maximum generation is not meet)<br />
Step 3: Output the optimization results.<br />
In the template, five key components <strong>of</strong> scatter search<br />
are labeled in bold font. The computation flow chart <strong>of</strong><br />
scatter search can be shown in Figure 2.<br />
Figure 2. Computation flow chart <strong>of</strong> scatter search
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 977<br />
The paper develop a new hybrid scatter search, which<br />
introduces the simulated annealing into the improvement<br />
method to enhance the local search ability <strong>of</strong> scatter<br />
search, to solve the cost driven job-shop scheduling<br />
problem. The detailed discussion <strong>of</strong> applying hybrid<br />
scatter search is shown as follows.<br />
A. Encoding scheme and fitness function<br />
One <strong>of</strong> the key issues in applying SS successfully to<br />
cost driven JSP is how to encode a schedule <strong>of</strong> the<br />
problem to a search solution. We utilize an operationbased<br />
representation [3] that uses an unpartitioned<br />
permutation with m -repetitions <strong>of</strong> job numbers. A job is<br />
a set <strong>of</strong> operations that has to be scheduled on m<br />
machines. In this formulation, each job number occurs m<br />
times in the permutation, i.e. as <strong>of</strong>ten as there are<br />
operations associated with this job. By scanning the<br />
permutation from left to right, the kth occurrence <strong>of</strong> a<br />
job number refers to the kth operation in the<br />
technological sequence <strong>of</strong> this job. A permutation with<br />
repetition job numbers merely expressed the order in<br />
which the operations <strong>of</strong> jobs are scheduled.<br />
For example, suppose a solution is given as {3 2 3 4 2<br />
4 2 1 1 3 2 3 1 4 4 1} in 4 jobs and 4 machines problem.<br />
Each job consists <strong>of</strong> three operations, and is thereby<br />
repeated four times. Third number <strong>of</strong> solution in this<br />
example is 3. Here, 3 implies second operation <strong>of</strong> job 3<br />
because number 3 has been repeated twice.<br />
Fitness function is used to evaluate the performance <strong>of</strong><br />
solutions. In the paper, the objective function <strong>of</strong> solution<br />
S is defined as:<br />
m n n<br />
m n n<br />
Fit(<br />
S)<br />
����<br />
tijk<br />
� X ijk ����<br />
( C jk � Cik<br />
� pik<br />
) �Wk<br />
� X ijk<br />
k�1<br />
i�1<br />
j�1<br />
k�1<br />
i�1<br />
j�1<br />
n<br />
��<br />
[ � i � max( 0,<br />
ei<br />
� Ci<br />
) � �i<br />
� max( 0,<br />
Ci<br />
� di<br />
) ] � M � feaS<br />
i�1<br />
�0<br />
where feaS � �<br />
�1<br />
positive number.<br />
B. Diversification generation method<br />
(12)<br />
if<br />
S is feasible<br />
and M is a very large<br />
otherwise<br />
The diversification generation method is used to<br />
generate a collection <strong>of</strong> diverse trial solutions, using an<br />
arbitrary trial solution (or seed solution) as an input. This<br />
element <strong>of</strong> the SS approach is particularly important,<br />
given the goal <strong>of</strong> developing a method that balances<br />
diversification and intensification in the search. This<br />
method was suggested by Glover [14], which generates<br />
diversified permutations in a systematic way without<br />
reference to the objective function.<br />
Assume that there are a n� m JSP, a given trial solution<br />
S used as a seed is representing by indexing its elements,<br />
so that they appears in consecutive order to yield<br />
S � {[ 1],<br />
[ 2],...,<br />
[ l]}<br />
, where l � m � n . Define the<br />
subsequence S ( h : t)<br />
, where t is a positive integer<br />
between 1 and h , to be given by:<br />
S( h : t)<br />
� ([ t],<br />
[ t � h],<br />
[ t � 2h],....,<br />
[ t � rh])<br />
,<br />
© 2011 ACADEMY PUBLISHER<br />
where r is the largest nonnegative integer such that<br />
t � rh � l . Then define the S (h)<br />
for h � n , to be:<br />
S( h)<br />
� { S(<br />
h : h),<br />
S(<br />
h,<br />
h �1),...,<br />
S(<br />
h : 1)}<br />
.<br />
To illustrate the strategy, suppose S is given by:<br />
S={[1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11],<br />
[12], [13], [14], [15],[16]}.<br />
If we choose h=4, then<br />
S(4:4)={[4], [8], [12],[16]},<br />
S(4:3)={[3], [7], [11], [15]},<br />
S(4:2)={[2], [6], [10], [14]},<br />
S(4:1)={[1], [5], [9], [13]},<br />
to give:<br />
S(4)={(4:4), (4:3), (4:2), (4:1)}<br />
={[4], [8], [12],[16] ,[3], [7], [11], [15], [2], [6], [10],<br />
[14], [1], [5], [9], [13]}.<br />
In this illustration, we have allowed h to take the value<br />
closest the square root <strong>of</strong> l. The value is interesting based<br />
on the fact that, when h equals the square root <strong>of</strong> l, the<br />
minimum relative separation <strong>of</strong> each element from each<br />
other in the new permutation is maximum, compared to<br />
the relative separation <strong>of</strong> exactly 1 in the trial solution S<br />
[14]. In general, for the goal <strong>of</strong> generating a diverse set <strong>of</strong><br />
trial solutions, preferable values for h range from 1 to l 2 .<br />
C. Improvement method<br />
Each <strong>of</strong> the new trial solutions which are obtained<br />
from the diversification generation method or solution<br />
combination method is subjected to the improvement<br />
method. This method aims to enhance the quality <strong>of</strong> these<br />
solutions. In the paper, we take two versions <strong>of</strong> local<br />
search meta-heuristics to improve trial solutions. A longterm<br />
SA-based improvement method is only applied to<br />
the best new trial solution, and a short-term swap-based<br />
local search is taken to enhance other new trial solutions.<br />
With the hybridization <strong>of</strong> these two local methods, we<br />
can get a compromise between solution quality and<br />
computational effort.<br />
C.1 Swap -based local search method<br />
A Swap-based local search method is taken to improve<br />
methods for HSS. In the method, swap operator is<br />
adopted to obtain neighbors.<br />
Swap operator: Let [i ] and [ j ] be two randomly<br />
selected positions whose job numbers are different in the<br />
trial solution S. A neighborhood <strong>of</strong> S is obtained by<br />
interchanging the job in position [i ] and [ j ] .<br />
For each new trial solution, the local search method<br />
takes it as the initial solution, and then searches in its<br />
neighborhood until there is no improvement. If the local<br />
search yields a better value than the one from the original<br />
solution, the new solution will replace the original<br />
solution. If no improvement has been found after the<br />
local search, no replacement will be made.<br />
C.2 SA-based improvement method<br />
Ever since it was introduced by Kirkpatrick [18], the<br />
SA algorithm has been applied to many combinatorial<br />
optimization problems. The SA approach can be<br />
interpreted as an enhanced version <strong>of</strong> local search or<br />
iterative improvement, which can avoid being trapped in
978 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
local minima by probabilistic jumping. In the paper, a<br />
SA-based local search method is developed as a longterm<br />
improvement method for HSS.<br />
In the SA-based local search method, swap operator<br />
described above is adopted to obtain neighbors. The new<br />
solution is accepted if the objective function is improved.<br />
Otherwise, the new solution is accepted with probability<br />
exp( � � / T ) , where � is the change <strong>of</strong> the objective<br />
function value and T is a control parameter.<br />
SA process can he controlled by the cooling schedule.<br />
The selection <strong>of</strong> initial temperature, cooling rate,<br />
termination temperature and temperature length<br />
influences the quality <strong>of</strong> the solutions. In SA optimization<br />
process, the temperature is gradually reduced. It is well<br />
known that specifies temperature with the equation<br />
Tk=λ·Tk-1 is <strong>of</strong>ten a good choice and it can provide a<br />
trade<strong>of</strong>f between computation time and good solutions.<br />
Tend is chosen to terminate the SA process, when the<br />
current temperature T
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 979<br />
Similarly, filling in the holes <strong>of</strong> string 2 with<br />
unselected operations <strong>of</strong> original trial solution 1, we can<br />
get:<br />
New trial solution 2: 3 2 3 1 4 2 2 1 3 4 4 4 1 3 2 1<br />
IV. COMPUTATIONAL EXPERIMENTS<br />
A. Instance data generation<br />
To illustrate the effectiveness <strong>of</strong> the algorithm<br />
described in this paper, we consider several instances<br />
originated from two classes <strong>of</strong> standard JSP test problems:<br />
Fisher and Thompson instances (FT06, FT10, FT20) [19],<br />
and Lawrence instances (LA01, LA02,…,LA31) [20].<br />
Originally, each instance only consists <strong>of</strong> the machine<br />
number and processing time ( p ik ) for each step <strong>of</strong> the<br />
job. Furthermore, to apply these instances to the cost<br />
driven JSP presented in this paper, some extra instance<br />
data should be generated.<br />
In this paper, a method with lower and upper bound on<br />
including transition cost ( t ijk ), machine waiting cost<br />
per unit time( k W ), earliness/tardiness penalties( i � , � i )<br />
and random selection is used to generate extra data <strong>of</strong><br />
problem instances. Table I shows the bounds used for<br />
problem data.<br />
TABLE I.<br />
LOWER AND UPPER BOUND FOR PROBLEM DATA<br />
Problem data k W t ijk i � i �<br />
Lower bound 1 1 1 1<br />
Upper bound 10 10 5 10<br />
The due date [ e i , d i ] <strong>of</strong> problem can be generated by<br />
the following formula:<br />
�<br />
�ei<br />
� 0.<br />
5�<br />
C<br />
�<br />
�<br />
�<br />
�<br />
�di<br />
� 0.<br />
5�<br />
C<br />
��<br />
max<br />
max<br />
� � �<br />
� � �<br />
m<br />
�<br />
k �1<br />
m<br />
�<br />
k �1<br />
p<br />
ik<br />
p<br />
ik<br />
(14)<br />
where C max is best makespan known so far, δ and σ<br />
are tightness factors <strong>of</strong> due dates, and we can obtain<br />
proper δ and σ for each instance by several runs.<br />
B. Experimental setup<br />
The algorithm for cost driven JSP mentioned above was<br />
programmed in Borland C++ and the experiments were<br />
executed on a Intel Pentium 2.8G with 512M RAM. The<br />
parameters <strong>of</strong> HSS use the following configuration:<br />
Parameters for SS:<br />
The number <strong>of</strong> trial solutions in Refset1 ( b 1 ) equals the<br />
number <strong>of</strong> jobs in the problem, and so is that in RefSet2.<br />
And after n� m generations for n-job m-machine problem,<br />
the algorithm is terminated.<br />
Parameters for SA-subprogram:<br />
© 2011 ACADEMY PUBLISHER<br />
Initial temperature T 0 in this paper is set by T0 � �fmax<br />
,<br />
where �fmax is the maximal difference in fitness value<br />
between any two neighboring solutions. It should be<br />
adjusted experimentally. Epoch length L is set to the<br />
number <strong>of</strong> ( n � 1)<br />
� m . Decreasing rate � is set as values<br />
0.98. Termination temperature T end is equal to 0.1. SA<br />
sub-program is terminated whenever there is no<br />
improvement in 20 successive generations, which enables<br />
a reduction in running time.<br />
C. Simulation results and comparison<br />
In order to illustrate the effectiveness <strong>of</strong> the hybrid<br />
method, we compare the proposed method with other<br />
evolutionary algorithms such as GA [11], SA and SS. The<br />
GA parameters: population size is set equal to HSS,<br />
crossover probability is 0.95, mutation probability is 0.05.<br />
The termination conditions <strong>of</strong> these algorithms are set<br />
equal to the HSS’s CPU time. The statistical performance<br />
<strong>of</strong> 20 independent runs <strong>of</strong> these algorithms are listed in<br />
table 2, including the optimum value known so far(BKS),<br />
the best objective value ( C * ) , the percentage value <strong>of</strong><br />
average objective value over BKS (%) and the average<br />
CPU time ( t ) <strong>of</strong> the HSS.<br />
Table II shows that the results obtained by HSS are<br />
much better than those obtained by GA, SA and SS. The<br />
superiority <strong>of</strong> the best optimization quality demonstrates<br />
the effectiveness and the global search property <strong>of</strong> the<br />
hybrid search, and the superiority <strong>of</strong> the average<br />
performance over 20 random runs shows that the hybrid<br />
probabilistic search is more robust than these algorithms.<br />
It can be seen that the HSS algorithm can get desirable<br />
solutions in a reasonable computation time even for<br />
problems with 30 jobs and 10 machines. For more largescale<br />
problems, we can trade <strong>of</strong>f between computation<br />
time and solution quality by adjusting the number <strong>of</strong><br />
generations or the parameters <strong>of</strong> SA-subprogram.<br />
V. CONCLUSION<br />
Cost optimization is an attractive and critical research<br />
and development area for both academic and industrial<br />
societies. This is also a multi-disciplinary subject with<br />
optimization, control, business intelligence, computer<br />
science and operation research, and so on. In the cost<br />
driven JSP model we proposed, the solutions are driven<br />
by business inputs, such as market demand and the costs<br />
<strong>of</strong> inventory and machine idle time during the product<br />
transition phase. And then, a hybrid optimization<br />
algorithm that combines SS with SA is proposed to solve<br />
the problem. This hybrid method combines the<br />
advantages <strong>of</strong> these two algorithms and mitigates the<br />
disadvantages <strong>of</strong> them. The obtained results indicate that<br />
this hybrid method is superior to GA, SA and SS, and is<br />
an effective approach for the cost driven JSP.
980 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Inst. Size BKS<br />
TABLE II.<br />
EXPERIMENT RESULTS.<br />
GA[11] SA SS HSS<br />
C * % C * % C * % C * % t<br />
FT06 6� 6 426 426 0 426 0 426 0 426 0 1.85<br />
FT10 10� 10 2732 2741 4.05 2848 6.53 2732 2.85 2732 1.61 45.21<br />
FT20 20� 5 2442 2495 2.32 2504 1.74 2491 1.45 2442 0.92 42.92<br />
LA01 10� 5 935 935 1.26 935 0 935 0 935 0 5.26<br />
LA02 10� 5 1028 1028 1.41 1028 0 1028 0 1028 0 5.90<br />
LA06 15� 5 1670 1670 1.22 1670 0.15 1670 0 1670 0 10.83<br />
LA07 15� 5 1789 1789 1.25 1789 0.25 1789 0 1789 0 11.57<br />
LA11 20� 5 1899 1899 2.85 1899 1.92 1899 1.27 1899 0.56 58.63<br />
LA12 20� 5 2013 2013 3.22 2013 2.01 2013 1.32 2013 0.56 42.44<br />
LA16 10� 10 3061 3176 3.16 3103 1.95 3061 1.57 3061 1.28 41.23<br />
LA17 10� 10 2912 3021 2.93 2952 1.78 2912 1.46 2912 1.44 36.45<br />
LA21 15� 10 3978 4122 3.83 413 5 3.05 4065 2.57 3978 1.61 132.62<br />
LA22 20� 10 3463 3521 4.92 3539 3.54 3463 3.15 3463 2.37 367.17<br />
LA26 20� 10 3781 3997 4.50 3844 3.54 3805 3.32 3781 2.61 331.54<br />
LA27 20� 10 4433 4538 5.89 4507 5.76 4482 4.37 4433 3.29 430.17<br />
LA31 30� 10 5704 5983 5.75 5875 4.63 5841 3.72 5704 2.86 725.12<br />
LA32 30� 10 5443 5709 5.41 5607 4.43 5574 3.36 5443 2.74 692.20<br />
LA36 15� 15 3523 3748 6.36 3675 4.31 3627 3.92 3523 2.35 542.94<br />
LA37 15� 15 3791 3922 6.83 3955 4.63 3903 4.29 3791 2.68 476.67<br />
ACKNOWLEDGMENT<br />
This research is supported by the Shandong Provincial<br />
Natural Science Foundation, China (No.ZR2010FQ009)<br />
and the National Nature Science Foundation <strong>of</strong> China<br />
(No. 61074150).<br />
REFERENCES<br />
[1] M. R. Garey, D. S. Johnson, and R. Sethi, “The<br />
Complexity <strong>of</strong> Flowshop and Jobshop Scheduling”,<br />
Mathematics <strong>of</strong> Operations Research, Vol. 1, No. 2,<br />
pp.117-129, 1976.<br />
[2] J. Adams, E. Balas, and D. Zawack, “The shifting<br />
bottleneck procedure for job shop scheduling”,<br />
Management Science, Vol. 34, pp.391–401, 1988.<br />
[3] L. W. Cai, Q. H. Wu, and Z. Z. Yong, “A genetic<br />
algorithm with local search for solving the job problems”,<br />
Lecture Notes in Computer Science, Vol. 1803, pp.363-<br />
365, 2000.<br />
[4] Y. Li, Y. Chen, “A genetic algorithm for job-shop<br />
scheduling”, <strong>Journal</strong> <strong>of</strong> S<strong>of</strong>tware, Vol 5, pp.269-274, 2010.<br />
[5] M. Kolonko, “Some new results on simulated annealing<br />
applied to the job shop scheduling problem”, European<br />
<strong>Journal</strong> <strong>of</strong> Operational Research, Vol. 133, No. 1, pp.<br />
123-13,6, 1999.<br />
[6] V. P. Eswaramurthy, and A. Tamilarasi, “Tabu search<br />
strategies for solving job shop scheduling problems”,<br />
<strong>Journal</strong> <strong>of</strong> Advanced Manufacturing Systems, Vol. 6, No.1,<br />
pp.59-75, 2007.<br />
© 2011 ACADEMY PUBLISHER<br />
[7] D. Y. Sha, and C. Y. Hsu, “A hybrid particle swarm<br />
optimization for job shop scheduling problem”,<br />
Computers & Industrial Engineering, Vol. 51, pp.791–<br />
808, 2006.<br />
[8] M. Lansiti, and R. Levien, “Strategy as Ecology’, Harvard<br />
Business Review, Vol. 3, pp.68-78, 2004.<br />
[9] E. M. Goldratt, and J. Cox, “The Goal - A Process <strong>of</strong><br />
Ongoing Improvement”, North River Press, Croton-on-<br />
Hudson, New York, 1984.<br />
[10] Y. Z. Lu, “Pr<strong>of</strong>it driven manufacturing enterprise<br />
optimization: Problem and Solution”, Plenary Presentation,<br />
Proceedings <strong>of</strong> 23rd Chinese Control Conference, August,<br />
2004.<br />
[11] L. W. Jiang, Y. Z. Lu, Y. W. Chen, “Cost Driven<br />
Solutions for Job-shop Scheduling with GA.”, Control<br />
Engineering <strong>of</strong> China. Vol. 44, pp. 72—74, 2007.<br />
[12] A. Allahverdi, J. N. D. Gupta, and T. Aldowaisan, “A<br />
Review <strong>of</strong> Scheduling Research Involving Set-up<br />
Considerations”, Omega, Vol. 27, No.2, pp.219-239, 1999.<br />
[13] D. R. Sule, and K. Y. Huang, “Sequency on two and three<br />
machines with set-up, processing and removal times<br />
separated”, International <strong>Journal</strong> <strong>of</strong> Production Research,<br />
Vol. 21, pp.723-732, 1983.<br />
[14] F. Glover, “Scatter search and path relinking”, In: Corne,<br />
D., Dorigo, M., Glover, F. (Eds.), New Ideas in<br />
Optimization. McGraw-Hill, pp. 297–316, 1999.<br />
[15] R. A. Russel, and W. C. Chiang, “Scatter search for the<br />
vehicle routing problem with time windows”, European<br />
<strong>Journal</strong> <strong>of</strong> Operational Research, Vol. 169, pp.606-622,<br />
2006.<br />
[16] A. Haq, M. Saravanan, A. Vivekraj, and T. Prasad, “A
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 981<br />
scatter search approach for general flowshop scheduling<br />
problem”, International <strong>Journal</strong> <strong>of</strong> Advanced<br />
Manufacturing Technology, Vol.31, No.7, pp.731-736,<br />
2007.<br />
[17] N. Mansour, C. Kehyayan, H. Khachfe, “Scatter search<br />
algorithm for protein structure prediction”, International<br />
<strong>Journal</strong> <strong>of</strong> Bioinformatics Research and Applications.<br />
Vol.5, No.5, 501-15, 2009.<br />
[18] S. Kirkpauick, C. D. Geian, and M. P. Vecchi,<br />
“Optimization by simulated annealing”, Science, Vol. 220,<br />
pp.671-680, 1983.<br />
[19] H. Fisher. and G. L. Thompson, “Industrial scheduling”,<br />
Englewood Cliffs, NJ: Prentice-Hall, 1963.<br />
[20] S. Lawrence, “Resource constrained project scheduling:<br />
An experimental investigation <strong>of</strong> heuristic scheduling<br />
techniques”, Graduate School <strong>of</strong> Industrial Administration,<br />
Carnegie Mellon University, Pittsburgh, PA, 1984.<br />
Bai Jie, was born in Shanghai, China in<br />
April 8, 1981. Received his bachelor<br />
and master degrees in control theory<br />
and engineering from Department <strong>of</strong><br />
Automation at Shanghai Jiaotong<br />
University in 2002 and 2005.<br />
Now he is a PH.D candidate in<br />
control theory and engineering <strong>of</strong><br />
Shanghai Jiaotong University. His<br />
research interest covers industrial production scheduling,<br />
computer integrated manufacturing, and intelligent<br />
optimization and application.<br />
© 2011 ACADEMY PUBLISHER<br />
Sun Kai, was born in Yutai city,<br />
Shandong province, China, in May 30,<br />
1979. Received his PH.D degree in<br />
control theory and engineering from<br />
Department <strong>of</strong> Automation at Shanghai<br />
Jiaotong University in 2009.<br />
Now he is a Lecturer at School <strong>of</strong><br />
Electronic Information and Control<br />
Engineering, Shandong Institute <strong>of</strong><br />
Light Industry, Shandong, China. His research interests are in<br />
optimization and scheduling.<br />
management.<br />
Yang Gen Ke, was born in Taiyuan city,<br />
Shanxi province, China, in June 5, 1963.<br />
Received his PH.D degree in system<br />
engineering from Department <strong>of</strong><br />
Automation at Xi’an Jiaotong University<br />
in 1998.<br />
Now he is a pr<strong>of</strong>essor at Department<br />
<strong>of</strong> Automation <strong>of</strong> Shanghai Jiaotong<br />
University. His research interests are in<br />
operation research and supply chain
982 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Multi-objective Genetic Algorithm for System<br />
Identification and Controller Optimization <strong>of</strong><br />
Automated Guided Vehicle<br />
WU Xing, LOU Peihuang and TANG Dunbing<br />
Nanjing University <strong>of</strong> Aeronautics and Astronautics, Nanjing, China<br />
Email: {Wustar5353, meephlou, d.tang}@nuaa.edu.cn<br />
Abstract— This paper presents a multi-objective genetic<br />
algorithm (MOGA) with Pareto optimality and elitist tactics<br />
for the control system design <strong>of</strong> automated guided vehicle<br />
(AGV). The MOGA is used to identify AGV driving system<br />
model and optimize its servo control system sequentially. In<br />
system identification, the model identified by least square<br />
method is adopted as an evolution tutor who selects the<br />
individuals having balanced performances in all objectives<br />
as elitists. In controller optimization, the velocity regulating<br />
capability required by AGV path tracking is employed as<br />
decision-making preferences which select Pareto optimal<br />
solutions as elitists. According to different objectives and<br />
elitist tactics, several sub-populations are constructed and<br />
they evolve concurrently by using independent reproduction,<br />
neighborhood mutation and heuristic crossover. The lossless<br />
finite precision method and the multi-objective normalized<br />
increment distance are proposed to keep the population<br />
diversity with a low computational complexity. Experiment<br />
results show that the cascaded MOGA have the capability to<br />
make the system model consistent with AGV driving system<br />
both in amplitude and phase, and to make its servo control<br />
system satisfy the requirements on dynamic performance<br />
and steady-state accuracy in AGV path tracking.<br />
Index Terms— multi-objective optimization, genetic<br />
algorithm, system identification, controller optimization.<br />
I. INTRODUCTION<br />
Automated guided vehicle (AGV) is a wheeled mobile<br />
robot with automatic guidance and driving systems. It can<br />
move along the designated routes and transport materials<br />
in flexible manufacturing systems [1]. To correct position<br />
and attitude error promptly in its movement, AGV servo<br />
control system should regulate the velocities <strong>of</strong> driving<br />
wheels at a frequency and accuracy required by its path<br />
tracking [2]. In the hierarchical control architecture, many<br />
sophisticated control laws are used for path tracking at<br />
the upper layer, but it is usual to adopt a PID control law<br />
for servo control at the bottom layer.<br />
How to construct a sufficiently accurate plant model is<br />
the first step for using most non-empirical control system<br />
design methods. Classical identification techniques such<br />
as least square method still have many limitations. If<br />
model construction is considered as an optimization <strong>of</strong><br />
identification accuracy instead <strong>of</strong> a mapping from plant to<br />
model, genetic algorithm (GA) can be used for it [3,4],<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.982-989<br />
e.g. a time-delay system model is identified by GA from<br />
step responses [3].<br />
Moreover, GA can also be used to optimize parameters<br />
<strong>of</strong> PID controller [5-10], e.g. a self-organization GA with<br />
cyclic mutation [6] and a real-coded adaptive GA with a<br />
variable crossover and mutation probability [8]. In many<br />
control systems, it is usual to adopt different controller<br />
parameters based on a trade<strong>of</strong>f in multiple performance<br />
objectives. A multi-objective GA (MOGA) is proposed to<br />
find an appropriate setting <strong>of</strong> PID controller by analyzing<br />
Pareto optimal surfaces [9]. A modified GA with elitist<br />
model and niching method is developed to guarantee a set<br />
<strong>of</strong> PID parameters with different trade<strong>of</strong>fs regarding<br />
multiple requirements [10].<br />
This paper presents a MOGA with Pareto optimality<br />
and elitist tactics for system identification and controller<br />
optimization <strong>of</strong> AGV. The remaining parts are organized<br />
as follows. Section II introduces the existing GAs used<br />
for multi-objective optimization. Section III presents the<br />
MOGA with Pareto optimality and elitist tactics in detail.<br />
Section IV describes AGV prototype and its test system.<br />
Section V applies the MOGA to experiments <strong>of</strong> system<br />
identification and controller optimization. Finally, section<br />
VI gives a brief conclusion.<br />
II. MOGA FOR SYSTEM IDENTIFICATION AND<br />
CONTROLLER TUNING<br />
System identification and controller tuning can both be<br />
converted into multi-objective optimization problems if<br />
the former is viewed as model parameter optimization by<br />
minimizing the error between model response output and<br />
plant response output, and the latter is considered as<br />
controller parameter optimization by minimizing the error<br />
between the actual output and the desired output. In this<br />
sense, they can be handled by one optimization method,<br />
such as a cascaded GA [11,12].<br />
A. Problem Description for Multi-Objective Optimization<br />
The selection <strong>of</strong> objective function has a significant<br />
influence on optimization results. In control engineering,<br />
it is usual to use rising time t , overshoot r<br />
� , stead-state<br />
error e and other time-domain or frequency-domain<br />
s<br />
criteria as control objectives. The purpose <strong>of</strong> multiobjective<br />
optimization is to find a vector X containing a
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 983<br />
set <strong>of</strong> variables x that can simultaneously minimize a<br />
i<br />
function vector F(X) containing a set <strong>of</strong> objective<br />
functions f ( xi<br />
) , which is formulated as<br />
min�<br />
f1( x1)<br />
f 2 ( x ) � f m ( xm<br />
) �<br />
�a b �<br />
min F(X) �<br />
2<br />
.<br />
s.t. x � ,<br />
(1)<br />
i<br />
i<br />
i<br />
Unfortunately, these objectives may conflict with each<br />
other, and these functions may not be minimized at the<br />
same time. For example, a solution with small rising time<br />
may be the one with higher overshoot. Pareto superiority<br />
and Pareto optimality are defined to compare different<br />
solutions in multi-objective optimization problems.<br />
Definition 1. Pareto Superiority<br />
Let X and Y denote two vectors in multi-objective<br />
optimization problems. If the function vector has the<br />
following relationship: fi ( yi<br />
) � f i ( xi<br />
) for all objectives<br />
(i=1,2,…,m), and fi ( yi<br />
) � fi<br />
( xi<br />
) for at least one<br />
objective, X is Pareto superior to Y.<br />
Definition 2. Pareto Optimality<br />
Let X ∈ [a,b] denote a vector in multi-objective<br />
optimization problems. If there is no other vector Y∈[a,b]<br />
Pareto superior to X, X is Pareto optimal.<br />
Multi-objective optimization problems usually involve<br />
the minimization <strong>of</strong> several conflicting criteria that can<br />
not be achieved simultaneously. Therefore a satisfactory<br />
trade<strong>of</strong>f must be found and a set <strong>of</strong> optimal solutions<br />
(instead <strong>of</strong> a single solution) must be provided. In this set,<br />
there is no solution superior to others when all objectives<br />
are taken into account. These solutions (also called nondominated<br />
solutions) comprise Pareto optimal set. The<br />
graphical expression <strong>of</strong> their function values is called<br />
Pareto front. Take the minimization <strong>of</strong> two objective<br />
functions f 1( x)<br />
and f 2 ( x)<br />
for example. If it assumes the<br />
area surrounded by a solid line and a dotted line in Fig.1<br />
is the value range <strong>of</strong> objective functions, then the solid<br />
line is Pareto front <strong>of</strong> this minimization problem, and<br />
point X is Pareto superior to point Y.<br />
Figure 1. Graphical expression <strong>of</strong> Pareto Superiority and Optimality.<br />
B. Improved GA for Multi-Objective Optimization<br />
Conventional GA is only suitable for single-objective<br />
optimization problems because its fitness function only<br />
contains one criterion. Fitness function need be modified<br />
to make it compatible with multi-objective optimization<br />
problems. The possible improved approaches <strong>of</strong> GA can<br />
be classified in three groups.<br />
(1) Aggregating approaches in which all objectives are<br />
combined into a single function, such as weighted sum<br />
approach [6,8]. It is not necessary to modify GA structure<br />
itself, and multi-objective optimization problems can be<br />
solved as the same as single-objective ones. However, it<br />
© 2011 ACADEMY PUBLISHER<br />
is difficult to select weights for different objectives and<br />
an improper selection may lead to optimization failures.<br />
(2) Non-aggregating approaches that are not Pareto<br />
based, e.g. some techniques based on population policies<br />
and special handling <strong>of</strong> the objectives are used to search a<br />
solution set. Vector evaluated GA (VEGA) [13] is a wellknown<br />
example <strong>of</strong> this group.<br />
(3) Pareto based approaches in which the amount <strong>of</strong><br />
individuals that are superior to the individual A is used as<br />
the rank <strong>of</strong> A [14]. Non-dominated Sorting GA (NSGA)<br />
divides the entire population into several groups with<br />
different ranks, and the individuals with the same rank<br />
have the same reproduction probability [15]. NSGA-II is<br />
an improved one that preserves the optimal individuals by<br />
using elitist tactics and replaces fitness sharing parameter<br />
with crowding distance [16].<br />
III. MOGA WITH PARETO OPTIMALITY AND ELITIST<br />
TACTICS<br />
In this paper, Pareto superiority or optimality is used to<br />
construct Pareto sub-population. Elitist selection tactics<br />
are used to preserve excellent individuals and guide the<br />
entire population evolution direction. The lossless finite<br />
precision method and the normalized increment distance<br />
are proposed to keep the population diversity with a low<br />
computational complexity. Multi-population evolution<br />
mechanism is presented to promote the development <strong>of</strong><br />
multiple sub-populations.<br />
A. Elitist Selection Tactics<br />
Because <strong>of</strong> probabilistic behavior existing in evolution,<br />
the best individuals may be lost in the next generation.<br />
Elitist tactics are used widely to guarantee the survival <strong>of</strong><br />
the best individuals in many GAs [9,10]. On another hand,<br />
current researches on MOGA mainly focus on how to get<br />
non-dominated solutions distributed uniformly in Pareto<br />
front, but almost neglect the influence <strong>of</strong> decision-making<br />
preferences on solution selection. Apart from preserving<br />
the best individuals, elitist selection tactics in this paper is<br />
used to inject decision-making preferences into MOGA,<br />
which limits the scope <strong>of</strong> non-dominated solutions to the<br />
area interested by decision-makers (shaded area in Fig.1).<br />
Elitist selection tactics are implemented by two ways.<br />
One way is to designate an individual having balanced<br />
performances in all objectives as the evolution tutor, and<br />
select these individuals that are Pareto superior to this<br />
tutor as elitists. If X is designated as the tutor in Fig.1, all<br />
individuals in the shaded area are potential elitists. They<br />
will evolve forward to Pareto front under the guidance <strong>of</strong><br />
evolution tutor while keeping balanced performances.<br />
The other way is to use decision-making preferences<br />
directly. If a Pareto solution is meaningless to a practical<br />
problem, eliminate it from Pareto sub-population (PSP).<br />
If it satisfies decision-making preferences, select it into<br />
elitist sub-population (ESP). In addition, an aggregated<br />
function is used to describe the overall performance <strong>of</strong><br />
Pareto solutions for multi-objective optimization.<br />
Decision-making preferences get into ESP construction<br />
via elitist selection tactics, pass to the next generation via<br />
ESP reproduction and mutation, and spread in the entire
984 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
population via heuristic crossover <strong>of</strong> elitists and other<br />
individuals. They can form an elitist guidance mechanism<br />
that leads the entire population to Pareto front interested<br />
by decision-makers.<br />
B. Diversity Keeping Techniques<br />
Diversity keeping techniques are used in GA to avoid<br />
the population premature problem [14,16-18]. Niching<br />
methods [14] penalize the crowded individuals by a cost<br />
function. The crowded distance between individuals is<br />
used in non-dominated sorting [16]. The vector norm<br />
function is used for multi-objective fitness [17]. These<br />
methods avoid the closeness and similarity <strong>of</strong> solutions<br />
but need much computing time. Finite precision method<br />
deletes the similar individuals by reducing the computing<br />
precision <strong>of</strong> objective functions intentionally [18]. Its cost<br />
is the reduction <strong>of</strong> computing precision.<br />
In this paper, the lossless finite precision method<br />
(LFPM) and the multi-objective normalized increment<br />
distance (MNID) are used as diversity keeping techniques<br />
at two layers, in order to decrease the distribution density<br />
<strong>of</strong> solutions and keep a high computing efficiency and<br />
precision. The former is used to eliminate the individual<br />
with a serious congestion and a low fitness and the latter<br />
is used to evaluate the elitist fitness that determines the<br />
corresponding parameters in genetic operations.<br />
Let all objective functions in optimization problems be<br />
{ i<br />
F , i=1,2,…,m}. Let the entire population in the current<br />
generation be { D , r=1,2,…,n}. Let all function values <strong>of</strong><br />
r<br />
individual r D be { 1 2<br />
f , r f ,…, r<br />
m<br />
f }. The steps <strong>of</strong> LFPM<br />
r<br />
are described below.<br />
(1) All individuals in the current generation are sorted<br />
in a descending order respectively according to each<br />
i<br />
objective function, and the resulting sequences are F<br />
={ i<br />
j<br />
f , j=1, 2,…,n}. The minimum result variation from<br />
large to small is specified as a fixed step A for each<br />
i<br />
objective function. Initialize the individual number to be<br />
compared in the first function as k=1.<br />
1<br />
(2) Consider the sorting sequence F for the first<br />
objective. Increase the individual number: k=k+1. If k>n,<br />
then terminate the algorithm, otherwise check whether<br />
f and that<br />
the difference between the function value <strong>of</strong> 1<br />
k<br />
<strong>of</strong> its former 1<br />
f is larger than k �1<br />
i<br />
A .<br />
function value <strong>of</strong> individual D (1≤k≤n).<br />
k<br />
1 (3) If f - k<br />
1<br />
f ≥ k�1<br />
i<br />
to step (2).<br />
(4) If 1<br />
k<br />
f - 1<br />
f < k�1<br />
i<br />
function to be compared as s=2.<br />
(5) Search the function value<br />
f is the first<br />
A , then preserve individual D . Jump<br />
k<br />
A , then set the number <strong>of</strong> objective<br />
1<br />
k<br />
s<br />
f <strong>of</strong> individual h<br />
D in k<br />
s<br />
the sequence F <strong>of</strong> the s-th objective. If sorting rank<br />
s<br />
h =1, then preserve individual k<br />
D , and jump to step (2).<br />
k<br />
Otherwise, check whether the difference between the<br />
s<br />
s<br />
function value <strong>of</strong> f and that <strong>of</strong> its former<br />
h<br />
f is larger<br />
h-<br />
1<br />
than A . s<br />
© 2011 ACADEMY PUBLISHER<br />
s<br />
fh�1 ≥ s<br />
s<br />
(6) If f - h A , then preserve individual D . k<br />
Jump to step (2).<br />
s s<br />
(7) If f - h fh� 1 < A , then increase the number <strong>of</strong><br />
s<br />
objective function to be compared as s=s+1. If s>m, then<br />
eliminate individual D due to its serious congestion and<br />
k<br />
jump to step (2). Otherwise, jump to step (5).<br />
It is seen that LFPM deletes the individual when the<br />
difference between its function result and its former’s<br />
result in each objective sorting is smaller than fixed steps,<br />
and it is more suitable for keeping population diversity in<br />
Pareto-based multi-objective optimization.<br />
MNID adopts the conception <strong>of</strong> crowded distance and<br />
vector norm. Elitists are sorted as different sequences<br />
according to each objective respectively. Each objective<br />
function value is converted to one component <strong>of</strong> a vector,<br />
shown as follow<br />
�<br />
� 1<br />
� � k k<br />
�(<br />
oi<br />
� oi�1)<br />
/ oi<br />
k<br />
Di k<br />
�1<br />
i � 1<br />
i � 1<br />
k<br />
k<br />
Where o and i�<br />
1 o are the function values <strong>of</strong> elitist<br />
i<br />
E and i�1<br />
E in the sequence sorted by objective i<br />
k . All<br />
objective function components are computed according to<br />
(2), and MNID is<br />
(2)<br />
1 2 2 2<br />
m 2<br />
�Di � ( �Di<br />
) � ( �Di<br />
) ��<br />
� ( �Di<br />
) (3)<br />
It is seen that relative increments <strong>of</strong> objective functions<br />
are used to compute the difference between the former<br />
elitist and the latter one on each objective, and it is more<br />
suitable for evaluating crowded degree in Pareto-based<br />
multi-objective optimization.<br />
C. Multi-population Evolution Mechanism<br />
Multi-population evolution mechanism (MEM) is used<br />
here to promote the development <strong>of</strong> different individuals.<br />
The entire current-generation population is divided into<br />
multiple sub-populations according to different objectives<br />
and elitist tactics. Single-objective sub-populations (SSP)<br />
are constructed by selecting the individuals based on each<br />
objective. Pareto optimality is used to organize Pareto sub<br />
-population (PSP) in which at least one objective function<br />
value <strong>of</strong> each individual is superior to that <strong>of</strong> others.<br />
Evolution tutor or decision-making preferences are used<br />
as elitist selection tactics for elitist sub-population (ESP).<br />
In the evolution process, sorting rank <strong>of</strong> individual is<br />
the base on which reproduction probability is calculated.<br />
Nonlinear normalized geometric sorting is used here to<br />
relieve the population premature. For the individual with<br />
rank r, the reproduction probability is<br />
r �1<br />
� p(<br />
r)<br />
� q0(<br />
1�<br />
q)<br />
�<br />
n<br />
�q0<br />
� q /[ 1�<br />
( 1�<br />
q)<br />
]<br />
(4)<br />
Where q is the probability parameter changing from 0<br />
to 1, q is the reproduction probability <strong>of</strong> the individual<br />
0<br />
with rank 1, n is the total number <strong>of</strong> individuals in the<br />
sorting sequence, p (r)<br />
is the reproduction probability <strong>of</strong>
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 985<br />
the individual with rank r , and the sum <strong>of</strong> p(r) <strong>of</strong> all<br />
individuals is 1.<br />
The larger q is, the larger reproduction probability the<br />
individual with high rank has, the more influence elitists<br />
have on ordinary individuals, and the heavier selection<br />
pressure the entire population is under. So the selection<br />
pressure can be kept in an appropriate range by changing<br />
the probability parameter q . The function values do not<br />
determine the reproduction probabilities directly, which<br />
can decrease the possibility <strong>of</strong> population premature.<br />
After reproduction, the neighborhood mutation with<br />
variable amplitude is used to produce new elitists. The<br />
mutation amplitude is associated with the reproduction<br />
number <strong>of</strong> individual. Let the reproduction number <strong>of</strong> the<br />
1<br />
r<br />
elitist with rank 1 be N . For the elitist having<br />
e<br />
N e<br />
copies, the amplitude <strong>of</strong> the i-th neighborhood mutation is<br />
B<br />
� 0<br />
�<br />
�B1<br />
� ( i �1)<br />
B<br />
i � 1<br />
r ( i)<br />
� 1<br />
r<br />
2 / N e 1 � i � N e<br />
Where 1 B and B are the initial value and incremental<br />
2<br />
value <strong>of</strong> mutation amplitude. When an elitist has more<br />
copies, the variable range <strong>of</strong> mutation amplitude is larger,<br />
and this operation can search better individuals around<br />
this elitist more carefully. If the elitist only has one copy,<br />
the mutation amplitude is 0, having the same effect as the<br />
elitist preservation method.<br />
Produce randomly a mutation factor � in the range<br />
e<br />
from 0 to 1 for the (i+1)-th copy <strong>of</strong> elitist E , and the<br />
r<br />
new individual after the i-th neighborhood mutation is<br />
(5)<br />
� E ( 1�<br />
( 0.<br />
5 ��<br />
) B ( i))<br />
(6)<br />
X i r<br />
e r<br />
The heuristic crossover is carried out between elitists<br />
and individuals. For the individual X with rank r in the<br />
r<br />
sorting sequence, the crossover amplitude is<br />
C � C rC / N<br />
(7)<br />
r<br />
1 �<br />
Where 1 C and C are the initial value and incremental<br />
2<br />
value <strong>of</strong> crossover amplitude, and N is the total number<br />
x<br />
<strong>of</strong> the sorting sequence.<br />
Produce randomly a crossover factor � in the range<br />
c<br />
from 0 to 1 for the elitist E and the individual i<br />
X , and r<br />
the new individual after the heuristic crossover is<br />
r<br />
c<br />
r<br />
i<br />
2<br />
Y � ( 1��<br />
C ) E ��<br />
C X (8)<br />
It is seen from (7) and (8) that the crossover amplitude<br />
is larger if the sorting rank <strong>of</strong> the individual X is lower,<br />
r<br />
and it is influenced by the elitist E to a larger extent in<br />
i<br />
the heuristic crossover. Decision-making preferences can<br />
spread to SSP from ESP by this operation.<br />
If some components <strong>of</strong> new individuals are beyond the<br />
parameter boundary after the neighborhood mutation and<br />
heuristic crossover, replace them by the boundary values.<br />
© 2011 ACADEMY PUBLISHER<br />
x<br />
c<br />
r<br />
r<br />
D. Algorithm Description<br />
In this paper, Pareto optimality is used to guarantee<br />
solutions with different trade<strong>of</strong>fs regarding multiple<br />
objectives. Elitist selection, neighborhood mutation and<br />
heuristic crossover are combined to expand the influence<br />
<strong>of</strong> decision-making preferences and make a directional<br />
search in Pareto front. LFPM, MNID and MEM are used<br />
to enhance the population fitness and diversity. The steps<br />
<strong>of</strong> the proposed MOGA with Pareto optimality and elitist<br />
tactics are detailed in Fig. 2.<br />
Figure 2. MOGA with Pareto Optimality and elitist tactics.<br />
IV. AGV PROTOTYPE AND ITS TEST SYSTEM<br />
This section describes a vision-based AGV prototype<br />
and its test system, as shown in Fig.3. A CCD camera is<br />
set in the vehicle center. Two driving wheels are placed<br />
on each side <strong>of</strong> its body symmetrically, and their velocity<br />
and direction are controlled by two sets <strong>of</strong> driving devices<br />
(drivers, motors, reducers, etc) respectively. Castors are<br />
distributed around the vehicle to support its weight.<br />
AGV movements at the desired linear and angular<br />
speed are achieved by changing the rotation velocities <strong>of</strong><br />
driving-wheel motors. Path errors are perceived by AGV<br />
vision navigation, and the speed difference between two<br />
driving wheels is calculated by path tracking to eliminate<br />
these errors. Desired driving-wheel velocities are got by<br />
synthesizing the speed difference and AGV moving speed.<br />
Actual velocities are detected by processing the encoder<br />
signals and the errors between them and desired values<br />
are the inputs <strong>of</strong> servo controller. PID controller regulates<br />
driving-wheel velocities by changing motor voltages. Its
986 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
performance is associated with its PID parameters, and<br />
parameter tuning needs the model <strong>of</strong> AGV driving system.<br />
In order to get the real-time data in the experiments <strong>of</strong><br />
system identification and controller optimization, remote<br />
control s<strong>of</strong>tware is developed on the host computer. The<br />
vehicular controller transmits the real-time data <strong>of</strong> speed<br />
difference, desired velocities and actual velocities to the<br />
host computer by using wireless communication devices.<br />
Then the host computer saves the experiment data to the<br />
database, on which different algorithms can be analyzed<br />
and compared effectively based.<br />
V. EXPERIMENTS AND ANALYSIS<br />
This section uses the MOGA to identify system model<br />
and optimize servo controller. The actual velocity <strong>of</strong><br />
driving wheel is recorded in the step response experiment,<br />
and the plant response curve is plotted according to them.<br />
Different GAs are used here and their optimization results<br />
are compared.<br />
A. System Identification by MOGA<br />
Give a step voltage to AGV motor driver, and record<br />
the real-time data <strong>of</strong> actual velocity in the start-up process.<br />
The plant response curve is plotted as the solid curve in<br />
Fig. 4. The second-order model identified by least square<br />
method (LSM) has the response curve as the dashed<br />
curve in Fig. 4, which is largely different from the plant<br />
response curve. GA is used to optimize model parameters<br />
in the following part. Let the perfect second-order model<br />
<strong>of</strong> driving system is<br />
G(<br />
z)<br />
b z � b<br />
* 1 2 � (9)<br />
2<br />
z � a1z<br />
� a2<br />
Where the object optimized by GA is the parameter<br />
vector X � [ a1<br />
a2<br />
b1<br />
b2<br />
] . The vector identified by<br />
LSM is<br />
LSM<br />
X � [�1.<br />
514 0.<br />
6704 0.<br />
005 0.<br />
1504]<br />
.<br />
According to the above result, these parameters are<br />
limited to the following range.<br />
� 2 � a 1 � �1,<br />
0 2 1 � � a , 1 . 0 0 1 � � b , 3 . 0 0 2 � � b .<br />
Firstly, conventional single-objective genetic algorithm<br />
(SOGA) is used here. ITSE (integral <strong>of</strong> the time and the<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 3. Vision-based AGV prototype.<br />
squared error) objective function is adopted regarding the<br />
fast rising response curve in the start-up process, shown<br />
F<br />
A<br />
�<br />
n<br />
�<br />
i�1<br />
* 2<br />
T ( y � y )<br />
(10)<br />
i<br />
*<br />
Where y and i y are the model response output and<br />
i<br />
the plant response output at sampling time T . i<br />
Set the following parameters for SOGA in terms <strong>of</strong><br />
model order and data number. Population scale is N=60.<br />
The maximum number <strong>of</strong> population generation is G=100.<br />
Crossover probability is P c � 0.<br />
6 . Mutation probability is<br />
P m � 0.<br />
2 . SOGA runs randomly 5 times, and the model<br />
parameters identified by SOGA are listed in table I.<br />
TABLE I.<br />
MODEL PARAMETERS IDENTIFIED BY SOGA<br />
Number Fitness Model parameters<br />
1 2.9908 -1.6670 0.7969 0.0470 0.0830<br />
2 2.4510 -1.5437 0.7050 0.0170 0.1433<br />
3 2.2540 -1.5069 0.6712 0.0265 0.1369<br />
4 2.7409 -1.6122 0.7535 0.0365 0.1048<br />
5 2.0128 -1.3975 0.5887 0.0204 0.1707<br />
In table I, the first model has the highest fitness, and its<br />
step response curve is shown as the solid-dotted curve in<br />
Fig.4. It is obvious that the solid-dotted curve approaches<br />
to the solid curve more closely than the dashed curve<br />
both at amplitude and at phase in the first wave top and<br />
bottom, but it has an increasingly larger phase error than<br />
the dashed curve from the second wave top.<br />
Figure 4. Step response curve <strong>of</strong> model optimized by SOGA.<br />
This phenomenon does not imply that GA is inferior in<br />
phase optimization to LSM. Careful analysis reveals the<br />
cause. The objective function (10) defines the amplitude<br />
error between the model response output and the plant<br />
response output. Minimizing the amplitude error <strong>of</strong> the<br />
perfect response curve can also achieve a minimum <strong>of</strong> the<br />
phase error. However, it is almost impossible to minimize<br />
two errors <strong>of</strong> the practical response curve simultaneously<br />
because <strong>of</strong> many distortions. Only using the objective<br />
function <strong>of</strong> amplitude error in GA unavoidably results in<br />
the lack <strong>of</strong> phase precision in this optimization process.<br />
So two objective functions including amplitude error and<br />
phase error need to be used, and SOGA needs to be<br />
replaced with the MOGA as well. The phase error is<br />
defined as<br />
i<br />
i
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 987<br />
F<br />
P<br />
�<br />
m<br />
�<br />
i�1<br />
* 2<br />
w ( O � O ) (11)<br />
*<br />
Where O and i O are the sampling number <strong>of</strong> the i-th<br />
i<br />
wave top or bottom for the model response output and the<br />
plant response output. w is the weight related to i.<br />
i<br />
The multi-objective aggregated function is<br />
FC 1FA<br />
2<br />
i<br />
i<br />
i<br />
� � ��<br />
F<br />
(12)<br />
Where 1 � and 2 � are the weights. Set 1 � = 2<br />
� =0.5 for<br />
a balanced optimization on two objectives.<br />
Since the model identified by LSM has a balanced<br />
precision both at amplitude and at phase, the first elitist<br />
selection tactics are used by employing the parameter<br />
LSM<br />
vector X as the evolution tutor, and leading the entire<br />
population to evolve towards the direction that is Pareto<br />
LSM<br />
superior to X . The selection tactics can decrease the<br />
negative influence <strong>of</strong> fixed weights to the aggregated<br />
function (12), which limits all objective items to a finite<br />
variable range.<br />
The MOGA proposed uses the same population scale<br />
and maximum number <strong>of</strong> population generation as SOGA.<br />
The probability parameter is q =0.1, the initial value and<br />
1<br />
incremental value <strong>of</strong> mutation amplitude is B =0.5 and<br />
e<br />
2<br />
B =0.5, and the initial value and incremental value <strong>of</strong><br />
e<br />
crossover amplitude is 1 C =0.5 and C =0.5. MOGA runs<br />
2<br />
randomly 5 times, and the model parameters identified by<br />
MOGA are listed in table II.<br />
TABLE II.<br />
MODEL PARAMETERS IDENTIFIED BY MOGA<br />
Number Amplitude<br />
error<br />
Phase<br />
error<br />
Model parameters<br />
1 28.2715 47 -1.6781 0.8396 0.0004 0.1599<br />
2 28.4033 47 -1.6762 0.8375 0.0022 0.1580<br />
3 28.3686 47 -1.6750 0.8367 0.0005 0.1605<br />
4 28.6065 47 -1.6706 0.8320 0.0035 0.1568<br />
5 28.7471 47 -1.6630 0.8252 0.0026 0.1588<br />
In table II, model parameters are optimized directly by<br />
minimizing two objective functions <strong>of</strong> amplitude error<br />
and phase error. Five groups <strong>of</strong> models are similar to each<br />
other, which reflect a better convergence <strong>of</strong> MOGA than<br />
that <strong>of</strong> SOGA. The step response curve <strong>of</strong> the first model<br />
identified by MOGA is shown as the solid-dotted curve in<br />
Fig.5. This curve approaches to the solid curve more<br />
closely than the dashed curve both at amplitude and at<br />
phase in the first three wave tops and bottoms, which<br />
shows that the model identified by MOGA has a balanced<br />
high precision both at amplitude and at phase.<br />
Driving system model is identified by MOGA as<br />
0.<br />
0004z<br />
� 0.<br />
1599<br />
G ( z)<br />
�<br />
(13)<br />
2<br />
z �1.<br />
6781z<br />
� 0.<br />
8396<br />
B. Controller Optimization by MOGA<br />
Ziegler-Nichols method is used to tune PID parameters<br />
for the second-order model (13).<br />
© 2011 ACADEMY PUBLISHER<br />
P<br />
k =0.6, P<br />
I k = 5.5245, k =0.0163.<br />
D<br />
The object optimized by GA is the parameter vector<br />
K � [ k P k D k I ] , and their ranges are<br />
0 � k P � 1,<br />
0 � k D � 0.<br />
2,<br />
0 � k I � 30 .<br />
Firstly, weighted sum genetic algorithm (WSGA) is<br />
used here. There objectives combine with each other to<br />
form a weighted sum function, shown as<br />
�<br />
F 3<br />
2<br />
� ( w1<br />
| e(<br />
t)<br />
| t � w2u<br />
( t)<br />
� w4<br />
| ey(<br />
t)<br />
|) dt � w t (14) r<br />
Where e (t)<br />
is the error between the response output<br />
and the desired output. u (t)<br />
is the control input. ey (t)<br />
is<br />
the overshoot error when the response output overshoots.<br />
The weights are set as following to avoid overshoot [6,8].<br />
w =0.999, 1<br />
2 w =0.001, w =2, 3 w =200 4<br />
WSGA uses the same population scale and maximum<br />
number <strong>of</strong> population generation as SOGA. It runs 5<br />
times to optimize the controller for driving system model<br />
(13), and PID parameters are listed in table III. These<br />
parameters are very similar, overshoot is almost zero, and<br />
rising time is equal to setting time.<br />
k<br />
Figure 5. Step response curve <strong>of</strong> model optimized by MOGA.<br />
P<br />
TABLE III.<br />
CONTROLLER PARAMETERS OPTIMIZED BY WSGA<br />
k<br />
D<br />
k F � /%<br />
I<br />
t /s t /s s<br />
0 0.0174 8.3978 8.0267 0.11 0.24 0.24<br />
0 0.0175 8.4764 8.0594 0.13 0.24 0.24<br />
0.0014 0.0178 8.6152 8.1217 0.18 0.22 0.22<br />
0 0.0174 8.4215 8.0379 0.12 0.24 0.24<br />
0.0018 0.0177 8.5431 8.0176 0.21 0.24 0.24<br />
Servo controller is designed by using the first group <strong>of</strong><br />
parameters with the smallest overshoot, the third group<br />
with the highest fitness, and the fifth group with the<br />
lowest fitness. The step response curves <strong>of</strong> driving system<br />
model (13) are shown as the solid-dotted curve, the solid<br />
curve and the dashed curve in Fig.6. Three response<br />
curves superpose with each other, and the driving system<br />
has a step response without any oscillation. Weight 4 w<br />
related to overshoot error is much larger than other<br />
weights in (14), and the severe punishment to overshoot<br />
decreases system response speed unavoidably.<br />
It is seen that weight selection has a great influence on<br />
the optimization results <strong>of</strong> WSGA, and it is difficult to<br />
get a compromise between speed and stability <strong>of</strong> system<br />
r
988 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
response. The MOGA with Pareto optimality and elitist<br />
tactics is proposed to avoid the negative influence. Rising<br />
t and integral absolute error (IAE)<br />
time t , settling time r<br />
s<br />
| e | are adopted as multi-objective functions, shown as<br />
F � y<br />
3<br />
F � t<br />
1<br />
err<br />
�<br />
r<br />
F � t<br />
2<br />
s<br />
n<br />
�<br />
i�1<br />
| e |<br />
Figure 6. Step response curve <strong>of</strong> controller optimized by WSGA.<br />
i<br />
(15)<br />
AGV servo control system is required to achieve a fast<br />
rising and settling response output for velocity regulating<br />
in path tracking [2], and decision-making preferences for<br />
Pareto optimal solutions are defined as following.<br />
(1) If t >1.5s, settling time <strong>of</strong> driving system is too<br />
s<br />
long to satisfy velocity regulating in path tracking. Delete<br />
the Pareto optimal solutions.<br />
Number<br />
1<br />
2<br />
3<br />
4<br />
Figure 7. Step response curve <strong>of</strong> controller optimized by MOGA.<br />
© 2011 ACADEMY PUBLISHER<br />
5<br />
k<br />
TABLE IV.<br />
CONTROLLER PARAMETERS OPTIMIZED BY MOGA<br />
k<br />
P D<br />
I<br />
0.0108 0.0398 18.9429<br />
0.0107 0.0380 17.6868<br />
0.0020 0.0383 17.8309<br />
0.0016 0.0401 19.1172<br />
0.0004 0.0402 19.2231<br />
0.0003 0.0387 17.9431<br />
0.0050 0.0382 17.7343<br />
0.0046 0.0402 19.1693<br />
0.0047 0.0383 17.7930<br />
0.0046 0.0399 19.0549<br />
k | e |<br />
t /s<br />
r<br />
3.2082 0.10<br />
3.2653 0.10<br />
3.2372 0.10<br />
3.1575 0.10<br />
3.1596 0.10<br />
3.2382 0.10<br />
3.2536 0.10<br />
3.1760 0.10<br />
3.2434 0.10<br />
3.1736 0.10<br />
(2) If t ≤0.1s and r t ≤0.2s, driving system achieves a<br />
s<br />
fast and smooth response for velocity regulating. Select<br />
the Pareto optimal solutions as elitists.<br />
(3) If no elitist exists, compare the multi-objective<br />
optimization performance <strong>of</strong> Pareto optimal solutions by<br />
the aggregated function<br />
F � t � t<br />
(16)<br />
C<br />
MOGA adopts the same parameters as the above subsection<br />
except using the maximum generation number<br />
when all elitists remain without any change continuously<br />
is G =10. MOGA runs 5 times to optimize the controller<br />
k<br />
for driving system model (13), and PID parameters are<br />
listed in table IV. It shows that MOGA can find multiple<br />
Pareto optimal solutions rather than only one according to<br />
decision-making preferences, which is different from<br />
WSGA essentially. The generation number <strong>of</strong> MOGA is<br />
only half to that <strong>of</strong> WSGA, and PID parameters have the<br />
similar components and performance, which shows that<br />
MOGA converges to Pareto front interested by decisionmaker<br />
without falling into the local minimum trap. Servo<br />
controller is designed by using two groups <strong>of</strong> parameters<br />
in the second test, and the step response curves <strong>of</strong> driving<br />
system model (13) are shown as the solid curve and the<br />
dashed curve in Fig.7. Two curves have a shorter rising<br />
time and settling time than those in Fig.6. Although their<br />
overshoots are a little larger than that <strong>of</strong> WSGA, the<br />
value <strong>of</strong> less than 10% can still ensure a smooth step<br />
response output and no more than one oscillation.<br />
t /s<br />
s � /s<br />
0.16 8.68<br />
0.10 4.98<br />
0.10 4.95<br />
0.16 8.73<br />
0.16 9.03<br />
0.10 4.97<br />
0.10 4.74<br />
0.16 9.08<br />
0.10 4.97<br />
0.16 8.71<br />
Regarding the difference between the second-order<br />
model (13) and the actual driving system, the first group<br />
<strong>of</strong> PID parameters is used to design servo controller on<br />
ARM LPC2220 and RTOS μC/OS-II. Plot the actual<br />
response curve <strong>of</strong> AGV driving system as the solid-dotted<br />
curve in Fig.7. Although rising time, settling time and<br />
overshoot <strong>of</strong> this actual curve have some increase, this<br />
influence is not so significant to decrease the controller<br />
performance obviously. In AGV movement control test,<br />
our servo controller still has the satisfactory performance<br />
<strong>of</strong> velocity regulating for path tracking.<br />
2<br />
r<br />
2<br />
s<br />
Generation numbers<br />
when convergence<br />
35<br />
50<br />
68<br />
43<br />
32<br />
VI. CONCLUSION
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 989<br />
A cascaded MOGA is used to identify AGV driving<br />
system model and optimize its servo control system in<br />
this paper. Pareto optimality is used in genetic algorithm<br />
to guarantee solutions with different trade<strong>of</strong>fs for multiobjective<br />
optimization. Elitist selection, neighborhood<br />
mutation and heuristic crossover are combined to expand<br />
the influence <strong>of</strong> decision-making preferences and make a<br />
directional search in Pareto front. LFPM, MNID and<br />
MEM are combined to enhance the fitness and diversity<br />
<strong>of</strong> the entire population. Experiment results show that the<br />
cascaded MOGA have the capability to make the system<br />
model consistent with AGV driving system well, and to<br />
make its servo control system satisfy the requirements on<br />
dynamic performance and steady-state accuracy in AGV<br />
path tracking<br />
ACKNOWLEDGMENT<br />
This work was supported in part by a grant from<br />
NUAA Research Funding (Grant No.NJ2010025) and<br />
Research Start-up Funding (Grant No.S1026-053).<br />
REFERENCES<br />
[1] Kelly A., Nagy B., Stager D., et al, “An infrastructure-free<br />
automated guided vehicle based on computer vision,”<br />
IEEE Robotics and Automation Magazine, Vol.14, No.3<br />
(September 2007), pp.24-34<br />
[2] Wu Xing, Lou Peihuang, “Optimal path tracking control<br />
based on motion prediction,” Control and Decision, Vol.<br />
24, No.4 (April 2009), pp. 565-569.<br />
[3] Shin G. W., Song Y. J., Lee T. B, “Genetic algorithm for<br />
identification <strong>of</strong> time delay systems from step responses,”<br />
International <strong>Journal</strong> <strong>of</strong> Control, Automation and Systems,<br />
Vol.5, No.1 (February 2007), pp.79-85.<br />
[4] Tan Xin, Yang Huaqian, “The optimization <strong>of</strong> nonlinear<br />
systems identification based on genetic algorithms,”<br />
Proceedings <strong>of</strong> International Conference on<br />
Computational Intelligence and Security, Guangzhou,<br />
China, October 2006, pp. 266-269.<br />
[5] Meng X. Z., Song B. Y., “Fast genetic algorithms used for<br />
PID parameter optimization,” Proceedings <strong>of</strong> IEEE<br />
International Conference on Automation and Logistic,<br />
Jinan, China, August 2007, pp.2144-2148.<br />
[6] Zhang J. H., Zhuang J., Du H. F., et al, “PID controller<br />
optimization based on the self-organization genetic<br />
algorithm with cyclic mutation,” Proceedings <strong>of</strong> the 6th<br />
International Conference on Artificial Intelligence,<br />
Aguascalientes, Mexico, November 2008, pp. 277-284.<br />
[7] Tan G. Z., Jiang B., Yang L. M., “A novel immune genetic<br />
algorithm-based PID controller design and its application<br />
to CIP-I intelligent leg,” Proceedings <strong>of</strong> the 3rd<br />
International Conference on Natural Computation, Haikou,<br />
China, August 2007, pp. 282-286.<br />
[8] Ding Y. M., Wang X. Y., “Real-coded adaptive genetic<br />
algorithm applied to PID parameter optimization on a 6R<br />
manipulator,” Proceedings <strong>of</strong> the 4th International<br />
Conference on Natural Computation, Jinan, China,<br />
October 2008, pp. 635-639.<br />
[9] Arruda L.V. R., Swiech M. C. S., Delgado M. R. B., et al,<br />
“PID control <strong>of</strong> MIMO process based on rank niching<br />
© 2011 ACADEMY PUBLISHER<br />
genetic algorithm,” Applied Intelligence, Vol.29, No.3<br />
(December 2008), pp. 290-305.<br />
[10] Wang Guoliang, Yan Weiwu, Shao Huihe, “Multiobjective<br />
optimization based on genetic algorithm for PID<br />
controller tuning,” <strong>Journal</strong> <strong>of</strong> Harbin Institute <strong>of</strong><br />
Technology, Vol.16, No.1 (February 2009), pp. 71- 74.<br />
[11] Dionisio S. P., Joao O. P. P., “Genetic algorithm based<br />
system identification and PID tuning for optimum adaptive<br />
control,” Proceedings <strong>of</strong> IEEE/ASME International<br />
Conference on Advanced Intelligent Mechatronics,<br />
Monterey, United states, July 2005, pp.801-806.<br />
[12] Valarmathi K., Devaraj D., Radhakrishnan T. K., “Realcoded<br />
genetic algorithm for system identification and<br />
controller tuning,” Applied Mathematical Modelling,<br />
Vol.33, No.8 (August 2009), pp.3392-3401.<br />
[13] Schaffer J. D., “Multiple objective optimization with<br />
vector evaluated genetic algorithms,” Proceedings <strong>of</strong> the<br />
1st International Conference on Genetic Algorithms,<br />
Hillsdale, Canada, 1985, pp.93-100.<br />
[14] Fonseca C. M., Fleming P. J., “Genetic algorithms for<br />
multiobjective optimization: formulation, discussion and<br />
generalization,” Australian Electronics Engineering,<br />
Vol.27, No.2 (February 1994), pp.416-423.<br />
[15] Srinivas N., Deb K., “Multiobjective optimization using<br />
nondominated sorting in genetic algorithm,” Evolutionary<br />
Computation, Vol.2, No.3 (1994), pp.221-248.<br />
[16] Deb K., Pratap A., Agarwal S., “A fast and elitist<br />
multiobjective genetic algorithm: NSGA-II,” IEEE<br />
Transaction on Evolutionary Computation, Vol.6, No.2<br />
(April 2002), pp.182-197.<br />
[17] Zhao Liang, Ju Gang, Lu Jianhong, “An improved genetic<br />
algorithm in multi-objective optimization and its<br />
application,” Proceedings <strong>of</strong> Chinese Society for Electrical<br />
Engineering, Vol.28, No.2 (January 2008), pp. 96-102.<br />
[18] Qi Rongbin, Qian Feng, Du Wenli, et al, “Multiobjective<br />
genetic algorithm based on elitist selection and individual<br />
migration,” Control and Decision, Vol.22, No.2 (February<br />
2007), pp.164-168.<br />
WU Xing was born in China in 1982. He received his Doctor<br />
Diploma on Mechanical Engineering from Nanjing University<br />
<strong>of</strong> Aeronautics and Astronautics in 2010. He holds a lecturer<br />
position in NUAA now, and acts as the key member <strong>of</strong> some<br />
important projects such as National Natural Science Foundation<br />
<strong>of</strong> China. His research interests include mobile robot, motion<br />
control and embedded system control.<br />
LOU Peihuang was born in China in 1962. He is the dean <strong>of</strong><br />
Jincheng College, NUAA now and holds a pr<strong>of</strong>essor position<br />
since 2001. He acts as the chief leader <strong>of</strong> some important<br />
projects such as Special Project <strong>of</strong> Jiangsu Province for the<br />
Transformation <strong>of</strong> scientific and technological achievements.<br />
His research interests include manufacturing system control and<br />
fault diagnosis. He has won the first prize for scientific and<br />
technological progress <strong>of</strong> Jiangsu province 1 time, the second<br />
prize 2 times and the third prize 5 times.<br />
TANG Dunbing was born in China in 1972. He is the<br />
deputy dean <strong>of</strong> Mechanical and Electrical Engineering<br />
Department, NUAA now and holds a pr<strong>of</strong>essor position since<br />
2005. He acts as the chief leader <strong>of</strong> some important projects<br />
such as National Natural Science Foundation <strong>of</strong> China. His<br />
research interests include creative design and complex system<br />
modeling. He has won the first and second prize for scientific<br />
and technological progress <strong>of</strong> Jiangsu province in 2008.
990 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
WebVR——Web Virtual Reality Engine Based<br />
on P2P network<br />
Zhihan Lv<br />
College <strong>of</strong> Information Science and Engineering, Ocean University <strong>of</strong> China, QingDao, China<br />
CNRS UPR9080/ IBPC, 13 rue Pierre et Marie Curie, F-75005, Paris, France<br />
Email: lvzhihan@gmail.com, lu@ibpc.fr<br />
Tengfei Yin, Yong Han, Yong Chen, Ge Chen*<br />
College <strong>of</strong> Information Science and Engineering, Ocean University <strong>of</strong> China, QingDao, China<br />
� Abstract: WebVR, a multi-user online virtual reality engine,<br />
is introduced. The main contributions are mapping the<br />
geographical space and virtual space to the P2P overlay<br />
network space, and dividing the three spaces by quad-tree<br />
method. The geocoding is identified with Hash value, which<br />
is used to index the user list, terrain data, and the model<br />
object data. Sharing <strong>of</strong> data through improved Kademlia<br />
network model is designed and implemented. In this model,<br />
XOR algorithm is used to calculate the distance <strong>of</strong> the<br />
virtual space. The model greatly improves the hit rate <strong>of</strong> 3D<br />
geographic data search under P2P overlay network. Some<br />
data preprocessing methods have been adopted to accelerate<br />
the data transfer. 3D Global data is used for testing the<br />
engine. The test result indicates that, without considering<br />
the client bandwidth limit, the more users, the faster<br />
loading.<br />
Keyword: Virtual Reality; P2P; WebVR; Web3D; GIS;<br />
Geocoding; Kademlia<br />
I. INTRODUCTION<br />
Geographic Information Science is developed on the<br />
basis <strong>of</strong> Geography, Cartography, Surveying and<br />
Computer Science Disciplines, the s<strong>of</strong>tware entity<br />
implied based on which is Geographic Information<br />
System (GIS). Virtual Reality (VR) technology is<br />
reflection <strong>of</strong> the real world to simulate and generate a<br />
three-dimensional virtual space with computer. The<br />
combination <strong>of</strong> Geographic Information System and<br />
Virtual Reality technology generates VRGIS, which not<br />
only possesses GIS function such as spatial data storage,<br />
process, query and analysis, but significant improves<br />
friendly interface and intuitive interaction combined with<br />
VR technology.<br />
With the developing <strong>of</strong> Internet era this century,<br />
theories and practices combined with Web demonstrates<br />
its unprecedented vigor and vitality, and WebVR-GIS<br />
becomes the inevitable outcome <strong>of</strong> this trend, which<br />
implies VR-GIS on Internet, providing data sharing,<br />
collaborative roaming and GIS analysis function, while<br />
shows the whole “global virtual environment” in a scene.<br />
One case <strong>of</strong> WebVR-GIS extension to multi-user is Web<br />
Virtual Environment, where each user has a virtual role.<br />
Virtual scene inflects with real environment and efficient<br />
information sharing makes each virtual role interactive<br />
entertainment and work together free from limits <strong>of</strong> time<br />
� *Corresponding author, email: webvr@vip.qq.com<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.990-998<br />
and spatial. It is necessary that virtual districts turn to<br />
virtual cities and virtual earth in virtual environment for<br />
the rapid increase <strong>of</strong> users and people explore desire for<br />
new things. Virtual earth is the most macro-scale<br />
implement <strong>of</strong> virtual environment. People are inspired by<br />
its real reflection <strong>of</strong> real world, and eagerly waiting its<br />
characteristics such as massive user collaborative<br />
interaction, massive virtual environment data sharing,<br />
global real geographic location reflection, etc.<br />
However, a series <strong>of</strong> new problems will arise while the<br />
level <strong>of</strong> virtual environment extends to earth. For<br />
example: (1) Global geographic location is complex, so<br />
the space partition methods <strong>of</strong> traditional GIS based on<br />
topological can’t meet its demand; (2) Global virtual<br />
scene data is enormous, which can’t be deposited from<br />
inside and external memory at one-time; (3) Global-scale<br />
nodes was enormous, frequent changes, unpredictable<br />
behavior, so the controllability faces tough challenges.<br />
The normal solutions for the three problems above are<br />
as follows: Common methods known as latitude and<br />
longitude region division model, map projection division<br />
model and Voronoi diagram region division used in<br />
geographic information system are used to divide GIS<br />
region (geographic area dividing) efficiently. The<br />
problem that loading enormous global virtual scene data<br />
is attributed to “global spatial data partition model”.<br />
Earth model division method in three-dimensional space<br />
include traditional “grid (cell) partition”, “G2PS model”,<br />
etc. The reasonable division and organization <strong>of</strong> earth<br />
model will reduce virtual scene data effectively. The third<br />
problem can be attributed as “distribute network model”,<br />
which is much related with hardware. The better solution<br />
is server cluster [1] technology. For example, the terrain<br />
and image database <strong>of</strong> Google Earth containing 70T in<br />
2007 is support with a huge “cloud storage” server cluster<br />
in Google Inc. “Second Life” use each computer to<br />
simulate 90 square meters <strong>of</strong> virtual scenes, which has<br />
5000 servers running now in Linden Inc.<br />
In practice, firstly, the effectiveness and feasibility <strong>of</strong><br />
“geographic area division” and “global special partition<br />
model” is eagerly to be improved. Secondly, global net<br />
supporting similar “cloud computing storage” hardware<br />
requirements coupled with higher maintenance costs will<br />
discourage many institutions. At the same time, the<br />
existing distributed network algorithms are with virtual<br />
network (P2P) on the basis <strong>of</strong> DHT algorithm based, the<br />
user ID <strong>of</strong> which is mostly based on logic distance, rather
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 991<br />
than related to geographic location, and even network<br />
address such as IP. It can reduce index time for user<br />
based on logic distance, but bandwidth waste and traffic<br />
congestion while file transferring cross-regional and<br />
cross-country has become a problem that can’t be<br />
ignored.<br />
Taking all the above into consideration, we propose a<br />
new space division method to encoding map global<br />
geographic environment model based on Hash, which<br />
makes it apply to divide real world, virtual scene and<br />
network, and be the spatial division method <strong>of</strong> overlap<br />
world in network virtual environment. The method<br />
belongs to the reasonable combination <strong>of</strong> geographic<br />
information science, virtual reality and virtual network,<br />
with higher innovation and better scalability. The<br />
universal <strong>of</strong> the method will take “earth virtual<br />
environment” to everyone, providing global users a<br />
platform in which interacting with 3D avatar in virtual<br />
reality environment. All these will not only shorten<br />
distance between people, but also provide a better data<br />
share and cooperation means for science research work,<br />
while provide data base and theoretical basis for<br />
interaction between people and environment.<br />
II. RESEARCH<br />
Through the problems on the P2P network <strong>of</strong> virtual<br />
geographic environment, the following points are worthy<br />
<strong>of</strong> studying:<br />
(1) A supporting mass, précising model <strong>of</strong> the real<br />
world environment, classification and indexing <strong>of</strong> space,<br />
mapping with the real location, enhancing fidelity, meet<br />
the deep planning needs on three-dimensional data.<br />
Cavagna R., C. Bouville, J. Royan developed a P2P<br />
model based on theory <strong>of</strong> space division <strong>of</strong> Voronoi<br />
diagram [2] , for the transfer <strong>of</strong> cites` three-dimensional<br />
scene, and planed to use for online games in the future .<br />
However, its support <strong>of</strong> streaming three-dimensional<br />
scene index structure PBTree [3] for the 2.5-dimensional<br />
data, with a high degree <strong>of</strong> property from<br />
two-dimensional vector data compression out <strong>of</strong> the<br />
three-dimensional model <strong>of</strong> the true model does not<br />
support the fine. M. Varvello, C. Diot, and EW Biersack<br />
combined the KAD network model with the<br />
mathematical model <strong>of</strong> the virtual environment [4] , and<br />
tested in Second life as a framework [5] . He still used the<br />
original data partitioning Second life and organization as<br />
the core model, without importing space division and the<br />
advanced theory <strong>of</strong> real-world mapping.<br />
(2) It can improve stability and data retrieval<br />
efficiency by using distributed Hash structure, which is<br />
based on space partition to virtual index structure in peer<br />
networks. In distributed network, frequently joining in or<br />
quitting <strong>of</strong> nodes will cause a large number <strong>of</strong> networks<br />
Churn, which on the overall robustness <strong>of</strong> the distributed<br />
network architecture puts forward higher requirements.<br />
Hu SY, TH Huang, SC Chang, et al. have developed a set<br />
<strong>of</strong> P2P model Flod [6] based on the Voronoi diagram<br />
theory <strong>of</strong> the space partition. The model has solved the<br />
problem <strong>of</strong> user neighbors distributed storage. Dynamic<br />
classification method applies to rapidly changing data in<br />
© 2011 ACADEMY PUBLISHER<br />
3D game scene. However, the P2P model based on the<br />
theory <strong>of</strong> division space is constantly in a dynamic, it will<br />
trigger the whole user list traversal each time the user<br />
moves, which costs a lot <strong>of</strong> computing performance.<br />
Under the conditions <strong>of</strong> the existing hardware, it can not<br />
be extended to the global field. The global scope <strong>of</strong> the<br />
real geo-spatial environment is relatively stationary; there<br />
is no need <strong>of</strong> dynamic division. Therefore, Approach<br />
with pre-partition manner can completely make the land<br />
division, without changing zoning process with the<br />
mobile node, greatly improving the system running in<br />
real-time. Using the distributed Hash structure, based on<br />
space partition to store, can realize retrieve distributed<br />
resources, achieve the load balancing requirements, and<br />
also improve the hit rate and data retrieval efficiency.<br />
(3) Creating a set <strong>of</strong> the space partition model applied<br />
to the real world, the global virtual environment and the<br />
global network structure can organically couple<br />
geographic addresses, network addresses and user<br />
identity. In the regular research, geo-coding, spatial data<br />
partition model and the virtual peer networks are as an<br />
independent branch <strong>of</strong> study. Use the virtual network<br />
architecture on the virtual environment scene to share<br />
data, while associate the information <strong>of</strong> virtual<br />
geographical world and real users, can effectively solve<br />
the model for low precision, a large amount <strong>of</strong> data,<br />
multi-user interaction delay, waste <strong>of</strong> network bandwidth<br />
optimization and other issues.<br />
III. SYSTEM OVERVIEW<br />
A. Data flow<br />
The design principle is trying to support more load<br />
quantity <strong>of</strong> online users, to reduce the data transmission<br />
quantity, to improve data compression ratio, and to<br />
increase the data download source. Different types <strong>of</strong><br />
data are preprocessed, and the result is stored to data<br />
servers.<br />
While the client is browsing scene, according to the<br />
neighborhood search strategy, it choose the source <strong>of</strong> data<br />
transfer, loading from server if the source can`t be find.<br />
Data<br />
Data<br />
Server<br />
Client<br />
Client<br />
Figure 1 Data Flow<br />
Client<br />
�. THE KEY ISSUES
992 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
A. Graph-based scene topology<br />
The block index based on Hash<br />
Terrain and<br />
scene block<br />
index<br />
Spherical Space<br />
Geocode<br />
d b<br />
The client cache<br />
WebVR rendering<br />
The scene graph is managed by hierarchical bounding<br />
box, using bounding sphere and bounding box to achieve<br />
the scene bounding volume hierarchy. The information is<br />
stored by a directed acyclic graph structure. A scene<br />
graph includes a root node, multi-level interior <strong>of</strong> the side<br />
nodes, and multiple terminal leaf nodes. The root and<br />
side nodes take charge <strong>of</strong> the construction <strong>of</strong> the level <strong>of</strong><br />
© 2011 ACADEMY PUBLISHER<br />
Model and<br />
other kinds<br />
<strong>of</strong> index<br />
Judge AOI<br />
Make the data in AOI into<br />
Download Request list<br />
Select nearby nodes<br />
Through the XOR<br />
Send Download Request List to selected nodes<br />
On-demand loading<br />
the nodes, and the completion <strong>of</strong> certain functions; the<br />
leaf node is saved to one or more object information can<br />
Geocode Based On Hash<br />
Node roaming<br />
Node<br />
Localization<br />
based on Hash<br />
Geocode<br />
Figure 2 Engine Architecture<br />
Node List<br />
Search node<br />
Observe User interaction<br />
and data transmission<br />
Summarize the<br />
mathematical model and<br />
test key data<br />
Join<br />
Withdraw<br />
be drawn. Each node maintains its own bounding volume,<br />
and so on, constitute a distinct level. This level bounding<br />
box diagram can speed up the correct information in the<br />
expression <strong>of</strong> the composition <strong>of</strong> the scene graph, and<br />
also expedite the reduction <strong>of</strong> scene objects, intersection<br />
tests, collision detection and a series <strong>of</strong> operations. This<br />
structure allows each node to have multiple parent
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 993<br />
nodes. When the same geometric object needs to be<br />
repeatedly referenced by more than one parent node<br />
pointing to the same child node, with each parent node<br />
pointing to a new child node <strong>of</strong> the tree than the total<br />
number <strong>of</strong> nodes, memorizing utilization and scene<br />
traversal steps reduced, rendering the final results remain<br />
unchanged.<br />
B. Data file partition<br />
The nodes in scene graph include terrain, objects and<br />
other types, based on different data types, different<br />
partition methods are adopted.<br />
Quadtree-based multi-scale geographic data block<br />
Geographic data includes terrain model using Tin<br />
Triangulation, the real image in the terrain covering the<br />
surface, and vector data. These geographic data are the<br />
surface data and little overlap in the vertical direction, so<br />
as to evenly split the region based on quadtree<br />
classification and index structure, each quadtree level<br />
represents a level <strong>of</strong> precision. An example <strong>of</strong> mature<br />
application is worldwind [7] . Quadtree construction<br />
processes the whole terrain as root, starting from the root<br />
node, checking whether the root partition to satisfy<br />
certain conditions. If it is not satisfied, not partition, it<br />
will be used as a leaf node preserved. Otherwise,<br />
recursively to the root node continuously divided into<br />
four equal sub-regional nodes, until it not split up any<br />
longer. The last step is drawing and rendering all the leaf<br />
nodes. The greater depth <strong>of</strong> division, the resolution will<br />
be higher. That is, each raising separate layer <strong>of</strong> depth,<br />
sampling density doubled. For the earth's surface, it is the<br />
need for separate ways after a projection from the plane<br />
to the latitude and longitude <strong>of</strong> the projection<br />
transformation.<br />
Figure 3 Quadtree block<br />
Object node as unit data block<br />
The scene contains a variety <strong>of</strong> objects nodes,<br />
including the following ones. (1) Construction<br />
information extruded from attributes information with a<br />
high degree <strong>of</strong> vector data; (2)3D model information<br />
import from 3DsMax; (3) geometry data. In the traversal<br />
<strong>of</strong> scene, each node <strong>of</strong> the outermost layer under the root<br />
is considered as a unit.<br />
C. Multi-scale data preprocessing<br />
The topographic data according to the different<br />
quadtree levels, divided the LOD data and stored it to<br />
external memory. For object nodes, each node object to<br />
the crude unit are generated from the refined precision <strong>of</strong><br />
the data L1 4 to L4, in which L2 and L3 as L1 generated<br />
© 2011 ACADEMY PUBLISHER<br />
by collapse <strong>of</strong> law on the simplified model, L4 Impostor<br />
generated by image cache node. The texture object node<br />
based on cell aspect ratios <strong>of</strong> 2:1 generated three<br />
simplified texture memory to external memory, L2-L4 in<br />
four levels corresponding to the simplified model. The<br />
texture data is compressed as DTX3 by GPGPU in the<br />
way <strong>of</strong> calling the CUDA library. External memory<br />
models in different scales in turn are called as needed,<br />
more efficient than single MIPMAP file to be transferred.<br />
D. Data request depends on culling result<br />
For large-scale scenes, when a large number <strong>of</strong> models<br />
are read into memory, the computer system will<br />
inevitably result in a huge burden and could lead to<br />
insufficient memory. At this point, we need a dynamic<br />
scheduling mechanism. Time is unidirectional and cannot<br />
accurately predict the behavior <strong>of</strong> users in the future, and<br />
therefore the data loading to pre-deployment <strong>of</strong><br />
space-related scheduling. Dynamic scheduling can<br />
produce the node when he was on the scene <strong>of</strong> some child<br />
nodes, while drawing long term without any participation.<br />
Child nodes can be automatically uninstalled, free<br />
memory space; On the contrary, he cannot load certain<br />
child nodes in memory, the dynamic scheduling <strong>of</strong> its<br />
control <strong>of</strong> the scene sub-tree.<br />
E. Memory release in time<br />
Using the time-related scheduling on the data<br />
withdraw after called. LOD node <strong>of</strong> a level <strong>of</strong> detail<br />
rendering scenes, if not involved in long-term, it will<br />
uninstall it, otherwise it is loaded. Design <strong>of</strong> smart<br />
pointers in the realization <strong>of</strong> the base class for all nodes,<br />
effectively prevent the memory leak caused by<br />
incomplete release.<br />
F. P2P-based data sharing<br />
Advanced Kademlia-based protocol scheduling discipline<br />
Node ID and files are geo-coded with HASH for an<br />
index purpose.<br />
1. Building geo-coding database<br />
Multiple data formats are involved in this research,<br />
such as vector data, DEM data, image Data, and<br />
three-dimensional model mesh and texture data. It is<br />
necessary to build up an index for data after creating the<br />
earth three-dimensional model. First, vector data,<br />
DEM data and image Data are partitioned in a multi-scale<br />
way according to space information. Each level is labeled<br />
by a 16-bit binary HASH which called “prefix”. After<br />
eight-times partition, each type <strong>of</strong> data can be expressed<br />
as a 128-bit binary HASH, which is made up <strong>of</strong> eight<br />
prefixes in the order left to right, the most significant byte<br />
is the first-scale data (big-endian byte order). The 128-bit<br />
binary HASH will be used as its geo-coding, which also<br />
mapping the sign <strong>of</strong> different scales. Second, as for the<br />
three-dimensional model mesh and texture data. The first<br />
7 partition are combined in the same way as vector data<br />
and DEM data. Accordingly, the result is a 112-bit binary<br />
HASH. Meanwhile, the three-dimensional model itself is<br />
also labeled by a 16-bit binary HASH, adding the<br />
preceding result then the three-dimensional model mesh
994 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
and texture data are expressed as a 128-bit binary HASH<br />
as well. Similarly, its 128-bit binary HASH will be used<br />
as its geo-coding, which mapping the sign <strong>of</strong> different<br />
scales. Next, the geo-coding and corresponding mappings<br />
<strong>of</strong> various data formats are put into different fields <strong>of</strong> a<br />
database and established an index. In addition, filename<br />
<strong>of</strong> data are renamed to relevant geo-coding.<br />
In a Kademlia network, node’s ID is created randomly<br />
and data’s ID is generated according to the contents <strong>of</strong><br />
documents<br />
[8] , therefore, there is time correlativity<br />
between data and their forms <strong>of</strong> expression. By contrast,<br />
the HASH ID <strong>of</strong> our study is calculated by using the<br />
space division method, which has space correlativity.<br />
128bit Hash<br />
128bit Hash<br />
160bit Hash<br />
Index <strong>of</strong> terrain<br />
2 . Constructing Peer-to-Peer network based on<br />
geocoding<br />
For each user, its final ID used for positioning is made<br />
up <strong>of</strong> two parts. The first part is the geo-coding that is the<br />
first eight geographical prefix, which is relevant to its<br />
current location; the second one is its 32-bit binary user<br />
ID.<br />
On user neighbor lists, every node keeps a list <strong>of</strong><br />
neighbor’s information( IP address, UDP port, Node ID).<br />
Those lists are stored in a quadtree. The quadtree is<br />
divided into 7 layers and each layer is marked with a<br />
prefix. Every neighbor’s 32-bit binary user ID is included<br />
in the leaf nodes.<br />
Logical distance. Given two 160-bit identifiers, A and<br />
B, we define the distance between them as their bitwise<br />
exclusive or (XOR) interpreted as an integer, d (A, B) =<br />
A�B . T he distance is related to their<br />
geographical position, during to the fact that the<br />
generation <strong>of</strong> node’s ID which depends on geographical<br />
division. Consequently, smaller distance means they are<br />
geographically closer.<br />
User neighbor lists generating. The locale-sensitive<br />
area <strong>of</strong> nodes is a circle with radius-r.<br />
The adjacent area is formed by the intersection <strong>of</strong> this<br />
circle and the spatial-quadtree leaf nodes. If the<br />
adjacent area has no nodes at all, the radius-r would<br />
become bigger constantly until there is node being<br />
contained in. At the same time, all the nodes being<br />
© 2011 ACADEMY PUBLISHER<br />
Index <strong>of</strong> nodes object<br />
Index <strong>of</strong> users<br />
contained in the adjacent area are added into the user<br />
neighbor lists.<br />
3. Joining and quitting network<br />
Every time when the node logs on to the network, it is<br />
expected to finish three tasks. They are generating its ID,<br />
generating its neighbor lists, and downloading required<br />
item <strong>of</strong> data according to its new location.<br />
When node quit the network, it is impossible for other<br />
nodes to receive this node’s response. As a result, this<br />
node will be regarded as <strong>of</strong>f-line. Thus, it is predicted<br />
that the stability <strong>of</strong> network won’t be affected.<br />
Figure 4 Index Structure<br />
4. Nodes shifting<br />
There are mainly four changes when a node is moving<br />
from one place to another. First, its ID will be<br />
regenerated quickly. Second, starting to search required<br />
data on the basis <strong>of</strong> the new position. Third, its neighbor<br />
lists will be regenerated as well. Fourth, the node has to<br />
inform its neighbors <strong>of</strong> leaving, send request to its new<br />
neighbors and ask for establishing friendship, which<br />
aiming to insert itself into other nodes’ neighbor lists as<br />
necessary.<br />
Figure 5 Node AOI change<br />
5. Neighborhood Selection<br />
Some s<strong>of</strong>tware for file sharing, such as emule would<br />
be in operative condition except at zero searching hit<br />
rates. We need to guarantee the occurrence rate <strong>of</strong> the<br />
geographical neighbors, or we may lose them. We can<br />
reach the conclusion that the node searching way <strong>of</strong> DHT<br />
cannot be used in virtual environment.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 995<br />
After trigger event <strong>of</strong> neighborhood selection<br />
occurring, the peer send its information, I for short, to<br />
any neighbor in its neighbors list. I, will go through the<br />
following process when any peer which receives the<br />
information. P for short, calculating its geocode ID by the<br />
first 128 bit hashID <strong>of</strong> the sending peer, then judge<br />
whether the receiving peer owns the same geocode by<br />
XOR algorithm. If the answer is yes, it will send I to its<br />
neighbors list, then any neighbor receives I executes P,<br />
which is a recursive process. It occurs under the<br />
following conditions. If the hashID <strong>of</strong> peer receiving<br />
information is different, from me at the geocode bits<br />
between 64 and 128, then P stops. The condition <strong>of</strong> all<br />
the searching process stops, the distance between peers<br />
executing P and starting peer is beyond the value <strong>of</strong><br />
(scene side/24). It is different from the method <strong>of</strong> limiting<br />
recursion frequency by depth in Kadmelia. In fact the<br />
whole process is distributed traversal.<br />
6. Data Searching<br />
When trigger event <strong>of</strong> data searching occurs, the peer<br />
will recalculate the logical distance between any peer in<br />
its neighbors list and itself. That is XOR value between<br />
its ID and other IDs. Then it sends data index value to the<br />
nearest n peers, after any peer online receives the<br />
information. It doesn’t only execute local research, but<br />
also sends geo-code info to its peers list, and so on.<br />
According to the small world theory, we could get the<br />
data we need when the recursion frequency is 6. We<br />
make the recursion frequency is 8 by default, which is<br />
used by the radius <strong>of</strong> data searching. When we get the<br />
data, we add the peer with the data into our local<br />
neighbors list. Then they begin to execute multi-source<br />
transmition.<br />
Server based on IOCP for large numbers <strong>of</strong> peers<br />
Because users’ list is stored among lots <strong>of</strong> clients,<br />
some clients may not add it into the network. One <strong>of</strong> the<br />
extreme situations is that, when a peer logins in, all <strong>of</strong> its<br />
users on the list are <strong>of</strong>fline, so it can’t take part in the<br />
network. We call this situation “Information isolated<br />
island”. In order to prevent this situation, we designed the<br />
Server based on IOCP. It stores the recent login user info,<br />
including the user’s name, IP address, and geocode info.<br />
After each user login in the network, it will send its info<br />
to server to store. Then it will totally break the link with<br />
server. After it login in the server, it will download the<br />
latest 20 users without any active user on its users’ list,<br />
and add them into its neighbors list. The model <strong>of</strong> IOCP<br />
and multi-thread can be the best way to use resources,<br />
which supports lots <strong>of</strong> users.<br />
Transmission Strategy<br />
After data searching, we begin to transmit data<br />
according to resource list by searching.<br />
1. Cache Mechanism<br />
The process is manifested in two aspects. First, on the<br />
sending side, when file block is being sent, the<br />
mechanism will control the data bytes sent by a single<br />
© 2011 ACADEMY PUBLISHER<br />
running thread all the time, which is a limiting process.<br />
Second, on the receiving side, it will create a buffer with<br />
the same size with source file before file transmission.<br />
Transmission is pigeon-holing. When the s<strong>of</strong>tware quits,<br />
it will remember every position <strong>of</strong> downloaded data in<br />
the file. When the s<strong>of</strong>tware restarts, it will execute<br />
Resume from break point.<br />
During the transmission process, model data is pressed<br />
by zip-standard while texture data is compressed by<br />
Jpeg2000-standard.<br />
2. File Distribution<br />
In order to improve the number <strong>of</strong> data resources, we<br />
don’t only store the whole data in neighbor peers, but<br />
also peers nearby after file data is cut into blocks. On the<br />
data distribution side, cutting and calculating every file<br />
into a lot <strong>of</strong> blocks with the same size by hash algorithm,<br />
each block can generate unique ID by file geocode and<br />
block content. Distribute and store data resource by the<br />
principle <strong>of</strong> file HashID being closed to geocode on<br />
clients. The process is prepared for data searching. On the<br />
receiving side, searching process is based on parameter <strong>of</strong><br />
file name. Ask some peers whose hashID is closed to the<br />
file name ID, if the answer is yes, then we can find some<br />
peers which owns the file, and ask for downloading.<br />
Figure 6 P2P file share structure<br />
Network adaptability<br />
Realize penetrating NAT by the way <strong>of</strong> holing, and<br />
penetrating firewall based on UPNP.<br />
�. IMPLEMENTATION AND TEST<br />
A. Implementation<br />
The component is encapsulated by Micros<strong>of</strong>t ATL<br />
library. Pack security certificate and the component into<br />
the style <strong>of</strong> CAB. Interact by the way <strong>of</strong> JavaScript<br />
calling components interfaces.
996 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
B. Test<br />
We use 1 ° × 1 ° global elevation and image data<br />
provided by Noaa web site to do data partitioning and<br />
geocoding index generatation for free. Among them,<br />
Qingdao City’s elevation and image data is high accuracy<br />
ASTER G-DEM data. Model data is part <strong>of</strong> building<br />
Data<br />
type<br />
Globe<br />
DEM<br />
Globe<br />
DOM<br />
Globe<br />
9 level<br />
LOD<br />
models <strong>of</strong> Qingdao. We use Sqlserver Database and do<br />
the geocoding <strong>of</strong> number and connectivity <strong>of</strong> parts <strong>of</strong> the<br />
database using C#. Rendering and the network part use<br />
C++. Some pages use script called component functions<br />
<strong>of</strong> JavaScript, and achieve interface effects based on exits<br />
library.<br />
Usually the rendering engine and networking engines<br />
are separately tested, such as people usually test the<br />
amount <strong>of</strong> data supported by rendering engine and its<br />
Qingdao<br />
DEM<br />
rendering efficiency, while testing network engine on the<br />
transmission efficiency. For P2P networks, Consistency,<br />
Persistency, Scalability and other properties are usually<br />
tested. Some researches on P2P networks <strong>of</strong> the virtual<br />
environment also test the changes in performance caused<br />
by AOI region changes.<br />
Because WebVR engine has compact architecture and<br />
Qingdao<br />
DOM<br />
Qingdao<br />
5 level<br />
LOD<br />
Building<br />
Model<br />
4 level<br />
LOD<br />
size 911M 683M 990M 25M 479M 15M 5G<br />
Time 87min 2min 900Min<br />
Figure 7 From Globe to City<br />
© 2011 ACADEMY PUBLISHER<br />
Table 1 Data preprocess time<br />
20<br />
15<br />
10<br />
it is tightly integrated with rendering, Data Division,<br />
Dispatch and network part, we just test the overall<br />
performance.<br />
The users number / User information server load<br />
Test the server bandwidth changes in the second<br />
period when users were respectively (1, 5, 10, 50, 100,<br />
and 200). They login in the user information server at the<br />
same time. The figure shows that the server only<br />
provided the neighbor list download in the initial phase.<br />
Then it disconnected with the node, no longer has traffic.<br />
The users number / Data server load<br />
Test the data server bandwidth changes in 10 seconds<br />
when the number <strong>of</strong> users were respectively (1, 5, 10, 50,<br />
100, 200), and they come into a new scene. The figure<br />
shows that when the node could not get the scene from<br />
other nodes, the data server provides data for download.<br />
5<br />
0<br />
0.1s 0.2s 0.3s 0.4s 0.5s 0.6s 0.7s 0.8s 0.9s 1s<br />
-5<br />
Figure 8 User information server load<br />
100<br />
80<br />
60<br />
40<br />
20<br />
0<br />
1s<br />
-20<br />
2s 3s 4s 5s 6s 7s 8s 9s 10s<br />
Figure 9 Data server load<br />
1<br />
5<br />
10<br />
50<br />
100<br />
200<br />
1<br />
5<br />
10<br />
50<br />
100<br />
200
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 997<br />
The users number / Data loading speed<br />
Test the loading times changes when users were<br />
respectively (1, 5, 10, 50, 100, 200), and a node comes<br />
into a new scene. It can be observed that with the<br />
increasing <strong>of</strong> the number <strong>of</strong> users, data loading speed<br />
increased.<br />
800<br />
700<br />
600<br />
500<br />
400<br />
300<br />
200<br />
100<br />
0<br />
1 5 10 50 100 200<br />
Data size / Loading time and Data size / Rendering frame<br />
Test the data load time and rendering frame rate<br />
changes <strong>of</strong> a node when data volume is respectively (50,<br />
100, 200, 500, 1000, 5000) MB. When the amount <strong>of</strong><br />
data increases by the various engine optimizations, it still<br />
maintains a smooth roaming rate.<br />
5<br />
4<br />
3<br />
2<br />
1<br />
Figure 10 Data loading speed<br />
0<br />
50 100 200 500 1000 5000<br />
Figure 11 Loading time<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
50 100 200 500 1000 5000<br />
�. CONCLUTION<br />
This paper introduced a multi-user online virtual<br />
reality engine. The main contributions are mapping the<br />
geographical space and virtual space to the P2P overlay<br />
network, and dividing the three spatial by<br />
business-oriented quad-tree method. The geographical<br />
code is identified with Hash value, which is used to index<br />
the user list, terrain data, and the model object data.<br />
Achieve sharing <strong>of</strong> data through improved Kadmelia<br />
model. In this model, XOR algorithm is used to calculate<br />
the distance <strong>of</strong> the virtual space.<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 12 Rendering frame<br />
ACKNOWLEDGMENT<br />
This research was supported by the Open Research<br />
Project <strong>of</strong> State Key Laboratory <strong>of</strong> Coal Resources and<br />
Safe Mining under Project SKLCRSM09KFB02 and<br />
Scientists and Engineers Serve Enterprises Program <strong>of</strong><br />
Ministry <strong>of</strong> Science and Technology under Project<br />
2009GJA00047<br />
REFERENCES<br />
[1] Anderson T. E., D. E. Culler, D. A. Patterson, et al.<br />
A case for NOW (<strong>Networks</strong> <strong>of</strong> Workstations), IEEE<br />
Micro, 15(1):54--64, February 1995.<br />
[2] Cavagna R., C. Bouville, J. Royan, P2P Network<br />
for very large virtual environment, Proceedings <strong>of</strong> the<br />
ACM symposium on Virtual reality s<strong>of</strong>tware and<br />
technology, 269-276, 2006.<br />
[3] Royan J., C. Bouville, P. Gioia, PBTree - A new<br />
progressive and hierarchical representation for<br />
network-based navigation in densely built urban<br />
environments, Annales des Télécommunications, 60,<br />
1394-1421, 2005.<br />
[4]M.Varvello, E. Biersack, and C. Diot. A networked<br />
virtual environment over KAD. In Proc. ACM CoNEXT<br />
conference (CoNEXT), pages 1-2, New York, NY, USA,<br />
December 2007.<br />
[5] M.Varvello, C.Diot, and E. W. Biersack. P2P<br />
Second Life: experimental validation using Kad. In<br />
Infocom 2009, 28th IEEE Conference on Computer<br />
Communications, pages 19-25, Rio de Janeiro, Brazil,<br />
April 2009.<br />
[6] Hu S. Y., T. H. Huang, S. C. Chang, et al., FLoD:<br />
A Framework for Peer-to-Peer 3D Streaming, In The<br />
27th Conference on Computer Communications (IEEE<br />
INFOCOM ‘08), 2008.<br />
[7] David G. Bell, Frank Kuehnel, Chris Maxwell,<br />
Randy Kim, Kushyar Kasraie, Tom Gaskins, Patrick<br />
Hogan, Joe Coughlan, NASA World Wind: Opensource<br />
GIS for Mission Operations. New York, 2007.<br />
[8] Maymounkov P., D. Mazires, Kademlia: A<br />
peer-to-peer information systems based on the XOR<br />
metric. In: Proceedings <strong>of</strong> IPTPS, Cambridge, USA,<br />
pp.53-65, Mar.2002,<br />
Zhihan Lv is a Ph. D candidate <strong>of</strong><br />
Marine Information Technology<br />
laboratory, Ocean University <strong>of</strong> China,<br />
China. His research interests include<br />
virtual reality, 3D Visualization,<br />
computer network and s<strong>of</strong>tware<br />
architecture.<br />
He has been an intership at the<br />
Immersion technology Co., LTD for four<br />
years and at the Key Lab <strong>of</strong> Marine<br />
Resource and Environmental Geology, First Institute <strong>of</strong><br />
Oceanography, SOA for four months. From 2010 Sep, he had<br />
been a visiting Ph.D student at French National Center for<br />
Scientific Research (CNRS) in France for one year, with the<br />
support <strong>of</strong> China Scholarship Council.
998 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Tengfei Yin is a master student <strong>of</strong><br />
Marine Information Technology<br />
laboratory, Ocean University <strong>of</strong> China,<br />
China. His research interests include<br />
Web based virtual reality and P2P<br />
network.<br />
.<br />
Yong Han is a pr<strong>of</strong>essor <strong>of</strong> virtual<br />
reality in the keylaboratory <strong>of</strong> Ocean<br />
Remote Sensing, Ministry <strong>of</strong> Education,<br />
Ocean University <strong>of</strong> China, China. His<br />
researchinterests include virtual reality,<br />
computer animation, and GIS.<br />
Yong Chen is a lecturer in college <strong>of</strong><br />
information science and engineering at<br />
Ocean University <strong>of</strong> China, China. His<br />
research interests include computer<br />
technique and computer graphics.<br />
He has gone to New York University<br />
for his postdoctoral fellow at 2009.<br />
Ge Chen is a pr<strong>of</strong>essor <strong>of</strong> physical<br />
oceanography at Ocean University <strong>of</strong><br />
China, China, the dean <strong>of</strong> the School<br />
<strong>of</strong> Information Science, Ocean<br />
University <strong>of</strong> China. His main research<br />
interests include marine remote<br />
sensing, virtual reality, and GIS, PhD<br />
supervisor.<br />
He has been to IFREMER in France<br />
for his postdoctoral fellow for two<br />
years.<br />
Pr<strong>of</strong>. Chen is an International Member <strong>of</strong> New York <strong>Academy</strong><br />
<strong>of</strong> Sciences, and an international member <strong>of</strong> AAAS. He have<br />
served as president <strong>of</strong> the international conference session six<br />
times.<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 999<br />
An Energy-Efficient Communication Protocol for<br />
Wireless Sensor <strong>Networks</strong><br />
Fengjun Shang<br />
College <strong>of</strong> Computer Science and Technology<br />
Chongqing University <strong>of</strong> Posts and Telecommunications, Chongqing 400065, China<br />
E-mail: shangfj@cqupt.edu.cn<br />
Abstract—WSNs (Wireless Sensor <strong>Networks</strong>) can collect<br />
reliable and accurate information in distant and<br />
hazardous environments, and can be used in National<br />
Defence, Military Affairs, Industrial Control,<br />
Environmental Monitor, Traffic Management, Medical<br />
Care, Smart Home, etc. The sensor whose resources are<br />
limited is cheap, and depends on battery to supply<br />
electricity, so it’s important for routing to efficiently<br />
utilize its power. In this paper, an energy-efficient Single-<br />
Hop Active Clustering (SHAC) algorithm is proposed for<br />
wireless sensor networks. The core <strong>of</strong> SHAC has three<br />
parts. Firstly, a timer mechanism is introduced to select<br />
tentative cluster-heads. By analyzing relation between<br />
time <strong>of</strong> timer and residual energy, it is known that time <strong>of</strong><br />
timer is inversely proportional to residual energy <strong>of</strong> nodes<br />
so a timer mechanism can balance the residual energy <strong>of</strong><br />
the whole network nodes which improves the network<br />
energy efficiency. Secondly, a cost function is proposed to<br />
balance energy-efficient <strong>of</strong> each node. Finally, an active<br />
clustering algorithm is proposed for single-hop<br />
homogeneous networks. Through both theoretical analysis<br />
and numerical results, it is shown that SHAC prolongs the<br />
network lifetime significantly against the other clustering<br />
protocols such as LEACH-C and EECS. Under general<br />
instance, SHAC may prolong the lifetime by up to 50%<br />
against EECS.<br />
Index Terms—wireless sensor network, active cluster,<br />
cost function, homogeneous, timer<br />
I. INTRODUCTION<br />
The rapid developments and technological advances in<br />
MEMS(Micro Electromechanical System) and wireless<br />
communication, has made possible the development<br />
and deployment <strong>of</strong> large scale wireless sensor networks.<br />
Wireless sensor network consists <strong>of</strong> hundreds to several<br />
thousands <strong>of</strong> small sensor nodes scattered throughout<br />
an area <strong>of</strong> interest. The potential applications <strong>of</strong> sensor<br />
networks are highly varied, such as environmental<br />
monitoring, target tracking, and battlefield surveillance.<br />
Sensors in such a network are equipped with sensing,<br />
data processing and radio transmission units.<br />
Distinguished from traditional wireless networks,<br />
sensor networks are characterized by severe power,<br />
computation, and memory constraints. Due to the strict<br />
energy constraints, energy efficiency for extending<br />
network lifetime is one <strong>of</strong> the most important topics.<br />
Sensor nodes are likely to be battery powered, and it is<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.999-1008<br />
<strong>of</strong>ten very difficult to change or recharge batteries for<br />
these nodes. Prolonging network lifetime for these<br />
nodes is a critical issue. Therefore, all aspects <strong>of</strong> the<br />
node, from hardware to the protocols, must be designed<br />
to be extremely energy efficient.<br />
Wireless sensor networking is a broad research area,<br />
and many researchers have done research in the area <strong>of</strong><br />
power efficiency to extend network lifetime. In order to<br />
achieve high energy efficiency and increase the<br />
network scalability, sensor nodes can be organized into<br />
clusters. The high density <strong>of</strong> the network may lead to<br />
multiple adjacent sensors generating redundant sensed<br />
data, thus data aggregation can be used to eliminate the<br />
data redundancy and reduce the communication load<br />
[1]. Hierarchical protocols aim at clustering the nodes<br />
so that cluster heads can do some aggregation and<br />
reduction <strong>of</strong> data in order to save energy.<br />
In this paper we assume that the sink is not energy<br />
limited (at least in comparison with the energy <strong>of</strong> other<br />
sensor nodes) and that the coordinates <strong>of</strong> the sink and<br />
the dimensions <strong>of</strong> the field are known. We also assume<br />
that the nodes are uniformly distributed over the field<br />
and they are not mobile. Under this model, we propose<br />
a new energy-efficient adaptive clustering algorithm.<br />
Our contributions have four parts. Firstly, a timer<br />
mechanism is introduced to produce tentative clusterheads<br />
so that our algorithm may prolong network<br />
lifetime. Secondly, an estimated average energy method<br />
is introduced to avoid additional communication<br />
between BS and cluster-head. Thirdly, a cost function is<br />
proposed to balance energy-efficient <strong>of</strong> each node. Last<br />
but not least, an active clustering algorithm is proposed<br />
in single-hop homogeneous network. Through both<br />
theoretical analysis and numerical results, it is shown<br />
that SHAC prolongs the network lifetime significantly<br />
against the other clustering protocols such as LEACH-<br />
C and EECS.<br />
Owning to constraining the resource <strong>of</strong> sensor node,<br />
clustering algorithm aiming at ad hoc networking can<br />
not be used directly, especially, the energy <strong>of</strong> WSN is<br />
limited, so new clustering algorithm must be researched.<br />
LEACH (low-energy adaptive clustering hierarchy) [1]<br />
is first proposed as clustering routing protocol in WSN.<br />
Its clustering idea is used in many clustering routing<br />
protocol, for example, TEEN (threshold sensitive<br />
energy efficient sensor network protocol) [2], HEED<br />
(hybrid energy-efficient distributed clustering) [3] etc.
1000 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
At the same time, there are some independent designing<br />
clustering routing protocol, for example, ACE<br />
(Algorithm for Cluster Establishment) [4], LSCP<br />
(Lightweight Sensing and Communication Protocols) [5]<br />
etc.<br />
The paper is organized as follows. In Section Ⅱ,<br />
related work is discussed. Section Ⅲ, describes our<br />
proposed clustering routing algorithm. In section Ⅳ,<br />
simulation results are presented while Section Ⅴ<br />
concludes the paper.<br />
II. RELATED WORKS<br />
Generally, clustering algorithms for WSNs can be<br />
categorized into two groups: Single hop and Multi-hop<br />
clustering. This section describes a number <strong>of</strong> existing<br />
clustering algorithms within each <strong>of</strong> the following<br />
categories.<br />
A. Single-hop Clustering Algorithm<br />
LEACH is one <strong>of</strong> the most popular hierarchical routing<br />
algorithms for sensor networks. The idea is to form<br />
clusters based on the received signal strength and use<br />
local cluster heads as routers to the sink. This is shown<br />
to save energy since the transmissions will only be done<br />
by such cluster heads rather than all sensor nodes. All<br />
the data processing such as data fusion and aggregation<br />
are local to the cluster. Cluster heads change randomly<br />
over time in order to balance the energy dissipation <strong>of</strong><br />
nodes. This decision is made by the node choosing a<br />
random number between 0 and 1.<br />
In recent years, a number <strong>of</strong> modifications have been<br />
proposed for the LEACH algorithm, for example,<br />
EECS [6], LEACH-B [7] etc. In EECS, in order to<br />
cluster, nodes in selecting cluster-head do not only the<br />
closest cluster-head but also the closest distant from<br />
cluster-head to BS to balance the load <strong>of</strong> network. But<br />
the EECS only balance energy in the area <strong>of</strong> clusterhead<br />
and it can not balance the energy in whole<br />
network.<br />
In Ref. [8], it is proposed to select cluster-head<br />
according to the residual energy <strong>of</strong> node. The main<br />
disadvantage <strong>of</strong> this algorithm requires the energy<br />
information <strong>of</strong> all nodes <strong>of</strong> the network not to be<br />
distributed implementing. SEP [9] mainly aims at twolevel<br />
heterogeneous network, that is, its initial energy<br />
has two kinds <strong>of</strong> level in this network. DEEC algorithm<br />
[10] aim at the general multi-level heterogeneous<br />
networks, But it can also adapted to operate in<br />
homogeneous sensor network. In Ref. [11], DCHS<br />
algorithm is proposed. In this algorithm, in order to<br />
En<br />
_ current<br />
extend the lifetime, a parameter<br />
is<br />
En<br />
_ max<br />
introduced. Furthermore, introducing a factor r s is a<br />
further modification <strong>of</strong> the threshold equation so that<br />
this may improve the performance <strong>of</strong> algorithm.<br />
© 2011 ACADEMY PUBLISHER<br />
HEED (hybrid energy-efficient distributed clustering)<br />
[3] periodically selects cluster heads according to a<br />
hybrid <strong>of</strong> the node residual energy and a secondary<br />
parameter, such as node proximity to its neighbors or<br />
node degree. HEED terminates in O(1) iterations,<br />
incurs low message overhead, and achieves fairly<br />
uniform cluster head distribution across the network. In<br />
order to balance the consuming energy, the above<br />
protocol periodically select cluster heads.<br />
B. Multi-hop Clustering Algorithm<br />
In PEGASIS [12], further improvement on energyconservation<br />
is suggested by connecting the sensors<br />
into a chain. Its shortcoming is that the algorithm must<br />
know the topology <strong>of</strong> network. In Ref. [13], the<br />
network is grouped into a number <strong>of</strong> clusters according<br />
to a randomly selected clustering scheme. TEEN [2] is<br />
well suited for time critical applications and is also<br />
quite efficient in terms <strong>of</strong> energy consumption and<br />
response time. It also allows the user to control the<br />
energy consumption and accuracy to suit the<br />
application. The main drawback <strong>of</strong> this scheme is that,<br />
if the thresholds are not reached, the nodes will never<br />
communicate and the user will not receive any data<br />
from the network at all and will not be ware <strong>of</strong> the<br />
overall operation or availability <strong>of</strong> the network. Thus,<br />
this scheme is not well suited for applications where the<br />
user needs to get data on a regular basis. Another<br />
possible problem with this scheme is that a practical<br />
implementation would have to ensure that there are no<br />
collisions in the cluster. TDMA scheduling <strong>of</strong> the nodes<br />
can be used to avoid this problem. This will however<br />
introduce a delay in the reporting <strong>of</strong> the time-critical<br />
data. APTEEN [14] combines the best features <strong>of</strong> both<br />
proactive and reactive networks and to provide periodic<br />
data collection as well as near real-time warnings about<br />
critical events. It is suitable for a network with evenly<br />
distributed nodes. But it is difficult to design APTEEN<br />
protocol so that it can not be applied. ECMR (energyconscious<br />
message routing) [15] is multi-hop routing<br />
protocol and calls for network clustering and assigns a<br />
less-energy-constrained gateway node that acts as a<br />
centralized network manager. Based on energy usage at<br />
every sensor node and changes in the mission and the<br />
environment, the gateway sets routes for sensor data,<br />
monitors latency throughout the cluster, and arbitrates<br />
medium access among sensors. But it does not support<br />
mobile node.<br />
In Ref. [16], an Energy-Efficient Unequal Clustering<br />
is proposed, which considers the hot spots problem in<br />
multi-hop wireless sensor networks. It partitions the<br />
nodes into clusters <strong>of</strong> unequal size, and clusters closer<br />
to the base-station have smaller sizes than those farther<br />
away from the base-station. Thus cluster heads closer to<br />
the base-station can preserve some energy for the intercluster<br />
data forwarding. Simulation results show that<br />
the unequal clustering mechanism balances the energy<br />
consumption well among all sensor nodes and achieves<br />
an obvious improvement on the network lifetime.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1001<br />
M-LEACH [17] aims at multi-hop sensor network and<br />
the complexity <strong>of</strong> algorithm is great. In Ref. [18],<br />
author proposes a novel energy efficient clustering<br />
scheme for single-hop wireless sensor networks, which<br />
better suits the periodical data gathering applications. It<br />
elects cluster heads with more residual energy in an<br />
autonomous manner through local radio communication<br />
with no iteration while achieving good cluster head<br />
distribution; furthermore, it introduces a novel distancebased<br />
method to balance the load among the cluster<br />
heads. But the algorithm does not consider the energy<br />
distribution.<br />
The clustering algorithm for single-hop networks have<br />
little delay and been well suited for time critical<br />
applications. However, its energy consumption is much<br />
higher between BS and node. The clustering algorithm<br />
for multi-hop network is complex and difficult to<br />
implement. In this paper, we propose a single-hop<br />
clustering algorithm which prolongs the lifetime <strong>of</strong><br />
network.<br />
In order to conserve node energy and prolong lifetime<br />
<strong>of</strong> the network, the previous research have been mainly<br />
focused on balancing energy consumption among<br />
cluster members and they do not consider energy<br />
consumption among cluster-heads. In this paper, we<br />
propose the SHAC algorithm for homogeneous and<br />
single-hop sensor network. According to the residual<br />
energy <strong>of</strong> node, SHAC algorithm selects tentative<br />
cluster-heads in order to improve the clustering idea <strong>of</strong><br />
LEACH. At the same time, SHAC algorithm keeps the<br />
distributed characteristic <strong>of</strong> algorithm and it does not<br />
require location information <strong>of</strong> all nodes <strong>of</strong> the network.<br />
Ⅲ. SHAC ROUTING ALGORITHM<br />
In this paper, a novel clustering idea is proposed<br />
called active clustering. Generally speaking, clusterheads<br />
are first selected based on the corresponding rule<br />
and then nodes are passive adding to that cluster-head,<br />
for example, LEACH selects them according to<br />
threshold, etc.. In our idea, nodes select actively<br />
cluster-heads according to cost function so that it can<br />
balance energy well. Our idea includes several parts as<br />
follows.<br />
In selecting tentative cluster-heads phase, a timer<br />
mechanism is introduced. Its aim is rational selecting<br />
cluster-heads according to residual energy. The high<br />
residual energy is high probability selected cluster-head<br />
so that this may balance the whole network energy.<br />
According to above idea, a clustering algorithm is<br />
proposed based on the average energy <strong>of</strong> whole<br />
network. This algorithm is similar to LEACH-C [8], but<br />
it avoids transmitting residual energy from nodes to BS.<br />
An estimation algorithm must be introduced so that this<br />
algorithm may avoid above energy consumption<br />
problem. After tentative cluster-heads are selected,<br />
according to the cost function, the tentative clusterheads<br />
select final cluster-heads according to prior, that<br />
is, the number <strong>of</strong> nodes adding to that cluster-head and<br />
then each tentative cluster-head knows final clusterhead.<br />
Lastly, the final cluster-head broadcasts<br />
© 2011 ACADEMY PUBLISHER<br />
information around nodes, because selecting clusterheads<br />
is minimal cost so that it may prolong the<br />
lifetime <strong>of</strong> network.<br />
A. Network Model<br />
Let us consider a sensor network consisting <strong>of</strong> N<br />
sensor nodes uniformly deployed over a vast field to<br />
continuously monitor the environment. We denote the<br />
th<br />
i sensor by s i and the corresponding sensor node<br />
set Node = { n1,<br />
n2,...,<br />
nN<br />
} , where Node = N . We make<br />
some assumptions about the sensor nodes and the<br />
underlying network model:<br />
1) There is a base-station (i.e., data sink) located far<br />
away from the square sensing field. Sensors and the<br />
base-station are all stationary after deployment.<br />
2) All nodes are homogeneous and have the same<br />
capabilities. Each node is assigned a unique identifier<br />
(ID).<br />
3) Nodes have no location information.<br />
4) All nodes are able to reach BS in one hop.<br />
5) Nodes can use power control to vary the amount <strong>of</strong><br />
transmission power which depends on the distance to<br />
the receiver.<br />
6) Links are symmetric. A node can compute the<br />
approximate distance to another node based on the<br />
received signal strength, if the transmitting power is<br />
given [16].<br />
We use a simplified model shown in figure 1 for the<br />
radio hardware energy dissipation. Both the free space<br />
2<br />
4<br />
( d power loss) and the multi-path fading ( d power<br />
loss) channel models are used in the model, depending<br />
on the distance between the transmitter and receiver.<br />
Transmission ( E Tx ) and receiving costs ( E Rx ) are<br />
calculated as follows 8 :<br />
⎪⎧<br />
2<br />
lEelec<br />
+ lε<br />
fsd<br />
, d < do<br />
ETx<br />
( l,<br />
d)<br />
= ⎨<br />
(1)<br />
4<br />
⎪⎩<br />
lEelec<br />
+ lε<br />
mpd<br />
, d > do<br />
Where d is the distance between the transmitter and the<br />
receiver.<br />
L bit<br />
packet<br />
Transmit<br />
Electronics<br />
L<br />
E elec<br />
E Tx<br />
L bit<br />
packet<br />
( L,<br />
d)<br />
Tx Amplifier<br />
2<br />
εLd<br />
ERxL Receive<br />
Electronics<br />
L<br />
E elec<br />
Figure 1 Radio Energy Dissipation Model<br />
To receive this message, the energy used by the radio<br />
can be expressed following:<br />
ERx ( l)<br />
= Eelecl<br />
(2)<br />
with l as the length <strong>of</strong> the message in bits, d as the<br />
distance between transmitter and receiver node. A<br />
d
1002 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
sensor node also consumes E DA (nJ/bit/signal) amount<br />
<strong>of</strong> energy for data aggregation. It is also assumed that<br />
the sensed information is highly correlated, thus the<br />
cluster-head can always aggregate the data gathered<br />
from its members into a single fixed length packet.<br />
B. SHAC Algorithm<br />
In the network deployment stage, the base-station<br />
broadcasts a “hello” message to all nodes at a certain<br />
power level. Using this approach, each node can<br />
compute the approximate distance to the base-station<br />
based on the received signal strength. It not only helps<br />
nodes to select the proper power level to communicate<br />
with the base-station, but also helps us to produce<br />
clusters <strong>of</strong> unequal size. Figure 2 gives an overview <strong>of</strong><br />
the SHAC mechanism, where the anomalistic polygon<br />
<strong>of</strong> unequal size represent our clusters <strong>of</strong> unequal size<br />
and the traffic among cluster heads illustrates our<br />
single-hop forwarding method.<br />
3<br />
4<br />
Node i<br />
200m×200m<br />
5<br />
Clusterhead<br />
1<br />
6<br />
2<br />
BS<br />
Figure 2 An overview <strong>of</strong> the SHAC mechanism<br />
From figure 2, The SHAC algorithm makes nodes cost<br />
maximum to have a lower chance <strong>of</strong> becoming a<br />
cluster-head than nodes cost minimum in order to<br />
reduce energy consumption, at the same time, in each<br />
area <strong>of</strong> cluster-head overlay, the highest residual energy<br />
is selected as final cluster-head. The process <strong>of</strong> SHAC<br />
algorithm is as follows. It firstly starts by selecting a<br />
tentative cluster-head. This decision is made by the<br />
timer <strong>of</strong> nodes. If the timer expires, then the sensor<br />
declares itself to be a tentative cluster-head. Thus each<br />
tentative cluster-head receives information from node<br />
adding its cluster-head and calculates the number <strong>of</strong><br />
node. Lastly, in selecting final cluster-heads phase,<br />
each tentative cluster-head selects the final cluster-head<br />
according to the prior so as to acquire final clusterheads.<br />
Once final cluster-head is selected, it broadcasts<br />
the information to the neighboring nodes. In forming<br />
cluster phase, each node adds the selected cluster-head<br />
according to cost function and then each node returns<br />
the information to selected cluster-head. In data<br />
transmitting phase, cluster member transmits data to<br />
© 2011 ACADEMY PUBLISHER<br />
cluster-head according to TDMA slot and then clusterhead<br />
converges the data and transmits to BS. Once the<br />
above process is completed, the algorithm begins to<br />
prepare next round work.<br />
C. Selecting Tentative Cluster-head<br />
Selecting tentative cluster-head is the basis for<br />
creating clusters. After deployment, each sensor sets a<br />
random waiting timer [20]. If the timer expires, then the<br />
sensor declares itself to be a cluster-head, a focal point<br />
<strong>of</strong> a new cluster. However, events may intervene that<br />
cause a sensor to shorten or cancel its timer. For<br />
example, whenever the sensor detects a new neighbor,<br />
it shortens the timer. On the other hand, if a neighbor<br />
declares itself to be a cluster-head, the sensor cancels<br />
its own timer and joins the neighbor’s new cluster.<br />
Because LEACH does not consider residual energy and<br />
distance <strong>of</strong> nodes, an average energy factor for SHAC<br />
and make the node with the highest level <strong>of</strong> energy to<br />
be first tentative cluster-head. The key parameter is as<br />
follows.<br />
E(<br />
i)<br />
residual<br />
Energyi<br />
= (3)<br />
E(<br />
r)<br />
where E ( i)<br />
residual is the residual energy <strong>of</strong> the i-th node,<br />
E (r)<br />
is the average energy <strong>of</strong> the node and r is the<br />
current round number. Energy factor is used to balance<br />
the whole network energy.<br />
Definition The number <strong>of</strong> rounds from first round to<br />
round which first node dies is called lifetime.<br />
Every node i maintains a variable x i , which is<br />
assigned a random value from 0 to 1, namely, x i<br />
=random(0,1). Obviously, x i is a random variable with<br />
uniform distribution on the interval [0, 1]. Each node i<br />
waits for a initiator timer according to an exponential<br />
random distribution i.e.<br />
−λit<br />
i xi<br />
= e<br />
(4)<br />
where λ i = Energyi<br />
.<br />
Formula (4) explains that t i is inversely proportional to<br />
E ( i)<br />
residual . Formula (4) may be written by<br />
ln( xi<br />
)<br />
ti<br />
= −<br />
(5)<br />
λ<br />
Substituting (3) into formula (5) and E (r)<br />
is invariant<br />
in that round, so Formula (5) is written by<br />
ln( xi<br />
)<br />
ti<br />
= −<br />
E(<br />
i)<br />
residual<br />
(6)<br />
We can find the relation between t i and E ( i)<br />
residual by<br />
setting the derivative t i with respect to E ( i)<br />
residual .<br />
dti<br />
dE(<br />
i)<br />
residual<br />
ln( xi<br />
)<br />
=<br />
2<br />
E(<br />
i)<br />
residual<br />
(7)<br />
Q 1 ≤ xi<br />
≤ 1<br />
i
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1003<br />
∴<br />
dt<br />
dE(<br />
i)<br />
i<br />
residual<br />
< 0<br />
, that is, t i is inversely<br />
proportional to E ( i)<br />
residual .<br />
From above analysis, we select short time as clusterhead,<br />
that is, we select high energy as cluster-head, so it<br />
is beneficial to prolong lifetime <strong>of</strong> network. After<br />
sensors are deployed, each sensor sets a random waiting<br />
timer. If the timer expires, then the sensor declares<br />
itself to be a cluster-head, a focal point <strong>of</strong> a new cluster.<br />
D. Estimating Average Energy<br />
It is important to estimate the average energy, but the<br />
disadvantage <strong>of</strong> this approach is that each node has to<br />
estimate the aggregate remaining energy in the network<br />
since this requires additional communication with the<br />
base-station and other nodes. In order to improve the<br />
approach, the average energy must be estimated [10].<br />
In SHAC, we assume that there are N nodes<br />
distributed uniformly in a M×M region. If there are k<br />
N<br />
clusters, there are on average nodes per cluster (one<br />
k<br />
N<br />
cluster head and −1<br />
non-cluster head nodes). Each<br />
k<br />
cluster head dissipates energy receiving signals from<br />
the nodes, aggregating the signals, and transmitting the<br />
aggregate signal to the BS. Since the BS is far from the<br />
nodes, we can assume that the energy dissipation<br />
4<br />
follows the multi-path model ( d power loss). Each<br />
non-cluster head node only needs to transmit its data to<br />
the cluster head once during a round. We can also<br />
assume that the distance to the cluster head is small, so<br />
the energy dissipation follows the free-space model<br />
2<br />
( d power loss). Hence, the total energy consumed<br />
during a single round can be estimated as:<br />
Eround = ECH<br />
+ Enon−CH<br />
(8)<br />
where E CH is tentative cluster-head consumption<br />
energy, Enon− CH is cluster member consumption<br />
energy. For tentative cluster-head E CH , the single<br />
round energy consumption is as follows.<br />
4 N N<br />
E CH = LEelec<br />
+ Lε<br />
mpd<br />
toBS + × LE DF + ( −1)<br />
× LEelec<br />
k<br />
k<br />
(9)<br />
For cluster member Enon− CH , the single round<br />
consumption energy is as follows.<br />
2<br />
Enon−CH<br />
= LEelec + Lε<br />
fsd<br />
toCH (10)<br />
where l is the number <strong>of</strong> bits in each data message,<br />
d toBS is the distance from the cluster head node to the<br />
BS, d toCH is the distance from the node to the cluster<br />
head, and we have assumed perfect data aggregation.<br />
If we know E round , we may estimate the average energy<br />
Etotal<br />
− rEround<br />
Er<br />
= (11)<br />
N<br />
© 2011 ACADEMY PUBLISHER<br />
N<br />
where Etotal<br />
= ∑ Ei<br />
is the initial energy <strong>of</strong> all the<br />
i=<br />
1<br />
nodes, E i is i th node energy, E round is single round<br />
energy consumed. Furthermore, let single round energy<br />
consumed to be uniform. On above condition, we may<br />
estimate E r as follows.<br />
In Ref. [8], the two parameters k and d toCH are given<br />
by:<br />
N ε fs M<br />
k = (12)<br />
2<br />
2π ε mp dtoBS<br />
M<br />
dtoCH = (13)<br />
2πk<br />
From Ref. [19], we may find the parameters d toBS to be<br />
equal to:<br />
M<br />
dtoBS = 0.<br />
765<br />
(14)<br />
2<br />
Substituting (12), (13), and (14), into (11) allows for<br />
estimation <strong>of</strong> Eround. Furthermore, we may estimate that<br />
Etotal<br />
− rEround<br />
Er<br />
=<br />
N<br />
and avoid additional communication between clusterhead<br />
and BS.<br />
E. Active Selecting Cluster-heads<br />
Clustering a wireless sensor network means<br />
partitioning its nodes into groups, each one with a<br />
cluster head and some ordinary nodes as its members.<br />
The task <strong>of</strong> being a cluster head is rotated among<br />
sensors in each data gathering round to distribute the<br />
energy consumption across the network. SHAC is a<br />
distributed cluster heads competitive algorithm, where<br />
cluster head is selected primarily based on the residual<br />
energy <strong>of</strong> each node.<br />
Firstly, several tentative cluster-heads are selected<br />
using the timer mechanism to compete for final cluster<br />
heads. Secondly, each tentative cluster-head broadcasts<br />
a COMPETE_HEAD_MSG, which includes residual<br />
energy(RE), distance from BS(DBS), broadcast radius R<br />
and ID <strong>of</strong> that tentative cluster-head and each node adds<br />
the cluster-head according to the cost function f(i,j).<br />
Thirdly, each tentative cluster-head broadcasts<br />
RECEIVE_NODE_MSG, which includes the number <strong>of</strong><br />
nodes adding this cluster-head and each tentative<br />
cluster-head receives the information and selects the<br />
final cluster-head according to prior knowledge, that is,<br />
selects the cluster-head according to number <strong>of</strong> nodes<br />
adding to that cluster-head, we select them about six<br />
cluster-heads. Lastly, the final cluster-heads broadcast<br />
FINAL_HEAD_MSG.<br />
F. Balancing Cluster Member Energy<br />
After cluster-heads are selected, the key problem is<br />
assigned each node to particular cluster-head. It is<br />
important to balance energy consumption in area<br />
around the cluster-head, for example, node i shown in
1004 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
figure 2 can use cluster-head 5 or 6, which must de<br />
decided by the cost function. Intuitively, node i add<br />
cluster-head 5, because it is close to cluster-head, but, it<br />
is not beneficial to balance the network energy<br />
consumption. The proposed cost function f ( i,<br />
j)<br />
is as<br />
follows.<br />
c(<br />
ni<br />
, CH j ) g(<br />
CH j , BS)<br />
f ( i,<br />
j)<br />
= w×<br />
+ ( 1−<br />
w)<br />
×<br />
(15)<br />
E<br />
E<br />
i<br />
The condition is cost function f ( i,<br />
j)<br />
is minimum, if<br />
node i uses the CH j , where E i is current energy <strong>of</strong><br />
the i-th node, ECH denotes current energy <strong>of</strong> the j-th<br />
j<br />
cluster-head and<br />
2<br />
d ( ni<br />
, CH j )<br />
c ( ni<br />
, CH j ) =<br />
2<br />
d<br />
d<br />
g ( n , CH ) =<br />
i<br />
j<br />
n _ CH<br />
4<br />
( CH j ,<br />
4<br />
dCH<br />
_ max<br />
BS)<br />
where d( CH j , BS)<br />
denotes the distance from j-th<br />
cluster-head to BS, E denotes residual energy <strong>of</strong> j-th<br />
CH j<br />
d _ max max{( CH j<br />
cluster-head, CH = , BS)}<br />
,<br />
d ( ni<br />
, CH j ) denotes the distance from i-th node to j-th<br />
cluster-head, d n _ CH = max{( ni<br />
, CH j )} .<br />
The cost function f includes both distance and<br />
energy factors. In intuition, it can balance the current<br />
energy consumption <strong>of</strong> area around the cluster-head.<br />
The explanation is as follows:<br />
The idea is to making f ( i,<br />
j)<br />
minimum in selecting<br />
cluster member, that is<br />
min{ f ( i,<br />
j)}<br />
= }<br />
) , (<br />
c(<br />
ni<br />
, CH j ) g CH j BS<br />
min{ w × + ( 1−<br />
w)<br />
×<br />
E<br />
E<br />
2<br />
i<br />
d ( ni<br />
, CH j )<br />
d ( CH j , BS)<br />
= min{ w×<br />
+ ( 1−<br />
w)<br />
×<br />
}<br />
4<br />
d × E<br />
d × E<br />
2<br />
n _ CH<br />
i<br />
CH<br />
4<br />
CH _ max<br />
j<br />
CH<br />
j<br />
CH j<br />
In order to be convenient to analysis, constant is<br />
introduced, that is<br />
2<br />
4<br />
d ( ni<br />
, CH j ) d ( CH j , BS)<br />
min{ a ×<br />
+ b ×<br />
}<br />
E<br />
E<br />
i<br />
CH j<br />
1<br />
1<br />
where a = , b = .<br />
2<br />
4<br />
dn<br />
_ CH dCH<br />
_ max<br />
Since at each round, there it is communication between<br />
cluster member and between cluster-head and BS, the<br />
above formula become:<br />
2<br />
4<br />
d ( ni<br />
, CH j ) d ( CH j , BS)<br />
min{ a × ∑<br />
+ b × ∑<br />
} (16)<br />
E<br />
E<br />
i<br />
i<br />
j<br />
CH j<br />
In above formula, factors are introduced, that is<br />
© 2011 ACADEMY PUBLISHER<br />
min{ a ×<br />
∑<br />
i<br />
2<br />
ε fsd<br />
( ni<br />
, CH j )<br />
+ b ×<br />
ε E<br />
fs<br />
i<br />
2<br />
∑<br />
j<br />
4<br />
ε mpd<br />
( CH j , BS)<br />
}<br />
ε E<br />
mp<br />
CH j<br />
In above formula ε d ( n , CH ) is the key for cluster<br />
fs<br />
i<br />
member consumption energy and ε mpd<br />
( CH j , BS)<br />
is<br />
the key for cluster member consumption energy.<br />
Generally, only cost function f ( i,<br />
j)<br />
is minimum, so<br />
cluster member and cluster-head consumption energy is<br />
minimum so that the network consumption energy is<br />
minimum. If so, this may prolong the lifetime <strong>of</strong><br />
network. Or else, if cluster member and cluster-head<br />
consumption energy is minimum so that cost function<br />
f ( i,<br />
j)<br />
must be minimum.<br />
G. Analysis Algorithm<br />
SHAC is a distributed cluster heads competitive<br />
algorithm, where cluster head selection is primarily<br />
based on the residual energy <strong>of</strong> each node. The pseudocode<br />
for an arbitrary node s i is given as follows. Our<br />
goal is selecting the final cluster-heads according to<br />
cost function. Therefore the distribution <strong>of</strong> cluster<br />
heads can be controlled over the network.<br />
Each tentative cluster head maintains a set ACH <strong>of</strong> its<br />
“adjacent” tentative cluster heads. Tentative head si is<br />
an “adjacent” node <strong>of</strong> sj if si is in sj’s competition range.<br />
Whether a tentative cluster head sj will become a final<br />
cluster head depends on the cost function <strong>of</strong> nodes only,<br />
i.e., the algorithm is distributed.<br />
Algorithm1 SHAC Pseudo-code<br />
BS broadcast average energy message AE_MSG;<br />
%Selecting ClusterHead<br />
For i=1:1:NodeNumber<br />
If the timer fires<br />
Broadcast ADV_Tentative_Head(R);<br />
End<br />
On receiving ADV_Tentative_Head(R) form si<br />
sj broadcast a JOIN_MSG(ID);<br />
End<br />
For i=1:1:NodeNumber<br />
If si is Tentative Clusterhead<br />
Broadcast COMPETE_HEAD_MSG(ID, R, RE, DBS);<br />
End<br />
End<br />
On receiving COMPETE_HEAD_MSG form sj<br />
If ∀ ni∉beTentativeHead with min(f(i,j))<br />
Add ni to sj;<br />
sj.Num=sj.Num+1;<br />
End<br />
For k=1:1: TentativeClusterNumber<br />
Broadcast RECEIVE_NODE_MSG(ID, R, RE, DBS);<br />
Other ∀ ni∉ beTentativeHead Receive RECEIVE_NODE_MSG;<br />
End<br />
For j=1:1: TentativeClusterNumber<br />
Selecting about six clusterheads with max(sj.Num);<br />
Broadcast FINAL_HEAD_MSG;<br />
On receiving a FINAL_HEAD_MSG from sj<br />
If si∈sj.ACH<br />
j<br />
4
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1005<br />
Broadcast QUIT_MSG;<br />
End<br />
On receiving a QUIT_MSG from si<br />
If si∈sj.ACH<br />
Remove si from sj.ACH;<br />
End<br />
End<br />
%Forming Cluster<br />
For i=1:1:NodeNumber<br />
If ni∈sj.ACH<br />
Broadcast HEAD_MSG(ID, Rmax, RE);<br />
Wait for JOIN_ClusterHead_MSG;<br />
Else<br />
Receiving all HEAD_MSG;<br />
Calculate the f(i,j);<br />
Select the ClusterHead with min{f(i, j)};<br />
Add ni to CHj and Broadcast the JOIN_ClusterHead_MSG;<br />
End<br />
End<br />
In this algorithm, each node broadcast information<br />
using same power, in order to save energy, this<br />
broadcasting radius is R. Firstly, according to waiting<br />
timer, if the timer expires, the tentative cluster-heads<br />
are produced. Secondly, the tentative cluster-heads<br />
broadcast the number <strong>of</strong> receiving node. Thirdly, it<br />
produces the final cluster-head according to sj.Num.<br />
Furthermore, each cluster-head broadcasts HEAD_MSG<br />
information, which has ID <strong>of</strong> node, Rmax and residual<br />
energy RE. According to receiving information, nodes<br />
calculate cost function f ( i,<br />
j)<br />
and select minimum<br />
f ( i,<br />
j)<br />
cluster-head. Lastly, nodes send<br />
JOIN_ClusterHead_MSG information and tell that<br />
cluster-head.<br />
After cluster-head is selected, according to cost<br />
function, nodes add themselves to the cluster-head.<br />
Furthermore, during the data transmission phase, the<br />
cluster-head structures a TDMA-based schedule, which<br />
determines when each cluster member can<br />
communicate with the cluster-head.<br />
According to Algorithm 1, the cluster head selection<br />
process is message driven, thus we must discuss its<br />
message complexity.<br />
Theorem The message complexity <strong>of</strong> the cluster<br />
formation algorithm is O(N).<br />
Pro<strong>of</strong>: At the beginning <strong>of</strong> the cluster head selection<br />
Area<br />
phase, BS broadcasts an AE_MSG, (Area is<br />
2<br />
πR<br />
selected that area <strong>of</strong> scenario) tentative cluster-heads<br />
broadcast ADV_Tentative_Head and cluster members<br />
Area<br />
broadcast N- JOIN_MSG. Next step, tentative<br />
πR<br />
2<br />
cluster heads are produced and each <strong>of</strong> them broadcasts<br />
a COMPETE_HEAD_MSG. Then each <strong>of</strong> them makes<br />
a decision by broadcasting a RECEIVE_NODE_MSG to<br />
calculate the total number <strong>of</strong> node adding that clusterhead.<br />
Furthermore each <strong>of</strong> them makes a decision by<br />
broadcasting a FINAL_HEAD_MSG to act as a final<br />
cluster head, or a QUIT_MSG to act as an ordinary<br />
node. Suppose k cluster heads are selected, they send<br />
© 2011 ACADEMY PUBLISHER<br />
out k HEAD_MSG, and then (N-k) ordinary nodes<br />
transmit (N-k) JOIN_ClsterHead_MSG. Thus the<br />
Area<br />
Area<br />
messages add up to N × +N × (1 - ) + 3N<br />
2<br />
2<br />
πR<br />
πR<br />
Area<br />
Area<br />
× + k + N -k +1= (3× + 2) × N + 1 in the<br />
2<br />
2<br />
πR<br />
πR<br />
cluster formation stage per round, i.e., O(N).<br />
The above theorem shows the message overhead <strong>of</strong><br />
SHAC is small. In HEED, the upper-bound <strong>of</strong> message<br />
complexity is Niter×N where Niter is the number <strong>of</strong><br />
iterations. Because we have avoided message iteration<br />
in the cluster-head selection algorithm, the control<br />
message overhead in SHAC is much lower than that in<br />
HEED. EECS algorithm is also O(N), but when it select<br />
cluster-head, it does not consider the cost <strong>of</strong> each node<br />
and cluster-head.<br />
Ⅳ. SIMULATIONS AND ANALYSIS<br />
We select a scenario to simulate our algorithm using<br />
MATLAB and the parameter set is shown in Table I.<br />
TABLE Ⅰ. SIMULATION PARAMETERS<br />
Parameter Value<br />
Network coverage (0,0)~(200,200)<br />
Base station location (100,350)<br />
N 1000<br />
Initial energy 1J<br />
50nJ/bit<br />
Eelec<br />
εfs<br />
εmp<br />
do<br />
EDA<br />
10pJ/bit/m 2<br />
0.0013pJ/bit/m 4<br />
87m<br />
5nJ/bit/signal<br />
Data packet size 4000 bits<br />
The key problem is how to select parameter w in SHAC<br />
algorithm, because w is weight value using balancing<br />
between cluster-head and cluster member. We select w<br />
from 0 to 1 to test it and the test results are shown in<br />
figure 3. From figure 3, it is known w=0.6 is the best,<br />
so we select w =0.6.<br />
Number <strong>of</strong> Rounds<br />
1000<br />
900<br />
800<br />
700<br />
600<br />
500<br />
400<br />
300<br />
200<br />
100<br />
0<br />
0 0.2 0.4 0.6 0.8 1 1.2<br />
Weight w<br />
Figure 3 Relation between w and number <strong>of</strong> rounds<br />
We compare between LEACH-C, EECS and SHAC.<br />
Figure 4 shows lifetime <strong>of</strong> network over the simulation<br />
time, where SHAC is the longest. SHAC prolongs the<br />
network lifetime significantly against the other<br />
clustering protocols such as LEACH-C and EECS.<br />
Under general instance, SHAC may prolong the<br />
lifetime at least 30%. Additionally, it may prolong the<br />
lifetime by up to 50% against EECS.
1006 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Number <strong>of</strong> Alive Nodes<br />
1000<br />
900<br />
800<br />
700<br />
600<br />
500<br />
SHAC<br />
EECS<br />
LEACH-C<br />
Comparing Number <strong>of</strong> Alive Node<br />
400<br />
600 700 800 900 1000 1100 1200 1300 1400<br />
Number <strong>of</strong> Rounds<br />
Figure 4 Lifetime <strong>of</strong> network over simulation time<br />
By observing the number <strong>of</strong> dead nodes from figure 4,<br />
it can be seen that there are no dead nodes in 1,000<br />
rounds <strong>of</strong> SHAC. In 1,000 rounds <strong>of</strong> EECS, there are<br />
30 nodes at least, which is 3% <strong>of</strong> number <strong>of</strong> total<br />
nodes, there are 100 nodes at least, which is 10% <strong>of</strong><br />
total number <strong>of</strong> node in 1,000 rounds in LEACH-C.<br />
The number <strong>of</strong> dead node shows balance network<br />
energy consumption. The less there are dead nodes,<br />
the better we can do balance network energy. SHAC<br />
both prolongs lifetime <strong>of</strong> network and reduces the<br />
number <strong>of</strong> dead nodes. Hence, SHAC more efficiency<br />
balances the energy consumption <strong>of</strong> network<br />
compared to the other two strategies. EECS introduces<br />
a cost function, but its performance is not better than<br />
SHAC, the reasons have three points. Firstly, it does<br />
not consider energy consumption <strong>of</strong> node in selecting<br />
tentative cluster-head. Secondly, it only considers the<br />
distance factor and omits residual energy <strong>of</strong> node and<br />
cluster-head. Lastly, it does not consider the cost <strong>of</strong><br />
each node and cluster-head. It has been considered the<br />
residual energy <strong>of</strong> node in LEACH-C, but overhead<br />
about transmitting information is bigger, so it has the<br />
worst performance.<br />
Number <strong>of</strong> Rounds<br />
500<br />
450<br />
400<br />
350<br />
300<br />
250<br />
200<br />
150<br />
100<br />
50<br />
0<br />
Comparing Number <strong>of</strong> Clusterhead<br />
EECS<br />
LEACH-C<br />
SHAC<br />
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17<br />
Number <strong>of</strong> Clusterheads<br />
Figure 5 Number <strong>of</strong> cluster-head versus number <strong>of</strong> rounds<br />
The number <strong>of</strong> cluster-head is shown figure 5. It is<br />
known that SHAC is stable. The fluctuation about<br />
number <strong>of</strong> cluster-head, EECS is similar to SHAC,<br />
LEACH-C is the biggest, because LEACH-C only<br />
uses random number and threshold scheme to select<br />
cluster-head so that the fluctuation is significant. The<br />
© 2011 ACADEMY PUBLISHER<br />
number <strong>of</strong> cluster-head in SHAC and EECS is<br />
converged to the optimized number <strong>of</strong> cluster-head,<br />
because two algorithms use competing scheme so that<br />
they control number <strong>of</strong> cluster-head, but SHAC<br />
algorithm introduces active scheme to select clusterheads<br />
so that energy-efficiency is the highest among<br />
above three algorithms.<br />
In this part, we investigate the energy efficiency <strong>of</strong><br />
SHAC, we compare the amount <strong>of</strong> cumulative residual<br />
energy by all nodes in three algorithms. 1400 rounds <strong>of</strong><br />
simulations are sampled and the amount <strong>of</strong> total<br />
cumulative residual energy by all nodes is shown in<br />
Figure 6. The residual energy by all nodes per round in<br />
SHAC is the highest compared to EECS, LEACH-C.<br />
And because the distribution <strong>of</strong> selected cluster-heads is<br />
uncontrollable in EECS and LEACH-C, there is a<br />
dramatic variation in the energy consumption <strong>of</strong> the<br />
cluster-heads. Due to the stability <strong>of</strong> cluster-heads<br />
topology in SHAC, the amount <strong>of</strong> energy spent by all<br />
nodes is almost the same and the lowest in each round.<br />
In Figure 6, SHAC achieves the most residual energy in<br />
the network and this further illustrates why SHAC<br />
achieves the longest lifetime compared to the other<br />
strategies.<br />
Total Cumulative Residual Energy in Network(J)<br />
Comparing Total Cumulative Residual Energy <strong>of</strong>Network over Number <strong>of</strong> Rounds<br />
750<br />
SHAC<br />
EECS<br />
700<br />
LEACH-C<br />
650<br />
600<br />
550<br />
500<br />
450<br />
400<br />
500 600 700 800 900 1000 1100 1200 1300 1400<br />
Number <strong>of</strong> Rounds<br />
Figure 6 Total cumulative residual energy <strong>of</strong> network versus number<br />
<strong>of</strong> rounds.<br />
It compares the number <strong>of</strong> information from all cluster<br />
member nodes to all cluster-heads in figure 7. This<br />
figure shows that, SHAC forwards the most number <strong>of</strong><br />
packets, when compared to the other strategies, from<br />
nodes to cluster-heads. Furthermore, from this figure it<br />
can be seen that LEACH-C and EECS experience a<br />
significant drop in performance after 850 th round,<br />
where SHAC continues to forward more packets.<br />
After first node is dead, the number <strong>of</strong> information<br />
about SHAC is bigger than LEACH-C and EECS,<br />
because the lifetime is longer than LEACH-C and<br />
EECS. After 850 th rounds, the number <strong>of</strong> information<br />
is significantly bigger than LEACH-C EECS, because<br />
dead nodes add quickly. The number <strong>of</strong> information<br />
from node to cluster explains details <strong>of</strong> real scenario.<br />
The more is the formation, the more are details, so<br />
SHAC is better in real application.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1007<br />
Number <strong>of</strong> Message Received<br />
x Number 106 <strong>of</strong> Total Message Received in Clusterhead Over Number <strong>of</strong> Rounds<br />
1.4<br />
1.3<br />
1.2<br />
1.1<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
SHAC<br />
EECS<br />
LEACH-C<br />
0.6<br />
700 800 900 1000 1100 1200 1300 1400<br />
Number <strong>of</strong> Rounds<br />
Figure 7 Number <strong>of</strong> information from node to cluster-head<br />
Figure 8 shows the total number <strong>of</strong> dead node, when<br />
the number <strong>of</strong> dead nodes is less than 5% <strong>of</strong> each<br />
round. From figure 8, we know that the mostly dead<br />
nodes are in center <strong>of</strong> scenario and it may explain the<br />
algorithm can balance energy consumption <strong>of</strong> network<br />
and has better energy efficiency.<br />
400<br />
350<br />
300<br />
250<br />
200<br />
150<br />
100<br />
50<br />
Base Station<br />
0<br />
0 50 100 150 200<br />
Figure 8 Distribution <strong>of</strong> dead nodes, where x axis is the length <strong>of</strong><br />
scenario and y axis is width <strong>of</strong> scenario<br />
Ⅴ . CONCLUSIONS AND FUTURE WORKS<br />
In this paper, a Single-Hop Active Clustering (SHAC)<br />
algorithm is proposed about wireless sensor networks<br />
by research current routing algorithms. The core <strong>of</strong><br />
SHAC has three parts. Firstly, a timer mechanism is<br />
introduced to select tentative cluster-heads. By<br />
analyzing relation between time <strong>of</strong> timer and residual<br />
energy, it is known that time <strong>of</strong> timer is inversely<br />
proportional to residual energy <strong>of</strong> nodes so a timer<br />
mechanism can balance the residual energy <strong>of</strong> the<br />
whole network nodes which improves the network<br />
energy efficiency. Secondly, a cost function is proposed<br />
to balance energy-efficient <strong>of</strong> each node. Last but not<br />
least, an active clustering algorithm is proposed in<br />
single-hop homogeneous network. Through both<br />
theoretical analysis and numerical results, it is shown<br />
that SHAC prolongs the network lifetime significantly<br />
against the other clustering protocols such as LEACH-<br />
C and EECS. Under general instance, SHAC may<br />
prolong the lifetime 30% at least, especially, it may<br />
prolong the lifetime up to 50% against EECS.<br />
In future research, we will consider NS2 simulation<br />
platform using event-driven mechanism to simulate<br />
© 2011 ACADEMY PUBLISHER<br />
performance <strong>of</strong> the SHAC algorithm. In LEACH-C,<br />
EECS and SHAC, we assume that data are transmitted<br />
at any moment, but for event-driven network, in no<br />
events, nodes do not consume energy and keep<br />
sleeping status. Once there is a event, the node is<br />
waked to collect data and communicate, so this can<br />
improve energy-efficient <strong>of</strong> sensor network so that this<br />
make SHAC is better to apply in real condition.<br />
ACKNOWLEDGEMENTS<br />
The author would like to thank the Chongqing Natural<br />
Science Foundation under Grant No. 2009BB2081 and<br />
the Science and Technology Research Project <strong>of</strong><br />
Chongqing Municipal Education Commission. The<br />
Project Sponsored by the Scientific Research<br />
Foundation for the Returned Overseas Chinese Scholars,<br />
State Education Ministry. The author would also like to<br />
thank to MATLAB s<strong>of</strong>tware.<br />
REFERENCES<br />
[1] W. Heinzelman, A. Chandrakasan, and H. Balakrishnan.<br />
Energy-efficient routing protocols for wireless microsensor<br />
networks. Proceeding 33rd Hawaii International Conference<br />
on System Sciences,Volume 8, pp.8020-8030, 2000.<br />
[2] Manjeshwar A, Grawal DP. TEEN: A protocol for enhanced<br />
efficiency in wireless sensor networks. Proceeding <strong>of</strong> the<br />
15th Parallel and Distributed Processing Symp. San<br />
Francisco: IEEE Computer Society, 2001, pp. 2009-2015.<br />
[3] Younis O, Fahmy S. Heed: A hybrid, energy-efficient,<br />
distributed clustering approach for ad-hoc sensor networks.<br />
IEEE Trans. on Mobile Computing, 2004, 3(4):660-669.<br />
[4] Chan H, Perrig A. ACE: An emergent algorithm for highly<br />
uniform cluster formation. Proceeding <strong>of</strong> the 1st European<br />
Workshop on Wireless Sensor <strong>Networks</strong>. LNCS 2920, Berlin:<br />
Springer-Verlag, 2004, pp. 154-171.<br />
[5] Fang Q, Zhao F, Guibas LJ. Lightweight sensing and<br />
communication protocols for target enumeration and<br />
aggregation. Proceeding <strong>of</strong> the 4th ACM Int’l Symp. on<br />
Mobile Ad Hoc Networking & Computing. ACM Press, 2003,<br />
pp. 165-176.<br />
[6] Ye M, Li C, Chen G, Wu J. EECS: An energy efficient<br />
cluster scheme in wireless sensor networks. Proceeding <strong>of</strong><br />
the IEEE IPCCC 2005. New York: IEEE Press, 2005, pp.<br />
535-540.<br />
[7] Depedri A, Zanella A, Verdone R. An energy efficient<br />
protocol for wireless sensor networks. Proceeding <strong>of</strong> the<br />
AINS 2003. Menlo Park, 2003, pp. 1-6.<br />
[8] Heinzelman WR, Chandrakasan AP, Balakrishnan H. An<br />
application-specific protocol architecture for wireless<br />
microsensor networks. IEEE Trans. on Wireless<br />
Communications, 2002, 1(4):660-670.<br />
[9] Smaragdakis G, Matta I, Bestavros A. SEP: A stable election<br />
protocol for clustered heterogeneous wireless sensor<br />
networks. Proceeding <strong>of</strong> the Int’l Workshop on SANPA,<br />
Boston.,no. 4, pp. 660-670, 2004.<br />
[10] Qing li, Zhou Qing-Xin, Wang Ming-Wen. A Distributed<br />
Energy-Efficient Clustering Algorithm for Heterogeneous<br />
Wireless Sensor <strong>Networks</strong> .Chinese <strong>Journal</strong> <strong>of</strong> S<strong>of</strong>tware,<br />
2006, 17(3): 481-489.
1008 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
[11] Handy MJ, Haase M, Timmermann D. Low energy adaptive<br />
clustering hierarchy with deterministic cluster-head selection.<br />
Proceeding <strong>of</strong> the 4th IEEE Conf. on Mobile and Wireless<br />
Communications <strong>Networks</strong>. Stockholm: IEEE<br />
Communications Society, 2002, pp. 368-372.<br />
[12] Lindsey S, Raghavenda CS. PEGASIS: Power efficient<br />
gathering in sensor information systems. In: Williamson DA,<br />
ed. Proceeding <strong>of</strong> the IEEE Aerospace Conf. Vol 3, New<br />
York: IEEE Press, 2002, pp.1125-1130.<br />
[13] Mhatre V, Rosenberg C. Design guidelines for wireless<br />
sensor networks: communication, clustering and aggregation.<br />
Ad Hoc Network <strong>Journal</strong>, 2004, 2(1):45-63.<br />
[14] Manjeshwar A, Agrawal DP. APTEEN: A hybrid protocol<br />
for efficient routing and comprehensive information retrieval<br />
in wireless sensor networks. Proceeding <strong>of</strong> the 2nd Int’l<br />
Workshop on Parallel and Distributed Computing Issues in<br />
Wireless <strong>Networks</strong> and Mobile Computing. Florida: IEEE<br />
Computer Society, 2002, pp.195-202.<br />
[15] Younis M, Youssef M, Arisha K. Energy-Aware routing in<br />
cluster-based sensor networks. Proceeding <strong>of</strong> the 10th IEEE<br />
Int’l Symp. on Modeling, Analysis and Simulation <strong>of</strong><br />
Computer and Telecommunications Systems. Fort Worth:<br />
IEEE Computer Society, 2002, pp.129-136.<br />
[16] Li Cheng-Fa, Chen Gui-Hai,Ye mao, etc..An uneven<br />
cluster-based routing protocol for wireless sensor network.<br />
Chinese journal <strong>of</strong> computers, 2007, 30(1): 27-36.<br />
[17] V. Mhatre and C. Rosenberg. Homogeneous vs.<br />
Heterogeneous Clustered Sensor <strong>Networks</strong>: A Comparative<br />
Study. Proceedings <strong>of</strong> 2004 IEEE International Conference<br />
on Communications, Paris, France, June 2004, Volume 6, pp.<br />
3646-3651.<br />
[18] Hsiao-Lan Hsu, Qilian Liang. An energy-efficient protocol<br />
for wireless sensor networks. In Vehicular Technology<br />
Conference, 2005, Volume: 4, pp. 2321-2325.<br />
[19] Bandyopadhyay S, Coyle EJ. An energy efficient<br />
hierarchical clustering algorithm for wireless sensor<br />
networks. In: Mitchell K, ed. Proceeding <strong>of</strong> the INFOCOM<br />
2003. Vol 3, New York: IEEE Press, 2003, pp.1713-1723.<br />
[20] Yongtao Cao,Chen He. A Distributed Clustering Algorithm<br />
with an Adaptive Back<strong>of</strong>f Strategy for Wireless Sensor<br />
<strong>Networks</strong>. IEICE TRANS. COMMUN. , 2006, Vol. E89-B(2):<br />
609-613.<br />
Fengjun Shang (1972- ), received<br />
the Diploma degree in Intelligent<br />
Instrument at Chengdu University<br />
<strong>of</strong> Technology, China, in 2001. He<br />
finished his Ph.D. in Instrument<br />
Science and Technology at the<br />
College <strong>of</strong> Opto-electronic<br />
Engineering, Chongqing University,<br />
China, in 2005. Since then he<br />
works as an associate pr<strong>of</strong>essor<br />
with the Institute <strong>of</strong> Computer Network Engineer in<br />
Chongqing University <strong>of</strong> Posts and Telecommunications,<br />
China. His research interests include sensor network, traffic<br />
engineer, network optimization and WiMAX.<br />
Email: shangfj@cqupt.edu.cn.<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1009<br />
Robust Cross-layer Design <strong>of</strong> Wireless<br />
Multimedia Sensor <strong>Networks</strong> with Correlation<br />
and Uncertainty<br />
Lei You Chungui Liu<br />
School <strong>of</strong> electronic and information engineering, Tianjin University, Tianjin, China<br />
Email: youlei@tju.edu.cn, cgliumail@163.com<br />
Abstract—Due to content-enriched sensing information and<br />
flexibility, Wireless Multimedia Sensor Network (WMSN)<br />
has a lot <strong>of</strong> potential applications. Low channel capacity,<br />
limited resource, correlation between sensing sources and<br />
uncertain factors make the design and optimization <strong>of</strong><br />
WMSN challenging. In a densely deployed WMSN, there<br />
generally exist correlation and redundancy in the<br />
multimedia information collected by sensors with<br />
overlapped sensing area. In this paper, we adopt a crosslayer<br />
method to deal with the robust lifetime optimization <strong>of</strong><br />
WMSN with correlated sources, which also has energy<br />
consumption uncertainty in the transmission <strong>of</strong> wireless<br />
links. To reduce the number <strong>of</strong> constraints <strong>of</strong> the source rate<br />
region, a pairwise Distributed Source Coding (DSC) scheme<br />
is proposed by matching sensor nodes based their<br />
correlation. A Distributed Pairwise Matching (DPM)<br />
algorithm is thus proposed. With a polyhedral set modeling<br />
the uncertain energy consumption, a robust cross-layer<br />
problem, which maximizes the lifetime <strong>of</strong> the WMSN<br />
through allocating source rate and flow rate on the link (i.e.,<br />
routing) simultaneously under all the possible uncertainty,<br />
is formulated. The counterpart <strong>of</strong> the problem is showed to<br />
be a convex problem with linear constraints, which can be<br />
divided into several separate subproblems in different layers<br />
by dual decomposition. Based on the subgradient method<br />
and DPM, a partially distributed optimization algorithm is<br />
proposed. The algorithm can be implemented in small scale<br />
WMSNs with sink node only responsible for the calculation<br />
and distribution <strong>of</strong> the value <strong>of</strong> the lifetime. Simulation<br />
results verify the performance <strong>of</strong> the proposed algorithm<br />
and show its robustness under uncertainty.<br />
Index Terms—robust cross-layer optimization, wireless<br />
multimedia sensor networks, correlation, uncertainty, dual<br />
decomposition, subgradient.<br />
I. INTRODUCTION<br />
Wireless multimedia sensor network (WMSN) is<br />
emerging with the availability <strong>of</strong> inexpensive visual and<br />
audio hardware such as CMOS cameras and microphones<br />
[1-3]. A sensor node in WMSN is typically comprised <strong>of</strong><br />
multimedia sensor module, microprocessor unit and radio<br />
module. WMSN can sense and transmit video and audio<br />
streams, still images, as well as scalar sensor data through<br />
multihop wireless links between sensors. WMSN is able<br />
to enhance the traditional sensor networks with contentenriched<br />
sensing information and has many new<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1009-1016<br />
applications such as ambient Intelligence [4]. An example<br />
<strong>of</strong> WSMN is illustrated in Figure 1. The multimedia<br />
sensor nodes can be connected wirelessly with neighbors<br />
and the sensed multimedia information (visual or audio)<br />
is transmitted through one or multiple hops to the sink<br />
node.<br />
Monitored<br />
objectives<br />
or events<br />
Servers<br />
Wreless multimedia sensor nodes<br />
Sink node<br />
Figure 1. An example <strong>of</strong> a wireless multimedia sensor networks<br />
Low capacity and limited resource make it challenging<br />
to deliver multimedia data in WMSNs. Due to densely<br />
deployed, multimedia information collected by sensors<br />
with overlapped sensing area is spatially correlated [3].<br />
Thus, eliminating redundancy by exploring the<br />
correlation between nodes can increase the efficiency <strong>of</strong><br />
WMSNs. It was shown that the distributed source coding<br />
(DSC) (Slepian-Wolf coding theorem) [5] can be used to<br />
eliminate the redundancy without explicit communication<br />
between correlated nodes. In addition, the encoder <strong>of</strong><br />
DSC is much simpler than that <strong>of</strong> the state-<strong>of</strong>-the-art<br />
predictive encoding algorithm [6]. Thus, DSC is quite<br />
suited for WMSN with low capacity and limited resource.<br />
Optimal DSC schemes are proposed in [14]. However,<br />
the number <strong>of</strong> constraints to determine the Slepian-Wolf<br />
coding rate region grows exponentially in the number <strong>of</strong><br />
correlated sources involved in the DSC.<br />
Cross-layer optimization <strong>of</strong> wireless communication<br />
networks has been an active research area recent years.<br />
Backpressure-based approaches [15] that determine<br />
routing and scheduling using queue backlog information<br />
are widely used for cross-layer design due to its<br />
optimality. Capacity (stability) region for a time-varying<br />
wireless network is formulated in [17], an optimal routing,<br />
scheduling and power control policy is proposed based on<br />
the backpressure approach in [15]. Joint optimization <strong>of</strong><br />
congestion control, routing and scheduling in wireless<br />
multi-hop networks is studied in [16] [18] using dual
1010 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
decomposition and sub-gradient method. A lot <strong>of</strong> papers<br />
consider the optimization <strong>of</strong> wireless (multimedia) sensor<br />
networks without considering uncertain parameters [7-11]<br />
using the similar methods.<br />
However, ignoring the uncertainty in the design <strong>of</strong><br />
WMSN may results in inefficient energy consumption<br />
and reduced QoS. In the case <strong>of</strong> variable and uncertain<br />
system parameters, a solution that is efficient with all the<br />
possible values <strong>of</strong> the parameters is preferable (although<br />
it may not be optimal for given values). The authors in<br />
[12] considered the uncertainty in the distance between<br />
sensor nodes in wireless sensor networks (WSN),<br />
modeled the uncertainty with polyhedral and ellipsoidal<br />
sets, formulated the robust optimization problems <strong>of</strong><br />
WSN and derived the robust counterpart problems that<br />
are proved to be convex. The problem in [12] is solved<br />
using a centralized algorithm in the programming solver,<br />
which is suitable for the analysis and initialization <strong>of</strong><br />
WSNs but not for the practical operation. In addition,<br />
there are <strong>of</strong>ten strict requirement for the suorce rates from<br />
sensor nodes to the sink in WMSNs, while it only<br />
required some percentage <strong>of</strong> information reaches the sink<br />
in [12].<br />
Taking a cross-layer approach and optimizing<br />
transmission and source rate allocation simultaneously,<br />
especially in the present <strong>of</strong> correlation and uncertainty,<br />
are important for extending the lifetime <strong>of</strong> WMSN. In<br />
this paper, we consider the robust optimization <strong>of</strong> the<br />
lifetime <strong>of</strong> WMSNs with correlated sources and under<br />
energy consumption uncertainty using cross-layer design.<br />
Our aim is to develop an algorithm that is able to be<br />
implemented in the practical operation <strong>of</strong> WMSN. For<br />
this, firstly, a Distributed Pairwise Matching (DPM)<br />
algorithm is proposed to group the sensor nodes in pairs<br />
in order to implement the DSC with reduced complexity.<br />
Then, the transmission energy consumption uncertainty in<br />
each sensor nodes was modeled as a polyhedral set, with<br />
which we showed that the robust cross-layer optimization<br />
problem can be transformed into a convex counterpart<br />
problem with linear constraints. The lifetime<br />
maximization problem <strong>of</strong> such a WMSN is solved by<br />
dual decomposition and subgradient method [13]. The<br />
algorithm we developed not only is (partially) distributed,<br />
but also keeps the layering and modularity architecture <strong>of</strong><br />
the communication network. The algorithm can be<br />
implemented in small scale WMSNs with sink node only<br />
responsible for the calculation and distribution <strong>of</strong> the<br />
value <strong>of</strong> the lifetime.<br />
II. SYSTEM MODEL AND PROBLEM FORMULATION<br />
We consider a static wireless multimedia sensor<br />
network, the sensor nodes <strong>of</strong> which are observing the<br />
same event using certain multimedia sensors (e.g., visual<br />
or audio sensor). We assume homogeneous sensor nodes,<br />
i.e., nodes are deployed with the same sensor. The<br />
WMSN is modelled as a directed graph G � ( N, L)<br />
with a<br />
node set N and a link set L. There are n +1 nodes in N ,<br />
including n sensor nodes (the set <strong>of</strong> which is denoted by<br />
N ) and a sink node. Link between nodes i and j is<br />
s<br />
© 2011 ACADEMY PUBLISHER<br />
denoted by (, i j ) . The maximum transmission rate <strong>of</strong> link<br />
(, i j) is c ij . The set <strong>of</strong> nodes directly connected to node i<br />
is denoted by V i .<br />
A. Pairwise Distributed Source Coding (Pairwise DSC)<br />
Due to densely deployed, multimedia information<br />
collected by nearby sensor nodes are <strong>of</strong>ten correlated and<br />
thus redundant. Since transmission <strong>of</strong> redundant<br />
multimedia data is energy- and bandwidth-consuming, we<br />
employ Slepian-Wolf distributed source coding (DSC) [5]<br />
to eliminate the redundancy. Denote Si as the source<br />
coding rate <strong>of</strong> sensor node i. To construct the sensed<br />
multimedia event in the sink without distortion, the<br />
source coding rates <strong>of</strong> all the sensor nodes must satisfy<br />
the Slepian-Wolf (S-W) source rate region [5] as follow,<br />
c<br />
� Si � H(Z Z ), Z � Ns<br />
(1)<br />
where<br />
i�Z<br />
Z C is the complement <strong>of</strong> Z . c<br />
H (Z Z ) is<br />
conditional entropy<br />
However, the S-W source rate region determined by (1)<br />
is difficult to be used in practical optimization <strong>of</strong><br />
WMSNs due to:<br />
1) The number <strong>of</strong> constraints in (1) grows<br />
exponentially in the number <strong>of</strong> sensor nodes;<br />
2) Global correlation information (joint entropy and<br />
conditional entropy <strong>of</strong> all the possible partitions <strong>of</strong> sensor<br />
nodes) is required.<br />
Here, we propose to implement the DSC in pairs, i.e.,<br />
group two nodes in pair, which will perform the DSC<br />
jointly based on their correlation (joint entropy Hij). (, )<br />
We thus propose the following distributed algorithm to<br />
match the nodes into pairs.<br />
Distributed Pairwise Matching (DPM) algorithm <strong>of</strong><br />
correlated sensor nodes<br />
Initialization: Setting the weight <strong>of</strong> node i and one <strong>of</strong> its<br />
neighbor j, Wij (, ) � Hij (, ) ; initializing the temporary<br />
set <strong>of</strong> neighboring nodes <strong>of</strong> sensor node i, �� i Ns,<br />
as<br />
tmp<br />
Bi � Vi<br />
.<br />
Algorithm:<br />
(1) Each node i sends a ‘matching’ massage to the<br />
neighbor j � arg min H( i, j)<br />
.<br />
tmp<br />
j�Bi (2) If node i receives a ‘matching’ massage from its<br />
neighbor j � arg min H( i, j)<br />
(to which it has sent a<br />
tmp<br />
j�Bi ‘matching’ massage), it sends a ‘matched’ massage<br />
to all its other neighbors. Thus, node i and node<br />
j are matched, i.e., ( i, j) are the pairs to perform<br />
DSC.<br />
(3) If node i receives a ‘matched’ massage from its<br />
neighbor j � arg min H( i, j)<br />
(to which it has sent a<br />
tmp<br />
j�Bi tmp tmp<br />
‘matching’ massage), it sets B � B / j and then<br />
i i
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1011<br />
sends a ‘matching’ massage to the<br />
neighbor j � arg min H( i, j)<br />
.<br />
tmp<br />
j�Bi (4) (1)-(3) repeat until all the node are matched or<br />
B � null<br />
tmp<br />
i<br />
In the DPM algorithm, sensor nodes are grouped<br />
through local negotiation with their neighbors. The sensor<br />
node tries to find a node which has the minimal joint<br />
entropy (maximum correlation) in his potential neighbors<br />
that have not be matched.<br />
With the pairs determined by DPM algorithm, pairwise<br />
DSC is performed between sensor nodes in each pairs.<br />
Note that there may be some nodes that cannot find any<br />
neighboring nodes as his partner. We assume the set <strong>of</strong><br />
such nodes is Q . Let P the set <strong>of</strong> all the pairs determined<br />
by the DPM algorithm. S-W region (1) can be now<br />
written as<br />
Si � H( i j), Sj � H( j i), Si �Sj � H( i, j), �( i, j) �P<br />
Sm� H( m), �m�Q (2)<br />
B. Uncertainty Model <strong>of</strong> Energy Consumption<br />
For the designing <strong>of</strong> WMSNs, optimization with<br />
respect to energy is <strong>of</strong> most important work. However,<br />
Energy consumption may be uncertain due to several<br />
reasons: the inaccurate distance measurement, system<br />
noise or error, changes <strong>of</strong> ambient environment and<br />
deviation <strong>of</strong> the electric circuits. Here, we consider the<br />
uncertainty in energy consumption <strong>of</strong> transmitting<br />
multimedia data on wireless links.<br />
Denote eij the energy consumed by transmitting one bit<br />
on the link ( i, j ) . et � ( eij) (, i j) �Lis<br />
the energy consumption<br />
vector on all links. We argue that there is little correlation<br />
in the energy consumption uncertainty among different<br />
sensor nodes in practice. Thus we can use the following<br />
polyhedral set � to model the uncertainty,<br />
�� �<br />
0<br />
�<br />
���e eij �eij �hij , � hij �� Ri, �i�Ns, h�0�<br />
(3)<br />
�� j�Vi ��<br />
which specifies a set <strong>of</strong> energy consumption vectors that<br />
are within a certain distance ( i R � ) from an nominal<br />
0 0<br />
vector, e � ( eij ) (, i j) �L<br />
, which can be seen as the estimated<br />
energy assumption. The parameter � �[0,1] controls the<br />
level <strong>of</strong> the uncertainty.<br />
In practical WMSN scenarios, we may know neither<br />
which energy consumption vectors in � should be use,<br />
nor the statistic characteristics <strong>of</strong> vectors in � . In this<br />
situation, we must assume that all the instances may<br />
happen. We should guarantee the robustness <strong>of</strong> our<br />
optimal solution under all the possible energy<br />
consumption vectors. Next, we will formulate such robust<br />
cross-layer optimization problem, and solve it using dual<br />
decomposition and subgradient method.<br />
© 2011 ACADEMY PUBLISHER<br />
C. Robust Optimization Problem Formation and its<br />
counterpart<br />
Our goal is to maximize the lifetime <strong>of</strong> a WMSN with<br />
correlated sources and uncertain energy consumptions.<br />
Assume that each node i has an initial energy <strong>of</strong> Ei � 0 .<br />
Let fij be the flow rate <strong>of</strong> link between node i and node<br />
j . The lifetime <strong>of</strong> a node i with flow rate vector<br />
f � ( fij ) (, i j) �L<br />
is<br />
Ti<br />
�<br />
Ei<br />
e f<br />
.<br />
�<br />
j�Vi ij ij<br />
The lifetime <strong>of</strong> an WMSN, T , is defined as the time<br />
when the first sensor node runs out <strong>of</strong> its energy, i.e.,<br />
T � minT . i<br />
i�N Robust cross-layer optimization problem <strong>of</strong> a WMSN<br />
with correlated source is to determine all the source<br />
coding rates Si, i�Ns and the flow rate vector f on the<br />
links, which satisfy the constraints (2) and (4)-(6) under<br />
all the possible energy consumption vectors in � while<br />
maximizes lifetime <strong>of</strong> the WMSN. Specifically, we have<br />
the following robust optimization problem,<br />
max T<br />
st .. Si � H( i j), Sj � H( ji),<br />
Si �Sj � H(, i j), �(, i j) �P<br />
Sm� H( m), �m�Q f � f �S , �i�N (4)<br />
� �<br />
ij ji i s<br />
j�Vi j�Vi �<br />
T( f e ) � E , �i�N , e��<br />
(5)<br />
j�Vi ij ij i s<br />
0 � fij �cij , �( i, j) �L<br />
(6)<br />
where<br />
� the constraints in (4) is the flow conservation law at<br />
each sensor nodes;<br />
� constraints in (5) is the energy conservation law at<br />
each node under all the possible energy<br />
�<br />
consumption vectors;<br />
constraints in (6) represent that the flow rate on a<br />
link should be less than the capacity <strong>of</strong> that link�<br />
The problem above is not a linear problem due to<br />
products in the constraints in (5). Note that the initial<br />
energy Ei � 0 implies T � 0 , and thus we can introduce a<br />
new variable q � 1/ T to obtain a equivalent linear<br />
problem which is<br />
max q<br />
st . . (2), (4), and (6)<br />
f e �qE , �i�NE�� (7)<br />
�<br />
j�Vi ij ij i s<br />
(P1)<br />
For convenience, we denote the above problem as P1.<br />
For robust optimization <strong>of</strong> the lifetime <strong>of</strong> a WMSN,<br />
constraints in (7) should be satisfied under all the possible<br />
energy consumption vectors, which is equivalent to
1012 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
� fe � max � fh � qE, �i�N(8) j�Vi ij<br />
0<br />
ij<br />
� hij ��Ri, �i, h�0<br />
j�Vi j�Vi ij ij i<br />
The dual <strong>of</strong> the linear programming term (i.e.,<br />
max � fh ij ij with variables h ij ) in (8) is<br />
� hij � Ri, i, h 0<br />
j�Vi � � �<br />
j�Vi min<br />
� Ry<br />
i i<br />
st .. y � f , �j�V, y �0<br />
i ij i i<br />
where yi is a dual variable. The constraints (8) is thus<br />
equivalent to<br />
�<br />
j�Vi fe ��Ry� qE<br />
ij<br />
0<br />
ij i i i<br />
y � f , y �0, �i�N, �j�V i ij i i<br />
Finally, we obtain the robust counterpart problem <strong>of</strong><br />
P1 as<br />
min<br />
q<br />
st . . (2), (4), and (6)<br />
0<br />
� fe ij ij ��Ry i i<br />
j�Vi � qEi, �i�N(9) (P2)<br />
y � f , �j�V , y �0,<br />
i ij i i<br />
III. SOLUTION WITH DUAL DECOMPOSITON<br />
In this section, we derive the distributed algorithm for<br />
the robust cross-layer optimization <strong>of</strong> a WMSN by<br />
solving P2 using dual decomposition and subgradient<br />
algorithm.<br />
A. Dual Problem<br />
The problem P2 is a linear (and thus convex) problem.<br />
We can solve it in its dual domain since strong duality<br />
holds for convex problem under mild condition (which is<br />
satisfied for this problem). For a detailed discussion <strong>of</strong><br />
this, refer [13]. However, the objective function <strong>of</strong> P2<br />
(with linear constraints) is not strictly convex and thus the<br />
dual function is not differentiable. We use a similar<br />
approach as [8] to change the objective <strong>of</strong> P2 as follow: 1)<br />
2<br />
Change the objective q to q since objective that<br />
minimizes q is equivalent to the one that minimizes<br />
,<br />
2<br />
q . 2)<br />
Add a small quadratic regularization term for each source<br />
rate and flow rate <strong>of</strong> each link to the objective. Then, the<br />
objectives <strong>of</strong> P3 becomes<br />
2 2 2<br />
q �� f ��<br />
S<br />
�� �<br />
ij i<br />
i�N j�Vi i�N By choosing � and � small enough, the solution <strong>of</strong> the<br />
regularized problem can be arbitrarily close to that <strong>of</strong> the<br />
original one.<br />
By introducing Lagrange multipliers �i for the flow<br />
conservation constraints in (4) and �i for the robust<br />
energy conservation constraints in (9) at each sensor<br />
nodes, the Lagrangian <strong>of</strong> P2 can be given by<br />
© 2011 ACADEMY PUBLISHER<br />
Lq ( , f, y;<br />
�� , )<br />
� q �� f ��<br />
S<br />
�� �<br />
2 2 2<br />
ij i<br />
i�N j�Vi i�N � � �<br />
� � ( f � f �S<br />
)<br />
i ij ji i<br />
i�N j�Vi j�Vi 0<br />
���i( � fe ij ij ��Ry i i �qEi)<br />
i�N j�Vi 2<br />
��i i � �i i �<br />
2<br />
i<br />
i�N i�N �<br />
2<br />
� ( � fij j�Vi 0<br />
� fij ( �ieij ��i��j) i�N � �Ri�iyi �( q �q E ) � ( S � S )<br />
�<br />
�<br />
�<br />
�<br />
�� �<br />
�<br />
�<br />
��<br />
Then, the objective function <strong>of</strong> dual problem can be<br />
written as<br />
D( �� , ) � min L( q, f, y;<br />
�� , )<br />
and the dual problem is<br />
(2) and (6), yi<br />
�0,<br />
yi� fij, q�0<br />
max D(<br />
�� , )<br />
st .. � �0, � � 0<br />
(DP1)<br />
B. Subgradient algorithm<br />
Since dual objective function is not strictly convex,<br />
subgradient method is used to solve DP1. To obtain the<br />
subgradient, we should solve the minimization problem<br />
in D( �� , ) given ��, , which can be decomposed into<br />
three independent subproblems, SP1, SP2, SP3 and SP4<br />
as follows.<br />
SP1:<br />
2<br />
min q � q � E<br />
SP2:<br />
SP3:<br />
st . q�<br />
0<br />
�<br />
�<br />
i�N i i<br />
� f � f �e �� �� ��R�<br />
y<br />
min (<br />
2<br />
ij ij ( i<br />
0<br />
ij i j ) i i i<br />
j�Vi st .. 0 � f �c, y � f , y �0<br />
ij ij i ij i<br />
�S ��S � � S ��S<br />
2 2<br />
min ( i i i ) ( j j j)<br />
st .. S � H( i j), S � H( j i),<br />
i j<br />
Si �Sj � H(, i j), �(, i j) �P<br />
SP4:<br />
2<br />
min ( �mSm ��Sm)<br />
st .. Sm � H( m), �m�Q The subproblems actually correspond to the problems<br />
should be handled by different layers:<br />
� SP1 is to optimize the lifetime <strong>of</strong> the WMSN, and is<br />
an application layer problem.<br />
� SP2 is to determine the flow rate <strong>of</strong> each link, which<br />
can be seen as a routing problem <strong>of</strong> network layer.<br />
Due to the uncertain energy consumption, SP2 is<br />
actually a robust routing problem.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1013<br />
� SP3 is to allocating the source rates between<br />
correlated sources <strong>of</strong> each pair, and is the rate<br />
control problem <strong>of</strong> transport layer. SP4 is also the<br />
source rate problem, whose solution can be readily<br />
obtained as Sm� H( m)<br />
.<br />
� � � � With the obtained value q , Si , fij and yi<br />
by solving<br />
SP1, SP2, SP3 and SP4, the subgradients <strong>of</strong> �D( �� , )<br />
the dual variables �i, �i can be given, respectively, by<br />
at<br />
f �<br />
�<br />
f �<br />
� �<br />
f �S<br />
,<br />
� �<br />
i ij ji i<br />
j�Vi j�Vi i �<br />
�<br />
i �� j�Vi �<br />
ij<br />
0<br />
ij ��i<br />
�<br />
i<br />
g q E f e R y<br />
k k<br />
k �1<br />
(10)<br />
For the k-th steps <strong>of</strong> the subgradient method, the dual<br />
variables are updated as<br />
k�1k �i �( �i ��k<br />
fi)<br />
� ,<br />
k�1k �i �( �i ��kgi)<br />
�<br />
(11)<br />
where ( x) � � x if x � 0 , and ( ) 0 x � � , otherwise. �k is a<br />
positive scalar step-size.<br />
According to [13], subgradient method is guaranteed to<br />
converge to the optimum if step-sizes �k appropriately as follows.<br />
are designed<br />
Theorem 1: Dual variables � and � converge to the<br />
� � optimal dual solutions ( � , � ) if the positive scalar stepsizes<br />
�k are chosen such that<br />
�<br />
lim� �0, � ��<br />
k ��<br />
�<br />
Remark1: Since strong duality holds, the corresponding<br />
� � �<br />
primal variables ( q , r , y ) are globally optimal variables<br />
<strong>of</strong> primal problem P2 (and also P1) for optimal dual<br />
� � variables ( � , � ).<br />
Remark2: We can see that through subgradient update<br />
algorithm, four subprolems are coordinated by Lagrange<br />
multipliers � , �<br />
and cooperatively work with each other to achieve the<br />
whole goal <strong>of</strong> maximizing the lifetime.<br />
� � (i.e., dual variables in the dual problem)<br />
IV. IMPLEMENTATION OF THE PROPOSED SOLUTION<br />
In this section, we would discuss how to implement the<br />
solution in the optimization in a practical wireless<br />
multimedia sensor network. We will show that the<br />
solution obtained in the last section implies a partially<br />
distributed algorithm to calculate the routing (i.e.<br />
determine the flow rate on each link) and the source rate<br />
in the present <strong>of</strong> correlated sources and energy<br />
consumption uncertainty to maximize the lifetime <strong>of</strong> the<br />
WMSN.<br />
Assume that a small scale wireless multimedia sensor<br />
network as illustrated in Figure.1. The sink node is<br />
partially responsible for the organization and<br />
management <strong>of</strong> the whole WMSN. The WMSN suffer the<br />
uncertainty <strong>of</strong> energy consumption which is modeled in<br />
(3) in section II. Assume that the time is slotted. Every<br />
time the correlation <strong>of</strong> the WMSN changes, for example<br />
© 2011 ACADEMY PUBLISHER<br />
the locations or orientations <strong>of</strong> the cameras in wireless<br />
visual sensor networks changes, the whole network<br />
would calculate the routing and source rates with the<br />
following partially distributed algorithm (Algorithm 1)<br />
that is derived from the solution in the last section.<br />
A. Distributed algorirthm<br />
With the pairwise DSC and the subgradient method,<br />
we propose the following distributed algorithm to the<br />
robust cross-layer optimization <strong>of</strong> a WMSN with<br />
correlated sources and uncertainty, which works as<br />
follow:<br />
Algorithm 1: Robust cross-layer optimization <strong>of</strong><br />
WMSN with correlated sources and uncertainty<br />
1) Pairwise matching <strong>of</strong> source sensor nodes using the<br />
DPM algorithm.<br />
0 0<br />
2) Initializing dual variables with �i , �i<br />
At the k-th step, nodes in the WMSN implement the<br />
k k<br />
following operations (dual variables are �i , �i in k-th<br />
step):<br />
3) The sink node calculate the optimal qk ( ) with the all<br />
k the �i , �� i N , which is sent to it from all the sensor<br />
nodes at the end <strong>of</strong> the last step,<br />
k<br />
qk ( ) � � �i<br />
Ei<br />
/2<br />
i�N and then sends the resulting qk ( )<br />
nodes.<br />
to all the sensor<br />
4) Each pair <strong>of</strong> sensor nodes �(, i j) �P solve the<br />
constrained quadratic programming SP3 to obtain<br />
optimal source rate for their pairwise DSC;<br />
5) Each sensor node m<br />
Sm� H( m)<br />
.<br />
�m�Q uses source rate<br />
6) Sensor node i, �� i N solve the quadratic<br />
programming SP2 to obtain optimal fij ( k) and yi( k)<br />
7) Each sensor node i, �� i N calculate the subgradient<br />
value i f and gi as (10), and updates the dual<br />
variables �i , �i as (11).<br />
k 1<br />
8) Every sensor sends the value <strong>of</strong> �i � to the sink node<br />
k 1<br />
and its neighboring nodes, and the value <strong>of</strong> �i � neighboring nodes.<br />
9) Go to step 3) and repeat until convergence.<br />
to its<br />
Remarks 1: To facilitate the distributed implementation<br />
<strong>of</strong> subgradient method in the above algorithm and reduce<br />
the massage pass in each step, we use a small constant<br />
step size � that is same for all the nodes and in all the<br />
steps. The subgradient method can converge to a small<br />
neighborhood <strong>of</strong> the optimal solution <strong>of</strong> DP1 with<br />
constant step size that is small enough [13].<br />
Remarks 2: The calculation <strong>of</strong> the source coding rate in<br />
step 4) can be implemented in any sensor node <strong>of</strong> the<br />
pairs, and the results are reported to the other node by<br />
massage passing. We assume that nodes in pairs know<br />
their joint entropy and conditional entropy.
1014 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
B. Sumary<br />
We show the whole picture <strong>of</strong> the robust cross-layer<br />
design <strong>of</strong> WMSN with correlated sources and uncertainty<br />
in Figure 2. The original problem is a lifetime<br />
maximization problem with constraints <strong>of</strong> link rate,<br />
source rate, energy conservation, and robustness<br />
guarantee. Though we can transform the problem into a<br />
linear programming, it still involves a lot <strong>of</strong> variables.<br />
Centralized algorithm like simplex or interior-point<br />
methods can be used to solve it. However, they are not<br />
efficient approaches to be implemented in the practical<br />
operation <strong>of</strong> the WMSNs. Sink node needs to gather all<br />
the information (topology, link, uncertainty, correlation<br />
and so on), calculates the solution centrally, and<br />
distributes the results to every node. This will not only<br />
shorten the lifetime <strong>of</strong> the whole network due to energy<br />
consumption in information collection and distribution,<br />
but also increase the delay in the delivery <strong>of</strong> multimedia<br />
data. It may <strong>of</strong>fset the benefits from cross-layer<br />
optimization.<br />
Robust lifetime optimization problem with<br />
correlated sources and uncertainty<br />
Application layer:<br />
Objective <strong>of</strong> the network<br />
Transport layer:<br />
Source rate control<br />
Network layer:<br />
Routing (flow rates on<br />
links)<br />
Goblal variables<br />
dual variables<br />
local variables<br />
dual variables<br />
local variables<br />
dual variables<br />
Cross-layer<br />
coordination<br />
by dual<br />
variables<br />
update<br />
Figure 2. Cross-layer design architecture <strong>of</strong> robust optimization <strong>of</strong><br />
WMSN with correlated sources and uncertainty<br />
In this work, we deal with the problem in its dual<br />
domain. The main advantage <strong>of</strong> doing this is that dual<br />
problem can be further divided into some separate<br />
subproblems, which correspond to the functions in<br />
different protocol layers. As shown in Figure 2, the<br />
subproblems are, respectively,<br />
� the application layer problem (system objective, i.e.,<br />
lifetime maximization),<br />
� transport layer problem (source rate control) and<br />
� network layer problem (routing).<br />
Dual problem can be solved using subgradient method.<br />
Local variables (except lifetime variable in Algorithm 1)<br />
are needed to update the subgradient, and thus it can be<br />
© 2011 ACADEMY PUBLISHER<br />
used to develop a distributed algorithm. Except<br />
application layer problem, the other two problems can be<br />
implemented distributedly in Algorithm 1.<br />
Dual variables are “bridges” connecting and<br />
coordinating the separate subproblems in different layers.<br />
As shown in Figure 2, through dual variables, our crosslayer<br />
design not only achieves the optimal objective, but<br />
also keeps the layering architecture developed for the<br />
computer networks, which has the advantages <strong>of</strong><br />
modularity.<br />
V. PERFORMANCE EVALUATION<br />
In this section, we conduct simulations to investigate<br />
the performance <strong>of</strong> the proposed algorithms. The<br />
simulations are implemented in MATLAB s<strong>of</strong>tware.<br />
In our scenario setup, we assume the following radio<br />
model for the nominal energy consumption in the<br />
0 t 2<br />
transmission on a wireless link: ei, j� e di,<br />
j . The unit <strong>of</strong><br />
t<br />
2<br />
e is J bytes m and di, j is the distance between<br />
transmitter i and receiver j . There are 30 sensor nodes<br />
and one sink, which are randomly deployed in an area <strong>of</strong><br />
2<br />
200� 200 m . The initial energy <strong>of</strong> every sensor node is<br />
10,000 t<br />
e .Set 100 t<br />
Ri�e and Ci, j�18,<br />
�( i, j) � L.<br />
For simplicity, we assume that the correlation between<br />
source nodes only depends on the distance between them.<br />
We use the following correlation model: Hij (, ) � Hd 0 i, j<br />
and Hij ( ) � H( ji) � Hd 1 i, j . In practical scenarios with<br />
camera sensors, more complex models, which may<br />
depend on not only distance but also orientation and focal<br />
length, can be used. We set H0 � 0.02 and H1 � 0.005 .<br />
A. Performance <strong>of</strong> DPM algorithm<br />
We compare the performance <strong>of</strong> DPM algorithm with a<br />
random matching (RM) algorithm. For random matching<br />
algorithm, each sensor node finds a partner in its<br />
neighbors randomly by massage passing. We replace the<br />
DPM in Algorithm 1 with the random matching<br />
algorithm for Algorithm 1 with RM, and show the<br />
resulting lifetime <strong>of</strong> Algorithm 1 with DPM and RM in<br />
Figure 3.<br />
The value <strong>of</strong> lifetime<br />
2.8<br />
2.6<br />
2.4<br />
2.2<br />
2<br />
1.8<br />
1.6<br />
Algorithm 1 with DPM<br />
Algorithm 1 with RM<br />
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9<br />
Uncertianty parameter �<br />
Figure 3. Comparison <strong>of</strong> DPM and RM
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1015<br />
We can see from Figure 3 that the lifetime obtained<br />
by Algorithm 1 with DPM is larger than that with RM<br />
(more than 23% for this scenario). In addition, the<br />
lifetime decreases with the increase <strong>of</strong> the uncertainty<br />
parameter. This is because the robust algorithm would<br />
sacrifice some performance to confront the uncertainty.<br />
This is desirable, especially when the uncertainty is<br />
relative large (as illustrated in Figure 4).<br />
B. Performance <strong>of</strong> Robustness<br />
We compare Algorithm 1 with a deterministic solution<br />
which does not consider the uncertainty in energy<br />
consumption. The deterministic solution optimizes<br />
0<br />
WMSN with the nominal energy consumption (i.e., e i, j).<br />
To show the benefit <strong>of</strong> the robust optimization, we<br />
calculate the following two ratios:<br />
L �L L �L<br />
R � R �<br />
L L<br />
det robust robust det<br />
0 un un wc<br />
e e e e<br />
1 , 2<br />
det det<br />
0<br />
e<br />
wc<br />
e<br />
det where L 0 is the optimal lifetime <strong>of</strong> deterministic<br />
e<br />
robust<br />
solution; L un is the lifetime obtained by robust<br />
e<br />
optimization considering the uncertainty; det<br />
L wc is the<br />
e<br />
lifetime <strong>of</strong> deterministic solution in the worst energy<br />
consumption.<br />
The ratio R1 is the relative decrease <strong>of</strong> optimal lifetime<br />
<strong>of</strong> robust solution in the nominal case, while ratio<br />
R2 reflects the relative increase <strong>of</strong> the lifetime <strong>of</strong> robust<br />
solution over deterministic solution in the waste case. The<br />
results are shown in Figure 4.<br />
We can observe from Figure 4 that R2 is larger than<br />
R1 and the difference increases with the increase <strong>of</strong> � .<br />
This means that robust solution is more desirable when<br />
uncertainty becomes large. It guarantees the optimality<br />
under all the instances <strong>of</strong> the uncertain energy<br />
consumption vectors with small performance loss<br />
(compared with the deterministic energy consumption).<br />
The value <strong>of</strong> ratios: R1 and R2<br />
0.12<br />
0.1<br />
0.08<br />
0.06<br />
0.04<br />
0.02<br />
0<br />
R1<br />
R2<br />
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9<br />
Uncertianty parameter �<br />
Figure 4. Comparison <strong>of</strong> the ratio R1 and R2<br />
© 2011 ACADEMY PUBLISHER<br />
VI. CONCLUSION AND FUTURE WORK<br />
Robust cross-layer design <strong>of</strong> wireless multimedia<br />
sensor network with correlated sources and uncertain<br />
energy consumption was investigated. To reduce the<br />
redundant information and complexity, distributed source<br />
coding was implemented in pairs which were determined<br />
by a distributed pairwise matching algorithm. A<br />
polyhedral set was used to model the uncertainty in<br />
energy consumption. Robust lifetime maximization<br />
problem was formulated and its counterpart was showed<br />
to be a convex problem with linear constraints. Dual<br />
decomposition and subgradient method were used to<br />
solve the problem, which leads to a cross-layer design<br />
approach and a partially distributed solution. Dual<br />
variables were used to coordinate the separate<br />
subproblems in each layer, and thus layering architecture<br />
and modularity were maintained in the cross-layer design.<br />
The obtained distributed algorithm can be implemented in<br />
small scale WMSNs with sink node responsible for the<br />
calculation and distribution <strong>of</strong> the value <strong>of</strong> the lifetime.<br />
Simulation results verified the performance <strong>of</strong> the<br />
proposed algorithms, especially the benefits <strong>of</strong> robust<br />
optimization under large uncertainty case.<br />
ACKNOWLEDGMENT<br />
This work is partially supported by Innovation<br />
Foundation <strong>of</strong> Tianjin University (Grant No.60302014),<br />
Tianjin Science and Technology Support Program (Grant<br />
No.09ZCKFGX0170), NSFC <strong>of</strong> China (Grant<br />
No.60972054), 863 Program <strong>of</strong> China (Grant<br />
No.2009AA011507, SQ2009AA01XK1485134), and<br />
Major Programs <strong>of</strong> Ministry <strong>of</strong> Science and Technology<br />
<strong>of</strong> China (Grant No.2009ZX03004-006).<br />
REFERENCES<br />
[1] I. F. Akyildiz, T. Melodia, K. R. Chowdhury, “A survey on<br />
wireless multimedia sensor networks,” Comput. Netw.<br />
(Elsevier), vol. 51, no. 4, Mar. 2007, pp. 921–960.<br />
[2] Akyildiz I. F., Melodia T., Chowdury K. R., "Wireless<br />
Multimedia Sensor <strong>Networks</strong>: Applications and Testbeds,"<br />
Proceedings <strong>of</strong> the IEEE (invited paper), vol. 96, no. 10,<br />
October 2008, pp. 1588-1605.<br />
[3] Akyildiz, I. F., Melodia, T., Chowdury, K. R., "Wireless<br />
Multimedia Sensor <strong>Networks</strong>: A Survey," IEEE Wireless<br />
Communications Magazine, vol. 14, no. 6, December 2007,<br />
pp.32-39.<br />
[4] Nakashima Hideyuki, Aghajan Hamid, Augusto Juan<br />
Carlos, Handbook <strong>of</strong> Ambient Intelligence and Smart<br />
Environments. Springer Press. 2010.<br />
[5] D. Slepian, J. K.Wolf, “Noiseless coding <strong>of</strong> correlated<br />
information sources,” IEEE Trans. Inform. Theory, vol. IT-<br />
19, Jul. 1973, pp. 471–480.<br />
[6] X. Guo, Y. Lu, F. Wu, W. Gao, S. Li, “Distributed<br />
multiviewvideo coding,” Proc. Visual Communications<br />
and Image Processing (VCIP), San Jose, CA, USA,<br />
January 2006.<br />
[7] A. Sankar, Z. Liu, “Maximum lifetime routing in wireless<br />
ad-hoc networks,” in Proc. IEEE INFOCOM 2004, pp.<br />
1089–1097. IEEE press. Hong Kong (2004).<br />
[8] R. Madan, S. Lall, “Distributed algorithms for maximum<br />
lifetime routing in wireless sensor networks,” in Proc.
1016 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
IEEE GLOBECOM 2004, pp.2185–2193. IEEE press.<br />
Dallas (2004).<br />
[9] Ilias Politis, Michail Tsagkaropoulos, Tasos Dagiuklas,<br />
Stavros Kotsopoulos, “Power Efficient Video Multipath<br />
Transmission over Wireless Multimedia Sensor <strong>Networks</strong>”.<br />
Mobile Netw Appl, Vol.13, 2008, pp.274–284<br />
[10] F. Ordonez, B. Krishnamachari, “Optimal information<br />
extraction in energy-limited wireless sensor networks,”<br />
IEEE J. Select. Areas Commun., vol. 22, no. 6, 2004, pp.<br />
1121–1129.<br />
[11] Jamal N. Al-Karaki, Raza Ul-Mustafa, Ahmed E. Kamal,<br />
“Data aggregation and routing in Wireless Sensor<br />
<strong>Networks</strong>: Optimal and heuristic algorithms”. Computer<br />
<strong>Networks</strong>, vol. 53, no. 7, 2009, pp. 945-960.<br />
[12] Wei Ye, Fernando Ord´o˜nez. “Robust Optimization<br />
Models for Energy-Limited Wireless Sensor <strong>Networks</strong><br />
under Distance Uncertainty”. IEEE transactions on<br />
wireless communications, vol. 7, no. 6, Jun. 2008, pp.<br />
2161-2169.<br />
[13] S. Boyd, L. Xiao, and A. Mutapcic, “Subgradient<br />
methods,” in lecturenotes <strong>of</strong> EE392o, Stanford University,<br />
Autumn Quarter 2003-2004.<br />
[14] R. Cristescu, B. Beferull-Lozano, and M. Vetterli,<br />
“Networked Slepian-Wolf: Theory, algorithms, and scaling<br />
laws,” IEEE Trans. Inform. Theory, vol. 51, no. 12, pp.<br />
4057–4073, Dec. 2005.<br />
[15] L. Tassiulas and A. F. Ephremides. , “Stability properties<br />
<strong>of</strong> constrained queueing systems and scheduling policies<br />
for maximum throughput in multihop radio networks, ”<br />
IEEE Transactions on Automatic Control, vol. 37, no. 12,<br />
Dec.1992, pp. 1936–1948.<br />
[16] X. Lin and N. Shr<strong>of</strong>f, “The impact <strong>of</strong> imperfect scheduling<br />
on cross-layer congestion control in wireless networks,”<br />
IEEE/ACM Trans. Networking, vol. 14, no. 2, April 2006,<br />
pp. 302–315.<br />
[17] M. Neely, E. Modiano, and C. Rohrs, “Dynamic power<br />
allocation and routing for time-varying wireless networks,”<br />
IEEE J. Select. Areas Commun., vol. 23, no. 1, Jan. 2005,<br />
pp. 89–103.<br />
[18] L. Chen, S. H. Low, M. Chiang, and J. C. Doyle, “Crosslayer<br />
congestion control, routing and scheduling design in<br />
ad hoc wireless networks,” in Proc. <strong>of</strong> IEEE Infocom, Apr.<br />
2006, pp.1-13.<br />
Lei You received the Ph.D degree from Beijing University <strong>of</strong><br />
Posts and Telecommunications, Beijing, China, in 2010. During<br />
2008.9-2009.8, He was a visiting PhD student with Uppsala<br />
University, Uppsala, Sweden. He is currently a lecture <strong>of</strong> the<br />
school <strong>of</strong> Electronic Information Engineering at Tianjin<br />
University. His research interests include wireless multi-hop<br />
networks and cross-layer optimization.<br />
Chungui Liu received the Ph.D degree in Computer<br />
Applications from Tianjin University, Tianjin, China, in<br />
2009. He is currently working as a Post-Doctor with<br />
school <strong>of</strong> Electronic Information Engineering at Tianjin<br />
University. His research interests include wireless mesh<br />
networks and Internet <strong>of</strong> Things (IoT).<br />
© 2011 ACADEMY PUBLISHER
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1017<br />
The E-Commerce Model <strong>of</strong> Health Websites: An<br />
Integration <strong>of</strong> Web Quality, Perceived<br />
Interactivity, and Web Outcomes<br />
Chung-Hung Tsai<br />
Department <strong>of</strong> Health Administration, Tzu Chi College <strong>of</strong> Technology, R.O.C.<br />
Email: tsairob@tccn.edu.tw<br />
Abstract—The study integrates web quality (system<br />
quality, information quality, and service quality),<br />
perceived interactivity (human-message, humanhuman),<br />
and web outcomes (web usage, web<br />
satisfaction, and web loyalty) to explore the ecommerce<br />
model <strong>of</strong> health websites. A survey <strong>of</strong> 1076<br />
users <strong>of</strong> health websites was conducted to validate the<br />
proposed model. The findings show that web quality<br />
has significantly positive effect on perceived<br />
interactivity, web usage, and web satisfaction<br />
separately, which in turn influence web loyalty. This<br />
study also confirms that perceived interactivity is an<br />
important mediator between web quality and web<br />
outcomes. This study emphasizes the importance <strong>of</strong><br />
both web quality and perceived interactivity in the<br />
progress towards success health websites. The findings<br />
may be used as theoretical base for future research<br />
and can also <strong>of</strong>fer empirical foresight to executives<br />
and managers <strong>of</strong> hospitals when they initially<br />
introduce and upgrade the health websites into their<br />
organizations.<br />
Index Terms—web quality, perceived interactivity, web<br />
usage, web satisfaction, web loyalty<br />
I. INTRODUCTION<br />
Ever since the Internet emerged in the 1990s, a great<br />
number <strong>of</strong> situations and modes <strong>of</strong> commercial<br />
competitions have had tremendous changes. Lots <strong>of</strong><br />
customer-oriented service industries have started to set up<br />
the platforms and portals on the Internet to serve the<br />
customers so that the customers can be connected with<br />
the services the organizations <strong>of</strong>fer at any time no matter<br />
how far they are or where they are. Moreover, many<br />
traditional commercial behaviors can also be conducted<br />
with the long-distance virtual transactions through the<br />
Internet. The Internet has changed not only the modern<br />
people’s living styles but also the transaction modes<br />
between contemporary businesses and customers.<br />
Therefore, the representative organizations, hospitals,<br />
which <strong>of</strong>fer the service <strong>of</strong> medical treatments, have also<br />
gradually valued this tendency and trend.<br />
Hospitals can take advantage <strong>of</strong> the websites to<br />
provide the patients and their families with the health<br />
information so that they can learn the latest knowledge <strong>of</strong><br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1017-1024<br />
medicine. The mechanism <strong>of</strong> the online registration can<br />
also be utilized to create another channel to seek medical<br />
advice. Alternatively, the conferences or seminars<br />
regarding the health education can be regularly held in<br />
the hospitals and post the contents onto the websites for<br />
the general public to browse the films and download the<br />
briefings and other materials. Lots <strong>of</strong> healthcare courses<br />
can even be <strong>of</strong>fered online. For instance, the prenatal<br />
healthcare course for the couples, the smoking cessation<br />
course, bodybuilding course and so on. These functions<br />
are no longer just the static introduction to each clinic and<br />
department. Instead, they have become the virtual<br />
clearinghouse for the health information [1]. At present,<br />
numerous websites <strong>of</strong> hospitals and medical institutions<br />
have been equipped with these features.<br />
Health website is not only a health communication<br />
channel but also a full representation <strong>of</strong> a service<br />
department to the customer. With the establishment <strong>of</strong> the<br />
reliable and popular health websites, the customers can be<br />
provided with the general healthcare information, and<br />
hospital-customer relationships can be further reinforced.<br />
In addition, studies have showed that the customers<br />
satisfied with a website would have higher level <strong>of</strong><br />
customer loyalty [2]. Thus, it’s critical to investigate how<br />
the websites quality to affect health web outcomes (e.g.<br />
web usage, web satisfaction, web loyalty).<br />
In addition, Internet has transformed the traditional<br />
physician-patient relationship because those who use the<br />
Internet frequently ask their physicians more specific<br />
questions and suggest specific illnesses and treatments [3].<br />
Lustria [4] also showed that the use <strong>of</strong> interactive<br />
technology could enhance learning and persuasion <strong>of</strong><br />
content on the basis <strong>of</strong> well health web quality. Therefore,<br />
it’s very crucial to explore the issue <strong>of</strong> interactivity <strong>of</strong><br />
health websites.<br />
Accordingly, the purpose <strong>of</strong> the study is to examine<br />
how web quality (system, information, and service<br />
quality) and perceived interactivity (human-message and<br />
human-human) affect the websites’ loyalty through users’<br />
attitude toward websites (website usage and satisfaction).<br />
The research model will be empirically tested using the<br />
structural equation modeling (SEM). Through the<br />
statistical analysis, we can investigate the interaction<br />
between technological and social factors, and furthermore<br />
find out the important antecedents <strong>of</strong> websites’ loyalty. In<br />
this way, we hope to provide the managers <strong>of</strong> hospitals
1018 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
and the administrators <strong>of</strong> the IS department with the<br />
insight and reference regarding the management <strong>of</strong><br />
hospital websites.<br />
II. LITERATURE REVIEW<br />
2.1 Web Quality<br />
Aladwani and Palvia [5] define website quality as<br />
users’ evaluation <strong>of</strong> a website’s features meeting users’<br />
needs and reflecting overall excellence <strong>of</strong> the web site.<br />
Hwang and Kim [6] on the other hand define website<br />
quality as the user’s perception on the customer service<br />
and privacy based on the website interface and functions.<br />
Liu and Arnett [7] conducted a survey on the top<br />
thousand businesses as listed in Fortune magazine, and<br />
found four factors that relevantly affected the success <strong>of</strong><br />
the website: (1) information and service quality, (2)<br />
system use, (3) playfulness, and (4) system design quality.<br />
In DeLone and McLean’s research [8] on e-commerce<br />
systems and the measurement <strong>of</strong> quality, it was found that<br />
other than system and information quality, the importance<br />
<strong>of</strong> customers support in e-commerce system was essential,<br />
thus emphasizing the importance <strong>of</strong> the quality <strong>of</strong> service.<br />
Ahn, Ryu, and Han [9] believe that the technologyfocused<br />
approach sees a website as an information system<br />
and focuses on system and information quality, while a<br />
service-focused approach see a website as a service<br />
provider and includes service quality. According to the<br />
arguments <strong>of</strong> the above literature, even though<br />
measurements for website quality may be changed<br />
according to research purpose or field, the main<br />
classification still uses the categories <strong>of</strong> system,<br />
information, and service quality.<br />
2.2 Perceived Interactivity<br />
Most scholars agree that interactivity is a strong and<br />
important marketing characteristic <strong>of</strong> the World Wide<br />
Web as compared to other traditional media (e.g.<br />
television, newspaper, magazines, etc.). After review <strong>of</strong><br />
29 articles on interactivity, McMillan and Hwang [10]<br />
classified the various scholarly definitions <strong>of</strong> interactivity<br />
in four categories: (1) process, (2) feature, (3) perception,<br />
and (4) a combination <strong>of</strong> the three. McMillan [11] found<br />
that website interactivity was based on two-way<br />
communication, levels <strong>of</strong> control, user activity, sense <strong>of</strong><br />
place, and time sensitivity. McMillan [12] further pointed<br />
out that there are three dimensions <strong>of</strong> perceived<br />
interactivity: two-way communication, controls <strong>of</strong><br />
navigation (or choices), and time to load/time to find. Wu<br />
[13] found that perceived interactivity was the<br />
psychological state experienced by the website visitor<br />
during the interaction process. It consists <strong>of</strong> three<br />
dimensions: perceived controls, perceived responsiveness,<br />
and perceived personalization. Ko, Roberts, and Cho [14]<br />
define perceived interactivity as “the degree to which<br />
people engage in a communication process by actively<br />
interacting with mediated messages and other people.”<br />
They found that the two most frequently occurring and all<br />
encompassing dimensions were human-message<br />
interactivity and human-human interactivity.<br />
© 2011 ACADEMY PUBLISHER<br />
From a healthy communication perspective on the<br />
World Wide Web, Cassell, Jackson, and Cheuvront [1]<br />
argued that the World Wide Web was suitable for<br />
persuasive communication. This type <strong>of</strong> communication<br />
is a form <strong>of</strong> social influence that can effectively affect<br />
internalization <strong>of</strong> specific attitudes and in turn affect<br />
behaviors. Also, Lustria [4] found that the users <strong>of</strong> web’s<br />
hypertext are able to freely browse through the system,<br />
and randomly access material, processing information<br />
according to individual mental models, and redefine<br />
learning structure and content. This research also showed<br />
that high levels <strong>of</strong> perceived interactivity promoted high<br />
levels <strong>of</strong> comprehension <strong>of</strong> the content <strong>of</strong> website. Thus,<br />
high levels <strong>of</strong> perceived interactivity <strong>of</strong> website would<br />
lead to strengthen the effect <strong>of</strong> online learning and<br />
persuasion so that the user feels the convenience,<br />
usefulness, and enjoyment <strong>of</strong> visiting the websites. Due<br />
to the aforementioned reasons, this study uses “perceived<br />
interactivity” to measure website interactivity. Besides,<br />
perceived interactivity includes two dimensions: humanmessage<br />
interactivity and human- human interactivity.<br />
2.3 The relationship between website quality and<br />
perceived interactivity<br />
Wu [13] suggested a conceptual structure for the<br />
antecedent and consequential variables in interactivity.<br />
That is, website factors (actual interactivity, vividness,<br />
and design), site-visitor factors (personality traits, product<br />
knowledge, and web skills), and situational factors (visit<br />
motivation, access speed, and visit location) are three<br />
types <strong>of</strong> factors that influence perceived interactivity.<br />
Song and Zinkhan [15] manipulated 16 different<br />
versions <strong>of</strong> a website in an experiment directed at college<br />
students to prove that speed and message format<br />
(personalization <strong>of</strong> messages) would positively influence<br />
perceived interactivity. Because the response time <strong>of</strong> the<br />
website is a characteristic <strong>of</strong> system quality, and website<br />
messages are one kind <strong>of</strong> information quality, it is<br />
predicted that website quality may affect perceived<br />
interactivity. Therefore, this research proposes the<br />
following hypothesis:<br />
H1a: Website quality has a positive effect on<br />
perceived interactivity.<br />
According to information systems success model<br />
proposed by DeLone and McLean [16], system quality,<br />
information quality, and service quality are related to<br />
usage and user satisfaction <strong>of</strong> an information system.<br />
DeLone and McLean [8] showed that D&M IS Success<br />
Model also could be adapted to the measurement <strong>of</strong> the ecommerce<br />
systems.<br />
Hwang and Kim [6] also showed that website quality<br />
would positively influence affective reaction <strong>of</strong> users,<br />
which is a subjective perception or judgment about<br />
whether such interaction will change their core affect or<br />
emotion toward the website. Therefore, this research<br />
proposes the following hypothesis:<br />
H1b: Website quality has a positive effect on website<br />
usage.<br />
Hic: Website quality has a positive effect on website<br />
satisfaction.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1019<br />
2.4 The relationship between website quality and<br />
perceived interactivity<br />
Jee and Lee [17] found that perceived interactivity has<br />
positive impact on the user’s attitudes toward the website<br />
and intent <strong>of</strong> purchase. McMillan, Hwang, and Lee [18]<br />
found that “perceived interactivity” was a better predictor<br />
<strong>of</strong> user attitudes toward the website than “actual<br />
interactivity”. Similarly, Song and Zinkhan [15] also<br />
found that perceived interactivity was positively<br />
correlated with website loyalty and attitude. Therefore,<br />
this research proposes the following hypotheses:<br />
H2a: Perceived interactivity has a positive effect on<br />
website usage.<br />
H2b: Perceived interactivity has a positive effect on<br />
website satisfaction.<br />
2.5 The relationship between website use, website<br />
satisfaction, and website loyalty<br />
DeLone and McLean [8,16] argued that system usage<br />
and user satisfaction both affect the user’s net benefit.<br />
Besides, Dick and Basu [19] found that sustainable<br />
loyalty could only be achieved when the customer enjoys<br />
a high level <strong>of</strong> positive attitudes (satisfaction) toward the<br />
product as well as a high level <strong>of</strong> repetitive patronage.<br />
III. RESEARCH METHOD<br />
In order to provide the public with correct and current<br />
online health information, Department <strong>of</strong> Health (DOH)<br />
hold the activities <strong>of</strong> excellent awards <strong>of</strong> health<br />
information websites to assess the websites <strong>of</strong> all<br />
hospitals in Taiwan since 2002. We mail the invitation<br />
letters to the executives <strong>of</strong> hospitals who have obtained<br />
the excellent awards to express our need for the research<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 1. The Proposed Research Model<br />
In the research <strong>of</strong> virtual community websites, Kuo<br />
[20] found that there was a relationship between<br />
continuous usage and overall satisfaction, and continuous<br />
usage, satisfaction, and loyalty were also related. Otim<br />
and Grover [21] proved that product satisfaction<br />
influences customers repeat purchase intention (loyalty)<br />
in the context <strong>of</strong> website service. Kassim and Abdullah<br />
[22] also proved that customer satisfaction has positive<br />
direct effect on loyalty in e-commercial environments.<br />
Casalo, Flavian, and Guinaliu [2] also found that<br />
customers satisfy with previous interactions with the bank<br />
websites has positive effect on customer loyalty, and<br />
website usability has also positive effect on customer<br />
loyalty. Therefore, this research proposes the following<br />
hypotheses:<br />
H3a: Website usage has a positive effect on website<br />
satisfaction.<br />
H3b: Website usage has a positive effect on website<br />
loyalty.<br />
H4: Website satisfaction has a positive effect on<br />
website loyalty.<br />
Based on the review <strong>of</strong> the literature, figure 1 presents<br />
the conceptual framework from which the proposed<br />
research model is formed.<br />
purpose. Of these hospitals contacted, five teaching<br />
hospitals (located in northern, central, southern Taiwan)<br />
were willing to participate in the survey. Before we can<br />
conduct the survey, it must be approved by IRB<br />
(Institutional Review Board) <strong>of</strong> hospitals. Distribution<br />
and collection <strong>of</strong> survey questionnaires was coordinated<br />
with the help <strong>of</strong> the executives, and information systems<br />
managers <strong>of</strong> the hospitals.<br />
We used a self-report questionnaire to empirically<br />
validate the proposed research model. The questionnaire
1020 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
was pilot tested using 30 hospitals’ patients who had<br />
prior experience in online websites. These items were<br />
revised according to the feedback. After the revision, the<br />
survey was conducted to a convenient sample <strong>of</strong> 1200<br />
patients for four months. Of the 1200 samples, the<br />
samples with incomplete responses and missing data<br />
were deleted. Finally, the eligible samples <strong>of</strong> 1076<br />
patients were yielded, and the total response rate is<br />
89.67%.<br />
IV. RESULT<br />
The data analysis proceeded according to a two-step<br />
approach [23]. First, we assessed the measurement model,<br />
which consists <strong>of</strong> the six latent factors, including the<br />
assessment <strong>of</strong> reliability, discriminant validity, and<br />
convergent validity <strong>of</strong> the scales. Second, we validated<br />
the structural model, which represents the series <strong>of</strong> path<br />
relationships linking the six constructs.<br />
4.1 Sample Characteristics<br />
Of these respondents, 653 respondents are women<br />
(60.7%), 37.6% are age 30 and below. The education<br />
levels <strong>of</strong> mostly respondents are university (40.8%). The<br />
majority <strong>of</strong> respondents’ career belongs to service<br />
industry (24.1%). Mostly respondents lived in northern<br />
Taiwan (49.8%). The times using the internet is mostly<br />
10 times and above per week (32.1%), while 1~3 times<br />
and above per day (46.6%). Table I presents descriptive<br />
statistics for the seven constructs in the study. The mean<br />
scores for seven constructs are all almost on the middle<br />
point <strong>of</strong> 5-point Likert-type scales, and show a<br />
reasonable dispersion in their distributions across the<br />
ranges.<br />
Table I Sample demographics<br />
Construct Mean<br />
Standard<br />
Deviation<br />
Minimum Maximum<br />
Web Quality 3.96 0.50 2.33 5.00<br />
System Quality 4.03 0.61 2.00 5.00<br />
Information Quality 3.97 0.51 2.00 5.00<br />
Service Quality 3.90 0.61 2.00 5.00<br />
Perceived Interactivity 4.07 0.61 2.50 5.00<br />
Human-Message 4.12 0.66 1.00 5.00<br />
Human-Human 4.01 0.69 2.00 5.00<br />
Web Usage 3.83 0.75 1.67 5.00<br />
Web Satisfaction 3.98 0.63 2.00 5.00<br />
Web Loyalty 3.93 0.74 1.50 5.00<br />
4.2 Measurement Model Results<br />
To validate the measurement model, three types <strong>of</strong><br />
validity were assessed: content validity, convergent<br />
validity, and discriminant validity. Content validity was<br />
done by interviewing senior system users and pilottesting<br />
the instrument. And the convergent validity was<br />
validated by examining Cronbach’s α, composite<br />
reliability and average variance extracted from the<br />
measures [24]. As shown in Table II, the Cronbach’s α <strong>of</strong><br />
every subscales range from 0.81 to 0.92, which are above<br />
the acceptability value 0.7 [25]. Besides, the composite<br />
reliability values range from 0.81 to 0.90, and the<br />
average variances extracted by our measures range from<br />
0.52 to 0.76, are all within the commonly accepted range<br />
greater than 0.5 [24]. In addition, all measures are<br />
significant on their path loadings at the level <strong>of</strong> 0.001.<br />
Therefore, the convergent validities <strong>of</strong> all seven<br />
constructs are confirmed.<br />
Discriminant validity <strong>of</strong> the sub-dimensions <strong>of</strong><br />
2<br />
constructs was validated by comparing the � values <strong>of</strong><br />
the CFA with original sub-dimensions <strong>of</strong> every construct<br />
© 2011 ACADEMY PUBLISHER<br />
against other CFAs which every possible combination <strong>of</strong><br />
two dimensions (the correlation coefficient <strong>of</strong> two<br />
dimensions assigned to be 1) was examined. As shown in<br />
2<br />
Table III, the � values <strong>of</strong> the CFA with original subdimensions<br />
<strong>of</strong> web quality (system quality, information<br />
quality, and service quality) and perceived interactivity<br />
(human-message and human-human) were significantly<br />
better than any possible union <strong>of</strong> any two dimensions.<br />
Therefore, the discriminant validities <strong>of</strong> the subdimensions<br />
<strong>of</strong> the two constructs are confirmed.<br />
Besides, according to Fornell and Larcker [26],<br />
discriminant validity can also be tested among all<br />
constructs by comparing the average variance extracted<br />
(AVE) <strong>of</strong> each construct with the squared correlation <strong>of</strong><br />
that construct and all the other constructs. As shown in<br />
Table IV, all squared correlations between two<br />
constructs are less than the average variance extracted <strong>of</strong><br />
both constructs. Therefore, the results confirm that the<br />
discriminant validity <strong>of</strong> constructs in the study is<br />
satisfactory.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1021<br />
Web Quality<br />
Table II Construct Reliability and Convergent Validity<br />
Construct Cronbach’s α Composite Reliability<br />
Average Variance<br />
Extracted<br />
Web Quality 0.92 0.90 0.76<br />
System Quality 0.85 0.85 0.60<br />
Information Quality 0.82 0.81 0.52<br />
Service Quality 0.87 0.87 0.63<br />
Perceived Interactivity 0.88 0.90 0.76<br />
Human-Message 0.86 0.87 0.68<br />
Human-Human 0.81 0.82 0.60<br />
Web Usage 0.86 0.86 0.68<br />
Web Satisfaction 0.81 0.81 0.60<br />
Web Loyalty 0.90 0.90 0.67<br />
Table III Discriminant Validity <strong>of</strong> Sub-Dimensions <strong>of</strong> Web Quality and Perceived Interactivity<br />
Model<br />
2<br />
� d.f. Δ 2<br />
�<br />
1.Not Restricted 440.878 51 -<br />
2.System Quality and Information Quality Assigned to 1 908.829 52 467.951***<br />
3. System Quality and Service Quality Assigned to 1 778.147 52 337.269***<br />
4. Information Quality and Service Quality Assigned to 1 845.428 52 404.550***<br />
Perceived Interactivity<br />
1.Not Restricted 56.569 13 -<br />
2. Human-Message and Human-Human Assigned to 1 403.170 27 346.601***<br />
*** p
1022 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Table V Fit Indices for the Structural Model<br />
Structural Model Statistic Fit Indexes<br />
Recommended<br />
Threshold<br />
Figure 2 illustrate the results <strong>of</strong> the structural model<br />
with the estimated standardized path coefficients and<br />
path significance among constructs (non-significant<br />
paths as dotted lines). As predicted, all proposed<br />
hypotheses are supported. Table VI illustrates the<br />
squared multiple correlations (R 2 ) <strong>of</strong> all endogenous<br />
variables in the model. The estimated standardized path<br />
coefficients indicate the strengths <strong>of</strong> the relationships<br />
between the dependent and independent variable.<br />
Meanwhile the R 2 value represents the proportion <strong>of</strong><br />
variance that is explained by the predictors <strong>of</strong> the<br />
variable in the model.<br />
As expected, web quality (β=0.940) has significant<br />
effects on perceived interactivity, accounting for 88.4%<br />
<strong>of</strong> the variance in the construct. Web quality (β=0.292)<br />
and perceived interactivity ( β =0.546) have all<br />
© 2011 ACADEMY PUBLISHER<br />
2<br />
� 1458.789 -<br />
2<br />
� / d.f. 4.355 < 5<br />
GFI 0.90 > 0.9<br />
RMSEA 0.056 < 0.08<br />
AGFI 0.88 > 0.8<br />
NFI 0.93 > 0.9<br />
RFI 0.92 > 0.9<br />
IFI 0.95 > 0.9<br />
TLI 0.94 > 0.9<br />
CFI 0.95 > 0.9<br />
Figure 2 Final Proposed Model<br />
significant effects on web usage, accounting for 68.2% <strong>of</strong><br />
the variance in the construct. Besides, web quality (β<br />
=0.239), perceived interactivity ( β =0.454), and web<br />
usage ( β =0.287) have significant effects on web<br />
satisfaction, accounting for 87.5% <strong>of</strong> the variance in the<br />
construct. Web usage (β=0.292), and web satisfaction<br />
(β=0.627) are both significant predictors <strong>of</strong> web loyalty,<br />
accounting for 78.9% <strong>of</strong> the variance in the construct.<br />
The results <strong>of</strong> the structural model show that web<br />
quality (system quality, information quality, and service<br />
quality), perceived interactivity (human-message and<br />
human-human) are two key aspects affecting web<br />
outcomes <strong>of</strong> hospitals’ websites (web usage, web<br />
satisfaction, and web loyalty). The results also<br />
demonstrate that web quality has significant impact on<br />
web outcomes mediated by perceived interactivity.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1023<br />
V. DISCUSSION<br />
This study proposed a research model to better<br />
understand the e-commerce model <strong>of</strong> health websites.<br />
The model considered the relationships among web<br />
quality (system quality, information quality, and service<br />
quality), perceived interactivity (human-message and<br />
human-human), and web outcomes (web usage, web<br />
satisfaction, and web loyalty). Moreover, the model aims<br />
to interpret that perceived interactivity is an important<br />
mediator between web quality and web outcomes. The<br />
results <strong>of</strong> this study are discussed below.<br />
The results <strong>of</strong> this study suggest that web quality<br />
consists <strong>of</strong> three dimensions: (1) system quality, (2)<br />
information quality, and (3) service quality. Furthermore,<br />
positive perceptions <strong>of</strong> health web quality predict<br />
customers’ perceived interactivity, web usage, and web<br />
satisfaction. Previous studies having found similar results.<br />
Hwang and Kim [6] proposed a conceptual framework to<br />
interpret how web quality influences affective reaction.<br />
Also, Ha and Stoel [27] proposed the extended<br />
technology acceptance model <strong>of</strong> online shopping<br />
techniques to show high quality e-shopping sites should<br />
result in the perception that one’s experience is enjoyable<br />
and trust in e-shopping. It implies that online customers<br />
are more likely to feel positive affective reaction (usage<br />
and satisfaction) when they feel the health website is<br />
well-designed, knowledgeable, and responsive.<br />
This study also confirms that perceived interactivity is<br />
an important mediator between web quality and web<br />
outcomes. Perceived interactivity has significantly<br />
positive effect on health web usage and web satisfaction.<br />
The findings also support previous empirically research<br />
(e.g. [15]). The finding proved the mediating role <strong>of</strong><br />
perceived interactivity in affecting the effect <strong>of</strong> web<br />
quality on online customers’ perception <strong>of</strong> web usage<br />
and web satisfaction. It is consistent with the results <strong>of</strong><br />
previous research (e.g. [28,29]). Interestingly, the<br />
empirical evidence supports again the mediating role <strong>of</strong><br />
perceived interactivity.<br />
The integrative viewpoint implies that an online health<br />
website is not only an information system but also but<br />
also a service provider/department to the customer.<br />
Accordingly, in the developing and maintaining phase <strong>of</strong><br />
online health websites, system engineers and managers<br />
should value simultaneously system functions, health<br />
contents, and follow up service in order to pursue better<br />
web quality. On other hand, perceived interactivity plays<br />
important roles in web outcomes. It implies that online<br />
customers are more likely to continue to use a health<br />
website when they feel the health web is playful and<br />
© 2011 ACADEMY PUBLISHER<br />
Table VI The Squared Multiple Correlations (R 2 )<br />
Construct R 2<br />
Perceived Interactivity 0.884<br />
Web Usage 0.682<br />
Web Satisfaction 0.875<br />
Web Loyalty 0.789<br />
interactive. Some strategies that managers <strong>of</strong> hospitals<br />
could use to increase the level <strong>of</strong> perceived interactivity,<br />
such as quick navigation, personalized web page,<br />
transmission <strong>of</strong> relevant messages, value-added search<br />
mechanism, bulletin boards, multi-media contents, etc.<br />
REFERENCES<br />
[1] M. M. Cassell, C. Jackson, and B. Cheuvront,<br />
“Health Communication on the Internet: An<br />
effective channel for health behavior change?,”<br />
<strong>Journal</strong> <strong>of</strong> Health Communication, Vol. 3, pp. 71-79,<br />
1998.<br />
[2] L. V. Casalo, C. Flavian, and M. Guinaliu, “The role<br />
<strong>of</strong> satisfaction and website usability in developing<br />
customer loyalty and positive word-<strong>of</strong>-mouth in the<br />
e-banking services,” International <strong>Journal</strong> <strong>of</strong> Bank<br />
Marketing, Vol. 26, No. 6, pp. 399-417, 2008.<br />
[3] K. Altinkemer, P. De, and Z. D. Ozdemir, “Toward<br />
a Consumer-to-Healthcare provider (C2H) electronic<br />
marketplace,” Communications <strong>of</strong> the Association<br />
for Information Systems, Vol. 18, pp. 413-430, 2006.<br />
[4] M. L. A. Lustria, “Can interactivity make a<br />
difference? effects <strong>of</strong> interactivity on the<br />
comprehension <strong>of</strong> and attitudes toward online health<br />
content,” <strong>Journal</strong> <strong>of</strong> the American Society for<br />
Information Science and Technology, Vol. 58, No. 6,<br />
pp. 766-776, 2007.<br />
[5] A. M. Aladwani, and P. C. Palvia, “Developing and<br />
validating an instrument for measuring userperceived<br />
web quality,” Information and<br />
Management, Vol. 39, pp. 467-476, 2002.<br />
[6] Y. Hwang, and D. J. Kim, “Customer self-service<br />
systems: the effects <strong>of</strong> perceived web quality with<br />
service contents on enjoyment, anxiety, and e-trust,”<br />
Decision Support Systems, Vol. 43, pp. 746-760,<br />
2007.<br />
[7] C. Liu, and K. P. Arnett, “Exploring the factors<br />
associated with web site success in the context <strong>of</strong><br />
electronic commerce,” Information and<br />
Management, Vol. 38, pp. 23-33, 2000.<br />
[8] W. H. DeLone, and E. R. McLean, “Measuring ecommerce<br />
success: applying the DeLone and<br />
McLean Information Systems Success Model,”<br />
International <strong>Journal</strong> <strong>of</strong> Electronic Commerce, Vol.<br />
9, No. 1, pp. 31-47, 2004.<br />
[9] T. Ahn, S. Ryu, and I. Han, “The impact <strong>of</strong> web<br />
quality and playfulness on user acceptance <strong>of</strong> online<br />
retailing,” Information and Management, Vol. 44,<br />
pp. 263-275, 2007.
1024 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
[10] S. J. McMillan, and J. Hwang, “Measures <strong>of</strong><br />
perceived interactivity: an exploration <strong>of</strong> the role <strong>of</strong><br />
direction <strong>of</strong> communication, user control, and time<br />
in shaping perceptions <strong>of</strong> interactivity,” <strong>Journal</strong> <strong>of</strong><br />
Advertising, Vol. 31, pp. 29-42, 2002.<br />
[11] S. J. McMillan, “Interactivity is in the eye <strong>of</strong> the<br />
beholder: function, perception, involvement, and<br />
attitude toward the web site,” in Proceedings <strong>of</strong> the<br />
2000 Conference <strong>of</strong> the American Acadenn/ <strong>of</strong><br />
Advertising. Mary A. Shaver, ed. East Lansing, MI:<br />
Michigan State University, pp. 71-78, 2000.<br />
[12] S. J. McMillan, “A four-part model <strong>of</strong> cyberinteractivity:<br />
some cyber-spaces are more interactive<br />
than others,” New Media and Society, Vol. 4 (June),<br />
pp. 271-291, 2002.<br />
[13] G. Wu, “Conceptualizing and measuring the<br />
perceived interactivity <strong>of</strong> websites,” <strong>Journal</strong> <strong>of</strong><br />
Current Issues and Research in Advertising, Vol. 28,<br />
No. 1, pp. 87-104, 2006.<br />
[14] H. K. Ko, M. S. Roberts, and C. Cho, “Crosscultural<br />
differences in motivations and perceived<br />
interactivity: a comparative study <strong>of</strong> American and<br />
Korean internet users,” <strong>Journal</strong> <strong>of</strong> Current Issues<br />
and Research in Advertising, Vol. 28, No. 2, pp. 93-<br />
104, 2006.<br />
[15] J. H. Song, and G. M. Zinkhan, “Determinants <strong>of</strong><br />
perceived web site interactivity,” <strong>Journal</strong> <strong>of</strong><br />
Marketing, Vol. 72, pp. 99-113, 2008.<br />
[16] W. H. DeLone, and E. R. McLean, “The DeLone<br />
and McLean Model <strong>of</strong> Information Systems Success:<br />
a ten-year update,” <strong>Journal</strong> <strong>of</strong> Management<br />
Information Systems, Vol. 19, No. 4, pp. 9-30, 2003.<br />
[17] J. Jee, and W. N. Lee, “Antecedents and<br />
consequences <strong>of</strong> perceived interactivity: an<br />
exploratory study,” <strong>Journal</strong> <strong>of</strong> Interactive<br />
Advertising, Vol. 3, No. 1, [available at http://<br />
www.jiad.org], 2002.<br />
[18] S. J. McMillan, J. Hwang, and G. Lee, “Effects <strong>of</strong><br />
structural and perceptual factors on attitudes toward<br />
the website,” <strong>Journal</strong> <strong>of</strong> Advertising Research, Vol.<br />
43, pp. 400-409, 2003.<br />
[19] A. S. Dick, and K. Basu, “Customer loyalty: toward<br />
an integrated conceptual framework,” <strong>Journal</strong> <strong>of</strong> the<br />
<strong>Academy</strong> <strong>of</strong> Marketing Science, Vol. 22, No. 2, pp.<br />
99-113, 1994.<br />
[20] Y. F. Kuo, ”A study on service quality <strong>of</strong> virtual<br />
community websites,” Total Quality Management<br />
and Business Excellence, Vol. 14, No. 4, pp. 461-<br />
473, 2003.<br />
[21] S. Otim, and V. Grover, “An empirical study on<br />
web-based services and customer loyalty, European<br />
<strong>Journal</strong> <strong>of</strong> Information Systems, Vol. 15, pp. 527-<br />
541, 2006.<br />
[22] N. M. Kassim, and N. A. Abdullah, “Customer<br />
loyalty in e-commerce settings: an empirical study,”<br />
Electronic Markets, Vol. 18, No. 3, pp. 275-290,<br />
2008.<br />
© 2011 ACADEMY PUBLISHER<br />
[23] J. C. Anderson, and D. W. Gerbing, “Structural<br />
Equation Modeling in practice: a review and<br />
recommended two-step approach,” Psychological<br />
Bulletin, Vol. 103, No. 3, pp. 411-423, 1988.<br />
[24] J. F. Hair, R. E. Anderson, R. L. Tatham, and W. C.<br />
Black, Multivariate Data Analysis, NJ: Prentice Hall,<br />
1998.<br />
[25] J. C. Nunnally, Psychometric Theory, 2 nd Ed., New<br />
York, NY: McGraw-Hill, 1978.<br />
[26] C. Fornell, and D. F. Larcker, “Evaluating Structural<br />
Equation Models with unobservable variables and<br />
measurement error,” <strong>Journal</strong> <strong>of</strong> Marketing Research,<br />
Vol. 18, pp. 39-50, 1981.<br />
[27] S. Ha, and L. Stoel, “Consumer e-shopping<br />
acceptance: antecedents in a technology acceptance<br />
model”, <strong>Journal</strong> <strong>of</strong> Business Research, Vol. 62, No.<br />
5, pp. 565-571, 2009.<br />
[28] A. Basso, D. Goldberg, S. Greenspan, and D.<br />
Weimer, “First impressions: emotional and cognitive<br />
factors underlying judgments <strong>of</strong> trust e-commerce,”<br />
In Proceedings <strong>of</strong> the 3rd ACM Conference on<br />
Electronic Commerce, 137-143, Tampa, FL, USA,<br />
2001.<br />
[29] G. Wu, “The mediating role <strong>of</strong> perceived<br />
interactivity in the effect <strong>of</strong> actual interactivity on<br />
attitude toward the website,” <strong>Journal</strong> <strong>of</strong> Interactive<br />
Advertising, Vol. 5, No. 2, [available at<br />
http://www.jiad.org], 2005.<br />
Chung-Hung Tsai is assistant pr<strong>of</strong>essor and director <strong>of</strong><br />
Department <strong>of</strong> Health Administration at Tzu Chi College <strong>of</strong><br />
Technology. He received his Ph. D. degree from National<br />
Dong-Hwa University. He is currently one member <strong>of</strong> the<br />
editorial board and reviewer <strong>of</strong> <strong>Journal</strong> <strong>of</strong> Healthcare<br />
Management.<br />
His current research areas are knowledge management system,<br />
health information system, e-commerce, and<br />
telemedicine/telecare/telehealth system management. His<br />
academic papers have been published in Technological<br />
Forecasting and Social Change (SSCI), International <strong>Journal</strong> <strong>of</strong><br />
Information Technology and Management (EI), Key<br />
Engineering Material (EI), International <strong>Journal</strong> for Quality<br />
Research (SCIndeks), <strong>Journal</strong> <strong>of</strong> e-Business (TSSCI), <strong>Journal</strong><br />
<strong>of</strong> Technology Management, MIS Review, <strong>Journal</strong> <strong>of</strong><br />
American <strong>Academy</strong> <strong>of</strong> Business (ABI), Electronic Commerce<br />
Studies, <strong>Journal</strong> <strong>of</strong> Business Administration, <strong>Journal</strong> Customer<br />
Satisfaction, and <strong>Journal</strong> <strong>of</strong> Health Management.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1025<br />
A New Method <strong>of</strong> Time-frequency Synthesis <strong>of</strong><br />
Harmonic Signal Extraction from Chaotic<br />
Background<br />
Erfu Wang<br />
Key Laboratory <strong>of</strong> Electronics Engineering, College <strong>of</strong> Heilongjiang Province<br />
School <strong>of</strong> Electronic Engineering, Heilongjiang University, Harbin, China<br />
Email: efwang_612@163.com<br />
Zhifang Wang, Jing Ma and Qun Ding<br />
Key Laboratory <strong>of</strong> Electronics Engineering, College <strong>of</strong> Heilongjiang Province<br />
School <strong>of</strong> Electronic Engineering, Heilongjiang University, Harbin, China<br />
Email: {zhifang.w@gmail.com, majing20041499@126.com, qunding@yahoo.cn}<br />
Abstract—The separation <strong>of</strong> chaos and signal is an<br />
important problem <strong>of</strong> chaos signal processing. In recent<br />
years, the time-frequency analysis method is more and more<br />
mature. It can extract the time-domain character and<br />
frequency-domain character at the meantime. Timefrequency<br />
method can mainly carry out the problem <strong>of</strong><br />
extraction from continuous chaos system background;<br />
achieve separation between chaos and signal according to<br />
different time-frequency character <strong>of</strong> chaos signal, noise<br />
signal and harmonic signal. So it can get useful signal from<br />
chaotic background. This paper first introduced the basic<br />
theory <strong>of</strong> time-frequency methods. Use the wavelet method<br />
and empirical mode decomposition method to analyze the<br />
extraction performance <strong>of</strong> harmonic signal from chaos<br />
background according to the different noise situation. After<br />
compare the wavelet method and empirical mode<br />
decomposition method, we summarize a new<br />
complementary synthesis method <strong>of</strong> harmonic signal<br />
extraction combine the wavelet threshold and empirical<br />
mode decomposition according to the experiments and<br />
simulation. Computer simulation verified that the methods<br />
have high availability.<br />
Index Terms—Harmonic Signal, Time-frequency Analysis,<br />
Extraction, Chaos, wavelet<br />
I. INTRODUCTION<br />
Chaos is widespread in the various domains, such as<br />
chaos secure communication, ECM and heart computer<br />
signal processing[1,2]. Chaos theory attracts a lot <strong>of</strong><br />
attention by the scholars in the last decade.<br />
Many researchers introduced several methods <strong>of</strong><br />
detecting, separating and extracting signals according to<br />
the chaos’ different characters. Leung uses smallest phase<br />
space capacity method to estimate polynomial parameter<br />
insert chaos[3]. Native Fuping Wang and some other ones<br />
use chaos attractor geometric properties, realize the<br />
separation between chaos interference and weak signal by<br />
the concept <strong>of</strong> differential manifold tangent space[4,5].<br />
Haykin research the extraction <strong>of</strong> small objectives signal<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1025-1032<br />
from the ocean noise which is proved as chaos noise by<br />
the way <strong>of</strong> artificial nerve network[6~9]. Short research<br />
the extraction from chaos communication system in the<br />
way <strong>of</strong> chaos forecast method according to the<br />
characteristic <strong>of</strong> short-time forecast. These methods<br />
carved out a new field <strong>of</strong> chaos signal processing, but<br />
lack <strong>of</strong> systematicness. Some methods are rigorous and<br />
weak applicability, and demand target signal smaller than<br />
chaos background signal[10~13]. Time-frequency<br />
analysis theory is more popular in recent years. Timefrequency<br />
method can mainly carry out the problem <strong>of</strong><br />
extraction from continuity chaos system background,<br />
achieve separation between chaos and signal according to<br />
different time-frequency character <strong>of</strong> chaos signal, noise<br />
signal and harmonic signal[14~17]. So it can get useful<br />
signal from chaos background. When the signal-to-noise<br />
ratio (SNR) is not so weak the extraction effect will be<br />
perfect.<br />
Comparing wavelet method with empirical mode<br />
decomposition(EMD) method, according to the<br />
performance analysis <strong>of</strong> harmonic signal extraction from<br />
chaos background in different noise level and signal-tonoise<br />
ratio(SNR), finally give the extraction produce<br />
about complementary scheme. After theory analysis and a<br />
lot <strong>of</strong> simulation, the paper will give the corresponding<br />
results and analysis.<br />
II. BASIC THEORY OF TIME-FREQUENCY METHODS<br />
Time-frequency methods are good at deal with nonstationary<br />
signal. It can extract the time-domain character<br />
and frequency-domain character at the meantime. Which<br />
is the most classical wavelet transform method, EMD as a<br />
new method <strong>of</strong> time-frequency signal processing in some<br />
application can get better than the wavelet transform.<br />
A. Chaotic System<br />
We choose the chaotic system Lorenz system to<br />
simulate and experiment. Lorenz system is a three-
1026 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
(a) Wavelet method<br />
(b) EMD method<br />
Figure 1. The recovery harmonic signal by wavelet method and EMD<br />
method when NL=10%<br />
dimensional continuous dynamic system, its nonlinear<br />
state equations are defined as followed:<br />
� dx<br />
�<br />
� �a(<br />
x � y)<br />
dt<br />
�<br />
� dy<br />
� � �xz<br />
� bx � y<br />
(1)<br />
� dt<br />
� dz<br />
� xy � cz<br />
��<br />
dt<br />
In the following part <strong>of</strong> the simulations and<br />
experimentations we choose a=10, b=28 and c=8/3 as the<br />
parameters. We get x0=y0=z0=0.1 and the step length is<br />
0.01.We iterate 4000 points and use the 1900 to 4000<br />
from x axis to be the chaotic background sequence.<br />
B. Wavelet Transform<br />
Wavelet transform started to develop as a timefrequency<br />
Analysis method from the anaphase 20th<br />
century. To the present signal x(t) ∈ L2(R), the<br />
continuous wavelet transform(CWT) <strong>of</strong> signal x(t) is<br />
defined as:<br />
© 2011 ACADEMY PUBLISHER<br />
�<br />
� �<br />
� � �<br />
� 1<br />
� t b<br />
WT x ( a,<br />
b)<br />
x(<br />
t)<br />
� [ ] dt<br />
��<br />
a<br />
a<br />
(2)<br />
x(<br />
t),<br />
� a,<br />
b ( t)<br />
In the Eq.2: a>0 is the scale factor and b is the shifted<br />
factor. 1 t � b<br />
� a,<br />
b(<br />
t)<br />
� � [ ] The equation is called wavelet<br />
a a<br />
primary function is the shift and scale dilation <strong>of</strong> the<br />
generating wavelet. Wavelet transform is a kind <strong>of</strong><br />
correlated calculation between the originality signals and<br />
the group <strong>of</strong> wavelet functions after dilation essentially.<br />
The analysis process based on wavelet transform is the<br />
decomposition and reconfiguration process virtually. The<br />
primary wavelet Daubechies is continuous, orthogonal<br />
and easy to implement. So the wavelet analysis part use<br />
the Daubechies. The db6 is chosen due to its good<br />
localized character and orthogonal character.<br />
C. Wavelet threshold de-noising theory<br />
Because <strong>of</strong> wavelet transform is linear, wavelet<br />
transformation coefficient is additive. When elect wavelet<br />
matching with signal to conduct wavelet<br />
transformation ,signal energy mainly focus on wavelet<br />
coefficient <strong>of</strong> a few sparse and amplitude relatively large<br />
on some frequency band, and wavelet transform <strong>of</strong> white<br />
noise is still white noise, which is widely distributed in<br />
each dimension <strong>of</strong> time axis and amplitude is not big. So<br />
we can set a threshold, using the threshold to adjust<br />
wavelet coefficients according to certain rules. After<br />
adjustment, various wavelet coefficients reconstruct<br />
signal according to the inversion algorithm for getting<br />
target signal. This is the wavelet threshold de-noising<br />
theory basis. This paper choose heursure threshold.<br />
D. EMD Method<br />
EMD (Empirical Mode Decomposition) is a<br />
trenchancy instrument to analyze non-linear and nonstationary<br />
signal and is instituted by N. E. Huang etc<br />
originally. EMD method is based on the concept <strong>of</strong><br />
instantaneous frequency on the basis <strong>of</strong> in-depth study,<br />
and the corresponding Hilbert transformation is close<br />
related to the method. Decompose non-linear and nonstationary<br />
signals can obtain a series IMF (Intrinsic Mode<br />
Function) which express the signal character and time<br />
scale. IMF is narrowband stationary signal. IMF must<br />
satisfied the following two conditions:<br />
(1)The discrepancy <strong>of</strong> zero-crossing and the extreme<br />
point is zero or one;<br />
(2)At every point, the mean <strong>of</strong> local maximum<br />
envelope and local minimum envelope is zero;<br />
Only decompose the signal into some IMF, analyze<br />
every instantaneous frequency can reveal true physics<br />
sense <strong>of</strong> original signal.<br />
III.UNDER DIFFERENT CIRCUMSTANCES OF TIME-<br />
FREQUENCY METHOD<br />
A. The Influence <strong>of</strong> noise level(NL) on the Extraction<br />
Effect<br />
The Lorenz system, Gauss noise and s(n)=Asin(2πfn)<br />
mixed the composite signal, and fix A=5,f=5Hz. The
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1027<br />
(a) Wavelet method<br />
(b) EMD method<br />
Figure 2. The recovery harmonic signal by wavelet method and EMD<br />
method when NL=30%<br />
TABLE I. EXTRACTION EFFECT COMPARISON IN DIFFERENT NL<br />
NL 10% 30% 50% 70% 100%<br />
R-wavelet 0.9608 0.8370 0.6850 0.5271 0.4273<br />
R-EMD 0.9451 0.8285 0.8404 0.6237 0.5545<br />
noise level (NL) means that the standard ratio <strong>of</strong> Gauss<br />
noise to chaotic interference. Observe wavelet transform<br />
method and EMD method on harmonic signal extraction<br />
effect after changing NL.<br />
Two methods <strong>of</strong> extraction effect when NL=10%,<br />
NL=30% and NL=70%.<br />
From Fig. 1, Fig. 2and Fig. 3 we can see that two<br />
methods are perfect when there is definite noise in the<br />
chaotic background and NL is small. The waveform<br />
almost distorted by wavelet method when the noise is<br />
much bigger, then this method can not extract harmonic<br />
signal. EMD method is still useful and the effect is<br />
perfect. The correlation coefficient is 0.5271 when use<br />
wavelet method, and the correlation coefficient is 0.7286<br />
when use EMD method. From Fig. 3 we can know that<br />
when the NL=70%, the recovery level has a large<br />
improvement and the extraction performance is more<br />
perfect.<br />
© 2011 ACADEMY PUBLISHER<br />
(a) Wavelet method<br />
(b) EMD method<br />
Figure 3. The recovery harmonic signal by wavelet method and EMD<br />
method when NL=70%<br />
When NL are 10%,30%,50%,70%,100%, signal is<br />
extracted from chaotic background with noise by use the<br />
wavelet method and EMD method. TABLE I gives<br />
quantitative comparison directly. R-wavelet means the<br />
correlation coefficient between recovery harmonic signal<br />
and original harmonic signal when use wavelet method.<br />
R-EMD means the correlation coefficient between<br />
recovery harmonic signal and original harmonic signal<br />
when use EMD method.<br />
We can get a conclude from TABLE I by quantitative<br />
results compared <strong>of</strong> wavelet method and EMD<br />
method .When NL50%, EMD method is better than<br />
wavelet method. The extraction effect is perfect.<br />
B. The Influence <strong>of</strong> SNR on the Extraction effect<br />
The Lorenz system, Gauss noise and s(n)=A*sin(2πfn)<br />
mixed the composite signal, and fix A=5,f=5Hz.<br />
According to the influence <strong>of</strong> NL on the extraction<br />
performance, and fix NL=30%. SNR means the energy<br />
ratio <strong>of</strong> harmonic signal to chaotic interference and Gauss<br />
noise. Compare the extraction performance <strong>of</strong> wavelet<br />
method and EMD method after changing Harmonic
1028 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
(a) Wavelet method<br />
(b) EMD method<br />
Figure 4. The recovery harmonic signal by wavelet method and EMD<br />
method when SNR=-5<br />
TABLE II. EXTRACTION EFFECT COMPARISON IN DIFFERENT SNR<br />
SNR -1 -5 -10 -20 -30<br />
R-wavelet 0. 8949 0.7945 0.5907 0.2381 0.0271<br />
R-EMD 0.8582 0.8398 0.7438 0.2589 0.1789<br />
Signal Amplitude A and SNR.<br />
Two methods <strong>of</strong> extraction effect when SNR=-5 and<br />
SNR=-20.<br />
From Fig. 4 and Fig. 5 we can see that two methods are<br />
perfect when there is a little noise in the chaotic<br />
background. When SNR=-5 the wavelet method is more<br />
perfect, the waveform almost distorted by wavelet<br />
method when the SNR is much lower (SNR=-20), then<br />
this method can not extract harmonic signal. EMD<br />
method is more perfect.<br />
When SNR are -1,-5,-10,-20,-30, we use <strong>of</strong> wavelet<br />
method and EMD method for extraction method <strong>of</strong><br />
harmonic signal from chaotic background with noise.<br />
TABLE II directly gives the quantitative <strong>of</strong> correlation<br />
coefficient comparison.<br />
We can get conclude from TABLE II by quantitative<br />
results compared <strong>of</strong> wavelet method and EMD method.<br />
© 2011 ACADEMY PUBLISHER<br />
(a) Wavelet method<br />
(b) EMD method<br />
Figure 5. The recovery harmonic signal by wavelet method and EMD<br />
method when SNR=-20<br />
When NL=30% in noise level and SNR>-5, wavelet<br />
method is better than EMD method. This wavelet<br />
decomposition stability plays a leading role. When<br />
SNR
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1029<br />
(a)NL=10%<br />
(b)NL=30%<br />
Figure 6. The extraction performance <strong>of</strong> Wavelet threshold and EMD method<br />
TABLE III. DE-NOISING PERFORMANCE OF WT(WAVELET<br />
TRANSFORM) AND EMD<br />
SNR Wavelet EMD<br />
Parameter set first give wavelet primary need not<br />
function and series<br />
first set<br />
Convergence<br />
rate<br />
slow<br />
fast<br />
Stability stable unstable<br />
SNR higher lower<br />
can infer that WT and EMD are complementary in denoising<br />
performance from the simulation result<br />
mentioned above.<br />
Due to complementary characteristics <strong>of</strong> the Wavelet<br />
method and EMD in all aspects, it was switched between<br />
wavelet method and EMD method for choosing better<br />
extracted performances <strong>of</strong> harmonic signal in the chaos <strong>of</strong><br />
background according to different noise level and SNR<br />
condition.<br />
IV. A NEW METHOD OF HARMONIC SIGNAL EXTRACTED<br />
W e w i l l d iscuss a comp r e h e n s i v e method<br />
complementary advantages which combines wavelet<br />
transform, threshold de-noising and EMD method<br />
according to complementary characteristics <strong>of</strong> the<br />
© 2011 ACADEMY PUBLISHER<br />
(c)NL=70%.<br />
(d)NL=100%.<br />
TABLE IV. THE CORRELATION COEFFICIENT OF MIXED<br />
METHOD AFTER EVERY STEP SEPARATION IN THE DIFFERENT NL<br />
NL 10% 30% 50% 70% 100%<br />
r-xbyz 0.9809 0.8500 0.6764 0.5665 0.4385<br />
R-EMD 0.9772 0.7554 0.8241 0.7614 0.5692<br />
Wavelet method and EMD in the noise level and SNR<br />
aspect and two methods also can be used in de-noising<br />
speed and stability.<br />
A. Simulation experiment and analysis<br />
Experimental one: The Lorenz system, Gauss noise<br />
and s(n)=A*sin(2πfn) mixed the composite signal, and fix<br />
A=5,f=5Hz. Using wavelet method for mixed signal, then<br />
using EMD method , and observe the separation<br />
performance.<br />
We can give the separation performance when<br />
NL=10%, NL=30%, NL=70% and NL=100%. “original<br />
signal” is harmonic signal .“recovery-xbyz” means that<br />
extraction <strong>of</strong> signal after the wavelet threshold de-noising.<br />
“recovery-EMD” is recovery harmonic signal after the<br />
EMD method.<br />
TABLE IV is that the correlation coefficient <strong>of</strong> mixed<br />
method after every step separation in the different NL.
1030 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
(a)NL=10%<br />
(c)NL=70%.<br />
(b)NL=30%<br />
(d)NL=100%<br />
Figure 7. The extraction effect <strong>of</strong> Wavelet threshold and EMD method<br />
TABLE V. THE CORRELATION COEFFICIENT OF MIXED<br />
METHOD AFTER EVERY STEP SEPARATION IN THE DIFFERENT NL.<br />
NL 10% 30% 50% 70% 100%<br />
r-xbyz 0.9627 0.8722 0.8091 0.6856 0.5684<br />
R-EMD 0.9579 0.8555 0.7868 0.6682 0.5435<br />
From Fig. 6, we known that we can use EMD method<br />
after reconstruction <strong>of</strong> wavelet threshold de-noising.<br />
When noise level(NL
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1031<br />
Time-frequency complementary advantages scheme<br />
based EMD and Wavelet.<br />
Design <strong>of</strong> Time-frequency complementary advantages<br />
scheme based EMD and Wavelet.<br />
(1)Estimate noise level and SNR level for original<br />
mixed signals ,choose appropriate parameter and series in<br />
noise .<br />
(2)Do EMD de-noising for mixed signals to obtain all<br />
IMF components ,extract harmonic signal components<br />
first step ,observe correlation coefficient <strong>of</strong> signal<br />
recovered this moment. Using EMD could work in lower<br />
SNR and considering advantage <strong>of</strong> quicker speed<br />
convergence .<br />
(3)Do the second de-noising using wavelet transform<br />
perfect stability and high-precision, then reduce noise by<br />
heursure again.<br />
(4)Make the signal that is removed noise twice as<br />
harmonic signal extracted.<br />
V. CONCLUSIONS<br />
This paper first introduce the basic theory <strong>of</strong> timefrequency<br />
methods. Comparing wavelet method with<br />
empirical mode decomposition(EMD) method, according<br />
to the performance analysis <strong>of</strong> harmonic signal extraction<br />
from chaos background in different noise level and<br />
signal-to-noise ratio(SNR). We summarize a new<br />
synthesis about wavelet threshold and empirical mode<br />
decomposition(EMD) complementary <strong>of</strong> new harmonic<br />
signal extraction by experimental simulation. Computer<br />
simulation verified that the methods are high availability.<br />
Finally give the extraction procedure about<br />
complementary scheme.<br />
ACKNOWLEDGMENT<br />
This work is supported by the National Science<br />
Foundation <strong>of</strong> China (no.60672011), Open Fund <strong>of</strong> Key<br />
Laboratory <strong>of</strong> Electronics Engineering College <strong>of</strong><br />
Heilongjiang University (No. D22D20100027), the<br />
technology research project <strong>of</strong> Education Department <strong>of</strong><br />
Heilongjiang Province (No. 11511381) and Dr. Start<br />
funds <strong>of</strong> Heilongjiang University.<br />
REFERENCES<br />
[1] Richer M, Schreiber T and Kaplan D T, “Fetal EGG<br />
extraction with nonlinear state-space projection,” IEEE<br />
Trans. Biom. Eng. vol. 45(1), pp. 133~137, 1998.<br />
[2] Leung Henry and Huang Xing-ping, “Sinusoidal frequency<br />
estimation in chaotic noise”, ICASSP, vol. 2,<br />
pp.1344~1347, 1995.<br />
[3] Wang Fuping, Guo Jingbo, Wang Zanji and Xiao Dachuan ,<br />
“Harmonic Signal Extraction from Strong Chaotic<br />
interference,” Acta Physica Sinica. vol. 50(6), pp.<br />
1019~1023, 2001.<br />
[4] Haykin S and Li X B, “Detection <strong>of</strong> signal in chaos,”<br />
Proceeding <strong>of</strong> IEEE, vol. 83(1), pp. 94~122, 1995.<br />
[5] Short K M and Parker A T, “ Unmasking a hyperchaoti<br />
communication scheme,” Physical Review E, vol. 58, pp.<br />
1159~1162, 1998.<br />
© 2011 ACADEMY PUBLISHER<br />
[6] Donald B.Percival and Andrew T.Walden, “Wavelet<br />
Methods for Time Series Analysis,” Cambridge University<br />
Press,2000.<br />
[7] Huang N E,Shen Z and Long S R, “The empirical mode<br />
decomposition and the Hilbert spectrum for nonlinear and<br />
non-station time series analysis,”Proceeding <strong>of</strong> the Royal<br />
Society <strong>of</strong> London A, vol. 454, pp. 903~995, 1998.<br />
[8] Newland D E, “Wavelet analysis <strong>of</strong><br />
vibration,”part1,2.<strong>Journal</strong> <strong>of</strong> Vibrationand Acoustics, vol.<br />
116, pp. 409~425, 1994.<br />
[9] Sun yankui, “Wavelet analysis and its application,”China<br />
Machine Press. pp. 219~243, Oct. 2005.<br />
[10] Li Hong Guang and Meng Guang, “Harmoic signal<br />
extraction from chaotic interference based on empirical<br />
mode decomposition”, vol.53, July, 2004.<br />
[11] Wang guoguang and Wang shuxun, “Research on Methods<br />
<strong>of</strong> Extracting signals from Chaos,” Jilin University, 2007.<br />
[12] Zhang Defeng, “Matlab and Wavelet analysis,” China<br />
Machine Press, Jan. pp.65-92, 2010.<br />
[13] Ying Tan, Jun Wang and J. M. Zurada. Nonlinear Blind<br />
Source Separation Using a Radial Basis Function Network.<br />
IEEE Transactions on Neural <strong>Networks</strong>. vol. 12, pp.<br />
124~134, 2001.<br />
[14] S. A. chard, D. T. Pham and C. Jutten. Criteria Based on<br />
Mutual Information Minimization for Blind Source<br />
Separation in Post-nonlinear Mmixtures. Elsevier signal<br />
processing. vol. 85, pp.965~974, 2005.<br />
[15] W. L. Woo, S. S. Dlay. Nonlinear Blind Source Separation<br />
Using a Mixture RBF-FMLP Network. IEE Proceedings,<br />
Vision, Image and Signal Processing. vol. 152, pp.173~183,<br />
2005.<br />
[16] A. Ziehe, M. Kawanabe and S. Harmeling. Separation <strong>of</strong><br />
Post-nonlinear Mixtures Using ACE and Temporal<br />
Decorrelation. Proceeding <strong>of</strong> Independent Component<br />
Analysis and Blind Signal Separation (ICA’2001).<br />
pp.433~438, 2001.<br />
[17] Kun Zhang, Laiwan Chan. Extended Gaussianization<br />
Method for Blind Separation <strong>of</strong> Post-nonlinear Mixtures.<br />
Neural Computation. pp.425~452, 2005<br />
Erfu Wang Heilongjiang province,<br />
China, 1980. Received PhD degree and<br />
M.S. degree in Harbin Institute <strong>of</strong><br />
Technology, China, in 2009 and 2005<br />
respectively, and B.S. degree in Jilin<br />
University, China. The interest fields are<br />
blind source separation, array signal<br />
processing, wideband wireless<br />
communication,etc.<br />
Zhifang Wang Henan province, China,<br />
1979. Received PhD degree and M.S.<br />
degree in Harbin Institute <strong>of</strong> Technology,<br />
China, in 2009 and 2005 respectively, and<br />
B.S. degree in Henan University, China.<br />
The interest fields are biometric,<br />
development <strong>of</strong> media systems, image<br />
analysis, cluster analysis,etc.
1032 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
© 2011 ACADEMY PUBLISHER<br />
Jing Ma Heilongjiang province, China,<br />
1985. Received B.S. degree in<br />
Heilongjiang University, China, in 2008.<br />
The interest fields are chaotic theory,<br />
encryption system, signal separation ,etc.<br />
Qun Ding Heilongjiang province,<br />
China, 1957. Received PhD degree<br />
and M.S. degree in Harbin Institute <strong>of</strong><br />
Technology, China, in 2007 and 1997<br />
respectively. The interest fields are<br />
secure communication, information<br />
security, Pattern recognition, etc.<br />
Currently, she is a pr<strong>of</strong>essor <strong>of</strong><br />
Heilong Jiang University, China.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1033<br />
Provable Data Possession <strong>of</strong><br />
Resource-constrained Mobile Devices in Cloud<br />
Computing<br />
Jian Yang 1,2<br />
1College <strong>of</strong> Electronic and Information Engineering, Tongii University, Shanghai 201804, China<br />
2College <strong>of</strong> Mathematics and Computer Science, Dali University, Dali, Yunnan 671003, China<br />
Email: sbjc1215@126.com<br />
Haihang Wang 1 , Jian Wang 1,3 , Chengxiang Tan 1 and Dingguo Yu 1<br />
1College <strong>of</strong> Electronic and Information Engineering, Tongii University, Shanghai 201804, China<br />
3College <strong>of</strong> Electronics & Information Engineering, Henan University <strong>of</strong> Science & Technology, Luoyang,<br />
China<br />
Email: wanghh@sh163.net, wangjian_migi@sina.com, cxtan@trimps.ac.cn, zjydg@163.com<br />
Abstract—Benefited from cloud storage services, users can<br />
save their cost <strong>of</strong> buying expensive storage and application<br />
servers, as well as deploying and maintaining applications.<br />
Meanwhile they lost the physical control <strong>of</strong> their data. So<br />
effective methods are needed to verify the correctness <strong>of</strong> the<br />
data stored at cloud servers, which are the research issues<br />
the Provable Data Possession (PDP) faced. The most<br />
important features in PDP are: 1) supporting for public,<br />
unlimited numbers <strong>of</strong> times <strong>of</strong> verification; 2) supporting<br />
for dynamic data update; 3) efficiency <strong>of</strong> storage space and<br />
computing. In mobile cloud computing, mobile end-users<br />
also need the PDP service. However, the computing<br />
workloads and storage burden <strong>of</strong> client in existing PDP<br />
schemes are too heavy to be directly used by the<br />
resource-constrained mobile devices. To solve this problem,<br />
with the integration <strong>of</strong> the trusted computing technology,<br />
this paper proposes a novel public PDP scheme, in which the<br />
trusted third-party agent (TPA) takes over most <strong>of</strong> the<br />
calculations from the mobile end-users. By using bilinear<br />
signature and Merkle hash tree (MHT), the scheme<br />
aggregates the verification tokens <strong>of</strong> the data file into one<br />
small signature to reduce communication and storage<br />
burden. MHT is also helpful to support dynamic data<br />
update. In our framework, the mobile terminal devices only<br />
need to generate some secret keys and random numbers<br />
with the help <strong>of</strong> trusted platform model (TPM) chips, and<br />
the needed computing workload and storage space is fit for<br />
mobile devices. Our scheme realizes provable secure storage<br />
service for resource-constrained mobile devices in mobile<br />
cloud computing.<br />
Index Terms—bilinear signature, merkle hash tree, provable<br />
data possession, mobile computing, cloud computing,<br />
trusted computing<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1033-1040<br />
I. INTRODUCTION<br />
In cloud computing, multi-tenants share the external<br />
resources <strong>of</strong> computing and storage, which allows<br />
enterprises and individuals get on-demand computing or<br />
storage services from cloud service providers(CSP), such<br />
as Amazon’s S3 and Google’s App Engine, and no<br />
longer maintain their local physical machines. However,<br />
In this new model, the users put their data on the cloud<br />
storage servers maintained by service providers, which<br />
deprives the users <strong>of</strong> their control <strong>of</strong> the physical<br />
possession <strong>of</strong> data, even though they are the owners <strong>of</strong><br />
the data. In this case, some new security needs and<br />
problems have arisen. At the same time, when one’s data<br />
are outsourced, he wants to know whether the data is<br />
truly stored at the correct servers and be intact as stated in<br />
the Service Level Agreement (SLA). In addition, in<br />
multi-layer cloud services framework, higher layer cloud<br />
applications need lower layer services (such as storage<br />
service or virtual machine image service).In this<br />
circumstance, higher layer CSP needs effective methods<br />
to verify the basic storage services provided by lower<br />
CSP. It is the problem that the important direction <strong>of</strong> the<br />
current research field in cloud computing - provable data<br />
possession (PDP) – wants to solve. On the basis <strong>of</strong> PDP,<br />
if servers ensure the clients can retrieve correct data files<br />
with the help <strong>of</strong> erasure codes, such as Reed-Solomon<br />
codes, the services are named as Pro<strong>of</strong>s <strong>of</strong> Retrievability<br />
(POR).<br />
The first solutions to this issue are proposed by<br />
Deswarte et al. [1] and Filho et al.[2], which both use<br />
RSA-based functions to hash the whole data file for every<br />
verification challenge. Obviously, both <strong>of</strong> them are
1034 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
inefficient for big data files, which need more time to<br />
compute and transfer their hash values. Ateniese et al. [3]<br />
proposed a formal definition and related operations <strong>of</strong> the<br />
PDP model. They use some homomorphic tokens in the<br />
encoded file to help verify whether the file was tampered<br />
without any legal authentication. Juels et al. [4] formally<br />
put forward the protocol structure and security framework<br />
<strong>of</strong> the POR model and proposed a method to detect<br />
unauthorized changes <strong>of</strong> data by adding some “sentinels”<br />
randomly in the original files. But the above schemes<br />
only have limited times <strong>of</strong> verification operations and do<br />
not support public verification. On the basis <strong>of</strong> the<br />
security model in [4], Shacham et al. [5] designed an<br />
improved scheme to realize public data possession<br />
verification by using bilinear signature. But in their<br />
framework, the number <strong>of</strong> the authentication tokens<br />
stored on the server is proportional to the number <strong>of</strong> data<br />
blocks, and it only is fit for static file storage. Similar to<br />
[5], Wang et al. [6] uses Merkle hash tree (MHT) to build<br />
verification tags stored at the servers and support<br />
dynamic data update. They treat the leaf nodes <strong>of</strong> MHT<br />
as the left-to-right sequence so that the locations <strong>of</strong> error<br />
data can be detected.<br />
From another perspective, with the prevalence <strong>of</strong> the<br />
3G and 4G wireless communication networks, the mobile<br />
devices, such as mobile phones, PDA, also want to share<br />
the benefits introduced by the cloud on-demand storage<br />
and computing service. But the traditional mobile<br />
terminals are resource-constrained devices (with low<br />
CPU frequency and small memory) and can not use the<br />
existing PDP schemes directly, which require clients<br />
encode files with erasure codes, divide encoded files into<br />
blocks and sign on every data blocks. Those operations<br />
on large files are intolerable in our mobile computing<br />
clients. How to solve this problem is the main goal <strong>of</strong> our<br />
work.<br />
A. Our Contribution<br />
Combined with bilinear map and the well-studied<br />
authentication structure-MHT, this paper addresses to<br />
design a storage service model with public provable data<br />
possession in mobile computing environment. To achieve<br />
public verifiability in this environment, we need a trusted<br />
third-party auditor (TPA) to do most <strong>of</strong> the calculation<br />
works done by the clients in other PDP schemes.<br />
Specifically, our contribution in this paper can be<br />
summarized as the following two aspects:<br />
� We use trusted computing technology for the<br />
mutual authentication between the end-users and<br />
the TPA so that the most computations <strong>of</strong> the<br />
end-users can be done by TPA. The end users<br />
only need to generate some passwords and a<br />
small amount <strong>of</strong> random numbers, which can be<br />
done by the TPM chips embedded in their<br />
machines.<br />
� We improve the existing PDP schemes<br />
(especially the work proposed in [6]) with<br />
bilinear mapping signature to make them fit for<br />
the mobile computing environment. To the best<br />
<strong>of</strong> our knowledge, our scheme is the first storage<br />
framework in mobile cloud computing to support<br />
© 2011 ACADEMY PUBLISHER<br />
the stateless verification, public provable data<br />
possession, dynamic data update, and is provably<br />
secure in random oracle model.<br />
B. Paper Structure<br />
The rest <strong>of</strong> the paper is organized as follows. Section II<br />
introduces the related works. Then we provide the<br />
description <strong>of</strong> our scheme in Section III, including model<br />
structure, notation and preliminaries, important functions<br />
and detailed implement <strong>of</strong> our scheme. Section IV gives<br />
the security analysis and performance evaluation,<br />
followed by Section V which gives the concluding<br />
remark <strong>of</strong> the whole paper and overviews the related<br />
work finally.<br />
II. RELATED WORKS<br />
Ateniese et al. [3] proposed the first formal definition<br />
<strong>of</strong> the PDP model and related operations, functions. In<br />
their system some homomorphic verifiable tags are used<br />
to verify data file. Juels et al. [4] proposed a formal<br />
definition <strong>of</strong> POR and its security model. After being<br />
encrypted and divided into small data blocks, which are<br />
encoded with Reed-Solomon codes, the data file is added<br />
into some "sentinels” to detect whether it was intact.<br />
However, the both schemes don not support dynamic data<br />
update and can only verify limited times because that the<br />
two schemes only have finite number <strong>of</strong> the “sentinels” in<br />
a file. When the finite “sentinels” are exhausted, the file<br />
must be send back to the owner to re-compute new<br />
“sentinels”. In their improved work, Ateniese et al. [7]<br />
proposed a new scheme with homomorphic linear<br />
authenticators (HLA), <strong>of</strong> which communication<br />
complexity is independent <strong>of</strong> the file length. Though the<br />
scheme supports infinite times <strong>of</strong> verification, it can not<br />
verify publicly.<br />
Chang et al. [8] proposed a remote identity check<br />
scheme. By using redactable signature (a kind <strong>of</strong><br />
homomorphic signature scheme proposed in [9]), a<br />
redactor can calculate a effective signature on a redacted<br />
message x’ without knowing the private key. This idea<br />
can be used in a third-party verification scheme and<br />
verifier does not need to know the private key. Shacham<br />
et al. [5] proposed two POR schemes: the first uses<br />
bilinear signature, and is provable secure and efficient in<br />
the random oracle model. The second depends on<br />
pseudo-random function and bilinear map, and is<br />
provable secure in the standard model. Both schemes rely<br />
on the homomorphic property--aggregating the<br />
verification pro<strong>of</strong>s into a small value. However, the<br />
above two methods are not aware <strong>of</strong> the user’s privacy<br />
preserving in public audit.<br />
An effective public PDP scheme should have the two<br />
important following properties [10]:<br />
� To allow a TPA to verify the correctness and<br />
integrity <strong>of</strong> the data without retrieving a copy <strong>of</strong><br />
the whole data or introducing additional on-line<br />
burden to the clients.<br />
� To avoid introducing new vulnerabilities to the<br />
privacy <strong>of</strong> the data.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1035<br />
Wang et al. [6, 10] discussed the privacy protection in<br />
public audit. In their framework, the third-party audit<br />
protocol with the privacy preserving is independent <strong>of</strong><br />
data encryption. By using homomorphic authenticator<br />
and random masking, the scheme conceals the content <strong>of</strong><br />
the original data from TPA and TPA can perform<br />
multiple auditing tasks in a batch manner.<br />
Erway et al. [11] is the first to support dynamic data<br />
update by using rank-based verification skip list in cloud<br />
servers. The most similar PDP scheme to ours was<br />
proposed by Wang et al. [6], which is also our work’s<br />
basis. Their scheme supports public data possession<br />
verification and dynamic data update at the same time. It<br />
improves the previous PDP models by using the classical<br />
MHT and enables TPA to complete privacy-preserved<br />
data integrity check with the support <strong>of</strong> dynamic data<br />
update. This method does not require user's real-time<br />
participation in the cloud, and avoids the leakage <strong>of</strong><br />
user’s privacy. However, it still requires the user to<br />
calculate initial verification tokens <strong>of</strong> the files, which is<br />
not suitable for this paper’s application environment.<br />
If the TPA takes over a large number <strong>of</strong> computing<br />
works <strong>of</strong> the mobile end-users, first <strong>of</strong> all, the TPA<br />
should achieve mutual authentication with the end-users,<br />
and establish a secure transmission channel on this basis.<br />
A feasible method is to take the advantage <strong>of</strong> the trusted<br />
computing technology, which is a mature technology, but<br />
has few successful applications applied in the cloud<br />
computing and few concrete frameworks designed. EMC<br />
China lab collaborated with Fudan University, Huazhong<br />
University <strong>of</strong> Science and Technology, Tsinghua<br />
University and Wuhan University to carry out a research<br />
projects on trusted virtual infrastructure, named Daoli<br />
[12]. The research project is committed to tenants’<br />
isolation and protection platform provider away from<br />
attacks by malicious tenants in multi-tenant cloud<br />
computing environments. Combination trusted computing<br />
with virtualization technologies to enhance the security <strong>of</strong><br />
computing platforms, makes cloud service providers<br />
being able to provide virtual private cloud (VPC) services<br />
in public cloud. In [13], the authors discussed how to use<br />
TCB for security enhancements in Xen—a famous<br />
open-sourced virtual machine monitor can be used in<br />
cloud computing—and described how this method is used<br />
to achieve "trusted virtualization" and enhance the<br />
security <strong>of</strong> the virtual TPM. They moved the new VM<br />
creation function to a small trusted VM out <strong>of</strong> Dom0.<br />
This method has two main goals: the one is to reduce and<br />
bound the size <strong>of</strong> TCB in Xen-based systems, especially<br />
remove the user space in Dom0 from the TCB, to<br />
enhance security. Another one is that if we suppose TCB<br />
were security, the new VM would have maintained the<br />
same attributes <strong>of</strong> the security and integrity with physical<br />
machine. The privacy manager discussed in [14] uses<br />
TPM to manage privacy keys required in privacy<br />
protection. Trusted computing technology to the IaaS<br />
cloud computing systems was introduced in [15]. In<br />
Eucalyptus, for example, the authors used a Trusted<br />
Coordinator (TC) (maintained by an external trusted<br />
entity) to combine unreliable cloud Manager (CM) with a<br />
© 2011 ACADEMY PUBLISHER<br />
number <strong>of</strong> trusted nodes in order to form its main<br />
architecture. This framework ensures the safety <strong>of</strong> the<br />
customers VM, allows users to verify the IaaS service<br />
providers and determine whether the services are security<br />
before they start VM.<br />
However, about the combination with trusted<br />
computing and cloud computing, the schemes mentioned<br />
above did not enhance the security and integrity <strong>of</strong> the<br />
cloud storage services with the provable data possession,<br />
which is a critical measure to enhance the user's<br />
confidence in using the cloud storage services, especially<br />
in wireless mobile computing environment. To the best <strong>of</strong><br />
our knowledge, our scheme is the first to explore the<br />
application <strong>of</strong> PDP scheme in mobile computing<br />
environment combined with the trusted computing.<br />
III. OUR SCHEME<br />
A. Model Structure<br />
System participants: as shown in Figure 1, in our<br />
resource-constrained public provable data possession<br />
scheme <strong>of</strong> the cloud storage service, there are three main<br />
participants: 1)client, i.e. mobile end-user, which has a<br />
TPM chip and stores data files in the cloud, and expects<br />
to get trusted storage validation; 2) trusted third party<br />
auditor (TPA), which is credible for clients and take the<br />
main file encryption and authentication tasks required<br />
during the process <strong>of</strong> the scheme; 3) cloud storage service<br />
provider (CSP), which has a large capacity <strong>of</strong> storage and<br />
provides the users with storage services and the pro<strong>of</strong> <strong>of</strong><br />
data possession when needed.<br />
In the mobile computing environment, the client’s<br />
computing and storage capacity is very limited, but it has<br />
the ability to use its TPM chips to produce and store<br />
secret keys. The TPA is acted by a service agent located<br />
between the mobile access point and the gateway <strong>of</strong> the<br />
IP network services. The TPA should have high<br />
performance <strong>of</strong> computing but limited storage space,<br />
which is only for small part <strong>of</strong> the information <strong>of</strong> the<br />
clients and the current session message. In addition, it<br />
should connect securely with the client for providing<br />
services. Through Internet, the CSP provides<br />
high-capacity, redundant storage services. Generally<br />
Fig.1. System structure model<br />
speaking, the CSP is an unsafe, even malicious entity.<br />
That is to say, for some financial benefits, it is possible to
1036 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
read, tamper or delete the user's original files, even to<br />
forge the pro<strong>of</strong> <strong>of</strong> the data possession. So the files should<br />
be encrypted before sending to the CSP. In this paper, we<br />
assume all parties communicate through secure, reliable,<br />
authenticated channels for all phases。<br />
Security goal: In this system, the client and the TPA<br />
need to establish a secure link. On the basis <strong>of</strong> security <strong>of</strong><br />
the underlying communication link between the mobile<br />
terminal and the TPA, firstly, through the Diffie-Hellman<br />
key exchange protocol, the scheme generates a symmetric<br />
key for data exchange. Except for some necessary keys<br />
and random values, the client does not perform additional<br />
computation works, which are taken over by the TPA.<br />
The scheme is secure under the random oracle model and<br />
our security model is based on the one proposed in [6].<br />
Design goal: The scheme has three design goals<br />
shown as followed.<br />
� public provable data possession: to allow a<br />
verifier, not just the data owner or users, to have<br />
the capability to verify the correctness <strong>of</strong> the<br />
stored data on demand;<br />
� stateless verification: to generate the pro<strong>of</strong>s <strong>of</strong><br />
data possession according to the challenge<br />
produced randomly by the verifier, not to the<br />
persistent information maintained by some<br />
entities;<br />
� resource-constrained mobile environment support:<br />
to allow end-user just to generate and store the<br />
keys by utilizing TPM chip instead <strong>of</strong> doing lots<br />
<strong>of</strong> computing works, which are done by the TPA<br />
now;<br />
� trusted computing technology applying: to build a<br />
trusted channel between the TPA and the client,<br />
then use it to transfer the related messages and<br />
authorize the TPA to complete those works for<br />
reduce the client’s workload.<br />
B. Notation and Preliminaries<br />
Diffie-Hellman protocol. Diffie-Hellman key<br />
exchange protocol is one <strong>of</strong> the most famous schemes in<br />
cryptography, which mainly utilize the discrete log<br />
problem to safely exchange a shared symmetric key<br />
through an unsecured channel. By using this method, the<br />
end-user and the TPA can share a symmetric key so that<br />
the data files, asymmetric keys and the information <strong>of</strong> the<br />
verification can be transferred safely after being<br />
encrypted by this key.<br />
Merkle hash tree. A Merkle Hash Tree (MHT) is a<br />
well-studied authentication structure [16], which can<br />
efficiently and securely prove that a set <strong>of</strong> data blocks are<br />
undamaged and unaltered. It is constructed as a binary<br />
tree where the leaves in the Merkle hash tree are the hash<br />
values <strong>of</strong> authentic data. During the process <strong>of</strong> the<br />
verification, the verifier only needs check whether the<br />
root value <strong>of</strong> the tree is tampered. In the archive [6], their<br />
scheme treats the leaf nodes as the left-to-right sequence<br />
for getting the positions <strong>of</strong> error data blocks. For the sake<br />
<strong>of</strong> simplicity, in our scheme, MHT tree leaf nodes only<br />
store the hash <strong>of</strong> data blocks.<br />
© 2011 ACADEMY PUBLISHER<br />
Bilinear map. A bilinear map is a map e:G1×G1→G2,<br />
where G1 is a Gap Diffie-Hellman group and G2 is a<br />
multiplicative cyclic group with a big prime order p. It<br />
has the following properties:<br />
� Computable: there exists an efficiently<br />
computable algorithm for computing the map;<br />
� Bilinear: for all h1,h2∈G1 and a,b∈Zp, �<br />
a b<br />
ab<br />
e ( h1<br />
, h2<br />
) � e(<br />
h1,<br />
h2<br />
)<br />
;<br />
Non-degenerate:<br />
e(<br />
g,<br />
g)<br />
� 1<br />
, where g is a<br />
generator <strong>of</strong> G1.<br />
C. Some Important Procedures:<br />
Improved from [4, 6, 17], our secure storage system<br />
with provable data possession contains the following<br />
functions:<br />
KenGen(1 k ) → (pk,sk) : This algorithm takes as input<br />
initial secure parameter 1 k , and returns the public key pk<br />
and the private key sk. In our scheme, the end-user<br />
generates and maintains the keys in TPM chip.<br />
Encapek(F) → F’: The TPA uses this algorithm to<br />
encrypt the raw file F with the seal key ek and encode it<br />
with erasure codes. It returns the sealed file F’.<br />
SigGen_Clientsk (F’) → Sigsk (H(R)): This algorithm is<br />
run by end-users. It takes as input the hash value <strong>of</strong> the<br />
root <strong>of</strong> the MHT and outputs its signature as a metadata.<br />
SigGen_TPA(F’) → Ф: This algorithm is run by TPA.<br />
It takes as input each data blocks {mi} <strong>of</strong> the sealed file<br />
F’, and outputs the signature collection Ф= { σi } on<br />
{mi}.<br />
GenPro<strong>of</strong>(chal, F’, Sigsk(H(R)), Ф) → (P): This<br />
algorithm is run by the storage server. It takes as input the<br />
verification challenge message “chal” generated by TPA,<br />
the stored file F’, the metadata signature and the signature<br />
set Ф, and returns the possession pro<strong>of</strong> P.<br />
Verify(P, chal) → {TRUE|FALSE}: By run this<br />
algorithm, according to the random challenge chal, the<br />
pro<strong>of</strong> P returned from the server and some metadata <strong>of</strong><br />
the end-user, the TPA verify the correctness <strong>of</strong> the data<br />
file, and outputs TRUE if the integrity <strong>of</strong> the file is<br />
verified as correct, or FALSE otherwise.<br />
Decapdk(F’) →F: The end-user request to extract a file<br />
F, and the TPA retrievals the corresponding sealed file F’<br />
from the cloud, decodes and decrypts the file to get F<br />
with the decryption key (dk), then sends F to the<br />
end-user.<br />
D. Detailed Implementation<br />
For implementing the public provable data possession<br />
<strong>of</strong> the cloud storage service in our background, first <strong>of</strong> all,<br />
a trusted communication channel should be build between<br />
the TPA and the mobile end-user, which needs mutual<br />
remote identification. This issue will be discussed more<br />
detail later in this paper. After mutual identification, by<br />
utilizing the Diffie-Hellman key exchange protocol, the<br />
TPA and the end-user negotiate a symmetric key for the<br />
exchange <strong>of</strong> other information in this scheme.<br />
Now we start to present the main idea <strong>of</strong> our scheme.<br />
According to the model defined in [4, 6, 10, 17], we<br />
assume the raw data file F is first encrypted by the key
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1037<br />
(ek) and encoded into F’ using erasure codes, then the F’<br />
is divided into N blocks {mi}, where mi∈Zp and 1≤i≤N.<br />
Let e: G1×G1→G2 be a bilinear map, which has a big<br />
prime order p and a generator g <strong>of</strong> group G1. Let H:<br />
{0,1}*→ G1 is a hash function. The main procedure <strong>of</strong><br />
our scheme is as follows:<br />
Setup: In this phase, we assume the end-user has<br />
already completed the remote identification with the TPA<br />
by using TPM and set up a trusted communication<br />
channel. Through this channel, the phase setup is<br />
executed as Algorithm 1 described:<br />
Algorithm 1 Setup:<br />
1.C→T:g, g α<br />
2.T→C:g β<br />
{ F}<br />
3.C→T: ��<br />
g<br />
{ H ( R)<br />
} , { dk } ��<br />
4.T→C: ��<br />
g<br />
5.C→T:Sigsk (H(R)) = (H(R)) α<br />
4.T→S : Sigsk (H(R)), F’={mi}, Ф={σi}, where 1≤i≤N, σi<br />
=[H(mi)u mi ] β<br />
C represents for the Client, T for the trusted third-party<br />
agent and S for the cloud storage server<br />
The first two steps <strong>of</strong> algorithm 1 are to complete<br />
Diffie-Hellman key exchange. After this, the client and<br />
��<br />
the TPA share a symmetric key g . Then the client<br />
encrypts the original file F with this key and sends it to<br />
the TPA.<br />
Received the original file F from the client, a pair <strong>of</strong><br />
asymmetric keys (ek, dk) are generated by invoking<br />
KenGen(*), where (ek) is for encrypting the file and (dk)<br />
for decrypting after retrieval. By calling Encapek(F), the<br />
TPA encrypts the file, divides it into small blocks and<br />
encodes the data blocks with erasure codes. After doing<br />
these, the TPA gets N blocks {mi} (1≤i≤N) as the stored<br />
file F’. Then, the TPA calculates H(R), the hash value <strong>of</strong><br />
the root <strong>of</strong> the MHT, <strong>of</strong> which the leaves are the hash <strong>of</strong><br />
the corresponding mi. At last <strong>of</strong> this turn, the TPA sent<br />
the value H(R) and dk to the client encrypted with the<br />
��<br />
shared key g .<br />
By running SigGen_Client sk (F’), the client signs H(R)<br />
and sent the signature <strong>of</strong> H(R) back to the TPA. The TPA<br />
calls SigGen_TPA(F’) to calculate the signature<br />
collections <strong>of</strong> each blocks <strong>of</strong> F’, then sends {Sigsk (H(R)),<br />
F’, Ф} to the cloud storage servers. In algorithm 1, u is a<br />
element in G1 chosen randomly by the TPA.<br />
Integrity verification: This phase starts from the client<br />
or the TPA by sending a verification challenge to the<br />
cloud storage service provider (CSP). According to the<br />
challenge, the CSP computes the pro<strong>of</strong> <strong>of</strong> verification and<br />
sends it back to the TPA. After verifying the pro<strong>of</strong>, the<br />
TPA sends the result to the client. The detail process <strong>of</strong><br />
verification is shown as the algorithm 2:<br />
© 2011 ACADEMY PUBLISHER<br />
g<br />
Algorithm 2 Verify:<br />
1.T→S:chal={(i, vi)}, where 1≤i≤c<br />
2.S→T:pro<strong>of</strong>={μ,ω,(H(mi),Ωi), Sigsk (H(R))} , where 1≤i≤c<br />
3.T→C:Verify(μ,ω, (H(mi),Ωi) ,Sigsk (H(R)) ,chal)={true , false}<br />
The challenge message “chal” is generated by TPA.<br />
The TPA chooses c random numbers in the set [1, N] to<br />
constitute a sequence subset I. For each i∈I, the TPA<br />
picks a random element vi∈Zp. The challenge message<br />
“chal” sent to CSP is composed <strong>of</strong> the number i and the<br />
corresponding vi. On receiving the “chal”, the CSP runs<br />
GenPro<strong>of</strong>(chal, F’, Sigsk (H(R)), Ф) to generate the<br />
pro<strong>of</strong>, which includes the corresponding hash value H(mi)<br />
<strong>of</strong> the data blocks {mi} for every i∈I and the additional<br />
information Ωi for rebuilding the root H(R) <strong>of</strong> the MHT.<br />
In addition, the CSP also computes the following two<br />
values as a part <strong>of</strong> the pro<strong>of</strong>:<br />
�<br />
c<br />
� �<br />
i�1<br />
c<br />
�<br />
i�1<br />
v m � Z<br />
i<br />
i<br />
p<br />
and<br />
vi<br />
� � � i �G1<br />
.<br />
After receiving the pro<strong>of</strong>, the TPA runs Verify(*), and<br />
sends the result back to the client. The main goal <strong>of</strong> the<br />
function Verify(*) is to test whether the following two<br />
equations is correct:<br />
�<br />
e(Sig sk (H(R)), g) ? e(H(R), g )<br />
……… (1)<br />
c<br />
�<br />
vi<br />
� ��<br />
e(<br />
�, g ) ? e(<br />
�( H ( mi<br />
) u , g )<br />
i�1<br />
……. (2)<br />
If so, the Verify(*) returns TRUE; otherwise FALSE.<br />
File retrieval: Before retrieval, the client and the TPA<br />
should negotiate a symmetric session key Ks through<br />
Diffie-Hellman key exchange protocol. Then, shown as<br />
algorithm 3, the client sends the decryption key (dk)<br />
encrypted by Ks to the TPA and the TPA request the CSP<br />
for extracting the file F’. The CSP sends F’ to the TPA.<br />
Then the TPA runs Decapdk(F’) to get the raw file F and<br />
send F to the end-user through a secure communication<br />
channel.<br />
Algorithm 3 Retrieval:<br />
1.C→T:Request(F), {dk}Ks<br />
2.T→S:Request(F’)<br />
3.S→T:F’<br />
4. T→C:F=Decapdk(F’)<br />
Noted that, we do not include the process <strong>of</strong><br />
negotiation <strong>of</strong> the symmetric session key Ks in the above<br />
description <strong>of</strong> algorithm 3.<br />
E. Discussions<br />
Remote identification: As stated above, the client and<br />
the TPA use a symmetric key generated through the<br />
Diffie-Hellman protocol to encrypt the original file, the<br />
secret information, the hash value <strong>of</strong> the root <strong>of</strong> the
1038 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Merkle hash tree and some control information<br />
transferred between them. But before that, both <strong>of</strong> them<br />
need remote mutual identification. On one hand, based on<br />
trusted computing idea, we assume that the TPA is<br />
trusted in mobile computing environment. That is to say,<br />
the client does not need to identify the TPA. On the other<br />
hand, the client is attested by answering the challenge <strong>of</strong><br />
TPA with a message signed by the Attestation Identity<br />
Key (AIK), which is created and maintained by the TPM<br />
chip embedded in the client’s device. There are lots <strong>of</strong><br />
schemes to solve this problem. Because this paper<br />
focuses on the mobile computing, we can adopt the<br />
scheme proposed by Yang et al. [18] to realize the remote<br />
identification in wireless environment.<br />
Verification equation: During the verification phase, if<br />
the information stored at the cloud has not been tampered,<br />
replaced or deleted, according to the properties <strong>of</strong> bilinear<br />
map, the correctness <strong>of</strong> Eq. (1) can be elaborated as<br />
follows:<br />
e(Sig sk (H(R)), g)<br />
�<br />
e((H(R))<br />
�<br />
� e(H(R), g<br />
�<br />
, g)<br />
And the one <strong>of</strong> equation (2) can be proved as follows:<br />
c<br />
�<br />
i�1<br />
c<br />
�<br />
i�1<br />
c<br />
�<br />
i�1<br />
�<br />
v<br />
e(<br />
� , g ) � e(<br />
� , g<br />
� e(<br />
� e(<br />
� e(<br />
� e(<br />
c<br />
�<br />
i�1<br />
c<br />
�<br />
i�1<br />
i i<br />
�<br />
)<br />
m � v<br />
i<br />
i �<br />
�( H ( m ) u ) � , g )<br />
i<br />
v �<br />
i mivi<br />
�<br />
�H ( m ) � u � , g )<br />
( H ( m )<br />
( H ( m )<br />
i<br />
i<br />
i<br />
v<br />
v<br />
i<br />
i<br />
c<br />
)<br />
�<br />
i�1<br />
) �u<br />
m v<br />
�<br />
) u , g<br />
i i<br />
��<br />
, g<br />
Data confidentiality: according the pro<strong>of</strong> in [6], that<br />
the equation 1 and 2 are satisfied can ensure the data<br />
stored at the cloud storage servers are correct and intact<br />
under the random oracle model. In additional, for the<br />
privacy and the confidentiality, the TPA encrypts the raw<br />
file with the key (ek) and decrypt with the key (dk) when<br />
it sends and retrieves the file respectively. The<br />
asymmetric key pair (ek, dk) is created by the TPA.<br />
Because the TPA maybe need provider the verification<br />
service for multi-user in a real circumstance, according to<br />
the TPM main specification v1.2 [19], in our scheme<br />
there are two ways to maintain the key pair. One way is<br />
to use TPM_Bind to bind the key pair and store it on the<br />
TPA. Another solution is, as shown in algorithm 1, that<br />
the decryption key (dk), which encrypted by the<br />
symmetric key shared between the client and the TPA,<br />
can be sent to the client and sent back to the TPA during<br />
extracting the file.<br />
Supporting dynamic data update: the scheme in [6]<br />
has already supported dynamic data update with the<br />
operation on the Merkle hash tree, so our scheme also has<br />
© 2011 ACADEMY PUBLISHER<br />
)<br />
��<br />
)<br />
this property. Because in our mobile computing<br />
environment, a lot <strong>of</strong> calculations on mobile device are<br />
not meaningful, we do not intend to discuss this issue in<br />
detail.<br />
IV. SECURITY AND PERFORMANCE ANALYSIS<br />
A. Security Analysis:<br />
As mentioned above, an important basis <strong>of</strong> our scheme<br />
is building a secure information transferring channel<br />
between the client and the TPA. The TPA, which we<br />
assume is trusted, can authenticate the client based on<br />
hardware TPM chip embedded in the device <strong>of</strong> the client.<br />
By using Diffie-Hellman key exchange, the two entities<br />
can share a symmetric key to ensure the data files and<br />
signatures transferred between them are more secure.<br />
Furthermore, the trusted computing technology can also<br />
help them to avoid Man-in-the-middle attack during the<br />
process <strong>of</strong> the Diffie-Hellman key exchange.<br />
The file transferred between the TPA and the CSP is<br />
encrypted and encoded file F’, which avoid the CPS<br />
knowing the content <strong>of</strong> file and ensure the privacy and<br />
confidentiality <strong>of</strong> the raw file. The signature structure <strong>of</strong><br />
MHT is the basis <strong>of</strong> keeping the integrity <strong>of</strong> data. That<br />
the root signature <strong>of</strong> the MHT is computed by the<br />
end-user can avoid the leakage <strong>of</strong> the user’s private key.<br />
As for the security analysis <strong>of</strong> the bilinear map in PDP<br />
schemes, archives [5-6, 10] have exhaustive description,<br />
so we will not discuss it in detail.<br />
B. Performance Analysis:<br />
According algorithm 1 and 2, we can demonstrate the<br />
overall workload <strong>of</strong> the computing and storage <strong>of</strong> each<br />
parties in our scheme as followed:<br />
� Mobile terminal: stores the private key α and<br />
decryption key dk <strong>of</strong> file; computes the public<br />
�<br />
key g and the signature Sigsk(H(R)) <strong>of</strong> the<br />
MHT root H(R).<br />
�<br />
� TPA: stores the user’s public key g ;<br />
encodes/decodes, encrypts/decrypts the file,<br />
computes the data blocks signature collection Ф,<br />
and verifies the two equations during verification.<br />
� CSP: stores the signature Sigsk(H(R)), the<br />
encoded F’ and the Ф; generates the verification<br />
information μ and ω, and computes the Ωi for<br />
recovering the MHT.<br />
It should be noted that we do not include the<br />
workloads for generating the needed random numbers<br />
and nonce <strong>of</strong> every entity in our scheme.<br />
As described above, through a trusted channel between<br />
the end-user and the TPA, the heavy works, such as<br />
encoding (decoding), encryption (decryption) <strong>of</strong> the file,<br />
can be moved to the TPA to be done, and the end-user<br />
only need to generate keys and sign the root <strong>of</strong> the MHT.<br />
So our scheme realized the goal: the storage space and<br />
the computing ability needed for end-user during<br />
verifying the data possession are as small as possible,<br />
which is fit for the mobile computing device, such as<br />
mobile phone and PDA, etc.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1039<br />
By using the MHT idea, the data blocks signatures are<br />
aggregated into the signature <strong>of</strong> the root <strong>of</strong> the MHT to<br />
verify the data integrity with a minimum <strong>of</strong> storage space.<br />
The root signature is put into the cloud server, which<br />
realized the stateless verification. The MHT can also help<br />
to implement the dynamic data update as described in [6].<br />
V. CONCLUSIONS<br />
When the resource-constrained mobile devices use the<br />
cloud storage services deployed on traditional IP<br />
networks, the end-users are most concerned about<br />
whether the CSP stores their files correctly and dutifully.<br />
On the basis <strong>of</strong> the existing researches on PDP and POR,<br />
a secure storage scheme with public provable data<br />
possession <strong>of</strong> the mobile devices in cloud computing is<br />
proposed in this paper. Through the remote<br />
authentification <strong>of</strong> the mobile end-user by using trusted<br />
computing technology, a trusted third-party agent can<br />
undertake most computing workload <strong>of</strong> the client in<br />
traditional PDP, which makes our scheme be fit for the<br />
mobile computing environment. Combined with MHT<br />
and bilinear signature, improved the framework and<br />
related algorithms from the existing PDP schemes, our<br />
scheme realized public PDP with the support <strong>of</strong> dynamic<br />
data update. The scheme is provable secure under the<br />
random oracle model as proved in [6]. To the best <strong>of</strong> our<br />
knowledge, our scheme is the first to explore the<br />
application <strong>of</strong> PDP scheme in mobile computing<br />
environment combined with the trusted computing.<br />
Our future works include building a prototype system<br />
to test the performance <strong>of</strong> our scheme and exploring the<br />
application <strong>of</strong> other PDP framework applied in secure<br />
storage services <strong>of</strong> mobile computing environment.<br />
ACKNOWLEDGEMENT<br />
The authors wish to thank those reviewers and editors<br />
<strong>of</strong> this paper. This work is supported by the Science and<br />
Technology Support Program <strong>of</strong> Science and Technology<br />
Commission <strong>of</strong> Shanghai Municipality (No. 072712036),<br />
the National Natural Science Foundation <strong>of</strong> China (No.<br />
60803096), the Fundamental Research Funds for the<br />
Central Universities and Dalian IT teacher’s project.<br />
REFERENCES<br />
[1] Deswarte, Y., J.-J. Quisquater, and A. Saidane. “Remote<br />
integrity checking”. In Proc. <strong>of</strong> Conference on Integrity and<br />
Internal Control in Information Systems. 2003.<br />
[2] Filho, D.L.G. and P.S.L.M. Baretto. “Demonstrating data<br />
possession and uncheatable data transfer”, In IACR ePrint<br />
archive. 2006.<br />
[3] Ateniese, G., et al., “Provable data possession at untrusted<br />
stores”, in Proceedings <strong>of</strong> the 14th ACM conference on<br />
Computer and communications security. 2007, ACM:<br />
Alexandria, Virginia, USA. p. 598-609.<br />
[4] Juels, A. and J. Burton S. Kaliski, “Pors: pro<strong>of</strong>s <strong>of</strong><br />
retrievability for large files”, in Proceedings <strong>of</strong> the 14th ACM<br />
conference on Computer and communications security. 2007,<br />
ACM: Alexandria, Virginia, USA. p. 584-597.<br />
© 2011 ACADEMY PUBLISHER<br />
[5] Shacham, H. and B. Waters, “Compact Pro<strong>of</strong>s <strong>of</strong><br />
Retrievability”, in Proceedings <strong>of</strong> the 14th International<br />
Conference on the Theory and Application <strong>of</strong> Cryptology and<br />
Information Security: Advances in Cryptology. 2008,<br />
Springer-Verlag: Melbourne, Australia. p. 90-107.<br />
[6] Wang, Q., et al., “Enabling Public Verifiability and Data<br />
Dynamics for Storage Security in Cloud Computing”, in<br />
Computer Security – ESORICS 2009, M. Backes and P. Ning,<br />
Editors. 2009, Springer Berlin / Heidelberg. p. 355-370.<br />
[7] Ateniese, G., S. Kamara, and J. Katz, “Pro<strong>of</strong>s <strong>of</strong> Storage<br />
from Homomorphic Identification Protocols”, in Advances in<br />
Cryptology – ASIACRYPT 2009, M. Matsui, Editor. 2009,<br />
Springer Berlin / Heidelberg. p. 319-333.<br />
[8] Chang, E.-C., and J. Xu, “Remote Integrity Check with<br />
Dishonest Storage Server”, in Proceedings <strong>of</strong> the 13th<br />
European Symposium on Research in Computer Security:<br />
Computer Security. 2008, Springer-Verlag: Malaga, Spain. p.<br />
223-237.<br />
[9] Johnson, R., et al., “Homomorphic Signature Schemes”, in<br />
Proceedings <strong>of</strong> the Cryptographer's Track at the RSA<br />
Conference on Topics in Cryptology. 2002, Springer-Verlag. p.<br />
244-262.<br />
[10] Wang, C., et al., “Privacy-preserving public auditing for<br />
data storage security in cloud computing”, in Proceedings <strong>of</strong> the<br />
29th conference on Information communications. 2010, IEEE<br />
Press: San Diego, California, USA. p. 525-533.<br />
[11] C. Chris Erway, A.K., Charalampos Papamanthou,<br />
Roberto Tamassia, “Dynamic provable data possession”, in<br />
Proceedings <strong>of</strong> the 16th ACM conference on Computer and<br />
communications security. 2009, ACM: Chicago, Illinois, USA.<br />
p. 213-222.<br />
[12] Mao, W.B. “Talking About the Cloud Computing”. 2009<br />
2009-03-03; Available from:<br />
http://blog.csdn.net/wenbomao/archive/2009/03/03/3952761.as<br />
px and http://www.daoliproject. org.<br />
[13] Murray, D.G., G. Milos, and S. Hand, “Improving Xen<br />
security through disaggregation”, in Proceedings <strong>of</strong> the fourth<br />
ACM SIGPLAN/SIGOPS international conference on Virtual<br />
execution environments. 2008, ACM: Seattle, WA, USA. p.<br />
151-160.<br />
[14] Pearson, S., Y. Shen, and M. Mowbray, “A Privacy<br />
Manager for Cloud Computing”, in Proceedings <strong>of</strong> the 1st<br />
International Conference on Cloud Computing. 2009,<br />
Springer-Verlag: Beijing, China. p. 90-106.<br />
[15] Santos, N., K.P. Gummadi, and R. Rodrigues, “Towards<br />
trusted cloud computing”, in Proceedings <strong>of</strong> the 2009<br />
conference on Hot topics in cloud computing. 2009, USENIX<br />
Association: San Diego, California. p. 3-3.<br />
[16] Cao, T.J., Y.P. Zhang, and C.J. Wang, “Secure Protocols”.<br />
2009, Beijing, China: BUPT Press. 2.<br />
[17] Bowers, K.D., A. Juels, and A. Oprea, “Pro<strong>of</strong>s <strong>of</strong><br />
retrievability: theory and implementation”, in Proceedings <strong>of</strong><br />
the 2009 ACM workshop on Cloud computing security. 2009,<br />
ACM: Chicago, Illinois, USA. p. 43-54.<br />
[18] Li, Y., M. Jian-Feng, and Z. Jian-Ming, “Trusted and<br />
anonymous authentication scheme for wireless networks”.<br />
<strong>Journal</strong> <strong>of</strong> China Institute <strong>of</strong> Communications, 2009. 30(9): p.<br />
29-35.<br />
[19] Group, T.C., “TPM Main Specification Level 2 Version<br />
1.2”, Revision 103. 2007, TCG.
1040 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Jian Yang is a lecturer <strong>of</strong> Dali<br />
University in China. He was<br />
responsible for teaching database<br />
technology and E-commerce in<br />
Faculty <strong>of</strong> Computer. He was born on<br />
Dec in 1976. He received his master<br />
degree <strong>of</strong> Engineering in Kunming<br />
University <strong>of</strong> Science on May in 2005.<br />
Now he is a PhD student in Tongji<br />
University. He has published a dozen <strong>of</strong> papers on<br />
computer magazines. His major research interests include<br />
network security, trusted computing, cloud computing.<br />
Mr. Yang is a student member <strong>of</strong> China Computer<br />
Federation.<br />
Haihang Wang is a pr<strong>of</strong>essor and a<br />
supervisor <strong>of</strong> Ph.D. student in Tongji<br />
University at Shanghai. He was born on<br />
Mar. in 1965 and received his Ph.D.<br />
degree in Zhejiang University in 1994.<br />
Now He is engaged in teaching<br />
E-commerce and Project Management <strong>of</strong><br />
Information System. His research interests<br />
include intelligent information system,<br />
network security, E-commerce and supply chain management,<br />
enterprise information technology, mechatronics and<br />
automation, production and operations management. He has<br />
already published 60 papers on international magazines and<br />
conferences.<br />
© 2011 ACADEMY PUBLISHER<br />
Jian Wang is a lecturer <strong>of</strong> Henan<br />
University <strong>of</strong> Science and Technology in<br />
China. She earned her master degree in<br />
Huazhong University <strong>of</strong> Science &<br />
Technology in 2005. Now she is a Ph.D.<br />
student in Tongji University. Her research<br />
interests include trusted computing and<br />
network security.<br />
Chengxiang Tan was born in 1965 and<br />
earned his Ph.D. in North-west<br />
Polytechnic University in 1994. He is a<br />
pr<strong>of</strong>essor and a supervisor <strong>of</strong> Ph.D student<br />
in Tongji University at Shanghai. He is<br />
engaged in teaching information security,<br />
digital forensics and the theory <strong>of</strong><br />
information asurance. His major research<br />
interests include Network and information<br />
security, wireless and mobile services<br />
security support, multi-network integration, and digital crime<br />
investigation and forensic.<br />
Dingguo Yu was born in 1976, now is<br />
a PhD candidate <strong>of</strong> Tongji University,<br />
China. He received BS degree in<br />
mathematics in 1998 from Zhejiang<br />
Normal University, China, and received<br />
MS degree in computer application<br />
technology in 2005 from Tongji<br />
University. His current research interests<br />
include network & information security<br />
and mobile computing.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1041<br />
Image Compression Based on Improved FFT<br />
Algorithm<br />
Juanli Hu*<br />
Computer Engineering Department, Zhongshan Polytechnic, Zhongshan, China<br />
Email: hjlfoxes@163.com<br />
Jiabin Deng and Juebo Wu #<br />
Computer Engineering Department, Zhongshan Polytechnic, Zhongshan, China<br />
# Shenzhen Angelshine Co., LTD, Shenzhen, China<br />
Email: hugodunne@yahoo.com.cn and wujuebo@gmail.com<br />
Abstract—Image compression is a crucial step in image<br />
processing area. Image Fourier transforms is the classical<br />
algorithm which can convert image from spatial domain to<br />
frequency domain. Because <strong>of</strong> its good concentrative<br />
property with transform energy, Fourier transform has<br />
been widely applied in image coding, image segmentation,<br />
image reconstruction. This paper adopts Radix-4 Fast<br />
Fourier transform (Radix-4 FFT) to realize the limit<br />
distortion for image coding, and to discuss the feasibility<br />
and the advantage <strong>of</strong> Fourier transform for image<br />
compression. It aims to deal with the existing complex and<br />
time-consuming <strong>of</strong> Fourier transform, according to the<br />
symmetric conjugate <strong>of</strong> the image by Fourier transform to<br />
reduce data storage and computing complexity. Using<br />
Radix-4 FFT can also reduce algorithm time-consuming, it<br />
designs three different compression requirements <strong>of</strong> nonuniform<br />
quantification tables for different demands <strong>of</strong><br />
image quality and compression ratio. Take the standard<br />
image Lena as experimental data using the presented<br />
method, the results show that the implementation by Radix-<br />
4 FFT is simple, the effect is ideal and lower time-consuming.<br />
Index Terms—Image Compression, Fourier Transform,<br />
Quantization Table List, Compression Ratio, Coding and<br />
Decoding<br />
I. INTRODUCTION<br />
Image coding is a kind <strong>of</strong> method by using image<br />
source coding to achieve data compression, in order to<br />
ensure the quality <strong>of</strong> images and try to reduce code rate.<br />
Through image coding, it can get the goal for saving<br />
bandwidth or space, and it may also be provided for<br />
multimedia computer processing [1].With rapid<br />
development <strong>of</strong> multimedia and communication<br />
technology, it requires a higher demands for data storage<br />
and data transmission, especially in large volumes <strong>of</strong><br />
digital image communication, which greatly restricts the<br />
development <strong>of</strong> image communication. Therefore, more<br />
and more attentions are focused on image compression<br />
*Corresponding author.<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1041-1048<br />
techniques [2].<br />
Nowadays, there are many high compression rate<br />
methods for digital image that can be divided into three<br />
types: Waveform method, the second-generation coding,<br />
Fractal coding etc [3].<br />
A continuous tone image compression standard JPEG<br />
[4] is a common tool for static image compression, which<br />
allows both nondestructive compression and loss<br />
compression and it is representative <strong>of</strong> compression wave<br />
technology. Image is coded by prediction scheme with<br />
the conservative compression 2:1. In loss compression<br />
mode, the compression ratio can reach 5-20 times for<br />
most natural graphics providing better quality in the<br />
circumstances. However, its main defects are big<br />
distortion when doing compression with high<br />
ratio(blocking effect and mosaic noise) and lack <strong>of</strong> bits<br />
flow control and weak repair. On the condition that<br />
compression ration reaches 30-40, it may emerge a<br />
stronger blocking effect. In accordance with JPEG with<br />
high compression ratio, many improvement methods are<br />
proposed to overcome such obstacles like DCT zerotree<br />
coding and layer type DCT zerotree coding [5]. But in<br />
high compression ratio, the situation is still block-effect<br />
fatal weakness [6].<br />
Karhunen-Loeve Transform, also known as the<br />
characteristic vector Transform, principal component<br />
Transform or Hotelling, is a good way to process image<br />
with transformation matrix determined by the specific<br />
statistical characteristics <strong>of</strong> image(covariance matrix).<br />
The biggest advantage is the correlation between the<br />
transform domain can be removed totally, that is, owning<br />
well decorrelation [7]. Usually, K-L transform is used to<br />
eliminate the correlation among smaller matrix elements,<br />
such as to remove the spectrum correlation in many bands<br />
<strong>of</strong> remote sensing image compression. In practical<br />
application, because the transformation matrix is bigger<br />
and when the processing matrix is bigger, the covariance<br />
matrix is not easily either.<br />
For the past few years, since Wavelet Transform has<br />
the ability and characteristics <strong>of</strong> local signal analysis in<br />
time and frequency, it has been widely used in image<br />
denoising, image reinforcing and image compression.
1042 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Wavelet decomposition is a subset <strong>of</strong> the sub-bands<br />
decomposition for image, and it provides a type <strong>of</strong> multiresolution<br />
representation method.<br />
The basic ideas for image coding based on wavelet is<br />
conducting multistage wavelet decomposition to image<br />
by fast wavelet transform algorithm <strong>of</strong> Mallat tower, and<br />
then quantifying each layer <strong>of</strong> wavelet coefficients and<br />
coding. JPEG2000 is a new generation <strong>of</strong> image<br />
compression international standards developed by<br />
ISO/IEC organizations. It applies wavelet transform<br />
(DWT) as the basic algorithm, in combination with<br />
embedded coding technology, the result can not only<br />
reach in higher image quality and higher compression<br />
efficiency, but also meet the needs to the mobile and<br />
network environment and the interoperability and<br />
scalability. However, linear filter inherent ringing effect<br />
will appear if the method in the compression ratio reaches<br />
about 50.<br />
Structure Coding is the second generation coding<br />
technology for image, which is full consideration <strong>of</strong> the<br />
human visual physiological and psychological<br />
characteristics. Its principle is based on the meaning<br />
element <strong>of</strong> visual sense to describe images, such as the<br />
outline and texture. The second generation encoding<br />
technology can be divided into two categories: directional<br />
decomposition technique and face outline/characteristics.<br />
The objective <strong>of</strong> directional decomposition is to detect<br />
and express the edge information <strong>of</strong> image more<br />
accurately and more effectively, in order to apply proper<br />
separation and coding. Outline/feature oriented<br />
technology can make different reflection depending on<br />
different characteristics according to visual system. It<br />
extracts the main feature firstly, and then carries out the<br />
corresponding coding. Such method can give a better<br />
store for the image edge pr<strong>of</strong>ile information. Thus, the<br />
image quality remains a high level when high<br />
compression ratio.<br />
In the 1980s, Bamsley and Jacquin put IFS(Iterated<br />
Function Systems) into image compression. Fractal<br />
compression introduces the characteristics <strong>of</strong> self<br />
similarity to image compression. Through the iteration<br />
function <strong>of</strong> fractal image compression system, it realizes<br />
the segmentation <strong>of</strong> the original image, and then to map<br />
each subimage into an iterative function. Sub-images are<br />
stored by iterative function, the simpler the iterative<br />
function is, the bigger the compression ratio will be [8].<br />
Bamsley put forward Local IFS theory in 1988, which<br />
solved the problem <strong>of</strong> the parts and the whole without<br />
self similarity. After that, many studies indicated that the<br />
fractal image compression encoding remains very good<br />
quality when compression ratio is in higher lever (70-80).<br />
The main problem is the complexity in the phase <strong>of</strong><br />
image coding. Currently, the research on fractal<br />
technology focuses on the hybrid coding, fractal inverse<br />
problem and fractal convergence problems with<br />
improvement, fractal technology and other compression<br />
techniques (e.g., wavelet transform) [9].<br />
Because <strong>of</strong> the good nature <strong>of</strong> Fourier transform, it has<br />
a wide range <strong>of</strong> applications in the image coding, image<br />
segmentation, image reconstruction and other areas. DFT<br />
© 2011 ACADEMY PUBLISHER<br />
transform with good energy concentration, due to the<br />
inconvenient operations and the great amount <strong>of</strong><br />
calculation, hasn’t long been widely used in the image<br />
compression. For its complex algorithm and timeconsuming<br />
disadvantages, in this algorithm, the author<br />
makes use <strong>of</strong> the fast Fourier transform (FFT) to realize<br />
the limited distortion coding technology <strong>of</strong> the image.<br />
Investigating the feasibility <strong>of</strong> image compression using<br />
the Fourier transform, we utilize the conjugate symmetry<br />
to reduce the data storage. When reducing the timeconsuming,<br />
and for the different requirements <strong>of</strong> the<br />
compression ratio and the image quality, the Radix-4<br />
algorithm adopts three different quantization tables.<br />
Through the standard image compression ratio, root mean<br />
square signal noise ratio and the decoding consumption,<br />
we can draw that the Fourier transform method <strong>of</strong> the<br />
image compression not inferior to the JPEG compression<br />
system can get an ideal compression. Based on the<br />
conjugate symmetry <strong>of</strong> the Fourier transform, we can<br />
halve the data storage. By means <strong>of</strong> the Radix-4 Fourier<br />
transform, we can greatly reduce the time-consuming.<br />
II. FOURIER SPECTRUM ANALYSIS AND DESIGN OF IMAGE<br />
In the image processing, it <strong>of</strong>ten tends to do<br />
corresponding transformation for image in converting<br />
domain when facing to the problems that is complex and<br />
hard to deal with, in order to concentrate the energy on<br />
minority transform coefficient. Fourier Transform is a<br />
classical method to convert image from space domain to<br />
frequency domain, and it also the foundation <strong>of</strong> image<br />
processing titled as the second language for image<br />
description. It provides another perspective for image<br />
observation and the image can be transformed into gray<br />
distribution images to frequency distribution<br />
characteristics.<br />
In frequency domain, the more the frequency is, the<br />
faster the original signal changes. While the original<br />
signal is less changing as the less frequency. When<br />
frequency is 0, it means the dc signal has no change.<br />
Therefore, the size <strong>of</strong> the frequency reflects the signal<br />
changes. Most <strong>of</strong> the energy concentration in the image is<br />
located on the low dc and regional. Take Lena (512 x 512<br />
as shown in Fig. 1) for instance, the Fig. 2 shows the<br />
result <strong>of</strong> the Fourier transform. It is evidently seen from<br />
that the energy distribution <strong>of</strong> Lena image is focused in<br />
the low frequency part, to lower the frequency with the<br />
increase. Due to the real input signal, its distribution <strong>of</strong><br />
spectrum is axis-symmetric on 0 1 ,ω ω , so it can only take<br />
account into the half parts.<br />
This method executes image partitioning at the<br />
beginning to divide the data into non overlapping blocks.<br />
And then each part will be mapped by 2-dimension FFT<br />
where the coefficients are not related after the<br />
transformation and the energy <strong>of</strong> coefficient matrix is<br />
gathered in low-frequency area. After that, the<br />
quantization table will be designed to non-uniform<br />
according to the different requirements, and the<br />
quantitative results retain the coefficient <strong>of</strong> low frequency<br />
part while drop the high frequency coefficient. "Z"
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1043<br />
coding is done after the quantized data in this step.<br />
Finally, the entropy coding is conducted by Huffman<br />
coding so as to realize image compression. The entire<br />
process is shown in Fig. 3.<br />
Figure 1. Lena picture.<br />
Figure 2. The Fourier spectrum.<br />
Figure 3. The flowchart <strong>of</strong> image compression.<br />
III. MAPPING TRANSFORMATION OF IMAGE BY FFT<br />
Discrete Fourier Transform requires data discretization<br />
at the beginning in order to apply into computer<br />
technology. Discrete Fourier Transform is widely used in<br />
image processing and digital signal processing. Assume<br />
f (x)<br />
is number sequence with N length, and onedimensional<br />
discrete Fourier transform is defined as<br />
follows:<br />
F(<br />
u)<br />
N 1<br />
= ∑ −<br />
x=<br />
0<br />
f ( x)<br />
exp(<br />
− j2πux<br />
)<br />
N<br />
where u = 0, 1,...<br />
N −1<br />
. As shown in above formula,<br />
one F ( u)<br />
needs N times plural multiplication and<br />
2<br />
N −1<br />
times plural additio n; N F (u ) needs N times<br />
plural multiplication and<br />
N * ( N −1)<br />
times plural<br />
addition. Obviously, the bigger N is, the more calculation<br />
will<br />
be.<br />
Digital image f (m, n) is described as a matrix M rows<br />
by N columns [f (m, n)] in computer, and the following<br />
© 2011 ACADEMY PUBLISHER<br />
(1)<br />
definitions are the 2-dimension DFT and inverse<br />
transform <strong>of</strong> image matrix respectively.<br />
F(<br />
s,<br />
t)<br />
N−1<br />
1 ⎡<br />
* ⎢<br />
N n=<br />
0 ⎣<br />
1<br />
M<br />
tn<br />
∗exp(<br />
−j2π<br />
)<br />
N<br />
f ( m,<br />
n)<br />
M−1<br />
= ∑ ∑<br />
m=<br />
0<br />
sm ⎤<br />
f ( m,<br />
n)<br />
exp( −j2π<br />
)<br />
M<br />
⎥<br />
⎦<br />
(2)<br />
N−1<br />
1 ⎡<br />
* ⎢<br />
N t=<br />
0 ⎣<br />
M−1<br />
1<br />
sm ⎤<br />
F(<br />
s,<br />
t)<br />
exp( −j2π<br />
)<br />
M s 0<br />
M<br />
⎥<br />
= ⎦ (3)<br />
tn<br />
* exp( −j2π<br />
)<br />
N<br />
= ∑ ∑<br />
Two-dimensional Fourier transform can be seen as two<br />
times <strong>of</strong> one-dimensional, namely the Fourier transform<br />
<strong>of</strong> image sequence according to the ranks respectively.<br />
Obviously, the Fourier transform with N points need N *<br />
N times and N*(N-1) additions. Therefore, the bigger the<br />
image is the longer consuming the calculation will be and<br />
it is very important to choose one kind <strong>of</strong> fast algorithm.<br />
A. Figures and Tables<br />
Because the Fourier transform operation contains a lot<br />
<strong>of</strong> repetition computation, people studied many fast<br />
Fourier transform algorithm.<br />
In 1965, J.W.Cooley and J.W.Turky proposed fast<br />
Fourier transform algorithm. According to the<br />
composition <strong>of</strong> basic wing operation, the algorithm is<br />
divided into 2-radical, 4-radical, 8-radical, 16-radical<br />
arbitrary factor FFT algorithm. Currently, 2-radical and<br />
4-radical are the most widely used. The affiliate<br />
multiplication and wig <strong>of</strong> FFT are proportional to<br />
N log N , where the bigger N can save more<br />
2<br />
computation.<br />
FFT algorithm are as follows: Firstly, by using the<br />
basic idea <strong>of</strong> the three characteristics <strong>of</strong> rotating factors<br />
W<br />
kn<br />
N<br />
, namely, the periodicity, symmetry and about sex,<br />
the original N points <strong>of</strong> DFT long sequence are<br />
decomposed into two or more short sequences, and<br />
merging DFT operation fitting for combination. Secondly,<br />
recombine the DFT <strong>of</strong> the original sequence after<br />
calculation <strong>of</strong> short sequences, in order to improve the<br />
speed with less computation [10]. These can be divided<br />
into two kinds <strong>of</strong> decomposition methods:<br />
1) Transfer a large Discrete Fourier Transform<br />
computation into a group <strong>of</strong> short length, known as<br />
Decimation-In-Time-FFT (DIT-FFT).<br />
2) Decompose Fourier series X (k), known as<br />
Decimation-In- Frequency-FFT (DIF-FFT).<br />
The basic principle <strong>of</strong> radix-4 DIT-FFT [11] is as<br />
follows: Divide FFT with N points into 4 sequences and<br />
compute the DFT separately. Likewise, turn N/4 points<br />
into fine particle size, and so on. For multi-points (4m),<br />
multi-stage decomposition can be established in a similar<br />
way. For instance, based on radix-4 DIT-FFT, the series x<br />
(n) with 256 length can be obtained:
1044 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
x(<br />
k)<br />
= X(<br />
k k k k<br />
W<br />
W<br />
( K<br />
N<br />
3<br />
K0<br />
4 n3<br />
N<br />
3<br />
3 4 +<br />
W<br />
2<br />
2<br />
1<br />
0<br />
)<br />
2<br />
( K1<br />
4+<br />
K0<br />
) 4 n<br />
N<br />
2<br />
0<br />
3<br />
3 2 1 0<br />
n = 0n<br />
= 0n<br />
= 0n<br />
= 0<br />
K 4 + K 4+<br />
K ) n<br />
3<br />
0 1 2 3<br />
W<br />
2<br />
3<br />
( K2<br />
4 + K1<br />
4+<br />
K0<br />
) 4n1<br />
N<br />
3<br />
∑∑∑∑<br />
x(<br />
n n n n ) .<br />
Whilst the time complexity <strong>of</strong> radix-2 FFT is:<br />
3<br />
.<br />
2<br />
1<br />
0<br />
(4)<br />
Figure 4. The flowchart <strong>of</strong> radix-4 butterfly<br />
1<br />
= N log N<br />
(6)<br />
2<br />
m F 2<br />
In all kinds <strong>of</strong> FFT algorithms, radix-2 FFT is the<br />
simplest one but its calculation is more complex than<br />
using radix-4 FFT. People usually measure a performance<br />
<strong>of</strong> the algorithm by the times <strong>of</strong> addition and<br />
multiplication. For example, the radix-4 FFT is described<br />
as: each butterfly is 3 plural by N plural points, having<br />
log4N levels. Each level has N/4 radix-4 butterflies.<br />
Generally speaking, it shows the higher cardinal the less<br />
computation, but to judge whether an algorithm good or<br />
not is not only to consider the calculation but also the<br />
complexity. From radix-2 to radix-4, the number <strong>of</strong><br />
multiplication and addition has a big jump and the count<br />
<strong>of</strong> radix-4 FFT general reduces about 1/4 by radix-2.<br />
From 4 to 8, or 8 to 16, the number has not a obvious<br />
change. Regarding algorithm complexity, radix-2 is<br />
easiest to control and use while radix-4 is harder to<br />
control, but radix-4 is the analogy as radix-2. Compared<br />
with radix-4, the jump <strong>of</strong> radix-8 and radix-16 is very<br />
obvious. Considering speed and control complexity,<br />
radix-4 has the highest realization ratio in FFT. Therefore,<br />
this paper selects radix-4 FFT algorithm to realize FFT<br />
processor design.<br />
© 2011 ACADEMY PUBLISHER<br />
The radix-4 butterfly processing is shown as Fig. 4.<br />
The time complexity <strong>of</strong> radix-4 FFT is as following<br />
where there are N/4 FFT 4-points and L levels totally.<br />
N 3<br />
mF 3× × ( L −1)<br />
≈ N log2<br />
N<br />
4 8<br />
= (5)<br />
B. The feature and data store <strong>of</strong> conjugate symmetry <strong>of</strong><br />
Fourier transform<br />
By using Fourier transform, the result is plural number,<br />
such as matrix 8*8 will be converted to 8*8 plural matrix.<br />
It seems that the data is double than original one, but the<br />
fact is that the real data quantity does not increase as a<br />
result <strong>of</strong> the conjugate symmetry properties. After DFT<br />
transform, the matrix 8*8 will be:<br />
*<br />
F( s,<br />
t)<br />
= F ( 8 − s,<br />
8 − t)<br />
Where reflects the cycle around s=4 or t=4 in a space<br />
spectrum with conjugate symmetry properties as shown<br />
in Fig. 5.<br />
Figure 5. The conjugate symmetry <strong>of</strong> Fourier transform<br />
(7)
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1045<br />
In such process, the real number includes F(0, 0), F(0,<br />
4), F(4, 0) and F(4, 4), the others is conjugate symmetric.<br />
F (0, 0) is "dc" component, F(s, 4) and F (4, t) are high<br />
frequency area and F(4, 4) is the highest. The shaded<br />
parts are the valid data.<br />
The method <strong>of</strong> data store after FFT is closely related to<br />
quantization table, and this paper adopts the following<br />
way to save data. Taking image into 8*8 sub-blocks, it<br />
⎡ F(<br />
0,<br />
0)<br />
⎢<br />
⎢<br />
F(<br />
1,<br />
0).<br />
R<br />
⎢F(<br />
2,<br />
0).<br />
R<br />
⎢<br />
⎢F(<br />
3,<br />
0).<br />
R<br />
⎢ F(<br />
4,<br />
0)<br />
⎢<br />
⎢F(<br />
1,<br />
4).<br />
R<br />
⎢F(<br />
2,<br />
4).<br />
R<br />
⎢<br />
⎢⎣<br />
F(<br />
3,<br />
4).<br />
R<br />
F(<br />
0,<br />
1).<br />
R<br />
F(<br />
1,<br />
0).<br />
I<br />
F(<br />
2,<br />
0).<br />
I<br />
F(<br />
3,<br />
0).<br />
I<br />
F(<br />
4,<br />
1).<br />
R<br />
F(<br />
1,<br />
4).<br />
I<br />
F(<br />
2,<br />
4).<br />
I<br />
F(<br />
3,<br />
4).<br />
I<br />
F(<br />
0,<br />
1).<br />
I<br />
F(<br />
11,<br />
). R<br />
F(<br />
2,<br />
1).<br />
R<br />
F(<br />
31,<br />
). R<br />
F(<br />
4,<br />
1).<br />
I<br />
F(<br />
51,<br />
). R<br />
F(<br />
6,<br />
1).<br />
R<br />
F(<br />
7,<br />
1).<br />
R<br />
F(<br />
0,<br />
2).<br />
R<br />
F(<br />
1,<br />
1).<br />
I<br />
F(<br />
2,<br />
1).<br />
I<br />
F(<br />
3,<br />
1).<br />
I<br />
F(<br />
4,<br />
2).<br />
R<br />
F(<br />
5,<br />
1).<br />
I<br />
F(<br />
61,<br />
). I<br />
F(<br />
7,<br />
1).<br />
I<br />
IV. DESIGN AND IMPLEMENT OF QUANTIZATION TABLE<br />
Quantization is the most important step in the<br />
compression method. Its function is to map the<br />
continuous transform coefficient into limited data set.<br />
Quantitative can be divided into the scalar quantization<br />
and vector quantization. The scalar quantification can be<br />
divided into uniform and nonuniform quantification and<br />
visual quantification in detail. Visual quantification is in<br />
the process <strong>of</strong> considering quantification <strong>of</strong> different<br />
visual band which has different sensitive degree <strong>of</strong><br />
properties, in order to apply the large quantization step to<br />
high frequencies. Vector quantization is an image block<br />
coding method, and the process is to map vector into<br />
predefined code book where selection <strong>of</strong> books is<br />
according to the statistic characteristics before<br />
quantitative. Vector quantitative is superior to scalar<br />
quantitative theoretically, but the cost <strong>of</strong> actual<br />
implementation is too big. Combined with other methods<br />
to realize in hardware is the development direction in the<br />
future. Considering the overall performance <strong>of</strong> the<br />
algorithms, the scalar quantitative is adopted in this<br />
algorithm.<br />
Subjective evaluation method is to observe an image<br />
directly and justify the degree <strong>of</strong> distortion from feeling.<br />
After that, the evaluation <strong>of</strong> quality score is given by the<br />
weighted average scores from all comments and the<br />
results are the subjective evaluation results, including<br />
absolute scale and relative scale.<br />
This evaluation system is consistent with the visual<br />
perception in relative to the objective evaluation method,<br />
which is reliable. But it cannot be used in convenient way:<br />
Firstly, it cannot be used in the process <strong>of</strong> image coding<br />
quality evaluation and control. Secondly, the subjective<br />
assessment <strong>of</strong> the testee is susceptible to the effects <strong>of</strong><br />
subjective factors, such as age, and deviation, education<br />
level, personality, cultural background, etc. Thus, this<br />
© 2011 ACADEMY PUBLISHER<br />
can apply symmetry way to compensate the missing<br />
blocks if they fail to meet the sub-block. If the image is<br />
true color with 24 bits, RGB should be used to code.<br />
Then, FFT is carried out for all <strong>of</strong> the matrix 8*8, and the<br />
result can be saved half according to conjugate symmetric,<br />
that is 8*8 real number.<br />
F(<br />
0,<br />
2).<br />
I<br />
F(<br />
1,<br />
2).<br />
R<br />
F(<br />
2,<br />
2).<br />
R<br />
F(<br />
3,<br />
2).<br />
R<br />
F(<br />
4,<br />
2).<br />
I<br />
F(<br />
5,<br />
2).<br />
R<br />
F(<br />
6,<br />
2).<br />
R<br />
F(<br />
7,<br />
2).<br />
R<br />
Figure 6. The matrix 8*8<br />
F(<br />
0,<br />
3).<br />
R<br />
F(<br />
1,<br />
2).<br />
I<br />
F(<br />
2,<br />
2).<br />
I<br />
F(<br />
3,<br />
2).<br />
I<br />
F(<br />
4,<br />
3).<br />
R<br />
F(<br />
5,<br />
2).<br />
I<br />
F(<br />
6,<br />
2).<br />
I<br />
F(<br />
7,<br />
2).<br />
I<br />
F(<br />
0,<br />
3).<br />
I<br />
F(<br />
1,<br />
3).<br />
R<br />
F(<br />
2,<br />
3).<br />
R<br />
F(<br />
3,<br />
3).<br />
R<br />
F(<br />
4,<br />
3).<br />
I<br />
F(<br />
5,<br />
3).<br />
R<br />
F(<br />
6,<br />
3).<br />
R<br />
F(<br />
7,<br />
3).<br />
R<br />
F(<br />
0,<br />
4)<br />
⎤<br />
F(<br />
1,<br />
3).<br />
I<br />
⎥<br />
⎥<br />
F(<br />
2,<br />
3).<br />
I⎥<br />
⎥<br />
F(<br />
3,<br />
3).<br />
I⎥<br />
F(<br />
4,<br />
4)<br />
⎥<br />
⎥<br />
F(<br />
5,<br />
3).<br />
I⎥<br />
F(<br />
6,<br />
3).<br />
I⎥<br />
⎥<br />
F(<br />
7,<br />
3).<br />
I⎥⎦<br />
paper adopts the objective and subjective evaluation<br />
method <strong>of</strong> combining ways.<br />
The scalar qualitative contains uniform quantification<br />
and nonuniform quantification. The uniform<br />
quantification is to quantify the range <strong>of</strong> values for input<br />
signal by equidistance division. In comparison with<br />
uniform quantification, the nonuniform quantification has<br />
two advantages. One is the output <strong>of</strong> nonuniform<br />
quantification can obtain higher average signal<br />
quantization noise power when the input is non-uniform<br />
probability density. The other is the quantization noise<br />
power and value is in proportion to signal sampling value.<br />
Thus, the signal quantitative signal-to-noise ratio can be<br />
improved.<br />
This algorithm adopts non-uniform quantification. In<br />
this process, the low frequency signal has more important<br />
information, so it requires higher quantitative precision<br />
than the low energy. Since the high frequency signal is<br />
not sensitive to eye, the low-frequency coefficient is set<br />
in a smaller value while the high frequency part set larger.<br />
In this paper, three varieties <strong>of</strong> quantization tables are<br />
utilized to set image data according to the different<br />
requirements for quality. We quantify the results after<br />
FFT to control the compression ratio for different image<br />
quality, as shown in table 1~3.<br />
TABLE I.<br />
THE QUANTIZATION TABLE WITH HIGH COMPRESSION RATIO<br />
8 20 20 32 32 40 40 60<br />
20 20 32 32 40 40 60 60<br />
32 32 40 40 60 60 80 80<br />
40 40 60 60 80 80 80 80<br />
60 80 80 100 100 100 100 100<br />
40 40 60 60 80 80 80 80<br />
60 60 40 40 60 60 80 80<br />
80 80 32 32 40 40 60 60
1046 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
TABLE II.<br />
THE QUANTIZATION TABLE WITH MIDDLE COMPRESSION RATIO<br />
8 16 16 16 16 32 32 40<br />
16 16 20 20 32 32 40 40<br />
16 16 32 32 40 40 60 60<br />
32 32 40 40 60 60 60 60<br />
40 40 40 60 60 80 80 100<br />
32 32 40 40 60 60 60 60<br />
40 40 32 32 40 40 60 60<br />
60 60 20 20 32 32 40 40<br />
TABLE III.<br />
THE QUANTIZATION TABLE WITH LOW COMPRESSION RATIO<br />
8 12 12 12 12 20 20 32<br />
12 12 16 16 20 20 32 32<br />
12 12 20 20 32 32 40 40<br />
20 20 32 32 40 40 40 40<br />
32 32 32 40 40 40 40 60<br />
20 20 32 32 40 40 40 40<br />
32 32 20 20 32 32 40 40<br />
40 40 16 16 20 20 32 32<br />
This algorithm applies "Z" run length encoding to<br />
coefficients after quantification and Huffman entropy to<br />
coding results in order to achieve image compression.<br />
Entropy coding is a kind <strong>of</strong> lossless coding and the<br />
common methods are composed <strong>of</strong> Run-length encoding<br />
(RLE), Lempel-Ziv-Welch(LZW), Shannon, Huffman<br />
and Arithmetic coding. The basic idea <strong>of</strong> entropy coding<br />
is using short code to express the signal with larger<br />
probability and long code for less signal. It gains a shorter<br />
length in statistic to reduce the average code data <strong>of</strong> space<br />
and improve the compression ratio.<br />
V. EVALUATION ANALYSIS FOR CODE EFFICIENCY AND<br />
COMPRESSION QUALITY<br />
The meaning <strong>of</strong> image quality includes two layers:<br />
One is to reconstruct the distortion degree <strong>of</strong> image with<br />
the original image reconstruction and deviation degree.<br />
The other is the readability <strong>of</strong> the image, that is, the<br />
information from the image who gets. Normally, the eye<br />
is the information receiver <strong>of</strong> image, but the visual<br />
Lena Image<br />
quality<br />
Good<br />
(lossless)<br />
Bitmap File<br />
Size (bytes)<br />
Compressed<br />
file size<br />
TABLE IV.<br />
LENA IMAGE TE ST RESULTS<br />
Compression<br />
ratio<br />
system is very limited and cannot understand the<br />
distortion degree <strong>of</strong> image with readability and<br />
quantitative description. So, image quality assessment<br />
needs an objective evaluation method besides the<br />
subjective evaluation methods.<br />
In objective evaluation methods,<br />
the mean square error<br />
(MSE) and mean square signal to noise ratio (SNR) are<br />
commonly used. SNR is defined as: if the compressed<br />
image is represented as the superposition <strong>of</strong> the original<br />
image and noise, that is<br />
f ( x,<br />
y)<br />
= g ( x,<br />
y)<br />
+ e(<br />
x,<br />
y)<br />
(8)<br />
∑∑<br />
∑∑<br />
2<br />
2<br />
( SNR ) RSM = f ( x,<br />
y)<br />
/ e ( x,<br />
y)<br />
(9)<br />
For a digital image compression coding system, we can<br />
use<br />
redundant, coding efficiency and compression ratio to<br />
measure the source characteristics and encoding /<br />
decoding performance. The compression efficiency <strong>of</strong><br />
image is usually measured with the compression ratio. the<br />
higher the compression ratio, the greater the image<br />
compression, and vice versa. In addition, it till needs to<br />
consider the complexity <strong>of</strong> the algorithm, including time<br />
complexity and space complexity. In this paper the<br />
algorithm coding efficiency is taken as the main<br />
consideration, and the coding efficiency can be reflected<br />
through the time-consuming <strong>of</strong> the encoding and<br />
decoding.<br />
It should be noted that the MSE and SNR reflect the<br />
overall<br />
differences between the original image and<br />
reconstructed image, and do not reflect the local<br />
differences. Sometimes in the same signal to noise ratio,<br />
visual effects will still have some differences, primarily<br />
due to the uniformity <strong>of</strong> the error. In general, if the<br />
uniformity <strong>of</strong> the error is high, the visual effect is good,<br />
otherwise the visual effects is bad. In most cases, we can<br />
use the PSNR to evaluate the image quality, but<br />
sometimes the results may be inconsistent with the<br />
subjective evaluation. But sometimes the outcome may<br />
have deviation from the results <strong>of</strong> subjective evaluation.<br />
This algorithm uses standard image (Lena) as the data<br />
analysis standard, the results shown in Table 4. Images<br />
512x512 pixel 24-bit True Color. Compression results<br />
map, as shown in Fig. 7~ Fig. 10.<br />
SNR<br />
Coding time -<br />
consuming<br />
(seconds)<br />
Decoding timeconsuming<br />
(seconds)<br />
786486 765279 1.03 Infinite 0.131 0.47<br />
Better 786486 94588 8.31 20.12 0.280 0.2115<br />
Medium 786486 77396 10.16 17.88 0.258 0.211<br />
Lower 786486 62245 12.64 15.71 0.280 0.211<br />
© 2011 ACADEMY PUBLISHER<br />
Test environment: CPU: Pentium® Dual-core 2.5 G; Memory: 2G DDR; OS: Windows XP.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1047<br />
Figure 7. Lossless compression<br />
Figure 8. Compression ratio by 8.31<br />
Figure 9. Compression ratio by 10.16<br />
Figure 10. Compression ratio by 12.64<br />
© 2011 ACADEMY PUBLISHER<br />
It can be seen that Fourier transform limited<br />
compression system is better than distortion coding,<br />
design coding and decoding time <strong>of</strong> the three. But the<br />
compression ratio is about 10 and most above 0.25<br />
seconds with decoding time about 0.2 and visual<br />
compression ratio <strong>of</strong> more than 12.64 where the image<br />
has a slight blocking effect.<br />
VI. CONCLUSIONS<br />
Fourier transform has a good transformation energy<br />
concentration, but for a long time, the results <strong>of</strong> Fourier<br />
transform will be plural, leading to the high complexity<br />
<strong>of</strong> encoding and long time <strong>of</strong> encoding/decoding, so can’t<br />
be widely used. This paper uses Fourier transform<br />
compression coding system based on 4FFT algorithm.<br />
Can be seen from Figure 6,7,8,9, image compression is<br />
better, image compression ratio is at 12.64 and above, a<br />
slight blocking effect will appear. Can be seen from<br />
Table 4, when the compression ratio is at about 10 and<br />
above, the time <strong>of</strong> compression is between 0.25-0.28<br />
seconds, with the increase <strong>of</strong> compression ratio,<br />
compression time <strong>of</strong> encoding/decoding change slightly.<br />
Decoding time is less than the encoding time at about<br />
0.21 seconds. Fourier transform image compression<br />
method can also produce ideal compression results.<br />
According to the conjugate symmetry <strong>of</strong> Fourier<br />
transform data, data storage can be reduced by half.<br />
Through the use <strong>of</strong> Radix-4 FFT, the algorithm can<br />
greatly reduce the time-consuming <strong>of</strong> algorithm.<br />
Reference to the direction <strong>of</strong> future research: in the<br />
compression rate, we can adopt several Fourier transform<br />
to get more focused information and reduce the<br />
correlation between pixels; do segmentation coding<br />
according to the characteristics <strong>of</strong> spectrum in different<br />
region; After FFT transform, plural data store the matrix,<br />
high-frequency information are set into the lower right<br />
corner <strong>of</strong> the matrix, low-frequency information are set<br />
into the top left corner, in order to improve its<br />
compression ratio. We can use the assembly language in<br />
FFT module to improve the encoding/decoding speed, or<br />
use hardware or other fast FFT methods to achieve. The<br />
biggest problems <strong>of</strong> Fourier transform and JPEG<br />
compression method are the severe block effects in the<br />
high compression ratio. In future work, we should<br />
consider combining the human visual characteristic to<br />
compress.<br />
ACKNOWLEDGMENT<br />
The authors wish to thank their colleagues and external<br />
advisors for their help and support, in particular, Shuliang<br />
Wang (Ph.D. Pr<strong>of</strong>essor, Wuhan University, China).<br />
This paper is supported by National 973<br />
(2007CB310804), National Natural Science Fund <strong>of</strong><br />
China (60743001), Best National Thesis Fund (2005047),<br />
and Natural Science Fund <strong>of</strong> Hubei Province (CDB132).
1048 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
REFERENCES<br />
[1] W. B. Penneder, J. L. Mitchell. JPEG: Still Image Data<br />
Compression Standard, Van Nostrand Reinhoid. New York:<br />
A cademic Press,1993.<br />
[2] X. C. Zhu, F. Liu, D. Hu. Digital image processing and<br />
communication. Peking University Press, 2002,108-132.<br />
[3] J. H. Xu. Image processing and analysis. Beijing: science<br />
press, 1992.79-86.<br />
[4] G. A. Rong. Computer image processing. Tsinghua<br />
university press, 2000.89-98<br />
[5] P. Kauff, K. Schuur. Shape-adaptive DCT with blockbased<br />
DC separation and Delta DC correction [J].IEEE<br />
Trans. Circuits Syst. Video Technol., 1998, 8 (3):237-242.<br />
[6] M. Bi, S. H. Ong, Y H Ang. Comment on "Shape-adaptive<br />
DCT for generic coding <strong>of</strong> video [J].IEEE Trans. Circuits<br />
Syst. Video Technol.,1996 6(6): 686-688.<br />
[7] Huazhong university <strong>of</strong> science and technology department.<br />
The probability and statistics. Higher Education<br />
Press,1999.104-128<br />
[8] O. Egger, P. Fleury, T. Ebrahimi. Shape-adaptive wavelet<br />
transform for zerotree coding [C]. Proc. Eur. Workshop<br />
Image Analysis and Coding for TV, HDTV and<br />
Multimedia Application Rennes France, 1996: 201–208.<br />
[9] [9] O. Egger. Region representation using nonlinear<br />
techniques with applications to image and video coding<br />
[D].Ph.D. dissertation, Swiss Federal Institute <strong>of</strong><br />
Technology (EPFL), Lausanne, Switzerland, 1997.<br />
[10] X. D Duan, L. Z. Gu. High-performance Radix-4 FFT<br />
processing Design. Computer Engineering, Vol.34 No.24 ,<br />
2008, 238-243.<br />
[11] K. Miyase, S. Kajihara. Optimal Scan Tree Construction<br />
with Test Vector Modification for Test<br />
Compression[C]//Proc. <strong>of</strong> IEEE Asian Test Symposium. [S.<br />
l.]: IEEE Press, 2003.<br />
© 2011 ACADEMY PUBLISHER<br />
Juanli Hu is currently lecturing at the<br />
Computer Engineering Department <strong>of</strong><br />
Zhong Shan Polytechnic. She has been<br />
working as Director <strong>of</strong> Teaching<br />
(Fundamental Computing Research)<br />
since 2006. She received her Bachelor<br />
Degree in Computer Engineering from<br />
Xi'an University <strong>of</strong> Technology in 2000, followed by a<br />
Master Degree in Computer Engineering from Xi'an<br />
University <strong>of</strong> Technology in 2005. Her research interests<br />
include signal, information processing and data mining.<br />
Some <strong>of</strong> her research papers are indexed by IEEE CS.<br />
Jiabin DENG is a lecturer at the<br />
Computer Engineering Department <strong>of</strong><br />
Zhong Shan Polytechnic. He received<br />
a Bachelor Degree in Computer<br />
Engineering from Hubei University<br />
in 2005, then a Master Degree in<br />
Computer Engineering from Wuhan<br />
University in 2007. His research<br />
interests include artificial intelligence,<br />
data mining, and complex network. His research papers<br />
are also indexed by IEEE CS etc.<br />
Juebo Wu was born in China and has<br />
obtained B.A. and M.A respectively in<br />
2005 and 2007 from International<br />
School <strong>of</strong> S<strong>of</strong>tware, Wuhan University,<br />
China. At present, he is working for his<br />
PH.D. candidate <strong>of</strong> Mapping and<br />
Remote Sensing in State Key<br />
Laboratory <strong>of</strong> Information Engineering<br />
in Surveying(Wuhan University, China)<br />
and will graduate in the summer 2010.<br />
Juebo Wu's primary research is spatial data mining,<br />
GIS and s<strong>of</strong>tware engineering, etc. In recent years, he has<br />
published more than 10 papers indexed by EI/ISTP and<br />
won two computer s<strong>of</strong>tware copyright.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1049<br />
Correlative Peak Interval Prediction and Analysis<br />
<strong>of</strong> Chaotic Sequences<br />
Qun Ding<br />
Electronic Engineering Key Laboratory <strong>of</strong> Universities in Heilongjiang Province, Heilongjiang University, Harbin,<br />
China, Email: qunding@yahoo.com<br />
Lu Wang, Guanrong Chen<br />
Electronic Engineering Key Laboratory <strong>of</strong> Universities in Heilongjiang Province, Heilongjiang University, Harbin,<br />
China,<br />
Department <strong>of</strong> Electronic Engineering, City University <strong>of</strong> Hong Kong, Hong Kong, China<br />
Abstract—The paper proposes a digital circuit design for<br />
the logistic-map module used in chaotic stream ciphers,<br />
analyzes the factors that may affect the output <strong>of</strong> the<br />
sequences, and develops a calculation method for estimating<br />
the output sequential correlative peak interval. With the<br />
respective tests using different initial values, the values <strong>of</strong><br />
parameter u and the computational precisions, extensive<br />
experiments have been carried out. A<br />
formula for calculating correlative peak interval is<br />
proposed. Moreover, the relationships among precision,<br />
parameter u and correlative peak interval is provided. To<br />
ensure the security <strong>of</strong> the plaintext which is encrypted by<br />
the output sequence <strong>of</strong> the logistic-map, a proper precision<br />
could be chosen according to the formula. It provides a<br />
theoretic basis for the actual application <strong>of</strong> the chaos<br />
cryptology. The basic theory and methods have a significant<br />
implication on the statistical analysis and practical<br />
applications <strong>of</strong> the digital chaotic sequences. A diagram that<br />
presents the relationship among precision, parameter u and<br />
correlative peak interval has been generated for analysis.<br />
Index Terms—discrete chaotic systems, correlative peak<br />
interval, finite precision, encryption<br />
I. INTRODUCTION<br />
Chaos theory has been studied extensively for many<br />
years, that the analysis <strong>of</strong> chaotic characteristics and<br />
practical applications has significant meaning in the<br />
research <strong>of</strong> chaos [1,2]. Chaos has many prominent<br />
features, such as the output <strong>of</strong> the system is sensitive<br />
depended on the initial conditions; the output <strong>of</strong> the<br />
system has the feature <strong>of</strong> long-term unpredictability; the<br />
orbit is irregularity; the output <strong>of</strong> the system has<br />
random-like behaviors. Since the properties <strong>of</strong> the chaos<br />
are desirable by cryptographic applications, more<br />
attention has been paid on the research <strong>of</strong> cryptography<br />
with chaos. Chaos theory has had many applications on<br />
the cryptography, such as encrypting still image with<br />
chaotic maps; generating pseudorandom sequences with<br />
chaotic sequences instead <strong>of</strong> the conventional<br />
m-sequences; using chaotic maps as key generators in<br />
ciphers design. Particularly, stream cipher based on<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1049-1056<br />
Email: wanglu.daily@gmail.com,eegchen@cityu.edu.hk<br />
chaotic maps can solve some difficult problems in<br />
nonlinear sequential cipher [3,4. Consequently it provides<br />
a new approach for information security application by<br />
increasing the complexity <strong>of</strong> deciphering.<br />
Chaotic stream ciphers are desirable for encryption<br />
devices and secure communication systems. However,<br />
when a chaotic device or system is realized by a<br />
computer which the precision is finite, the resultant<br />
discrete dynamics are different from that <strong>of</strong> the original<br />
analog system. Although there are some methods that can<br />
be used to improve the quality <strong>of</strong> the discredited chaotic<br />
systems, such as small-perturbation algorithms or<br />
multiple cascading chaotic systems [5,6], the finite<br />
precision <strong>of</strong> the computer is the main problem in<br />
application <strong>of</strong> chaos. Therefore it is difficult to set up an<br />
output sequence mathematical model which restricts the<br />
application <strong>of</strong> chaos. With the improvement <strong>of</strong><br />
computational accuracy and operational speed <strong>of</strong><br />
large-scale integrated circuits, the intrinsic degenerating<br />
phenomena <strong>of</strong> various chaotic characteristics can be<br />
studied more precisely which may promote the<br />
applications <strong>of</strong> chaos theory, especially in cryptography<br />
and secure communications[7,8].<br />
II. CHAOS THEORY<br />
Bifurcation chart is a description <strong>of</strong> state variant<br />
based on parameter space. In the bifurcation chart, the<br />
range <strong>of</strong> chaotic parameter the process <strong>of</strong> bifurcation and<br />
the period window are very clear. The general form <strong>of</strong> me<br />
to itself mapping is:<br />
xn+ 1 f( μ,<br />
xn), xn R<br />
= ∈ (1)<br />
f : I → I is the differentiable function , μ is a<br />
parameter。If from a certain initial value, the value <strong>of</strong><br />
xn state repeats infinite cycle among p( p≥ 1)<br />
states,<br />
and then (2) is a periodic orbit. If the period is 1, it means<br />
it is a fixed point.
1050 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
x′ 1, x′<br />
2,<br />
�, x ′ p<br />
(2)<br />
According to the method <strong>of</strong> linear stability, the<br />
conditions for stable periodic orbits is<br />
p<br />
∏ f′ ( μ,<br />
x′<br />
t ) ≤ 1<br />
(3)<br />
t=<br />
1<br />
and the periodic orbit p is called super-stable.<br />
We discuss the bifurcation chart with the example <strong>of</strong><br />
logistic map. Logistic map is defined as [9]:<br />
[ ] [ ]<br />
xn+ 1 μx(1 x), μ 0, 4 xn<br />
0,1<br />
= − ∈ ∈<br />
(4)<br />
from (4) we get the fixed point:<br />
1<br />
O: x= 0; A: x=<br />
1−<br />
.The stability <strong>of</strong> fixed point is<br />
μ<br />
determined by the gradient f ′ ( x)<br />
<strong>of</strong> y = f( x)<br />
,<br />
which is:<br />
f ′ ( x) = μ − 2xμ<br />
(5)<br />
Consequently, the stability <strong>of</strong> fixed points depends<br />
on the parameters μ .From the behavior <strong>of</strong> iterative<br />
equation (4), it relies on the steepness <strong>of</strong> the parabola<br />
sensitively which has the same meaning with nonlinearity.<br />
Therefore, when the parameter μ becomes larger from<br />
zero, the iterative process (4) has different dynamic<br />
behavior [10]:<br />
When 0< μ < 1,<br />
fix an initial value x0 in [0,1], and<br />
then iteration process moves to a fixed point quickly<br />
xn → 0 , due to f ′ (0) = μ < 1,<br />
so it exist stable fix<br />
point O.<br />
When μ = 1 , f ′ (0) = 1 , collapse bifurcation<br />
occurs.<br />
When 0< μ ≤ 3=<br />
μ1<br />
, there are two fixed points O<br />
and A. Because <strong>of</strong> f ′ (0) = μ > 1 , so the point O is<br />
1<br />
unstable. For the point <strong>of</strong> A, f ′ (1 − ) = 2 − μ < 1.<br />
Therefore the iterative process <strong>of</strong> initial value x 0 moves<br />
away from fixed point O and closer to the fixed point A.<br />
For example, when μ = 2 , after the iteration xn → 0.5 ,<br />
this called period 1, such can be seen from figure 1(a).<br />
When μ = 3 , as f′ ( A) = 2− μ =− 1 , so it<br />
generates fork-type bifurcation<br />
When 3< μ ≤ 1+ 6 = μ2<br />
, as f ′ (0) = μ > 1 ,<br />
it is still unstable; For the point A,<br />
1<br />
f ′ (1 − ) = 2 − μ > 1,<br />
so the point A changes from<br />
μ<br />
stable to unstable, such can be seen from figure 1(b)<br />
© 2011 ACADEMY PUBLISHER<br />
μ<br />
which is called period 2.<br />
When 3.449 < μ < 3.545 = μ3<br />
, the two values <strong>of</strong><br />
period 2 changes unstable again and generates a couple <strong>of</strong><br />
new fixed points, so x n jumps among the four values,<br />
such can be seen from figure 1(c) which is called period 4.<br />
Until μ > 3.57 = μ∞ , the time<br />
series x0, x1, x2, �, xn,<br />
� are like the random number<br />
in [0, 1], so it is called chaos, such can be seen from figure<br />
1(d)[11,12].<br />
(a)period 1 Time series<br />
(b)period 2 Time series<br />
(c) period 4 Time series
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1051<br />
(d) Chaotic Time series<br />
Figure 1. Different dynamic behavior <strong>of</strong> logistic map<br />
Nonlinear equations change the topology structure <strong>of</strong><br />
system trajectory through the above dynamic behavior<br />
which will cause the overall shape <strong>of</strong> the system changes<br />
suddenly and produce the phenomenon <strong>of</strong> bifurcation [13].<br />
This is a necessary process <strong>of</strong> the generation <strong>of</strong> chaos, and<br />
it is also the source that chaos fits cryptographic<br />
properties. Therefore, it has become a new trend <strong>of</strong><br />
cryptography to design new cryptogram program based on<br />
chaotic systems. With the large scale application <strong>of</strong><br />
integrated circuit in the chaotic encryption, we need a<br />
uniform standard to evaluate the models <strong>of</strong> digital chaotic<br />
system which still has the characteristic <strong>of</strong> randomness<br />
[14]. Through the study we found that although the<br />
application <strong>of</strong> chaos cannot fully rely on its features, it<br />
couldn’t separate its properties from the equation, such as:<br />
correlation function, bifurcation maps. Consequently, we<br />
try to construct chaotic and design the characteristics <strong>of</strong><br />
chaotic bifurcation with FPGA hardware platform design.<br />
Thus, the method <strong>of</strong> digital chaos has been proposed,<br />
which will test our security systems and play an important<br />
role in promoting the application <strong>of</strong> chaos [15].<br />
Ⅲ. CIRCUIT DESIGN FOR LOGISTIC PSEUDORANDOM<br />
SEQUENCES<br />
The signal also has the statistical property <strong>of</strong> the<br />
correlation that the interval between the relevant peak<br />
value is equivalent and the interval is usually<br />
unpredictable. Although there is correlation, as the<br />
interval between the peak value is large and the circuit is<br />
simple which is suitable for integrated circuit design, it<br />
still has practical value in some encryption occasions [16].<br />
In this paper, we will define the correlative feature <strong>of</strong> the<br />
digital chaotic sequence which is the equivalent interval<br />
between the peak value as the correlative peak interval.<br />
The estimate <strong>of</strong> correlative peak interval is an important<br />
© 2011 ACADEMY PUBLISHER<br />
parameter during the research <strong>of</strong> the chaotic application.<br />
The correlative peak interval <strong>of</strong> the simple<br />
one-dimensional logistic chaotic sequence is researched<br />
in this paper.<br />
The well-known logistic map is (4), if the chaotic<br />
logistic map is used in a cryptosystem; we have to ensure<br />
the output sequence is pseudorandom. To ensure the<br />
logistic map in the region <strong>of</strong> chaos, the key parameter u<br />
has to be chosen exactly.<br />
The circuit design diagram <strong>of</strong> the new method is<br />
shown as Figure 2 ,the theory <strong>of</strong> schematic circuit could<br />
be seen from [17,18], which has many modules, such as<br />
operational amplifier (for u), adder, multiplier, and delay<br />
units. It can generate modulated chaotic output sequences<br />
via control-shifting and extraction. A 128-bit data<br />
sequence goes through a data processor which is as the<br />
initial key .The selector controls the selection. If the SEL<br />
is 1, then pass the initial value; else if the SEL is 0, RES<br />
is 0 and EN is 1, then generate a logistic-map output by<br />
repeated iterations.<br />
Figure 2. Design diagram <strong>of</strong> the chaotic logistic-map module<br />
After the circuit design <strong>of</strong> logistic chaotic equation<br />
which is based on FPGA[19], we get the output <strong>of</strong> the<br />
simulation that is shown as Figure 3. It could see that<br />
when the control signals RES、SEL、EN are valid, a total<br />
<strong>of</strong> 128 initial key signal which is from Input1 to Input8<br />
will generate the output signal <strong>of</strong> the chaotic sequence.<br />
After a large number <strong>of</strong> auto-correlation tests, the result<br />
indicates that there is periodic interval between the peak<br />
value <strong>of</strong> the sequence which is shown as Figure 4. The<br />
correlation peak with the same interval and the interval is<br />
usually unpredictable. This paper carries out the<br />
theoretical analysis and calculation on the correlative<br />
peak interval.
1052 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
As a key sequence generator, we must first consider<br />
the selection <strong>of</strong> the initial key. Through the experiment<br />
test and analysis, we can see that the statistical<br />
characteristics <strong>of</strong> the output sequence are affected by the<br />
computing accuracy, equation parameters and initial<br />
values[20]. As the computing precision in the system is<br />
extremely limited and precision is the main reason for<br />
degradation <strong>of</strong> chaotic dynamics, it is not suitable as the<br />
key input in chaotic encryption algorithm. According to<br />
the research <strong>of</strong> chaotic dynamics, when μ is in the<br />
interval <strong>of</strong> [3.5699456, 4], Logistic map is in the chaotic<br />
state and the output sequence is non-periodic and<br />
non-convergent. However, it can be seen from Lyapunov<br />
index curve that the interval is not always in chaos state,<br />
and when μ =4, the map is a full shot in the unit<br />
interval [0, l] that the chaotic sequence has the<br />
characteristic <strong>of</strong> periodicity. Therefore μ can not be the<br />
initial key input <strong>of</strong> chaotic encryption .When the initial<br />
value has a tiny deviation, the orbit will separate with<br />
exponential speed. So it is impossible to have a<br />
long-term prediction on the Behavior <strong>of</strong> the system. Just<br />
as the chaotic system is sensitive with the initial value,<br />
when the chaotic system is assigned with different initial<br />
value, we can get a series different and not related<br />
chaotic sequence. Therefore, we choose the initial value<br />
<strong>of</strong> chaotic systems as the chaotic key input.<br />
© 2011 ACADEMY PUBLISHER<br />
(a) the initial value is 0.118<br />
Figure 3. Simulation output sequences <strong>of</strong> the logistic map<br />
(b) the initial value is 0.1378<br />
(c) the initial value is 0.216<br />
Figure 4. Diagram <strong>of</strong> correlative peak interval<br />
Ⅳ. THE CORRELATIVE PEAK INTERVAL OF OUTPUT<br />
SEQUENTIAL<br />
The correlative peak interval could be got by<br />
calculating the auto-correlation function. When the<br />
6<br />
precision is obtained between 13 and 44, 2× 10 dots are<br />
used to test the output with different values <strong>of</strong> u.<br />
Experiments have confirmed that under the condition <strong>of</strong>
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1053<br />
a fixed computational precision, the parameter u is<br />
changed. If the parameter u is fixed, the correlative peak<br />
interval will change according to the change <strong>of</strong><br />
computational precision. The change <strong>of</strong> the correlative<br />
peak interval is nonlinear and irregular. In practical<br />
applications, the parameter u and the precision are fixed<br />
according to the requirements. An example <strong>of</strong> a cyclic<br />
period curve versus the precision is shown as Figure 5.<br />
The parameter u is fixed which is equal to 3.617.<br />
Figure 5. A curve changing with the precision (u =3.617)<br />
Using the least-squares method for curve fitting can<br />
determine a formula, i.e. a fitting model, f(x). For the<br />
original data shown in Figure 5, for instance, one may<br />
construct an exponential fitting model, a piecewise<br />
fitting model, or a polynomial fitting model. For the first<br />
model, the iterated value <strong>of</strong> the exponent diverges; for<br />
the second model, piecewise fitting is <strong>of</strong>ten inconvenient<br />
to use as a mathematical model (for example, if the test<br />
data are divided into two parts then the simulation results<br />
are as shown in Figure 6, therefore, the polynomial<br />
fitting model is chosen in this investigation.<br />
© 2011 ACADEMY PUBLISHER<br />
(a) Interval function on [13,35]<br />
(b) Interval function on [36,41]<br />
Figure 6. Curves obtained by piecewise fitting.<br />
Linear, quadratic, cubic and quartic polynomial fitting<br />
have been performed and compared, which is shown<br />
as Figure 7.<br />
(a) linear fitting<br />
(b) quadratic fitting
1054 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Figure 7.<br />
(c) cubic fitting<br />
(d) quartic fitting<br />
Polynomial fitting models<br />
It is clear that the quartic fitting gets the best result in<br />
approximating the test data. Therefore the quartic fitting<br />
model is chosen as:<br />
f ( x) a bx cx dx ex<br />
2 3 4<br />
= + + + + (6)<br />
All the parameters <strong>of</strong> that equation can be ensured by<br />
the least square method.<br />
ϕ(<br />
a , a , a , a , a ) = ( y − y )<br />
0 1 2 3 4<br />
N<br />
∑<br />
i=<br />
1<br />
N<br />
∑<br />
i=<br />
1<br />
* 2<br />
i i<br />
= ( y−a−ax−ax −ax−ax i 0 1 1i 2 2i 3 3i 4 4i<br />
To make ϕ(a 0,a 1,a 2,a 3,a 4)<br />
minimal, the partial<br />
derivatives <strong>of</strong> the equation ϕ (a 0,a 1,a 2,a 3,a 4)<br />
to<br />
a, �,a<br />
0<br />
4<br />
should be 0. That is:<br />
© 2011 ACADEMY PUBLISHER<br />
)<br />
2<br />
(7)<br />
N ⎧ ∂ϕ<br />
⎪ =−2 ∑(<br />
yi −a0 −ax 1 1i −ax 2 2i − ax 3 3i - ax 4 4i)<br />
= 0<br />
a0<br />
i=<br />
1<br />
⎪<br />
∂<br />
⎪ N ∂ ϕ<br />
⎪ = −2 ∑(<br />
yi −a0 −ax 1 1i −ax 2 2i − ax 3 3i - ax 4 4i) x1i<br />
= 0<br />
⎨∂a1<br />
i=<br />
1<br />
⎪<br />
��<br />
⎪<br />
N ⎪ ∂ ϕ<br />
⎪ = −2 ∑(<br />
yi −a0 −ax 1 1i −ax 2 2i −ax<br />
3 3i - ax 4 4i) x4i<br />
= 0<br />
⎩∂a4<br />
i=<br />
1<br />
(8)<br />
After the readjustment <strong>of</strong> formula (8), we get the<br />
equation:<br />
N N N N<br />
⎧<br />
⎪Na0<br />
+ ∑x1ia1 + ∑x2ia2 + � + ∑x4ia4 = ∑yi<br />
i= 1 i= 1 i= 1 i= 1<br />
⎪<br />
N N N N N<br />
⎪<br />
⎪∑x1ia1+<br />
∑x1ix1ia1+ ∑x1ix2ia2+ � + ∑x1ix4ia4= ∑x1iyi<br />
⎨ i= 1 i= 1 i= 1 i= 1 i= 1<br />
⎪��<br />
⎪<br />
N N N N N<br />
⎪<br />
⎪∑x<br />
a + ∑x x a + ∑x x a + � + ∑x x a = ∑x<br />
y<br />
⎩<br />
(9)<br />
Formula (9) is called normal equation, which is linear<br />
equations about a,a, 0 1 �,a4,<br />
and it could be expressed<br />
with matrix as:<br />
4i 1 4i 1i 1 4i 2i 2 4i 4i 4 4i i<br />
i= 1 i= 1 i= 1 i= 1 i= 1<br />
⎡<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
�<br />
⎢<br />
⎢<br />
⎣<br />
� � ��<br />
⎤<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥⎦ ⎢�⎥ ⎢ ⎥<br />
a<br />
⎡<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎢<br />
⎢��<br />
⎢<br />
⎢<br />
⎢⎣<br />
(10)<br />
N<br />
N<br />
∑ x1i N<br />
∑ x2i<br />
N<br />
�∑<br />
x4i N<br />
∑ yi<br />
N<br />
∑ x1i<br />
i= 1<br />
i= 1<br />
N<br />
2<br />
∑ x1i i= 1<br />
i= 1<br />
�<br />
i= 1<br />
N<br />
�∑<br />
x1ix4i i= 1<br />
⎡a0⎤ ⎢ ⎥<br />
a1<br />
⋅ ⎢ ⎥ =<br />
i= 1<br />
N<br />
∑ x1iyi i= 1<br />
N<br />
∑ x 4i<br />
i= 1<br />
N<br />
∑ x4ix1i �<br />
N<br />
2<br />
�∑<br />
x4i ⎣ 4 ⎦ N<br />
∑ x4iyi<br />
i= 1 i= 1 i= 1<br />
It can be certified that the coefficient matrix <strong>of</strong><br />
equations (10) is a symmetric positive definite matrix, so<br />
it exists a unique solution. According to the test data,<br />
Matlab program is used to derive the coefficients. The<br />
approximate formula is<br />
2 3<br />
4<br />
f ( x) = 558900 − 103800x+ 69280x − 1977x + 20.41x<br />
(11)<br />
Generally, the accurate <strong>of</strong> the fitting depends on the<br />
polynomial degree. But if the polynomial degree is too<br />
high, it requires increasing calculations and causes<br />
severer oscillations at the two ends <strong>of</strong> the resultant curve.<br />
The parameter u <strong>of</strong> correlative peak interval formula<br />
chosen to the test data is equal to 3.617. To verify its<br />
generality, 8 different values <strong>of</strong> u have been tested and<br />
analyzed. The precision-u-correlative peak interval<br />
diagram is shown as Figure 8.<br />
From this figure, we can see that if the parameter u is<br />
in the chaotic regions and the computational accuracy is<br />
higher than 35 the logistic map has a long-correlative<br />
peak interval. Therefore a good parameter region can<br />
obtain long-correlative peak interval, which is desirable<br />
⎤<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥<br />
⎥⎦
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1055<br />
by many cryptographic applications.<br />
Figure 8. Diagram <strong>of</strong> precision-u-correlative peak interval<br />
The chaotic sequence is non-cycle theoretically, but<br />
under the situation <strong>of</strong> finite precision and approximation<br />
computing, chaotic sequence will become a cycle<br />
sequence. Another more easily observed phenomenon is<br />
that after discretization chaos has strong correlation,<br />
which has a greater threat to the period and will directly<br />
affects the confidential strength.<br />
With the relationship between the correlative peak<br />
interval and the computational precision, we could get<br />
some long periods sequences within the available limited<br />
hardware and computing resources. A suitable parameter<br />
u, precisions and correlative peak interval can effectively<br />
generate pseudorandom output sequences, which is<br />
acceptable by encryption devices and secure<br />
communication systems.<br />
Ⅴ. SUMMARY<br />
This paper has implemented and analyzed a new<br />
design <strong>of</strong> a logistic-map module by calculating its<br />
Lyapunov exponent to determine the suitable regions <strong>of</strong><br />
the chaotic parameter u. With the respective tests using<br />
different initial values, the values <strong>of</strong> parameter u and the<br />
computational precisions, extensive experiments have<br />
been carried out. The correlative peak interval has also<br />
been tested and analyzed with a fixed initial value,<br />
alterative parameter u, and different precisions. An<br />
approximate formula for calculating the correlative peak<br />
interval is provided. Some conditions for correlative<br />
peak interval affected by the chaotic sequences are also<br />
derived. Moreover, through fitting the relation between<br />
the correlative peak interval and the computational<br />
precisions, it is possible to define the best regions for<br />
both correlative peak interval and precisions in the<br />
design <strong>of</strong> key sequence generators based on FPGA.<br />
Finally, a diagram that presents the relationship among<br />
precision, parameter u and correlative peak interval has<br />
been generated for analysis. The basic theory and<br />
methods have a significant implication on the statistical<br />
analysis and practical applications <strong>of</strong> the digital chaotic<br />
sequences.<br />
© 2011 ACADEMY PUBLISHER<br />
ACKNOWLEDGMENT<br />
This work is supported by the National Natural<br />
Science Foundation <strong>of</strong> China (no. 60672011).<br />
REFERENCES<br />
[1] F. Dachselt and W. Schwarz. Chaos and cryptography.<br />
IEEE Transactions on Circuits and Systems – Part I, 2001,<br />
48(12): 1498-1509.<br />
[2] M. S. Baptista. Cryptography with chaos. Physics Letters<br />
A. 1998, 240: 50-54.<br />
[3] Wei Xiang and Fangqi Chen, Sliding Mode Control<br />
Strategies for the Hyperchaotic MCK System, ICIC<br />
Express Letters, vol.3, no.3(A), pp.283-288, 2009.<br />
[4] Yang Yang, Xiangzhong Bai, Zhenguo Tian and Huan<br />
Wang, Chaotic Motion in Fluid-solid Interaction<br />
Problem <strong>of</strong> the Elastic Cylinder, ICIC Express Letters,<br />
vol.3, no.3 (A), pp.439-444, 2009.<br />
[5] Junyan Yi, Gang Yang, Shangce Gao and Zheng Tang,<br />
Transiently Chaotic Neural Network Based on Switched<br />
Cooling and Its Application to Maximum Clique<br />
Problem, International <strong>Journal</strong> <strong>of</strong> Innovative Computing,<br />
Information and Control, vol.5, no.6, pp.1569-1586,<br />
2009.<br />
[6] C. Y. Chee and D. L. Xu. Chaotic encryption using<br />
discrete-time synchronous chaos. Physics Letters A.<br />
2006, 348(3-6): 284-292.<br />
[7] Q. Ding, Y. Zu, F. Zang, and X. Peng. Discrete chaotic<br />
circuit and the property analysis <strong>of</strong> output sequence.<br />
International Symposium on Communications and<br />
Information Technologies. Beijing, China, Oct. 12-14,<br />
2005, 2: 1009-1012.<br />
[8] T. Gao and Z. Chen. Image encryption based on a new<br />
total shuffling algorithm. Chaos, Solitons and Fractals.<br />
2008, 38: 213-220.<br />
[9] S. Pan, J. Huang and G. Wang. Modern DSP Technology.<br />
Xi’an Electronic and Science University Press, Xi’an,<br />
China. 2003, 57-91.<br />
[10] W. Zheng. Random Signal Analysis. Harbin Industrial<br />
Univ ersity Press, Harbin, China, 1999, 66-76.<br />
[11] K. Li, Y. C. Soh and C. Zhang. A frequency aliasing<br />
approach to chaos-based cryptosystems. IEEE<br />
Transactions on Circuits and Systems – Part I. 2004,<br />
51(12): 2470-2475.<br />
[12] H. Zhou and X. Ling. Sequence m-perturbation<br />
implement <strong>of</strong> finite precision chaotic system. Electronic<br />
<strong>Journal</strong>s. 1997, 25(7):95-97.<br />
[13] H. Zhou, J. Yu and X. Ling. The design <strong>of</strong> chaotic feed<br />
forward type stream cipher. Electronic <strong>Journal</strong>s. 1998,<br />
26(1): 98-101.<br />
[14] R. Huang. Chaos and its Applications. Wuhan University<br />
Press, Wuhan, China. 2003, 128-138.<br />
[15] S. Pan, J. Huang and G. Wang. Modern DSP<br />
Technology. Xi’an Electronic and Science University<br />
Press, Xi’an, China. 2003, 57-91<br />
[16] W. Zheng. Random Signal Analysis. Harbin Industrial<br />
University Press, Harbin, China, 1999, 66-76.<br />
[17] Q. Ding, J. Pang, J. Fang, and X. Peng. Designing <strong>of</strong><br />
chaotic system output sequence circuit based on FPGA<br />
and its possible applications in network encryption cards.<br />
International <strong>Journal</strong> <strong>of</strong> Innovative Computing,<br />
Information and Control, 2007, 3(2): 449-456.<br />
[18] F. Belkhouche, U. Qidwai, I. Gokcen and D. Joachim.<br />
Binary image transformation using two-dimensional<br />
chaotic maps. IEEE International Conference on Pattern<br />
Recognition. Cambridge, UK, 2004, 4: 823-826.<br />
[19] C. Y. Chee and D. L. Xu. Chaotic encryption using<br />
discrete-time synchronous chaos. Physics Letters A.
1056 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
2006, 348(3-6): 284-292.<br />
[20] M. I. Sobhy, A-E. R. Shehata. Chaotic algorithms for<br />
data encryption. IEEE International Conference on<br />
Acoustics, Speech, and Signal Processing. Salt Lake<br />
City, UT, USA. 2001, 2: 997-1000.<br />
Qun Ding, born at Harbin, Heilongjiang<br />
Province in 1957. She got instrument<br />
and technology science doctor’s degree<br />
at Harbin Institute <strong>of</strong> Technology in<br />
2007. Now she is the dean <strong>of</strong> electronic<br />
engineering college, doctoral director,<br />
the director <strong>of</strong> Heilongjiang electronic<br />
engineering key laboratory, the director<br />
<strong>of</strong> Heilongjiang signal and information<br />
key laboratory, the councilman <strong>of</strong><br />
Heilongjiang communication institute, the panel judge <strong>of</strong><br />
national 863 programs and national nature science fund. Her<br />
major research field is the security <strong>of</strong> information and<br />
encryption communication.<br />
© 2011 ACADEMY PUBLISHER<br />
Lu Wang, born at Harbin, Heilongjiang<br />
Province in 1984. She got her bachelor’s<br />
degree in Heilongjiang University at<br />
Harbin, Heilongjiang Province, China in<br />
2008. The major <strong>of</strong> the undergraduate is<br />
communication engineering. Now she is<br />
at the graduate stage at Heilongjiang<br />
University at Harbin, Heilongjiang<br />
Province, China. The research field is<br />
the encryption <strong>of</strong> communication. And<br />
the major is communication engineering.<br />
Guanrong Chen,got applied<br />
mathematics doctor’ degree at Texas<br />
A&M University in 1987. Now he is the<br />
chair pr<strong>of</strong>essor <strong>of</strong> Hong Kong City<br />
University and the Changjiang chair<br />
pr<strong>of</strong>essor <strong>of</strong> Peking University. His<br />
main research field is nonlinear system<br />
<strong>of</strong> control theory, dynamics analysis<br />
research and its application in<br />
complicated network.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1057<br />
An Energy Efficient Dynamic Clustering Protocol<br />
based on Weight in Wireless Sensor <strong>Networks</strong><br />
Ming Zhang 1,2<br />
1 Department <strong>of</strong> S<strong>of</strong>tware,Nanjing University <strong>of</strong> posts & Telecommunications Nanjing,China.<br />
Email: lyg690916@163.com<br />
Suoping Wang 1<br />
2 Department <strong>of</strong> Electronic Engineering, Huaihai Institute <strong>of</strong> Technology, Lian yungang, China<br />
Email: wangsp@njupt.edu.cn<br />
Abstract—Nodes in most wireless sensor networks (WSNs)<br />
are powered by batteries with limited energy. Prolonging<br />
network lifetime and saving energy are two critical issues<br />
for WSNs. Clustering is an effective technique to improve<br />
the energy efficiency and prolong network lifetime <strong>of</strong><br />
wireless sensor networks. In this paper, an energy efficient<br />
dynamic clustering protocol (EEDCP) based on weight for<br />
wireless sensor networks is proposed, which is able to<br />
dramatically prolong network lifetime and save energy. In<br />
the EEDCP, we introduce the typical energy model to<br />
compute energy consumption, virtual grid technology to<br />
construct cluster and a long sleeping state to reduce energy<br />
consumption. In addition, we use the value <strong>of</strong> weight to<br />
measure the size <strong>of</strong> residual energy instead <strong>of</strong> voting, which<br />
can significant reduce the voting times and the number <strong>of</strong><br />
transmitting information. Further, simulation experiments<br />
are conducted to compare the EEDCP with some wellknown<br />
clustering algorithms and simulation results show<br />
that the proposed method overcomes the existing methods in<br />
the aspects <strong>of</strong> energy consumption and network lifetime in<br />
wireless sensor networks.<br />
Index Terms—Wireless sensor networks (WSNs), Energy<br />
efficient, Network lifetime, Dynamic clustering, Weight<br />
I. INTRODUCTION<br />
Wireless sensor networks (WSNs) have been<br />
blooming recently, which are being widely used in<br />
various areas such as reconnaissance, disaster relief,<br />
intelligent transportation, surveillance, environmental<br />
monitoring, healthcare, target tracking, and more. WSNs<br />
are extremely useful to collect information in harsh or<br />
hostile environment. A WSN has two important and<br />
interesting characteristics that are different from<br />
traditional wireless networks [1]. First, after the event<br />
occurs, multiple sensors nodes (denoted as data source<br />
nodes) around this event will sense the event, and then<br />
send the data back to one sensor node (denoted as sink<br />
node). Hence, communication mode in WSN occurs from<br />
multiple data source nodes to one data sink node. This is<br />
a type <strong>of</strong> multipoint-to-point, rather than the traditional<br />
Corresponding author:Ming Zhang,E-mail:lyg690916@163.com<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1057-1064<br />
point-to-multipoint communication in WSNs[2] .<br />
A wireless sensor network (WSN) is composed <strong>of</strong> a<br />
large number <strong>of</strong> sensor nodes that are densely deployed<br />
near an area <strong>of</strong> interest and are connected by a wireless<br />
interface. Since each sensor is limited in terms <strong>of</strong><br />
processing capability, wireless bandwidth, battery power<br />
and storage space, in most applications, it is impossible to<br />
replenish power resources, a major constraint <strong>of</strong> WSN<br />
lifetime is energy consumption. Energy savings<br />
optimization is thus a major challenge for the success <strong>of</strong><br />
WSNs. Typical tasks <strong>of</strong> a sensor node in a sensor network<br />
are to collect data, perform data aggregation, and then<br />
transmit data. Among these tasks, monitoring and<br />
transmitting data require much more energy than<br />
processing it [3]. Therefore, in wireless sensor network, a<br />
significant focus has been put on increasing energy<br />
efficiency [4]. Generally, there are two basic approaches<br />
to the problem <strong>of</strong> saving energy in WSN. The first one is<br />
scheduling some sensor nodes to go into an active mode<br />
while enabling the other sensor nodes to go into a lowpower<br />
sleep mode [5,6]. The second approach is to select<br />
the optimization routing algorithm, eliminating redundant<br />
energy consumption.<br />
Hence, proper energy efficient dynamic routing<br />
protocols should be designed to increase the lifetime <strong>of</strong><br />
the network greatly. In this paper, an energy efficient<br />
dynamic clustering protocol (EEDCP) based on weight is<br />
proposed for wireless micro sensor networks to facilitate<br />
the achievement <strong>of</strong> low energy dissipation. From the<br />
simulation results, it is illustrated that the EEDCP<br />
achieves an order <strong>of</strong> magnitude increase in system<br />
lifetime when compared to the general – purpose<br />
approaches. Moreover, for a given quality, the overall<br />
residual is reduced by an order <strong>of</strong> magnitude.<br />
The rest <strong>of</strong> this paper is organized as follows. Section<br />
II gives the detailed related work done. Section III<br />
presents the system model for our architecture, such as<br />
network model, energy model and node state transition<br />
model. Section IV gives the detailed description <strong>of</strong> the<br />
algorithm and structure. Section V gives the experimental<br />
results and section Ⅵconcludes the paper.<br />
II. RELATED WORK
1058 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
There are many energy-saving routing algorithms used<br />
in wireless sensor networks, and new ideas for routing are<br />
announced in recent years. In this section, we review<br />
some <strong>of</strong> the most effective algorithm.<br />
In low-energy adaptive clustering hierarchy LEACH)<br />
[7], the authors discuss an energy efficient algorithm.<br />
Various algorithms developed after that is based on this<br />
algorithm. In order to determine the cluster head, LEACH<br />
uses randomization technique Gossiping [8] is the<br />
improvements <strong>of</strong> flooding algorithm (Flooding), it can<br />
effective resolve the implosion and information overload<br />
problem which lead to energy loss, but it can not solve<br />
some <strong>of</strong> the data overlap and too much delay, but it can<br />
not balance the energy <strong>of</strong> nodes. Hybrid energy-efficient<br />
distributed clustering (HEED)[19] is based on LEACH<br />
thinking, the important difference is the choice <strong>of</strong> cluster<br />
head and cluster head formation. In PEGASIS (Power-<br />
Efficient Gathering in Sensor Information Systems),<br />
author tried to foster the past technique [10]. This new<br />
mechanism is a chain-based power efficient protocol<br />
based on LEACH [11]. It assumes that each node must<br />
know location information about all other nodes at first.<br />
PEGASIS starts with the farthest node from the base<br />
station. The chain can be constructed easily by using a<br />
greedy algorithm. The chain leader aggregates data and<br />
forwards it to the base station. In order to balance the<br />
overhead involved in communication between the chain<br />
leader and the base station, each node in the chain takes<br />
turn to be the leader.<br />
A clustering-based routing protocol called base station<br />
controlled dynamic clustering protocol (BCDCP)[12],<br />
which utilizes a high energy base station to set up cluster<br />
heads and perform other energy-intensive tasks, can<br />
noticeably enhance the lifetime <strong>of</strong> a network. United<br />
voting dynamic cluster routing algorithm based on<br />
residual-energy in wireless sensor networks (UVDC)[13],<br />
which periodically selected cluster head according to<br />
residual energy among the nodes located in the event area,<br />
so the voting cost <strong>of</strong> UVDC is gigantic and the large<br />
redundant nodes will waste limited energy. Sensor<br />
protocols for information via negotiation (SPIN)[14] is<br />
also improved the flooding algorithm, before transferring<br />
data, it only transmit data to needed neighbor nodes<br />
which using meta-data to reduce redundant information to<br />
save energy consumption. Directed diffusion (DD)[11]<br />
periodic automatic forms the enhanced path, because <strong>of</strong><br />
node energy and topology changes, the enhanced path<br />
will be different in different period, the most <strong>of</strong> data from<br />
the source to the cluster is transmitted by the enhanced<br />
path, thus reduce the energy consumption <strong>of</strong> nonenhanced<br />
nodes. And GBR (Gradient-Based Routing) is<br />
also proposed as a variant <strong>of</strong> directed diffusion [15]. The<br />
key idea in GBR is to memorize the number <strong>of</strong> hops<br />
when the interest in diffused through the whole network.<br />
In GBR, three different data dissemination techniques<br />
have been discussed [11] (i) Stochastic Scheme, where a<br />
node picks one gradient at random when there are two or<br />
more next hops that have the same gradient, (ii) Energybased<br />
scheme, where a node increases its height when its<br />
energy drops below a certain threshold, so that other<br />
© 2011 ACADEMY PUBLISHER<br />
sensors are discouraged from sending data to that node,<br />
and (iii) Stream-based scheme, where new streams are<br />
not routed through nodes that are currently part <strong>of</strong> the<br />
path <strong>of</strong> other streams. The main objective <strong>of</strong> these<br />
schemes is to obtain a balanced distribution <strong>of</strong> the traffic<br />
in the network, thus increasing the network lifetime.<br />
Threshold sensitive energy efficient sensor network<br />
protocol (TEEN)[16] is designed for responsive<br />
applications, it determine whether to send data by setting<br />
up a reasonable s<strong>of</strong>t and hard threshold to compare with<br />
the monitoring data .It only transmit the interest<br />
information to users to effectively reduces the network<br />
traffic and thus reduce network energy consumption.<br />
Guojun Wang, et al., have proposed a local updatebased<br />
routing protocol in WSNs with a mobile sink [17].<br />
The protocol proposed by the authors saves the energy for<br />
sensor networks and makes the sink keep continuous<br />
communications to sensors by confining the destination<br />
area into a local area for updating the sink location<br />
information as the sink moves. Hayoung Oh and Kijoon<br />
Chae have presented a sensor routing scheme [18], EESR<br />
(Energy-Efficient Sensor Routing) that provides energyefficient<br />
data delivery from sensors to the base station.<br />
Their scheme divides the area into sectors and locates a<br />
manager node to each sector.<br />
Besides these algorithms mentioned above, there exist<br />
several other algorithms [19], such as: Soro et al. [20]<br />
proposed an unequal clustering size model for network<br />
organization, which can lead to more uniform energy<br />
dissipation among cluster head nodes, thus increasing<br />
network lifetime. Ye et al. [21] proposed a clustering<br />
algorithm, which achieves good cluster head distribution<br />
with no iteration and introduces a weighted function for<br />
the plain node to make a decision for joining a proper<br />
cluster.<br />
Ⅲ.SYSTEM MODEL<br />
A. Network model<br />
In this paper, we consider the wireless sensor networks<br />
where N nodes in field A are homogenous and energy<br />
constrained and the sensor network has the following<br />
properties [22]:<br />
(1) This network is a static densely deployed network.<br />
It means a large number <strong>of</strong> sensor nodes are densely<br />
deployed in a two-dimensional geographic space, forming<br />
a network and these nodes do not move any more after<br />
deployment.<br />
(2) There exists only one Sink node, which is deployed<br />
at a fixed place outside the WSNS.<br />
(3) The energy <strong>of</strong> sensor nodes cannot be recharged. It<br />
means sensor node will die if its energy be exhausted.<br />
(4) Sensor nodes are location-aware, i.e. sensor node<br />
can get its location information through other<br />
mechanisms such as GPS or position algorithms (in order<br />
to describe the position <strong>of</strong> node uses (Xi,Yj) ,the Sink<br />
node as (Xsink,Ysink)).<br />
(5) The radio power can be controlled, i.e., a node can<br />
vary its transmission power depending on the distance to<br />
the receiver [23].
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1059<br />
In table 1,we give the symbols definition in this paper.<br />
TABLE 1.<br />
Symbols<br />
NOTATION<br />
Definition<br />
N The set <strong>of</strong> sensor nodes<br />
x,y Coordinates <strong>of</strong> the node<br />
di,j The distance <strong>of</strong> node i and node j<br />
Etx The consumption energy <strong>of</strong> node I transmitted a<br />
packet.<br />
Erx The consumption energy <strong>of</strong> node I receive a<br />
packet<br />
Er(i) The residual energy <strong>of</strong> node i<br />
Ep The consumption <strong>of</strong> processing in cluster head<br />
Ti Listening timer<br />
Ts Sleeping timer<br />
Tj Sensing timer<br />
r(i) The number <strong>of</strong> active node I receives voting<br />
information<br />
s(i) The number <strong>of</strong> active node I sends voting<br />
information<br />
v(t) The total number <strong>of</strong> sending and receiving<br />
voting information in head<br />
Gi The ith virtual grid in a cluster<br />
k The size <strong>of</strong> a packet<br />
Weight The value <strong>of</strong> residual energy divided by Er(i)for<br />
each active node<br />
M,N The number <strong>of</strong> virtual grid in a cluster<br />
First, we use the virtual grid ideas to divide the field A<br />
into many same square, namely, there are many clusters,<br />
and each node can directly communicate with other nodes<br />
in a cluster. Then the cluster was divided into M×N small<br />
area (the value <strong>of</strong> M,N are determined by the cluster’s<br />
size, assume that there are M×N grid in a cluster, each<br />
grid is named Gk(k=1.. M×N).<br />
Figure 1. Virtual grid model<br />
Fig.1 shows the virtual grid ideas, in order to<br />
conveniently describe, we suppose that the value <strong>of</strong> M<br />
and N are equal to three, each small square call as a<br />
virtual grid, nodes are randomly distributed into this<br />
virtual grid, such as CH1 has 9 virtual grid, namely, G1,<br />
G2, G3,G4,G5,G6,G7,G8 and G9,we suppose virtual grid<br />
G5 as a cluster head grid, and the red pentagram as the<br />
cluster head., for arbitrary adjacent virtual grid G1 and<br />
G2,each node in G1 can communicate with all nodes in<br />
© 2011 ACADEMY PUBLISHER<br />
G2, and vice versa. In a cluster, the red dot as the<br />
working node in each grid and each node can<br />
communicate with cluster head, and we suppose the<br />
number <strong>of</strong> simultaneous working node is one in a virtual<br />
grid (red dot), other nodes are sleeping (black dot). In<br />
order to guarantee the network normal working and<br />
prolong network lifetime, one sleeping node in a virtual<br />
grid will be awaken at the right time so as to instead <strong>of</strong><br />
the energy-exhausted node or disabled node [24].<br />
B. Energy model<br />
We adopt a simplified power model <strong>of</strong> radio communication<br />
in document [25], namely, in order to send a k-bit<br />
packet information and the sending distance is di,j, the<br />
sending energy consumption is<br />
ETx( k, d) = Eelec × k + ε amp × k× d× d<br />
(1)<br />
The distance <strong>of</strong> node I and node j is di,j:<br />
2<br />
2<br />
| d i, j | = ( xi − xj) + ( yi − yj)<br />
The receiving energy consumption is<br />
(2)<br />
ERx( k) = Eelec × k<br />
(3)<br />
Where Eelec is the energy/bit consumed by the sender<br />
amp<br />
and receiver electronics, J/bit, Eelec=50nJj/bit,. ε is<br />
amp<br />
the J/(bit × m2), ε =100pJ/bit/m2.we commonly<br />
assume that the sending distance and d2 is directly<br />
proportional for shorter distance, while the sending<br />
distance and d4 is directly proportional for longer<br />
distance, so we can see the directly sending to long<br />
distance is consumed more energy than multi-hop<br />
sending.<br />
But the differentiation from the document [26], we<br />
consider the processing consumption in order to<br />
proximity real scene, the energy consumption <strong>of</strong> cluster<br />
head is Ep<br />
m<br />
E ( k , m ) = 1 / 3 × E × k<br />
∑<br />
P elec i<br />
i = 1<br />
So the residual energy <strong>of</strong> cluster head is:<br />
n1<br />
n2<br />
Er() i = Er() i − ∑ ETx( kn, dn) −∑ERx( kl) −Ep(<br />
k, m), n1, n2∈N n = 1<br />
l=<br />
1<br />
(4)<br />
(5)<br />
Where n1,n2 are the cluster head respectively sending<br />
and receiving times before time Ti.<br />
The residual energy <strong>of</strong> ordinary node is:<br />
Er() i = Er() i −ETx( kn, dn) −ERx(<br />
kl)<br />
(6)<br />
C. Node state transition model<br />
The energy dissipation in wireless sensor networks has<br />
three models: sensor model; procession model; wireless<br />
radio model [27]. In order to maximum lifetime and<br />
minimum routing, nodes in the EE-MLMR have various<br />
operation modes with different levels <strong>of</strong> activation and,<br />
thus, different levels <strong>of</strong> energy consumption. We put<br />
forward the new state conversion model which have flag
1060 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
<strong>of</strong> valve which depend on the EPGR state model. In this<br />
model, each node has six operation modes [28]: mode 1:<br />
sleeping-sensing <strong>of</strong>f and radio <strong>of</strong>f; mode 2: sensing -<br />
sensing on and radio <strong>of</strong>f; mode 3: receiving sensing on<br />
and radio receiving; mode 4: transmitting -sensing on and<br />
radio transmitting; mode 5: listening - sensing on; mode<br />
6:long sleeping- sensing <strong>of</strong>f and radio <strong>of</strong>f forever, no<br />
responding.<br />
Figure 2. The transition model os sensor nodes<br />
Where Ts is a sleeping timer, Ti is listening timer, Tj is<br />
sensing timer.<br />
The Fig. 2 shows the ‘‘commands’’ performed along<br />
the path (state transition) between states [29]. It means<br />
that whenever a node changes its state-based energy<br />
dissipation model to current state it performs tests and<br />
actions until the new state is reached. “sleeping’’<br />
determines whether the node will sleep or not; The<br />
‘‘receiving’’ test depends also on the characteristics <strong>of</strong><br />
the event. Its value is influenced by the degree <strong>of</strong><br />
cooperation needed by the application. The ‘‘sensing’’<br />
test is called only if there is no event in the area <strong>of</strong> the<br />
node. If no event happens, this test will depend on the<br />
degree <strong>of</strong> coverage needed by the application. In the<br />
“transmitting” state, if flag=0,then sensing <strong>of</strong>f and radio<br />
<strong>of</strong>f and node convert to long sleeping; if flag=1,then<br />
convert to the transmitting. The “listening” test<br />
determines whether a new sensing event is present; the<br />
long sleeping denotes the node never respond any event.<br />
“Timer’’ is an action that starts a timer. The outcome <strong>of</strong><br />
each test depends on a probabilistic parameter associated<br />
with the test. These transitions try to capture the behavior<br />
<strong>of</strong> a sensor node, specially in terms <strong>of</strong> energy<br />
consumption.<br />
Ⅳ.ENERGY EFIICIENT DYNAMIC CLUSTERING<br />
PROTOCOL BASED ON WEIGHT<br />
At any time, only one node within a virtual grid stays<br />
active to be a coordinator, while the others fall into<br />
sleeping mode. Doing this significantly reduces the<br />
energy consumption because nodes in the idle state spend<br />
much more energy as compared with the sleeping state.<br />
In our protocol, we use weight as the selection criteria<br />
for new cluster head, when the residual energy <strong>of</strong> the<br />
cluster head is lower than threshold, the EEDCP will be<br />
implemented. First, it compute the total residual<br />
© 2011 ACADEMY PUBLISHER<br />
energy(Et) <strong>of</strong> all active nodes from the cluster head<br />
member table, then respectively compute the weight <strong>of</strong> all<br />
active nodes as shown in formula (7). Second, we select<br />
the minimal weight active node as the new cluster head<br />
and inform all member nodes and the cluster head<br />
neighbor. All received information nodes will reply<br />
related information and update its table. Lastly, the old<br />
cluster head will select a new active node replace him and<br />
then goto long sleeping.<br />
weight( i)<br />
=<br />
k<br />
∑ (7)<br />
j = 1<br />
E r ( j)<br />
E r ( i )<br />
Where k is the number <strong>of</strong> active nodes in the<br />
cluster,Er(i) is the residual energy <strong>of</strong> active node.<br />
The energy efficient dynamic voting cluster (EEDCP)<br />
based on weight has four steps: Initialization, active node<br />
selection, dynamic clustering based on weight phase,<br />
sensing and sending. When the residual energy <strong>of</strong> cluster<br />
head is lower than threshold, dynamic clustering based on<br />
weight will happens, namely, cluster head will initiate<br />
clustering process from the active node. if new cluster<br />
head is come into being, clustering is formed. The<br />
detailed process described in the followings:<br />
Step1:Initialization: the whole area was divided into<br />
many same squares, namely, there are many clusters, and<br />
each node can directly communicate with other nodes in a<br />
cluster. Then the cluster was divided into M×N small<br />
area (the value <strong>of</strong> M,N are determine by the cluster’s size,<br />
there are M×N grid in a cluster, each grid is named<br />
Gk(k=1.. M × N). The first cluster head is randomly<br />
selected from the active nodes and each active node has a<br />
neighbor table (as shown in table 2.)to record its member<br />
information and all directly connect active node ID and<br />
energy information, then cluster head will send<br />
information to other cluster head, so as to form a cluster<br />
neighbor table(as shown in table 3.).<br />
Step2:Sensing and sending: when there is a event, the<br />
active node will collect event information, then compute<br />
its residual energy, if Er is lower than threshold then goto<br />
active node selection; else it will send event information<br />
and the residual energy to the cluster head.<br />
Step3:Active node selection: for each grid, if its active<br />
node residual energy is lower than threshold, the active<br />
node will select a new node as the new active node from<br />
its neighbor table which has maximum residual energy,<br />
and send the new active node ID and energy information<br />
to cluster head and its directly connect active nodes.<br />
Step4:Dynamic clustering based on weight: Cluster<br />
head is responsible for receiving data, gathering data,<br />
sending data to next top, computing its residual energy<br />
and maintaining its all table. After some time, if the<br />
residual energy <strong>of</strong> cluster head is lower than threshold,<br />
the cluster head will implement the dynamic clustering<br />
process. First, it compute the total residual energy <strong>of</strong> all<br />
active nodes and the weight (is equal to the total residual<br />
energy <strong>of</strong> all active nodes in the cluster divided by the<br />
residual energy <strong>of</strong> each active node.) <strong>of</strong> each active nodes;<br />
Second, old head will use active node selection and select<br />
a new active node and update its member table; Third, old<br />
head will select a minimal weight active node as the new
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1061<br />
cluster, if there are more than one active node has the<br />
same weight, it will randomly select one as the new<br />
cluster head, then, the cluster head will inform its<br />
member nodes and neighbor nodes about the new cluster<br />
head, such as: ID,(xi,yi) etc, all active node in this cluster<br />
and the neighbor nodes received this information will<br />
update its table and record the new cluster head and<br />
replay a ACK to old head; Fourth, the old head will send<br />
the information to the new cluster head, including its<br />
member table and neighbor table. Lastly, the old cluster<br />
head goes to sleep.<br />
TABLE 2.<br />
THE ACTIVE NODE STRUCTURE<br />
Name Propetries<br />
ID Identification <strong>of</strong> nodes<br />
xi,yi Position od node i<br />
Er(i) Residual energy <strong>of</strong> node i<br />
flag 0-cluster head;1-node itself;2-member node;3neighbor<br />
node<br />
state 0-active;1-sleeping<br />
TABLE 3.<br />
THE CLUSTER HEAD STRUCTURE<br />
Name Properties<br />
ID Identification <strong>of</strong> nodes<br />
xi,yi Position od node i<br />
Er(i) Residual energy <strong>of</strong> node i<br />
flag 0-cluster head;1-node itself;2-member node;3neighbor<br />
node<br />
weight The value <strong>of</strong> node residual energy divided by total<br />
residual energy<br />
state 0-active;1-sleeping<br />
For example,at a time, if cluster head (such as G5)<br />
residual energy is lower than threshold in CH1 and other<br />
active node residual energy is show as Fig.3.<br />
Figure 3. The residual energy <strong>of</strong> active nodes in cluster CH1<br />
Fig.4 is the voting relationship and the voting<br />
results .In the Fig.4,s(i) is the times <strong>of</strong> active node i<br />
sending information and r(i) is the times <strong>of</strong> active node i<br />
receiving information, we can see that G2 has the<br />
maximum ballot, namely, r(G2)=5(the residual energy <strong>of</strong><br />
G2 is 0.8 and is the maximum),so the new cluster head is<br />
G2.but in this voting process, the total times <strong>of</strong> voting is<br />
20,namely, all active nodes have sent 20 times voting<br />
information and the receiving voting information times is<br />
20, the total times is 4o, namely v(t)=40,in the voting<br />
© 2011 ACADEMY PUBLISHER<br />
process, each active node has consumed sending energy,<br />
receiving energy and computing energy, so the total<br />
consumption energy is large.<br />
At a time,the active nodes residual energy in the<br />
cluster CH1 as shown in Fig.5,the result is shown in Fig.6<br />
using volting algorithm.From Fig.6,we can see that there<br />
are two active nodes G2 and G8 have the same ballot,<br />
namely, r(G2) and r(G8) are equal to 5 ,in this case,<br />
generally randomly select one <strong>of</strong> the active nodes as the<br />
new cluster head which the maximum residual energy<br />
active node may not be the new cluster head, such as if<br />
we select G2 as the new cluster, we find the maximum<br />
residual energy active node is G8,so it can not ensure the<br />
residual energy active node must be as the new cluster<br />
head which lead to a decline in the network lifetime.<br />
Figure 4. The voting relation and the voting times<br />
Figure 5. The residual energy <strong>of</strong> active nodes in cluster CHi<br />
But using our proposed method, we can use minimal<br />
times to generate the new cluster head, significantly save<br />
energy and ensure the residual energy active node as the<br />
new cluster head, for example, for the same case as<br />
shown in Fig.3,the total computing times is 10 times (one<br />
is compute the total energy, the others are compute the<br />
weight <strong>of</strong> each active node)and each active nodes does<br />
not require any messages to send and receive in this<br />
process, the old head can choosees a new cluster head<br />
G2 as shown in Fig.7,and we can see that the w(2) is the<br />
minimum, namely, G2 is the maximum energy active<br />
node. At the same circumstances in Fig.5,using our<br />
method, we can see from the Fig.8 that the minimal<br />
weight active node(G8) is the only new cluster head , so
1062 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
we can reduce the energy consumption in the new cluster<br />
head generation, always select the maximum active node<br />
as the new head and extend the network lifetime.<br />
Figure 6. The voting relation and the times <strong>of</strong> sending and<br />
receiving for each active node.<br />
Figure 7. The weight <strong>of</strong> each active node using EEDCP<br />
Figure 8. The weight <strong>of</strong> each active node using EEDCP<br />
© 2011 ACADEMY PUBLISHER<br />
Ⅴ.SIMULATION RESULTS<br />
A. Simulation Parameters:<br />
We have implemented our proposed protocol in NS-2<br />
(ver. 2.31). We considered a 600 node random network<br />
deployed in an area <strong>of</strong> 360 X 360 m. Initially the nodes<br />
are placed randomly in the specified area. The only Sink<br />
node is assumed to be situated 100 meters away from the<br />
above specified area. At the same time, we considered<br />
specified area is divided into 90 X 90 m square area<br />
called cluster and each cluster is divided into 30 X 30 m<br />
area called virtual grid. Obviously, the first set <strong>of</strong> cluster<br />
heads are taken randomly. The initial energy <strong>of</strong> all the<br />
nodes assumed as 5 joules. The radio range is varies from<br />
30m to 120m.Each data packet has 64 bytes, and the<br />
others are 36 bytes long. Summary <strong>of</strong> parameters and<br />
defined values are shown in Table 4.<br />
TABLE 4.<br />
SIMULATION PARAMETERS AND VALUES<br />
Simulation parameters value<br />
N(total nodes) 600nodes<br />
A(network size) 360×360 m<br />
Cluster size 90×90 m<br />
Virtual grid size 30×30 m<br />
Number <strong>of</strong> sink 1<br />
Eelec 50nJ/bit<br />
ε<br />
amp<br />
0.0013pJ/bit/m 2<br />
Data packet size 64 bytes<br />
Other packet size 32 bytes<br />
Simulation times 150 seconds<br />
Threshold energy 0.2w<br />
E(i i i l ) 5J l<br />
B. Experimental results and analysis<br />
From the diagram <strong>of</strong> Fig.9, we can see that there are<br />
considerable differences on the average energy<br />
consumption among the three algorithms. EEDCP has the<br />
minimum energy consumption, and with the increase <strong>of</strong><br />
number nodes in each cluster, the consumption is slowly<br />
increase, For UVDC, because <strong>of</strong> there are large redundant<br />
nodes and the consumption <strong>of</strong> voting is gigantic, so the<br />
average energy consumption is rapidly increase. For<br />
LEACH, because <strong>of</strong> randomly rotating the role <strong>of</strong> a<br />
cluster head among all the nodes, the average energy<br />
consumption is approximate linear increase.<br />
.<br />
Figure 9. The average energy consumption <strong>of</strong> LEACH, UVDC<br />
and EEDCP
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1063<br />
In this second series <strong>of</strong> experiments, we compare the<br />
three energy efficient clustering algorithm LEACH,<br />
UVDC and EEDCP with regard to the network lifetime,<br />
when the number <strong>of</strong> nodes in a cluster from 15 to 40. The<br />
LEACH algorithm that does never take energy into<br />
account and always randomly rotating the role <strong>of</strong> a<br />
cluster head among all the nodes. Simulation results are<br />
illustrated in Figure 10, assuming that the initial energy<br />
<strong>of</strong> the nodes is uniformly as 5J<br />
Figure 10. Comparison <strong>of</strong> network lifetime with LEACH, UVDC<br />
and EEDCP<br />
As expected, LEACH provides the smallest network<br />
lifetime. This shows that the random selection <strong>of</strong> the<br />
cluster head is not sufficient to save energy. UVDC<br />
provides better results than LEACH, but in the voting<br />
cluster head process, UVDC consumed large energy in<br />
voting and sending information. The main conclusion <strong>of</strong><br />
these experiments is that EEDCP significantly<br />
outperforms LEACH and UVDC whatever the number <strong>of</strong><br />
nodes in a cluster. Moreover, EEDCP prolongs the<br />
network lifetime <strong>of</strong> 21% compared with LEACH for a<br />
different number <strong>of</strong> nodes in a cluster. Notice that in the<br />
same conditions, UVDC prolongs the network lifetime <strong>of</strong><br />
only 6%.<br />
Ⅵ.CONCLUSIONS In WSNs, it is significant to prolong network lifetime<br />
so that more data can be collected by the sensor(s) to<br />
transmit to sink node. It is well known that, efficiently<br />
use <strong>of</strong> energy is critical for network lifetime. Although<br />
some routing algorithms like voting dynamic cluster<br />
routing algorithm based on residual-energy (UVDC) can<br />
dynamic clustering, they usually place too heavy burden<br />
<strong>of</strong> voting information in cluster which consumed large<br />
valuable energy.<br />
In this paper, an energy efficient dynamic clustering<br />
protocol (EEDCP) based on weight for wireless sensor<br />
networks is proposed, which is bale to dramatically<br />
prolong network lifetime and save energy. In the EEDCP,<br />
we introduce the typical energy model to compute energy<br />
consume, virtual grid technology to construct the cluster<br />
and a long sleeping state to reduce energy consumption.<br />
In addition, we use the value <strong>of</strong> weight to measure the<br />
size <strong>of</strong> residual energy instead <strong>of</strong> voting ,which can<br />
significant reduce the voting times and the number <strong>of</strong><br />
transmitting information. Further, simulation experiments<br />
are conducted to compare the EEDCP with some well-<br />
© 2011 ACADEMY PUBLISHER<br />
known clustering algorithms and simulation results show<br />
that the proposed methods overcomes the existing<br />
methods in the aspects <strong>of</strong> energy consumption and<br />
network lifetime in wireless sensor networks.<br />
ACKNOWLEDGMENT<br />
We acknowledge the support <strong>of</strong> University Natural<br />
Science Foundation <strong>of</strong> Jiangsu Province (No.<br />
08KJD520003),we sincerely thank the anonymous<br />
reviewers for their constructive comments and<br />
suggestions.<br />
REFERENCES<br />
[1] Frank Yeong-Sung Lin , Hong-Hsu Yen and Shu-Ping Lin,<br />
A Novel Energy-Efficient MAC Aware Data Aggregation<br />
Routing in Wireless Sensor <strong>Networks</strong>, Sensors,2009(9),<br />
1518-1533.<br />
[2] Krishnamachari, B.; Estrin, D.; Wicker, S. Modeling Data-<br />
Centric Routing in Wireless Sensor <strong>Networks</strong>. USC<br />
Computer Engineering Technical Report CENG 02-14,<br />
2002.<br />
[3] Pottie, G.J.; Kaiser, W.J. Wireless integrated network<br />
sensors. Commun. ACM 2000, 43, 51-58.<br />
[4] Qin Wang, WoodwardYang, “Energy Consumption Model<br />
for Power Management in Wireless Sensor <strong>Networks</strong>”.<br />
IEEE Communications Society,2007.pp:142-151.<br />
[5] Cardei, M.; Du, D.Z. Improving wireless sensor network<br />
lifetime through power aware organization. Wirel. Netw.<br />
2005, 11, 333-340.<br />
[6] Carle, J.; Simplot, D. Energy-efficient area monitoring by<br />
sensor networks. IEEE Comput. 2004, 37, 40-46.<br />
[7] Heinzelman W, Chandrakasan A, Balakrishnan H. Energy<br />
efficient communication protocol for wireless microsensor<br />
networks Proceedings <strong>of</strong> the 33rd Annual Hawaill<br />
International Conferenc On System Sciences, Jan 4-7,<br />
2000, Maui, HI, USA. Los Alamitos CA, USA: IEEE<br />
Computer Society, 2000: 223.<br />
[8] Santi P , Simon J . Silence is Golden with Hi}gh<br />
Probability : Maintaining a Connected Backbone in<br />
Wireless Sensor <strong>Networks</strong>.In Proceeding <strong>of</strong> 1st European<br />
Workshop on WirelessSensor <strong>Networks</strong>(EWSN 2004),<br />
Jan 2004,106—121.<br />
[9] Sungju Lee, Jangsoo Lee , Hongjoong Sin,et. Al, An<br />
Energy-Efficient Distributed Unequal Clustering Protocol<br />
for Wireless Sensor <strong>Networks</strong>, Proceedings <strong>of</strong> World<br />
<strong>Academy</strong> <strong>of</strong> Science,Engineering and Technology volume<br />
36 december 2008 issn 2070-3740.<br />
[10] Manjeshwar A , Agrawal D . TEEN : A Protocol for<br />
Enhanced Efficiency in Wireless Sensor <strong>Networks</strong> . In<br />
Proceeding <strong>of</strong> the lth International Workshop on Parallel<br />
and Distributed Computing Issues in Wireless <strong>Networks</strong><br />
and Mobile Computing’01.2001:23-27.<br />
[11] Intanagonwiwat C,Govindan R,Estdn D,et a1.Directed<br />
Diffusion for Wireless Sensor Networking [J].IEEE/<br />
ACM Transactions on Networking,2003,1 1(1):2-16.<br />
[12] Muruganathan S D, Ma DCF, Bhasin PI, et al. A<br />
centralized energy-efficient routing protocol for wireless<br />
sensor networks. IEEE Communications Magazine,<br />
2005,43(3): 8 – 13.<br />
[13] Guo Bin,Li Zhe.United voting dynamic cluster routing<br />
algorithm based on residual-energy in wireless sensor<br />
networks.<strong>Journal</strong> <strong>of</strong> Electronics & Information Technology.<br />
2007,29(12).pp:3006-3010.
1064 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
[14] Heinzelman W,chandrakasan A,BalakrishnanH.An<br />
Application-Specific Protocol Architecture for Wireless<br />
Microsensor <strong>Networks</strong>.IEEE Transactions on Wireless<br />
Communications,October 2002,1(4):660—670.<br />
[15] C. Schurgers and M. B. Srivastava, “Energy Efficient<br />
Routing in Wireless Sensor <strong>Networks</strong>,” in proc. IEEE<br />
Military Communications Conf. vol. 1, pp. 357-361. 2001.<br />
[16] Manjeshwar A , Agrawal D . TEEN : A Protocol for<br />
Enhanced Efficiency in Wireless Sensor <strong>Networks</strong> . In<br />
Proceeding <strong>of</strong> the lth International Workshop on Parallel<br />
and Distributed Computing Issues in Wireless <strong>Networks</strong><br />
and Mobile Computing’01.2001:23-27.<br />
[17] Xinyun Fan,Fubao Wang etc,Wireless sensor network<br />
routing protocols. Computer Measurement & Control,<br />
2005,1 3 (9) :1010-1013.<br />
[18] LindseyS,Raghavendra C . PEGASlS : Power-Efficient<br />
Gathering in Sensor Information Systems.In Proceeding<br />
<strong>of</strong> the IEEE Aerospace Conference . Montana : IEEE<br />
Aerospace and Electronic Systems Society,2002:1 125-<br />
l 130.<br />
[19] Guojun Wang , Tian Wang, Weijia Jia, Minyi Guo, Hsiao-<br />
Hwa Chen, Mohsen Guizani “Local Update-Based Routing<br />
Protocol in Wireless Sensor <strong>Networks</strong> with Mobile Sinks”<br />
This full text paper was peer reviewed at the direction <strong>of</strong><br />
IEEE Communications Society subject matter experts for<br />
publication in the ICC 2007 proceedings.<br />
[20] Hayoung Oh and Kijoon Chae “An Energy-Efficient<br />
Sensor Routing with low latency, scalability for Smart<br />
Home <strong>Networks</strong>” International <strong>Journal</strong> <strong>of</strong> Smart Home,<br />
Vol. 1, No. 2, July, 2007.<br />
[21] Chan H, Perrig A (2004) ACE: An emergent algorithm for<br />
highly uniform cluster formation. In:Proceedings <strong>of</strong> the<br />
first European workshop on sensor networks (EWSN),<br />
2004 3.<br />
[22] Ye M, Li CF, Chen GH, Wu J (2004) EECS: an energy<br />
efficient clustering scheme in wireless sensor networks. In:<br />
Proceedings <strong>of</strong> the IEEE international workshop on<br />
strategies for energy efficiency in ad hoc and sensor<br />
networks (IWSEEASN’05), April 2004.<br />
[23] Ming Liu · Jiannong Cao · Yuan Zheng,et.al. An energyefficient<br />
protocol for data gathering and aggregation in<br />
wireless sensor networks, J Supercomput (2008) 43: 107–<br />
125.<br />
[24] Heinzelman WR, et al(2002) An application—specific<br />
protocol architecture for wireless microsensor networks.<br />
IEEE Trans Wireless Commun 1(4):660–670.<br />
[25] ZHOU Si-Wang, LIN Ya-Ping etc. A Wavelet Data<br />
Compression Algorithm Using Ring Topology for Wireless<br />
Sensor <strong>Networks</strong>.<strong>Journal</strong> <strong>of</strong> S<strong>of</strong>tware. 2007.18(3).669-680.<br />
[26] Gandham S R,Dawande M,Prakash R,etc.Energy efficient<br />
schemes for wireless sensor networks with multiple mobile<br />
base station[A].In:GLOBECOM 2003,IEEE<br />
Comunnications Society [C].SanFrancisco,USA:2003.377-<br />
381.<br />
[27] D.Braginsky,D.Estrin,Rumor routing algorithm for sensor<br />
networks, WSNA’ 02, Atlanta, GA, September, 2002.<br />
[28] V.Rodoaplu and T.H.Meng, Minimum energy mobile<br />
wireless networks, IEEE J.Select.Areas Communi., vol.17,<br />
no 8,pp.1333-1334,1999.<br />
[29] Ming Zhang, Suoping Wang et.al, An Novel Energy-<br />
Efficient Minimum Routing Algorithm (EEMR) in<br />
Wireless Sensor <strong>Networks</strong>, WICOM2008,135-438.<br />
© 2011 ACADEMY PUBLISHER<br />
Ming Zhang is an associate<br />
pr<strong>of</strong>essor with the School <strong>of</strong><br />
Electronic Engineering, Huaihai<br />
Institute <strong>of</strong> Technology, Lian<br />
yungang,China.He received his<br />
Master degree in Computer Science<br />
and Technology from Soochow<br />
University, Jiangsu, China in 2002.<br />
Since 2006, he has been pursuing his<br />
Dr. degree in the Department <strong>of</strong> S<strong>of</strong>tware at Nanjing<br />
University <strong>of</strong> Posts & Telecommunications. His current<br />
research interests include wireless sensor networks,<br />
wireless networks and s<strong>of</strong>tware technology.<br />
Suoping Wang is currently as<br />
Pr<strong>of</strong>/Head <strong>of</strong> Wujiang School,<br />
Nanjing University <strong>of</strong> Posts &<br />
Telecommunications,Nanjing, Chaina.<br />
He graduated from Department <strong>of</strong><br />
Radio, Tsinghua University, Beijing,<br />
China in 1970.He received his<br />
Master degree in Department <strong>of</strong> Communications and<br />
electronic Engineering,Nanjing University <strong>of</strong> posts &<br />
Telecommunications,Nanjing, China in 1981.His research<br />
focuses on Real-time Systems, Wireless Network,<br />
Network Communication Theory and Technology.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1065<br />
Performance <strong>of</strong> UWB Systems with Direct-<br />
Sequence Bipolar Pulse Amplitude Modulation<br />
and RAKE Reception over IEEE 802.15.3a<br />
Channel<br />
Jingjing Wang 1, 2 Hao Zhang 2<br />
1 College <strong>of</strong> information Science & Technology, Qingdao University <strong>of</strong> Science & Technology, Qingdao, China<br />
2 Department <strong>of</strong> Electrical Engineering, Ocean University <strong>of</strong> China, Qingdao, China<br />
Email: kathy1003@163.com<br />
Abstract—Direct-Sequence Pulse Amplitude Modulation<br />
(DS-PAM) has been widely proposed for Ultra-Wideband<br />
(UWB) communication systems because it provides better<br />
performance with low computational complexity. UWB<br />
signals suffer from severe multi-path interference when<br />
employed in indoor fading environments. But using RAKE<br />
reception can make use <strong>of</strong> the rich multi-path <strong>of</strong> UWB<br />
systems to improve system performance. In this paper we<br />
present the performance <strong>of</strong> a RAKE receiver employing<br />
maximal ratio combining (MRC) in a DS UWB system with<br />
BPAM modulation. Performance in a practical multi-path<br />
fading Channel (IEEE 802.15.3a Channel) is considered to<br />
analyze the performance <strong>of</strong> DS-PAM UWB systems with<br />
different RAKE receivers. The bit error rate (BER) <strong>of</strong><br />
ARake, PRake, and SRake over DS-BPAM UWB systems is<br />
simulated. The results indicate that ARake has the best<br />
performance, SRake is better than PRake when the number<br />
<strong>of</strong> fingers is the same.<br />
Index Terms—Performance, Ultra-Wideband, Direct-<br />
Sequence, Pulse Amplitude Modulation, IEEE 802.15.3a,<br />
RAKE Receiver<br />
I. INTRODUCTION<br />
In wireless communications, electromagnetic waves<br />
with an instantaneous bandwidth greater than 25% <strong>of</strong> the<br />
center operating frequency or an absolute bandwidth <strong>of</strong><br />
1.5 GHz or more are referred to as Ultra-Wideband<br />
(UWB) signals[1][2]. The basic concept <strong>of</strong> UWB is to<br />
transmit and receive an extremely short duration burst <strong>of</strong><br />
radio frequency (RF) energy to implement high data rate<br />
transmission. UWB is a promising technology for future<br />
high speed wireless communication systems. Pulse<br />
amplitude modulation (PAM), pulse position modulation<br />
(PPM) and on/<strong>of</strong>f keying (OOK) modulation are the most<br />
commonly used modulation schemes in UWB systems.<br />
PPM modulation uses the precise collocation <strong>of</strong> impulses<br />
in time to convey information, while PAM and OOK use<br />
the impulse amplitude for this purpose. UWB systems<br />
with PAM and PPM modulation have been extensively<br />
This work is supported by Outstanding Youth Fund <strong>of</strong> Shandong<br />
province under Grant no.JQ200821 and New Century Educational<br />
Talents Plan <strong>of</strong> Chinese Education Ministry under Grant no. NCET-08-<br />
0504<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1065-1071<br />
investigated [3][4][5][6].<br />
Wireless communications systems typically operate in<br />
multi-path fading channels. In addition to the direct path<br />
signal (if present), many reflected path signals can arrive<br />
at the receiver with different delays and attenuations,<br />
resulting in fading and inter symbol interference.<br />
Employing a RAKE receiver is an efficient means <strong>of</strong><br />
overcoming these effects to achieve better performance<br />
[7][8]. Actually, one <strong>of</strong> the advantages <strong>of</strong> broadband<br />
wireless communication systems such as code division<br />
multiple access (CDMA) and UWB is the capability <strong>of</strong><br />
utilizing multi-path signals to improve system<br />
performance and capacity. Since UWB systems can<br />
resolve many paths they are rich in multi-path diversity,<br />
so the use <strong>of</strong> RAKE diversity combining can be very<br />
effective. Considering the reasons given above, a RAKE<br />
receiver is an essential component <strong>of</strong> future UWB<br />
communication systems [9][10].<br />
The performance <strong>of</strong> PPM and PAM with Multiple<br />
Receive antennas has been investigated [11][12] [13].<br />
And [14] presented the Performance <strong>of</strong> UWB systems<br />
with PPAM and RAKE Reception. But most conclusions<br />
are derived over additive white Gaussian channels and<br />
Rayleigh, Ricean fading channels. IEEE 802.15.3a<br />
channel which could express the actual indoor fading<br />
channel well is neglected.<br />
Compared with TH-PPM, DS-PAM provides a lower<br />
BER for the same Signal to Noise Ratio (SNR) and the<br />
computational complexity. Compared with PPAM, DS-<br />
PAM provides a less complicated hardware.<br />
In this paper, we consider a DS-BPAM UWB system<br />
over IEEE 802.15.3a channel model with a RAKE<br />
receiver, and the performance with different RAKE<br />
receivers is analyzed.<br />
The remainder <strong>of</strong> the paper is organized as follows. In<br />
Section II, UWB system model is described. Section III<br />
introduces the channel model and three primary<br />
parameters that are important to characterize multi-path.<br />
Section IV introduces the IEEE802.15.3a channel model.<br />
Section V presents the error probability analysis <strong>of</strong> a DS-<br />
BPAM UWB system with a RAKE receiver over IEEE<br />
802.15.3a channel. The performance <strong>of</strong> a DS-BPAM<br />
UWB system with different RAKE receivers is examined<br />
and some conclusions are given in Section VI.
1066 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
b<br />
Code<br />
Repetition<br />
Coder<br />
(Ns, 1)<br />
Binary<br />
+1 Series<br />
II. UWB SYSTEM MODEL<br />
Transmission<br />
Coder<br />
C is a binary code<br />
Figure 1 shows the transmission scheme for a PAM-<br />
DS-UWB signal [15]. Input data sequence is a binary<br />
sequence = (…, b0, b1, bk, bk+1,). And each bit is<br />
repeated Ns times by the first module, Code Repetition<br />
Coder. Then a new binary sequence is generated<br />
which is (…, b0, b0 , …, b0, b1, b1, …, b1, …, be, be, …, be ).<br />
The second module transforms the sequence to a<br />
sequence which only includes positive and negative<br />
element. The transmission encoder applies a binary code<br />
composed <strong>of</strong> ± and period Np to the binary sequence d<br />
which equal to a . The sequence d enters the PAM<br />
modulator which generates a sequence <strong>of</strong> unit pulses<br />
whose position is . Then, the output <strong>of</strong> the modulator<br />
enters the pulse shaper filter and the result s<br />
b<br />
*<br />
a<br />
*<br />
a<br />
a<br />
1<br />
⋅c<br />
jTS<br />
(t)<br />
is<br />
transmitted.<br />
The PAM-DS-UWB signal s(t)<br />
at the output <strong>of</strong> the<br />
transmitter can be expressed as<br />
T<br />
a*<br />
s( t)<br />
ETX<br />
d j p(<br />
t − jTS<br />
)<br />
= ∑ ∞<br />
j=<br />
−∞<br />
Where S is the frame time, i.e. average pulse<br />
repetition period, p(t) is the energy-normalized waveform<br />
<strong>of</strong> the basic pulse and E is the transmitted energy per<br />
TX<br />
pulse. E is assumed to be 1 in Figure 1.<br />
TX<br />
At the receiver, the received signal is PAM<br />
demodulated. After detecting procedure in demodulator,<br />
DS code which is identical to that utilized at the<br />
transmitter is employed to recover the transmitted<br />
sequences. Then, the output <strong>of</strong> code repetition decoder is<br />
decoded to estimate original data sequence.<br />
III. CHANNEL MODEL<br />
For a UWB system with multi-path fading, the<br />
following discrete impulse response <strong>of</strong> the channel is<br />
considered:<br />
∑ − L 1<br />
l=<br />
0<br />
(1)<br />
h( t)<br />
= a δ ( t −τ<br />
)<br />
(2)<br />
τ<br />
where al<br />
is the channel gain for the l-th path, l is the<br />
delay for the l-th path, L is the number <strong>of</strong> resolvable<br />
paths, and δ (⋅)<br />
is the Dirac delta function. If the relative<br />
delay <strong>of</strong> two paths is less than a pulse width, they cannot<br />
be identified by a RAKE receiver. Thus we<br />
τl −τk ≥Tp<br />
assume , ∀l ≠ k T p , where is the width <strong>of</strong> p(t).<br />
© 2011 ACADEMY PUBLISHER<br />
l<br />
l<br />
a<br />
The primary parameters that are important to<br />
characterize multi-path: the total multi-path gain, the root<br />
mean square delay spread, and the power delay pr<strong>of</strong>ile.<br />
The following sub-sections describe in more detail each<br />
<strong>of</strong> these components.<br />
A. The Total Multi-Path Gain<br />
The total multi-path gain G measures the total amount<br />
<strong>of</strong> energy collected over the N received pulses when a<br />
pulse with unitary energy is transmitted. The G parameter<br />
can be determined as follows:<br />
∑ − L 1<br />
al<br />
l=<br />
0<br />
2<br />
G =<br />
(3)<br />
Given the G value, the impulse response can be written:<br />
Where 0 1 ,..., L−<br />
∑ − L 1<br />
l = 0<br />
h ( t ) = G α δ ( t − τ )<br />
(4)<br />
l<br />
α α are the energy-normalized channel<br />
gain parameters verifying:<br />
∑ − L 1<br />
l=<br />
0<br />
2<br />
l<br />
l<br />
α = 1<br />
(5)<br />
Note that G ≤1<br />
and is related to the attenuation<br />
suffered by the transmitted pulses during propagation. In<br />
multi-path environments, G decreases with distance<br />
according to the following law:<br />
G0<br />
G = (6)<br />
γ<br />
d<br />
Where G is the reference value for power gain evaluated<br />
0<br />
at d=1 m and γ is the exponent <strong>of</strong> the power or energy<br />
attenuation law. The G0<br />
value can be evaluated as<br />
follows:<br />
0 / 10<br />
10 A −<br />
G = (7)<br />
0<br />
Where A (in dB) represents the path loss at a reference<br />
0<br />
PAM<br />
Modulator<br />
Figure1. Transmission scheme for a PAM-DS-UWB signal.<br />
d<br />
d=a.c<br />
Pulse<br />
Shaper<br />
p(t)<br />
distance d0=1 m, that is, 10 log 10(<br />
/ ) , is the<br />
E E<br />
A =<br />
E<br />
0 TX RX 0<br />
s( t)<br />
d j p(<br />
t − jTS<br />
)<br />
= ∑ ∞<br />
j=<br />
−∞<br />
RX 0<br />
energy <strong>of</strong> a single pulse at d0. Values for both and 0<br />
are suggested in [16] for different propagation<br />
environments: A 47dB<br />
0 and = 1.<br />
7<br />
= γ for a LOS<br />
environment, and A 51dB 0 and = 3.<br />
5<br />
= environment.<br />
γ for a NLOS<br />
A γ
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1067<br />
B. The Root Mean Square Delay Spread<br />
τ<br />
rms<br />
=<br />
L−1<br />
2<br />
∑τ<br />
l<br />
l=<br />
0<br />
G<br />
a<br />
2<br />
l<br />
⎛<br />
⎜<br />
− ⎜<br />
⎜<br />
⎜<br />
⎝<br />
L−1<br />
∑<br />
τ a<br />
l<br />
l=<br />
0<br />
Equation (8) measures the effective duration <strong>of</strong> the<br />
channel impulse response. It is a fundamental parameter<br />
for evaluating the presence <strong>of</strong> Inter Symbol Interference<br />
(ISI) at the receiver. If the time interval separating two<br />
pulses is smaller than τ rms , ISI is present.<br />
C. The Power Delay Pr<strong>of</strong>ile(PDP)<br />
The Power Delay Pr<strong>of</strong>ile (PDP) <strong>of</strong> an impulse response<br />
given by Equation (4) is a graphical representation that<br />
shows time <strong>of</strong> arrival <strong>of</strong> the different contributions versus<br />
received power. Time <strong>of</strong> arrival <strong>of</strong> a generic path is<br />
usually indicated relative to the LOS contribution, which<br />
has a time <strong>of</strong> arrival fixed at 0.<br />
G<br />
2<br />
l<br />
2<br />
⎞<br />
⎟<br />
⎟<br />
⎟<br />
⎟<br />
⎠<br />
IV. THE IEEE802.15.3A CHANNEL MODEL<br />
The IEEE 802.15.3a channel model is based on the<br />
Saleh-Valenzuela (S-V) model in [17], where the impulse<br />
response is composed <strong>of</strong> exponentially decaying signal<br />
clusters to model the dense multi-path components. The<br />
UWB indoor channel model is then [18]<br />
− L 1 K ( l)<br />
∑∑<br />
l=<br />
0 k = 0<br />
(8)<br />
h( t)<br />
= X α δ ( t −T<br />
−τ<br />
)<br />
(9)<br />
l,<br />
k<br />
where<br />
α } are the coefficients <strong>of</strong> the k-th multi-path<br />
{ l, k<br />
contribution <strong>of</strong> the n-th cluster,<br />
T<br />
{ l } is the time <strong>of</strong> arrival <strong>of</strong> the l-th cluster,<br />
{ τ lk }is the delay <strong>of</strong> the k-th multi-path contribution<br />
within the l-th cluster,<br />
{X} is a log-normal random variable representing the<br />
amplitude gain <strong>of</strong> the channel.<br />
The proposed model describes four different<br />
measurement environment named CM1, CM2, CM3 and<br />
CM4 as shown in Table 1. CM1 describes a line-<strong>of</strong>-sight<br />
(LOS) channel with a distance from transmitter to<br />
receiver less than 4 meters. CM2 describes a non-LOS<br />
channel with the same range (0-4m). CM3 describes a<br />
non-LOS channel for distances between 4 and 10m. CM4<br />
describes an extreme NLOS multi-path channel.<br />
The parameters are defined as:<br />
Λ, inter-cluster (cluster) average arrival rate;<br />
λ , intra-cluster (ray) average arrival rate;<br />
Γ , inter-cluster (cluster) average decay rate;<br />
γ , intra-cluster (ray) average decay rate;<br />
σ ξ , average cluster lognormal standard deviation;<br />
σ ζ , average ray lognormal standard deviation;<br />
σ g , channel amplitude gain standard deviation.<br />
© 2011 ACADEMY PUBLISHER<br />
l<br />
lk<br />
TABLE I.<br />
PARAMETERS FOR IEEE802.15.3A CHANNEL MODEL<br />
Channel<br />
Model<br />
CM 1<br />
Λ λ Γ σ ξ σ ζ σ g<br />
LOS<br />
(0-4m)<br />
CM 2<br />
0.0233 2.5 7.1 4.3 3.3941 3.3941 3<br />
NLOS<br />
(0-4m)<br />
CM 3<br />
0.4 0.5 5.5 6.7 3.3941 3.3941 3<br />
NLOS<br />
(4-10m)<br />
CM 4<br />
0.0667 2.1 14 7.9 3.3941 3.3941 3<br />
Extreme<br />
NLOS<br />
0.0667 2.1 24 12 3.3941 3.3941 3<br />
V. ERROR PROBABILITY ANALYSIS OF DS-<br />
BPAM UWB SYSTEM OVER IEEE802.15.3A<br />
CHANNEL MODEL WITH RAKE RECEIVER<br />
In multi-path fading channels, many reflected path<br />
signals can arrive at the receiver with different delays and<br />
attenuation, resulting in fading and inter-symbol<br />
interference. Employing a RAKE receiver is an efficient<br />
means <strong>of</strong> overcoming these effects.<br />
m 1(t)<br />
r(t) τ + Ts<br />
Z1 …<br />
m 2(t)<br />
…<br />
m N(t)<br />
∫<br />
τ<br />
dt<br />
correlator<br />
τ + T s<br />
∫<br />
τ<br />
τ + T s<br />
∫<br />
τ<br />
d t<br />
dt<br />
Z 2<br />
Z N<br />
ω 1<br />
ω 2<br />
…<br />
ω N<br />
Detector<br />
Figure 2. The structure <strong>of</strong> a RAKE receiver with N correlators.<br />
Estimated<br />
symbol<br />
The typical structure <strong>of</strong> a RAKE receiver is shown in<br />
Figure 2, which consists <strong>of</strong> a series <strong>of</strong> correlators and a<br />
detector. The correlators or matched filters are also called<br />
fingers. Each RAKE finger is matched to a particular<br />
multi-path component in order to combine them<br />
coherently. A reference or template signal matched to the<br />
incoming received signal is used by the RAKE receiver.<br />
Each finger <strong>of</strong> the RAKE receiver uses a delayed version<br />
<strong>of</strong> the template signal to match the delay to a specific<br />
multi-path component. In order to enable symbol-rate<br />
sampling, the received signal is correlated with a symbollength<br />
template signal, and the output <strong>of</strong> the correlator is<br />
sampled once per symbol. If the receiver uses all L<br />
received paths, it is called All-RAKE (ARake)[19].<br />
However, the number <strong>of</strong> multi-path components that can<br />
be utilized in a typical RAKE combiner is limited by<br />
power issues, design complexity, and the quality <strong>of</strong> the<br />
channel estimation. Thus, in practice, only a subset <strong>of</strong> the<br />
resolved multi-path components is used, giving rise to the<br />
γ<br />
Z
1068 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Partial RAKE (PRake) and Selective RAKE (SRake)<br />
receivers which have a limited number <strong>of</strong> fingers. The<br />
PRake receiver uses the first M arriving paths out <strong>of</strong> the L<br />
resolvable multi-path components, while SRake searches<br />
for the M best paths [19].<br />
The transmitted symbol can be simply described as<br />
formula (1).The combining algorithm commonly used in<br />
RAKE receivers is maximal ratio combining (MRC)<br />
which is known to maximize the signal-to-noise ratio<br />
(SNR) for diversity channels [20].<br />
The received signal over IEEE 802.15. 3a channel can<br />
be simply described as<br />
∑ − L 1<br />
l=<br />
0<br />
r( t)<br />
= α s(<br />
t −τ<br />
) + σ n(<br />
t)<br />
where n(t) is white noise.<br />
The template signal is<br />
where<br />
l<br />
L 1<br />
= ∑ −<br />
m ( t)<br />
ω m(<br />
t −τ<br />
)<br />
R<br />
j=<br />
0<br />
m( t τ j ) = d j p(<br />
t − jTS<br />
−τ<br />
j )<br />
l<br />
j<br />
n<br />
j<br />
(10)<br />
(11)<br />
− (12)<br />
and ωj is the RAKE combining weight <strong>of</strong> j-th branch, L<br />
is the number <strong>of</strong> RAKE receiver branches. For BPAM, d<br />
∈{0,1}. ω = [ω0, ω1, … , ωL−1] are the RAKE combining<br />
weights. If MRC technique is used, the amplitudes <strong>of</strong> the<br />
received multi-path components (MPCs) are estimated<br />
and used as weighing vector ω in each finger. In case <strong>of</strong><br />
ARake, the combining weights are chosen as ω = α,<br />
where α = [α0, α1, … , αL−1] are the fading coefficients <strong>of</strong><br />
the channel. If the set <strong>of</strong> indices <strong>of</strong> the M best fading<br />
coefficients with largest amplitude is denoted by S, then<br />
the combining weights ω <strong>of</strong> an SRake are chosen as [21],<br />
⎧α<br />
l, l∈S ω = ⎨<br />
⎩ 0, l∉S , (13)<br />
Similarly, for PRake using the first M multi-path<br />
components, the weights <strong>of</strong> MRC combining are given by<br />
[20],<br />
⎧αl<br />
, l = 0, …, M −1<br />
ω = ⎨<br />
(M≤L) (14)<br />
⎩ 0, l = M, …L−1 The output <strong>of</strong> combiner is<br />
τ + T s<br />
Z = ∫ r( t) m ( t) dt , (15)<br />
τ<br />
Assuming a perfect match <strong>of</strong> the received signal with<br />
the template signal, zero inter-frame and inter-symbol<br />
interference, and symbol rate sampling at the output <strong>of</strong><br />
RAKE fingers, then Equation (15) can be rewritten in<br />
discrete time as<br />
© 2011 ACADEMY PUBLISHER<br />
R<br />
∑ − L 1<br />
b<br />
l=<br />
0<br />
Z = E ω α + n<br />
(16)<br />
where Eb is the energy per bit.<br />
+∞<br />
σ n R<br />
−∞<br />
n= ∫ n() t m () t dt<br />
l<br />
l<br />
(17)<br />
n is the noise at the output <strong>of</strong> the correlator which is<br />
2<br />
approximately distributed as n ~ N(0, σn<br />
).<br />
To determine the BER at the output <strong>of</strong> the RAKE, the<br />
output SNR needs to be evaluated. From Equation (15),<br />
the approximate signal energy and the noise variance at<br />
the output <strong>of</strong> RAKE are evaluated as<br />
(<br />
2<br />
) =<br />
L−1<br />
b( ∑<br />
l=<br />
0<br />
l<br />
2<br />
l)<br />
, (18)<br />
E signal E ω α<br />
2<br />
=<br />
L−1<br />
2<br />
n ∑<br />
l=<br />
0<br />
E( noise ) σ ωl<br />
2<br />
. (19)<br />
In case <strong>of</strong> BPAM, for a given SNR per Bit γb, the<br />
approximate expression <strong>of</strong> BER conditioned on a<br />
particular channel realization is given by [22],<br />
⎛ L−1<br />
2<br />
Eb(<br />
ωα<br />
⎞<br />
l 0 l l<br />
Pe| α ( γ b)<br />
Q( SNR) Q⎜<br />
∑ )<br />
=<br />
= ≈ ⎟<br />
⎜ 2 L−1<br />
2<br />
⎜<br />
⎟<br />
σn ω ⎟<br />
⎝ ∑l=<br />
0 l ⎠<br />
(20)<br />
where Q(.) is the standard Q function.<br />
However, to obtain the error probabilities when<br />
channel fading coefficients α are random, we must<br />
average the Pe(γb) over the probability density function<br />
<strong>of</strong> γb [21],<br />
∞<br />
P = P( γ ) p( γ ) dγ<br />
. (21)<br />
∫<br />
b<br />
e e b b<br />
0<br />
By evaluating the probability distribution function <strong>of</strong><br />
output SNR, average BER can be obtained. It is difficult<br />
to obtain a closed-form expression <strong>of</strong> (21). However, this<br />
average can be evaluated numerically, or by employing<br />
Monte-Carlo simulations.<br />
Figure 3 shows the equivalent RAKE receiver structure<br />
with BPAM based on discrete-time channel models.<br />
The adoption <strong>of</strong> a Rake considerably increases the<br />
complexity <strong>of</strong> the receiver. This complexity increases<br />
with the number <strong>of</strong> multi-path components analyzed and<br />
combined before decision, and can be reduced by<br />
decreasing the number <strong>of</strong> components processed by the<br />
receiver. However, a deduction <strong>of</strong> the number <strong>of</strong> paths<br />
leads to a decrease <strong>of</strong> energy collected by the receiver. It<br />
is important to catch a good compromise between the two<br />
elements.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1069<br />
L−1<br />
∑<br />
rt () = αst ( − τ ) + nt ()<br />
l l<br />
l=<br />
0<br />
m(t)<br />
τ + Ts<br />
∫<br />
τ<br />
dt<br />
Correlator<br />
t = kΔt<br />
Detector<br />
ωN-1 ZN-1 ω2 ω1<br />
Z Z 2<br />
1<br />
Estimate<br />
Bits<br />
Figure 3. Equivalent RAKE receiver structure with BPAM<br />
VI. NUMERICAL RESULTS AND CONCLUSIONS<br />
In this section, the performance <strong>of</strong> a DS-BPAM UWB<br />
system with RAKE receiver over an IEEE 802.15.3a<br />
channel is presented.<br />
Figure 4. Bit error rate for DS-BPAM UWB over CM1 <strong>of</strong> IEEE<br />
802.15.3a channel with different RAKE receivers.<br />
Figure 4 shows the bit error rate for DS-BPAM over<br />
CM1 <strong>of</strong> IEEE 802.15.3a channel with different RAKE<br />
receivers. This shows that there is almost a 1dB gain with<br />
an ARake receiver over an SRake receiver (the number <strong>of</strong><br />
fingers S=5) at a BER <strong>of</strong> 10 -2 . Also, there is almost a<br />
1.5dB gain with a SRake receiver over a PRake receiver<br />
(the number <strong>of</strong> fingers S=L=5) at a BER <strong>of</strong> 10 -2 .<br />
Figure 5 shows the bit error rate for DS-BPAM UWB<br />
over a CM2 <strong>of</strong> IEEE 802.15.3a channel with different<br />
RAKE receivers. In this case, there is almost a 1dB gain<br />
with an ARake receiver over a 5 finger SRake receiver at<br />
a BER <strong>of</strong> 10 -1 . Also, there is about a 1.7dB gain with an<br />
SRake receiver over a PRake receiver (S=L=5) at a BER<br />
<strong>of</strong> 10 -1 .<br />
Figures 6 and 7 show the bit error rates for DS-BPAM<br />
UWB over CM3 and CM4 <strong>of</strong> IEEE 802.15.3a channel<br />
with different RAKE receivers respectively. It is<br />
obviously that the bit error rates <strong>of</strong> a DS-BPAM UWB<br />
system have the similar trend over CM3 and CM4<br />
channels with CM1 and CM2.<br />
Z N<br />
Δ t<br />
Δt<br />
ω N<br />
Zout Z>0 b=0;<br />
= jT + N Δt<br />
Z
1070 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Figure 8. Bit error rate for DS-BPAM UWB over CM1-CM4 <strong>of</strong> IEEE<br />
802.15.3a channel with ARAKE receiver<br />
Figure 9. Bit error rate for DS-BPAM UWB over CM1-CM4 <strong>of</strong> IEEE<br />
802.15.3a channel with SRAKE(S=5) receiver<br />
Figure 10. Bit error rate for DS-BPAM UWB over CM1-CM4 <strong>of</strong> IEEE<br />
802.15.3a channel with PRAKE (P=5) receiver<br />
From Figure 4-7, the conclusion can be drawn that All-<br />
RAKE (ARake) receiver has the best performance,<br />
because it uses all the multi-path components which the<br />
receiver can identify. With the same number <strong>of</strong> fingers,<br />
SRake has a better performance than PRake. Because<br />
SRake has a selection process, the complexity <strong>of</strong> channel<br />
estimation and channel tracking is same with ARake, but<br />
© 2011 ACADEMY PUBLISHER<br />
the number <strong>of</strong> branches is smaller than that <strong>of</strong> ARake.<br />
PRake only considers the first arrival components and<br />
without having a selection process. For all types <strong>of</strong><br />
RAKE reception, the bigger number <strong>of</strong> fingers, the better<br />
performance will be achieved.<br />
Figure 8 to 10 compare the BER performance <strong>of</strong> DS-<br />
BPAM UWB over the CM1 to CM4 <strong>of</strong> IEEE 802.15.3a<br />
channel with ARAKE, SRAKE(S=5), PRAKE (P=5)<br />
receiver, respectively. As expected, the system<br />
performance over a CM1 channel is the best, while the<br />
BER performance <strong>of</strong> DS-BPAM deteriorates sharply<br />
when signals are transmitted over a CM4 channel.<br />
ACKNOWLEDGMENT<br />
The authors would like to thank the anonymous<br />
reviewers for their constructive comments and questions<br />
that greatly improved the paper.<br />
REFERENCES<br />
[1] OSD/DARPA Ultra-Wideband Radar Review Panel,<br />
“Assessment <strong>of</strong> Ultra-Wideband Technology,” Defense<br />
Advanced Research Projects Agency, July, 1990.<br />
[2] J.D. Taylor, Introduction to Ultra-Wideband Radar<br />
Systems, CRC Press, 1995.<br />
[3] R.A. Scholtz, “Multiple access with time-hopping impulse<br />
modulation,” Proc. IEEE Military Commun. Conf., pp. 11–<br />
14, Oct. 1993.<br />
[4] M.Z. Win and R.A. Scholtz, “Impulse radio: How it<br />
works,” Proc. IEEE Commun. Letts., vol. 2, pp. 36–38,<br />
Feb. 1998.<br />
[5] F. Ramirez-Mireles and R.A. Scholtz, “System<br />
performance analysis <strong>of</strong> impulse radio modulation,” Proc.<br />
IEEE Radio and Wireless Conf., pp. 67–70, Aug. 1998.<br />
[6] F. Ramirez-Mireles and R.A. Scholtz, “Multiple-access<br />
performance limits with time hopping and pulse position<br />
modulation,” Proc. IEEE Military Commun. Conf., pp.<br />
529–533, Oct. 1998.<br />
[7] J.G. Proakis, Digital Communications, 3rd Ed., McGraw<br />
Hill, New York, NY, 1995.<br />
[8] R. Price and P.E. Green Jr., “A communication technique<br />
for multi-path channels,” Proc. IRE, vol. 46, pp. 555–570<br />
Mar. 1958.<br />
[9] B. Mielczarek, M.O. Wessman and A. Svensson,<br />
“Performance <strong>of</strong> coherent UWB Rake receivers with<br />
channel estimators,” Proc. IEEE Vehic. Tech. Conf., pp.<br />
1880–1884, Oct. 2003.<br />
[10] S. Imada and T. Ohtsuki “Pre-RAKE diversity combining<br />
for UWB systems in IEEE 802.15 UWB multi-path<br />
channel,” Proc. Int. Workshop on Ultra Wideband Systems,<br />
pp. 236–240, May 2004.<br />
[11] H. Zhang and T. A. Gulliver, “Performance and capacity <strong>of</strong><br />
PAM and PPM UWB time-hopping multiple access<br />
communications with receive diversity,” EURASIP J.<br />
Applied Signal Processing, pp. 306-315, Mar. 2005.<br />
[12] H. Zhang and T. A. Gulliver, “Performance and capacity <strong>of</strong><br />
PAM and PPM UWB systems with multiple receiver<br />
antennas,” Proc. IEEE Pacific Rim Conf. on Commun.,<br />
Computers and Signal Processing, pp. 740-743, Aug. 2003.<br />
[13] H. Zhang and T. A. Gulliver, “Capacity <strong>of</strong> time-hopping<br />
PPM and PAM UWB multiple access communications<br />
over indoor fading channels,” EURASIP J. Wireless
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1071<br />
Commun. and Networking, vol. 2008, [online].<br />
Available:http://www.hindawi.com/journals/wcn/2008/273<br />
018.html [accessed: Sept.12, 2009]<br />
[14] Wei Li, T. Aaron Gulliver and Hao Zhang, “Performance<br />
<strong>of</strong> ultra-wideband transmission with pulse position<br />
amplitude modulation and rake reception” 2005<br />
IEEE/ACES International Conference on Wireless<br />
Communications and Applied Computational<br />
Electromagnetics,pp.1-4<br />
[15] Benedetto, M. G. D. and Giancola, G. 2004 Understanding<br />
Ultra Wide Band Radio Fundamentals. Prentice Hall, New<br />
Jersey.<br />
[16] Ghassemzadeh,S.S., L.J. Greenstein, A. Kavčić, T.<br />
Sveinsson, and V. Tarokh, “an empirical indoor path loss<br />
model for uwb channels”, <strong>Journal</strong> <strong>of</strong> Communication and<br />
<strong>Networks</strong>, VOL.5, pp.303-308, Dec. 2003<br />
[17] J. G. Proakis Digital Communications, 4th ed.<br />
Boston:McGraw-Hill,2001.<br />
[18] J. Foerster, ed., Channel modeling sub-committee report<br />
final, IEEE 802.15 Working Group for Wireless Personal<br />
Area <strong>Networks</strong>(WPANs), IEEE P802.15-02/490r1-SG3a,<br />
Feb.2003.<br />
[19] D. Cassioli, M. Z.Win, F. Vatalaro, and A. F. Molisch,<br />
“Performance <strong>of</strong> low-complexity RAKE reception in a<br />
realistic UWB channel,” Proc. IEEE Int. Conf. Commun.,<br />
pp. 763-767, May 2002.<br />
[20] S. Tantikovit, A. U. H. Sheikh, and M. Z. Wang,<br />
“Combining schemes in RAKE receiver for low spreading<br />
factor long-code W-CDMA systems,” IEE Elect. Letts., vol.<br />
36, no. 22, pp. 1872–1874, Oct. 2000.<br />
[21] S. Gezici, H. Kobayashi, H. V. Poor, and A. F. Molisch,<br />
“Performance evaluation <strong>of</strong> impulse radio UWB systems<br />
with pulse-based polarity randomization in asynchronous<br />
multi-user environments,” Proc. IEEE Wireless Commun.<br />
and Networking Conf., pp. 908-913, Mar. 2004.<br />
[22] H. Hashemi, “Impulse Response Modeling <strong>of</strong> Indoor<br />
Radio Propagation Channels,” IEEE JSAC, Vol. 11, No. 7,<br />
Sept. 1993,pp,967-968<br />
Jingjing Wang was born in Anhui, China, in 1975. She<br />
received her B.S. degree in industrial automation from<br />
Shandong University, Jinan, China, in 1993, the M.Sc. degree<br />
from control theory and control engineering, Qingdao<br />
University <strong>of</strong> Science & Technology, Qingdao, China in 2002.<br />
From 1997 to 1999, she was the assistant engineer <strong>of</strong> Shengli<br />
Oilfield, Dongying, China. From 2002 to now, she is an<br />
associate pr<strong>of</strong>essor at the College <strong>of</strong> information Science &<br />
Technology, Qingdao University <strong>of</strong> Science & Technology. Her<br />
research interests include 60GHz wireless communication, and<br />
ultra wideband radio systems.<br />
Hao Zhang was born in Jiangsu, China, in 1975. He<br />
received his B.S. degree in telecom engineering and industrial<br />
management from Shanghai Jiaotong University, Shanghai,<br />
China, in 1994, the M.B.A. degree from New York Institute <strong>of</strong><br />
Technology, Old Westbury, NY, in 2001, and the Ph.D. degree<br />
in electrical and computer engineering from the University <strong>of</strong><br />
Victoria, Victoria, BC, Canada, in 2004.<br />
From 1994 to 1997, he was the Assistant President <strong>of</strong> ICO<br />
(China) Global Communication Company, Beijing, China. He<br />
was the Founder and CEO <strong>of</strong> Beijing Parco Company, Ltd.,<br />
Beijing, China, from 1998 to 2000. In 2000, he joined Micros<strong>of</strong>t<br />
Canada, Vancouver, BC, as a S<strong>of</strong>tware Engineer, and was Chief<br />
Engineer at Dream Access Information Technology, Victoria,<br />
BC, Canada, from 2001 to 2002. He is currently an Adjunct<br />
© 2011 ACADEMY PUBLISHER<br />
Assistant Pr<strong>of</strong>essor with the Department <strong>of</strong> Electrical and<br />
Computer Engineering, University <strong>of</strong> Victoria. His research<br />
interests include ultra wideband radio systems, MIMO wireless<br />
systems, and spectrum communications.
1072 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Data Accuracy Estimation for Spatially<br />
Correlated Data in Wireless Sensor <strong>Networks</strong><br />
under Distributed Clustering<br />
Jyotirmoy Karjee , H.S Jamadagni<br />
Centre for Electronics Design and Technology, Indian Institute <strong>of</strong> Science, Bangalore, India<br />
kjyotirmoy@cedt.iisc.ernet.in, hsjam@cedt.iisc.ernet.in<br />
Abstract—Objective-The main purpose <strong>of</strong> this paper is to<br />
construct a distributed clustering algorithm such that each<br />
distributed cluster can perform the data accuracy at their<br />
respective cluster head node before data aggregation and<br />
transmit the data to the sink node.Design<br />
approach/Procedure – We investigate that the data are<br />
spatially correlated among the sensor nodes which form the<br />
clusters in the spatial domain. Due to high correlation <strong>of</strong><br />
data, these clusters <strong>of</strong> sensor nodes are overlapped in the<br />
spatial domain. To overcome this problem, we construct a<br />
distributed clustering algorithm with non-overlapping<br />
irregular clusters in the spatial domain. Then each<br />
distributed cluster can perform data accuracy at the cluster<br />
head node and finally send the data to the sink node.<br />
Findings- Simulation result shows the associate sensor nodes<br />
<strong>of</strong> each distributed cluster and clarifies their data accuracy<br />
pr<strong>of</strong>ile in the spatial domain. We demonstrate the<br />
simulation results for a single cluster to verify that their<br />
exist an optimal cluster which give approximately the same<br />
data accuracy level achieve by the single cluster. Moreover<br />
we find that as the distance from the tracing point to the<br />
number <strong>of</strong> sensor node increases the data accuracy<br />
decreases. Design Limitations – This model is only applicable<br />
to estimate data accuracy for distributed clusters where the<br />
sensed data are assumed to be spatially correlated with<br />
approximately same variations. Practical implementation –<br />
Measure the moisture content in the distributed agricultural<br />
field. Inventive/Novel idea- This is the first time that a data<br />
accuracy model is performed for the distributed clusters<br />
before data aggregation at the cluster head node which can<br />
reduce data redundancy and communication overhead.<br />
Index Terms—Wireless sensor networks, distributed<br />
clusters, data accuracy, spatial correlation<br />
I. INTRODUCTION<br />
Wireless sensor network has made a drastic change in<br />
communications for the last several years. One <strong>of</strong> the<br />
vital tasks <strong>of</strong> wireless sensor network is to sense or<br />
measure the physical phenomenon <strong>of</strong> data such as<br />
measurement <strong>of</strong> humidity, temperature, seismic event etc<br />
from the environment [1]. Physical phenomenon <strong>of</strong> data<br />
is measured or sense by a device called sensor nodes<br />
which are capable to sense, process and communicate the<br />
data through out the network. Since most <strong>of</strong> the data are<br />
spatially correlated [2] among them, the sensor nodes<br />
form clusters in the sensor field to reduce data collection<br />
cost [3]. According to literature survey, LEACH [4] gives<br />
a clear idea about how dynamically cluster and cluster<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1072-1083<br />
head are created according to a priori probability. Finally<br />
cluster head aggregate all the data and send it to the sink<br />
node. Similarly SEP [5] demonstrates the formation <strong>of</strong><br />
cluster in heterogeneous sensor networks. Since data<br />
correlation in wireless sensor networks shows Gaussian<br />
distribution with zero mean, literature [6] shows the<br />
spatial correlation among data is high in sensor networks<br />
but it lags the practical implementation <strong>of</strong> analyzing the<br />
correlated data for transmitting the packets for<br />
communication. Literature [7] proposes a grid based<br />
spatial correlation clustering method where the entire<br />
cluster is equipped in a grid sensor field. However this<br />
type <strong>of</strong> model rarely happens in an original scenario in<br />
wireless sensor networks. Moreover literature [8]<br />
proposes a disk-shaped circular cluster where sensor<br />
nodes are grouped into disjoint sets each managed by a<br />
designated cluster head which lags the practical shape <strong>of</strong><br />
a cluster. As most cases the cluster formation are<br />
irregular in shape for the spatial domain. Hence in this<br />
paper we propose a foundation <strong>of</strong> distributed clustering<br />
algorithm which is much more practical than the previous<br />
work done in the spatial domain. In our model, we<br />
propose a spatially correlated distributed irregular non<br />
overlapping cluster formation in the spatial domain.<br />
These distributed irregular cluster formation in the spatial<br />
domain is much more practical model in original scenario<br />
than the previous literature discussed above.<br />
Most <strong>of</strong> the work done till today is based upon the<br />
fact that the sink node or the base station is responsible<br />
for estimating the data accuracy for physically sensed<br />
data by sensor nodes [9, 10, 11] .Therefore it is applicable<br />
for one hop communication where the raw data are<br />
sensed and measured by the sensor nodes and directly<br />
transmitted to the sink node. Again we propose a model<br />
[12] for data accuracy where we have considered two hop<br />
communications in which physical phenomenon <strong>of</strong><br />
sensed data is transmitted via intermediate node called<br />
cluster head (CH)[18]node. But in this paper we propose<br />
a distributed clustering algorithm where each cluster can<br />
perform data accuracy at their respective CH node and<br />
finally send the data to the sink node. Each distributed<br />
cluster is responsible for sensing and measuring the<br />
physical phenomenon <strong>of</strong> data in the sensor region.<br />
The main goal <strong>of</strong> this paper is to estimate data<br />
accuracy for each distributed cluster before data<br />
aggregation [19] at their respective CH node which can<br />
reduce the data redundancy and communication
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1073<br />
overhead. However to the best understanding <strong>of</strong> the<br />
authors, there is no work done so far on verifying the data<br />
accuracy for distributed cluster before data aggregation<br />
[21, 22] at their respective CH node. Since from the<br />
literature survey we have seen that most <strong>of</strong> the work done<br />
till today is that data from cluster <strong>of</strong> sensor nodes directly<br />
send to CH node for aggregation without verifying its<br />
accuracy. Hence it is important that the most precise or<br />
accurate data send by the distributed cluster can<br />
aggregate at their respective CH node before transmitting<br />
to the sink node and not aggregating all the redundant<br />
data at CH node. The data send by each distributed<br />
cluster should first verify its accuracy level at their<br />
respective CH node then only the data get aggregates and<br />
finally send to the sink node. Since CH node verifies the<br />
data accuracy for their respective distributed cluster, it<br />
may reduce the power consumption and increase the<br />
lifetime <strong>of</strong> the networks.<br />
Another important reason for estimating data accuracy<br />
for each distributed cluster before data aggregation at<br />
their respective CH node, if some <strong>of</strong> the sensor nodes in<br />
the distributed cluster get malicious [20]. If some <strong>of</strong> the<br />
sensor nodes become malicious in the distributed cluster,<br />
then it can sense and read inaccurate data. These<br />
inaccurate data send by the malicious nodes gets<br />
aggregated with the other correct data results in<br />
inaccurate (incorrect) data aggregation at the CH node <strong>of</strong><br />
their respective cluster and finally send to the sink node.<br />
This may increase the power consumption, data<br />
redundancy and communication overhead in the<br />
distributed network. It results very high or low variations<br />
<strong>of</strong> the estimated data accuracy value compare to the<br />
actual variations <strong>of</strong> estimated data accuracy value at the<br />
CH node. Hence to overcome this problem, it is important<br />
to estimate the data accuracy at CH node for distributed<br />
cluster before data aggregation and send the accurate data<br />
to the sink node. In our model we assume that the sensed<br />
data are spatially correlated with approximately the same<br />
variations in each distributed cluster and the sensor nodes<br />
are appropriate to sense the correct data. We verify<br />
estimated data accuracy with approximately same<br />
variations at the CH node for each distributed cluster.<br />
In our model, each distributed cluster is responsible to<br />
sense the physical phenomenon <strong>of</strong> data such as moisture<br />
content <strong>of</strong> soil in the sensor region. Once the data<br />
accuracy is processed by CH node for each distributed<br />
cluster, it transmits the estimated accurate data to the sink<br />
node. From the literature survey, it is clear that only the<br />
sensor nodes are responsible to sense the physical<br />
phenomenon <strong>of</strong> data and not the sink node. But in our<br />
model not only sensor nodes are responsible to sense the<br />
physical phenomenon <strong>of</strong> data but the CH node can also<br />
do the sensing phenomenon in each distributed cluster.<br />
We investigate how each distributed cluster can sense the<br />
physical phenomenon <strong>of</strong> data to estimate the data<br />
accuracy in the sensor field. Literature [9, 13] has given<br />
some approaches regarding jointly sensing nodes which<br />
gives an idea about how the raw data is sensed by the<br />
jointly sensing nodes and how the number <strong>of</strong> jointly<br />
sensing nodes affects the data accuracy. However they<br />
© 2011 ACADEMY PUBLISHER<br />
address this problem if only sensing nodes are<br />
responsible to retrieve physical phenomenon <strong>of</strong> data<br />
where they investigate to find a proper number and<br />
positions <strong>of</strong> jointly sensing nodes. But in our model, we<br />
consider both the sensor nodes and the CH node which<br />
forms each distributed cluster in the sensor field are<br />
sensing the physical phenomenon such as humidity or<br />
moisture content <strong>of</strong> the soil. Since we verify data<br />
accuracy for each distributed cluster in the sensor field,<br />
there exit an optimal cluster which gives approximately<br />
the same data accuracy level achieve by each cluster.<br />
Rest <strong>of</strong> the paper is given as follows. In section II,<br />
we construct a data correlation model for sensor nodes in<br />
spatial domain. These data correlation can give rise to<br />
overlapping <strong>of</strong> clusters in the sensor region. Hence to<br />
overcome this problem, we propose a distributed<br />
clustering algorithm with non overlapping irregular<br />
clusters in the spatial domain. Then we perform data<br />
accuracy for each distributed cluster at CH node before<br />
data aggregation in the sensor region. In section III, we<br />
verify simulation results for distributed clusters. We<br />
demonstrate results how each distributed cluster are<br />
formed with their respective associate nodes and their<br />
data accuracy. Then we show the performance model <strong>of</strong> a<br />
single cluster with respect to data accuracy. Finally we<br />
conclude our work in section IV.<br />
II. SYSTEM MODEL<br />
In this section, sensor nodes deployment strategies are<br />
done where the sensor nodes form distributed clusters<br />
which are capable to perform data accuracy in the spatial<br />
domain. We propose an algorithm for distributed clusters<br />
which perform data accuracy at the cluster head node<br />
where the data are spatially correlated and finally send<br />
the data to the sink node. Let a set <strong>of</strong> sensor nodes are<br />
deterministically deployed uniformly over a sensor region<br />
Z. These set <strong>of</strong> sensor nodes forms the cluster head nodes<br />
[18] for the distributed clusters equipped with additional<br />
energy resource [5]. Since CH node perform the data<br />
accuracy for the respective distributed clusters, we set the<br />
CH node with additional energy resource and distributed<br />
deterministically in the sensor field. Again another set <strong>of</strong><br />
sensor nodes are randomly deployed over the sensor<br />
region Z and are called normal nodes [5]. Normal nodes<br />
form the distributed cluster along with their respective<br />
CH node which can sense and measure the spatially<br />
correlated data and estimate the data accuracy at the CH<br />
node.CH node has more energy resource than the normal<br />
nodes because CH nodes has to estimate the data<br />
accuracy for the cluster. Thus CH nodes and normal<br />
nodes form the total set <strong>of</strong> sensor nodes represented as L<br />
with Z ⊆ R 2 where ||L|| can be represented as total<br />
number <strong>of</strong> sensor nodes. They are capable for sensing and<br />
measuring the spatially correlated data in the sensor<br />
region Z. For example, we measure the moisture content<br />
<strong>of</strong> soil at different locations <strong>of</strong> sensor region Z. Generally<br />
there are much more variations in measurement <strong>of</strong><br />
moisture content at different locations in the sensor field.<br />
Some places the water (or moisture) content in the soil
1074 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
are more than other different places where the water<br />
(moisture) content is less. Thus there are variations <strong>of</strong><br />
monitoring the measurement <strong>of</strong> moisture content in the<br />
soil at different places in the sensor region Z.<br />
A. Data Correlation for sensor nodes in Spatial Domain<br />
We consider reference values for higher concentration<br />
<strong>of</strong> moisture content at different places <strong>of</strong> sensor region Z.<br />
Suppose the reference values are called tracing points<br />
[20] and can be represented as S i where i=1, 2, 3…n are<br />
the number <strong>of</strong> tracing points at different locations in the<br />
sensor field with higher variations .The tracing points can<br />
be located at the different places <strong>of</strong> sensor field where the<br />
moisture content is high. For example, water (or<br />
moisture) content in the soil can be higher at different<br />
locations <strong>of</strong> the sensor field. It is considered as reference<br />
values for tracing points at different locations in sensor<br />
the region Z. Although the data are spatially correlated in<br />
the sensor region, there are variations in measurement for<br />
concentratation <strong>of</strong> data (moisture content) at different<br />
places in sensor region Z. The higher concentratation <strong>of</strong><br />
data has higher variations with respect to lower variations<br />
<strong>of</strong> data at different places. In spatial domain, data<br />
correlation depends upon the distance between the tracing<br />
points to the sensor nodes and the distance between<br />
jointly sensing nodes [13]. Thus we have two points to<br />
note in our work. Firstly, data correlation decreases as the<br />
distance between the tracing points (or reference values)<br />
to the sensor nodes increases. Secondly, data correlation<br />
decreases as the distance between jointly sensing nodes<br />
increases. Thus data correlation is more when the sensor<br />
nodes are close to each other.<br />
Since these tracing points has higher concentratation<br />
<strong>of</strong> moisture content with higher variations, the sensor<br />
nodes can sense the higher variation <strong>of</strong> tracing points (or<br />
reference values )at different locations in sensor field .<br />
There may be higher or lower variations <strong>of</strong> data<br />
(moisture) measurement in spatial domain where the data<br />
are spatially correlated in the sensor field. Thus if the<br />
distance from the tracing point to the sensor nodes<br />
increases, the variations <strong>of</strong> the data correlation also get<br />
decreases.<br />
We represent a single tracing point where S i for i=1<br />
sensed by the sensor nodes Si and Sj where they sense and<br />
do measurement over a window frame <strong>of</strong> time T to<br />
capture the continuous data sample with Si={ si1 , si2, si3,<br />
……..sin } and Sj={sj1 , sj2, sj3, ……..sjn} respectively. The<br />
data correlation is strong when the tracing point is sensed<br />
by the sensor nodes Si and Sj located near to each other.<br />
The data correlation decreases as sensor node Si and Sj are<br />
far apart from tracing point. We compute the mean <strong>of</strong> the<br />
sampled data <strong>of</strong> sensor nodes as follows<br />
_ n 1<br />
Si= ∑ sik<br />
n k = 1<br />
and<br />
_ n 1<br />
Sj= ∑ sjk<br />
n k = 1<br />
Variance <strong>of</strong> the sample data collected by nodes Si and Sj<br />
can be given as<br />
n _<br />
1<br />
2<br />
var( S ) = ( s −Si)<br />
− ∑ (1)<br />
i ik<br />
n 1 k = 1<br />
© 2011 ACADEMY PUBLISHER<br />
And<br />
n<br />
_<br />
1<br />
2<br />
var( S ) = ( s −S<br />
j )<br />
− ∑ (2)<br />
j jk<br />
n 1 k = 1<br />
The covariance is given as<br />
n _ _<br />
1<br />
cov( Si, Sj) = ( sik −Si)( sjk −S<br />
j)<br />
( n −1) k=<br />
1<br />
∑ (3)<br />
The correlation coefficient ( ρ S , S)<br />
for correlation<br />
i j<br />
between data sensed by the sensor nodes Si and Sj for the<br />
tracing points can be given by<br />
cov( Si, S j)<br />
ρ Si, S =<br />
j var( S ).var( S )<br />
ρ<br />
i j<br />
1<br />
n _ _<br />
∑(<br />
sik −Si)( sjk−Sj) ( n−1)<br />
k=<br />
1<br />
Si, Sj n _ 1 2<br />
∑( sik −Si) n−1k= 1<br />
n _ 1<br />
2<br />
∑(<br />
sjk −Sj)<br />
n−1k=<br />
1<br />
=<br />
⎡ ⎤⎡ ⎤<br />
⎢ ⎥⎢ ⎥<br />
⎣ ⎦⎣ ⎦<br />
The equation-no 4 shows the data correlation<br />
coefficient for nodes Si and Sj in the spatial domain.<br />
Similarly from the co-variance model [16], we get the<br />
correlation coefficient ( ρ Si, S)<br />
for the data in spatial<br />
j<br />
domain.<br />
2 2<br />
[ i, j] = cov[ i, j] = σ i [ i, j] = σ i.<br />
ρ[<br />
i, j]<br />
S S<br />
(4)<br />
ES S S S corrS S S S<br />
cov[ S , S ] E[ S , S ]<br />
ρ[<br />
Si, Sj]<br />
= = (5)<br />
σ σ<br />
i j i j<br />
2 2<br />
i<br />
S<br />
i<br />
S<br />
Again from the power exponential model [16,17], we<br />
get the correlation coefficient function between node Si<br />
(xi, yi) and node Sj (xj, yj) as follows<br />
ρ[<br />
S , S ] e<br />
⎛ d ⎞<br />
−⎜ ⎟<br />
θ2<br />
⎝θ1⎠ = (6)<br />
i j<br />
We define a threshold τ which can determine whether<br />
the data are spatially correlated among the sensor nodes<br />
to trace the higher variations <strong>of</strong> data (called as tracing<br />
points) in the spatial domain. θ is called a ‘Range<br />
1<br />
parameter’ which controls how fast the spatially<br />
correlated data decays with the distance. θ is called a<br />
2<br />
‘Smoothness parameter’ which controls the geometrical<br />
properties <strong>of</strong> wireless sensor field.<br />
If ρ[ S , S ] ≥ τ , Data are strongly correlated in<br />
i j<br />
spatial domain for nodes Si and Sj .<br />
If ρ[ Si, S j]<br />
< τ , Data are weakly correlated in spatial<br />
domain for nodes Si and Sj.<br />
From equation no. (4), (5) and (6), we can derive the<br />
correlation coefficient ρ [ Si, S j]<br />
<strong>of</strong> data for nodes Si and<br />
Sj represented as follows:
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1075<br />
θ2<br />
⎛d⎞ cov[ S , S ] −<br />
i j<br />
⎜<br />
θ<br />
⎟<br />
1<br />
ρ[<br />
S<br />
i<br />
, S<br />
j<br />
] = = e<br />
⎝ ⎠<br />
2<br />
σ<br />
Si<br />
(7)<br />
When the data are strongly correlated for nodes Si and<br />
Sj in the spatial domain we have<br />
θ2<br />
⎛ d ⎞<br />
cov[ Si, S ] −⎜ ⎟<br />
j ⎝θ1⎠ [ Si, Sj] e<br />
2<br />
σ i<br />
S<br />
ρ = = ≥τ<br />
From the equation no (8), we can derive as following<br />
or<br />
or<br />
or<br />
θ2<br />
⎛ d ⎞<br />
−⎜ ⎟<br />
⎝θ1⎠ ≥<br />
e τ<br />
2<br />
d<br />
log( )<br />
1<br />
θ<br />
⎛ ⎞<br />
−⎜ ⎟ ≥ τ<br />
⎝θ⎠ 2<br />
d 1<br />
log<br />
1<br />
θ<br />
⎛ ⎞ ⎛ ⎞<br />
⎜ ⎟ ≤ ⎜ ⎟<br />
⎝θ⎠ ⎝τ⎠ d<br />
2<br />
2 2 ⎛ 1<br />
θ ⎛ ⎞⎞<br />
≤ θ 2<br />
1 log ⎜ ⎟<br />
(8)<br />
⎜<br />
τ<br />
⎟<br />
⎝ ⎝ ⎠⎠<br />
(9)<br />
where the Euclidean distance between the node Si (xi, yi)<br />
and node Sj (xj, yj) as follows<br />
d = ( x − x ) + ( y − y )<br />
2 2 2<br />
i j i j<br />
Put the value <strong>of</strong><br />
2<br />
d in equation no. (9) ,we get<br />
2 2 2 ⎛ 1<br />
θ ⎛ ⎞⎞<br />
( x ) ( ) 2<br />
i − xj + yi − yj ≤θ1⎜log⎜ ⎟<br />
τ<br />
⎟<br />
⎝ ⎝ ⎠⎠<br />
(10)<br />
Compare equation no. (10) with equation <strong>of</strong> circle<br />
with cluster head at the centre with the radius <strong>of</strong> the<br />
cluster r, we get<br />
2 2 2<br />
( x − x ) + ( y − y ) = r<br />
(11)<br />
i j i j<br />
From equation no. (10) and (11) , we get<br />
r<br />
θ<br />
θ ≤ 2<br />
1 log ⎜ ⎟<br />
2 2<br />
2<br />
⎛ ⎛1⎞⎞ ⎜<br />
τ<br />
⎟<br />
⎝ ⎝ ⎠⎠<br />
(12)<br />
The equation no. (12), shows the relation between the<br />
radius <strong>of</strong> the cluster and the threshold value <strong>of</strong> spatially<br />
correlated data. The radius <strong>of</strong> the cluster depends upon<br />
the threshold value τ , θ1 andθ 2 .If the value <strong>of</strong> threshold<br />
τ increases, the radius <strong>of</strong> the cluster from the CH node<br />
located at the centre <strong>of</strong> the cluster get decreases. So we<br />
have taken the appropriate value <strong>of</strong> θ 1 , θ2 and the<br />
threshold valueτ to maintain a good correlation <strong>of</strong> data<br />
between sensor nodes for the clusters.<br />
© 2011 ACADEMY PUBLISHER<br />
2<br />
B. Distributed Cluster Formation in Spatial Domain<br />
We consider a square field <strong>of</strong> area with Z=Z1 x Z2<br />
where the cluster head (CH) node are deterministically<br />
deployed uniformly and the normal nodes are deployed<br />
randomly in the sensor field Z which form the distributed<br />
cluster. Since the number <strong>of</strong> cluster head node deployed<br />
in the sensor region is known, we get the same number <strong>of</strong><br />
clusters as the number <strong>of</strong> cluster head nodes. We are<br />
interested in measuring the moisture content pr<strong>of</strong>ile in<br />
each cluster embedded in the sensor field Z. Thus we<br />
assume that every cluster has a single tracing point. Every<br />
cluster in the sensor field is responsible for sensing and<br />
measuring the physical phenomenon <strong>of</strong> data for the<br />
tracing point value. The highly correlated data among the<br />
sensor nodes and the CH node forms the cluster. The CH<br />
node located at the centre <strong>of</strong> each cluster performs the<br />
estimation <strong>of</strong> data accuracy and finally send the data to<br />
the sink node. The number <strong>of</strong> tracing points is equal to<br />
the number <strong>of</strong> cluster head nodes. Hence in our model<br />
numbers <strong>of</strong> sensor (normal) nodes are considered to be<br />
more than the number <strong>of</strong> cluster head nodes.<br />
=Tracing point in each distributed cluster<br />
=Cluster head node in each distributed cluster<br />
Figure 1: Overlapping clusters in sensor region<br />
Thus in the square sensor field Z ,every cluster are<br />
embedded in the sensor field which are capable to sense<br />
their respective tracing point (to measure the high<br />
variation <strong>of</strong> correlated data ) distributed uniformly as<br />
shown in Figure-1.Thus the known number <strong>of</strong> clusters<br />
formed in the sensor region Z can be represented as N as<br />
follows<br />
⎛⎢Z1 ⎥⎢Z 2 ⎥ ⎢Z1 ⎥⎢Z 2 ⎥⎞<br />
⎜⎢ 1 1<br />
2r ⎥⎢2r ⎥<br />
+<br />
⎢<br />
+ + ≤<br />
2r ⎥⎢2r ⎥⎟<br />
Number <strong>of</strong><br />
⎝⎣⎦⎣ ⎦ ⎣ ⎦⎣ ⎦⎠<br />
⎛ ⎡ Z1 ⎤⎡Z 2 ⎤ ⎡Z1 ⎤⎡Z 2 ⎤⎞<br />
Clusters ≤ ⎜ ⎢<br />
1 1<br />
2r ⎥⎢2r ⎥<br />
+<br />
⎢<br />
+ +<br />
2r ⎥⎢2r ⎥⎟<br />
⎝ ⎢ ⎥⎢ ⎥ ⎢ ⎥⎢ ⎥⎠<br />
Since the sensor field is square, Z1=Z2=W
1076 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
⎢ ⎥<br />
⎢ ⎥<br />
⎢ 2<br />
⎥<br />
⎢ W W ⎥<br />
+ + 1<br />
⎢ ⎛ 2 ⎞ θ 1<br />
⎥<br />
⎢ 2θ 1<br />
2 2 θ 2<br />
⎛ ⎞<br />
⎜ ⎛ ⎛ ⎞⎞<br />
log ⎥<br />
θ<br />
1<br />
log ⎟ 1 ⎜ ⎟<br />
⎢ ⎜ ⎜ ⎜ ⎟<br />
τ<br />
τ<br />
⎟ ⎟<br />
⎝ ⎠ ⎥<br />
⎢ ⎜ ⎝ ⎝ ⎠⎠<br />
⎟<br />
⎣ ⎝ ⎠<br />
⎥⎦<br />
≤ Number _ <strong>of</strong><br />
⎡ ⎤<br />
⎢ ⎥<br />
⎢ ⎥<br />
⎢ W 2<br />
W<br />
⎥<br />
Clusters≤<br />
⎢ + + 1 ⎥<br />
⎢ ⎛ 2 ⎞ θ<br />
2<br />
1<br />
2 θ<br />
⎛ ⎞ ⎥<br />
⎢ ⎜ 2 ⎛ ⎛ 1 ⎞ ⎞ ⎟ θ log<br />
2 θ log<br />
1 ⎜ ⎟<br />
τ<br />
⎥<br />
⎢ ⎜ 1 ⎜ ⎜ ⎝ ⎠<br />
τ ⎟ ⎟ ⎟<br />
⎝ ⎝ ⎠ ⎠<br />
⎥<br />
⎢ ⎜ ⎟<br />
⎢ ⎝ ⎠<br />
⎥<br />
(13)<br />
The equation no. (13) shows the relation between the<br />
number <strong>of</strong> clusters and the threshold used for data<br />
correlation. If the threshold increases, the number <strong>of</strong><br />
clusters with in the sensor field will get increases and<br />
vice versa. Thus we should choose appropriate threshold<br />
for clusters to perform data correlation in the spatial<br />
domain. Since the data are spatially correlated among the<br />
sensor nodes, there exist overlapping <strong>of</strong> clusters in the<br />
sensor region Z as shown in Figure 1. Equations no (12)<br />
and (13) derives how the clusters are overlapped among<br />
them in the sensor region Z. Hence it is important to find<br />
a distributed algorithm for clusters that can separate out<br />
the clusters from each other in the sensor region.<br />
Overlapping <strong>of</strong> cluster can sense the same correlated data<br />
among the sensor nodes and send the overlapped data to<br />
the sink node. It is like utilizing the same resource among<br />
the sensor nodes .Hence it leads to wastage <strong>of</strong> energy<br />
resource among the clusters and increases the data<br />
redundancy. Here we propose a distributed algorithm for<br />
cluster to overcome this problem for spatially correlated<br />
data and form non-overlapped irregular clusters in the<br />
sensor region Z.<br />
______________________________________________<br />
Algorithm I: Distributed clustering algorithm for<br />
spatially correlated data in sensor field Z.<br />
______________________________________________<br />
• Let U be the set <strong>of</strong> cluster head (CH) nodes<br />
deterministically deployed uniformly in sensor<br />
region Z.<br />
• Let V be the set <strong>of</strong> sensor (normal) nodes<br />
randomly deployed in sensor region Z.<br />
• Let d(a,b) be the Euclidian distance between<br />
node a and b.<br />
• Let dv be the distance from node v to the nearest<br />
CH node.<br />
• Initialize dv= ∞<br />
• Initialize CHv=0<br />
• for v ∈V<br />
• for u ∈U<br />
• if d(v,u)
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1077<br />
Node<br />
2<br />
Here we consider the mathematical analysis for data<br />
accuracy for the single cluster in the sensor region Z.<br />
Thus every cluster distributed in the sensor region can<br />
verify its data accuracy before data aggregation at the CH<br />
node. Once this procedure is being done by CH nodes for<br />
all the distributed clusters, they send the data to the sink<br />
node. Each sensor node i in the distributed cluster M can<br />
measure and observe the physically sensed data Si for<br />
tracing point S with observation noise Ni. Hence the<br />
observation and measurement made by the sensor node i<br />
in a given cluster is given by<br />
X = S + N where i ∈ M (14)<br />
i i i<br />
The sensor node i can sense and measure the observe<br />
sample Xi and transmits Xi to cluster head node sharing<br />
wireless additive white Gaussian noise (AWGN) channel<br />
[9,14]. Hence the observation and measurement received<br />
by the CH node from other sensor nodes in the cluster<br />
with transmission noise N over the AWGN channel is<br />
ti<br />
given by<br />
Node<br />
1<br />
Node<br />
m-1<br />
S<br />
Y = X + N<br />
2 2 t2<br />
Y = X +<br />
1 1<br />
N<br />
t1<br />
X<br />
CH<br />
= S<br />
CH<br />
+ N<br />
CH<br />
Y = X +<br />
m−1 m−1<br />
N<br />
tm<br />
Figure 2: Data accuracy model for distributed cluster<br />
Y = X + N = S + N + N<br />
i i ti i i ti<br />
Where i ∈ M and i ∉CH (15)<br />
We adopt uncoded transmission with finite number <strong>of</strong><br />
sensor nodes for optimal point-to-point transmission [10]<br />
and consider the encoding power constraint value P, the<br />
measured value received by the CH are given by<br />
P<br />
(16)<br />
Z = Y = α(<br />
S + N + N )<br />
i 2 2 2 i i i ti ( σ + σ + σ )<br />
Si Ni Nti<br />
where i ∈ M and i ∉CH<br />
P<br />
and α = 2 2 2<br />
( σ + σ + σ )<br />
Si Ni Nti<br />
CH node can sense and measure the tracing point S by<br />
finding the estimate <strong>of</strong> each physical phenomenon Si for<br />
node i. We take minimum mean square estimation<br />
(MMSE) for optimal decoding phenomenon [15] for<br />
uncoded transmission .CH node can find the MMSE for<br />
sensing and measuring the physical phenomenon Si<br />
© 2011 ACADEMY PUBLISHER<br />
−1<br />
CH<br />
extracted by sensor node i with observed sample Zi<br />
represented as<br />
ˆ<br />
E[ S Z ]<br />
S = Z<br />
i i<br />
i 2<br />
E[ Z ] i<br />
i<br />
where i ∈ M and i ≠ CH (17)<br />
Since the sensor node i can sense and measure the<br />
physical phenomenon Si <strong>of</strong> S , we take independent<br />
identically distributed (i.i.d) Gaussian random variable<br />
2<br />
2<br />
σ i.e E[S]=0 , var[S]= σ<br />
S<br />
S<br />
with zero mean and variance<br />
for tracing points . Similarly for sensing and measuring<br />
2<br />
phenomenon <strong>of</strong> Si, we assume E[Si]=0 , var[Si]= σ .<br />
We also have taken the observation noise Ni and<br />
transmission noise N ti with an independent identically<br />
distributed Gaussian random variable with variances<br />
2<br />
σ<br />
2<br />
, σ respectively with zero means.<br />
Ni Nti<br />
Hence E[Ni]=0 ,<br />
E[ N ]=0,var[Ni]= 2<br />
σ ,var[ N ]= 2<br />
σ respectively.<br />
t i<br />
Thus,<br />
N i<br />
2<br />
= i i Si ESZ [ ] ασ<br />
t i<br />
N ti<br />
2 2 2 2 2<br />
= + +<br />
i Si Ni Nti<br />
EZ [ ] α ( σ σ σ )<br />
Thus the estimation <strong>of</strong> ˆ Si is given by<br />
2<br />
σ Si = + +<br />
i 2 2 2 i i ti<br />
( σ + σ + σ )<br />
Si Ni Nti<br />
Sˆ ( S N N )<br />
2<br />
Si +<br />
2<br />
σ Si 2<br />
Ni +<br />
2<br />
Nti<br />
β = i<br />
( σ σ σ )<br />
where i ∈ M and i ∉ CH<br />
Si<br />
(18)<br />
for 0< β
1078 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Thus the estimation <strong>of</strong> ˆ SCH is given by<br />
Where β<br />
σ<br />
Sˆ ( S N )<br />
CH<br />
=<br />
2<br />
S CH<br />
2 2<br />
( σ + σ<br />
SCH NCH<br />
)<br />
CH<br />
+<br />
CH<br />
CH<br />
=<br />
σ<br />
2<br />
SCH 2<br />
S CH<br />
+<br />
2<br />
N CH<br />
( σ σ )<br />
(21)<br />
for 0< β
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1079<br />
0 at d= ∞ . We have taken power exponential model [17]<br />
i.e.<br />
PE .<br />
V i, j<br />
K ( d ) =<br />
e<br />
θ<br />
−(<br />
d 2<br />
i, j/<br />
θ1)<br />
, θ θ 1 2<br />
> 0; ∈ (0, 2] θ is the<br />
1<br />
‘Range parameter’ and θ is the ‘Smoothness parameter’.<br />
2<br />
Using (15) and (23) in (25), we perform the normalized<br />
data accuracy with spatial correlation model for every<br />
distributed cluster in the sensor region given as follows:<br />
1 ⎡ M −1<br />
θ<br />
( / ) 2<br />
θ ⎤<br />
−d ( / ) 2<br />
( ) (2 ( Si , )<br />
θ<br />
1<br />
−d<br />
D M 1) 2<br />
( SCH , )<br />
θ<br />
= ⎢β 1<br />
A<br />
i ∑ e − + βCHe<br />
⎥<br />
m ⎢⎣ ⎥<br />
i = 1<br />
⎦<br />
⎡ ⎤<br />
1 ⎢ ⎥<br />
−<br />
M−1M−1 θ M−1<br />
( / ) 2<br />
θ<br />
( / ) 2<br />
2 ⎢ββ ( −d (,) ij<br />
θ<br />
1<br />
−d<br />
i 1) (2<br />
( CHi ,)<br />
θ<br />
1<br />
i∑∑e −+ βCH βi∑e + βCH)<br />
⎥<br />
⎢ ⎥<br />
⎣ i= 1 j≠ i i=<br />
1<br />
⎦<br />
m<br />
Node<br />
1<br />
ρ(<br />
S , S )<br />
CH<br />
1<br />
ρ(<br />
S , S )<br />
i j<br />
ρ(<br />
S , S )<br />
CH<br />
CH<br />
Node<br />
2<br />
ρ(<br />
SS , )<br />
(26)<br />
The equation no. (26) shows that the normalized data<br />
accuracy D ( M ) for each cluster depends upon m sensor<br />
A<br />
nodes and factors i β and β respectively. Since we get<br />
CH<br />
a normalized data accuracy at each CH node for each<br />
cluster, we construct a spatial correlation model given by<br />
equation no. (26) for each individual distributed cluster in<br />
the sensor region. The spatial correlation model for each<br />
distributed cluster can be explained as follows:<br />
� Each sensor node i can sense a tracing point<br />
S in each distributed cluster where i ∈ M<br />
and i ∉CH node<br />
� CH node itself can sense the tracing point S<br />
in each distributed cluster.<br />
CH<br />
ρ(<br />
S , S )<br />
i j<br />
ρ(<br />
S , S )<br />
ρ(<br />
S , S )<br />
i j<br />
CH<br />
ρ(<br />
SS , )<br />
1<br />
ρ(<br />
SS , )<br />
2<br />
Node<br />
3<br />
ρ(<br />
SS , )<br />
3<br />
= Data sensed by node i from point event<br />
= Spatial data correlation between node i ,j<br />
= Data transmitted to the CH node<br />
= Data transmitted to sink node<br />
Figure 3: Spatial correlation model for distributed cluster<br />
© 2011 ACADEMY PUBLISHER<br />
2<br />
Sink<br />
3<br />
S<br />
� A spatial correlation between node i, j in<br />
each distributed cluster where i,j ≠ CH<br />
node.<br />
� Each sensor node i transmits the sensed data<br />
to the CH node in each distributed cluster<br />
where i ∈ M and i ∉ CH.<br />
Thus each distributed cluster formed in the sensor<br />
region has different set <strong>of</strong> sensor nodes. Hence each<br />
cluster can perform the normalized data accuracy at the<br />
CH node before data aggregation. The purpose <strong>of</strong><br />
verifying the data accuracy for each cluster is to confirm<br />
that the most accurate data send by m set <strong>of</strong> sensor nodes<br />
can aggregate at the CH node rather than aggregating all<br />
the redundant data at the CH node. To visualize the<br />
correlation model for distributed cluster, we take an<br />
example where m=4 sensor nodes and out <strong>of</strong> m sensor<br />
nodes one node is chosen as a CH node as shown in<br />
Figure 3. Once we estimate the data accuracy at the CH<br />
node for each distributed cluster, the most accurate data<br />
get aggregated and finally send to the sink node.<br />
III. SIMULATION RESULTS<br />
In the first simulation setup , twenty five CH nodes are<br />
deterministically deployed uniformly and hundred sensor<br />
(normal) nodes are deployed randomly in a wireless<br />
sensor field <strong>of</strong> 120 m X 120 m based sensor topology as<br />
shown in Figure 1 . Each CH node performs the data<br />
accuracy for their respective cluster. Hence each cluster<br />
can sense and measure a single tracing point randomly<br />
located in each cluster region. Once each cluster can<br />
sense and measure their respective tracing point, it<br />
performs the data accuracy at CH node and finally<br />
transmits the data to the sink node.<br />
#CH Associated Nodes (Normal Nodes) Data<br />
Nodes<br />
Accuracy<br />
CH1 2 21 59 78<br />
0.837847<br />
CH2 1 6 7 8 13 35 43 76 92<br />
93<br />
0.843960<br />
CH3 11 17 22 69 84 98 0.866797<br />
CH4 10 46 62<br />
CH5 4 40 58 87<br />
CH6 15 25 32 33 53 73 81<br />
CH7 36 41 57 61 74 80 83<br />
CH8 13 19 28 31 49 85 95<br />
CH9 9 29 37 38<br />
0.694458<br />
0.833017<br />
0.820673<br />
0.862045<br />
0.793657<br />
0.882088<br />
CH10 20 23 63 75 77 79 88 91 97 0.857425<br />
CH11 44 51 66 86 99<br />
CH12 45 50 55 89<br />
CH13 5 18 24 47 48 52 82<br />
CH14 27 30 34 39 71 100<br />
0.820772<br />
0.809979<br />
0.813055<br />
0.756650
1080 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
CH15 26 60 72<br />
CH16 70<br />
CH17 65 94<br />
CH18 68 90<br />
0.787127<br />
0.714302<br />
0.854421<br />
0.873163<br />
CH19 42 0.705224<br />
CH20 56 0.759352<br />
CH21 67 0.730805<br />
CH22 12 0.799681<br />
CH23 96 0.739846<br />
CH24 14 16 0.894157<br />
CH25 54 64 0.843685<br />
Table 1: Data Accuracy for each distributed cluster<br />
According to our proposed algorithm-I discuss<br />
previously, each CH node can form the cluster with their<br />
associated sensor nodes. Once the sensor nodes are<br />
associated with each CH node, they form distributed<br />
clusters in the sensor region Z. Thus twenty five CH<br />
nodes can form twenty five individual non-overlapping<br />
distributed clusters. Each distributed cluster can perform<br />
the data accuracy at their respective CH node as shown in<br />
Table-1. Similarly in the second simulation set up as<br />
shown in Table-2, we perform hundred runs for each CH<br />
nodes associated with their respective sensor nodes and<br />
find their average data accuracy for each cluster.<br />
#CH<br />
Nodes<br />
Average<br />
Data Accuracy<br />
#CH<br />
Nodes<br />
Average<br />
Data Accuracy<br />
CH1 0.8494 CH14 0.7327<br />
CH2 0.8731 CH15 0.7778<br />
CH3 0.8765 CH16 0.9662<br />
CH4 0.8734 CH17 0.8001<br />
CH5 0.8468 CH18 0.7706<br />
CH6 0.8401 CH19 0.9662<br />
CH7 0.8364 CH20 0.8111<br />
CH8 0.7975 CH21 0.8135<br />
CH9 0.9033 CH22 0.9736<br />
CH10 0.8615 CH23 0.9047<br />
CH11 0.7942 CH24 0.8343<br />
CH12 0.8171 CH25 0.8352<br />
CH13 0.7796<br />
Table 2: Average Data Accuracy for each distributed cluster<br />
© 2011 ACADEMY PUBLISHER<br />
In the third simulation set up, we take a single circular<br />
cluster <strong>of</strong> m=4 sensor nodes which can sense and<br />
measure a tracing point. We put m sensor nodes in a<br />
deployed circular cluster and a tracing point S located at<br />
the centre <strong>of</strong> the deployed circular cluster. i.e dS,i (where<br />
i=1,2,3)and dS,CH are equidistance as shown in the Figure-<br />
4. Here we have fixed the number <strong>of</strong> m sensor nodes and<br />
vary the distance from the tracing point S to m sensor<br />
nodes. As we increase the radius <strong>of</strong> the deployed circular<br />
cluster for dS,i and dS,CH with same proportion , D ( M ) A<br />
decreases i.e. the distance from the tracing point S to the<br />
m sensor nodes increases as shown in Figure 5. We put<br />
θ = {50,100} and θ =1 for our statistical data<br />
1<br />
2<br />
performance for the normalized data accuracy DA( M ) .<br />
Data Accuracy D A (M)<br />
Node<br />
3<br />
0.96<br />
0.94<br />
0.92<br />
0.9<br />
0.88<br />
0.86<br />
0.84<br />
0.82<br />
0.8<br />
CH<br />
Node<br />
Node<br />
2<br />
Node<br />
1<br />
Figure 4: Deployed sensor nodes in circular cluster topology<br />
S<br />
m=4 ,θ 1 =50<br />
m=4,θ 1 =100<br />
0.78<br />
1 2 3 4 5 6 7 8 9 10<br />
Radius <strong>of</strong> the deployed circle<br />
Figure 5: Data accuracy versus radius <strong>of</strong> the circular cluster<br />
In the fourth simulation setup, the distance from the<br />
tracing point S to m sensor nodes is fixed in the deployed<br />
circular cluster <strong>of</strong> radius =5 metre. We increase the<br />
number <strong>of</strong> sensor nodes with a fixed distance from the<br />
tracing point S i.e we increase m sensor nodes with fixed<br />
deployed circular cluster <strong>of</strong> radius 5 metre. At first, we<br />
put m=2 (one CH node and one sensor node) which
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1081<br />
shows that the data accuracy is very poor with its value in<br />
between 0.6 to 0.75 for θ ={50,100,200,400}.The reason<br />
1<br />
is that there is only one sensor node which shows that the<br />
third condition <strong>of</strong> spatial correlation model given in<br />
section II(D) doesn’t satisfies the DA( M ) at the CH node .<br />
But if we put m=3 (one cluster head and two sensor<br />
nodes), there is a drastic improvement <strong>of</strong> DA( M ) since all<br />
the conditions for spatial correlation model are satisfied.<br />
The Figure-6 also shows that five to eight nodes are<br />
sufficient to perform the D ( M ) for the cluster, if the<br />
A<br />
distance from tracing point to m sensor nodes with<br />
deployed circular cluster <strong>of</strong> radius is 5 metre.<br />
For the simplicity <strong>of</strong> our model, we perform the fifth<br />
simulation set up where we have simulated a wireless<br />
sensor field (900 metre 2 ) <strong>of</strong> 5m X 5m grid based single<br />
cluster topology with a fixed tracing point (S) at the<br />
centre and a CH node on the corner edge with 47 sensor<br />
nodes distributed uniformly in the grid based cluster<br />
topology as shown in Figure 7.Our assumptions is that<br />
cluster <strong>of</strong> m sensor nodes are in the sensing range <strong>of</strong> the<br />
tracing point (S). Initially we put m=4(one cluster head<br />
node and three sensor nodes located at the four extreme<br />
corner <strong>of</strong> sensor field).We verified that D ( M = 4) is<br />
A<br />
0.6333 when θ =50 as shown in Figure 8. If we increase<br />
1<br />
θ = 400, then D ( m = 4) =0.911.<br />
1<br />
Data Accuracy D A (M )<br />
1<br />
0.95<br />
0.9<br />
0.85<br />
0.8<br />
0.75<br />
0.7<br />
0.65<br />
A<br />
θ 1 =50<br />
θ 1 =100<br />
θ 1 =200<br />
θ 1 =400<br />
2 3 4 5 6 7 8 9 10 11 12<br />
Number <strong>of</strong> Sensor Nodes<br />
Figure 6: Data accuracy versus number <strong>of</strong> sensor nodes in a<br />
cluster<br />
This shows that θ control as how fast the spatially<br />
1<br />
correlated data decays with distance between sensor<br />
nodes and the tracing point. Hence it is always suitable to<br />
take the value <strong>of</strong> θ large for large sensor field to get<br />
1<br />
DA( M ) in an efficient way. Now we increase cluster <strong>of</strong> m<br />
sensor nodes with increment <strong>of</strong> four sensor nodes every<br />
time concentrating towards tracing point till m sensor<br />
nodes are able to sense and measure the tracing point S in<br />
© 2011 ACADEMY PUBLISHER<br />
the region. As we increase the sensor nodes, the data<br />
accuracy DA( M ) also get increases. Hence for 900 metre 2<br />
sensor field, 15 to 20 sensor nodes are sufficient to give<br />
DA( M ) <strong>of</strong> 0.944 for θ =400 and 1<br />
DA( M ) remains<br />
approximately constant still we increase the number <strong>of</strong><br />
sensor nodes for the cluster. We plot in the Figure-8 for<br />
the DA( M ) versus node density for a cluster. Node density<br />
is defined as the number <strong>of</strong> sensor nodes per unit area in a<br />
single cluster. Hence it is needless to choose so many<br />
sensor nodes to achieve data accuracy for the cluster in<br />
sensor field to sense and measure a tracing point.<br />
Data Accuracy D A (M )<br />
1<br />
0.95<br />
0.9<br />
0.85<br />
0.8<br />
0.75<br />
0.7<br />
0.65<br />
CH node<br />
θ 1 =50<br />
θ 1 =100<br />
θ 1 =200<br />
θ 1 =400<br />
0 0.01 0.02 0.03<br />
Node density<br />
0.04 0.05 0.06<br />
Figure 8: Data accuracy vs. node density in a single cluster<br />
In the sixth simulation setup, we take a single cluster <strong>of</strong><br />
m sensor nodes randomly deployed in a region (30 X 30<br />
= 900 metre 2 ) that sense and measure a tracing point. We<br />
fix the tracing point at x,y (15,15) coordinate and CH<br />
node at x,y (0,0) coordinate with 99 sensor nodes<br />
S<br />
Figure 7: Sensor nodes deployed in grid topology
1082 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
randomly deployed in the region. For each run we verify<br />
DA( M ) with respect to randomly deployed cluster <strong>of</strong> m<br />
sensor nodes. Finally we verify for 100 runs and find the<br />
average DA( M ) for the cluster <strong>of</strong> m sensor nodes. Figure 9<br />
shows that if the value <strong>of</strong> θ =400, 1 DA( M ) is 0.944 for 10<br />
to 15 sensor nodes. If we continuously increase the<br />
number <strong>of</strong> sensor nodes the DA( M ) remains approximately<br />
same. Hence it is useless to deploy sensor nodes beyond<br />
15 sensor nodes because 10 to 15 sensor nodes are<br />
sufficient to give approximately the same DA( M ) for the<br />
cluster with 1 θ =400. Again if we constantly increasesθ , 1<br />
average DA( M ) also get increases for the cluster <strong>of</strong> m<br />
sensor nodes. But after certain approximate value <strong>of</strong> θ 1<br />
the DA( M ) remains approximately constant for the cluster.<br />
If we continuously increase the value <strong>of</strong> θ the average<br />
1<br />
DA( M ) remains approximately constant since it achieve the<br />
saturation level in the cluster. Finally the output graph<br />
shows distortion in the signal due to additive white<br />
Gaussian noise components.<br />
Average Data Accuracy D A (M)<br />
1<br />
0.95<br />
0.9<br />
0.85<br />
0.8<br />
0.75<br />
0.7<br />
0.65<br />
0.6<br />
θ 1 =50<br />
θ 1 =100<br />
θ 1 =200<br />
θ 1 =400<br />
0.55<br />
0 10 20 30 40 50 60 70 80 90 100<br />
Number <strong>of</strong> Sensor Nodes<br />
Figure 9: Average data accuracy versus number <strong>of</strong> sensor nodes in a<br />
single cluster<br />
Since the data are spatially correlated in the sensor<br />
region, we propose a distributed algorithm with non<br />
overlapping irregular cluster for the spatially correlated<br />
data in the sensor region. Each distributed cluster can<br />
perform DA( M ) before data aggregation at their respective<br />
CH node. Hence it is important to sense and measure the<br />
most appropriate (accurate) data send by each distributed<br />
cluster at the CH node rather than aggregating all the<br />
redundant data at their respective CH node. Thus it can<br />
reduce the data redundancy. Since the data accuracy is<br />
performed by each distributed cluster, we verified from<br />
the simulation results that there exists a minimal set <strong>of</strong><br />
sensor nodes with optimal cluster which is sufficient to<br />
© 2011 ACADEMY PUBLISHER<br />
give approximately the same DA( M ) as achieved by the<br />
each distributed cluster. Therefore the time complexity<br />
done at each CH node <strong>of</strong> respective distributed cluster for<br />
aggregating the most accurate data send by their<br />
respective optimal cluster will be less. Thus we find an<br />
optimal cluster from each distributed cluster which can<br />
reduce the data redundancy and communication<br />
overhead.<br />
In the fifth simulation setup, a grid based single cluster<br />
is formed where we deployed m=48 sensor nodes<br />
uniformly. We examine that 15 to 20 nodes are sufficient<br />
to perform DA( M ) =0.944 for 1 θ =400 in 900 metre2 cluster<br />
region. Similarly in sixth simulation setup a cluster with<br />
m=100 sensor nodes are randomly deployed in 900<br />
metre 2 region and we get 10 to 15 sensor nodes are<br />
sufficient to perform DA( M ) =0.944 for θ =400. Therefore<br />
1<br />
it is unnecessary to choose so many sensor nodes in 900<br />
metre 2 region as DA( M ) remains approximately same as it<br />
achieve the saturation level still we increase m sensor<br />
nodes in the cluster. Hence we have P minimal set <strong>of</strong><br />
sensor nodes with optimal cluster which is sufficient to<br />
give approximately the same DA( M ) by M set <strong>of</strong> sensor<br />
nodes in each distributed cluster as shown by Venn<br />
diagram in Figure 10.<br />
Figure 10: Venn diagram for optimal cluster in each distributed cluster<br />
IV . CONCLUSIONS<br />
In this paper we investigate that the data are spatially<br />
correlated among sensor nodes and form clusters in the<br />
sensor region. Since the data are highly correlated in the<br />
spatial domain, the sensor nodes form regular<br />
overlapping clusters among them in the sensor region.<br />
Overlapping <strong>of</strong> cluster can sense and measure the same<br />
correlated data among the clusters. Thus to overcome this<br />
situation, we constructed a distributed clustering<br />
algorithm with data accuracy model .We perform data<br />
accuracy for each distributed cluster. We find that the<br />
most accurate data send by the distributed cluster can<br />
aggregate at the CH node rather than aggregating all the<br />
redundant data at their respective CH node. We<br />
demonstrate by simulation that the data accuracy for a<br />
single cluster depend on number <strong>of</strong> sensor nodes and their<br />
exist an optimal cluster which is adequate to sense and<br />
measure the tracing point to perform approximately the<br />
same data accuracy level achieve by single cluster.<br />
Finally we conclude that the data accuracy performed for<br />
each distributed cluster can reduce the data redundancy<br />
and communication overhead.<br />
M<br />
P
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1083<br />
REFERENCES<br />
[1] I.F Akyuildz ,W.Su , Y. Sankarasubramanian and E.<br />
Cayirci, “A survey on sensor <strong>Networks</strong> ”,IEEE<br />
Communcations Magazine, vol .40 , pp.102-114 ,Aug<br />
2002.<br />
[2] S.S. Pradhan , K. Ramchandran ,“Distributed Source<br />
Coding : Symmetric rates and applications to sensor<br />
networks”, in procecding <strong>of</strong> the data compressions<br />
conference 2000,pp.363-372.<br />
[3] A. Abbasi and M. Younis , “A survey on clustering<br />
algoirthms for wireless sensor networks ” Computer<br />
communications , vol-30 , n-14-15 ,pp-2826-2841, 2007.<br />
[4] W.B Heinzelman , Anantha P. Chandrakasan, “An<br />
Application Specific Protocl Architecture for Wireless<br />
Microsensor <strong>Networks</strong> ” , IEEE transactions on wireless<br />
communications , vol. no 4 , Oct-2002.<br />
[5] Georgios Smaragdakis, Ibrahim Matta ,Azer Bestavros , “<br />
SEP : A stable Electon Prtocol for cluster hetrerogenous<br />
wireless sensor networks ”<br />
[6] Chongqing Zhang , Binguo wang , Shen Fang , Zhe Li , “<br />
Clustering Algorithms for wireless sensor networks using<br />
spatial data correlation ”, International conference on<br />
information and Automation , pp-53-58 ,june 2008.<br />
[7] Zhikui chen , Song Yang , Liang Li and Zhijiang Xie , “ A<br />
clustering Approximation Mechinism based on Data<br />
Spatial Correlation in Wireless sensor <strong>Networks</strong> “,<br />
Proceedings <strong>of</strong> the 9 th international conferenses on<br />
wireless telecommunication symposium -2010.<br />
[8] Ali Dabirmoghaddam , Majid Ghaderi , Carey Williamson<br />
, “Energy Efficient Clustering in wireless Sensor<br />
<strong>Networks</strong> with spatially correlated dara “ IEEE infocom<br />
2010 proceedings.<br />
[9] Kang Cai, Gang Wei and Huifang Li,“Information<br />
Accuracy versus Jointly Sensing Nodes in Wireless<br />
Sensor <strong>Networks</strong>” IEEE Asia Pacific conference on<br />
curcuit and systems 2008 ,pp.1050-1053.<br />
[10] M.Gastpar, M. Vetterli, “ Source Channel Communication<br />
in Sensor <strong>Networks</strong> ”, Second International Workshop on<br />
Information Processing in Sensor <strong>Networks</strong> (IPSN’2003).<br />
[11] Varun M.C,Akan O.B and I.F Akyildiz, “ Spatio-<br />
Temporal Correlation : Theory and Applications Wireless<br />
Sensor <strong>Networks</strong>” , Computer Network <strong>Journal</strong> (Elsevier<br />
Science ), vol. 45 , pp.245-259 , june 2004.<br />
[12] Jyotirmoy karjee , H.S Jamadagni , “Data Accuracy<br />
Estimation for Cluster with Spatially Correlatd Data in<br />
Wireless Sensor <strong>Networks</strong> ” ,to be published in the<br />
proccedings ICISCI-2011, Harbin ,China<br />
[13] Huifang Li, Shengming Jiang ,Gang Wei ,“Information<br />
Accuracy Aware Jointly Sensing Nodes Selection in<br />
Wireless Sensor <strong>Networks</strong> ”,MSN 2006 , LNCS 4325 ,<br />
pp.736-747.<br />
[14] T.J. Goblick ,“ Theoritical Limitions on the transmission<br />
<strong>of</strong> data from analong sources”,IEEE Transaction Theory ,<br />
IT-11 (4) pp.558-567 ,1965.<br />
[15] V.Poor ,“ An Introduction to Signal Detection and<br />
Estimation ”,Second edition , Springer ,Berlin 1994.<br />
[16] J.O. Berger , V.de Oliviera and B.Sanso ,“ Objective<br />
Bayesian Anylysis <strong>of</strong> Spatially correlated data<br />
”J.Am.Statist. Assoc. Vol-96,pp.1361-1374,2001.<br />
[17] De Oliveira V, Kedan B and Short D.A , “ Bayesian<br />
predication <strong>of</strong> transformed Gaussian random fields”<br />
<strong>Journal</strong> <strong>of</strong> American statistical Association 92, pp.1422-<br />
1433.<br />
[18] L.Guo , F chen , Z Dai , Z. Liu „”Wireless sensor network<br />
cluster head selection algorithm based on neural<br />
© 2011 ACADEMY PUBLISHER<br />
networks” , PP-258-260 , International conference on<br />
Machine vision and human machine interference, 2010.<br />
[19] T.Minming , N Jieru , W Hu , Liu Xiaowen “ A data<br />
aggregation Model for underground wireless sensor<br />
network” Vol-1, pp-344-348 , WRI world congress on<br />
computer science and information engineering, 2009 .<br />
[20] Jyotirmoy karjee , Sudipto Banerjee , “ Tracing the<br />
Abnormal Behavior <strong>of</strong> Malicious Nodes in MANET ”,<br />
Fourth International conference on wireless<br />
communications , networking and Mobile Computing<br />
,pp-1-7 Dalian-china -2008 .<br />
[21] C.Y. cho , C.L Lin , Y.H Hsiao , J S wang , K.C yong “<br />
Data aggegation with spatially correlated grouping<br />
Techninques on cluster based WSNs” , SENSORCOMM<br />
,pp-584-589, venice- 2010.<br />
[22] Shirshu Varma , Uma shankar tiwary , “ Data Aggregation<br />
in Cluster based wireless sensor <strong>Networks</strong> ”Proceedings<br />
<strong>of</strong> the first International confernce on Intelligent human<br />
computer interaction , page-391-400 , part-5 , 2009.<br />
Jyotirmoy Karjee received his B.E<br />
(Electronics), M.E (Information<br />
Technology) specialization in Network<br />
Security in 2003 and 2005 respectively.<br />
He worked in Prakriti Inbound Pvt. Ltd<br />
as a s<strong>of</strong>tware engineer for a year and<br />
worked as a lecturer in Sikkim Manipal<br />
Institute <strong>of</strong> Technology, Sikkim till<br />
2008. He is currently pursuring his<br />
Ph.D degree at Centre for Electronics<br />
Design and Technology, Indian Instutute <strong>of</strong> Science, Bangalore.<br />
His current research interests include data accuracy estimation<br />
and data aggregation in wireless sensor networks.<br />
Pr<strong>of</strong>. H.S Jamadagni received his<br />
M.E and Ph.D degree in Electrical &<br />
Communication Engineering from<br />
Indian Institute <strong>of</strong> Science ,Bangalore.<br />
Currently He is the pr<strong>of</strong>essor at Centre<br />
for Electronics Design and<br />
Technology, Indian Institute <strong>of</strong><br />
Science. He is one <strong>of</strong> the main<br />
coordinators for the intel higher<br />
education program and was the key mentors for various intel<br />
workshops in india. His current research work includes in the<br />
areas <strong>of</strong> embedded systems, VLSI for wireless networks and<br />
wireless sensor networks.
1084 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Networking as a Service: a Cloud-based Network<br />
Architecture<br />
Tao Feng, Jun Bi, Hongyu Hu and Hui Cao<br />
Network Research Center, Tsinghua University<br />
Department <strong>of</strong> Computer Science, Tsinghua University<br />
Tsinghua National Laboratory for Information Science and Technology (TNList), Beijing, China<br />
fengt09@mails.tsinghua.edu.cn, junbi@tsinghua.edu.cn, huhongyu@cernet.edu.cn and cao-h06@mails.tsinghua.edu.cn<br />
Abstract—With the rapid development and integration <strong>of</strong><br />
the Internet, wireless communication network and the<br />
Internet <strong>of</strong> Things, the Internet faces many challenges as a<br />
bearer network: a large volume <strong>of</strong> information exchange,<br />
multi-level QoS and smoothly switching multiple access<br />
protocols. The Internet should be able to provide a variety<br />
<strong>of</strong> network capacities in a more dynamic and on-demand<br />
way, not just limited network resource provision through<br />
virtualization. The elastic network is expected to adapt to<br />
network changes by enabling network protocols selection<br />
and combination dynamically. Cloud computing illustrates a<br />
new Internet-based model <strong>of</strong> IT resources (hardware,<br />
s<strong>of</strong>tware, data) provision, delivery and consumption as a<br />
service. Therefore, networking as a service can provide<br />
guaranteed quality <strong>of</strong> service and good quality <strong>of</strong> experience<br />
to users who do not care about any network configuration<br />
and network management. In this paper, we propose a novel<br />
idea <strong>of</strong> networking as a service by combining the service<br />
provision model <strong>of</strong> cloud computing with the openness <strong>of</strong> the<br />
network protocol. The related conception and stakeholders<br />
<strong>of</strong> networking as a service is depicted. Cloud-based network<br />
architecture is design to present the provision, delivery and<br />
consumption <strong>of</strong> networking as a service and discuss the key<br />
features <strong>of</strong> cloud-based network. Finally, a prototype <strong>of</strong><br />
cloud-based network is implemented by extending<br />
OpenFlow architecture.<br />
Index Terms—network capacity; networking as a service;<br />
cloud computing; network architecture<br />
I. INTRODUCTION<br />
With the rapid development and integration <strong>of</strong> the<br />
Internet, wireless communication network and the<br />
Internet <strong>of</strong> Things, the Internet faces many challenges as<br />
a bearer network in the future: a large volume <strong>of</strong><br />
information exchange, multi-level QoS, smoothly<br />
switching multiple access protocols, mobility and<br />
management. The design philosophy <strong>of</strong> the current<br />
Internet [12] limits the flexibility <strong>of</strong> the network<br />
architecture to meet new requirements. For instance, the<br />
end-to-end argument proposes that a network simply<br />
forwards packets between end-systems while complex<br />
data processing function is implemented on the end-<br />
This is the extended version <strong>of</strong> our paper at ICISCI’10.<br />
Corresponding author: Tao Feng, fengt09@mails.tsinghua.edu.cn<br />
© 2011 ACADEMY PUBLISHER<br />
doi:10.4304/jnw.6.7.1084-1090<br />
systems [33]. At present, the Internet is expected to<br />
handle more complex and customized forwarding<br />
capacity in the company <strong>of</strong> more and more mobile<br />
devices connected and new client/server paradigm <strong>of</strong><br />
cloud computing. Even the current Internet provides a lot<br />
<strong>of</strong> new features or services that go beyond forwarding in<br />
order to deal with more tussles [13], such as<br />
heterogeneous network resources, personalized delivery<br />
service, trust network access and low-cost network<br />
maintenance, and this shift towards more network<br />
capacities will continue.<br />
New network features or services are difficult to be<br />
introduced into the current network because a network<br />
protocol is locked in a vendor device. Network service<br />
should be decoupled from specific data transport<br />
technologies so that new features or services can be<br />
deployed freely. In addition, the Internet should be able to<br />
provide a variety <strong>of</strong> network capacities in a more<br />
dynamic and on-demand way, not just limited network<br />
resource provision through virtualization [7]. The elastic<br />
network is expected to adapt to network changes by<br />
enabling network protocols selection and combination<br />
dynamically. Cloud computing [1] illustrates a new<br />
Internet-based model <strong>of</strong> IT resources (hardware,<br />
s<strong>of</strong>tware, data) provision, delivery and consumption as a<br />
service. Therefore, network capacity on demand can<br />
provide guaranteed quality <strong>of</strong> service and good quality <strong>of</strong><br />
experience to users who do not care about any network<br />
configuration and network management.<br />
There will be little place either for static network<br />
configurations like the current network stack or for<br />
manual optimization and tuning as enforced at the interlayer<br />
boundaries <strong>of</strong> the current network in such a<br />
dramatically dynamic network operational circumstance.<br />
On the opposite end, the revolution that removes such<br />
constraints and at the same time maximizes the ondemand<br />
capacity <strong>of</strong> network will play a key role in the<br />
evolution towards future network.<br />
In this paper, we propose a novel idea <strong>of</strong> networking as<br />
a service by combining the service provision model <strong>of</strong><br />
cloud computing with the openness <strong>of</strong> the network<br />
protocol. The related conception and stakeholders <strong>of</strong><br />
networking as a service is depicted. Cloud-based network<br />
architecture is design to present the provision, delivery<br />
and consumption <strong>of</strong> network protocol as a service and
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1085<br />
discuss the key features <strong>of</strong> cloud-based network. Finally,<br />
a prototype <strong>of</strong> cloud-based network is implemented by<br />
extending OpenFlow architecture.<br />
II. RELEATED WORK<br />
The design <strong>of</strong> the current Internet architecture has been<br />
rethinking and some architectural principles for new<br />
Internet architectures have been proposed for some years<br />
on structuring a new generation <strong>of</strong> network protocols<br />
[14], adding mechanisms to the core <strong>of</strong> the Internet [15]<br />
and exploring specific architectural issues [16]. The trend<br />
<strong>of</strong> the Internet towards a commercial development has<br />
changed the underlying hypotheses <strong>of</strong> trust and economic<br />
incentives [17].<br />
Several new features have been proposed and<br />
implemented in the Internet with the development <strong>of</strong> the<br />
Internet, which was not considered in the initial design <strong>of</strong><br />
network architecture. Since the Internet does not support<br />
the dynamic deployment <strong>of</strong> new protocols, on-demand<br />
composition <strong>of</strong> network protocols and pay-as-you-go<br />
business model <strong>of</strong> network capacity, these features<br />
needed to be added as special processing functions to<br />
future network. Some <strong>of</strong> the current research gives a ray<br />
<strong>of</strong> hope for network capacity on demand. The following<br />
list highlights such features and functions:<br />
A. The openness <strong>of</strong> the network protocol<br />
The openness <strong>of</strong> the network protocol refers to the<br />
future that the introduction, deployment and operation <strong>of</strong><br />
network protocol or service can be achieved on minimum<br />
cost by standardizing network interfaces and enhancing<br />
interoperability <strong>of</strong> network protocol. The openness <strong>of</strong> the<br />
network protocol is a prerequisite for network protocol as<br />
a service. The realization <strong>of</strong> the openness future makes<br />
the network protocol custom and flexible adoption<br />
according to different application scenarios and user<br />
requirements. Open architectures and analogous work on<br />
the openness <strong>of</strong> networks have contributed ideas for the<br />
programming behavior <strong>of</strong> a node [29, 30] and flow-driven<br />
modification <strong>of</strong> the data plane services [31]. To manage<br />
the complexity <strong>of</strong> new protocol in the network, a working<br />
group <strong>of</strong> IETF has attempted to define Open Pluggable<br />
Edge Services (OPES) [19]. In such architecture, a set <strong>of</strong><br />
data flow operations that are implemented on nodes<br />
throughout the network can be specified in end-systems.<br />
Currently, there are two ways to achieve the openness <strong>of</strong><br />
the network protocol in the control plane: the out-box<br />
openness and the in-box openness. OpenFlow [2] is one<br />
<strong>of</strong> the implementations <strong>of</strong> the out-box openness.<br />
OpenFlow provides a way to control network device by<br />
network protocols running outside <strong>of</strong> a network device.<br />
OpenFlow achieves a variety <strong>of</strong> network behaviors on the<br />
switch by controlling the flowtable such as routing,<br />
firewall, and so on. On the other hand, the JUNOS SDK<br />
[3] is another way to open network protocol. The JUNOS<br />
SDK enables developers to innovate on top <strong>of</strong> JUNOS<br />
and Juniper <strong>Networks</strong> platforms, so developers can<br />
create, deploy, and validate innovative applications<br />
tailored to specific needs.<br />
© 2011 ACADEMY PUBLISHER<br />
B. The modularity <strong>of</strong> the network protocol<br />
Modularity is central tenets in the design and<br />
implementation <strong>of</strong> hardware and s<strong>of</strong>tware system. In the<br />
paper <strong>of</strong> [16], the modularity <strong>of</strong> the network architecture<br />
is defined that breaks a network system into parts,<br />
normally to permit independent construction and<br />
replacement, reuse <strong>of</strong> parts, and so on. Early works on<br />
modular protocols have provided some solutions on<br />
protocol decomposition, configurable frameworks and<br />
process model. [8] proposed an x-kernel environment and<br />
mechanisms for communication between microprotocols.<br />
[9] is a configurable communication<br />
framework that provides a runtime platform <strong>of</strong> protocols<br />
consisting <strong>of</strong> standard, reusable services. [10] proposed a<br />
process-per-protocol model, a process or thread<br />
shepherds a message through the protocol stack. Some<br />
work [11] has proved the success <strong>of</strong> a modular platform<br />
with widespread deployment for reconfiguration <strong>of</strong> the<br />
entire data plane <strong>of</strong> a router system. Active networks [18]<br />
provided a powerful and very general approach to module<br />
packet processing function.<br />
C. Service-oriented network protocol and network<br />
architecture<br />
The research on service-oriented network protocol<br />
composition and network architecture enables dynamic<br />
adjustment <strong>of</strong> network features possible according to the<br />
requirement <strong>of</strong> users and applications. Service-oriented<br />
network architecture provides mechanisms for composing<br />
custom protocol stack [23], such as the SILO architecture<br />
[22]. The key technologies on service-oriented network<br />
protocol include abstraction <strong>of</strong> network service, network<br />
protocol composition and service path selection. A<br />
number <strong>of</strong> previous research projects have addressed<br />
some general thinking about how to specify a network<br />
service. [20] emphasize specifically on the middle boxes<br />
in the network such as traversing firewalls and network<br />
address translators, it can be seen as a step towards<br />
managing connections involving general services. [21,<br />
35] provided a more general method that specifies<br />
services very similar to pipeline abstractions. Service<br />
socket [28] is a user-level abstraction that has<br />
implemented some networking applications and services<br />
in networks. Some approaches focus on the<br />
decentralization <strong>of</strong> service composition. The SpiderNet<br />
project [24] provides the ability <strong>of</strong> service composition<br />
by a decentralized approach in P2P networks. A similar<br />
research about service composition [25] discuss the<br />
challenges that how to support for the service compositio<br />
on top <strong>of</strong> the Internet Indirection Infrastructure (i3). In<br />
[26, 27], the path selection is also done in a distributed<br />
manner, but end-systems and other entities along the path<br />
may specify specific service requirements.<br />
D. The appropriate mechanism <strong>of</strong> the provision, delivery,<br />
and consumption <strong>of</strong> network protocol<br />
The appropriate mechanism <strong>of</strong> the provision, delivery,<br />
and consumption <strong>of</strong> network protocol is becoming an<br />
important foundation for network protocol as a service. A<br />
network protocol developed by JUNOS SDK developers<br />
needs to deploy and run on each network device <strong>of</strong>
1086 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Juniper. The development can gain development fees<br />
based on a s<strong>of</strong>tware license. Cloud computing is different<br />
from JUNOS SDK. The devices and services are<br />
centralized deployment and running in data centers.<br />
Service developers can publish and sell their own<br />
s<strong>of</strong>tware services to the cloud service provider.<br />
III. NETWORKING AS A SERVICE<br />
A. Conceptualisation<br />
Networking as a service refers to a new Internet-based<br />
model that communication service provider (CSP) can<br />
deliver network protocols on-demand and reliably to the<br />
user based on SLA. The service consumer can use the<br />
service as pay as you go and achieve a good quality <strong>of</strong><br />
experience.<br />
From the perspective <strong>of</strong> service, the abstraction <strong>of</strong><br />
network function and the layer <strong>of</strong> network protocol stack<br />
will be re-organized and divided into three layers: service<br />
specification, network capacity, network behavior. In this<br />
vision, network service in different abstract forms will<br />
regard as middle ground for the continuous resolution <strong>of</strong><br />
tussles between providers and users. At the top level <strong>of</strong><br />
abstraction the service specification which defines data<br />
transmission parameters <strong>of</strong> user information needs to<br />
satisfy end user requirements. At the intermediate level<br />
<strong>of</strong> abstraction network capacity in accord with service<br />
specification is set up with network protocol composition.<br />
The dynamics <strong>of</strong> network capacity construction provides<br />
a utility function for the composition <strong>of</strong> horizontal service<br />
across the network and vertical service within a node. In<br />
this case, it is the service that utilizes the network and<br />
drives the customization <strong>of</strong> network capacity. As we<br />
move to lower levels this customization process is<br />
mapped to network behavior, access technologies and<br />
resource management policies such as forwarding,<br />
filtering, dropping, and so on.<br />
Figure 1. Cloud-based network.<br />
Cloud-based network (CBN), as shown in Fig. 1, is a<br />
form <strong>of</strong> implementation <strong>of</strong> networking as a service, which<br />
learns from concepts and ideas <strong>of</strong> cloud computing and<br />
service-oriented architecture. CBN provides the ability to<br />
© 2011 ACADEMY PUBLISHER<br />
deploy and run network protocols in the cloud, configure<br />
network resources and compose network protocol<br />
dynamically according to the user's service requirements<br />
and SLA, accordingly generate network control rules to<br />
manipulate forwarding behavior <strong>of</strong> a network device.<br />
CBN transforms network protocol to network service<br />
with zero-configuration [4] and zero-maintenance for<br />
network users. Each CSP may set up a CBN or the<br />
federation <strong>of</strong> CBNs to serve for the costumers.<br />
Protocol service instance (PSI) is a set <strong>of</strong> network<br />
protocols corresponding to each service requirement. PSI<br />
is the minimum unit <strong>of</strong> a network service in CBN.<br />
B. Stakeholders in Cloud-based Network<br />
In the current Internet business model, network-related<br />
stakeholders consist <strong>of</strong> end users, communication service<br />
providers and network equipment providers formed. In<br />
this case, end users pay for communications service<br />
providers to apply for network access services by<br />
communications service providers. Communications<br />
service providers are regard as a “pipeline” manager<br />
since it is almost impossible that communication service<br />
providers can deploy a new protocol to provide<br />
customized or value-added network services because the<br />
protocols are embedded in the device by network<br />
equipment providers. The business model <strong>of</strong> the Internet<br />
will be changed with the emergence <strong>of</strong> networking as a<br />
service. Communications service providers will enhance<br />
the ability to control the network. The role <strong>of</strong> network<br />
equipment providers will be subdivided. Users will apply<br />
for appropriate network services according to their need<br />
and consume the service in the way <strong>of</strong> pay-as-you-go<br />
with business development, thereby reduce the cost <strong>of</strong><br />
network investment and maintenance.<br />
Communication Service Providers: Communication<br />
service providers are responsible for the management and<br />
maintenance <strong>of</strong> network-based cloud and the provision <strong>of</strong><br />
a guaranteed quality <strong>of</strong> network services to service<br />
consumers. CSP can gain service revenue from service<br />
consumers according to the period, quality, quantity and<br />
scale <strong>of</strong> network service.<br />
Protocol Developers: Protocol developers can develop<br />
various network protocols with API and specification <strong>of</strong><br />
CBN. After passed a test, a network protocol can be<br />
published to the CBN. Protocol developers can gain<br />
license fees from CSP according to the scale <strong>of</strong><br />
deployment and frequency <strong>of</strong> running <strong>of</strong> the network<br />
protocols.<br />
Network Equipment Providers: In the CBN, the<br />
various components and interfaces <strong>of</strong> network devices<br />
will be standardized. Thus, the function <strong>of</strong> network<br />
equipment providers will be refined and divided into<br />
network components providers and network equipment<br />
integrators. Network component providers will focus on<br />
improving the performance, capacity <strong>of</strong> network<br />
components, while network equipment integrators will<br />
focus on improving the stability and reliability <strong>of</strong> network<br />
equipment composed by network components provided<br />
by network component providers.
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1087<br />
Figure 2. The reference architecture <strong>of</strong> Cloud-based network<br />
Network Service Consumers: Network service<br />
consumers, including personal and business users,<br />
purchase network services in a “pay as you go” model.<br />
They do not need to purchase expensive network<br />
equipment, do not need to care about the network<br />
configuration and maintenance. They do only need to put<br />
forward the requirements <strong>of</strong> network services according<br />
to the development <strong>of</strong> the business. And then use it.<br />
IV. CLOUD-BASED NETWORK REFERENCE ARCHITECTURE<br />
Cloud-based network reference architecture, as shown<br />
in Fig. 2, is divided into four layers: network resource<br />
pool, network operation interface, network runtime<br />
environment and network protocol service. Network<br />
operation interface is implemented in network device to<br />
manipulate network resource. Network runtime<br />
environment is a platform for network protocol<br />
deployment and operation. Network protocol service<br />
generates control rules and call for network operation<br />
interface to control and manage network resource.<br />
A. Network Resource Pool<br />
Network resource pool (NRP) is the network resource<br />
such as ports, bandwidth, queue, address, which can be as<br />
a basic service related with packet forwarding. Examples<br />
are Amazon EC2 for IP address and bandwidth<br />
assignment. Instead <strong>of</strong> some components <strong>of</strong> raw network<br />
hardware, NRP typically <strong>of</strong>fers the combination <strong>of</strong> these<br />
resources as a service through unified configuration and<br />
management.<br />
B. Network Operation Interface<br />
Network operation interface (NOI) is open and<br />
standardized API in order to configure and manage NRP.<br />
NOI provides three types <strong>of</strong> operating functions:<br />
parameter configuration for the network resource,<br />
forwarding control and event report. Parameter<br />
© 2011 ACADEMY PUBLISHER<br />
configuration function provides to set or get the max or<br />
minimum bandwidth limit, the numbers <strong>of</strong> queues, and IP<br />
address <strong>of</strong> a port, which can construct a user-oriented<br />
network topology. Forwarding control function provides<br />
abilities to output, drop, and filter packets according to<br />
the rules. Event report function provides an alarm or trap<br />
information when some network resource is down or<br />
overload.<br />
C. Network Runtime Environment<br />
Each <strong>of</strong> protocol set is called protocol service instance<br />
(PSI) which can be set up and running as a plug-in in<br />
network runtime environment (NRE). There is always a<br />
daemon running as a default PSI that provides a basic<br />
network layer protocol such as IP. NRE is responsible<br />
for billing, resource allocation, assessment, interconnect<br />
and reliability assurance for each PSI.<br />
Scheduling: According to the network service requests<br />
and current SLA state <strong>of</strong> the user, the scheduling function,<br />
firstly, will search related network resources and network<br />
protocols. If the requests are satisfied, the scheduling<br />
function will reconfigure network resource properties and<br />
generates PSI to control the rule <strong>of</strong> packet forwarding by<br />
issuing to the network equipment. Meanwhile, the<br />
scheduling function will inform the user the service is<br />
working and start the billing.<br />
Plug-in: Plug-in function enables network protocols to<br />
deploy, start, stop, upgrade and uninstall without reboot<br />
the system, just like OSGi, which is a module system and<br />
service platform for the Java programming language that<br />
implements a complete and dynamic component model,<br />
something that does not exist in standalone Java/VM<br />
environments. Thus, network protocol in PSI can be<br />
dynamically adjusted and smoothly switched by Plug-in<br />
feature.<br />
Pricing: To regard network capacity as a service, there<br />
will be a new billing model instead <strong>of</strong> bit-per pricing or<br />
online-time pricing: network capacity-based pricing and<br />
flow-per pricing. Pricing function enables to calculate the<br />
cost <strong>of</strong> service consumers according to the amount <strong>of</strong><br />
involved network protocols and run time <strong>of</strong> each PSI.<br />
Evaluation: Evaluation function provides the ability to<br />
monitor the operational status and network resource<br />
usage <strong>of</strong> PSI. The evaluation feature can determine<br />
whether the service provided by the PSI match the SLA<br />
through checking the service request.<br />
Interconnection: Interconnection function provides a<br />
communication mechanism and interface between multi<br />
PSIs. A PSI can send and receive the status information<br />
<strong>of</strong> a protocol through an interconnection interface such as<br />
JSON when it needs access or negotiates some<br />
information <strong>of</strong> protocols in another PSI. Eventually,<br />
network clouds can be interconnected by the multi PSIs<br />
interconnection.<br />
Migration: Migration function provides the mobility <strong>of</strong><br />
network services in the PSI level against network failures<br />
and high load. The PSI migration process includes: PSI<br />
state capture, marshaling PSI state and PSI service<br />
relocation.
1088 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
D. Network Protocol Service<br />
With the openness <strong>of</strong> network, a variety <strong>of</strong> new<br />
network protocol will be designed and implemented.<br />
How to identify and manage the new network protocols is<br />
a new problem in the future. Network protocol service<br />
consists <strong>of</strong> three functions: the description, management<br />
and composition <strong>of</strong> network protocol.<br />
Service Description <strong>of</strong> Network Protocol: Service<br />
description <strong>of</strong> network protocol is a structured language<br />
such as XML that provides a model for describing the<br />
capabilities <strong>of</strong> a network protocol. The NRE can choose<br />
appropriate network protocols to set up a PSI that can<br />
meet the user’s demands. Therefore, a network protocol<br />
should be able to accurately express properties and<br />
forward capacities <strong>of</strong> network protocol.<br />
Service Lifecycle Management <strong>of</strong> Network Protocol:<br />
The feature <strong>of</strong> service lifecycle management <strong>of</strong> network<br />
protocol provides the functions to manage network<br />
protocol versions, registration, certification and licensing.<br />
Service Composition <strong>of</strong> Network Protocol: The feature<br />
<strong>of</strong> service composition <strong>of</strong> network protocol provides the<br />
ability to generate new, more powerful network protocol<br />
service by composing protocols with different functions.<br />
The service composition <strong>of</strong> network protocols may learn<br />
from context aware service composition [5] and semantic<br />
web service composition [6].<br />
V. IMPLEMENTATION<br />
OpenFlow is an ideal way to build a network cloud.<br />
OpenFlow [36] is an open standard that allows network<br />
researchers to run experimental protocols in production<br />
network. It provides an open protocol to program the<br />
flowtable in a network device. The protocols<br />
implemented in a server outside control network devices<br />
by OpenFlow protocol, which is embedded in a device<br />
currently. It is in the process <strong>of</strong> being implemented by<br />
major switch vendors and used today by universities to<br />
deploy innovative networking technology. Thus,<br />
openness <strong>of</strong> network protocol in OpenFlow provides the<br />
possibility <strong>of</strong> network protocols as a service.<br />
NOX [34] is an open-source OpenFlow controller<br />
intended to simplify the development <strong>of</strong> s<strong>of</strong>tware for<br />
controlling or monitoring networks composed <strong>of</strong><br />
OpenFlow switches. Programs written within NOX<br />
(using either C++ or Python) have flow-level control <strong>of</strong><br />
the network. This means that they can determine which<br />
flows are allowed on the network and the path they take.<br />
In addition, NOX provides abilities to access to the<br />
network state including the network topology and the<br />
location <strong>of</strong> all detected hosts.<br />
Apache Hadoop is an open source distributed<br />
processing framework. The framework split dataset into<br />
manageable blocks in order to compute large datasets. It<br />
is in charge <strong>of</strong> the whole process by launching protocol<br />
instances, processing the protocol messages across many<br />
machines where the protocol is physically deployed and,<br />
at the end, aggregating the set <strong>of</strong> forwarding rules output<br />
into a final result [32].<br />
© 2011 ACADEMY PUBLISHER<br />
Figure 3. An initial prototype <strong>of</strong> Cloud-based network.<br />
In our laboratory, an initial prototype <strong>of</strong> Cloud-based<br />
network, as shown in Fig. 3, has been implemented by<br />
extending OpenFlow architecture. The implementation is<br />
divided into two levels: controllers cloud plan and<br />
network data plan. Controllers cloud plan provides the<br />
functions <strong>of</strong> NRE and NPS in Cloud-based Network<br />
Reference Architecture. A master controller with Apache<br />
Hadoop is responsible for distributing the data stream to<br />
three slave servers with different protocols. The LAMP<br />
on the master controller is responsible for network<br />
protocol registration and lookup. Xen deployed on each<br />
slave server make multi slave server as a slave server<br />
cluster. Network data plan consists <strong>of</strong> six OpenFlow<br />
switches to receive control information <strong>of</strong> flows from the<br />
master server in controllers cloud plan.<br />
VI. FUTURE WORK<br />
In this article, we have demonstrated the conception<br />
and role <strong>of</strong> on-demand provision <strong>of</strong> network capacity in<br />
order to achieve networking as a service. We have<br />
designed a novel future network architecture leaned from<br />
cloud computing and service-oriented architecture: cloudbase<br />
network.<br />
Based on the cloud-based network, we have built a<br />
prototype by extending the OpenFlow architecture and<br />
virtualization technology to verify on-demand provision<br />
<strong>of</strong> network capacity.<br />
Moving forward, there are some future works in cloudbased<br />
network architecture. First <strong>of</strong> all, based on the<br />
above prototype implementation, we will evaluate the<br />
performance and latency <strong>of</strong> network protocol as a service<br />
in the cloud-based network. And we will research an<br />
accurate expression <strong>of</strong> the demand for network services.<br />
A formal network protocol description language will be<br />
designed to configure network services automatically<br />
according to an accurate expression <strong>of</strong> network service<br />
requirements. Then, we will extend the current prototype<br />
implementation to multi-clouds interconnection by<br />
designing a cloud interconnection communication
JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011 1089<br />
protocol. In the multiple clouds based network prototype,<br />
we will research the capability <strong>of</strong> service optimization to<br />
provide a service consumer with the nearest service<br />
delivery. Finally, we will research the migration ability <strong>of</strong><br />
PSI in multiple clouds to improve the reliability <strong>of</strong> the<br />
cloud-based network.<br />
ACKNOWLEDGMENT<br />
This work was supported by National Science<br />
Foundation <strong>of</strong> China under Grant 61073172, Program for<br />
New Century Excellent Talents in University, and<br />
National Basic Research Program ("973" Program) <strong>of</strong><br />
China under Grant 2009CB320501.<br />
REFERENCES<br />
[1] NIST Definition <strong>of</strong> Cloud Computing v15,<br />
csrc.nist.gov/groups/SNS/cloud-computing/cloud-defv15.doc<br />
[2] N. McKeown, T. Anderson, H. Balakrishnan,G. Parulkar,<br />
L. Peterson, J. Rexford, S. Shenker, and J. Turner.<br />
Openflow: enabling innovation in campus networks.<br />
SIGCOMM Comput. Commun. Rev.,38(2):69–74, 2008.<br />
[3] J. Kelly, W. Araujo, and K. Banerjee, “Rapid service<br />
creation using the JUNOS SDK,” in ACM PRESTO, 2009.<br />
[4] S. Cheshire and D. H. Steinberg, “Zero Configureation<br />
Networking: The Definitive Guide,” O’Reilly, 2006,.<br />
[5] Vukovic M Context Aware Service Composition. PhD<br />
thesis, University <strong>of</strong> Cambridge, 2006<br />
[6] RAO, J. Semantic Web service composition via logicbased<br />
program synthesis. PhD thesis. Department <strong>of</strong><br />
Computer and Information Science, Norwegian University<br />
<strong>of</strong> Science and Technology, 2004.<br />
[7] N. M. K. Chowdhury and R. Boutaba, “A Survey <strong>of</strong><br />
Network Virtualization”, Technical Report, David R.<br />
Cheriton School <strong>of</strong> Computer Science, University <strong>of</strong><br />
Waterloo, Waterloo, Ontario, Canada, Tech. Rep. CS-<br />
2008-25, Oct 2008.<br />
[8] N. Hutchinson, L. Peterson, The x-kernel: An architecture<br />
for implementing network protocols, IEEE Transactions on<br />
S<strong>of</strong>tware Engineering 17 (1991) 64–76.<br />
[9] M. Zitterbart, B. Stiller, A. Tantawy, A model for flexible<br />
high-performance communication subsystems, IEEE<br />
<strong>Journal</strong> on Selected Areas in Communications 11 (1993)<br />
507–518.<br />
[10] D. Schmidt, T. Suda, Transport system architecture<br />
services for high-performance communications systems,<br />
IEEE <strong>Journal</strong> on Selected Areas in Communications 11<br />
(1993) 489–506.<br />
[11] R. Morris, E. Kohler, J. Jannotti, M. Kaashoek, The click<br />
modular router, SIGOPS Operating Systems Review 33<br />
(1999) 217–231.<br />
[12] D. D. Clark, “The design philosophy <strong>of</strong> the DARPA<br />
internet protocols,” in Proc. <strong>of</strong> ACM SIGCOMM 88,<br />
Stanford, CA, Aug. 1988, pp. 106–114.<br />
[13] D.D. Clark, J. Wroclawski, K.R. Sollins, and R. Braden,<br />
“Tussle in Cyberspace: Defining Tomorrow’s Internet,”<br />
Proc. ACM SIGCOMM 2002, pp. 347-356.<br />
[14] D. D. Clark and D. L. Tennenhouse, “Architectural<br />
considerations for a new generation <strong>of</strong> protocols,” in Proc.<br />
<strong>of</strong> ACM SIGCOMM 90, Philadelphia, PA, Sept. 1990, pp.<br />
200–208.<br />
[15] M. S. Blumenthal and D. D. Clark, “Rethinking the design<br />
<strong>of</strong> the internet: the end-to-end arguments vs. the brave new<br />
© 2011 ACADEMY PUBLISHER<br />
world,” ACM Transactions on Internet Technology, vol. 1,<br />
no. 1, pp. 70–109, 2001.<br />
[16] D. Clark, K. Sollins, J. Wroclawski, D. Katabi, J. Kulik, X.<br />
Yang, B. Braden, T. Faber, A. Falk, V. Pingali, M.<br />
Handley, and N. Chiappa, “New Arch: future generation<br />
internet architecture,” Tech. Rep., Dec. 2003.<br />
[17] S. Shenker, D. Clark, D. Estrin, and S. Herzog, “Pricing in<br />
computer networks: reshaping the research agenda,”<br />
SIGCOMM Computer Communication Review, vol. 26,<br />
no. 2, pp. 19–43, 1996.<br />
[18] Tennenhouse, D. L., and Wetherall, D. J. Towards active<br />
network architecture. ACM SIGCOMM Computer<br />
Communication Review 26, 2 (Apr. 1996), 5–18.<br />
[19] Barbir, A., Reinaldo, P., Chen, R., H<strong>of</strong>mann, M., and<br />
Hilarie, O. An architecture for open pluggable edge<br />
services (OPES). RFC 3835, Network Working Group,<br />
Aug. 2004.<br />
[20] Guha, S., and Francis, P. An end-middle-end approach to<br />
connection establishment. In SIGCOMM ’07: Proceedings<br />
<strong>of</strong> the 2007 conference on Applications, technologies,<br />
architectures, and protocols for computer communications<br />
(Kyoto, Japan, Aug. 2007), pp. 193–204.<br />
[21] Keller, R., Ramamirtham, J., Wolf, T., and Plattner, B.<br />
Active pipes: Program composition for programmable<br />
networks. In Proc. <strong>of</strong> the 2001 IEEE Conference on<br />
Military Communications (MILCOM) (McLean, VA, Oct.<br />
2001), pp. 962-966.<br />
[22] Rudra Dutta, G. N. R., Baldine, I., Bragg, A., and<br />
Stevenson, D. The SILO architecture for services<br />
integration, control, and optimization for the future<br />
internet. In Proc. <strong>of</strong> IEEE International Conference on<br />
Communications (ICC) (Glasgow, Scotland, June 2007),<br />
pp. 1899–1904.<br />
[23] Baldine, I., Vellala, M., Wang, A., Rouskas, G., Dutta, R.,<br />
and Stevenson, D. A unified s<strong>of</strong>tware architecture to<br />
enable cross-layer design in the future internet. In Proc. <strong>of</strong><br />
Sixteenth IEEE International Conference on Computer<br />
Communications and <strong>Networks</strong> (ICCCN) (Honolulu, HI,<br />
Aug. 2007).<br />
[24] Gu, X., Nahrstedt, K., and Yu, B. SpiderNet: An integrated<br />
peer-to-peer service composition framework. In Proc. <strong>of</strong><br />
Thirteenth IEEE International Symposium on High-<br />
Performance Distributed Computing (HPDC) (Honolulu,<br />
HI, June 2004), pp. 110–119.<br />
[25] Lakshminarayanan, K., Stoica, I., and Wehrle, K. Support<br />
for service composition in i3. In Proc. <strong>of</strong> the 12th annual<br />
ACM international conference on Multimedia (New York,<br />
NY, Oct. 2004), pp. 108–111.<br />
[26] Huang, X., Ganapathy, S., and Wolf, T. A scalable<br />
distributed routing protocol for networks with data-path<br />
services. In Proc. <strong>of</strong> 16th IEEE International Conference<br />
on Network Protocols (ICNP) (Orlando, FL, Oct. 2008).<br />
[27] Fu, X., Shi, W., Akkerman, A., and Karamcheti, V. CANS:<br />
composable, adaptive network services infrastructure. In<br />
Proc. <strong>of</strong> the 3rd USENIX Symposium on Internet<br />
Technologies and Systems (USITS) (San Francisco, CA,<br />
Mar. 2001), pp. 135–146.<br />
[28] Schmitt, M., Acharya, A., and Ibel, M. Service Sockets: A<br />
uniform user-level interface for networking applications.<br />
Tech. Rep. TRCS99-39, University <strong>of</strong> California, Santa<br />
Barbara, Santa Barbara, CA, Dec. 1999.<br />
[29] D.J. Wetherall, J. Guttag, and D.L. Tennenhouse, “ANTS:<br />
A Toolkit for Building and Dynamically Deploying<br />
Network Protocols,” Technical Report, MIT, 1997, in Proc.<br />
OPENARCH’98.<br />
[30] T. Meyer, L. Yamamoto, C. Tschudin, An artificial<br />
chemistry for networking, in: Bio-Inspired Computing and
1090 JOURNAL OF NETWORKS, VOL. 6, NO. 7, JULY 2011<br />
Communication, First Workshop on Bio-Inspired Design<br />
<strong>of</strong> <strong>Networks</strong> (BIOWIRE 2007), pp. 45–57.<br />
[31] D. S. Alexander, W. A. Arbaugh, M. W. Hicks, P. Kakkar,<br />
A. D. Keromytis, J. T. Moore, C. A. Gunter, S. M. Nettles,<br />
and J. M. Smith, “The SwitchWare active network<br />
architecture,” IEEE Network, vol. 12, pp. 29–36, Aug.<br />
1998.<br />
[32] VARIA, J.2009. Cloud Architectures. Amazon Web<br />
Services.<br />
[33] T. Wolf, “Service-centric end-to-end abstractions in nextgeneration<br />
networks,” in Proc. <strong>of</strong> Fifteenth IEEE<br />
International Conference on Computer Communications<br />
and <strong>Networks</strong> (ICCCN), Arlington, VA, Oct. 2006, pp. 79–<br />
86.<br />
[34] NOX: An OpenFlow Controller. http://noxrepo.org/wp/.<br />
[35] S. Shanbhag and T. Wolf, “Implementation <strong>of</strong> end-to-end<br />
abstractions in a network service architecture,” in Fourth<br />
Conference on emerging Networking EXperiments and<br />
Technologies (CoNEXT), Madrid, Spain, 2008<br />
[36] The OpenFlow Switch Consortium.<br />
http://www.openflowswitch.org/<br />
Tao Feng was born in Shandong, China, in 1979. He received<br />
the M.S. degree in communication engineering in 2006. He is<br />
currently working towards the Ph.D. degree in network<br />
technology form Tsinghua University, Beijing, China.<br />
He has been selected for the APAN’31 fellowship program.<br />
His research interests include future network, data center<br />
network, QoS and network management.<br />
© 2011 ACADEMY PUBLISHER<br />
Jun Bi received the B.S., M.S., and Ph.D. degree in computer<br />
science from Tsinghua University, Beijing, China. His<br />
dissertation studied Internet routing protocols and high<br />
performance routers and won the best dissertation award from<br />
Tsinghua University.<br />
From 1999 to 2000, he was a postdoctoral scholar <strong>of</strong> High<br />
Speed Network Department in Bell Laboratories Research,<br />
Lucent Technologies, New Jersey, USA. From 2000 to 2003, he<br />
was a research scientist <strong>of</strong> Bell Labs Research Communication<br />
Science Division and Bell Labs Advanced Communication<br />
Technologies Center. His research interests include Next<br />
Generation Internet Architecture and Protocols, High<br />
Performance Routers/Switches, Source Address Validation,<br />
Internet Routing, IPv4/IPv6 Transition, etc.<br />
Pr<strong>of</strong>. Jun Bi is a full pr<strong>of</strong>essor and director <strong>of</strong> Network<br />
Architecture & IPv6 Research Division, Network Research<br />
Center <strong>of</strong> Tsinghua University.<br />
Hongyu Hu was born in Hubei, China, in 1976. She received<br />
Ph.D.degree in Beijing Institute <strong>of</strong> Technology, Beijing, China.<br />
She current is a Post Doctor in Network Research Center,<br />
Tsinghua University, Beijing, China.<br />
Her current research interests include future Internet, QoS<br />
routing and IP multicast.<br />
Hui Cao was born in Shandong, China, in 1980. She received<br />
the M.S. degree in communication engineering in 2006. She is<br />
currently working towards the Ph.D. degree in trust computing<br />
technology form Tsinghua University, Beijing, China.<br />
Her research interests include trust computing, social<br />
network and complex network.
Aims and Scope.<br />
Call for Papers and Special Issues<br />
<strong>Journal</strong> <strong>of</strong> <strong>Networks</strong> (JNW, ISSN 1796-2056) is a scholarly peer-reviewed international scientific journal published monthly, focusing on theories,<br />
methods, and applications in networks. It provide a high pr<strong>of</strong>ile, leading edge forum for academic researchers, industrial pr<strong>of</strong>essionals, engineers,<br />
consultants, managers, educators and policy makers working in the field to contribute and disseminate innovative new work on networks.<br />
The <strong>Journal</strong> <strong>of</strong> <strong>Networks</strong> reflects the multidisciplinary nature <strong>of</strong> communications networks. It is committed to the timely publication <strong>of</strong> highquality<br />
papers that advance the state-<strong>of</strong>-the-art and practical applications <strong>of</strong> communication networks. Both theoretical research contributions<br />
(presenting new techniques, concepts, or analyses) and applied contributions (reporting on experiences and experiments with actual systems) and<br />
tutorial expositions <strong>of</strong> permanent reference value are published. The topics covered by this journal include, but not limited to, the following topics:<br />
• Network Technologies, Services and Applications, Network Operations and Management, Network Architecture and Design<br />
• Next Generation <strong>Networks</strong>, Next Generation Mobile <strong>Networks</strong><br />
• Communication Protocols and Theory, Signal Processing for Communications, Formal Methods in Communication Protocols<br />
• Multimedia Communications, Communications QoS<br />
• Information, Communications and Network Security, Reliability and Performance Modeling<br />
• Network Access, Error Recovery, Routing, Congestion, and Flow Control<br />
• BAN, PAN, LAN, MAN, WAN, Internet, Network Interconnections, Broadband and Very High Rate <strong>Networks</strong>,<br />
• Wireless Communications & Networking, Bluetooth, IrDA, RFID, WLAN, WMAX, 3G, Wireless Ad Hoc and Sensor <strong>Networks</strong><br />
• Data <strong>Networks</strong> and Telephone <strong>Networks</strong>, Optical Systems and <strong>Networks</strong>, Satellite and Space Communications<br />
Special Issue Guidelines<br />
Special issues feature specifically aimed and targeted topics <strong>of</strong> interest contributed by authors responding to a particular Call for Papers or by<br />
invitation, edited by guest editor(s). We encourage you to submit proposals for creating special issues in areas that are <strong>of</strong> interest to the <strong>Journal</strong>.<br />
Preference will be given to proposals that cover some unique aspect <strong>of</strong> the technology and ones that include subjects that are timely and useful to the<br />
readers <strong>of</strong> the <strong>Journal</strong>. A Special Issue is typically made <strong>of</strong> 10 to 15 papers, with each paper 8 to 12 pages <strong>of</strong> length.<br />
The following information should be included as part <strong>of</strong> the proposal:<br />
• Proposed title for the Special Issue<br />
• Description <strong>of</strong> the topic area to be focused upon and justification<br />
• Review process for the selection and rejection <strong>of</strong> papers.<br />
• Name, contact, position, affiliation, and biography <strong>of</strong> the Guest Editor(s)<br />
• List <strong>of</strong> potential reviewers<br />
• Potential authors to the issue<br />
• Tentative time-table for the call for papers and reviews<br />
If a proposal is accepted, the guest editor will be responsible for:<br />
• Preparing the “Call for Papers” to be included on the <strong>Journal</strong>’s Web site.<br />
• Distribution <strong>of</strong> the Call for Papers broadly to various mailing lists and sites.<br />
• Getting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors. Authors should be<br />
informed the Instructions for Authors.<br />
• Providing us the completed and approved final versions <strong>of</strong> the papers formatted in the <strong>Journal</strong>’s style, together with all authors’ contact<br />
information.<br />
• Writing a one- or two-page introductory editorial to be published in the Special Issue.<br />
Special Issue for a Conference/Workshop<br />
A special issue for a Conference/Workshop is usually released in association with the committee members <strong>of</strong> the Conference/Workshop like<br />
general chairs and/or program chairs who are appointed as the Guest Editors <strong>of</strong> the Special Issue. Special Issue for a Conference/Workshop is<br />
typically made <strong>of</strong> 10 to 15 papers, with each paper 8 to 12 pages <strong>of</strong> length.<br />
Guest Editors are involved in the following steps in guest-editing a Special Issue based on a Conference/Workshop:<br />
• Selecting a Title for the Special Issue, e.g. “Special Issue: Selected Best Papers <strong>of</strong> XYZ Conference”.<br />
• Sending us a formal “Letter <strong>of</strong> Intent” for the Special Issue.<br />
• Creating a “Call for Papers” for the Special Issue, posting it on the conference web site, and publicizing it to the conference attendees.<br />
Information about the <strong>Journal</strong> and <strong>Academy</strong> <strong>Publisher</strong> can be included in the Call for Papers.<br />
• Establishing criteria for paper selection/rejections. The papers can be nominated based on multiple criteria, e.g. rank in review process plus<br />
the evaluation from the Session Chairs and the feedback from the Conference attendees.<br />
• Selecting and inviting submissions, arranging review process, making decisions, and carrying out all correspondence with the authors.<br />
Authors should be informed the Author Instructions. Usually, the Proceedings manuscripts should be expanded and enhanced.<br />
• Providing us the completed and approved final versions <strong>of</strong> the papers formatted in the <strong>Journal</strong>’s style, together with all authors’ contact<br />
information.<br />
• Writing a one- or two-page introductory editorial to be published in the Special Issue.<br />
More information is available on the web site at http://www.academypublisher.com/jnw/.
(Contents Continued from Back Cover)<br />
An Energy-Efficient Communication Protocol for Wireless Sensor <strong>Networks</strong><br />
Fengjun Shang<br />
Robust Cross-layer Design <strong>of</strong> Wireless Multimedia Sensor <strong>Networks</strong> with Correlation and<br />
Uncertainty<br />
Lei You and Chungui Liu<br />
The E-Commerce Model <strong>of</strong> Health Websites: An Integration <strong>of</strong> Web Quality, Perceived Interactivity,<br />
and Web Outcomes<br />
Chung-Hung Tsai<br />
A New Method <strong>of</strong> Time-frequency Synthesis <strong>of</strong> Harmonic Signal Extraction from Chaotic<br />
Background<br />
Erfu Wang, Zhifang Wang, Jing Ma, and Qun Ding<br />
Provable Data Possession <strong>of</strong> Resource-constrained Mobile Devices in Cloud Computing<br />
Jian Yang, Haihang Wang, Jian Wang, Chengxiang Tan, and Dingguo Yu<br />
Image Compression Based on Improved FFT Algorithm<br />
Juanli Hu, Jiabin Deng, and Juebo Wu<br />
Correlative Peak Interval Prediction and Analysis <strong>of</strong> Chaotic Sequences<br />
Qun Ding, Lu Wang, and Guanrong Chen<br />
REGULAR PAPERS<br />
An Energy Efficient Dynamic Clustering Protocol Based on Weight in Wireless Sensor <strong>Networks</strong><br />
Ming Zhang and Suoping Wang<br />
Performance <strong>of</strong> UWB Systems with Direct-Sequence Bipolar Pulse Amplitude Modulation and<br />
RAKE Reception over IEEE 802.15.3a Channel<br />
Jingjing Wang and Hao Zhang<br />
Data Accuracy Estimation for Spatially Correlated Data in Wireless Sensor <strong>Networks</strong> under<br />
Distributed Clustering<br />
Jyotirmoy Karjee and H.S Jamadagni<br />
Networking as a Service: a Cloud-based Network Architecture<br />
Tao Feng, Jun Bi, Hongyu Hu, and Hui Cao<br />
999<br />
1009<br />
1017<br />
1025<br />
1033<br />
1041<br />
1049<br />
1057<br />
1065<br />
1072<br />
1084