Wireless Sensor and Actuator Networks for Lighting Energy ...

Wireless Sensor and Actuator Networks for Lighting Energy Efficiency and User
Satisfaction
by
Yao-Jung Wen
B.S. (National Taiwan University) 1999
M.S. (University of California, Berkeley) 2004
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
Engineering-Mechanical Engineering
in the
Graduate Division
of the
University of California, Berkeley
Committee in charge:
Professor Alice M. Agogino, Chair
Professor David M. Auslander
Professor Edward Arens
Fall 2008

Wireless Sensor and Actuator Networks for Lighting Energy Efficiency and User
Satisfaction
© 2008
by Yao-Jung Wen

Abstract
Wireless Sensor and Actuator Networks for Lighting Energy Efficiency and User
Satisfaction
by
Yao-Jung Wen
Doctor of Philosophy in Engineering-Mechanical Engineering
University of California, Berkeley
Professor Alice M. Agogino, Chair
Buildings consume more than one third of the primary energy generated in the
U.S., and lighting alone accounts for approximately 30% of the energy usage in
commercial buildings. As the largest electricity consumer of all building electrical
systems, lighting harbors the greatest potential for energy savings in the commercial
sector. Fifty percent of current energy consumption could be reduced with energyefficient
lighting management strategies. While commercial products do exist, they are
poorly received due to exorbitant retrofitting cost and unsatisfactory performance. As a
result, most commercial buildings, especially legacy buildings, have not taken
advantage of the opportunity to generate savings from lighting. The emergence of
wireless sensor and actuator network (WSAN) technologies presents an alternative that
circumvents costly rewiring and promises better performance than existing commercial
lighting systems.
The goal of this dissertation research is to develop a framework for wirelessnetworked
lighting systems with increased cost effectiveness, energy efficiency, and
1

user satisfaction. This research is realized through both theoretical developments and
implementations. The theoretical research aims at developing techniques for harnessing
WSAN technologies to lighting hardware and control strategies. Leveraging
redundancy, a sensor validation and fusion algorithm is developed for extracting
pertinent lighting information from the disturbance-prone desktop-mounted
photosensors. An adaptive sensing strategy optimizes the timing of data acquisition and
power-hungry wireless transmission of sensory feedback in real-time lighting control.
Exploiting the individual addressability of wireless-enabled luminaires, a lighting
optimization algorithm is developed to create the optimal lighting that minimizes
energy usage while satisfying occupants’ diverse lighting preferences.
The wireless-networked lighting system was implemented and tested in a
number of real-life settings. A human subject study conducted in a private office
concluded that the research system was competitive with the commercial lighting
system with much fewer retrofitting requirements. The system implemented in a sharedspace
office realized a self-configuring mesh network with wireless photosensors and
light actuators, and demonstrated a 50% energy savings and increased performance
when harvesting daylight through windows is possible. The cost analysis revealed a
reasonable payback period after the system is optimized for commercialization and
confirms the marketing feasibility.
______________________________
Professor Alice M. Agogino
Dissertation Committee Chair
2

To my wife for having faith in me as I pursued my doctoral degree.
To my parents and grandparents for their unconditional love and support.
i

Acknowledgements
I would like to thank my research advisor Professor Alice Agogino for her
immense support and advice during the course of this dissertation research. I would also
like to acknowledge the support I have received from my research partners, members in
my research laboratory, friends in my department, and the funding agencies that made
this research possible. In particular, Dr. J. Granderson was a great mentor and a
valuable partner throughout the research. MS recipient J. Bonnell’s productive and
creative work was invaluable to the presented research. Dr. K. Goebel, undergraduates
J. Kim, A. Lahijaniain, M. Podust, I. Gurin, and K. Yeates have all made contributions
that benefited this research. Dr. J.-L. Wu and Ph.D. candidate C.-H. Jiang were
incredible consultants for programming and analysis approach. General Electric Global
Research, the UC Office of the Microelectronics Innovation and Computer Research
Opportunities (MICRO), VTT Technical Research Centre of Finland, the Energy
Innovation Small Grant of the Public Interest Energy Research (PIER) program, and the
UC Berkeley Chancellor’s Green Campus Fund were the funding agencies that have
kept the research project developing. The generous support from P. Black, R. Abesamis,
E. Ito, M. Largoza, and S. Castro of the UC Berkeley Physical Plant - Campus Service
were instrumental in making the system implementation of this research happen.
ii

Table of Contents
Chapter 1 Introduction ............................................................................................1
1.1 Research Goal ..................................................................................................1
1.2 Research System Architecture ..........................................................................2
1.3 Research Method..............................................................................................4
1.3.1 Theoretical Development...........................................................................4
1.3.2 Implementations ........................................................................................6
1.3.3 Verification of Theoretical Works with Implementations ...........................8
1.4 Research Scope...............................................................................................10
1.5 Research Contributions...................................................................................11
1.6 Dissertation Outline........................................................................................12
Chapter 2 Motivation.............................................................................................14
2.1 Energy Efficiency of Lighting in Commercial Buildings.................................14
2.1.1 Environmental Impact of Commercial Buildings......................................15
2.1.2 Lighting Control Strategies and Potential Savings....................................16
2.1.3 Challenges of Current Technologies.........................................................18
2.2 Wireless Sensor Networks ..............................................................................20
2.2.1 Characteristics of Wireless Sensor Nodes.................................................20
2.2.2 Characteristics of Wireless Sensor Networks............................................22
2.2.3 Popular Applications of Sensor Networks ................................................24
2.2.4 Wireless Sensor and Actuator Networks...................................................26
Chapter 3 Related Research and Literature Review ............................................28
iii

3.1 Next-Generation Energy-Efficient Lighting Systems ......................................28
3.2 Light Sensing and Actuation Using Wireless Networks ..................................30
3.3 Sensor Fusion .................................................................................................33
3.4 Fuzzy Set and Fuzzy Logic.............................................................................37
3.5 Optimization and Linear Programming...........................................................42
Chapter 4 Mote-FVF: Fuzzy Validation and Fusion for Sensor Networks..........45
4.1 Rationale ........................................................................................................45
4.2 Fuzzy Approach for Sensor Validation and Fusion .........................................46
4.3 Algorithm and Mathematical Detail................................................................47
4.4 Simulation and Experiment Results ................................................................55
4.4.1 Hardware and Environment Setup............................................................55
4.4.2 Tuning the Parameters of the mote-FVF Algorithm..................................56
4.4.3 Simulation and Real-time Testing Results................................................57
Chapter 5 Autonomous Sensing with Adaptive Sensing Rate ..............................61
5.1 Rationale ........................................................................................................61
5.2 Algorithm and Mathematical Detail................................................................63
5.3 Simulation and Experiment Results ................................................................74
Chapter 6 Optimal Lighting Actuation .................................................................80
6.1 Rationale ........................................................................................................80
6.2 Open-loop Lighting Optimization Algorithm..................................................81
6.3 Lighting Optimization Algorithm with Sensory Feedback...............................87
6.4 Simulation and Experiment Results ................................................................89
6.4.1 Open-loop Lighting Optimization Algorithm ...........................................91
iv

6.4.2 Lighting Optimization Algorithm with Sensory Feedback........................94
Chapter 7 Wireless-Enabled Lighting Component Design and Integration........96
7.1 Overview........................................................................................................96
7.2 Mote Photosensor Design ...............................................................................96
7.3 Mote-based Ballast Actuation Module Design ................................................99
7.3.1 Design Requirement Identification...........................................................99
7.3.2 Design Iterations....................................................................................101
7.4 Initial Implementation on Prototyping Luminaire Structure...........................107
7.4.1 Prototyping Luminaire Structure ............................................................108
7.4.2 Desktop Illuminance Regulation with Distributed Mote-FVF Algorithm109
Chapter 8 System Verification on Human Subjects............................................115
8.1 Overview......................................................................................................115
8.2 System Setup................................................................................................117
8.2.1 Hardware Implementation in Small Private Office .................................119
8.2.2 Lighting Controller and Overriding Mechanism.....................................121
8.3 Sensor Placement Test..................................................................................124
8.3.1 Objective ...............................................................................................124
8.3.2 Testing Procedure ..................................................................................124
8.3.3 Results and Discussion...........................................................................125
8.4 Comparison of System Performance and User Interface................................133
8.4.1 Objective ...............................................................................................133
8.4.2 Testing Procedure ..................................................................................134
8.4.3 Results and Discussion...........................................................................136
v

Chapter 9 System Implementation and Verification in a Shared-Space Office.141
9.1 Overview......................................................................................................141
9.2 Integrated Intelligent Lighting System ..........................................................142
9.3 Energy Savings Assessment..........................................................................155
9.3.1 User Interface and Power Measurement .................................................155
9.3.2 Long-term Energy Savings.....................................................................157
9.4 Daylight Response........................................................................................161
9.4.1 Experiment Setup...................................................................................161
9.4.2 Artificial Daylighting Generation...........................................................162
9.4.3 Testing results........................................................................................164
9.4.4 Discussion .............................................................................................177
9.5 Commercial Implication ...............................................................................179
9.5.1 Payback Period – Current and Projected Payback Period........................179
9.5.2 Commercialization Bottleneck ...............................................................184
Chapter 10 Conclusion and Future Research.......................................................188
10.1 Conclusion ...................................................................................................188
10.1.1 Wireless-enabled Sensor and Actuator .................................................189
10.1.2 Sensor Validation and Fusion for Wireless Sensor Networks................190
10.1.3 Autonomous Sensing with Adaptive Rate.............................................191
10.1.4 Optimized Lighting Actuation and Control...........................................192
10.1.5 Human Subject Testing........................................................................193
10.1.6 Integrated Wireless-enabled Intelligent Lighting System......................194
10.2 Contribution .................................................................................................195
vi

10.2.1 Theoretical Contributions.....................................................................196
10.2.2 Applications-oriented Contribution ......................................................197
10.3 Future Research............................................................................................198
10.3.1 Preference Modeling and Acquisition...................................................199
10.3.2 Integration with Other Sensing and Actuation Options.........................200
10.3.3 Integration with Other Lighting Energy Management Strategies...........201
10.3.4 Integration with the Building Management System ..............................202
10.3.5 Application of the Proposed Framework in Other Domain ...................203
References padding ...............................................................................................206
Appendix padding ................................................................................................224
Appendix A
Electronic Circuit Schematics........................................................224
A.1 Circuit Schematics of Photosensor Board for MICA and MICA2.............224
A.2 Circuit Schematics of Photosensor Board for Tmote Sky..........................225
A.3 Circuit Schematics of the First Generation Wireless Actuation Module....226
A.4 Circuit Schematics of the Second Generation Wireless Actuation Module227
A.5 Circuit Schematics of the Third Generation Wireless Actuation Module ..228
Appendix B Human Subject Test .......................................................................229
B.1 Complete Sensor Placement Test Data .....................................................229
B.2 Comparison of System Performance and User Interface Test Procedure ...233
B.3 Summary of the Responses from the Questionnaires ................................235
B.4 Human Subject Test Approval Letter........................................................237
B.5 Human Subject Test Protocol Narrative....................................................239
B.6 Human Subject Test Consent Form ..........................................................248
vii

B.7 Human Subject Test Questionnaire...........................................................251
B.8 Human Subject Test Interview Script .......................................................255
Appendix C Cost Analysis .................................................................................256
C.1 Average Illuminance Calculation Sheet....................................................256
C.2 Example of System Payback Period Calculation.......................................258
viii

List of Figures
Figure 1-1 Research system architecture........................................................................3
Figure 2-1 Primary energy consumption in commerical buildings. ..............................15
Figure 2-2 Primary energy expenditure in commercial buildings. ................................15
Figure 2-3 Energy consumption in commercial sector. ................................................16
Figure 2-4 Energy usage in office buildings.................................................................16
Figure 2-5 Wireless sensor platforms and sensor boards..............................................22
Figure 4-1 Mote-FVF algorithm architecture...............................................................48
Figure 4-2 Gaussian correlation curve. ........................................................................50
Figure 4-3 Membership functions for defining the center of validation curve...............52
Figure 4-4 Fuzzy centered validation curve. ................................................................52
Figure 4-5 Membership functions for determining the adaptive parameter . ..............54
Figure 4-6 Mote-FVF with median value majority voting scheme. ..............................58
Figure 4-7 Mote-FVF with Gaussian correlation majority voting scheme. ...................58
Figure 4-8 Comparison of variations of sensor validation and fusion algorithm...........60
Figure 5-1 Mechanism for adaptive sampling [92].......................................................63
Figure 5-2 Architecture of adaptive sensing rate algorithm..........................................63
Figure 5-3 Prediction performance of Kalman filtering................................................66
Figure 5-4 Adaptive Wiener filter................................................................................66
Figure 5-5 Prediction performance of adaptive Wiener filtering...................................68
Figure 5-6 Prediction performance of double exponential smoothing...........................69
Figure 5-7 Membership functions for determining sensing rate....................................71
ix

Figure 5-8 Membership functions for determining sensing rate with damping. ............73
Figure 5-9 Adaptive sensing with Kalman filtering predictive model...........................75
Figure 5-10 Adaptive sensing with Wiener filtering predictive model..........................75
Figure 5-11 Adaptive sensing with double exponential smoothing predictive model....76
Figure 5-12 Damped adaptive sensing with Kalman filtering predictive model............77
Figure 5-13 Damped adaptive sensing with Wiener filtering predictive model.............78
Figure 5-14 Damped adaptive sensing with double exponential smoothing predictive
model. .................................................................................................................78
Figure 6-1 Architecture of the optimal lighting actuation algorithm.............................82
Figure 6-2 Workplane level illuminance distribution model.........................................83
Figure 6-3 Lighting optimization with sensory feedback. ............................................88
Figure 6-4 Pseudo code of the lighting optimization algorithm with sensory feedback.90
Figure 6-5 Floor plan of the experiment office.............................................................91
Figure 6-6 Optimal lighting for scenario 1...................................................................92
Figure 6-7 Optimal lighting for scenario 2...................................................................93
Figure 6-8 Optimal lighting with sensory feedback......................................................95
Figure 7-1 Spectral response of Hamamatsu 7686 photodiode [100]............................97
Figure 7-2 Calibration curve of the mote photosensor..................................................99
Figure 7-3 Advance Mark VII characterization curve. ...............................................101
Figure 7-4 First generation wireless ballast actuation module....................................103
Figure 7-5 First generation wireless ballast actuation module operation logic............103
Figure 7-6 Second generation wireless ballast actuation module................................104
Figure 7-7 Second generation wireless ballast actuation module operation logic........104
x

Figure 7-8 Voltage divider. .......................................................................................105
Figure 7-9 Third generation wireless ballast actuation module...................................106
Figure 7-10 Third generation wireless ballast actuation module operation logic.........106
Figure 7-11 Prototyping Luminaire Structure. ...........................................................109
Figure 7-12 Long term desktop illuminance regulation..............................................112
Figure 7-13 Desktop illuminance setpoint tracking....................................................114
Figure 8-1 Private office layout.................................................................................118
Figure 8-2 Installation of the wireless ballast actuation module. ................................120
Figure 8-3 Deployment of system components. .........................................................120
Figure 8-4 GUI provided by the research system. ......................................................123
Figure 8-5 Handheld remote controller provided by the commercial system. .............123
Figure 8-6 Sensor placement of subject No. 2............................................................127
Figure 8-7 Sensor placement of subject No. 3............................................................128
Figure 8-8 Sensor placement of subject No. 7............................................................128
Figure 8-9 Sensor placement of subject No. 9............................................................129
Figure 8-10 Sensor placement of subject No. 10........................................................130
Figure 9-1 Floor plan of the small shared-space office...............................................143
Figure 9-2 Wireless-enabled dimming luminaire. ......................................................144
Figure 9-3 System operation diagram. .......................................................................145
Figure 9-4 Office lighting after retrofitted with the research system...........................147
Figure 9-5 Actuation message structure.....................................................................150
Figure 9-6 Grouping message structure. ....................................................................150
Figure 9-7 Status report message structure.................................................................152
xi

Figure 9-8 Beacon message structure.........................................................................153
Figure 9-9 Sensor message structure..........................................................................154
Figure 9-10 Sensor management message structure. ..................................................155
Figure 9-11 User interface on the Kiosk [107]...........................................................156
Figure 9-12 Lighting preset setting interface [107]. ...................................................157
Figure 9-13 Average hourly percent utilization of the lighting system [107]..............159
Figure 9-14 Average hourly power consumption [107]..............................................159
Figure 9-15 Daylight simulating structure. ................................................................163
Figure 9-16 Daylight harvesting test on four occupants. ............................................165
Figure 9-17 Simulated daylight output for the first case.............................................165
Figure 9-18 Sensor readings in the first case..............................................................167
Figure 9-19 Percentage of energy consumption in the first case.................................168
Figure 9-20 Daylight harvesting test on seven occupants...........................................169
Figure 9-21 Simulated daylight output for the second case. .......................................169
Figure 9-22 Sensor readings in the second case (1)....................................................171
Figure 9-23 Sensor readings in the second case (2)....................................................172
Figure 9-24 Percentage of energy consumption in the second case. ...........................172
Figure 9-25 Measured daylight fluctuation. ...............................................................173
Figure 9-26 Light output from the daylight simulating structure. ...............................174
Figure 9-27 Sensor readings in the third scenario (1).................................................175
Figure 9-28 Sensor readings in the third scenario (2).................................................176
Figure 9-29 Percentage of energy consumption in the third scenario..........................177
Figure 9-30 Energy savings vs. payback period with current system cost...................183
xii

Figure 9-31 Energy savings vs. payback period with projected system cost. ..............183
Figure 9-32 Dimmable ballast unit prices vs. payback periods...................................185
Figure 9-33 Photodiode unit prices vs. payback periods.............................................186
Figure 9-34 Wireless platform unit prices vs. payback periods. .................................187
xiii

Chapter 1
Introduction
1.1 Research Goal
This research concerns the development of a wireless-enabled energy-efficient
lighting system for commercial office buildings. The system resulting from this research
is expected to meet the following three criteria with respect to the current technologies:
(1) increased cost effectiveness, (2) increased energy efficiency, and (3) increased user
satisfaction and lighting quality.
Buildings consume more than one third of the total primary energy generated in
the US, two thirds of which is electricity [1]. Lighting alone accounts for 30% of the
energy usage in office buildings, and it contributes to the largest proportion of energy
consumption among all electrical systems [2]. The statistics imply that lighting
dominates the potential of energy savings from the commercial sector, and research on
energy-efficient lighting systems is necessary for achieving the highest possible
reduction in energy consumption.
Energy-efficient lighting systems have been introduced to the market, but have
not been widely adopted. Despite their occasional unsatisfying performance, the
primary barrier hindering the existing products from popularization is the exorbitant
retrofitting effort and rewiring cost, especially for legacy buildings. The emergence of
wireless sensor network technologies provides an alternative that circumvents costly
rewiring [3] and promises even better performance than existing commercial lighting
systems.
1

The integration of wireless sensor network technologies into energy-efficient
lighting systems is immature and still largely under development. Although wirelessenabled
lighting systems benefit from low retrofitting overhead, advanced technologies
are required to take full advantage of the wireless sensing and actuation capabilities and
to guarantee the performance. This research aims at developing a framework for
bridging new wireless sensor network technologies and typical lighting systems to
achieve cost effectiveness and energy efficiency.
1.2 Research System Architecture
The architecture of the wireless-enabled energy efficient lighting system is
shown in Figure 1-1. The system comprises wireless photosensors, wireless-enabled
ballast actuators, and intelligent algorithms for processing sensor readings and
determining the optimal lighting, which minimizes energy usage while satisfying
occupants’ lighting preferences.
The wireless photosensors mounted on the desktops measure the task
illuminance on the worksurface, which may be contributed by the combination of
artificial light and daylight. The photosensors are deployed directly onto or around the
desktops in the vicinity of the occupants by taking advantage of the ultimate miniature
size of the wireless sensor network technologies. Unlike the typical ceiling-mounted
photosensors, the desktop-mounted sensors see the light that the occupants perceive,
and hence result in better estimations of the actual illuminance.
2

Figure 1-1 Research system architecture.
The tiny sensors on the desktops are, however, prone to disturbances from the
occupants. Moreover, they are powered by batteries, which have limited life spans.
Therefore, the robustness of the sensor readings is guaranteed by leveraging redundancy
along with an intelligent sensor fusion algorithm to extract pertinent values from
multiple sensors while isolating faulty readings.
The intelligent lighting optimization algorithm takes into account the
representative fused sensor values, the occupancy status, and the present occupants’
lighting preferences to determine the optimal light settings for the luminaires. The
3

esulting lighting is optimal in the sense that it minimizes energy usage while satisfying
the lighting preferences specified by each occupant.
The optimal light settings are transmitted wirelessly to the wireless-enabled
ballast actuators, which in turn dim the lights or toggle the lights on/off to deliver the
desired lighting.
1.3 Research Method
The work in this research was realized with a combination of theoretical
developments, hardware integration, and implementations. The theoretical research –
including sensor fusion, adaptive sensing and lighting optimization – is to develop tools
for engaging the wireless sensor network technologies with the lighting hardware and
control strategies. The wireless capabilities of the components in the lighting system
were enabled through the integration of wireless platforms with lighting hardware, such
as photosensors and ballast actuators. A series of implementations were realized for
testing and verifying each aspect of this research as well as the performance of the
integrated system.
1.3.1 Theoretical Development
This part of the research applies and incorporates various concepts, including
fuzzy logic, prediction theory, smoothing method and optimization, to develop a
theoretical framework for harnessing the wireless sensor network technologies to
lighting control systems. The resulting three algorithms are: fuzzy sensor validation and
4

fusion, autonomous sensing with adaptive rate, and the optimal lighting actuation. The
algorithms are designed to be lightweight and computational tractable in view of
practicability for real-time operation in lighting control systems.
The fuzzy sensor validation and fusion algorithm is developed for extracting
pertinent sensor readings from multiple sensors in close proximity while rejecting
readings from faulty or disturbed sensors. In the research system the novel idea of
deploying small photosensors directly on the workplanes for better estimating desktop
illuminances also introduces the increased possibility of the sensors being disturbed by
the occupants. Moreover, since steady power for the sensors is usually not conveniently
available on the desktops, the photosensors have to run on batteries, which adds to the
unreliability of the sensors. Given these uncertainties, the reliability of the lighting
measurement needs to be guaranteed by leveraging redundancy along with sensor
validation and fusion. The sensor validation and fusion algorithm utilizes the theories of
fuzzy logic and exponentially weighted moving average (EWMA) to validate each
sensor reading and draw the most representative value from multiple sensors.
The autonomous sensing with adaptive rate algorithm adds another layer of
intelligence to each sensor node by dynamically adapting the sensing rate to the change
of the physical stimulus for better resolution of sensor information. In the application of
lighting control, the readings from the photosensors have to be transmitted in real time
for the control system to respond promptly to the daylight changes. Daylight may
remain steady for a long period of time, but may change rapidly during the day. Given
the fact that the most costly operation of a wireless sensor node is radio transmission [4,
5], it is energy-efficient to adjust the sensing rate so that the sensor nodes perform data
5

acquisition and wireless transmission in accordance with the change of daylight. The
adaptive sensing algorithm employs prediction theories and the simple exponential
smoothing method to detect the change of daylight, and adapts the sensing rate with a
set of fuzzy rules.
The optimal lighting actuation algorithm optimizes the light settings for each
luminaire to minimize the overall energy usage while delivering the desired lighting
specified by present occupants. People have diverse lighting preferences and
requirements [6, 7], which are not possible to satisfy with a single universal light
setting. Exploiting the individual controllability of the wirelessly enabled luminaires,
each lighting fixture can be actuated at a different level to deliver optimal lighting that
meets all the occupants’ requirements. While the lighting is driven by the occupants’
preferences, energy consumption is also minimized during the optimization. The
optimal lighting actuation algorithm applies the theory of optimization and formulates
lighting control into a linear programming problem with minimizing energy
consumption as the objective function and satisfying occupants’ lighting preferences as
the constraints.
1.3.2 Implementations
This part of the work concerns designing the hardware for enabling wireless
capabilities of the lighting control components as well as building prototyping systems
for verifying different aspects of the theoretical framework. Photosensor boards are
designed to attach to the wireless platforms to function as wireless photosensors, and
wireless actuation modules integrated with wireless platforms are developed for
6

interfacing and controlling the dimmable ballasts. The research system is realized
through a series of implementations from a prototyping luminaire structure, to a private
office, to a small share-space office with increased scale and complexity. Each of the
realizations serves a different purpose and makes a unique contribution to the research.
Photosensors and ballast actuators are an integral component of the research
system. Both the photosensors and the ballast actuators are integrated with wireless
platforms known as ‘smart motes’. The wireless photosensor possesses a response
similar to a human eye so as to accurately estimate the illuminance on the desktop and
transmit the readings over the radio. The ballast actuator is an add-on actuation module
interfacing with the dimmable ballast. The actuation module receives wireless actuation
commands and translates them into ballast control signals to dim the lights or toggle the
lights on/off.
The first implementation of the research system was realized on a prototyping
single-luminaire structure equipped with a dimmable ballast. Due to the simplicity of
the structure, it is highly configurable for various purposes and has served as a testbed
for verifying ideas, developing prototypes, and testing algorithms along the course of
this research. The wireless photosensors, the developed sensor validation and fusion
algorithm, the wireless ballast actuators, and a simple closed-loop wireless lighting
control system were tested using this structure.
A private office with four luminaires was retrofitted with the research system as
the second implementation. A sensor network for measuring desktop illuminance and an
actuator network for actuating the four dimmable luminaires were realized. Both
networks were implemented with simple networking protocols and were coordinated by
7

a central base computer for sensor data collection and control action management. This
office was implemented particularly for a set of human subject tests designed for
comparing the research system and a representative commercial daylighting system.
The third implementation was realized in a small shared-space office with 19
overhead fluorescent lights. The luminaires were retrofitted with wireless ballast
actuators, and an advanced networking protocol was enforced for a practical selfconfiguring
multihop wireless actuator network. Individual control of each of the
luminaires was also enabled to deliver a lighting condition satisfying all the occupants.
This implementation was meant to test the complete integrated energy-efficient lighting
system developed in this research and also serve as the testbed for future development.
1.3.3 Verification of Theoretical Works with Implementations
The theoretical advancements developed in this research were integrated with
the implementations for evaluating the functionalities and performances. In addition to
assessing each of the research components individually, three integrative tests were
conducted on the implemented systems, including lighting control with distributed
sensor fusion, user testing, and integrated optimal lighting control.
The study of lighting control with distributed sensor fusion was first conducted
on the prototyping luminaire structure. The fuzzy sensor validation and fusion algorithm
originally designed for processing on a central computing base was deployed to the
photosensor nodes so as to avoid massive and costly wireless transmissions of every
single sensor reading. The sensor readings were transmitted inside the cluster of sensors
with relatively short distance and low power, and were fused by the sensor node in
8

charge. Only the fused value had to be sent out of each sensor cluster for control
purposes. The luminaire structure was made into an illuminance regulating system,
where the wireless ballast actuator was embedded with a rule-based control law that
regulates the desktop illuminance at a designated level using the fused sensor readings
as feedback. In other words, the resulting system was fully autonomous without any
central computing unit coordinating the sensing and actuation actions.
A user test, to be conducted in the private office implementation, was designed
to study the following two topics: (1) the impact on the accuracy and pertinence of the
fused value delivered by the sensor validation and fusion algorithm if the sensor
locations are selected by the occupants, and (2) the comparison of the research system
with a representative commercial daylighting system in terms of various aspects
regarding overriding interfaces and system performances. Ten people were recruited to
participate in the study. The results from the first test were to indicate how to better
incorporate the sensor validation and fusion into the research lighting system. The
conclusion drawn from the second test was to show how to make the performance of the
research system more appealing from a user’s standpoint.
The performance of the integrated optimal lighting control was evaluated in the
small shared-space office implementation, where most of the components developed
through this research were integrated together for an overall assessment. Photosensors
were deployed onto the workplanes to form a wireless sensor network. The sensor
readings were transmitted to the central base station for calculating the light settings for
each luminaire optimized for minimal energy usage and maximal user satisfaction. The
9

optimal light settings were in turn sent to the ballast actuators in the luminaires so as to
deliver the desired lighting.
1.4 Research Scope
The development of an energy-efficient wireless lighting system involves
incorporating the knowledge of both wireless sensor networks and lighting control
strategies, each of which is supported by a devoted research community. This
dissertation is meant for bridging these two technologies and demonstrating the
feasibility and benefits of such an integrated lighting system. The work within the scope
of this dissertation is broken into the following three primary sets:
(1) Development of application level technologies for harnessing wireless sensor and
actuator networks to lighting management systems.
(2) Feasibility demonstration of interfacing lighting control components with
wireless sensor and actuator technologies.
(3) Implementation of the integrated energy-efficient lighting system.
The development of application level technologies takes full advantage of the
wireless sensor and actuator networks for lighting management, including wireless
communication capability, massive deployment potential, and individual addressability
and controllability. The emphasis on application level technologies implies that the
concerns of lower level wireless networking protocols and architecture, such as media
access control (MAC), duty cycle optimization, synchronization, etc., are beyond the
scope of this research. The development of efficient lower level networking protocols is
10

a popular active research area [8-13], upon which the technologies developed in this
research are built.
The work of interfacing lighting control components with wireless sensor and
actuator network technologies is meant to demonstrate practicability and realize the
research system as defined in the third research area. Although the prototyping lighting
components are also designed to be compact with the most suitable, economical and
environmental-friendly elements, it is clear that they can be further optimized if the
components were to be mass-produced for commercialization.
1.5 Research Contributions
This dissertation introduces a framework for energy-efficient wireless lighting
systems in hopes of providing high quality office lighting while mitigating global
warming by reducing lighting energy consumption. This research has made both
theoretical and applications oriented contributions.
The theoretical contributions have been made mainly through the theoretical
developments of this research. The algorithms developed are optimized and practical for
real-time applications. The fuzzy sensor validation and fusion algorithm and the
autonomous sensing with adaptive rate algorithm can be generalized for various sensing
applications. For example, the sensor fusion algorithm has been applied to structural
health monitoring for space vehicles in [14]. The optimal lighting actuation algorithm
can also be extended for systems with multiple actuators to achieve the optimal
performance. Moreover, this research also pioneered in exploring and demonstrating the
11

possibility of wireless actuator networks, which uses wireless sensor network
technologies for actuation.
This dissertation also shows the feasibility of the wireless networked lighting
system through the implementations during the research, and it demonstrates the
possibility of delivering satisfying lighting while minimizing energy usage. The
wireless-integrated lighting components promise a lower retrofitting cost and hence a
shorter payback period that is more likely to meet the five-year payback period criterion
acceptable to facility managers for adopting energy efficient products [15].
Furthermore, the ability to simultaneously minimize energy consumption and satisfy
each occupant’s lighting preferences points to the potential of the system being widely
adopted to generate savings.
1.6 Dissertation Outline
This chapter served as an overview of this dissertation and addressed the
research goals, introduced the system architecture, laid out the research methods,
defined the scope, and summarized the contributions. Having provided the big picture,
the following chapters detail each aspect of this dissertation research.
Chapter two and three lay down the foundation and background of this
dissertation research. Chapter two states the motivation of this research from an energy
standpoint and introduces wireless sensor networks, the core technology applied in this
research on energy-efficient and cost-effective lighting systems. Chapter three reviews
related research and literature on techniques utilized in the course of this research.
12

Chapters four through six present the theoretical development of bridging
wireless sensor/actuator networks and lighting systems. Chapter four details the sensor
validation and fusion technology developed for guaranteeing pertinent sensory
information from networked and redundantly deployed wireless sensors. Chapter five
proposes the sensing strategy that adapts the sensing rate to the change of the stimulus,
the daylight, so as to facilitate more efficient sensor data acquisition and wireless
communication. Chapter six describes the algorithm for lighting actuation, which
optimizes energy usage and satisfies occupants’ lighting preferences.
Chapters seven through nine depict the implementation and verification of the
developed system. Chapter seven details the design of wireless light sensors and
actuators as well as the initial implementation of the wirelessly enabled lighting system
on a prototyping luminaire structure. Chapter eight describes the implementation of
both the research lighting system and a representative commercial lighting system in a
small private office and compares various aspects of systems on human subjects.
Chapter nine presents a larger scale implementation in a shared-space office in which
real occupants work on a daily basis and demonstrates potential energy savings that
could be achieved with the research lighting system.
Chapter ten concludes the dissertation with detailed discussions of the
contributions and commercial implications of this research and points out future
research directions that could build on the research presented in this dissertation.
13

Chapter 2
Motivation
The motivations of this dissertation research are described in this chapter.
Lighting has been one of the major energy consumers among all electric systems in
buildings, and is a promising source for generating energy savings. The challenges of
current lighting management strategies are discussed to signify the need for new
lighting control technologies. Wireless sensor networks, the core technology that is
applied in this research on energy-efficient and cost-effective lighting systems are also
introduced in this chapter.
2.1 Energy Efficiency of Lighting in Commercial Buildings
Lighting has become one of the most important sources for generating energy
savings in commercial buildings. In fact, lighting accounts for 26% primary energy
consumption and 22% of primary energy expenditure in commercial buildings, topping
all other building systems as shown in Figure 2-1 and Figure 2-2 [16]. 40-50% of
energy savings may be achieved with a combination of energy efficient lighting
management strategies [17-19]. Commercial energy-efficient lighting products exist but
are poorly received due to exorbitant retrofitting costs and unsatisfying performances.
Therefore, new technologies are necessary for revising and popularizing energyefficient
lighting control systems to actually deliver the expected energy savings.
14

Figure 2-1 Primary energy consumption in commerical buildings.
Figure 2-2 Primary energy expenditure in commercial buildings.
2.1.1 Environmental Impact of Commercial Buildings
Lighting consumes more than 2,000 terawatt-hours (TWh) of electricity
globally, which corresponds to about 1,800 million metric tons of carbon dioxide (CO 2 )
emissions per year, and 48% of the lighting electricity is attributed to the commercial
sector [18]. In the United Stats, the 640 TWh lighting electricity demand in buildings
alone contributes to about 115 million metric tons of carbon dioxide emissions per year
[16], which accounts for almost 40% of greenhouse gas emissions [20].
Office buildings are the single largest energy user in the commercial sector as
shown in Figure 2-3, and can therefore play a major role in increasing energy efficiency
15

[21]. Lighting accounts for 30% of energy usage in office buildings, which assumes the
largest portion of energy usage as shown in Figure 2-4, and hence dominates the
potential of energy savings [2].
Figure 2-3 Energy consumption in commercial sector.
Figure 2-4 Energy usage in office buildings.
2.1.2 Lighting Control Strategies and Potential Savings
Various lighting control strategies have been developed, including daylight
harvesting, light level tuning, occupancy sensing, time switching and load shedding.
The maximum potential lighting energy savings is estimated to be 40-50% with a
combination of properly functioning control strategies [17-19].
16

Light level tuning generates energy savings by reducing the electric light level
away from the recommended standard according to occupants’ lighting preferences.
Researchers have identified significantly diverse lighting preferences and requirements
among individuals [6, 7], and an average savings between 15% and 25% has been
observed by allowing the lights to be tuned to occupants’ choices [22, 23].
Occupancy sensing strategies switches off lights automatically after a short
predefined period when no human presence is detected by the occupancy sensor. This
technology is by far the most implemented energy-efficient lighting control strategy,
which is adopted by 60% of the commercial offices in both new and existing
constructions. The average savings of using scheduling alone is 25% on average [19],
and the estimated potential savings for open-plane offices is in the range of 5-35%
according to various government organizations and manufacturers [24].
A time switching strategy toggles or dims the lights according to a predefined
schedule, and is usually implemented on building energy management systems and
lighting automation panels [25]. This control strategy is good for premises with fixed
business hours such as libraries, retail stores, museums, etc. However, modern flexible
working patterns make it difficult to set a fixed schedule for implementing this
technology in offices [26].
Load shedding is an important modern energy management practice in which
electricity consumers voluntarily cut down energy usage during peak demand hours in
order to prevent brownout or blackout due to power capacity shortages. This technology
is usually in effect for a short period of time during the day, and is not designed to
generate significant contributions to energy savings.
17

Table 2-1 summarizes each of the control strategies [19, 25, 26].
Table 2-1 Summary of energy efficient lighting management strategies.
Lighting control
strategy
Strategy definition
Potential energy
savings
Adoption percentage
in commercial offices
New
Retrofit
Daylight harvesting
(daylight-response,
daylight-linking)
Automatically dim the electric
light or toggle the electric light off
in response to increased daylight
level
35-40% in daylit
areas
12% averaged over
the entire area
11.7% 7.5%
Light level tuning
Adjust electric light level
according to occupants’ preference
15-25% - -
Occupancy sensing
Automatically turn off lights after
space is vacated
25% compared to
manual switching
61.7% 59.7%
Time switching
Automatically dim or switch off
lights according to a predefined
schedule
- 54.3% 41.5%
Load shedding
(demand response)
Voluntarily reduce light levels
during peak demand period N/A - -
2.1.3 Challenges of Current Technologies
Although the lighting control strategies introduced in the previous section may
introduce more than 40% of energy savings, most of them have not been widely
adopted in commercial buildings, resulting in a missed opportunity for significant
energy savings. In fact, each of the control strategies has specific shortcomings that lead
to unaffordable costs or unsatisfying performances.
The difficulty of implementing a daylight harvesting system lies in the
complexity of design and commissioning. Poor operations can easily result from
improper zoning and inappropriate sensor locations [26]. A single photosensor can only
18

e used to control the lights in areas receiving similar amount of daylight so as to ensure
the same light level within the controlled zone. The locations of the photosensors need
to be carefully chosen according to the type of control algorithms to maintain the
required task illuminance. Commissioning the daylighting systems requires expertise
and is time-consuming, especially for calibrating the photosensors to respond correctly
to the light changes on the workplanes [27]. Various factors have to be taken into
account in order to result in good performance. These include the reflectance of the
furniture and interior, the ceiling/task illuminance ratio, the incident angle of daylight,
the maximum possible reception of daylight, etc. The extra commissioning expense on
top of the high installing/retrofitting cost makes the already expensive daylight
responsive systems even more economically unattractive.
Light level tuning only works for private offices with a single occupant, since
current technology on the market cannot balance the competing lighting preferences of
multiple occupants. Therefore, it is not possible to satisfy each occupant’s lighting
preference with this control strategy in shared-space offices, let alone generate energy
savings. For the occupancy sensing technology, unwanted switching off while people
are present has always threatened to render the sensing system obsolete. People would
rather rollback to manual switching of the lights than be bothered by the sudden falseoffs,
and consequently energy savings are jeopardized. The initial cost is also identified
as a barrier to the wide adoption of this technology [19, 25, 26].
In summary, exorbitant initial costs, complicated implementation, and
commonly poor performances prohibit daylight harvesting control strategy from
widespread adoption. Moreover, a lack of proper technologies renders energy savings
19

from light level tuning impossible for open-space offices. In response, this dissertation
research aimed at developing a framework for lighting systems exploiting wireless
sensor and actuator network technologies to circumvent excessive retrofitting costs. The
resulting lighting system generates energy savings by harvesting daylight while
satisfying the diverse lighting preferences of each occupant in a shared-space office.
2.2 Wireless Sensor Networks
The idea of ‘Smart Dust’ wireless sensor networks was first proposed by a
research team in the University of California at Berkeley in 1999 as a futuristic mesh
network comprised of dust-sized processing and communication units based on MEMS
(Micro-Electro-Mechanical Systems) technology [28]. These smart sensors would be
capable of being massively deployed to any kind of environment and to self-organize
into networks for monitoring purposes. Although not yet at the micro-scale, the wireless
sensor network technologies have been identified as a promising solution for advanced
building operation systems to circumvent costly installation and rewiring, particularly in
legacy buildings [3].
2.2.1 Characteristics of Wireless Sensor Nodes
A wireless sensor node in general comprises a microcontroller, a radio
transceiver, a sensor or a suite of sensors, memory spaces, and an energy source. The
operating system residing in the microcontroller coordinates all the other components to
accomplish sensing and wireless communication activities. The energy source generally
20

has very limited energy and provides power to all the components of a sensor node.
Since the wireless sensor nodes usually rely on limited power sources such as batteries,
operating at very low duty cycle becomes the most distinguishing characteristic of the
nodes. In other words, the sensor nodes will constantly turn off components not in use,
such as the radio transceiver and sensors, and enter into, and stay at, the low-power
mode (usually referred to as the sleeping mode) as often as possible.
In typical operation, the microcontroller wakes up from the low-power mode
and acquires sensor readings from the sensor suite either periodically or upon request.
Depending on the nature of the sensing tasks, the microcontroller may decide to send
out the readings as soon as they are acquired or temporarily store them in the memory
for sending out in batch later on. When the microcontroller is ready for wireless
transmission, it powers up the radio transceiver to the transmission mode and sends out
the queued data. The radio transceiver may then be switched to the receiving mode to
listen to acknowledgement or other messages. As soon as the microcontroller makes
sure there is no more wireless communication required, it turns off the radio transceiver
and enters the low-power mode again.
While truly miniature platforms are still under development, millimeter-scale
‘motes’ and the surrounding network technologies have been maturing over the past
years and are now commercially available [29-33]. Several generations and variations
of smart motes exist by putting together commercially available electronic components,
including a microcontroller, a wireless transceiver, memories, and possibly some
sensors. These units can be configured with a variety of sensors appropriate to specific
sensing applications. In particular, three different generations of mote platforms are
21

used in the progress of this research and are configured with customized sensor boards
to function as photosensors. All three modules operate on TinyOS, an operating system
designed specifically for resource-constrained wireless sensors [34]. Figure 2-5 (a), (b),
and (c) show the first, second, and third mote module used in this research respectively
with the customized sensor board attached. They have very similar functionalities with
minor variations on microcontroller, radio transceiver, and programming interfaces.
Table 2-2 compares the specification of the three mote modules.
(a) MICA (b) MICA2 (c) Tmote Sky
Figure 2-5 Wireless sensor platforms and sensor boards.
2.2.2 Characteristics of Wireless Sensor Networks
Wireless sensor networks are distinguishable from ad hoc networks by the
following features [10]:
• The number of sensor nodes in a sensor network can be significantly higher.
• Sensor nodes are densely deployed.
• Sensor nodes are prone to failures.
22

• The topology of a sensor network may change frequently.
• Sensor nodes mainly use broadcast communication paradigm.
• Sensor nodes have limited power, computational capability and memory.
• Sensor nodes may not have global identification.
The emphasis on sensor in the term wireless sensor networks makes them very
application-specific. The type and specification of sensors vary with applications, which
also affect how a sensor network is deployed.
Table 2-2 Comparison of sensor platform specifications.
Mote Module MICA MICA2 Tmote Sky
Processor Performance
Speed 4 MHz 4 MHz 8MHz
Flash 128K bytes 128K bytes 48K bytes
SRAM 4K bytes 512K bytes 1024K bytes
EEPROM 4K bytes 4K bytes 10 K bytes
Analog to Digital Converter 10 bit ADC 10 bit ADC 12 bit ADC
Processor Current Draw
Active 5.5 mA 8 mA 0.5 mA
Sleep

are expected to autonomously configure into a connected network to relay information
to the data sink. Under circumstances where the wireless connection is compromised
due to noisy links, malfunctioning nodes or moving nodes, the network should be able
to hop onto different carrier frequencies or rebuild routing paths to ensure the
connectivity. Furthermore, as a consequence of the dense deployment and the failureprone
nature of the wireless sensor nodes, the information these nodes can observe has
become much more important than the reading from any particular node.
2.2.3 Popular Applications of Sensor Networks
Technologies surrounding wireless sensor networks have become attractive
research areas. Active research topics include the development of networking protocol
stack, hardware, operating systems, communication privacy and security, localization,
simulator, data aggregation and processing methods, energy scavenging, applications
and deployment, and so on.
The progress of MEMS and CMOS (complementary metal-oxidesemiconductor)
processing technology is the key to the ultimate dust-sized wireless
sensor nodes that integrate the components onto a single chip [35]. Researchers have
been devoted to the design and fabrication of components such as low power RF
transceivers [36] and sensors [37, 38] based on these technologies. Before the
technology becomes mature enough to deliver the single-chip dust sensors, researchers
and manufacturers have developed a variety of matchbox-sized macro wireless sensor
platforms, generally referred to as motes, by integrating off-the-shelf components.
24

MICA mote [31], Tmote Sky [29], BTnodes [39], Intel Mote [40], Dust Networks [41],
Sensicast [42] are a few examples.
The networking protocol for wireless sensor networks follows the typical ISO
OSI (open system interconnection) seven layer model – physical, data link, network,
transport, session, presentation and application layers. The first three layers (physical,
data link, and network) have attracted the most attention since they are more generic to
all applications whereas the other layers are more application-specific. The key issue for
the physical layer is standardization. Physical layer protocols that support wireless
sensor networks include Bluetooth (IEEE 802.15.1), IEEE 802.15.4, IEEE 802.15.4a,
ultra-wideband (UWB), etc.; IEEE 802.15.4 currently dominates the market and is the
building block of other mesh networking technologies such as Zigbee. Research on the
data link layer mostly focuses on MAC (medium access control) protocols for high
communication throughput under ultra low duty cycle operation such as LEACH [12],
Sensor-MAC (S-MAC) [9], DMAC [8], B-MAC [13], etc. Network layer concerns
energy-efficient routing and dissemination protocols for constructing self-configuring
networks. Since routing in wireless sensor networks also has to build on low power
operation, which is defined in the MAC protocol, it cannot be designed independent of
the MAC protocol. LEACH [12] and PEGASIS [43] are two renowned examples.
Wireless sensor networks have become a promising solution for numerous
sensing and monitoring applications where their wired counterparts are not applicable
due to large coverage or a harsh environment for running wires. The potential areas of
application include, but are not limited to, environmental monitoring, health care,
positioning and animal tracking, entertainment, logistics, transportation, home and
25

office, and industrial applications [44]. Mainwaring et al. investigated the system
requirements for monitoring a variety of critical variables in habitats with a wireless
sensor network, and deployed the developed system consisting of 32 sensor nodes to the
Great Duck Island off the coast of Maine [45]. Stankovic et al. proposed a wireless
sensor network architecture for smart homecare that fulfills the requirements of the next
generation medical care [46]. Fogel et al. considered the use of wireless sensor
networks for tracking the location of pallets of materials in a warehouse in addition to
typical entry/exit logging to meet volumetric and spatial constraints critical to material
storing management [47]. Chen et al. built a prototype of intelligent transportation
system (ITS), exploiting a wireless sensor network for traffic information collection and
communication to overcome the limitations of its wired counterpart [48]. Huang et al.
designed a series of wireless platforms with different versatilities that could be
integrated with a variety of sensors and implemented in a smart living space that
monitors various parameters in a room with the developed wireless sensor networks
[49]. General Motors has also started to explore wireless sensor network technology as
a solution to measuring the health of manufacturing equipment, and is expecting 10-
20% savings on production costs [50].
2.2.4 Wireless Sensor and Actuator Networks
The addition of wireless actuators into wireless sensor networks has greatly
broadened the domain of application for wireless sensor network technologies. Besides
passively monitoring the environment, a sensor and actuator network can actively
interact with the physical world where the actuators perform actions based on the data
26

collected from the sensors. A variety of applications can potentially benefit from the
integrated sensor and actuator network technologies such as disaster relief systems,
building automation systems, street parking systems, etc.
Wireless sensor and actuator networks distinguish themselves from typical
wireless sensor networks in two aspects: real-time requirement and coordination [51].
These two characteristics not only present research challenges but also affect the design
of the networks. Typical wireless sensor networks may tolerate larger latency in
exchange for better information integrity; however, real-time communication is usually
the key to ensuring system stability and performance in wireless sensor and actuator
networks. In most applications, actuators usually have access to a power source or they
are at least equipped with a bigger energy source, which impacts the low-power-centric
design for wireless sensor networks and could help with real-time communication.
Furthermore, the scope of the effect of each networked actuator may be different and
can overlap; hence coordination among actuators is necessary for yielding the desired
outcome. Depending on the applications, the coordination can be centralized between
the central base and the actuators, or decentralized in between actuators or between
sensors and actuators.
The lighting system proposed in this dissertation is a realization of a wireless
sensor and actuator network, where the wireless photosensors measure the desktop
illuminance and the wireless ballast actuators dim the lights based on the sensory
information. Specifically, the illuminance on each single worksurface in an office is
affected by more than one luminaire, and thus coordination among ballast actuators is
critical to deliver the desired lighting.
27

Chapter 3
Related Research and Literature Review
In this chapter, research and commercial products of energy-efficient lighting
automation systems that have similar scopes of this dissertation research are reviewed.
Attempts on light sensing and actuation with wireless sensor and actuator network
technologies are also discussed. These two topics are important as they set the
cornerstone of this dissertation research. In addition, mathematical techniques and
methodologies that form the basis of the theoretical development in this dissertation
research are briefly introduced. These include sensor fusion, fuzzy logic, and linear
programming.
3.1 Next-Generation Energy-Efficient Lighting Systems
Research and commercial products for advanced lighting automatic control
systems need to fulfill two goals: (1) optimize users’ comfort inside the room, and (2)
minimize the energy used while allowing a good inside comfort [52]. These lighting
systems are superior to typical ones in that the luminaires can be addressed and
controlled individually regardless of how they were physically wired.
The Integrated Building Environmental Communications System (IBECS)
piggybacks on the facility’s existing Ethernet using microLAN bridges to control and
monitor the lighting equipment that is attached to it [53]. Lighting devices, including
the ballasts, sensors, etc., are connected to a microLAN bridge, and are individually
addressable through a network interface. The control command and sensor information
is transferred over the Ethernet to client browsers or databases, and each microLAN can
28

e hundreds of feet long, consisting of 10-100 devices. While utilizing existing network
and residing control software on microLAN bridges reduces the cost of lighting
controls, costly and distant rewiring is still inevitable to interface lighting devices to the
microLAN bridges.
The Digital Addressable Lighting Interface (DALI) is a type of dimmable ballast
that utilizes a stand-alone communication protocol [54]. A maximum number of 64
DALI units can be individually addressed within a system and up to 16 scenes/presets
can be stored, which presents a scalability challenge for DALI systems. While all the
DALI-complying control units are guaranteed to work with DALI ballasts, controllers
from different manufacturers do not necessarily communicate with one another [55].
Also, the need for low-voltage wires for communicating sensor data and control signals
between each DALI-compatible device and the controller implies significant rewiring
overhead.
LonWorks (Local Operating NetWorks) is a solution for a wide range of
networked control systems with a dedicated LonTalk communication protocol, and
lighting automation is one of the most important applications [56]. A variety of
communication media are supported, including power line, coax cable, twisted pair,
radio frequency, etc. The addressing mechanism is very similar to the Internet Protocol
(IP), and one subnet can connect to up to 127 nodes. Also, 63 groups can be assigned in
one domain. The concept of LonWorks is to lower the cost of the devices by reducing
the engineering time required to rebuild communication protocols and provide a wide
range of compatible products through the promotion of the open protocol [57].
29

Although radio frequency is a supported communication medium, most LonWorks
products are wired.
3.2 Light Sensing and Actuation Using Wireless Networks
Light sensing and actuation has been one of the major application areas for
wireless sensor and actuator network technologies, especially for building management
where energy savings and user comfort are the priorities. Research on adopting wireless
sensor and actuator network technologies to lighting applications are reviewed in this
section.
O’Reilly et al. examined and analyzed the idea of harvesting daylight by using
wireless sensor networks to measure workplane illuminances and modifying DALI
ballasts to accommodate wireless communication [58]. A concrete result has not yet
been reported.
Dust Networks, a wireless sensor network service provider, combined forces
with a ballast manufacturer (SVA Lighting) and the Lawrence Berkeley National
Laboratory (LBNL) to develop a wireless-integrated dimming ballast [59]. The
environmental sensors previously developed by LBNL and power meters were
wirelessly enabled with Dust Networks technology to work with the wireless dimming
ballast. A control software and user interface were designed to integrate the wirelessenabled
hardware into an energy-efficient lighting control system, which supports
manual control, demand response, daylight harvesting, and scheduling. Their analysis
has shown a 20% reduction in retrofitting cost compared to a wired system. Information
30

about how the system is going to be realized in large-scale implementation and how it
will take user satisfaction into account has not been reported yet.
Adura Technologies provides lighting control service using wireless actuator
network technologies as the most economical energy-efficient solution for legacy
lighting systems [60, 61]. They developed a wireless actuation module for interfacing
with traditional non-dimmable ballasts to perform bi-level switching, which gives the
option of turning on different numbers of fluorescent tubes inside a luminaire. The
configuration of the lights and the switching patterns, schedules, and management
strategies can be determined by the customers’ choices.
Singhvi et al. used a utility function to model an occupant’s lighting preference
at a given location for a given light setting, and formulated lighting control as an
optimization problem that minimizes energy consumption while satisfying occupants’
lighting preferences [62]. The key to the efficiency of the control algorithm relies on
making the optimization problem tractable assuming that a single lamp only illuminates
the small area around it so as to simplify the optimization problem. A wireless sensor
network was deployed for illuminance measurement, and the lamps were actuated by
the X10 system through the power line [63]. The control algorithm was only verified on
a testbed using incandescent lamps, and thus its feasibility under the combination of
fluorescent light and daylight, where each light source has a much larger influence area,
has not been proven. To obtain and update the utility functions of occupants’ lighting
preferences remains as future research.
Li proposed a heterogeneous architectural concept for wireless light sensing and
actuation where sensors and actuators communicate in separate communication
31

channels with different capabilities and reliabilities [64]. The sensor network and the
actuator network were implemented separately using distinct routing protocols and
addressing schemes, and are coordinated by a central server. The system in the case
study was mainly for manual user control, and energy issues were not explicitly
addressed. Users controlled the lights through ambient networks [65] that were
interfaced with the wireless sensor and actuator network by a gateway server.
Granderson developed a framework for a daylighting system using wireless
sensing and actuation that simultaneously minimizes energy consumption, balances
users’ diverse lighting preferences, and increases facilities managers’ satisfaction [66].
Her research targeted shared-space commercial office buildings, and it formulated
lighting control as a Bayesian decision-making problem using an influence diagram
[67], which hierarchically arranges the information necessary to affect personalized
lighting. This lighting control system assumed that lights in the same region are
identically controlled, and the resulting decision is a single optimal light setting for the
entire region.
In addition to office lighting applications, Park et al. designed and implemented
an intelligent lighting control system, called Illuminator, for entertainment and media
production [68]. High fidelity wireless light sensors were developed and implemented
to form a sensor network for collecting stage lighting information [69]. Professional
stage lighting systems were interfaced with the wireless sensor network to deliver the
desired lighting. Given a light setup and user constraints, the Illuminator system
recommends sensor deployment, characterizes the lights, and finds the best light
actuation profiles satisfying user constraints based on light characteristics and sensor
32

eadings. The search for the best light actuation profiles is accomplished by minimizing
the total cost of the user constraints.
3.3 Sensor Fusion
Sensor fusion is a subset of information fusion that combines sensor readings
and/or data derived from multiple sources to result in representative information better
than that which each individual sensor may reveal in terms of accuracy, reliability or
completeness. The scope of sensor fusion covers the extraction of data from a set of
homogeneous or heterogeneous sensors and/or history data resulting from the same
stimulus and the inference of information from sensors monitoring different aspects of
the object of interest.
Sensor fusion has been applied to a wide variety of applications, including
military applications such as guidance for autonomous vehicles, battlefield surveillance,
automated thread recognition, etc., and nonmilitary applications such as the monitoring
of manufacturing processes, condition-based maintenance of complex machinery,
robotics, and so on. Sensor data may be combined or fused at different levels from raw
data level to a feature vector level or a decision level [70]. Raw sensor data that are
commensurate, i.e., measure the same physical phenomena, can be directly fused into
more representative data at the same observational level. Features must be extracted
from multiple sensors responding to different physical stimuli and combined into a
concatenate feature vector for information inference at the feature vector level fusion.
Decision level fusion involves decision-making based on the preliminary determination
of the object of interest from each sensor.
33

In practice, fusion of sensor data could generate worse information than that
from a single robust, well-tasked and high fidelity sensor that directly measures the
phenomenon of interest [70]. Poor information results from the attempt to combine
accurate data with biased or corrupted data, which closely resembles the Byzantine
generals problem [71]. The problem discusses the dilemma a system may encounter
when receiving conflicting information from malfunctioning components through the
metaphor of the Byzantine Army Generals camped with their troops preparing to attack
an enemy city [72]. Communicating only by messengers, the generals must agree upon
a common battle plan – to attack or to retreat, while one or more of them may be traitors
trying to confuse others with false messages. It has been proven that an agreement
among all generals can be reached if and only if more than two-thirds of the generals
are loyal. In other words, two-thirds of the sensors must be well-functioning in order to
produce a pertinent data by equally weighing and combining each sensor data through
sensor fusion. Although one may prefer to gather information from a single robust and
high fidelity sensor than perform sensor fusion on multiple sensors, such a sensor could
be unaffordable or may not even exist, or the phenomenon of interest is not directly
measurable in reality.
Popular techniques that have been applied to sensor fusion applications include,
but are not limited to Kalman filters, probabilistic data association filters (PADF),
hidden Markov models (HMM), maximum likelihood, fuzzy logic, neural networks,
Dempster-Shafer theory, etc. Hybrids of the sensor fusion methods have also been
implemented in some complex applications. Alag et al. developed a methodology for
sensor validation, fusion, and fault detection for equipment monitoring and diagnostics
34

utilizing a variant of a Kalman filter [73]. The performance of the methodology was
verified by monitoring the temperature in a gas turbine power plant. Durrant-Whyte et
al. derived a decentralized form of Kalman filter for multiple sensors to perform sensor
fusion locally with limited communication among the sensors [74]. The efficiency of
this decentralized sensor fusion architecture largely depends on the communication
topology. Alag et al. used a modified Kalman filtering approach for real time sensor
validation and PADF method for sensor fusion on the longitudinal control of automated
vehicles [75]. Data from the sonar, radar and optical sensors mounted on each car were
fused for accurate longitudinal distance between the cars. Bernardin et al. developed a
human hand grasp recognition technique by fusing a grasp-type measure from finger
joint angles and contact-point information from tactile sensors with a hidden Markov
model [76]. The HMM was trained using expectation maximization (EM) algorithm and
was able to attain 90% recognition accuracy with very little training data for twelve
grasp classes. Zhou et al. proposed a maximum likelihood approach for fusion of
multiple sensor data with correlated noise and unknown scaling factor mapping the
sensor readings to physical representations [77]. The approach was based on a
parametric modeling of the noise covariance and formulated in the transformed noise
subspace. Wu et al. built a sensor fusion architecture based on Dempster-Shafer theory
for context-aware computing applications by fusing together the evidence from both
video and audio sensors [78]. Goebel et al. developed a fuzzy logic based architecture
for real time sensor validation and fusion called FUSVAF. The performance of this
algorithm was demonstrated on vehicle-following tasks on automated highways [79],
and on the control of gas turbine power plants [80].
35

The emergence of wireless sensor networks presents the need of sensor
validation and fusion for extracting pertinent information from massively deployed yet
disturbance-prone sensor nodes. Durrant-Whyte et al. extended the decentralized
Kalman filter for sensor fusion derived in [74] so that sensor fusion can be performed
locally on each sensor node without a fully connected network topology [81]. Yuan et
al. discussed the trade-off between fusing large amounts of sensor data and the incurred
latency when data are propagated towards the sink, and proposed a multi-level fusion
synchronization protocol [82]. The protocol synchronized the fusion processes of the
nodes at different levels in the data propagation tree so that maximum numbers of
sensor data (or fused data) can get fused along the way towards the sink while
minimizing the overall latency. Kumar et al. developed an architectural framework for
distributed data fusion called DFuse, which consists of a data fusion application
programming interface (API) and a distributed algorithm for energy-aware role
assignment [83]. DFuse was implemented and evaluated on a network of personal
digital assistant (PDA) with wireless capabilities; however, it can not yet be deployed to
the more resource-limited wireless mote platforms. Jiang et al. developed a fusion
algorithm based on a likelihood ratio-based test (LRT) scheme for fusing binary
decisions made locally by each sensor node and transmitted through a fading (noisy)
channel [84]. Various models of fading channels were investigated and the optimal LRT
was derived under each scenario for data fusion.
36

3.4 Fuzzy Set and Fuzzy Logic
A fuzzy set is a generalization of classical set theory that assesses the grades of
membership of elements in a set. In classical set theory, an element can only either
belong or not belong to a set; however, fuzzy set theory determines the membership of
an element in a set characterized by a membership function [85].
The framework of fuzzy sets provides an intuitive way to work with problems
where imprecision is caused by the lack of clear criteria of membership rather than
random uncertainties. In the real world, a class of objects may not have a crisp
boundary to define its members. Take “the class of all real numbers which are much
greater than 1” mentioned in [85] as an example: while 10,000 may belong to the set for
sure, ambiguity arises in the cases of smaller numbers such as 10. In this case, “a
number much greater than 1” defines a fuzzy variable, or a linguistic variable, and the
grade of membership of each real number may depend on the circumstance of different
applications.
Overview of fuzzy set theory
Let X be a space, and any element in X is denoted by x, i.e. xX. A fuzzy set A
in X is specified by a membership function f A (x) corresponding to each of the elements
x in X, where f A (x)[0,1] characterizes the “grade of membership” of x in A. The closer
the value of f A (x) to 1, the higher the grade of membership of x in A.
Empty set:
The fuzzy set A is empty if and only if its membership function is zero in X.
37

A is empty f A
(x) = 0, x X (3.1)
Equality:
Two fuzzy sets A and B are equal if and only if the membership functions are
identical for all x in X.
Complement:
Containment:
A = B f A
(x) = f B
(x), x X (3.2)
The complement of the fuzzy set A is denoted as A´ and is defined by
f A
(x) = 1 f A
(x), x X (3.3)
The fuzzy set A is contained in fuzzy set B or is a subset of B, if and only if f A (x)
f B (x) for all x in X.
Union:
A B f A
(x) f B
(x), x X (3.4)
The union of two fuzzy sets A and B with the membership functions f A (x) and
f B (x) respectively is a fuzzy set C ( C = A B ) with corresponding membership
function:
Intersection:
f C
(x) = max{ f A
(x), f B
(x)}, x X (3.5)
The intersection of two fuzzy sets A and B with the membership functions f A (x)
and f B (x) respectively is a fuzzy set C ( C = A B ) with corresponding membership
function :
De Morgan’s laws:
f C
(x) = min{ f A
(x), f B
(x)}, x X (3.6)
38

De Morgan’s laws in the traditional set theory also hold in fuzzy set theory with
the operations of complement, union, and intersection.
(A B) = A
B
(A B) = A
B
(3.7)
Distributive laws:
theory.
The distributive laws in the traditional set theory also apply to the fuzzy set
C (A B) = (C A) (C B)
C (A B) = (C A) (C B)
(3.8)
Algebraic product:
The algebraic product of two fuzzy sets A and B with the membership functions
f A (x) and f B (x) is denoted by AB whose membership function is
f AB
(x) = f A
(x) f B
(x), x X (3.9)
Algebraic sum:
The algebraic sum of two fuzzy sets A and B with the membership functions
f A (x) and f B (x) is denoted by A+B whose membership function is
f A+ B
(x) = f A
(x) + f B
(x), x X (3.10)
The algebraic sum is only meaningful if f A+B (x) is less than or equal to unity for all x in
X.
Absolute difference:
The absolute difference of two fuzzy sets A and B with the membership
functions f A (x) and f B (x) is denoted by |A-B| whose membership function is
f A B
(x) = f A
(x) f B
(x) , x X (3.11)
39

Convex combination:
the relation
The convex combination of fuzzy sets A, B and is denoted by (A, B; ) with
(A, B;) = A + B (3.12)
where ´ is the complement of . The resulting membership function of this operation
is
Cartesian product and fuzzy relation:
f (A,B;)
(x) = f
(x) f A
(x) + ( 1 f
(x)) f B
(x), x X (3.13)
The Cartesian product of two fuzzy sets A in space X and B in space Y with
membership functions f A (x) and f B (y) respectively is a fuzzy set in the product space
XY with membership function
f A B
(x, y) = min{ f A
(x), f B
(y)}, x X, y Y (3.14)
A fuzzy relation in space X is a fuzzy set A in product space XX with membership
function f A (x 1 , x 2 ) for all x 1 , x 2 in X.
Fuzzy sets induced by mappings:
Let B be a fuzzy set in space Y with membership function f B (y), and T be a
mapping from X to Y. The fuzzy set A in space X induced by the inverse mapping T -1
has the membership function f A (x) defined by
for all x in X mapped into Y by T.
f A
(x) = f B
(y), y Y (3.15)
40

Fuzzy systems
Fuzzy engineering begins with three steps [86]. The first is to pick the input and
output variables. Then pick fuzzy subsets for the variables. Finally relate the output sets
to the input sets through a set of fuzzy rules. Consider a simple air conditioner with only
four settings: off, low, medium, and high, which constitute the fuzzy sets in the output
space, and the motor speed of the compressor is the fuzzy variable. The air temperature
is of course the fuzzy variable in the input space of the air conditioner, which can
belong to four fuzzy sets: cold, cool, warm and hot. Consequently, a set of fuzzy rules
may be defined as follows.
IF air temperature is cold, THEN set the air conditioner to off;
IF air temperature is slightly warm, THEN set the air conditioner to low;
IF air temperature is warm, THEN set the air conditioner to medium;
IF air temperature is hot, THEN set the air conditioner to high.
The fuzzy system also maps the input to the output in three steps [86]. The input
variable is first mapped to all the IF parts of the fuzzy rules in parallel to determine how
much this input belongs to each input fuzzy set. This step is sometimes referred to as
fuzzification. The second step scales the THEN parts of the matched input sets and sums
them up into a final output set. In the final step, the system calculates the final crisp
output from the output set, which is also known as defuzzification and can be
accomplished through various techniques.
Fuzzy logic is the main technology exploited in the development of the sensor
validation and fusion algorithm in Chapter 4 and the adaptive sensing algorithm in
Chapter 5. Fuzzy approach was chosen in this research because it is useful for situations
41

where no crisp boundary can be defined between sets, such as the trustworthiness of the
sensor readings for sensor fusion and the fastness of the stimulus change for the
adapting sensing rate. In addition, fuzzy technique is generally light-weighted in coding
and computation, and is thus suitable for being realized in the resource-limited,
wirelessly networked sensors.
3.5 Optimization and Linear Programming
Optimization or mathematical programming is a systematic process of choosing
the values of the variables in a specific set to minimize or maximize a real function. An
optimization problem is usually expressed in the following form.
minimize f 0
(x)
subject to f i
(x) b i
, i = 1,…,m,
(3.16)
where x = [x 1 , x 2 , …, x n ] T is the vector of optimization variables, f 0 (x) is the objective
function, f i (x), i=1,…,m, are the constraint functions, and b i ’s are the bounds of the
constraints. A vector x is feasible if it satisfies the constraints f i (x) b i , i=1,…,m, and
the problem is feasible if there exists at least one feasible vector x. Vector x * is an
optimal solution to the optimization problem if it has the smallest objective value
among all vectors that satisfy the constraints.
f 0
(x * ) f 0
(z)
for all z satisfying f i
(z) b i
, i = 1,…, m.
(3.17)
The problem is solvable if x *
exists. If the objective function and the constraint
functions are convex, then the optimal solution to the problem is unique and is usually
referred to as global optimum.
42

The optimization problem in (3.16) is called a linear programming problem if
the objective function and the constraint functions are linear, i.e.
f i
(x + y) = f i
(x) + f i
(y), i = 0,…,m. (3.18)
The standard form of a linear programming problem is formulated as (3.19):
minimize
subject to
c 1
x 1
+ c 2
x 2
++ c n
x n
a 11
x 1
+ a 12
x 2
++ a 1n
x n
b 1
a 21
x 1
+ a 22
x 2
++ a 2n
x n
b 2
(3.19)
a m1
x 1
+ a m2
x 2
++ a mn
x n
b m
x 1
, x 2
, x n
0,
or in matrix representation:
minimize
subject to
c T x
Ax b
x 0,
(3.20)
where
c = [ c 1
c 2
c n ] T
x = [ x 1
x 2
x n ] T
a 11
a 12
a 1n
a 21
a 22
a
2n
A =
a m1
a m2
a mn
b = [ b 1
b 2
b n ] T .
Linear programming problems have been well-studied, and various algorithms
have been proposed with different levels of efficiency and complexity. The earliest and
yet still most popular solver for linear programming problems is the simplex algorithm
43

developed by Dantzig [87]. The inequality constraints form a polyhedron, inside which
is the feasible region, and it has been proven that the optimal solution always occurs at
the vertex of the polyhedron. The main idea of the simplex algorithm is to find an
admissible solution at a vertex of the polyhedron, and then keep on moving along the
edges of the polyhedron to the adjacent vertex with lower objective value until no
adjacent vertex with lower objective value can be found.
It is, however, possible for the simplex algorithm to be caught in an infinite loop
and render the worst-case solution time to be infinite. The cycling problem is very rare
in practice, and variants of simplex algorithm that do not cycle exist [88]. Although
simplex algorithm is very efficient and usually converges to the optimal solution in
polynomial time, the worst-case complexity can be exponential if it does not cycle,
which inspired the development of the family of interior point methods. As opposed to
the simplex method, which travels along the edge of the polyhedron, the concept of the
interior point methods is to move through the interior of the feasible region; this
guarantees to converge to the optimal solution in polynomial time.
Linear programming is the core technique utilized in the optimal lighting
actuation system presented in Chapter 6. Given the setup of the lighting control
problem, it is very straightforward to formulate energy usage as the objective function
and users’ lighting preferences as the constraints to the linear programming problem.
Furthermore, by treating the light actuator outputs as the variables, the problem is
tractable and practical for real-time implementation since the number of variables is
limited.
44

Chapter 4
Mote-FVF: Fuzzy Validation and Fusion for
Sensor Networks
The research discussed in this chapter and the following two chapters are
devoted to better harnessing the wireless sensor and actuator network technologies to
the energy-efficient lighting systems. Wireless sensor and actuator network
technologies present the potential for low-cost, easy-retrofitting, versatile and energyefficient
lighting systems; however, the expected superiorities can only be achieved
with the complement of pertinent sensing information and proper control actions.
Successful lighting control relies on good sensory information from the light
sensors. The miniature size, limited energy, and the desktop-deployed photosensor
make the sensor nodes prone to disturbances and unpredictable failures. Redundancy as
well as sensor validation and fusion are introduced to ensure the reliability and
pertinence of the sensory information. The sensor validation and fusion algorithm based
on fuzzy logic, named mote-FVF, is developed for extracting pertinent sensor data from
a cluster of redundant sensors while rejecting readings from disturbed or malfunctioning
sensors.
4.1 Rationale
Wireless sensor networks promise more useful information about the monitored
environment via massive deployment of miniature sensor nodes. The global information
that can be extracted from the distributed sensors in the data-centric operation of a
45

sensor network is far more valuable than what each single sensor node reveals,
highlighting the need for an effective sensor fusion mechanism.
In practical applications, sensor information is always corrupted to some degree
by noise and sensor degradation, which vary with operating conditions, environmental
conditions, and other factors. The inexpensive, less reliable and massively distributed
MEMS mote sensors will be even more prone to individual failure. The interference
caused by other wireless equipment, which has become more and more ubiquitous,
makes the communication environment harsher for wireless sensor networks. Moreover,
the fact that the limited energy of the wireless sensor node will eventually run out adds
another layer of uncertainty to the sensor network. Efficiently detecting and isolating
bad sensor readings boosts the reliability of the mote sensor network, and hence
motivates the use of sensor validation.
In this research, the wireless photosensors are deployed on or around the
desktops in the proximity of the occupants for better estimating the workplane
illuminance. In addition to the uncertainties and disturbances described above, it is
inevitable that the tiny sensors are more likely to be accidentally touched or shaded by
the occupants. Therefore, redundancy is implemented and sensor validation and fusion
is enforced in order to ensure the accuracy and reliability of sensory information.
4.2 Fuzzy Approach for Sensor Validation and Fusion
Fuzzy approaches have proven to be effective and robust in a number of
challenging sensor validation and fusion applications [79, 80]. They are particularly
useful in applications where it is difficult, if not impossible, to construct a precise
46

mathematical model, such as in the case of self-organizing wireless sensor networks
that dynamically change over time. In addition, fuzzy logic is a suitable approach for
intelligent data aggregation in sensor networks, as the readings from the distributed and
miniature sensor nodes can always carry local disturbances and biases to some degree
and there is no crisp criterion for judging their trustworthiness.
4.3 Algorithm and Mathematical Detail
The mote-FVF algorithm makes use of a fuzzy exponential weighted moving
average (FEWMA) time series predictor, dynamic validation curves, and a fusion
scheme that incorporates confidence values for the measurements, the predicted value,
the measurements, and the system state [89]. Input to the mote-FVF algorithm depends
on the configuration of the sensor network and can draw on raw sensor measurements
or intra-network processed data. The output can be used for other intra-network data
fusion, the machine level controller, or supervisory control. In the case of this research
lighting system, the output is the representative desktop illuminance, which is used for
calculating and delivering optimal lighting. Figure 4-1 shows the architecture of the
mote-FVF algorithm. There are three units in the mote-FVF algorithm: validation,
fusion and prediction.
47

Figure 4-1 Mote-FVF algorithm architecture.
Validation unit
The sensor validation part of the mote-FVF algorithm validates each incoming
reading by assigning it a confidence value, which is determined by a dynamic validation
curve. The validation curve is generated from the specific sensor characteristics, the
predicted value, the correlation among incoming readings, and the physical limitation of
the sensor value. The assignment takes place in a validation gate. If sensor readings
show a change beyond the gate, the readings are flagged as erroneous and are assigned a
confidence value of ‘0’. A maximum confidence value of ‘1’ will be assigned to
readings according to their coincidence with the center of the gate, which is determined
by considering the predicted value and the correlation among all readings. The
correlation among sensor readings is a majority-voting scheme, which can be achieved
by the following two approaches: median value or Gaussian correlation.
48

Median value approach
This method of analyzing the behavior of a cluster of sensor readings first filters
out the obvious false readings based on the physical limitation of the sensors, and then
takes the median of the rest of the readings. Taking the median rather than the mean
value prevents bias induced by unreasonable readings, which are far away from the
majority but not filtered out at the first step. Generally, the median value should reveal a
reasonable estimation of the majority of sensor readings.
Gaussian correlation approach
This approach is more robust than the majority median value approach described
above, but requires more computational power. The first step of this approach is also to
filter out the obvious false readings based on the physical limitation of the sensors. For
the remaining readings, this method first generates a Gaussian function centered on the
reading with an appropriately fine-tuned standard deviation, designated as PDF n (x), and
then calculates the reading corresponding to the maximum value of the normalized
summation of all Gaussian functions as the estimate as shown in Figure 4-2. The
normalized summation of all Gaussian functions is calculated by
n
k =1
PDF k
(x)
n
,x = 0,......,maximum sensor output , (4.1)
where n is the total number of reasonable readings (i.e., readings remaining after
filtering out obviously false readings) [90].
49

Figure 4-2 Gaussian correlation curve.
Validation curve
There are a number of suitable validation curves; one useful curve is a piecewise
Gaussian curve of the form v(z) = e
xz 2
a w
, where parameter a w can be chosen
separately for the left and right curves according to the characteristics of sensors, z is
the sensor reading, and x is the center or precisely the split point of the left and right
sections of the validation curves. The curves should be normalized in order to scale the
confidence values from 0 to 1. The confidence values are then computed as follows.
50

e
=
e
x z
2
a left
1 e
x z
2
a right
1 e
0 z < v left
e
xv 2
left
a left
xv
left
a left
e
2
v left
< z x
xv 2
right
a right
xv
right
a right
2
x < z v right
0 z > v right
, (4.2)
where is the confidence value corresponding to a particular sensor reading, v left and
v right are the left and right validation gate borders respectively, a left and a right are the
parameters for the left and right validation curves [80].
x is determined by a fuzzy rule and provides a trade-off between the predicated
value and the majority of sensor readings. Denoting the estimated majority sensor
readings from either method above as m(x), and the variation between m(x) and the
predicted reading ˆx as diff(x), the fuzzy rules are applied as follows, starting from the
initial condition that x coincides with ˆx ,
IF diff(x) small THEN move x toward the m(x) a small amount;
IF diff(x) medium THEN move x toward the m(x) a medium amount;
IF diff(x) large THEN move x toward the m(x) a large amount.
The fuzzy membership function is designed using standard triangular shaped
functions and maximum overlap as shown in Figure 4-3 [91]. There are two parameters
to be tuned, m dif for fuzzification and m mov for defuzzification. Figure 4-4 depicts how
the center of the validation curve shifts between the value of majority voting and the
51

predicted reading. Note that the offset from the predicted reading takes the relationship
among sensor readings into account.
Figure 4-3 Membership functions for defining the center of validation curve.
Figure 4-4 Fuzzy centered validation curve.
Fusion unit
The fused value x f is calculated by taking the average of measurements weighted
by corresponding confidence values plus the predicted value weighted by , an adaptive
parameter representing the system state, and a constant scaling factor . The equation is
52

x f
=
n
1
n
1
z i
(z i
) + ˆx
(z i
) +
53
. (4.3)
The scaling factor is introduced to include a fraction of the predicted value to
account for the possible situation when no valid readings remain after the validation
procedure, and the algorithm will maintain its robustness for a temporary failure of all
sensors. Since the only purpose of the term containing is to deal with the situation of
sensor failure, is typically large to prevent the predicted value from dominating the
fused value, and it has to be tuned to the system at hand. In addition, the adaptive
parameter carries information about the state of the system and is used in both the
fusion unit and the prediction unit. If the system is in steady state, is set to a large
value in order to weight the past history more as the variation in measurements are very
likely caused by noises; on the other hand, if the system is in a transient state, is set to
a small value in order to weight the predicted value less so as to reduce the lag induced
by past history. A mechanism that distinguishes transient from steady state operations
looks at the prediction error ( xk ˆ( ) x ( k)
) and adjusts dynamically according to the
system state is given by the set of fuzzy rules below [92].
IF prediction error small THEN large;
IF prediction error medium THEN medium;
IF prediction error large THEN small.
f
The membership functions are also designed using triangular shaped functions
with maximum overlap such that only two parameters have to be specified: m e for the

fuzzification and m for the defuzzification. The membership functions are shown in
Figure 4-5.
Figure 4-5 Membership functions for determining the adaptive parameter .
Prediction unit
A time series predictor generates the predicted value for the next time step with
the adaptive parameter tuned to optimize the trade-offs between responsiveness,
smoothness, stability, and lag of the predictor. The standard exponential weighted
moving average predictor has the form
ˆx(k + 1) = ˆx(k) + (1 )x(k) , (4.4)
where xk+ ˆ( 1) is the predicted value of the next time step, xk ˆ( ) is the predicted value
of the current step, and x(k) is the upgraded current state. Combining the equation with
the output of the fusion unit in (4.3), the predictor in the mote-FVF algorithm is of the
form
ˆx(k + 1) = ˆx(k) + (1 )x f
(k) , (4.5)
54

where x f (k) is the current upgraded fused value [79, 80].
4.4 Simulation and Experiment Results
The mote-FVF algorithm was tested with a small-scale sensor network
consisting of six mote sensors. The sensed target was the illuminance on the testbed
under the prototyping dimmable luminaire structure (see Chapter 7.4 for detail). The
MICA motes were integrated with the photosensor boards as shown in Figure 2-5(a) to
measure the illuminance. The test environment possessed the following characteristics:
• There could be a large change in illuminance, such as when lights are turned on
and off. Thus it is important to test the algorithm’s response to such a large
discontinuous change.
• The algorithm should be tested over the operating range of the lighting
environment – from 0 lux (completely dark) to 1000 lux (bright lighting
conditions in an indoor office).
• Although zero-valued states are usually considered to be an error or rest
condition for many systems, in the lighting environment zero can be either an
error or a meaningful state when the environment is completely dark. The
algorithm should be capable of distinguishing between them.
4.4.1 Hardware and Environment Setup
The illuminance on the testbed varied from 50 lux to 1000 lux in 18 equal
intervals, and the true illuminance was measured with a high fidelity light meter. Six
55

mote sensors were arranged in a 3-by-2 matrix; each was programmed to acquire 10
illuminance readings every 5 seconds at the rate of 50 milliseconds per sample and to
send the data packet containing the readings along with the source mote ID to the base
station. To simplify the test, the sensors were configured into a single-hop centralized
network. In addition to the base mote, the base station consisted of a computer, which
ran a Matlab ® program with a Java I/O interfacing program to receive the incoming data
packets forwarded from the base mote, process the data, and carry out the mote-FVF
algorithm. The sensed illuminances transmitted in the mote network were digital values
from the ADC (analog-to-digital converters) on the mote. A separate calibration was
conducted in order to map the digital values to lux units.
4.4.2 Tuning the Parameters of the mote-FVF Algorithm
The parameters requiring tuning were a left and a right for the validation curve, the
fuzzification parameter m dif and the defuzzification parameter m mov for shifting the
center of the validation curve, the constant scaling factor , the fuzzification parameter
m e and the defuzzification parameter m for generating the adaptive parameter .
Without loss of generality, a left and a right were assigned to be the same to get a
symmetric validation curve as suggested by prior experiences and experiments.
Moreover, the values of a left and a right for each mote sensor were assigned identically, as
they were all of the same type and were expected to have similar characteristics. m dif
and m mov were adjusted so that the validation curve could shift enough to catch at least
some information of the possible largest change in a single time step without becoming
56

too sensitive to failures. Parameters m e and m were tuned so as to optimize the
response of the algorithm to the lighting environment. m e , m , m dif and m mov were tuned
separately for the two majority voting schemes. The constant scaling factor was
chosen to be large in order not to result in an obvious lag in the fused value by
weighting the previous predicted value too heavily.
4.4.3 Simulation and Real-time Testing Results
Figure 4-6 shows the real-time implementation of the mote-FVF algorithm using
the median value majority voting scheme on the six-mote sensor network. The small
cross signs are the raw sensed data from each mote, the dashed line is the true
illuminance measured by the high fidelity light meter, and the solid line is the fused
value. The figure reveals that the algorithm accurately followed the sensed data and
reflected the illuminance pertinently with a maximum error of 3.36% regardless of the
lag in the transient mode. Similarly, the algorithm with Gaussian correlation majority
voting scheme run off-line on the same set of data, shown in Figure 4-7, revealed a
maximum error of 4.01%. These errors include calibration errors when mapping raw
digital readings to units of lux and the variation of illuminance in the spatial position on
the testbed. From a physical point of view, the 4.01% error represents about a 30 lux
difference, to which a human being is insensitive. Furthermore, the choice of majority
voting mechanism didn’t seem to have a noticeable impact on the performance of the
algorithm.
57

Figure 4-6 Mote-FVF with median value majority voting scheme.
Figure 4-7 Mote-FVF with Gaussian correlation majority voting scheme.
A comparison of several different variations of the sensor validation and fusion
algorithm is shown in Figure 4-8. Part of the data set gathered in the real-time testing
was used to run off-line simulations for four configurations, and some modifications
were introduced to the data to simulate possible sensor failures. Plot (a) shows the
58

sensed data points as cross signs and the true illuminance with a dashed line. In plots (b)
to (e), the red dashed line displays the true illuminance and the black solid line shows
how the fusion algorithms follow the true illuminance. Plot (b) and (c) are the mote-
FVF algorithms with the two majority voting schemes, median value and the Gaussian
correlation, respectively. Both show good performances, even when facing sensor
failures. Plot (d) shows that without involving validation and prediction, a majority
voting algorithm is much more sensitive to sensor failures than the mote-FVF. Plot (e)
is the fusion algorithm without majority voting mechanism in the validation process,
which is clearly not capable of dealing with large discontinuous changes by fixing the
center of the validation curve at the predicted value.
The mote-FVF algorithm has shown its ability to extract pertinent readings from
multiple sensors in real time. Meanwhile, good sensor information also relies on a
proper sensing rate to capture the changes of the monitored environment. Sensing and
radio transmissions, especially the later, contribute to the largest portion of energy
consumption of the battery-powered wireless mote sensors. Therefore, it is unrealistic to
always implement a high sensing rate for good resolution of sensor information without
jeopardizing the sensors’ life span. The next chapter proposes an adaptive sensing
strategy to optimize the timing for sensing and radio transmission to ensure good
sensing resolution and long life span of the mote sensors.
59

Figure 4-8 Comparison of variations of sensor validation and fusion algorithm.
60

Chapter 5
Autonomous Sensing with Adaptive Sensing Rate
The algorithm of autonomous sensing with adaptive rate is based on prediction
theory and fuzzy logic. It takes advantage of the computational capability of the smart
motes and adds another layer of intelligence to the sensors. Wireless sensor nodes
embedded with the algorithm dynamically adapt their sensing rate to the change of the
physical stimulus, and thus avoid unnecessary sensing and wireless transmission while
retaining high resolution of the overall sensory information for real-time lighting
control purposes.
5.1 Rationale
For wireless sensor networks that monitor phenomena with slow dynamics, it is
sometimes not clear how to set an appropriate sensing rate so that the sensors return
data with good resolution without consuming unnecessary resources by over-sampling.
A fixed or uniform sensing rate that results in good resolution when the system is in
transient state is usually considered unnecessarily high during steady state. High sensing
rates in battery-powered wireless sensor nodes imply that energy is wasted turning the
sensors on/off and processing sensor readings. Although data acquisition generally
requires only a negligible amount of energy in the operation of wireless sensor nodes,
wireless transmission of the acquired readings for real-time control purposes will
contribute to significant power drain. On the other hand, a low uniform sensing rate
could cause an undesirable delay in response to changes even though it consumes less
energy sensing, processing, and communicating as compared to a higher sensing rate.
61

The daylighting system application in this research is a good example for the
dilemma of choosing a fixed optimal sensing rate. The available daylight in a room may
stay relatively steady for a long period of time during the day, but change rapidly at
dawn and dusk or when there are clouds passing by. Therefore, tuning the sensing rate
is a nontrivial problem. Since it is difficult to find a universally optimal sensing rate for
all situations, it is reasonable to ask, “Is there an efficient method by which to
dynamically adapt the sensing rate so that it is high when the system is at transient state
and low when at steady state?”
Marbini et al. proposed a generic mechanism for adaptive sampling as replicated
in Figure 5-1 [93]. However, no conclusive suggestion was made on how each critical
element – “model” and “control loop” blocks – should be implemented. Jain et al.
implemented the idea by applying a Kalman filter in the “model” block and proposed a
two-fold module – source side module and central server module – for the “control
loop” block in Figure 5-1 [94]. The source side module allows the sensor node to
dynamically adjust the sensing rate within a predetermined range according to the
prediction error of the model using a customized function. If the new sensing rate
determined by the source side module is beyond the predetermined range, the central
server module manages the requests for a new sensing rate in a centralized fashion. The
proposed adaptive strategy in [94] is suitable for query-based sensor network
configuration, but is probably not ideal for real-time feedback control applications.
62

Figure 5-1 Mechanism for adaptive sampling [93].
5.2 Algorithm and Mathematical Detail
The algorithm developed follows the idea in [93] and [94] with a better choice
of predictive model and adaptive rules. The architecture of the algorithm is shown in
Figure 5-2. In each iteration, before a sensor reading becomes available the predictive
model generates a one-step prediction of the incoming sensor reading yk ˆ( ), which is
compared to the actual sensor reading y(k) to calculate the prediction error e(k) as in
(5.1). The parameter of the predictive model is updated according to e(k) for predicting
the next incoming sensor reading yk+ ˆ( 1) . Meanwhile, the prediction error serves as an
index of how fast the environment is changing so that the fuzzy sensing rate adaptor can
adjust the sensing rate to ensure the resolution of the sensed information.
e(k) = y(k) ŷ(k) (5.1)
Figure 5-2 Architecture of adaptive sensing rate algorithm.
63

The key components are the predictive model used to generate the one-step
sensor reading prediction and the fuzzy sensing rate adaptor, which uses the prediction
error to determine sensing rate. The predictor parameters update block simply
highlights that the parameters of the predictive model get updated with each sensing
iteration according to the prediction error, but the actual parameter update is usually an
integrated part of the predicting process.
Predictive model
Three different prediction or forecasting approaches were adopted and evaluated
– Kalman filtering, adaptive Wiener filtering, and double exponential smoothing
methods.
Kalman filtering
The model is represented in state-space equations in (5.2):
x k +1
= Ax k
+ w k
, (5.2)
y k
= Cx k
+ v k
where
x k is the state of the system at time k;
A is the state transition matrix;
w k is the model disturbance at time k with covariance matrix Q;
y k is the measurement at time k;
C is the measurement matrix;
v k is the measurement noise at time k with covariance matrix R.
64

The time update equations and measurement update equations can be derived as (5.3)
and (5.6) respectively.
k
x k +1
= Ax k
k
(5.3)
k
P k +1
= AP k k A T + Q (5.4)
K k +1
= P k k +1
C T CP k
k +1
C T + R
k
x +1 k
k +1
= x k +1
( ) 1 (5.5)
k
+ K k +1 ( y k +1
Cx k +1 ) (5.6)
k
P +1 k
k +1
= ( I K k +1
C)P k +1
(5.7)
where
k
x k +1
is the predicted state of the system for time k+1 using measurements up to k;
x k k is the estimated state of the system at time k given measurements up to k;
k
P k +1
k
( )( x k +1
ˆx k +1 ) T
= E
k
x k +1
ˆx
k +1
using measurements up to time k;
is the error covariance matrix for time k+1
P k k
= E
k
x k
ˆx
k
k
( )( x k
ˆx k ) T
is the error covariance matrix of time k given
measurements up to time k;
K k+1 is the Kalman gain for time k+1.
The prediction and corresponding prediction error generated by the Kalman
filtering approach on a set of daylight data is shown in Figure 5-3, where the blue solid
line in (a) is the true daylight data sampled every ten seconds and the green dashed line
is the prediction. The mean square error (MSE) is 1800.2, which is large because the
Kalman filtering method was started with random guesses of initial values. The MSE is
22.7 if the initial guesses are not used.
65

Figure 5-3 Prediction performance of Kalman filtering.
Adaptive Wiener filtering
A one-step predictive Wiener filter can be formulated as shown in Figure 5-4,
where
Y k = [y(k-1) y(k-2) … y(k-p)] T is the vector of the past p measurements;
W(k) = [w 1 (k) w 2 (k) … w p (k)] T is the vector of the p filter parameters updated to time
k-1, and p is the order of the filter.
Figure 5-4 Adaptive Wiener filter.
66

The one-step prediction of the measurement for time k, yk ˆ( ), given the previous p
measurements can be obtained using (5.8).
The prediction error at time k is expressed as (5.9).
p
ŷ(k) = W (k) T Y k
= w i
(k)y(k i)
(5.8)
i=1
p
e(k) = y(k) ŷ(k) = y(k) w i
(k)y(k i)
(5.9)
The least mean square (LMS) algorithm is used to update the filter parameter w i (k), and
the updating rule is shown in (5.10), where μ is the step size of the LMS algorithm.
i=1
W (k + 1) = W (k) μe(k)Y k
or
w i
(k + 1) = w i
(k) μe(k)y(k i), i = 1,2,..., p
(5.10)
The prediction and the corresponding prediction error generated by the adaptive
Wiener filtering approach with two coefficients (p=2) on the same set of daylight data
as that used in Figure 5-3 is shown in Figure 5-5. The blue solid line in (a) is the true
daylight data sampled every ten seconds and the green dashed line is the prediction. The
MSE is 17.7 when the initializing random guesses are accounted for. The MSE is 6.1 if
the initial guesses are deducted.
67

Figure 5-5 Prediction performance of adaptive Wiener filtering.
Double exponential smoothing
Since the change of daylight over a day usually presents a linear trend of
increasing or decreasing over time, double exponential smoothing method is the most
reasonable choice of predictive models in the exponential smoother family. The model
is shown in (5.11), where b 1 and b 2 are constants and k is random noise with zero mean
and variance
2
.
y k
= b 1
+ b 2
k + k
(5.11)
The first order and second order smoothed statistics at time k can be calculated in (5.12)
and (5.13), where is the smoothing constant.
S k
= y k
+ (1 )S k 1
(5.12)
S [2] [2]
k
= S k
+ (1 )S k 1
(5.13)
The one-step prediction equation can then be derived as in (5.14).
68

ŷ k +1
= 2 +
1
S
1+
k
1
S k
[2]
(5.14)
The prediction and corresponding prediction error generated by double
exponential smoothing approach on the same set of daylight data as that used in Figure
5-3 is shown in Figure 5-6, where the blue solid line in (a) is the true daylight data
sampled every ten seconds and the green dashed line is the prediction. Unlike Kalman
and Wiener filtering approaches, which may be started with random initial guesses, the
double exponential soothing method has to be initiated by smoothing over the first few
data points. In this case, the first 24 data points were used for initialization, and the
MSE is 91.9.
Figure 5-6 Prediction performance of double exponential smoothing.
Compare the performance of the above three predictive models (Figure 5-3,
Figure 5-5 and Figure 5-6). Each model can provide a reasonably good prediction.
69

Among them, adaptive Wiener filtering approach generally results in the smallest
prediction error with the least effort tuning the parameters.
Fuzzy sensing rate adaptor
The adaptation of sensing rate is based on the simple logic that large prediction
errors indicate that the environment is undergoing changes, and thus the sensing rate
should be high to catch the changes. On the other hand, if the prediction errors are
small, then the changes of the environment are likely to fall into the region where the
predictive model is able to correctly predict. Hence, the sensing rate could be
reasonably lowered without compromising the resolution of the sensory information.
However, it is hard to quantitatively determine whether the prediction error at a
certain single instance is large or small without the context of what the overall
prediction errors look like. Therefore, a simple exponential moving average smoother is
implemented. The smoothed prediction error at time k is generated by smoothing over
all the prediction errors up to time k as shown in (5.15), where is the smoothing
constant, e(k) is the last prediction error and S k and S k-1 are the smoothed statistics at
time k and k-1 respectively.
ê(k + 1) e smoothed
(k) S k
= e(k) + (1 )S k 1
(5.15)
The prediction error at time k+1 is then compared to the previous smoothed prediction
error using (5.16) to determine if the prediction error is large or small.
(k + 1) = e(k + 1) ê(k + 1) (5.16)
A set of fuzzy rules is implemented to adapt the sensing rate:
70

IF (k) small, THEN sensing rate low;
IF (k) medium, THEN sensing rate moderate;
IF (k) large, THEN sensing rate high.
The membership functions for determining the sensing rate are shown in Figure
5-7. The membership functions on the left are for fuzzification, where m is the
fuzzification parameter and max is the upper limit beyond which (k) is
deterministically considered as large. The membership functions on the right are for
defuzzification, where m SR is the defuzzification parameter, and SR min and SR max are the
minimum and maximum allowed sensing rate respectively. While m and m SR need to be
tuned, max can be specified from the nature of the sensed environment and SR min and
SR max are determined according to the specifications of the sensing tasks.
Figure 5-7 Membership functions for determining sensing rate.
The promising performance of the adaptive sensing algorithm is verified through
simulation in the next section as shown in Figure 5-9, Figure 5-10, and Figure 5-11. It
is, however, not impossible to encounter a circumstance where (k) becomes very small
71

even though the sensed environment is still under dramatic change. If this kind of
situation occurs, the algorithm will automatically lower the sensing rate in response,
which in turn causes loss of valuable sensory information. Therefore, it may be more
logical to gradually lower the sensing rate so as to mitigate information loss in case the
sensed environment is not steady for certain. On the contrary, the sensing rate should be
increased sharply to guarantee the resolution of the sensed data as soon as any
transience is detected. For this purpose, the fuzzy adaptive rules are extended to
introduce “damping” when adapting the sensing rate so that if the sensing rate tends to
lower, it will be decreased slowly. This works as follows:
IF (k) small, THEN sensing rate low AND damping heavy;
IF (k) medium, THEN sensing rate moderate AND damping moderate;
IF (k) large, THEN sensing rate high AND damping light.
The extended membership functions for the revised fuzzy rules are shown in
Figure 5-8, where the membership functions in the left two plots are the same as the
previous design and the membership functions in the plot to the right are for
defuzzification of the damping ratio. One additional defuzzification parameter, m damp ,
needs to be tuned for proper damping. The damping ratio will be in the range of [0, 1].
72

Figure 5-8 Membership functions for determining sensing rate with damping.
At each sensing instance, two values are generated from (k) using the fuzzy
rules – an updamped sensing rate SR undamped (k) and a damping ratio (k). SR undamped is
exactly the sensing rate in the undamped version of the adaptation mechanism. The
damped sensing rate, denoted as SR damped (k), is calculated using (5.17), where
SR damped (k-1) is the preceding sensing rate.
SR damped
(k) = SR damped
(k 1) + (k) ( SR undamped
(k) SR damped
(k 1) ) (5.17)
Equation (5.17) can be derived from (5.18), which interprets the current sensing rate
SR(k) as the previous sensing rate SR(k-1) adjusted by the difference between the two
consecutive sensing rates. Modifying (5.18) by multiplying the second term by a proper
damping ratio, the formula in (5.17) is obtained.
SR(k) = SR(k) SR(k 1) + SR(k 1)
= SR(k 1) + ( SR(k) SR(k 1) )
(5.18)
73

5.3 Simulation and Experiment Results
Each of the predictive models was integrated with the fuzzy sensing rate adaptor
and simulated using the same set of daylight data as that used in Figure 5-5. The
daylight data were sampled at a fixed rate of ten seconds/sample using one mote
photosensor. A total of 1851 consecutive data points were collected.
The fuzzy parameters were tuned as follows: m was tuned to 8 and max was set
to 20; SR min was arbitrarily set to 600 seconds/sample (10 minutes/sample); SR max was
set to the highest possible resolution of the daylight data – 10 seconds/sample, the rate
at which the daylight data was sampled; m SR was tuned to 250 seconds/sample; m damp
was tuned to 0.3 for the damped version of the algorithm.
Figure 5-9, Figure 5-10, and Figure 5-11 show the performance of the sensing
rate adaptation mechanism with Kalman filtering, adaptive Wiener filtering, and the
double exponential smoothing model respectively before applying any damping. The
solid blue line in (a) of each figure is the daylight data sampled uniformly at 10
seconds/sample, and the dashed green line is the daylight information reconstructed
from the adaptively sensed data. Compared to the original 1851 samples, only 164, 174,
and 262 points were sampled with respect to each predictive model during the entire
sensing task. Plot (b) of each figure details how the sensing rates were adapted over
time. All of the predictive models resulted in reasonably good abilities of preserving the
original daylight information.
74

Figure 5-9 Adaptive sensing with Kalman filtering predictive model.
Figure 5-10 Adaptive sensing with Wiener filtering predictive model.
75

Figure 5-11 Adaptive sensing with double exponential smoothing predictive model.
A close examination of the adaptation of the sensing interval in plot (b) of
Figure 5-9 to Figure 5-11 reveals that the sensing rate changed quite randomly. There is
no clear correlation between two consecutive sensing intervals. The sensing rate can be
very high in one instance, but determined to be very low in the following instance. For
example, around the 157 th minute in Figure 5-10(b), the sampling interval was adapted
to 450 seconds/sample although the previous ones were much lower. This caused
undesirable loss of information as seen in Figure 5-10(a) – the gap between the solid
blue line and the dashed green line reflects the fact that the adapted sensing interval
failed to capture the daylight change in the blue valley.
Figure 5-12, Figure 5-13, and Figure 5-14 show the performance of the damped
sensing rate adaptor with Kalman filtering, adaptive Wiener filtering, and the double
exponential smoothing model respectively. Like the figures for the undamped version,
the solid blue line in (a) of each figure is the daylight data sampled uniformly at 10
76

seconds/sample, and the dashed green line is the daylight information reconstructed
from the adaptively sampled data. Compared to the original 1851 samples, 236, 227,
and 358 data points were sampled with respect to each predictive model during the
entire sensing task. In other words, only less than one-fifth of the sensing effort was
required to pertinently represent daylight change in that period. Plot (b) of each figure
details how the sensing rates were adapted over time.
Figure 5-12 Damped adaptive sensing with Kalman filtering predictive model.
77

Figure 5-13 Damped adaptive sensing with Wiener filtering predictive model.
Figure 5-14 Damped adaptive sensing with double exponential smoothing predictive model.
At the cost of more samples due to the additional damping, each adaptive
sensing task resulted in even better fidelity in capturing the daylighting information than
its undamped counterpart. In plot (b) of Figure 5-12 to Figure 5-14, it is obvious that the
78

sampling intervals did not randomly jump up and down anymore but increased
gradually and decreased sharply. Also, the maximum adapted sampling intervals were
never as large as those without damping. Finally, it’s worth emphasizing that the
reduced sensing effort means less frequent wireless communication for real-time
lighting control, which translates into longer life spans of the battery-powered wireless
photosensors.
79

Chapter 6
Optimal Lighting Actuation
The optimal lighting actuation algorithm is based on linear programming
optimization. Taking advantage of individual addressability and controllability of the
wireless-enabled luminaires, this algorithm generates the optimal settings for each
luminaire, which results in minimum energy usage and maximum user satisfaction.
6.1 Rationale
The wireless capabilities of the luminaires in the research system are enabled
with mote actuation interfaces. In addition to the obvious benefit of circumventing
costly rewiring when retrofitting the existing lighting system with the research system,
each luminaire is individually addressable, and hence can be controlled independently.
Moreover, the luminaires are made dimmable after they are integrated with the wireless
actuation interface. In other words, instead of controlling all the lights with one or a few
switches as in a traditional lighting configuration, each of the wirelessly controllable
luminaires acts like an independent dimmer/switch in the research system. Given the
greatly increased degree of freedom, it is possible to generate more energy savings by
efficiently turning off the luminaires in any unoccupied area.
Another superiority of individually addressable luminaires is the ability to better
satisfy occupants’ lighting preferences and requirements by delivering the desired
amount of light from luminaires proximate to each occupant. Studies conducted in
typical office environments have shown a positive correlation between lighting
satisfaction and the productivity of the occupants [95]. Researchers have also identified
80

the significantly diverse preferences and requirements for lighting among individuals as
well as by the same person for different tasks [6, 7]. Therefore, individually addressable
luminaires present the promising potential to optimize user’s comfort in line with the
recommendations in the Daylight and Electric Lighting Control Systems Design Guide
of the International Energy Agency [52].
Furthermore, satisfying occupants’ lighting preferences may also contribute to
significant energy savings. It has been a common misunderstanding that satisfying
personal visual comfort is the contrary to energy savings. Nonetheless, recent studies
have revealed that it could be energy efficient on average if occupants are granted the
option of working under their ideal light settings [22, 96].
In summary, the challenge of designing a lighting system with individually
addressable luminaires lies in figuring out how to optimize energy usage and user
satisfaction given that occupants have diverse or even competing lighting preferences.
The research presented in this chapter proposes a lighting optimization algorithm
capable of delivering the optimal lighting that minimizes overall energy usage while
satisfying occupants’ diverse or even competing lighting preferences and requirements.
6.2 Open-loop Lighting Optimization Algorithm
Architecture
Figure 6-1 shows the architecture of the optimal lighting actuation algorithm.
The overall lighting in an office is considered as a linear combination of the light
contributions from each of the luminaires. In the illuminance model generator, a model
81

of the workplane level illuminance for the entire room is generated for each of the
luminaires. Figure 6-2 shows an example of one such model, where the x and y axes are
the room’s dimensions. The lighting optimizer calculates the optimal linear combination
of the individual illuminance models that minimize the entire lighting output, and hence
the energy consumption, while meeting the present occupants’ lighting preferences even
under possible conflicts. The optimal settings of each luminaire are wrapped into
actuation command packets and sent to each luminaire wirelessly. The wireless-enabled
luminaires subsequently translate actuation commands to corresponding signals to
dim/lighten the lights or toggle the lights on/off.
Figure 6-1 Architecture of the optimal lighting actuation algorithm.
The light setting control implemented in the lighting optimizer is formulated as a
linear programming problem. The objective is to minimize the resulting illuminances at
the workplane level, and the constraints are the lighting preferences of the presented
occupants. Since power consumption is proportional to the light output from the
luminaire [97], minimizing the light output is equivalent to minimizing energy usage.
82

Figure 6-2 Workplane level illuminance distribution model.
Illuminance model generator
The room is first geographically discretized into a grid of squares with
predefined resolution, and utilizing RADIANCE Synthetic Imaging System [98], the
workplane level illuminance at the center of each small square is calculated with respect
to each luminaire. This process requires basic knowledge of the room’s configuration,
including room dimensions, luminaire locations, surface reflectance, etc., and only has
to be done once as long as there is no change to the room configuration. The procedure
of generating the illuminance model is inspired by SPOT [99], a sensor placement
software that helps designers to optimize photosensor locations for better daylight
harvesting performance. The model generated is represented in a matrix where each of
the elements is the workplane level illuminance corresponding to each square. Take an
83

office with K luminaires, for example, and suppose it is discretized into a grid of mn
squares. The generated illuminance models are K m-by-n matrices, l 1 , l 2 ,…,l K ,
associated with each of the K luminaires indicated by the superscript number. Equation
(6.1) is the symbolic representation of the i th illuminance models matrix, l i .
i
i
l 11
… l 1n
l i =
i
i
l m1
l mn
(6.1)
Lighting optimizer
The illuminance of the room at the workplane level (E) is represented as the
linear combination of each model as shown in (6.2), where d i is the light level output of
each luminaire.
i
i
e 11
… e 1n
K
l 11
… l 1n
K
E=
= d
i
l i = d i
(6.2)
e m1
e i=1
i=1 i
i
mn
l m1
l mn
For mathematical manipulation, each matrix is rearranged into a column vector
by concatenating the columns, denoted 1 , 2 ,…,
K
l l l , and (6.2) can be rewritten as a
pure matrix operation expressed in (6.3). The operator L, with the rearranged vectors of
the models as its columns, defines the transformation from a vector of light output level
d into the resulting workplane level illuminance E of the room. E is the vector of
concatenated columns of E due to the rearrangement of the illuminance models.
84

E =
e 11
e m1
e 12
e m2
e 1n
e mn
1
l 11
1
l
m1
1
l
12
| | |
= l 1 l 2
l
K 1
d=
l m2
| | |
1
l
1n
1
l
mn
2
l 11
2
l m1
2
l 12
2
l m2
2
l 1n
12
l mn
K
l 11
K
l
m1
K
l
12
K
l m2
l 1n
K
K
l mn
d a
d b
d l
Ld (6.3)
The objective is to find an optimal set of light output levels d so as to result in a
properly illuminated room E that satisfies each occupant’s lighting preference. The
occupants’ lighting preferences are also specified in correspondence with the small
squares in the grid of discretized room, which are the locations of their workplanes.
Since the personally specified points of interest are most likely confined to a small area
rather than the entire room, it is unrealistic and unnecessary to artificially generate the
entire E for finding the optimal light setting. Therefore, the reduced-order vector E sub
,
which contains only the specified illuminances at the points of interest, is considered.
Likewise, the order of the operator L is reduced to obtain the corresponding matrix L sub
and (6.3) is then condensed to (6.4), where e pq ,e rs ,…,e xy are the desired illuminances at
the specified locations. The goal then becomes finding the optimal set of light output
levels d that satisfy the occupants’ lighting preferences E sub
.
85

E sub
=
e pq
e rs
e xy
1
= L sub
d =
1
1
l pq
l rs
l xy
2
l pq
2
l rs
2
l xy
K
l pq
K
l rs
K
l xy
d 1
d 2
d K
(6.4)
This problem is formulated into a linear programming problem as shown in (6.5)
. By minimizing the 1-norm of vector d, the summation of the output levels from each
luminaire is minimized, which translates into minimizing the energy usage of the
resulting settings. The lighting preferences of the occupants are posed as the constraints
to the linear programming problem. The physical dimming capabilities of each
luminaire, DimLevel min and DimLevel max , are also considered as constraints.
min d
1
, subject to
L sub
d = E sub
,
(6.5)
DimLevel min
d DimLevel max
However, this optimization problem may not have a solution, depending on how
each of the points of interest is specified in the equality constraint. Infeasible problems
are most likely caused by the occupants’ conflicting lighting preferences. In case of
infeasibility, the equality constraints are relieved to inequality constraints, thus allowing
some tolerances. The relief of the equality constraints makes sense in that the physical
meaning of this is to allow the lights at the points of interest to be regulated in certain
tolerable ranges ( tol ) instead of demanding the exact illuminances. Studies have shown
that people are insensitive to 20% of illuminance changes and are willing to accept an
illuminance change up to 30% [100]. In summary, the algorithm starts with the original
linear programming problem with equality constraints and gradually expands the
86

tolerable range by relieving the equality constraints into inequality constraints as shown
in (6.6).
min d
1
, subject to
L sub
d E sub
+ tol
,
L sub
E
L sub
d E sub
tol
,
L sub d
sub
+ tol
E sub
+ tol ,
(6.6)
DimLevel min
d DimLevel max
6.3 Lighting Optimization Algorithm with Sensory Feedback
The basic optimal lighting actuation algorithm is a single-iteration operation that
assumes perfect illuminance model and no extraneous source of light. It is, however, not
possible to capture every detail of the room configuration and account for
manufacturing tolerances of the lights to generate accurate workplane level illuminance
models. More importantly, daylight will always be an uncontrollable external light
source for a daylight harvesting system, so the algorithm must be able to incorporate
available daylight for optimal lighting actuation. Therefore, the open-loop lighting
optimization algorithm is extended by closing the control loop with sensory feedback to
compensate for uncertainties and external light sources.
Figure 6-3 shows the block diagram of the extended lighting optimization
algorithm. The behaviors of the illuminance model generator, lighting optimizer and
wireless-enabled luminaires are exactly the same as those described before. The
photosensors placed at each of the locations of interest sense the light, which may be a
mixture of electric light and daylight, and transmit the measurement back to the lighting
87

optimizer for a consistency check. The constraints of the linear program are revised in
response to the result from the consistency check for the next iteration of optimization.
Figure 6-3 Lighting optimization with sensory feedback.
For the first iteration of lighting optimization before any sensory feedback, the
constraints derived from occupants’ preference are the ones shown in (6.4). Suppose the
actual illuminances on the points of interest measured by the photosensors are
S 1
=
s pq
(1) s rs
(1) s xy
(1)
T
, where the subscript of the bold-face capital s and the
number in the parentheses denotes the number of the iteration, then the constraints will
be modified to (6.7) for the next iteration of optimization (6.8). Each iteration brings the
delivered lighting closer to the desired lighting, and the algorithm terminates when the
difference between the desired and the sensed illuminance fall into a tolerable range ().
The pseudo code of the algorithm is illustrated in Figure 6-4.
88

L sub
d = E sub
+ ( E sub
S 1 )=
min d
1
, subject to
e pq
e rs
e xy
e pq
s pq
(1)
e
rs
s rs
(1)
+
T 1
(6.7)
e xy
s xy
(1)
L sub
d = T 1
,
(6.8)
DimLevel min
d DimLevel max
6.4 Simulation and Experiment Results
The optimal lighting actuation algorithm was simulated and verified using a
shared-space office, which is part of the implementation described in Chapter 9.2. The
30-by-19-by-143 office contains ten personal workstations and some shared
worksurfaces as illustrated in Figure 6-5. The workplanes are 29.25 above the floor.
There are twelve 2-lamp troffers hanging 4 from the ceiling evenly mounted in the
office as the 12 equally spaced rectangles represent in Figure 6-5. The numbers marked
on the corner of the troffers indicate the order in which the optimal light output for the
luminaires will be presented in the following testing. The reflectance of the floor, walls
and ceiling is measured at 10%, 50% and 30% respectively. As the illuminance model
generator in the optimal lighting actuation algorithm requires a wall as a boundary, the
model placed an artificial wall on the north side of the part of the room under study,
whereas the actual room is connected to a meeting area with no physical separator in
between. No window was considered at this point as the there is no window in this inner
office.
89

Figure 6-4 Pseudo code of the lighting optimization algorithm with sensory feedback.
90

Figure 6-5 Floor plan of the experiment office.
6.4.1 Open-loop Lighting Optimization Algorithm
Two representative scenarios were considered. It was assumed that the
occupants only specified their preferred task illuminance at their desktop. Preferences
concerning the surroundings of the workplanes could be accounted for by specifying
more points of interest in the lighting optimization algorithm. The office was divided
into a grid of 4-by-4 squares in the illuminance model generator. Other practical
factors were also taken into account: (1) the resulting optimal light settings were
represented in the percentage of the maximum output of the luminaires and rounded to
integers as the dimming resolution of the wireless ballast actuator is discrete (256
distinct levels); (2) the light levels were set to 5% whenever the optimized output was
less than 5% but greater than 0 since the effective output range of typical dimming
ballasts is 5-100%. The light would be turned off if the optimized light setting was 0.
The first scenario considered a sparsely occupied office where only four
occupants were present, each with a unique lighting preference. The purpose of this
91

simulation was to show that the optimized lighting not only managed to meet each
occupant’s preference but also efficiently conserved energy by not illuminating
unoccupied areas. The resulting illuminance at the workplane level is shown in Figure
6-6, and the optimal light settings determined by the algorithm for each of the 12
luminaires are [74%, 31%, 79%, 0%, 100%, 5%, 0%, 0%, 0%, 100%, 0%, 0%] ordered
by the numbers annotated in Figure 6-5. Compared to the original lighting configuration
of the office where the luminaires were wired to be turned on/off all together, only
32.4% of the light, and hence the energy, was used to achieve the occupants’ lighting
requirements. To verify the simulation, a light meter was placed at the desktop where
preferred illuminances were specified. The numbers at the bottom of each group of
three stacked numbers in the luminance contour plot in Figure 6-6 are the specified
preferred illuminances (in lux), those in the middle are the light levels resulting from
the optimization algorithm, and those at the top are the actual illuminances measured by
the light meter.
Figure 6-6 Optimal lighting for scenario 1.
92

The second scenario assumed a more densely occupied office with seven
occupants present, and the lighting preferences of some occupants sitting adjacent to
each other varied drastically. This simulation demonstrated the capability of the system
to balance diverse and conflicting lighting preferences while delivering reasonable
lighting to all occupants. The resulting workplane level illuminance is shown in Figure
6-7, and the corresponding optimal settings for each luminaire are [76%, 27%, 100%,
100%, 100%, 60%, 96%, 0%, 21%, 0%, 44%, 0%]. Only 52.0% of the light was used
compared to the original all-on/off lighting configuration. The optimized light setting,
the numbers in the middle of each group of three stacked numbers in Figure 6-7,
diverged from the specified values (the numbers at the bottom) due to the necessary
compromise of diversely specified preferences between adjacent occupants.
Nonetheless, the optimized lighting stayed within 15% of the specified illuminances, as
did the actual measurements.
Figure 6-7 Optimal lighting for scenario 2.
93

It is obvious from Figure 6-6 and Figure 6-7 that the measured illuminances
differed from the calculated optimal illuminances more significantly towards the north
side of the office (right side) where an artificial wall was placed in the room to simulate
an interior room. The additional absorption and reflection introduced by the virtual wall
might have rendered the illuminance models near the north side of the office less
accurate. Since the illuminance models were generated with respect to an empty room,
the furniture in the real office could have also contributed to the inaccuracies in the
models. Moreover, the lighting hardware, namely the dimming ballasts and fluorescent
lamps, might not guarantee a consistent mapping from the dimming signals to the actual
output luminous due to manufacturing tolerances. These uncertainties signified the need
of sensory feedback for better performances.
6.4.2 Lighting Optimization Algorithm with Sensory Feedback
The sensor information was incorporated for testing the optimal lighting
actuation algorithm with sensory feedback to eliminate possible modeling uncertainties
as well as to respond to daylight. Four occupants were present in the office, and the
photosensors were placed on their desktops to measure the task illuminances and send
the lighting information back for the next iteration of optimization. The algorithm was
set to terminate when the difference between the specified lighting preferences and the
actual illuminances were within five lux for all occupants. As shown in Figure 6-8, it
took four iterations for the light to converge to each occupant’s preference. The top
number of each stack of five numbers represents the preferred lighting specified by each
occupant, and the following four numbers are the sensor readings after the first, second,
94

third and fourth iteration respectively. The final optimal settings for each luminaire are
[60%, 39%, 47%, 0%, 100%, 0%, 0%, 0%, 0%, 82%, 0%, 0%] ordered by the numbers
annotated in Figure 6-5, and only 27.3% of the energy was used to deliver the lighting
compared to the original all-on/off lighting configuration.
Figure 6-8 Optimal lighting with sensory feedback.
Having discussed the mathematical details of the lighting optimization algorithm
as well as the sensor fusion and adaptive sensing algorithms in Chapter 4 and Chapter 5,
the next chapter describes the design of the wireless-enabled components that are used
for verifying the developed algorithms. The wireless-enabled components, including the
wireless photosensors and the wireless actuation modules, are also the key elements for
a series of implementation of the intelligent lighting system, which will be discussed in
later chapters.
95

Chapter 7
Wireless-Enabled Lighting Component Design
and Integration
7.1 Overview
In order to implement the research system, prototypes of the hardware
components including the mote photosensors and the mote-based ballast actuation
modules were designed and built. A sensor board that integrates with the commercially
available wireless mote platform was designed to serve as wireless photosensor for light
sensing purposes. An actuation module incorporating the mote platform was developed
for enabling the wireless capability of the luminaires. This wireless actuation module
interfaced with the dimmable ballast to dim the electric lights or toggle the lights on/off
wirelessly.
The mote photosensors and wireless ballast actuation modules along with the
algorithms described in the preceding chapters were implemented in rooms with
different scales, including a prototyping luminaire structure, a small private office, and
a small shared-space office. The implementations will be discussed in the subsequent
chapters.
7.2 Mote Photosensor Design
The design of the wireless photosensors involved photosensitive element
selection and auxiliary circuit design. Since the photosensors are meant for estimating
the light that the occupants perceive, the photosensitive element must have a similar
response to that of human eyes throughout the span of the light spectrum. The
96

corresponding ancillary circuitry amplifies the signal from the photosensitive element to
match the input range of the analog-to-digital converter (ADC) on the mote platform.
The Hamamatsu S7686 silicon photodiode was selected as the photosensitive
element. This specific photodiode possesses sensitivity close to that of human eyes and
is color-corrected to match the CIE curve. The CIE curve is a color model based on
human perception that was developed by the Commission Internationale de l’Eclairage
(the International Commission on Illumination) committee. The spectral response of the
photodiode is reproduced in Figure 7-1, in which the CIE curve is overlaid.
Figure 7-1 Spectral response of Hamamatsu 7686 photodiode [101].
Since the wireless photosensors are meant for being deployed on desktops to
measure workplane illuminance, the normal operating range of the photosensor is
97

determined to be 0-2000 lux. Consequently, the supporting circuitry was designed to
translate the electronic signal produced by the photodiode in the normal operating range
to span the full voltage range of the mote ADC input. The resulting photosensor boards
integrated with different mote platforms are shown in Figure 2-5, and the electronic
circuit schematics can be found in Appendix A.1 and A.2. Except for the spectral
response, the behaviors of the wireless mote photosensors, including the sensor reading
acquisition rate, data processing, the communication and networking scheme, etc., are
defined in the operating system of the mote.
Figure 7-2 illustrates the calibration curve of one of the photosensor boards. The
sensor board was calibrated against a high fidelity light meter. Although the operating
range of the photosensor was set to 0-2000 lux, the controlled lighting used in this
calibration can only provide at most 1000 lux on the test surface. The illuminance was
increased from 0 to 1000 lux with a 50 lux interval, and the digital reading from the
ADC on the mote platform was recorded. The blue dots show the data points of the
illuminances versus the ADC output readings, and the red line is a first-order
polynomial fitting. The plot confirms that the response of the photosensor is linear.
98

Figure 7-2 Calibration curve of the mote photosensor.
7.3 Mote-based Ballast Actuation Module Design
7.3.1 Design Requirement Identification
The dimmable ballasts with which the actuation modules interface are the 0-10V
dimming ballasts, the mainstream dimming technology for linear fluorescent lamps on
the market. In addition to being powered by the mains like typical non-dimmable
ballasts, the dimmable ballasts deliver 5% to 100% of light in response to 0-10VDC
voltage signals. Therefore, the main requirement for the actuation module is to be able
to output 0-10VDC voltage signals to the dimmable ballasts.
Although the dimmable ballasts can take any voltage signal from 0 to 10VDC
and dim the light continuously, the motes operate in a digital fashion and only output
finite numbers. Thus, to ensure the highest actuating resolution, it is necessary that the
99

discrete outputs from the motes span the full range of the 0-10VDC signals
recognizable to the dimmable ballasts.
A characterization was conducted on one of the most popular 0-10V dimmable
ballasts on the market, Mark VII by Advance Transformer [102], in order to gain indepth
understanding of the behavior and response of dimmable ballasts. A luminaire
equipped with a Mark VII dimmable ballast was hung 1.2 meters above a light meter.
The ballast control voltage was increased stepwise from 0 to 10VDC and then
decreased to 0 with an interval of 0.2VDC, and the meter readings were recorded
manually. The result is plotted in Figure 7-3, where the blue solid line and the green
dashed line represent the light output when the ballast control voltage was increased
from 0 to 10VDC and decreased from 10VDC to 0 respectively. Although the nominal
operation range of the dimmable ballast is 0 to 10VDC, the actual effective range was
identified to be 1-8VDC. Therefore, the output range of the actuation module was
designed in the range of 1-8VDC instead of 0-10VDC so that each dimming state would
be distinct. The fact that the two lines in the plot don’t coincide suggests that there
exists hysteresis or memory effect. The maximum discrepancy between the two lines at
the same voltage is 13.7%. It is not possible to definitively determine whether this effect
comes from the fluorescent lamps or the ballast or both. This doesn’t affect the design
of the actuation module, but it implies that it is unrealistic to derive one-to-one mapping
from ballast control voltages to the light outputs.
100

Figure 7-3 Advance Mark VII characterization curve.
It is desirable to be able to turn the lights on and off wirelessly in addition to
being able to dim the lights. Also, it is preferable that the modules are directly powered
from the mains since they always have access to the power line. This would minimize
the maintenance required to change batteries which is common to most other sensor
network applications. Furthermore, since the actuation module is an add-on to the
luminaire, ease of installation is an important factor to guarantee effortless retrofitting
and deployment.
7.3.2 Design Iterations
Three iterations of the mote-based actuation modules have been developed due
to the employment of different versions of mote platforms, the improvement of
101

actuation technologies, and additional features that better meet the design requirements.
Table 7-1 summarizes the three generations of mote-based ballast actuation modules.
Table 7-1 Comparison of wireless ballast actuation modules.
Generation 1 st Generation 2 nd Generation 3 rd Generation
Mote platform MICA MICA2 Tmote Sky
Dimming method
Outboard 4-bit DAC
Outboard DPP +
voltage divider
Onboard 8-bit DAC
Dimming resolution 16 100 256
On/off control No Yes Yes
Mote power Batteries Line power Line power
Failsafe No No Yes
Retrofitting wiring A lot 6 5
Implementation
Prototyping
luminaire structure
Small private office
Small shared-space
office
The first generation of the ballast actuation module shown in Figure 7-4 is a
proof of concept demonstrating the feasibility of actuating dimmable ballast with the
mote technologies. The mote platform used in this generation was MICA platform, one
of the earliest versions of mote platforms. The block diagram in Figure 4-3 shows the
logic of the actuation module and the electronic circuit schematics can be found in
Appendix A.3. The four general input/output (GIO) ports on the mote are connected to
an external digital-to-analog converter (DAC) chip, and the 16 different on/off
combinations of the four GIO ports are translated to 16 distinct DC voltage outputs,
which are amplified to 1-8VDC dimming signals to control the ballast.
102

Figure 7-4 First generation wireless ballast actuation module.
Figure 7-5 First generation wireless ballast actuation module operation logic.
This module was installed on the prototyping luminaire structure and
successfully showed the possibility of dimming the electric lights wirelessly with a
mote platform, but the mote still operated on batteries and the lights couldn’t be toggled
on/off using the module. Since the dimming resolution is low, the light change between
two consecutive dimming states is very noticeable to a naked eye.
The second generation of the ballast actuation module integrated with a newer
version of the mote platform, MICA2, improves the dimming resolution to 100 distinct
states by utilizing a 100-stage digitally programmable potentiometer (DPP) and a
103

Figure 7-8 Voltage divider.
Since the DPP can be adjusted among 100 different resistances, the voltage
divider can output 100 distinct voltages. The output of the voltage divider is amplified
to 1-8VDC ballast control signals. In addition to the increased dimming resolution, this
generation takes more fundamental design requirements into account. One of the four
GIO ports on the mote is dedicated to toggle a power relay, which in turn connects or
disconnects the power to the ballast and hence turns the lights on or off. The supply
voltage is further regulated to directly power the mote platform on the module,
eliminating the need of batteries. Moreover, only six lines are required to interface this
actuation module between the mains and the dimmable ballast. This iteration of
actuation module was used for the implementation of a private office as described in
Chapter 8.2.
The third generation of the ballast actuation module is integrated with a new
family of mote platform, Tmote Sky. In addition to several conveniently accessible GIO
ports like a MICA mote, a Tmote Sky is also equipped with an onboard 8-bit digital-toanalog
converter (DAC), which is suitable for actuation purposes. The onboard DAC
can deliver 256 different output voltage signals, and hence 256 distinct dimming states,
which doubles the actuation resolution a 100-stage DPP can possibly provide. The
105

outputs from the DAC are amplified to 1-8VDC ballast control signals to dim the lights.
Figure 7-9 shows the actuation module, and the block diagram in Figure 7-10 illustrates
the operation of the actuation module. The electronic circuit schematics can be found in
Appendix A.5.
Figure 7-9 Third generation wireless ballast actuation module.
Figure 7-10 Third generation wireless ballast actuation module operation logic.
106

Given the 256-level dimming capability, this module is competitive with the
most advanced digital dimming ballast on the market, the digital addressable lighting
interface (DALI). This iteration not only further improves the dimming resolution but
also retains all the promising features from the preceding generation, including the linepowered
mote and the ability to toggle the lights on/off. Only five wires need to be
connected to interface this module with the dimmable ballasts. A failsafe mechanism is
also introduced into this actuation module. Should the mote fail for any reason, the
module is automatically bypassed and the lights can be manually turned on or off using
the wall switch as if the actuation module never existed. This actuation module was
adopted in the implementation of the research system in a small shared-space office as
described in Chapter 9.2.
7.4 Initial Implementation on Prototyping Luminaire Structure
The mote photosensors and the wireless ballast actuation module were deployed
and tested on a prototyping luminaire structure. In addition, the mote-FVF algorithm
developed in Chapter 4 along with a simple rule-based control law were also
implemented to form a closed-loop desktop illuminance regulation system.
Furthermore, the mote-FVF algorithm and the control law were coded into the mote
photosensors and the mote-based ballast actuator respectively so that the closed-loop
system could autonomously perform fusion and use the fused data to dim or brighten
the light.
107

7.4.1 Prototyping Luminaire Structure
The prototyping structure shown in Figure 7-11 is made of PVC pipes and is
highly configurable for software and hardware testing. This luminaire structure has
served for the following purposes:
• A platform for testing and verifying the developed algorithms such as the mote-
FVF sensor validation and fusion algorithm presented in Chapter 4.
• A hardware testbed for the calibration of wireless mote photosensors, and the
development of wireless ballast actuation modules mentioned in the previous
section.
• A prototyping closed-loop lighting control system integrating the intelligent
algorithms, the wireless photosensors and the wireless ballast actuation module.
Setup
This luminaire structure comprises one 4-lamp linear fluorescent lighting fixture
and a dimmable ballast. The luminaire is hung about 1.2 meters from the foot of the
structure, and is a little lower than the typical troffer mounting height in regular offices
when placed on top of a desk. Instead of being powered by the mains, the electricity is
supplied by plugging into a regular power outlet. The terminals of the electricity power
line and the ballast are configurable so that they are able to interface with different
ballast actuation hardware.
108

Figure 7-11 Prototyping Luminaire Structure.
7.4.2 Desktop Illuminance Regulation with Distributed Mote-FVF Algorithm
A desktop illuminance regulation system is realized on the prototyping structure
by integrating the wireless mote photosensors, the wireless ballast actuation module,
and the mote-FVF algorithm. An experiment was designed to demonstrate the
feasibility of embedding the developed sensor fusion algorithm in each mote sensor.
The mote-FVF algorithm described in Chapter 4 has shown promise in extracting
pertinent sensor information from a cluster of sensors while isolating disturbed or faulty
sensors. However, it may create heavy wireless network traffic and significant energy
drain if all the sensor nodes have to transmit readings to a designated central unit, which
could be at a distance, for validation and fusion. If a cluster of mote sensors can perform
109

low-power, short-distance communication and execute intra-network sensor fusion, then
only a single fused reading has to be transmitted to a central processing unit for control
purposes from each cluster at each time stamp. The mote-FVF algorithm is believed to
be lean enough in its current form for a battery powered mote platform to carry out the
computations with its available resources.
Experiment setup
Six mote photosensors, each of which was programmed with a cluster leader
electing logic and the mote-FVF algorithm, were placed on the testbed under the
prototyping luminaire structure. These six mote sensors form a cluster and should
generate one fused sensor reading each time. Upon being powered up, a timer also starts
in each of the motes. If the timer expires in a mote, it claims itself as the cluster leader,
requests sensor readings from its fellow motes, executes the mote-FVF algorithm, and
sends out the fused readings. If a mote is asked for readings by another sensor before its
timer expires, it sends out its reading, resets the timer and becomes a cluster member.
This simple cluster leader electing logic guarantees that there is always one leader
responsible for performing sensor fusion. More sophisticated logic may be adopted to
select a cluster leader, such as one considering the remaining energy in each mote, but
is out of the scope of this test. The mote-FVF algorithm, which was originally coded in
Matlab ® , was rewritten in nesC as a function and incorporated into TinyOS, the
operating system of the motes. The majority-voting scheme required in the validation
unit of the mote-FVF algorithm was implemented with the median value approach due
110

to its simplicity. Although coded with a different language, the integrity of the mote-
FVF algorithm remained intact.
The prototyping luminaire structure was equipped with the mote-based ballast
actuation module so as to dim the lights wirelessly. Furthermore, the experiment system
was made into a self-contained illuminance regulation system that maintained the
illuminance on the testbed at a specified value without the presence of a central
computer to coordinate the control actions. The mote integrated with the actuation
module was embedded with a simple rule-based control law, and it listened to the fused
readings from the cluster of mote sensors. Upon receiving a fused value from the sensor
cluster, the actuation mote calculated and performed the control action based on the
control law in order to maintain the testbed illuminance at the setpoint. Although no
extraneous computer was required in the experiment system, a computer was still
employed to intercept the readings from each sensor and the fused values for data
gathering purpose. The default setpoint of 500 lux was hard-coded in the actuation
mote, which could be changed in real-time by sending a management command to the
actuation module specifying a new setpoint from the computer.
Experiment results
The first experiment lasted over 17 hours to show that the system can sustain
long-term unattended operations. The luminaire structure was located in a research lab
without windows to introduce natural daylight perturbation. In order to evaluate the
system’s ability to regulate the desktop illuminance at the setpoint, the overhead
111

luminaires were used to simulate extraneous lighting. The setpoint of this experiment
was specified at 550 lux at all times.
Figure 7-12(a) shows the fused values (blue solid line) and the readings from
one of the mote sensors (magenta dashed line) gathered at the computer. For clarity and
readability, the raw data from the other five sensors are omitted in this plot. The big
spikes indicate the times when the overhead lights were turned on or off and the overall
illuminance on the desktop was rapidly brought back to 550 lux. Due to the large
number of data points squeezed into the plot, the fused line in Figure 7-12(a) seems
thick and fluctuates significantly. Figure 7-12(b) focuses on the fused value from the
first half hour of Figure 7-12(a) to show the performance of the system with better
resolution. The desktop illuminance was regulated within 5% of the setpoint during the
entire experiment. The magenta dashed line, which is based on the raw readings from
sensor No.3, stopped around the fourth hour because the mote ran out of batteries. The
fact that the blue line, the fused values, was not affected by the malfunctioning sensor
verified the robustness of the mote-FVF algorithm under sensor failure.
(a)
(b)
Figure 7-12 Long term desktop illuminance regulation.
112

The second experiment was executed in the same environment as the first one
with the same setup. The setpoint was changed several times during the 47-minute
period in order to test if the illuminance on the desktop can be maintained at a new
setpoint. The test started with the setpoint set to 600 lux, and the setpoint was altered
every seven minutes in the sequence of 400 lux 500 lux 900 lux 600 lux,
which required both increases and decreases in illuminance. The setpoints were
specified by sending management commands to the actuation module from the
computer in real time.
The fused values and the raw sensor readings from one of the mote sensors
collected at the computer are shown with a blue solid line and magenta dashed line
respectively in Figure 7-13. Again, only one of the six sets of sensor readings is plotted
for clarity and readability. The illuminance on the worksurface was kept to within 1%
of the setpoint 95% of the time, with a maximum difference of 3%. The maximum
overshoot in transient states was 11%. The overshoot could easily be reduced with a
more sophisticated control law implementing smoother dimming or lightening, but was
not the focus of this test. At the end of the experiment, two mote sensors were
deliberately shaded in order to simulate a situation where the desktop-mounted sensors
were obscured by the occupants. The magenta dashed line in Figure 7-13 is the data set
of one of the disturbed sensors, where the huge deviation from the blue solid line
indicates the period the sensor was shaded. As reflected by the fused values in Figure
7-13, the worksurface illuminance was maintained at the setpoint even though two of
the six sensors in the network were collecting erroneous data. This demonstrates that the
mote-FVF algorithm ensures the robustness of the intended system under disturbances.
113

Figure 7-13 Desktop illuminance setpoint tracking.
114

Chapter 8
System Verification on Human Subjects
This chapter describes a set of tests designed for evaluating and comparing
various aspects of the research lighting system and a representative commercial lighting
system via feedback from human subjects. Both lighting systems are implemented in
the same small private office, but can operate independently. Test results and key
findings are incorporated into the development iterations of the research system.
The human subject tests presented in this chapter are the contributions to a
broader-scope study conducted in collaboration with the research team. The rest part of
this study is discussed in [66].
8.1 Overview
A prototyping research lighting system realized in a small private office was
evaluated and compared to a representative commercial lighting control product through
human subject testing. The implemented closed-loop lighting regulation system
comprised desktop-mounted wireless photosensors, wireless ballast actuation modules,
the mote-FVF sensor validation and fusion algorithm, a graphical user interface, and a
set of rule-based control laws. The chosen commercial lighting system was the Watt
Stopper LS-301, which consisted of a photosensor that responds to the daylight and
dims the electric lights, and a hand-held remote controller that can override the
automatically dimmed electric lights. Two aspects of the systems were examined: the
impact of sensor node locations on sensing accuracy and the overall user satisfaction.
115

Every subject was asked to participate in two consecutive sessions and worked under
both the research lighting system and the commercial lighting system.
In the first test, participants were permitted to place the wireless mote
photosensors at any location close to the desk, and the pertinence of the fused values
was compared to the readings recorded from a high fidelity light meter placed directly
at the center of subjects’ working area. Recommendations based on the results of this
test were made for future improvements and practical implementations.
The second test assessed user satisfaction with the overriding mechanisms
provided by both systems as well as the performances of the systems. The participants
were asked to set the lights to an ideal level with each of the overriding mechanisms
from the two systems. In the meantime, the fact that the lights were controlled by
different lighting systems in each consecutive session was concealed from the subjects,
and the participants were later asked for observations that distinguished the two tested
systems. The results have shown that the overriding system provided by the research
system was generally preferred, which implicated good design practice for lighting
overriding mechanisms. Furthermore, participants could not differentiate the
performance of the research system from that of the commercial system, which implied
that the research system could be superior to the commercial products when integrated
with intelligence.
The first test, the sensor placement test, was actually sandwiched between the
second test of comparing the research and commercial system to ensure the smoothness
of the test session and the best use of the participants’ time. In hopes of mitigating
possible bias, the order of each subject exposed to the research system and the
116

commercial system was randomized for the second test. If the research system was
presented to the participant first, the test session would start with the first test by letting
the participant choose the sensor locations, and then the second test on the research
system followed without interruption. During the intermission after the test on the
research system and before the test on the commercial system, the investigator would
finish the data gathering part of the sensor placement test while the participant took a
10-minute break outside the office. Likewise, if the commercial system was tested first,
the sensor placement test would be executed after the 10-minute intermission. When the
participant had determined the locations of the sensors, the test on the research system
began right away. The investigator only took sensor and meter measurements to
complete the sensor placement test after the subject had done the test and left. For the
sake of clarity, the two tests are divided and discussed in separate sections in Chapter
8.3 and 8.4. The complete protocol narrative and a list of all the materials used in this
human subject test are attached in Appendix B.
8.2 System Setup
Figure 8-1 shows the physical layout of the office space implemented with the
lighting systems. The office is approximately 10-by-14, with a 142 ceiling. It
contains various sizes of bookshelves, file cabinets and file holders. Two desks are
placed against the walls so that the office fits two occupants sitting back to back. There
is a single 6-by-3 window covered with a blind, as indicated on the right-most wall
in Figure 8-1. The walls are painted ivory with decorations of picture frames. The
ceiling is also ivory and the floor is covered with beige tiles, which are a bit darker in
117

color than the walls and ceiling. The four fixtures are mounted 1011 from the floor
(86 from the desk surface). The testing desk is 25 in height with a light gray surface.
Figure 8-1 Private office layout.
Since this implementation was meant for comparing the performance between
the research system and a commercial product, the two lighting control systems must
coexist, operate independently, and be able to be switched between each other. This
requirement introduced significant complexity when retrofitting this office with the two
lighting systems. There are two rows of luminaires, and each row comprises two
connected troffers as depicted in Figure 8-1. Each troffer was retrofitted with a 4-lamp
dimmable ballast and four T-8 lamps, and the four troffers were mounted to the six
available ceiling suspension rods.
118

8.2.1 Hardware Implementation in Small Private Office
For the research system, four second-generation mote-based wireless ballast
actuation modules were installed in each of the luminaires interfacing the dimmable
ballast and the mains as illustrated in Figure 8-2. Three mote photosensors were
deployed onto the desktop. A base station composed of a base mote, a graphical user
interface (GUI) and a Matlab ® program was integrated with a regular desktop PC.
Figure 8-3 demonstrates the deployment of all the components on the desktop to the
right. The wireless sensors and actuators were configured to a single-hop, centralized
broadcast-based network with the base station as the processing center. The base mote
forwarded sensor readings from wireless photosensors to the Matlab ® program and
disseminated actuation commands from the Matlab ® program to actuation modules.
When the system was in operation, the lights were regulated at the level
specified by the users via the GUI. The mote sensors periodically acquired the
illuminance readings and sent them to the base station. The base station fused the
readings with the developed sensor validation and fusion algorithm, calculated the
control action based on the fused sensor information according to the control laws, and
transmitted the actuation commands to actuation modules. The four actuation modules
dimmed or brightened the lights identically upon receiving the actuation commands.
119

Figure 8-2 Installation of the wireless ballast actuation module.
Figure 8-3 Deployment of system components.
For the commercial system, a photosensor was hung from the ceiling at about
the same height as the troffers and was hard-wired to all four dimmable ballasts. The
photosensor had to be calibrated before being fully functional. The calibration process
involved setting two setpoints using the handheld remote controller, one during the
daytime with daylight and the other during the nighttime without daylight, so that the
illuminances at the reference point were the same at both times. In this case, the
120

eference point on the desktop was set to 500 lux, a commonly acknowledged standard
for lighting design.
When the Watt Stopper system was in operation, the photosensor dimmed or
brightened the electric lights in response to the available daylight so as to maintain the
illuminance on the desktop at 500 lux. The user might temporarily override the setpoint
from 25% above the original setpoint (500 lux) to the minimum (all the lights were
dimmed to the lowest level) with the handheld remote controller. The original setpoint
would resume when the Auto button on the remote controller was pressed.
The wall unit was modified to a single overriding switch and eight sub-switches
for alternating between the two systems. The overriding switch controlled the power to
all the components in the ceiling. Four of the sub-switches determined whether the AC
power went to the commercial system or the four mote-based actuation modules. The
other four switches directed the DC control signals to the ballast from either the
commercial photosensor or the four mote-based actuation modules.
8.2.2 Lighting Controller and Overriding Mechanism
The research system provided a graphical user interface, shown in Figure 8-4
(a), for the participants to set and adjust their ideal task lighting. The research system
maintained the desktop illuminance at the setpoint specified using this interface. There
were two ways to specify the setpoint: the stepwise and the direct-value control. Using
the buttons in the right column of the interface, the users could increase/decrease the
light step by step with the Step Up and Step Down buttons, or brighten/dim the light to
the maximum/minimum with the Highest Level and Lowest Level buttons. As an
121

alternative, the setpoint could also be specified by dragging the sliding bar or by
directly entering a value between 1 and 100 in the Target Value field, and then hitting
the SET button. An actuation command would be generated and sent to the actuation
modules as soon as a button was clicked. Although the research system allowed turning
the lights on/off wirelessly, the participants were not granted this privilege in the test.
The step size for the stepwise control on the interface was also predefined and could not
be changed by the subjects.
Another computer interfaced with a base mote was set up at the Investigator
Station in Figure 8-1 so that the investigator could intervene in the lighting control,
which was necessary for the testing procedure. The investigator could also conveniently
initialize the lighting system for the human subject test using this secondary computer.
A slightly more versatile control interface was provided on the computer as shown in
Figure 8-4 (b). Besides the functionality of the GUI for the participants, this interface
displayed the current actuation level, and allowed the investigators to specify the step
size in the Step Size field for stepwise control and turn the lights off with the OFF
button.
The setpoint of the commercial system was calibrated at 500 lux on the testing
desk. The participants overrode the predetermined setpoint using a handheld remote
controller as shown in Figure 8-5. Only stepwise control was provided with the
commercial system, and the Auto button on the remote controller, which was not
required in the test, was disabled as it resets the setpoint to the predefined one (500 lux).
Since the handheld remote controller used infrared communication technologies, it had
to point at the commercial photosensor mounted in the ceiling to ensure the overriding
122

commands were received. In addition, a secondary handheld remote controller
controlling the same photosensor was placed at the investigator’s station so that the
investigator could disturb the lighting as required in the testing procedure.
(a)
(b)
Figure 8-4 GUI provided by the research system.
Figure 8-5 Handheld remote controller provided by the commercial system.
123

8.3 Sensor Placement Test
8.3.1 Objective
The objective of this test was to (1) identify common patterns for sensor
locations chosen by the occupants; (2) evaluate the impact of customized sensor
locations on fused illuminance values.
Desktop-mounted photosensors facilitate better estimations of human perception
of light; however, it is impossible to put the sensors at the very spot where an occupant
conducts work without intervening in the work. As a result, the sensors have to be
deployed in proximity to an occupant’s working area and measure the illuminances
surrounding the worksurface. Although overhead luminaires generally deliver evenly
distributed lighting, slight illuminance variations at every point on a desktop are
inevitable. Moreover, in practice the sensors should eventually be placed by the users
rather than in carefully assigned locations. Therefore, it is critical to study the effect of
sensor locations on estimations of target illuminances, the robustness of the sensor
validation and fusion algorithm, and the feasibility of the intended sensor deployment.
8.3.2 Testing Procedure
The testing procedure was divided into two parts: sensor placement, and data
collection and documentation. The sensor placement was performed by the participants.
Each subject was given a brief explanation regarding the phenomenon – illuminance –
that was to be sensed and how the sensor worked with the lighting system. Participants
124

were provided with three wireless photosensors and permitted to place them until
comfortable with the configuration of the worksurface.
The data collection and documentation was accomplished by the investigator
without the presence of the participants. The finalized sensor locations chosen by each
subject were photographed. A high-fidelity light meter was placed at the spot where the
participants conducted their work as a reference. Sensor readings were transmitted
wirelessly to the central computer station where they were then stored, and meter
readings were logged manually by the investigator. The data from sensors and the meter
were recorded every 10 seconds for 5 minutes.
This experiment was designed to last approximately 25 minutes: 5 minutes for
instructions and questions, and 20 minutes for placing sensors and data collection.
8.3.3 Results and Discussion
Ten subjects were recruited to participate in this experiment. The participant
group was comprised of roughly the same number of males and females who came from
different ethnic backgrounds. All had experience working in an office.
Impact of sensor locations on fused values
The high fidelity light meter was placed at the spot on the desktop where the
subject did the most reading, writing and typing, and hence should best represent the
lighting the participant had perceived. The statistics of the fused readings in the ten
cases is tabulated in Table 8-1.
125

Subject
No.
Table 8-1 Summary of the sensor placement experiment.
Average meter
reading
Average fused
reading
Deviation of fused reading from
meter reading
Mean
Standard
deviation
1 500.83 499.41 1.42 1.49
2 518.23 519.77 1.55 1.77
3 576.36 488.21 88.16 9.65
4 366.93 318.66 48.28 6.90
5 557.11 493.72 63.39 4.73
6 602.10 475.42 126.68 3.62
7 582.33 433.02 149.32 2.68
8 568.07 476.26 91.81 2.18
9 511.94 554.84 42.90 0.94
10 1026.09 1151.94 125.85 2.35
In two out of the ten cases, subjects No. 1 and 2, the fused values closely
matched the meter readings with less than 1.5% deviation. Figure 8-6 (a) plots the
readings of one such case along with a picture of the sensor locations in Figure 8-6 (b),
where the circled numbers indicate the identification number of the mote photosensors.
In both cases, one of the three sensor measurements was very close to the meter
readings, another one was constantly higher than the meter readings, and the other one
was constantly lower. Nonetheless, all the three sensor measurements were within 4.5%
of the meter readings on average. Although the sensors were not placed in exact same
pattern in both cases, they were spread out to surround the occupants’ working area but
still close to the center of the working area in general. These two were the ideal cases as
the locations of the sensors rendered the fused values to effectively represent the real
lighting conditions perceived by the participants. Also, all the sensor readings were in a
pertinent range of the real task illuminance except those that were perturbed or lost.
126

(a)
Figure 8-6 Sensor placement of subject No. 2.
(b)
In six out of the ten cases, the fused values were constantly lower than the meter
readings by 11.37-25.64%. In two of these cases, subjects No. 3 and 4, one sensor
measurement was close to the meter measurement and the other two were constantly
lower. Figure 8-7 (a) illustrates the readings from one such case with the corresponding
sensor locations in Figure 8-7 (b). Even though the measurement of one sensor was
close to the meter readings, they were weighted less by the fusion algorithm as the
readings were significantly higher than the other two. As a result, the fused values were
lower than the real illuminances. In the other four cases of subjects No. 5-8, all three
sensor measurements were constantly lower than the meter readings. Figure 8-8
demonstrates one of the cases along with the sensor locations. These results show that
the sensors are likely to be placed at spots where the illuminances are lower than the
work area of the occupants. Those places with lower illuminance may be under subtle
shadows of other objects on the desktop, or simply farther away from the overhead
luminaire since the illuminance distribution is not perfectly even as Figure 6-2 implies.
127

(a)
Figure 8-7 Sensor placement of subject No. 3.
(b)
(a)
Figure 8-8 Sensor placement of subject No. 7.
(b)
In the remaining two of the ten cases, the fused values were higher than the
meter readings. These results were not expected since it was intuitive to assume that the
lighting on the subjects’ working areas is always unblocked, and thus the sensors are
much more likely to be placed at darker spots than brighter spots. For the sensor
locations chosen by subject No. 9 as shown in Figure 8-9, the fused values were 8.38%
above the meter readings on average. Two of the sensors measured the illuminance
higher than the meter while one gave lower readings. One possible explanation of the
128

lower meter readings was that the meter head was not properly placed and was partially
shaded by some subtle shadows of the objects on the desktop since it was more
sensitive and had a larger range of view.
(a)
(b)
Figure 8-9 Sensor placement of subject No. 9.
Figure 8-10 shows the other case of subject No. 10 where the fused values were
higher than the meter readings by 12.27%. The measurement of one of the sensors was
significantly higher than others since the participant chose to place it on top of the
computer case as the circle marked 04 shows in Figure 8-10 (b). Notice that the subjects
were free to choose the sensor locations, not limited only to the desktop, as long as they
felt the locations were appropriate. However, the sensor data in this case has shown that
illuminance at a place elevating from the desktop is higher and may not be an ideal
location for sensors without a proper pre-calibrating procedure.
129

(a)
(b)
Figure 8-10 Sensor placement of subject No. 10.
Overall, the fused values matched the meter readings, and hence the lighting
perceived by the subjects, in only 20% of the cases. In 60% of the cases, the fused
values were constantly lower than the meter readings, and the fused readings were
constantly higher than the meter readings in the other 20% of the cases. The plots of
sensor readings and the pictures of sensor locations for all of the participants are listed
in Appendix B.1. Although in most cases it is not obvious to the participants and the
investigator why each individual sensor measurement deviated from the meter
measurement, several reasonable conjectures can be made as follows:
(1) The light on the desktop might not be uniformly distributed. This could be
caused by the design of the luminaires and their layout in a room. In the testing
office, the desk was located at the place where approximately half of its surface
was directly beneath one of the four troffers while the rest of the surface was not.
As the workplane level illuminance model in Figure 6-2 suggests, the linear
combination of four such models is not likely to produce evenly distributed
lighting on the desktop.
130

(2) Subtle and invisible shadows may exist on the testing desktop. The shadows
could be introduced by objects on the desktop such as the computer chassis,
monitor, decorations and laptop (if the subjects brought their own laptop to
work).
(3) Reflection of light from the walls or glossy surfaces of objects on the desktop
may be picked up by the nearby sensors. In the testing office, one side of the
desktop was attached to the wall where there was a poster with a glossy surface
and a picture frame with a glass cover. Both of them could reflect light and cause
the variations of sensor readings.
Although the above three factors could very well be prevented by carefully
designing the testing environment, no ordinary office can avoid any of them in reality.
Therefore, the testing results still provide valuable information for further improvement
to accommodate the impact of fused values deviating from the perceived illuminance
due to customized sensor locations.
Allowing the subjects to select the photosensor locations may compromise the
accuracy of the fused values estimating the lighting perceived by the occupants. The
fusion algorithm extracts a pertinent representation of the illuminance sensed by each
individual sensor; in other words, there is no way for the algorithm itself to figure out
the illuminance of the target surface if all the sensor readings deviate from that in the
first place. The most important implication of the testing results is that instead of the
absolute illuminance, it would be more efficient and realistic to identify occupants’
preferences and control the lights using the fused sensor readings when designing the
131

esearch lighting system. After all, it is the ability of the lighting system to deliver and
maintain the preferred lighting that concerns the occupants, not the actual numbers.
Common configuration of sensor locations
There was no single conclusive common configuration of sensor placement
found in this test; however, several interesting and similar patterns were observed. All
of the ten subjects spread the sensors on the desktop. The subjects were never advised
not to put sensors close to each other by the investigator, but all of them tried to put
each sensor at a different spot on the desktop. One reasonable guess is that participants
believed it would minimize the chance of multiple sensors being simultaneously
disturbed by placing the sensors apart. In addition to spreading them out, eight of the
ten subjects configured the sensor locations so that the sensors sort of surrounded the
working area. The subjects might have done this in hopes of getting representative
sensor readings of the light they perceived.
Eight subjects had at least one sensor sitting at the corner or edge of the desktop.
Overall, a total of five sensors were put near the left edge of the desk, three sensors
were put at the right corner, two sensors were put at the left corner, and two sensors
were pushed deep near the wall. One extreme case was that a sensor was put on the top
of the computer chassis. This showed that although the subjects were eager to get
representative measurements from the sensors, given the sensors’ current size, the
subjects also tended to get them out of the way while they were working.
Four out of the six subjects who worked on the provided desktop computer put
at least one sensor on or very close to the keyboard, especially on the side where the
132

function keys are. The function key side of a keyboard is indeed the least likely to be
accidentally shaded or disturbed. This strongly suggests that when the mote sensor
technology is mature enough for embedded sensing, the function-key side of the
keyboard is a good place to integrate with the sensors.
Two opinions common to almost all participants can be concluded from the
questionnaires. First, while one participant kept a neutral opinion, the other nine
subjects would like to be able to assign the locations for sensors if the sensors were to
be mounted in close proximity to their working area. Second, the sensors, at their
current size, were still a bit too big to be directly mounted on the desktop without
distracting the participants during the tests. A few subjects felt strongly against having
the additional sensors on the desktop, which supported the idea of integrating the
sensors into items already on the desktop, such as the keyboard, computer monitor,
desktop decorations, etc. when the technologies mature.
8.4 Comparison of System Performance and User Interface
8.4.1 Objective
This test was intended to (1) determine occupants’ preferences of interacting
with the research and the commercial lighting systems via manual override; (2) obtain
participants’ opinions on the performance of both systems. Several studies have pointed
out the importance of user control in automated office conditioning environments [103,
104]. The handheld remote control provided by the commercial lighting system was
compared to the GUI used to control the research mote-based system. In addition, both
133

the research and the commercial systems in this experiment automatically maintained
the desktop illuminance at the participants’ preferences, and the participants were asked
to comment on whether or not they noticed the difference between the two systems.
8.4.2 Testing Procedure
Subjects were offered a brief explanation of how the automatic lighting system
works in general and how it can be overridden with both methods of control. The
participants were only told that two types of overriding mechanisms were to be tested
without the knowledge that the lighting systems were actually different. Before the test
began, participants were able to play with the controllers to gain familiarity with their
operation. Each participant received a questionnaire as attached in Appendix B.7, and
was invited to answer or disregard any of the questions and write down any comments
or complaints that come to mind during the test. In order to simulate a real office
working condition, the participants were encouraged to bring their own work to the test,
preferably a mixture of paper- and computer-based tasks. An artificial task would be
provided should the participant fail to find a suitable task to bring in. Subjects might
work with either their own laptop or the desktop PC on the experiment worksurface that
run the research system; however, they could only override the research system with the
GUI provided by the desktop PC no matter which computer they chose to work on their
own tasks.
In order to avoid any possible bias caused by the order of the systems to which
the subjects were exposed, half of the participants started with the research system
while the other half started with the commercial system. Furthermore, the investigator
134

deliberately perturbed the lights during the test to encourage more frequent use of the
overriding mechanisms for pertinent comments from the participants. The test
procedure for the subjects who started with the research system is listed as follows and
that for participants who started with the commercial system is very similar and can be
found in Appendix B.2.
(1) Participants start with the initial desktop illuminance of 500 lux under the
research system, use the GUI controller to set the lighting to a comfortable level,
and begin to work on the task of their choice.
(2) At the 5 th , 10 th and 15 th minutes, the lights are dimmed or brightened by the
investigator. This is to encourage the participants to override to return the lights
to a comfortable level.
(3) At the 20 th minute the investigator stops the first session of the experiment and
asks the participants to take a 10-minute break outside the office. Meanwhile, the
investigator resumes the data collection and documentation part of the sensor
placement experiment as described in the previous section. Before the
participants enter the office again for the second 20-minute experiment, the
investigator switches the lights to the commercial photosensor-dimming system
and sets the initial desktop illuminance to 500 lux. Participants are asked to use
the remote controller to override the initial lighting to a comfortable level after
settling for the remaining experiment.
(4) At the 5 th , 10 th and 15 th minutes of the second session, the lights are again
dimmed or brightened by the investigator using a second remote controller
without notifying the participants.
135

(5) At the 20 th minute, the test ends. The participants are debriefed and interviewed,
and their questionnaires are collected.
The post-test interview is intended to gather feedback beyond the questionnaire
and to identify reasons for any unanticipated behaviors. The interview basically
followed the script provided in Appendix B.8. Some of the questions were very similar
to the questions in the in-test questionnaire to encourage criticism of either system and
to ensure that important research questions were asked. This experiment lasted
approximately 60 minutes: 5 minutes for instructions and questions, 20 minutes for each
working set, 5 minutes for debriefing and interviewing, and 10 minutes for the
intermission in between the two sessions.
8.4.3 Results and Discussion
Preferred manual overriding mechanism
Three aspects of the two manual overriding mechanisms were evaluated through
the questionnaire: (1) easiness of learning the mechanisms, (2) easiness of overriding
the lights with the mechanisms, and (3) preference for the mechanisms. The
participants’ opinions were quantified using a five-level scale in the questionnaire:
totally agree, partially agree, neither agree nor disagree, partially disagree, and totally
disagree. The results are summarized in Table 8-2, where the options of totally agree
and partially agree are assigned in the same category, and so are those of totally
disagree and partially disagree. The detailed answers to each question can be found in
Appendix B.3.
136

Table 8-2 Summary of the test on preferred overriding mechanism.
Easy to learn
Easy to use
Handheld remote
controller
Graphical user
interface (GUI)
Agree 10 10
Neutral 0 0
Disagree 0 0
Agree 8 10
Neutral 0 0
Disagree 2 0
Preferred mechanism 2 7
From Table 8-2, it can easily be concluded that both overriding mechanisms
were reasonably easy to learn and use for setting the participants’ preferred lighting.
The GUI provided by the research system was generally considered to be more versatile
than the handheld remote controller due to the various overriding options. The handheld
remote controller was designed to increase or decrease the light level of the Watt
Stopper system in a continuous manner following a gentle curve. As a result, users had
to aim at the photosensor and keep on pressing the button on the controller until the
light reached a satisfying level. Comments on the questionnaires indicated that the slow
response led two subjects to disagree on the easiness of use of the handheld remote
controller.
Seven participants preferred the GUI provided by the research system over the
handheld remote controller while two subjects had the opposite opinions. One
participant didn’t express a specific preference on either type of the overriding
mechanisms. People preferred the GUI mostly because they could easily control the
lights while working on the computer without being distracted by having to locate a
physical device for this purpose. On the contrary, people who rated the handheld remote
137

controller higher liked the fact that they could conveniently control the lights whenever
they felt like it instead of having to go to the virtual controller even when not working
on the computer. These two conflicting opinions suggested that it might be a good
practice to have both virtual and physical overriding mechanisms when designing a
lighting system.
As mentioned in the previous section, there are three different ways to override
the research lighting system with the GUI: click on the stepwise control buttons, drag
the sliding bars and press the SET button, or type in a desired level and click on the SET
button. The step buttons, like those on the handheld remote controller of the commercial
system, allow the participants to adjust the lights gradually while the other two options
enable the subjects to quickly jump to a specific light level. An interesting observation
was that even though these three control options were equally and conveniently
available on the GUI, all the participants would settle for only one of them during the
entire test instead of a combination of all three. Nevertheless, the fact that the
participants didn’t choose the same control option on the GUI to set the light level
implied that it might be beneficial to offer more than one interaction method on
overriding interfaces.
The most common suggestion made by the subjects for improving user
satisfaction was to include certain feedback on the overriding mechanism, such as the
current light level. This confirmed the well-known interface design guidelines that
highlight the importance of providing user feedback [105-107]. The feature of status
feedback was originally available on the GUI of the research system, such as the
138

Current Level on the investigator’s GUI in Figure 8-4 (b), but was left out on purpose to
avoid possible biases when evaluating the performances of the systems.
System performance comparison
The questions asked in the questionnaire regarding the performance of both
systems were: (1) whether the participants felt comfortable working under the lighting
systems, (2) whether the systems successfully maintained the lighting chosen by the
participants, and (3) if the two systems performed differently on the subjects. As
pointed out previously, other than the types of overriding mechanisms, the subjects
were not told how many different lighting systems they were exposed to or what
systems the overriding mechanisms controlled. The first question was quantified using a
five-level scale: totally agree, partially agree, neither agree nor disagree, partially
disagree, and totally disagree. The second question was answered using a 1-7 scale
where 1 means poor performance and 7 means excellent performance. Table 8-3
summarizes the answers to the first two questions gathered from the questionnaires.
Answers of totally agree and partially agree to the first question are assigned to the
same category in Table 8-3, as are those of totally disagree and partially disagree. For
the second question, answers below 3 and above 5 are considered bad performance and
good performance respectively in Table 8-3. Detailed responses from the subjects are
tabulated in Appendix B.3.
According to Table 8-3, it is fair to say that both systems delivered a
comfortable lighting environment and managed to maintain the light at the levels of the
subjects’ choices. One subject did not rate the ability of the commercial system to
139

maintain the specified lighting. Although the mote-based research system led the
commercial system by one point on both questions, it is not significant enough to
conclude that the research system is superior to the commercial system.
Table 8-3 Summary of the test on system performance.
Feel comfortable working
under the system
Performance of
maintaining preferred
lighting
Watt Stopper
commercial system
Mote-based
research
system
Agree 9 10
Neutral 0 0
Disagree 1 0
Good 9 10
Neutral 0 0
Bad 0 0
Six out of the ten subjects said the two systems performed differently despite the
type of the overriding mechanisms in response to the third question. However, all of the
remarks made by the six participants regarded the responses of the systems to the
overriding mechanisms, namely the light changed a lot more slowly on the commercial
system as mentioned before. Therefore, it was not the ability of automatically regulating
the light at a specified level that distinguished the two systems.
The most significant finding revealed from this part of the test is that the motebased
lighting system can perform as least as well as the commercial lighting system at
regulating the light at the specified level automatically. In addition, a more important
implication is that the research system can be superior when implemented in a sharedspace
office if integrated with the mechanism that delivers the light satisfying all the
occupants’ preferences as will be discussed in Chapter 9.
140

Chapter 9
System Implementation and Verification in a
Shared-Space Office
The integrated research system was implemented in a small shared-space office
to verify its ability to generate energy savings and to create a satisfying lighting
condition. The lighting system was realized with a true self-configuring multi-hop
wireless sensor and actuator network. The first phase of this implementation studied the
potential energy savings simply by allowing the occupants to specify and work under
their preferred lighting. The second phase of this implementation verified the system’s
ability to automatically minimize energy usage and satisfy occupants’ lighting
preferences while harvesting daylight, and it also demonstrated the additional energy
savings.
9.1 Overview
The lighting system in a shared-space office was retrofitted with a selfconfiguring
and multi-hop wireless ballast actuator network to illustrate possible energy
savings and improved personal lighting satisfactions by enabling individual addressing
and control of each luminaire. In the first phase, the performance of this individually
addressable lighting system was evaluated by allowing the occupants to manually
specify their preferred lighting through a web interface. A 50% potential energy savings
was demonstrated simply by enabling individual control of each luminaire before any
automatic lighting control strategy. Sensors were integrated into the system in the
second phase along with the optimal lighting actuation algorithm developed in Chapter
141

6 to form an intelligent wireless sensor and actuator network. The preferred lighting
specified by the occupants in the first phase was automatically delivered and maintained
by the system based on occupancy status. The performance of the system was assessed
by how closely the automatically generated lighting matched the preferred lighting
specified by each of the occupants.
9.2 Integrated Intelligent Lighting System
A shared-space office in an educational research building over twenty years old
was retrofitted with the research system. The office contains a hallway, a group meeting
area and a work area with twelve personal workstations. Figure 9-1 shows the floor plan
of the office, where the equally spaced shaded rectangles represent the locations of
overhead luminaires. The 817 square foot office space was lit by these nineteen 2-lamp
fluorescent light troffers originally configured to be controlled by a single switch. The
office was located in the interior of the floor, and hence had no windows.
The occupants were researchers and graduate students that have been working in
the office on a daily basis. Each of the occupants had a designated desk in the work area
(the left part of Figure 9-1), and there was also a larger worksurface in the center of the
work area shared by all the occupants for conducting studies and experiments. The
group meeting area (the right portion of Figure 9-1) was not utilized as frequently as the
work area, but was in use quite often as each occupant was free to schedule meetings or
hold discussions in the area.
142

Figure 9-1 Floor plan of the small shared-space office.
The rest of this section describes the hardware setup and network design of this
implementation. The hardware setup subsection depicts the construction of the lighting
control system, including how the components developed in Chapter 7 fit into the
system, how each component interact with one another, and how the office was
retrofitted. The network design subsection specifies the wireless messages used for
configuring a communication network, managing actuation activities, and collecting
sensor readings. For completeness, each of the topics will be discussed in detail.
However, it is important to recognize that each hardware component and network
message may potentially be realized in various ways, and it is the functionalities of the
components and messages that matter in this research.
143

Hardware setup
Each troffer was retrofitted with a 2-lamp dimmable ballast, a third generation
wireless ballast actuation module and two T-8 fluorescent lamps as shown in Figure
9-2. A local server was built for coordinating the wireless ballast actuator network,
controlling and monitoring the lights, and bridging the occupants and the system with a
web interface. Figure 9-3 illustrates the operation of this implementation. The
information flow is bi-directional between the client web browsers and the local server
and between the local server and the luminaires. The route from client web browser
through the local server to the luminaires is for actuation activities, while that from the
luminaires via the local server to web browsers is for status feedback.
Figure 9-2 Wireless-enabled dimming luminaire.
144

Figure 9-3 System operation diagram.
The mote platform sitting on the actuation module is programmed to listen to the
network for actuation commands, report back its current actuation status to the local
server, and coordinate with other motes to form a multi-hop network. Upon receiving an
actuation command addressed to it, the actuation module interprets the message, and
sets the level of the dimmable ballast or toggles the lights on/off. The mote also
periodically generates packets containing its current status and sends them back to the
base station on the local server. The purpose for the status feedback is twofold: to
monitor the actuation status, and hence energy usage of the entire system, and to
reinforce the wireless network links and compensate for lost or corrupted actuation
packets during wireless transmission. Furthermore, the motes autonomously configure
themselves into a network by identifying their parent mote en route to the base station
as soon as they are powered up. The network is dynamic and periodically updated to
avoid bad or interrupted communication links.
The local server integrates four subcomponents: a database, a web server, a
control application program, and a mote base station. Each of the elements was
145

implemented with free open source software in view of zero-overhead, compatibility,
and scalability. The database is realized with MySQL ® . Tables in the database store
user accounts, users’ lighting presets, actuation and status feedback histories, and sensor
readings. The web server adopts the Apache HTTP Server and serves as a bridge
between the users and the control application program. The graphical user interface,
created by Bonnell, is in the form of a website written in PHP [108] through which
occupants can specify their preferred light settings, resume one of their presets, or
override the current setting. The mote base station is composed of a mote plugged into
the USB port of the computer and listens to the wireless network. A generic
‘SerialForwarder’ program commonly used in the TinyOS community bridges the mote
with the control application program. The control application program written in Java ®
is the core of the system in which the control logic resides. It processes the actuation
requests and issues actuation commands. Each time an actuation command is generated,
a corresponding entry is logged into the associated table in the database. In order to
account for possible interruption or corruption of the wireless communication, the
program also listens to the status feedback from the actuation motes, compares the
status to the most recent actuation history, and resends the actuation packet if any
inconsistency is detected.
The implementation was designed with the goal of minimizing the amount of
costly retrofitting, while still achieving full functionality. Retrofitting simply included
replacing the original ballasts with dimmable ballasts, installing the mote-based ballast
actuation modules in the troffers, and replacing the original T12 lamps with T8
fluorescent tubes. Two professional electricians, who belong to the staff in charge of
146

campus-wide lighting maintenance, were recruited for the retrofitting task. Both
electricians had worked with dimmable ballasts before, but neither of them had prior
experience with this particular mote-based actuation module. It took each of the
electricians roughly one hour to complete retrofitting the first luminaire, but the
retrofitting time reduced to about twenty minutes per troffer for each electrician after
they had gotten up to speed and had developed their own method for installing the
modules. The wires used on the actuation modules were color coded to match the color
of the wires used for typical dimmable ballasts and power lines to avoid confusion and
enhance the ease of installation. Figure 9-4 is a snapshot of the retrofitted office with
the luminaires actuated at different levels.
Figure 9-4 Office lighting after retrofitted with the research system.
147

Network design
In addition to enabling the wireless capability of each luminaire, the
implemented system also realized a self-configuring multi-hop wireless actuator
network. Four wireless message packet types were implemented for this purpose:
actuation message, status message, grouping message, and beacon message.
Specifically, the actuation message and the grouping message are packet types that
support lighting actuation activities, and status message and beacon message are packet
types for actuation status reporting from each luminaire.
Two additional packet types, sensor message and sensor management message,
were defined for wireless photosensors in the second phase of the implementation when
sensors were integrated into the research system. The structure of each message type
was designed to have certain amount of flexibility for future expansion, and thus not
every single field in each message type was utilized in this implementation.
Actuation activity
The actuation message type has the structure shown in Figure 9-5. The source
ID field contains the unique identification number of the mote that sends out the
message, which can be the mote that either generates the message or relays the message.
The origin ID field carries the identification number of the mote that originates the
message, which is always the ID of the base station mote since the base server is the
only entity that issues actuation commands in this implementation. The destination ID
field specifies the mote to which this message is addressed. If it is a multicasting or
broadcasting message, the destination group ID field indicates the group to which this
148

message is addressed. The destination group ID field for multicasting messages to a
specific group will only be in effect if the destination ID field is set to 65535. Actuation
messages with both destination ID and destination group ID fields set to 65535 are
broadcasting messages to which all ballast actuators should respond. The sequence
number field carries a unique number for the recipient motes to distinguish between a
new and a repetitive message. The time to live (TTL) field specifies how many times a
message can be relayed in order to avoid infinite loops. The number decreases by one
each time a message packet gets relayed by an intermediate mote, and the mote stops
relaying the message when TTL reaches zero. As mentioned in chapter 7.3.2, the third
generation ballast actuation module has an embedded failsafe mechanism which
bypasses the mote actuator by default. In order for the mote-based ballast actuation
module to be in effect, it has to be activated with an actuation message with the Boolean
activate field set to true. The power off field is also a Boolean field for toggling the
lights on and off. The dimming level field specifies the actuation level of the destination
luminaire(s). The step control field allows stepwise dimming/brightening control with a
predefined step-size and is currently not implemented.
149

Figure 9-5 Actuation message structure.
The grouping message type has the structure shown in Figure 9-6, and is for
assigning the actuation mote to the specified group. The group ID field specifies the
unique group identification number to which the destination motes are to be assigned.
The time to live and sequence number field has the same functionality as those in the
actuation messages. The group member list field carries the unique identification
numbers of the mote actuators that are to be assigned to the group indicated in the group
ID field. The grouping feature has been implemented and available in the system, but
was not in use for the verification in the following sections.
Figure 9-6 Grouping message structure.
150

Both the actuation and the grouping messages are transmitted from the base
station to the destination mote(s) using a flooding-based network protocol. When an
actuation or grouping message is ready to transmit, the base station broadcasts the
message. Each of the motes that hears the message will rebroadcast the same message
unless it is repetitive. Meanwhile, each mote will respond accordingly if the message is
addressed to it. The chosen network protocol supports unicast, multicast and broadcast
so that multiple motes may react to the same message.
Status reporting activity
The status report message type has the structure shown in Figure 9-7. The
functionalities of the source ID, origin ID, sequence number and time to live fields in
the message are exactly the same as those in the actuation message type. The activate,
power off, dimming level and step control fields carry the corresponding status of the
mote actuator.
As mentioned before, the purpose of the status feedback from each actuation
mote is twofold: (1) to reinforce the communication links through a tree-based per-hop
acknowledgement network protocol, and (2) to monitor the true actuation status of each
mote actuator. Although the flooding-based actuation message communication protocol
seems to be robust, there is still the chance that the destination mote will miss the
message since no acknowledgement mechanism was implemented in the protocol. With
the status report from the mote actuators, the base server can check the consistency
between the reported status and the latest actuation history and resend the actuation
message when inconsistency is detected. The status report messages are transmitted
151

using a tree-based mechanism with the base station as the root of the tree. The mote in
the route to the root will relay a received status message to its parent one layer closer to
the root, and wait for the acknowledgement from its parent. Unacknowledged messages
are resent up to three times.
Figure 9-7 Status report message structure.
Network management activity
The beacon message type is used to create the communication tree for
transmitting the status report messages from ballast actuators as well as the sensor
messages from photosensors, and has the structure of Figure 9-8. The source ID field
specifies the unique identification number of the mote that sends out the beacon
message. The parent ID field contains the parent’s identification number of the mote.
The cost field indicates the cost of using the transmission route, which is calculated
from the link quality indicator (LQI) embedded in every received packet and the hop
counts in the hop count field. The hop count field keeps track of how many of the same
beacon message are relayed, and increases by one for each relay.
152

The construction or updating of the communication tree is achieved by each
mote identifying its parent, and it starts with the base station mote broadcasting a
beacon message with source ID set to 0, the designated identification number for the
base mote in this implementation. The cost and hop count fields are also initialized to 0.
Each mote that hears a beacon message will perform the following six tasks: (1)
calculate the cost from the LQI and the hop counts of the received beacon message; (2)
replace the source ID with its own ID; (3) replace the parent ID with the ID of its
current parent if one exists; (4) add the calculated cost to the original one in the cost
field; (5) increment hop count field by one; (6) rebroadcast the message with the
updated fields. Each mote updates its parent to the one with minimum cost every time a
beacon message is received and then rebroadcasts the message with updated fields. To
avoid loops, a mote is not allowed to select the mote as its parent when the parent ID
field in the beacon message is filled with its ID.
Figure 9-8 Beacon message structure.
Sensing activity
The structure of the sensor message type is defined in Figure 9-9. The source ID
field contains the unique identification number of the mote that sends out the message.
The origin ID field contains the identification number of the mote photosensor that
originates the message. The destination ID field specifies the final destination of this
153

sensor message, which is always the ID of the base station mote since the base server is
the sink that processes and stores sensor information in this implementation. The
function of the sequence number and time to live fields are exactly the same as those in
other message types. The Boolean fused field indicates whether this sensor reading is a
raw sensor data or a fused value generated by an intra-network sensor fusion process.
Recall that it is possible for the photosensors to perform intra-network sensor fusion as
discussed in Chapter 7.4. The reading field carries the sensor reading.
Figure 9-9 Sensor message structure.
The sensor management message type supports managing mote photosensors
over the wireless network from the base server. The management mechanism employs a
broadcast-based protocol like the one used in actuation activities, and all sensors should
respond to the same message. The sequence number and time to live fields have the
same functionalities as those in other message types. The Boolean query field indicates
whether this message is a request for sensor readings. It allows the base server to
request sensor readings using this message with the query field set to true. Sensors
receiving this message immediately acquire a reading and send it back to the base server
in addition to its routine fixed-rate sensing tasks. The Boolean update sensing rate field
154

determines if this message is for updating the sensing rate of the sensors. If this field is
set to true, the sensing rate of the recipient sensors will be updated to the new rate
specified in the sensing rate field. The Boolean reset field, if set to true, instructs the
recipient photosensors to revert all parameters to default values.
Figure 9-10 Sensor management message structure.
9.3 Energy Savings Assessment
The first phase of the system implementation was to assess long-term energy
savings by enabling the individual addressability and controllability of each luminaire.
During this phase, the overhead luminaires were manually controlled by the occupants
through a designated user interface. The power consumed by the new lighting system
was measured and compared to that of the original lighting configuration.
9.3.1 User Interface and Power Measurement
A web interface designed by Bonnell [108] bridges the occupants and the
lighting system. The occupants may define presets that can be resumed with the GUI,
shown in Figure 9-11, residing in the Kiosk by the entryway of the office. Each preset is
155

specified by setting the percentage of light output from each of the luminaires that
matters using the GUI in Figure 9-12, where the background is the floor plan of the
office. When a user resumes a preset, the GUI sends the request to the base server of the
lighting system, which then retrieves the light settings corresponding to the preset and
transmits the actuation commands to the involved wireless ballast actuators. Conflicting
lighting preferences are resolved by simple heuristics that apply the highest specified
setting to the luminaires. For example, if multiple present occupants have assigned
different light intensity settings to the same luminaire, the luminaire will be actuated at
the highest level of all the settings.
Figure 9-11 User interface on the Kiosk [108].
156

Figure 9-12 Lighting preset setting interface [108].
In order to rigorously determine the energy consumption and the usage patterns,
a customized power measurement instrument was installed [108]. This device measures
the power consumption of the 19 luminaires in the office along with the instant voltage
and current, RMS (root-mean-square) voltage and current, VA, and power factor for
further analysis.
9.3.2 Long-term Energy Savings
To evaluate the possible energy savings, the power consumption of the office
after the implementation of the mote-based research lighting system was compared to
that of the original lighting configuration where all the lights could only be turned
on/off together. The energy usage data collected by the power measurement device
157

etween February 2 nd , 2008 and July 7 th , 2008 with 3,408 hours of valid data was
considered.
Since the original ballasts and the light tubes were all replaced during the
retrofitting, the energy usage for the original lighting configuration was theoretically
calculated for the comparison. The power consumption for all on/off operation was
determined using the same office light usage data, but assumed that the lamps were only
operated all on (rated as the full energy consumption of the ballast) and off (zero
power). In this sense, the comparison took into account the operation overhead of the
motes and the actuation modules in each of the retrofitted troffers. In addition, the input
power of the ballasts used in this calculation was the maximum power of the new
dimmable ballasts rather than the less energy-efficient non-dimmable ballasts that the
research system replaced. This choice was made to eliminate the difference in
efficiencies between the old and new ballasts.
The overall energy savings for the analyzed period was 51.2%. A total of 638.02
kWh were actually consumed by the research system, and power usage of 1307.46 kWh
was calculated for the all on/off lighting configuration. Figure 9-13 shows the
percentage utilization of the office broken down by the hour, and reveals the occupancy
pattern during typical business hours. For example, the third bar from the left of the
histogram indicates that in all the valid days between the hour of 10am and 11am, 60%
of the time the office was occupied by at least one person. The corresponding hourly
lighting power consumption is plotted in Figure 9-14, where the red-shaded bars and the
green-shaded bars represent the energy usage of the research system and the original
lighting system respectively.
158

Figure 9-13 Average hourly percent utilization of the lighting system [108].
Figure 9-14 Average hourly power consumption [108].
The 51.2% energy savings gained by enabling individual control of lights
appeared to be higher than studies conducted by other researchers [22, 109]. This might
be because approximately one-third of the office was a group meeting area, which was
159

not occupied as often as the personal workstations in the rest of the office. This large
area could now be unlit, whereas with the original lighting configuration it couldn’t.
The occupational nature of the users also led to some extended periods of under-use of
the office during the day. Nonetheless, the most important implication of this savings is
that enabling individual lighting control had the tendency of saving significant amounts
of energy while improving occupants’ satisfaction working under their preferred
lighting.
One interesting observation made from comparing Figure 9-13 and Figure 9-14
is that the more often the office was occupied, the more energy savings could be
generated. In the hour between 8am and 9am, the power consumption of the research
system was almost the same as that of the original all on/off operation, whereas the
energy usage was cut by more than a half with the research system during the late
morning and the afternoon hours. Aside from the typically unoccupied group meeting
area discussed before, the most likely cause of this was the overhead of the mote-based
system. Since the motes and the actuation modules were directly powered by the mains
and would continue to operate all the time, their energy consumption would also be
reflected from the lighting energy usage data and become noticeable, especially when
all the lights were off. This observation suggests that better selection of energy efficient
components for the actuation modules is necessary to further reduce the operation
overhead when it comes to mass production and commercialization. Furthermore,
although the impact on power consumption will not be as significant, a better
communication protocol could be employed so as to put the motes in a low-power sleep
mode as often as possible to reduce the energy draw.
160

9.4 Daylight Response
The second phase of this implementation integrates the system with the wireless
photosensors developed in Chapter 7.2 and the lighting optimization algorithm
developed in Chapter 6.3 for harvesting daylight. This phase serves as the verification
of the research lighting system performance operating under the presence of extraneous
and uncontrollable light sources. The capacity to respond to daylight is a critical feature
for the developed system as daylight harvesting is one of the most energy efficient
lighting management strategies.
9.4.1 Experiment Setup
An important simplification of the occupants’ lighting preferences was made for
this assessment. Although an ideal lighting condition includes sufficient task
illuminance as well as desirable surrounding lighting, occupants’ lighting preferences
were strictly confined to the illuminances on their desktops for simplicity. In other
words, the desktop illuminance specified by an occupant in the first phase of the
assessment was considered as his/her lighting preference. Incorporating the lessons
learned from Chapter 8.3 – that sensor readings don’t always perfectly match the task
illuminances due to the locations of sensors – occupants’ lighting preferences were
defined by the sensor readings instead of illuminances directly read off of a light meter
placed at the center of the working areas.
Since the occupants mostly worked on their own desks instead of in the group
meeting area, only the part of the office with the workstations was implemented with
161

the automatic lighting optimization program. A lighting structure simulated a window
with daylight coming through. This was mounted on the cabinets facing south at the
northwest side of the working area in the office. The details of this lighting structure are
depicted in the next section.
Wireless photosensors were deployed onto the desktops to form a sensor
network. The readings from each sensor are relayed to the local server of the system as
described earlier in this chapter, where the optimal light settings are calculated and
actuation commands are issued to the wireless-enabled luminaires. Execution of the
lighting optimization algorithm is triggered when any occupant’s task illuminance
exceeds ±5% of the preferred lighting or balanced lighting if confliction exists.
Similarly, iterations of lighting optimization terminate only when all the occupants’ task
illuminances have been within 5% of the preferred lighting or balanced lighting if
confliction existed.
9.4.2 Artificial Daylighting Generation
As mentioned in Chapter 9.2, the research system was realized in an office
located in the interior of the building without any window. Thus, a separate structure
was constructed for simulating daylight.
The daylight simulating structure is made of PVC pipes with a 4' by 1' 4-lamp
fluorescent light troffer attaching to it as shown in Figure 9-15. The troffer is equipped
with a dimmable ballast for adjusting the intensity of the light output and full spectrum
lamps that deliver close-to-nature light. The light can be dimmed and brightened using a
handheld remote controller or a 0-10V adjustable voltage supply for finer resolution.
162

Furthermore, the structure allows the luminaire to tilt and thus simulate different
incident angles.
Figure 9-15 Daylight simulating structure.
The structure was mounted on top of the file cabinet that separates the work area
and the group meeting area in the office; it faced the work area to mimic a window with
daylight penetrating through it. The overall height of this daylight source from the floor
to the center of the troffer was 6.8'.
The light spectrum generated from the daylight simulating structure is close to
natural daylight, the intensity, however, is much weaker. The intensity of the simulated
daylight attenuates quickly with distance. When illuminated by the simulated daylight
alone, the horizontal illuminance on the closest desktop used in this test, approximately
ten feet away from the structure, is about 100 lux, and that on the farthest desktop is less
than 10 lux. In other words, the simulated daylight can at most contribute to roughly
one-third of the lighting required by the occupants. Although not as strong as true
daylight, the simulated daylight is good for testing the system performance under the
163

situation where part of the room receives daylight and part of the room receives almost
no daylight.
9.4.3 Testing results
Three different cases were tested to verify the performance of the integrated
lighting system. The first case considered four occupants sitting sparsely in the office
with a well-planned daylight changing sequence. The second case increased the number
of occupants to seven with a wider range of lighting preferences, and the simulated
daylight was also changed in accordance with a simple sequence. The last case
considered the same occupants as the second one, but the simulated daylight followed a
downscaled pattern of real daylight change.
Four-occupant case with simple daylight changing sequence
Four occupants were considered in the first test. As shown in Figure 9-16, the
user-specified task illuminances are denoted on the desks at which these occupants sat.
The change of percentage output over time in relation to the maximum light output from
the daylight simulating structure is shown in Figure 9-17. The light from the daylighting
simulating structure was turned on to the dimmest level four minutes after the test
started and was gradually brightened approximately every two minutes to the maximal
level. The simulated daylight output was then dimmed to 40% for two minutes, turned
back up to the maximum for three minutes, and was then gradually dimmed to the
minimal level before being turned off.
164

Figure 9-16 Daylight harvesting test on four occupants.
Figure 9-17 Simulated daylight output for the first case.
The readings from the four photosensors mounted on the occupants’ desktop are
shown as the blue lines in Figure 9-18 and are ordered by the degree of each sensor
being affected by the extraneous light. The occupant numbers labeled in the upper-right
corner of each subplot in Figure 9-18 correspond to the circled digits in Figure 9-16,
which signify the location of each occupant. The red solid lines in Figure 9-18 represent
165

the occupants’ preferred lighting, and the dashed cyan lines indicate the ±5% boundary,
beyond which new iterations of lighting optimization will be triggered.
Except for the initialization of the test, the sensor readings in the bottom two
plots of Figure 9-18 never exceed 5% of the preferred lighting, and were not
significantly affected by the extraneous light since the two desks were located deep in
the office away from the daylight simulating structure. The sensor readings in the upper
two plots of Figure 9-18 better reflect the reception of daylight on the two desks closer
to the lighting structure, especially the one where occupant number three sat. The
overhead lights in the office were able to quickly compensate for the available daylight
and keep the desktop illuminances well within 5% of the occupant-specified
preferences. From the observation made in the office during the test, the change of the
electric lights was actually very subtle and nearly unnoticeable to bare eyes.
166

Figure 9-18 Sensor readings in the first case.
Figure 9-19 shows the percentage of energy consumption of the system during
this test compared to the original lighting configuration where all the overhead lights
can only be turned on/off together with a single switch. Only 36.4% of energy was used
to deliver occupants’ preferred lighting for the first four minutes before harvesting
daylight. The energy usage was further reduced roughly in accordance with the amount
of available daylight, and the minimum of 31.0% energy usage was achieved when the
167

light output from the daylight simulating structure reached the maximum level. Notice
that each stair in Figure 9-19 indicates a new energy usage level after a set of new
iterations of lighting optimization was triggered and revised light settings were
delivered. Remember that the lighting optimization algorithm was not necessarily
triggered every time the simulated daylight changed, but only when the task illuminance
on any of the present occupant’s desk diverged from specified preference by 5%.
Figure 9-19 Percentage of energy consumption in the first case.
Seven-occupant case with simple daylight changing sequence
This test considered a more compactly occupied office where seven occupants
were present with a wide range of lighting preferences as shown in Figure 9-20. Figure
9-21 illustrates the change sequence of the simulated daylight. The simulated daylight
was turned on to the dimmest level four minutes after the test started and was gradually
brightened approximately every two minutes to the maximum. The simulated daylight
was then dimmed to 50% for two minutes, turned back up to the maximal level, and
168

then gradually dimmed. After reaching the minimum, the light output was brought up to
80% for two minutes and then returned to the minimum before being turned off.
Figure 9-20 Daylight harvesting test on seven occupants.
Figure 9-21 Simulated daylight output for the second case.
The sensor readings from each of the seven sensors on the occupants’ desktops
are shown in Figure 9-22 and Figure 9-23, and are ordered by the amount of daylight
reception. The occupant numbers labeled in the upper-right corner of each subplot
correspond to occupants’ locations annotated by the circled digits in Figure 9-20. The
169

solid red horizontal lines and dashed cyan lines in the same plots represent the
occupant’s preferred lighting and the ±5% boundary respectively. As soon as any sensor
reading deviated from the occupant’s specified preference by 5%, new iterations of
lighting optimization were triggered to bring the light back within the ±5% range. It is
obvious that the closer the desks to the daylight simulating structure, the more
susceptible to the extraneous light the sensors would be. Nonetheless, the light was
maintained well within 5% of each occupant’s preferred lighting during the entire test.
The percentage of energy usage of the system during the test compared to the
original all-on/all-off lighting configuration is shown in Figure 9-24. Only 33.3% of
energy was used to deliver occupants’ preferred lighting before daylight was available.
The energy consumption reduced while harvesting daylight, and the minimum energy
usage of 28.9% was achieved when the light output from the daylight simulating
structure was at the brightest level.
170

Figure 9-22 Sensor readings in the second case (1).
171

Figure 9-23 Sensor readings in the second case (2).
Figure 9-24 Percentage of energy consumption in the second case.
172

Seven-occupant case with simulated daylight fluctuation sequence
Given the promising results of the lighting system responding to the wellplanned
daylight changing sequences in the previous two cases, a more close-to-real
daylight fluctuation sequence was employed to evaluate the system performance. The
same seven occupants and lighting preferences as the preceding case (Figure 9-20) were
considered in this test. The daylight fluctuation data was measured using a photosensor
placed five feet away from a north-facing window on an ordinary day with a sensing
rate of ten seconds per sample as shown by the blue line in Figure 9-25. In order to
make it possible to replicate the sequence of daylight change with the daylight
simulating structure, the daylight fluctuation data was sampled every ten minutes and
scaled to fit the operation range of the simulating structure. The red crosses in Figure
9-25 indicate the data points considered in this test, and the light output from the
daylight simulating structure was adjusted accordingly every minute instead of every
ten minutes. As a result, the simulated daylight followed the trajectory in Figure 9-26.
Figure 9-25 Measured daylight fluctuation.
173

Figure 9-26 Light output from the daylight simulating structure.
The solid blues lines in Figure 9-27 and Figure 9-28 show the sensor readings
from the seven desktop-mounted sensors, which correspond with the occupants and
ordered by the level of daylight reception. The occupant numbers associate each
occupant’s location to the circled digits in Figure 9-20. The solid red horizontal lines
and dashed cyan lines in the same plots represent the occupant’s preferred lighting and
the ±5% boundary respectively. All seven sensor readings, especially the ones close to
the daylight simulating structure (plots in Figure 9-27), did fluctuate with the change of
simulated daylight, but were maintained within 5% of occupants’ preferred lighting.
New iterations of lighting optimization were triggered to quickly bring the task
illuminances back to occupants’ preferences once any sensor reading exceeded the 5%
boundary. Like the observations made in the previous two cases, the change of electric
lights was very subtle and nearly unnoticeable in this case.
174

Figure 9-27 Sensor readings in the third scenario (1).
175

Figure 9-28 Sensor readings in the third scenario (2).
The 5% boundary in the above paragraph was picked mainly for demonstrating
the performance of the system harvesting daylight. In reality, however, sensible minor
fluctuations of daylight may actually contribute to positive experiences, which make the
occupants feel more connected to the nature. Therefore, when the system is deployed to
a real daylit office, the value of being in tune with fluctuations from natural light should
also be taken into account to create an enjoyable working environment.
The percentage of energy consumption of the system during the test compared to
the original all-on/off lighting configuration is shown in Figure 9-29. 34.3% of energy
was used to deliver occupants’ preferred lighting when there was no extraneous light.
The energy usage decreased after harvesting daylight, and could be as low as 29.1%
176

under significant daylight. Again, each stair in Figure 9-29 represents a new energy
usage level after the lighting optimization algorithm revised the light settings to
compensate for available daylight.
Figure 9-29 Percentage of energy consumption in the third scenario.
9.4.4 Discussion
The testing results presented in the previous section demonstrate a promising
performance of the integrated system responding to daylight, and hence the potential of
further boosting energy savings. Meanwhile, important observations and valuable
experiences have been learned from the tests and are discussed in this section.
It was found that when a room is packed with too many occupants, lighting
optimization could become infeasible. This issue is caused by too many competing
lighting preferences relative to the degree of freedom for optimization. From a
mathematical standpoint, the twelve luminaires implemented in the office for the
previous testing cases translate to twelve variables in the linear programming problem.
The constraints of the problem are formulated from each occupant’s lighting preference.
177

The physical limitations of light outputs from the luminaires are also posed as
constraints to the linear programming problem. As a result, if the number of constraints
(present occupants) is close to or even exceeds that of the variables (number of
luminaires), it is very likely that the optimization simply becomes overly stringent and
yields no feasible regain for an optimal solution. The over-constraining issue signifies
the critical link between degrees of freedom and scalability. In short, the scalability
potential of the lighting system depends more on the ratio of the luminaires and the
occupants than the size of the office or the total number of the luminaires. In practice, a
well-lit shared-space office generally has the proper number of luminaires appropriately
arranged to deliver designed lighting and it usually accommodates a reasonable number
occupants. Therefore, scalability of the system will not be an issue for a typical office.
Higher-level system design is necessary to make the current lighting control
system ready for practical realization. Although not considered in the previous tests,
exceptional cases have to be accounted for. One extreme would occur when the
available daylight alone already exceeds occupants’ lighting preferences. There will be
no way for the lighting system to generate light settings that counterbalance excessive
light. This worst-case scenario points to future research of integrating the automatic
shading system discussed in Chapter 10.3.2. The antithesis of the over-lit condition
arises when the occupants’ lighting requirements cannot be satisfied even if the settings
for electric lights have maxed out. This will most likely happen in legacy buildings that
are originally under-lit in the first place. Furthermore, the testing scenarios in the
previous sections assumed no change on occupancy status during the test. In reality,
occupancy change will certainly trigger new iterations of lighting optimization, where
178

newly calculated light settings may be significantly different from the previous one, and
the light change may become noticeable to the occupants. The system needs to ensure a
smooth transition when encountering a change of occupancy in order not to irritate
people with a series of annoying light adjustments.
9.5 Commercial Implication
As the energy crisis and global warming continues to worsen, the proposed
lighting system could potentially generate over 50% energy savings and high user
satisfaction with both legacy buildings and new constructions. With the development of
wireless sensor and actuator network technologies over the years, this section provides
an analysis of the commercial implications of the research system, including the current
and projected payback period and potential bottlenecks.
9.5.1 Payback Period – Current and Projected Payback Period
A payback period analysis on the research intelligent wireless lighting system is
provided in this section. Since the energy savings are different from building to building
largely dependent on occupancy pattern, available daylight, etc., educated guesses have
to be made to estimate some of the parameters required in the analysis.
Consider a 150 square-meter open-plan office space originally designed with a
single wall switch to control all luminaires in the office. The number of luminaires and
ballasts required to light the space is calculated from the lumen method for average
illuminance [110]. The average illumination calculation sheet attached in Appendix C.1
179

indicates that 15 fixtures and ballasts are sufficient to light the office under the
following assumptions:
a. The office is 15m in length, 10m in width and 3m in height, and features typical
ceiling/wall/floor reflectance of 80/50/20% [20]. The luminaire mounting
height, which is defined as the height from the worksurface to the fixture, is
1.65m.
b. Each luminaire is equipped with one 4-lamp electronic ballast and four 32W T8
fluorescent lamps.
c. The lamps are generic T8 fluorescent light tubes, such as those manufactured by
Philips [111] and GE [112] with 3500K color temperature and 2660 design
lumens.
d. The input power to each 4-lamp ballast is 98W, in accordance with a
representative electronic instant-start ballast manufactured by Advance
Transformer Co. [113].
e. The office was originally designed such that the worksurface illuminance is 500
lux, meeting the IES (Illuminating Engineering Society) recommendations for
visual tasks of medium contrast or small size [110].
f. The lights are on for 14 hours (from 7am to 9pm) per day, 5 days a week; in
other words, 260 days per year.
g. The office is considered a clean environment for calculating the light loss factor
(LLF).
The annual energy consumption under the original non-dimming groupcontrolled
ballasts is estimated at 5,350.8 kWh, as calculated in (9.1).
180

Annual energy consumption
= ( Number of ballast) ( Ballast power) ( Daily operating hours)
( Operating days per year)
(9.1)
= 15 98 14 260 = 5, 350,800 Wh year = 5, 350.8 kWh year
The prototyping costs for each component of the research system are
summarized in Table 9-1, as are the projected prices for both retrofitting and new
installation after mass production and if the technology has matured. The difference
between retrofitting and new installation is the requirement of dimmable ballasts.
Dimmable ballasts are not included as a part of the cost for new installations since
ballasts are inevitable for fluorescent lighting systems, and it is a trend to adopt
dimmable ballasts for daylight harvesting capabilities. Moreover, daylight responsive
capabilities have become mandatory for new constructions in some states’ code of
regulations such as the Energy Efficiency Standards of California, and dimmable
ballasts are part of the construction cost. Most of the off-the-shelf electronic
components have price quotes available for large quantities, however the future price of
dimming ballasts and motes must be determined by educated guesses. The ultimate
price of the wireless platforms used in these calculations was $10. In fact, researchers
and developers of the wireless platforms optimistically estimate the cost to be about $1
in the future.
181

Item
Table 9-1 System commonpont price.
Quantity
required
Prototype
unit price
Projected unit
price for
retrofitting
Projected unit
price for new
installation
Light sensor module 3 per desk $25.00 $10.00 $10.00
Wireless ballast
actuation module
1 per ballast $77.00
$15.00
$15.00
4-lamp dimming ballast 1 per luminaire $95.00 $40.00 $0.00
Wireless motes 1 per module $78.00 $10.00 $10.00
Total cost for an office with 15 luminaires
and 15 occupants
$8,385 $1,875 $1,275
Suppose the retail price of electricity in commercial buildings is 9.62 cents per
kilowatt-hour, as reported by the Energy Information Administration in April 2008
[114]. The annual electricity cost for the manual non-dimmable control is $515. The
energy savings versus payback period of the research system with current and projected
prices are shown in Figure 9-30 and Figure 9-31 respectively.
The hardware cost for retrofitting a 15-person 150m 2
office space with the
prototype intelligent system is about $8,400, and there is no reasonable payback time, as
Figure 9-30 suggests, no matter how much savings the system can possibly generate.
This finding is not surprising in that the prototype was designed for proof of concept
instead of optimized for commercialization. The small-quantity purchase of the
prototype components also contributes to their high unit price. More importantly, the
application of wireless sensor and actuator network technologies to energy-efficient
lighting systems is a new concept and is still at very early stage, and significant price
reduction can be expected once the industry has recognize its marketing potential.
182

Figure 9-30 Energy savings vs. payback period with current system cost.
Figure 9-31 Energy savings vs. payback period with projected system cost.
The predicted hardware cost of the system drops to approximately $1,875 for
retrofitting at the maximum as shown in Table 9-1, and a reasonable payback period
183

ecome possible if the system generates more than 60% savings as the blue solid line
reveals in Figure 9-31. Since the system is capable of harvesting daylight and
considering occupancy status simultaneously, it is not unrealistic to expect energy
savings greater than 60% as demonstrated earlier in Chapter 9.4. Like most of the
energy-efficient lighting systems, the projected payback time for retrofitting, however,
stays at the high end of the three- to five-year period acceptable to facility managers for
adopting energy efficient products. The next section provides a detailed discussion on
the potential commercialization bottleneck of the proposed lighting system. An
exemplary analysis of energy savings and payback period for this particular office can
be found in Appendix C.2. The payback period for new installation is about 30% shorter
than for retrofitting without the overhead of dimmable ballasts as the green dashed line
suggests in Figure 9-31. In the meantime, it is important to recognize that in addition to
energy efficiency, the research system also emphasizes on satisfying occupants’ lighting
preferences. This desirable feature and unique advantage could potentially outweigh a
longer but reasonable payback period.
9.5.2 Commercialization Bottleneck
According to Table 9-1, the potential bottleneck for commercializing the
intelligent lighting system will be the dimmable ballasts. Although the prototyping
sensors and actuation modules currently cost more than the dimmable ballasts, the
prices of the components are expected to drop significantly, especially those for the
electronic components. However, the technologies of dimmable ballasts have been
mature and on the market for a while. Even though the notably fewer market demands
184

have partially contributed to the high unit price, the cost of a dimmable ballast is not
likely to fall below that of its non-dimmable counterpart, which is in the range of $20 to
$30.
Figure 9-32 shows the projected payback periods with respect to the unit price
of dimmable ballasts assuming that the lighting system generates 60% energy savings
on average. At the price of $40 per ballast as predicted in Table 9-1, the payback time
of the proposed lighting system will be a little more than six years. For spaces that are
already equipped with dimmable ballasts, such as those where daylight harvesting is
mandatory under California’s Title 24 energy code, the zero additional cost on
dimmable ballasts will results in a four-year payback period as suggested in Figure
9-32. As the necessity of dimmable ballasts for versatile energy-efficient lighting
management technologies has been recognized, it will soon become a trend for
dimmable ballasts to take over the market of their non-dimmable counterpart.
Figure 9-32 Dimmable ballast unit prices vs. payback periods.
185

As the wireless platform will be made compact by integrating all components
into a single chip, the only element that may not be further integrated and is not
showing strong potential of price reduction is the optical component for the
photosensors, i.e., the photodiodes. The photodiodes are estimated to stay at the range
of $10-$15 per piece. Figure 9-33 shows the impact of the photodiode unit price on the
payback period of the lighting system. It is assumed that the unit price of the dimmable
ballasts is $40, and 60% average energy savings is attained by the system. The payback
time of the lighting system does shorten with the price drop of photodiodes, but is not as
cost-sensitive as with that of dimmable ballasts. This again verifies that the price of
dimmable ballasts dominates the payback period of the proposed lighting system.
Figure 9-33 Photodiode unit prices vs. payback periods.
Lastly, the relationship between the unit price of the wireless platforms and the
payback periods is presented in Figure 9-34 assuming that the dimmable ballasts are
186

$40 per unit and the system achieves 60% energy savings on average. According to the
plot, the payback time can be reduced to within five years if the cost on wireless
platforms drops below $4 per unit even if the unit price of dimmable ballasts stays at
$40. Although the unit price of the wireless platforms does not have as significant
impact on the payback period as that of the dimmable ballasts, the price reduction of the
wireless platforms will certainly shorten the payback period.
Figure 9-34 Wireless platform unit prices vs. payback periods.
187

Chapter 10 Conclusion and Future Research
10.1 Conclusion
This section provides the overall conclusions and findings of the research
covered in this dissertation. This dissertation research has been devoted to the
development of technologies for energy-efficient intelligent lighting systems harnessing
wireless sensor and actuator network technologies. Prototypes of wireless-enabled
photosensors and ballast actuating interfaces were developed as a proof of concept for
wirelessly connected and individually addressable lighting components.
Starting with sensing technologies for massive-deployed, energy-constrained
and disturbance-prone small photosensors, the mote-FVF sensor validation and fusion
algorithm presents an efficient method extracting pertinent lighting information while
isolating faulty or disturbed sensor readings. The adaptive sensing strategy dynamically
adapts the sensing rates of the photosensors to changes in daylight, and hence optimizes
the timing for energy-hungry wireless communication for real-time feedback lighting
control. Moving from passively sensing to actively affecting the environment, the
optimal lighting actuation algorithm delivers a lighting condition that simultaneously
minimizes energy usage and satisfies occupants’ lighting preferences by leveraging the
individual addressability of the wirelessly networked ballast actuation modules.
The human subject test provides valuable insight and recommendations for
future implementation of the research system. The final integrated wireless-enabled
intelligent lighting system implementation demonstrates the feasibility and superiority
of the research system in terms of energy savings and user satisfaction. The encouraging
188

performance of the current research system also points to future research directions on
supervisory-level design for practical system realization and commercialization.
10.1.1 Wireless-enabled Sensor and Actuator
The design of photosensors with photodiodes integrating onto the wireless mote
platforms has demonstrated the integration of wireless sensor networks to lighting
technologies. The response of the photosensors was designed to closely match human
eyes and be sensitive to the range of normal indoor lighting. With the much smaller
size, lower unit price, and simpler deployment, the wireless photosensors have shown
promise to replace traditional ceiling-mounted photosensors by being placed directly on
the desktops for better estimating human perception of light.
The development of wireless ballast actuation modules has explored the
possibility of extending the monitoring role of typical wireless sensor networks to
actuation, which actively affects the environment. In addition to the most obvious
benefit of circumventing exorbitant and complicated rewiring with wireless links,
individual addressability of the wireless actuation modules leads to a whole new
direction of lighting control. With the capability of being dynamically configured and
actuated at different levels, the wirelessly enabled luminaires present the promising
potential to achieve energy efficiency and user satisfaction in a more versatile yet
economic way.
189

10.1.2 Sensor Validation and Fusion for Wireless Sensor Networks
The mote-FVF sensor validation and fusion algorithm exploits fuzzy logic and
time series exponential moving average techniques to extract pertinent information from
redundantly deployed sensors while rejecting faulty or disturbed sensors. The miniature
wireless photosensors afford being placed on the desktops for better estimating human
perception of light. Redundancy has to be enforced in order to account for the tendency
of the tiny desktop-mounted photosensors to be disturbed. The sensor validation and
fusion algorithm complements the redundant-deployment strategy and returns
representative fused values for lighting control purposes.
The developed sensor validation and fusion algorithm has shown its ability to
work with sensors monitoring daylight changes and can function appropriately under
any possible desktop illuminance change in a daylit environment. The fused values
were accurate within 5% of the gold standard light meter readings in the testing
experiments, even in situations where some of the sensors were experiencing failures
and disturbances. The maximum error ever measured between the fused value and the
real illuminance was well beyond the sensitivity of human eyes. The implementation of
a closed-loop lighting system using fused sensor readings as sensory feedback also
demonstrated good performance of keeping the desktop illuminated at the desired level
under the appearance of extraneous light and sensor failure.
Although the mote-FVF algorithm was designed for and tested on lighting
applications, it can easily be adapted to other applications where aggregated sensor
values may reveal more valuable information than row readings from individual
190

sensors. The research in [14] presents an example of extending the mote-FVF algorithm
to applications on monitoring the structure health of space vehicles.
10.1.3 Autonomous Sensing with Adaptive Rate
The sensing rate adaptation algorithm employs prediction models to forecast the
next incoming sensor reading based on past readings and utilizes fuzzy logic to adapt
the sensing rate to the change of the environment according to the prediction error. The
algorithm serves as a useful mechanism when monitoring stimulus with changing
dynamics, especially if sensing is an expensive action. In the lighting application with
energy-constrained photosensors operating at an ultra-low duty cycle, bringing up the
processor and data acquisition unit for sensing can sacrifice the life span of a sensor
node. More importantly, since real time sensory feedback is critical for lighting control
purposes, frequent power-hungry wireless communication is inevitable. The adaptive
sensing strategy optimizes the timing for sensing and wireless communication, and
ensures the resolution of the sensed information for good control performance.
Three different prediction models were evaluated for the algorithm, and all
resulted in adequate prediction performance with minimal parameter tuning. The fuzzy
rules efficiently adapted the sensing rate in response to the prediction errors to catch
change in the stimulus. A test of daylight monitoring on a typical sunny day reveals that
only less than 20% of the sensing action, and hence wireless communication, was
required by adaptive sensing to catch all the daylight changing information compared to
that obtained with a fixed sensing rate at ten seconds per sample. The sensing rate was
dynamically shifted between ten seconds and over five minutes per sample.
191

10.1.4 Optimized Lighting Actuation and Control
Leveraging the individual addressability of wireless-enabled luminaires, the
optimal lighting actuation algorithm generates a set of luminaire settings that results in
minimum energy usage while satisfying occupants’ lighting preferences and
requirements. Lighting actuation is formulated into a linear programming problem with
the overall light output from the luminaires as the objective function and occupants’
lighting preferences and the physical limitations of the lighting hardware as the
constraints. The solution to the linear programming problem is the set of optimal light
settings that minimizes overall energy consumption and meets each user’s lighting
preference.
When it is possible to satisfy every individual’s lighting requirement, the
algorithm has shown its power to deliver the optimal lighting in a shared-space office in
real time. Under the circumstances where occupants’ competing preferences prevent the
linear programming problem from finding a lighting configuration that satisfies
everyone, the algorithm generates a balanced lighting that best meets everyone’s
requirements.
Incorporating sensory feedback through iterations, the algorithm is able to
overcome uncertainties and inaccuracies and complement available extraneous light
with dimmable electric light to illuminate the desktops at occupants’ preferences. In
other words, the optimal lighting actuation algorithm is ready to work with sensors for
harvesting daylight and hence further increase energy savings.
192

Other than optimizing the energy consumption, the algorithm can potentially be
modified with different objective functions and constraints for other lighting
applications, such as stage lighting, mood lighting, and so on.
10.1.5 Human Subject Testing
Two topics related to the research system were studied on human subjects:
sensor placement and a comparison of the research system with a representative
commercial product. Both the research and the commercial lighting regulation systems
were implemented in the same small private office. Ten individuals were recruited to
participate in the test.
The sensor placement test studied the impact on the pertinence of the values
fused by the sensor validation and fusion algorithm if the sensor locations are
customized by the occupants. It also identified common patterns of sensor location
choices made by occupants. It was found that the fused sensor readings could deviate
from the task illuminance depending on how the sensors were placed. The lighting
might not be evenly distributed on the entire worksurface, and hence the readings from
each sensor could be inconsistent to some degree. Therefore, it is recommended that
occupants’ preferred lighting is expressed with respect to the fused sensor values for
lighting control purposes instead of referring to the pre-measured illuminance at the task
area.
No obvious pattern of sensor placement was observed from the test. However,
the fact that at least one sensor was pushed to the corner of the desk in almost all testing
cases implies that users move the sensors as far out as possible to maximize their
193

working area. Furthermore, participants tended to spread the sensors rather than place
them close to one another, and this natural practice helped reduce the chance of multiple
sensors being disturbed at the same time.
The comparison of the research and the commercial lighting regulation system
tested the preferred type of overriding mechanism as well as the performances of the
two systems. It was concluded from the participants’ feedback that overriding
mechanisms are necessary even in an automatically conditioned lighting environment,
which is consistent with the lighting system design guideline [52]. Various types of
overriding controllers need to be provided to satisfy occupants’ diverse usage patterns
and habits. The research system also showed competitive performance against
commercial daylight systems in regulating the desktop lighting at the occupant’s
preferred level.
10.1.6 Integrated Wireless-enabled Intelligent Lighting System
A self-configuring, multihop wireless photosensor and ballast actuator network
along with the backend supporting system were realized in a small shared-space office
with real occupants. Two energy-saving potentials of the integrated wireless-enabled
lighting system were evaluated: energy savings generated from allowing occupants to
specify their preferred lighting, and additional savings introduced by harvesting
daylight.
50% energy savings before harvesting daylight has been demonstrated in a longterm
assessment simply by allowing individuals to specify and work under their
preferred lighting. The percentage of savings was in comparison with the original
194

lighting configuration in the office where all the lights could only be toggled on/off
together. Occupants created presets of preferred lighting configurations, which were
retrieved and applied to the wirelessly networked luminaires upon entering the office
through a dedicated web-based GUI. The savings were generated merely from turning
off the lights above unoccupied areas and dimming the lights in accordance with
occupants’ actual needs.
A full-spectrum artificial light was introduced to simulate daylight and to
examine the performance of the intelligent lighting system responding to daylight. A
wireless photosensor network and the optimal lighting actuation algorithm were
implemented along with the wireless ballast actuator network. The integrated intelligent
lighting system has successfully shown smooth and prompt response to the change of
daylight while delivering light that satisfies occupants’ requirements. The illuminance
on each present occupant’s desk was maintained within 5% of the specified preference,
and less than 40% of energy was required in each test case to deliver the optimal
lighting compared to the original all-on/off lighting configuration. Daylight harvesting
alone contributed to about 5% additional savings on top of the savings from optimizing
present occupants’ lighting preferences. The savings from harvesting daylight could
potentially be more significant under real daylight since the simulated daylight was
much weaker than natural daylight.
10.2 Contribution
This dissertation research has proposed a framework for energy-efficient
lighting control in a shared-space office by harnessing wireless sensor and actuator
195

networks. In addition to the general notion that wireless technologies could significantly
bring down the retrofitting complexity and cost, this research explores in depth the ways
in which commercial lighting systems can benefit from wireless sensor and actuator
network technologies. This section summarizes the dissertation’s contributions to
theoretical and applications-oriented standpoints.
10.2.1 Theoretical Contributions
The theoretical contributions include the development of sensing and actuation
strategies for applying wireless network technologies to lighting systems. The
algorithms developed are lightweight and capable of being executed in real time on
wireless sensor and actuator networks. In addition to office lighting systems, a variety
of sensing and actuation applications could potentially benefit from this research.
The mote-FVF sensor validation and fusion algorithm is an application of fuzzy
logic that extracts information from redundant sensors. The fused information is more
valuable than that revealed by every single sensor. The intra-network sensor fusion has
verified the feasibility of in-network data aggregation with the mote-FVF algorithm.
The algorithm fits the massive-deployment nature of wireless sensor networks, and
addresses the issue of excessive and less-reliable sensor readings by condensing
information into fewer and trustworthy values.
The autonomous sensing with adaptive rate algorithm integrates time series
prediction and fuzzy logic so as to dynamically adapt the sensing rate to the change of
stimuli. This algorithm adds another layer of intelligence to the smart sensor nodes by
optimizing the timing for both data acquisition and the most power-hungry wireless
196

communications. The adaptive sensing algorithm contributes to relieving the energy
limitation issue intrinsic to wireless sensor networks from the application level while
retaining good resolution of the measurements.
The optimal lighting actuation algorithm exploits mathematical programming to
determine the most energy-efficient light settings given present occupants’ lighting
preferences. It has been recognized that densely deployed sensors could carry redundant
information about the monitored environment that can be aggregated using sensor
fusion techniques. Likewise, the lighting optimization algorithm was pioneered to
explore the idea that actions taken by each actuator in a wireless actuator network may
have overlapping effects on the environment and thus should be optimized for
efficiency. This algorithm presents a novel way for wireless networked actuators to
deliver the global optimal result, satisfying the specified objective and constraints.
Although realized exclusively for lighting control, the same framework may easily be
extended to actuator networks in other applications such as fire sprinkler systems,
agriculture irrigation systems, etc.
10.2.2 Applications-oriented Contribution
The application-oriented contributions made by this research lie in the
comprehensive study of integrating office lighting systems with wireless sensor and
actuator network technologies. Prototyping hardware and pilot implementations realized
in the progress of this research have demonstrated the practicability of such an
intelligent wireless lighting system.
197

Although wireless lighting products have already been on the market, the
functionality is confined merely to personal area manual lighting manipulation. The
dissertation pioneered the system level design of an energy-efficient wireless lighting
control system for shared-space offices. In addition to circumventing costly and
complicated retrofits as claimed by all wireless systems, the research also focused on
taking advantage of the individual addressability of wireless-enabled lighting
components. Possessing a similar addressability feature as its wired counterpart, DALI
[54] and LonWorks [56], the research system is even more flexible and less constrained
for configuring the lights to meet occupants’ preferences in an energy-efficient manner.
Human subject tests conducted in this research have verified the promising
performance of the wireless lighting system as compared to commercial products with
similar functionalities. The testing result also reveals the need for a variety of overriding
mechanisms even though the ultimate goal is an automatic lighting system. Users
always demand a certain degree of control over the environment in which they work,
and recognizing and respecting this desire could be the key for advanced wireless
lighting systems to successfully penetrate the market.
10.3 Future Research
The future research directions are discussed from the following three aspects:
further improvement of the research system, extension of the developed lighting system,
and application of the framework beyond the current domain. Enhancing the system
performance and increasing sensing and actuation capabilities are considered as further
improvements of the research system. Generating more energy savings through the
198

integration of the research system with other energy management strategies or systems
falls into the category of extending the developed lighting system.
10.3.1 Preference Modeling and Acquisition
Better acquiring and modeling occupants’ lighting preferences could further
improve user satisfaction of the research system. The lighting preferences of each
individual are considered as known and deterministic variables in the current system. In
reality, lighting preferences could change over time, and it is impractical to acquire
them with a stand-alone procedure. Therefore, a non-intrusive mechanism is necessary
for the system to learn and update users’ lighting preferences. The most promising way
to extract lighting preferences is by examining occupants’ use of overriding
mechanisms, analyzing the interactions between the occupants and the automatic
lighting system and employing machine-learning techniques.
Modeling lighting preferences is also critical to the performance of the lighting
system. The lighting optimization algorithm discussed in Chapter 6 simplifies the
problem by considering lighting preference as a single value representing an occupant’s
ideal lighting and trying to deliver the matched lighting. However, there may be a
specific range of illuminances that a person will consider ideal. In addition, people have
their own judgment on different levels of lighting; some may be more willing to receive
brighter light while some may prefer dimmer light when balancing the light is inevitable
to satisfy everyone’s requirements. The degree of willingness also varies with
individuals and tasks. A proper function that better captures each occupant’s lighting
preference needs to be derived to enhance the versatility of the research lighting system.
199

10.3.2 Integration with Other Sensing and Actuation Options
The ability to incorporate various sensing and actuation options is the key for
the developed lighting system to exploit other energy management strategies and
guarantee higher lighting quality. The intelligent lighting control framework proposed
in this dissertation makes use of photosensors to harvest daylight, and ballast actuators
to deliver the optimal lighting. In addition to daylight harvesting, the addition of
occupancy sensors will enable occupancy sensing strategies to facilitate more energy
savings. Leveraging the massive deploying nature of wireless sensor networks,
heterogeneous occupancy sensing technologies may be integrated for inferring pertinent
occupancy status to overcome the annoying false-offs, a problem from which most
commercial occupancy sensors suffer. Furthermore, the lighting optimization algorithm
discussed in Chapter 6, which requires occupants’ locations and presences as its inputs,
can also benefit from the integration of the networked occupancy detection
mechanisms.
In addition to the overhead dimmable fluorescent lights that this dissertation
research targets, other energy-efficient electric lighting alternatives exist, such as
compact fluorescent lamps (CFLs), light-emitting diode (LED) lamps, etc. Some of the
lighting alternatives can potentially be more cost-effective than linear fluorescent lights
since a specialized dimming ballast, which is identified as the commercialization
bottleneck in Chapter 9.5.2, is not required to dim the lights. Also, different light
sources may have various scopes and mounting options or locations. Generalizing the
developed wireless lighting control framework to work with other lighting alternatives
200

will broaden the application to different types of commercial or even residential
buildings.
Delivering satisfying task illuminances by properly actuating the electric lights
is a critical aspect of creating quality lighting. On the other hand, glare, excessive
natural light, and heat caused by direct sunlight are also factors that may jeopardize
lighting quality and comfort and should not be overlooked. Blinds and windows with
special treatment, such as electrochromatic windows, are mechanisms mostly used for
manually blocking out undesired extraneous light. Inappropriate use of the lightblocking
systems not only compromises lighting comfort but also sacrifices energy
savings. The actuation capability of the wireless sensor and actuator networks provides
a promising solution for integrating blind and window control into automatic lighting
systems. Advanced control strategies need to be developed to coordinate the control of
both the electric lights and the blinds/windows for maximum lighting quality as well as
energy savings.
10.3.3 Integration with Other Lighting Energy Management Strategies
As pointed out in Chapter 2.1.2, maximum savings will occur with a proper
combination of effective energy management strategies. The current research system
integrates daylight harvesting and light level tuning. While daylight harvesting
possesses the highest savings potential for open-plane offices receiving significant
daylight, other strategies, such as occupancy sensing discussed in the previous section,
will also help cut down the energy usage.
201

Demand response, which temporarily reduces energy consumption during peak
demand hours, is currently the most discussed energy management strategy for
preventing blackout or brownout. Although not aimed at generating significant energy
savings, demand response is the key technology with which modern energy-efficient
systems must be equipped or at least compatible. For lighting systems, load shedding
translates to reducing the light below the original setting by certain amounts. Given
occupants’ lighting preferences, preferably well-modeled ones as discussed in Chapter
10.3.1, it would be interesting to incorporate load shedding into the research system for
delivering the light that least compromises the users’ preferences during the peak hours.
10.3.4 Integration with the Building Management System
It has become increasingly common for large-scale buildings to employ building
management systems (BMS) to coordinate the mechanical and electrical systems and
manage energy usage. Being the largest electricity consumer among all the electrical
systems in buildings as pointed out in Chapter 2.1.1, it is crucial to integrate the lighting
system into building management systems.
This dissertation research has focused on the energy efficiency of building
lighting systems alone; however, energy usage of all the electrical and even mechanical
systems of a building may be interrelated in reality. For instance, the room temperature
could rise due to the incoming heat from the windows while harvesting daylight, and
consequently the air conditioning system has to compensate for it to ensure a
comfortable environment. As a result, some of the energy saved from the lighting
202

system could be reallocated to operating the air conditioning system, which in turn
diminishes the actual amount of energy savings.
A BMS has the potential of providing higher-level supervisory control to
optimize the energy usage among all the building systems. The current research lighting
system needs to be made compatible with the BMS, and be capable of communicating
with other electrical systems such as the heating, ventilation and air conditioning
(HVAC) system for maximum possible energy savings.
10.3.5 Application of the Proposed Framework in Other Domain
Although targeted at generating energy savings from the fluorescent lights in
commercial office buildings, the same framework could possibly be employed by other
applications with reasonable revision and extension.
Stage and theater lighting systems could potentially benefit from the intelligent
lighting control framework. Being another lighting application, the requirements of
lighting in the art and entertainment industry are very different from that in office
buildings. Energy conservation may not be a priority, but other parameters such as the
color and incident angle of light, background lighting, etc. can be critical in addition to
the illuminance at the targeted area. Furthermore, the lighting fixtures could be moving
or rotating along with the performers or in accordance with the scenes. Intelligently
coordinating the wireless sensors and lighting actuators in real time to deliver the
optimal lighting may revolutionarily simplify the current stage lighting control while
retaining the same or even better results.
203

Other than lighting, it would be interesting to investigate how the framework
developed can be used in applications where multiple actuators with overlapping scopes
are needed for monitored environments. Possible areas include disaster mitigation
systems, agricultural irrigation systems, and so on. In disaster mitigation systems such
as fire sprinkler systems [115, 116], the sprinklers could be coordinated to extinguish
the fire at the right spots detected by the sensor network and secure emergency exit
routes while preventing water damage to valuable assets in non-affected areas. In
agriculture irrigation systems, water may be efficiently supplied to different areas in the
field according to soil moisture monitored by the sensor network and types of crops
[117, 118].
The developed framework has shown great promise on lighting systems with
enhanced energy efficiency and user satisfaction. With a reasonable payback period
when technology matures and the system is optimized for commercialization, noticeable
savings can be achieved while creating a delightful lighting environment. The payback
will be generated from not only the substantially reduced use of energy, but also the
increased revenue resulting from more productive employees working under satisfying
lighting. As a positive impact on the environment, the introduction of the research
lighting system will result in significant reduction of greenhouse gas emissions and
hence alleviate global warming.
When applied to a broader range of applications, the framework developed in
this research may have the potential to introduce benefits that are unattainable with
typical practice. The ability to robustly extract pertinent information from massively
deployed sensors and promptly response with actions from highly coordinated actuation
204

entities could be tremendously critical for many monitoring and control tasks.
Performance art, disaster mitigation, and agriculture mentioned above are only a few
fields that this research could make positive impacts on and improve the efficiency,
effectiveness, and versatility of current practice.
205

References
[1] The Interlaboratory Working Group, "Scenarios for a Clean Energy Future:
Interlaboratory Working Group on Energy-Efficient and Clean-Energy
Technologies," NREL/TP-620-29379; ORNL/CON-476; LBNL-44029, 2000.
[2] "Energy Solution for Your Building," 2000. [Online]. Available:
http://www.eere.energy.gov/buildings/info/office/index.html. [Accessed: April
7, 2008].
[3] M. C. W. Kintner-Meyer and R. Conant, "Opportunities of Wireless Sensors and
Controls for Building Operation," in Proceedings of the 2004 ACEEE Summer
Study on Energy Efficiency in Buildings, Pacific Grove, CA, 2004, pp. 3.139-
3.152.
[4] V. Shnayder, M. Hempstead, B.-r. Chen, G. W. Allen, and M. Welsh,
"Simulating the Power Consumption of Large-Scale Sensor Network
Applications," in Proceedings of the 2nd International Conference on
Embedded Networked Sensor Systems, Baltimore, MD, 2004, pp. 188-200.
[5] O. Landsiedel, K. Wehrle, and S. Gotz, "Accurate Prediction of Power
Consumption in Sensor Networks," in Proceedings of the Second IEEE
Workshop on Embedded Networked Sensors (EmNetS-II), 2005, pp. 37-44.
[6] P. R. Tregenza, S. M. Romaya, S. P. Dawe, L. J. Heap, and B. Tuck,
"Consistency and Variation in Preferences for Office Lighting," Lighting
Research and Technology, vol. 6, pp. 205-211, 1974.
206

[7] J. A. Veitch and G. R. Newsham, "Preferred Luminous Conditions in Open-Plan
Offices: Research and Practice Recommendations," International Journal of
Lighting Research and Technology, vol. 32, pp. 199-212, 2000.
[8] G. Lu, B. Krishnamachari, and C. S. Raghavendra, "An Adaptive Energy-
Efficient and Low-Latency MAC for Data Gathering in Wireless Sensor
Networks," in Proceedings of the 18th International Parallel and Distributed
Processing Symposium, 2004, pp. 224-231.
[9] W. Ye, J. Heidemann, and D. Estrin, "An Energy-Efficient MAC Protocol for
Wireless Sensor Networks," in Proceedings of the 21st Annual Joint Conference
of the IEEE Computer and Communication Societies (INFOCOM 2002). New
York, NY, 2002, pp. 1567-1576.
[10] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, "Wireless Sensor
Networks: A Survey," Computer Networks, vol. 38, pp. 393-422, 2002.
[11] E. H. Callaway, Wireless Sensor Networks: Architectures and Protocols, 1st ed.
Florida: Routledge, USA, 2003.
[12] W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan, "Energy-Efficient
Communication Protocol for Wireless Micorsensor Networks," in Proceedings
of the 33rd Annual Hawaii International Conference on System Sciences,
Hawaii, 2000.
207

[13] J. Polastre, J. Hill, and D. Culler, "Versatile Low Power Media Access for
Wireless Sensor Netowkrs," in Proceedings of the 2nd International Converence
on Embedded Networked Sensor Systems, Baltimore, MD, 2004, pp. 95-107.
[14] J. Sauvageon, A. M. Agogino, A. F. Mehr, and I. Y. Tumer, "Comparison of
Event Detection Methods for Centralized Sensor Networks," in Proceedings of
the IEEE Sensors Applications Symposium (SAS 2006), Houston, TX, 2006.
[15] V. Fecteau, "Commercial and Institutional Building Energy Use," Office of
Energy Efficiency, Ottawa, Canada, Summary Report Survey 2000, December
2003.
[16] D&R International, 2007 Buildings Energy Data Book. Silver Spring: U.S.
Department of Energy, 2007.
[17] F. Rubinstein, J. Jennings, D. Avery, and S. Blanc, "Preliminary Results from
An Advanced Lighting Controls Testbed," Journal of the Illuminating
Engineering Society, vol. 28, pp. 130-141, 1999.
[18] E. Mills, "Why We're Here: The $230-billion Global Lighting Energy Bill," in
Proceedings of the Right Light 5, Nice, France, 2002, pp. 369-385.
[19] Architectural Energy Corporation and Lawrence Berkeley National Laboratory,
"Retrofit Fluorescent Dimming with Integrated Lighting Control - Economic
and Market Considerations," PIER Lighting Research Program, Sacramento, CA
September 11 2003.
208

[20] K. A. Karmel, High Performance Design Guild to Energy-Efficient Commercial
Buildings. Fayston, VT: AIA 2004.
[21] Energy Information Administration, "Commercial Buildings Energy
Consumption Survey (CBECS)," 2007. [Online]. Available:
http://www.eia.doe.gov/emeu/cbecs/contents.html. [Accessed: June 4, 2008].
[22] G. R. Newsham and J. A. Veitch, "Individual Control over Office Lighting:
Perceptions, Choices and Energy Savings," in Construction Technology
Updates, 1998.
[23] D. Maniccia, B. Rutledge, M. S. Rea, and W. Morrow, "Occupant Use of
Manual Lighting Controls in Private Offices," in Proceedings of the Illuminating
Engineering Society of North America 1998 Annual Conference, New York,
NY, 1998, pp. 490-512.
[24] B. VonNeida, D. Maniccia, and A. Tweed, "An Analysis of the Energy and Cost
Savings Potential of Occupancy Sensors for Commercial Lighting Systems,"
Journal of the Illuminating Engineering Society, vol. 30, pp. 111-125, 2001.
[25] C. DiLouie, "Lighting Automation Becomes the Norm," Lighting Strategies -
Online Exclusive, 2004. [Online]. Available:
http://www.buildings.com/articles/detail.aspx?contentID=1979. [Accessed: June
5, 2008].
[26] P. Littlefair, "Daylight Linked Lighting Control in the Building Regulations,"
Building Research Establishment Ltd CR399/99, June 1999.
209

[27] F. Rubinstein, D. Avery, J. Jennings, and S. Blanc, "On the calibration and
commissioning of lighting controls," presented at Right Light 4 Conference,
Copenhagen, Denmark, 1997.
[28] K. S. J. Pister, J. M. Kahn, and B. E. Boser, "Smart Dust: Wireless Networks of
Millimeter-Scale Sensor Nodes," 1999.
[29] Moteiv Corporation, "Tmote Sky Brochure," 2005. [Online]. Available:
http://www.sentilla.com/pdf/eol/tmote-sky-brochure.pdf. [Accessed: June 6,
2008].
[30] "SNM - The Sensor Network Museum," 2007. [Online]. Available:
http://www.btnode.ethz.ch/Projects/SensorNetworkMuseum. [Accessed: June 6,
2008].
[31] Crossbow Technology Inc., "Product Reference Guide - End to end Solutions
for Wireless Sensor Networks," 2007. [Online]. Available:
http://www.xbow.com/Support/Support_pdf_files/Product_Feature_Reference_
Chart.pdf. [Accessed: March 10, 2008].
[32] T. Bokareva, "Mini Hardware Survey," [Online]. Available:
http://www.cse.unsw.edu.au/~sensar/hardware/hardware_survey.html.
[Accessed: August 10, 2008].
[33] "Body Sensor Networks: Resource: Hardware Platforms," 2007. [Online].
Available: http://ubimon.doc.ic.ac.uk/bsn/index.php?m=206. [Accessed: August
10, 2008].
210

[34] P. Levis, S. Madden, J. Polastre, R. Szewczyk, K. Whitehouse, A. Woo, D. Gay,
J. Hill, M. Welsh, E. Brewer, and D. Culler, "Tinyos: An operating system for
sensor networks," in Ambient Intelligence, W. Weber, J. M. Rabaey, and E.
Aarts, Eds. New York: Springer Berlin Heidelberg, 2005, pp. 115-148.
[35] B. W. Cook, S. Lanzisera, and K. S. J. Pister, "Soc Issues for RF Smart Dust,"
Proceedings of the IEEE, vol. 94, pp. 1177-1196, 2006.
[36] B. W. Cook, A. Molnar, and K. S. J. Pister, "Low Power RF Design for Sensor
Networks," in Proceedings of the Radio Frequency Integrated Circuits (RFIC)
Symposium, 2005, pp. 357-360.
[37] A. Witvrouw, A. Mehta, A. Verbist, B. Du Bois, S. Van Aerde, J. Ramos-
Martos, J. Ceballos, A. Ragel, J. M. Mora, M. A. Lagos, A. Arias, J. M.
Hinoiosa, J. Spengler, C. Leinenbach, T. Fuchs, and S. Kronmuller, "Processing
of MEMS Gyroscopes on Top of CMOS ICs," in Proceedings of the
Proceedings of the IEEE International Solid-State Circuits Conference (ISCC),
2005, pp. 88-89.
[38] T. Denison, J. Kuang, J. Shafran, M. Judy, and K. Lundberg, "A Self-Resonant
MEMS-based Electrostatic Field Sensor with 4V/m/Hz Sensitivity," in
Proceedings of the Proceedings of the IEEE International Soild-State Circuits
Conference, 2006, pp. 1121-1130.
[39] Eidgenossishe Technische Hochschule Zurich, "BTnode rev3 - Product Brief,"
2006. [Online]. Available:
211

http://www.btnode.ethz.ch/pub/files/btnode_rev3.24_productbrief.pdf.
[Accessed: July 19, 2008].
[40] Intel, "Intel Mote 2 Overview," [Online]. Available:
http://www.intel.com/research/downloads/imote_overview.pdf. [Accessed: July
19, 2008].
[41] Dust Networks Inc., "SmartMesh-XR Evaluation Kit," 2005. [Online].
Available: http://www.dustnetworks.com/docs/Evaluation_Kit.pdf. [Accessed:
July 19, 2008].
[42] Sensicast System Inc., "SensiNet - The Wireless Sensor Network," 2006.
[Online].
Available:
http://www.sensicast.com/datasheet_form.php?deliver_resource=SensiNet%20B
rochure0906-101706.pdf. [Accessed: July 19, 2008].
[43] S. Lindsey and C. S. Raghavendra, "PEGASIS: Power-Efficient Gathering in
Sensor Information Systems," in Proceedings of the IEEE Areospace
Conference, 2002, pp. 3-1125 - 3-1130.
[44] R. Verdone, D. Dardari, G. Mazzini, and A. Conti, Wireless Sensor and
Actuator Networks: Technologies, Analysis and Design. London, UK: Elsevier
Ltd., 2008.
[45] A. Mainwaring, D. Culler, J. Polastre, R. Szewczyk, and J. Anderson, "Wireless
Sensor Networks for Habitat Monitoring," in Proceedings of the 1st ACM
212

International Workshop on Wireless Sensor Networks and Applications, Atlanta,
GA, 2002, pp. 88-97.
[46] J. A. Stankovic, Q. Cao, T. Doan, L. Fang, Z. He, R. Kiran, S. Lin, S. Son, R.
Stoleru, and A. Wood, "Wireless Sensor Networks for In-Home Healthcare:
Potential and Challenges," in Proceedings of the High Confidence Medical
Device Software and Systems (HCMDSS) Workshop, Philadelphia, PA, 2005.
[47] M. Fogel, N. Burkhart, H. Ren, J. Schiff, M. Meng, and K. Goldberg,
"Automated Tracking of Pallets in Warehouses: Beacon Layout and Asymmetric
Ultrasound Observation Models," in Proceedings of the IEEE International
Conference on Automation Science and Engineering, Scottsdale, AZ, 2007, pp.
678-685.
[48] W. Chen, L. Chen, Z. Chen, and S. Tu, "WITS: A Wireless Sensor Network for
Intelligent Transportation System," in Proceedings of the 1st International
Multi-Symposiums on Computer and Computational Sciences (IMSCCS'06),
Hangzhou, China, 2006, pp. 635-641.
[49] J. D. Huang, C. S. Yeh, C. S. Chen, C. K. Lee, and W. J. Wu, "Design and
Inplementation of a Wireless Sensor Network for Smary Living Spaces," in
Proceedings of the Sensors and Smart Structures Technologies for Civil,
Mechanical, and Aerospace Systems 2008, San Diego, CA, 2008, pp. 69323P-1
- 69323P-9.
213

[50] P. Hochmuth, "GM Cuts the Cords to Cut Costs," TechWorld, 2005. [Online].
Available:
http://www.techworld.com/mobility/features/index.cfm?featureid=1530.
[Accessed: July 21, 2008].
[51] I. F. Akyildiz and I. H. Kasimoglu, "Wireless Sensor and Actor Networks:
Research Challenges," Ad Hoc Networks, vol. 2, pp. 351-367, 2004.
[52] N. Morel, "Daylight and Electric Lighting Control Systems Design Guide,"
International Energy Agency Solar Heating & Cooling Programme Task 31,
April 17 2005.
[53] F. Rubinstein, S. Johnson, and P. Pettler, "IBECS: An Integrated Building
Environmental Communications System - It's Not Your Father's Network," in
Proceedings of the ACEEE Summer Study, Pacific Grove, CA, 2000.
[54] [DALI AG] Digital Addressable Lighting Interface Activity Group, DALI
Manual. Frankfurt am Main: DALI AG of ZVEI, Division Luminaires, 2001.
[55] L. Meyer, "An Introduction to Digital Addressable Lighting Interface (DALI)
Systems & Study of a DALI Daylighting Application," MS Thesis, Kansas State
University, Manhattan, 2007.
[56] Echelon Corporation, Introduction to the LonWorks System. Palo Alto, CA,
1999.
214

[57] LonMark International, "Fact Sheet: Control Technology for Future
Generations." [Online]. Available:
http://www.lonmark.com/connection/solutions/fact%5Fsheets/ControlTechnolog
yForFutureGenerations1.pdf. [Accessed: June 9, 2008].
[58] F. O'Reilly and J. Buckley, "Use of Wireless Sensor Networks for Fluorescent
Lighting Control with Daylight Substitution," in Proceedings of the Workshop
on Real-World Wireless Sensor Networks, Stockholm, Sweden, 2005.
[59] D. Teasdale, F. Rubinstein, D. S. Watson, and S. Purdy, "Adapting Wireless
Technology to Lighting Control and Environmental Sensing," July 2006.
[60] "Adura Technologies," 2008. [Online]. Available:
http://www.aduratech.com/index.php. [Accessed: June 10, 2008].
[61] Center for the Built Environment (CBE), "Development of a Prototype Wireless
Lighting Control System," 2008. [Online]. Available:
http://www.cbe.berkeley.edu/research/wireless_lighting.htm. [Accessed: June
10, 2008].
[62] V. Singhvi, A. Krause, C. Guestrin, James H. Garrett, Jr., and H. S. Matthews,
"Intelligent light control using sensor networks," in Proceedings of the 3rd
international conference on Embedded networked sensor systems, San Diego,
California, USA, 2005.
[63] "X10 Powerline Carrier (PLC) Technology," [Online]. Available:
http://www.x10.com/support/technology1.htm. [Accessed: June 10, 2008].
215

[64] S.-F. Li, "Wireless Sensor Actuator Network for Lighting Monitoring and
Control Application," in Proceedings of the IEEE Consumer Communications
and Networking Conference, Las Vegas, NV, 2006, pp. 974-978.
[65] B. Ahlgren, L. Eggert, B. Ohlman, and A. Schieder, "Ambient Networks:
Bridging Heterogeneous Network Domains," in Proceedings of the 16th Annual
IEEE International Symposium on Personal Indoor and Mobile Radio
Communications (PIMRC 2005), Berlin, Germany, 2005.
[66] J. Granderson, "Human-Centered Sensor-Based Bayesian Control: Increased
Energy Efficiency and User Satisfaction in Commercial Lighting," PhD
Dissertation, University of California, Berkeley, 2007.
[67] R. D. Shachter, "Evaluating Influence Diagrams," Operations Research, vol. 34,
pp. 871-882, 1986.
[68] H. Park, J. Burke, and M. B. Srivastava, "Design and Implementation of a
Wireless Sensor Network for Intelligent Light Control," in Proceedings of the
6th International Conference on Information Processing in Sensor Networks,
Cambridge, MA, 2007, pp. 370-379.
[69] H. Park, J. Friedman, P. Gutierrez, V. Samanta, J. Burke, and M. B. Srivastava,
"Illumimote: Multimodal and High-Fidelity Light Sensor Module for Wireless
Sensor Networks," Sensors Journal, IEEE, vol. 7, pp. 996-1003, 2007.
[70] D. L. Hall and J. Llinas, "An Introduction to Multisensor Data Fusion,"
Proceedings of the IEEE, vol. 85, pp. 6-23, 1997.
216

[71] S. Sahni and X. Xu, "Algorithms for Wireless Sensor Networks," International
Journal of Distributed Sensor Netoworks, vol. 1, pp. 35-56, 2005.
[72] L. Lamport, R. Shostak, and M. Pease, "The Byzantine Generals Problem,"
ACM Transactions on Programming Languages and Systems (TOPLAS), vol. 4,
pp. 382-401, 1982.
[73] S. Alag, A. M. Agogino, and M. Morjaria, "A Methodology for Intelligent
Sensor Measurement, Validation, Fusion, and Fault Detection for Equipment
Monitoring and Diagnostics," Artificial Intelligence for Engineering Design,
Analysis and Manufacturing, vol. 15, pp. 307-320, 2001.
[74] H. F. Durrant-Whyte, B. Y. S. Rao, and H. Hu, "Toward a Fully Decentralized
Architecture for Multi-Sensor Data Fusion," in Proceedings of the Proceedings
of 1990 IEEE International Conference of Robotics and Automation, Cincinnati,
OH, 1990.
[75] S. Alag, K. Goebel, and A. M. Agogino, "A Methodology for Intelligent Sensor
Validation and Fusion Used in Tracking and Avoidance of Objects for
Automated Vehicles," in Proceedings of the American Control Conference,
Seattle, WA, 1995, pp. 3647-3653.
[76] K. Bernardin, K. Ogawara, K. Ikeuchi, and R. Dillmann, "A Sensor Fusion
Approach for Recognizing Continuous Human Grasping Sequences Using
Hidden Markov Models," IEEE Transactions on Robotics, vol. 21, pp. 47-57,
2005.
217

[77] Y. Zhou and H. Leung, "A Maximum Likelihood Approach for Multisensor
Data Fusion Applications," in Proceedings of the SPIE on Sensor Fusion:
Architectures, Algorithms, and Applications II, Orlando, FL, 1998, pp. 186-194.
[78] H. Wu, M. Siegel, R. Stiefelhagen, and J. Yang, "Sensor Fusion Using
Dempster-Shafer Theory," in Proceedings of the 19th IEEE Instrumentation and
Measurement Technology Conference, Anchorage, AK, 2002, pp. 7-12.
[79] K. Goebel and A. M. Agogino, "An Architecture for Fuzzy Sensor Validation
and Fusion for Vehicle Following in Automated Highways," in Proceedings of
the 29th International Symposium on Automotive Technology and Automation
(ISATA), Florence, Italy, 1996, pp. 202-209.
[80] K. Goebel and A. Agogino, "Fuzzy Sensor Fusion for Gas Turbine Power
Plants," in Proceedings of the SPIE, Sensor Fusion: Architecture, Algorithms,
and Applications III, Orlando, Florida, 1999, pp. 52-61.
[81] H. Durrant-Whyte and M. Stevens, "Data Fusion in Decentralised Sensing
Networks," in Proceedings of the 4th International Conference on Information
Fusion, Montreal, Canada, 2001, pp. 302-307.
[82] W. Yuan, S. V. Krishnamurthy, and S. K. Tripathi, "Synchronization of Multiple
Levels of Data Fusion in Wireless Sensor Networks," in Proceedings of the
IEEE Global Telecommunications Conference, 2003, pp. 221-225.
[83] R. Kumar, M. Wolenetz, B. Agarwalla, J. Shin, P. Hutto, A. Paul, and U.
Ramachandran, "DFuse: A Framework for Distributed Data Fusion," in
218

Proceedings of the 1st International Conference of Embedded Networked Sensor
Systems, Los Angeles, CA, 2003, pp. 114-125.
[84] R. Jiang and B. Chen, "Fusion of Censored Decisions in Wireless Sensor
Networks," IEEE Transactions on Wireless Communications, vol. 4, pp. 2668-
2673, 2005.
[85] L. A. Zadeh, "Fuzzy Sets," Information and Control, vol. 8, pp. 338-353, 1965.
[86] B. Kosko, Fuzzy Engineering. Upper Saddle River, NJ: Prentice-Hall, 1997.
[87] G. B. Dantzig, Linear Programming and Extenstions, 3rd ed. Princeton, NJ:
Princeton University Press, 1968.
[88] R. J. Vanderbei, Linear Programming: Foundations and Extensions, 2nd ed.
Norwell, MA: Kluwer Academic Publishers, 2001.
[89] P. Prajitno and N. Mort, "A Fuzzy Model-based Multi-Sensor Data Fusion
System," Proceedings of SPIE on Sensor Fusion: Architectures, Algorithms, and
Applications, vol. 4385, pp. 301-312, 2001.
[90] M. Abdelrahman, P. Kandasamy, and J. Frolik, " A methodology for the fusion
of redundant sensors," in Proceedings of the American Control Conference,
Chicago, Illinois, 2000, pp. 2922-2926.
[91] J. Harris, An Introduction to Fuzzy Logic Applications. Norwell, MA: Kluwer
Academic Publishers, 2000.
219

[92] P. S. Khedkar and S. Keshav, "Fuzzy Prediction of Timeseries," in Proceedings
of the IEEE International Conference on Fuzzy Systems, San Diego, California,
1992, pp. 281-288.
[93] A. D. Marbini and L. E. Sacks, "Adaptive Sampling Mechanisms in Sensor
Networks," in Proceedings of the London Communications Symposium, London,
UK, 2003.
[94] A. Jain and E. Y. Chang, "Adaptive Sampling for Sensor Networks," in
Proceedings of the First Workshop on Data Management for Sensor Networks
(DMSN 2004), Toronto, Canada, 2004, pp. 10-16.
[95] C. DiLouie, "Lighting and Productivity: Missing Link Found?," Architectural
Lighting Magazine, vol. Sep/Oct, 2003. [Online]. Available:
http://www.archlighting.com/industrynews.asp?articleID=453031&sectionID=0.
[Accessed: October 1, 2007].
[96] T. Moore, D. J. Carter, and A. Slater, "A Qualitative Study of Occupant
Controlled Office Lighting," Lighting Research and Technology, vol. 35, pp.
297-314, 2003.
[97] C. O'Rourke, "Dimming Electronic Ballasts," in Specifier Reports, vol. 7:
National Lighting Product Information Program, 1999.
[98] RADIANCE Synthetic Imaging System. [Computer software]. Berkeley, CA:
Lawrence Berkeley National Laboratory, 2006. Retrieved March 10, 2008.
Available from http://radsite.lbl.gov/radiance/index.html.
220

[99] SPOT: Sensor Placement + Optimization Tool. [Computer Software]. Boulder,
CO: Architectural Energy Corporation, 2006. Retrieved March 5, 2008.
Available from http://www.archenergy.com/SPOT.
[100] Y. Akashi and J. Neches, "Detectability and Acceptability of Illuminance
Reduction for Load Shedding," Journal of the Illuminating Engineering Society,
vol. 33, pp. 3-13, 2004.
[101] Hamamatsu Photonics, "Si Photodiode S7686," [Online]. Available:
http://sales.hamamatsu.com/en/products/solid-state-division/si-photodiodeseries/si-photodiode/part-s7686.php.
[Accessed: Feburary 8, 2005].
[102] Advance Transformer Technical Staff, IZT-2S32-SC Electrical Specifications:
Advance Transformer, 2003.
[103] D. Loe and P. Davidson, "A Holistic Approach to Lighting Design," 1997.
[Online].
Available:
http://www.iaeel.org/IAEEL/NEWSL/1997/tva1997/DesAppl_b_2_97.html.
[Accessed: July 2, 2008].
[104] C. DiLouie, "Good Controls Design Key to Saving Energy with Daylighting,"
2005. [Online]. Available:
http://www.aboutlightingcontrols.org/education/papers/daylighting.shtml.
[Accessed: July 2, 2008].
221

[105] J. Nielsen and R. Molich, "Heuristic Evaluation of User Interfaces," in
Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems: Empowering People, Seattle, WA, 1990, pp. 249-256.
[106] J. Nielsen, "Enhancing the Explanatory Power of Usability Heuristics," in
Proceedings of the SIGCHI Conference on Human Factors in Computing
Systems: Celebrating Interdependence, Boston, MA, 1994, pp. 152-158.
[107] T. Mandel, The Elements of User Interface Design. New York, NY: John Wiley
& Sons, Inc., 1997.
[108] J. T. Bonnell, "Green Lighting: Wireless Lighting Systems Integration for
Significant Energy Savings," M.S. thesis, University of California, Berkeley,
2008.
[109] R. Embrechts and C. V. Bellegem, "Increased Energy Savings by Individual
Light Control," in Proceedings of the Right Light 4, Copenhagen, 1997, pp. 179-
182.
[110] The IESNA Lighting Handbook, Reference & Application, 9th ed. New York,
NY: The Illuminating Engineering Society of North America, 2000.
[111] F32 T8 TL735 ALTO Production Family Description: Philips, 2008.
[112] "GE Ecolux® Starcoat® T8 " 2008. [Online]. Available:
http://genet.gelighting.com/LightProducts/Dispatcher?REQUEST=COMMERCI
ALSPECPAGE&PRODUCTCODE=26667. [Accessed: September 11, 2008].
222

[113] Advance Transformer Technical Staff, REL-4P32-LW-SC Electrical
Specifications: Advance Transformer, 2004.
[114] J. Luna-Camara, "Electric Power Monthly," Energy Information Administration,
Washington, DC DOE/EIA-0226, July 2008.
[115] M. Bromann, The Design and Layout of Fire Sprinler Systems, 2 ed. Lancaster,
PA: CRC Press, 2001.
[116] R. Burke, Fire Protection: Systems and Response, 1 ed. Boca Raton, FL: CRC
Press, 2007.
[117] M. Delwiche, R. Coates, R. Evans, and L. Oki, "Progress Report for Slosson
Endowment: Precision Irrigation in Landscapes bu Wireless Network,"
University of California, Davis, Davis, CA 2007.
[118] J. McCulloch, P. McCarthy, S. M. Guru, W. Peng, D. Hugo, and A. Terhorst,
"Wireless Sensor Network Deployment for Water Use Efficiency in Irrigation,"
in Proceedings of the Workshop on Real-World Wireless Sensor Networks,
Glasgow, Scotland, 2008, pp. 46-50.
[119] Advance Transformer Technical Staff, IZT-4S32 Electrical Specifications:
Advance Transformer, 2003.
223

Appendix
Appendix A
Electronic Circuit Schematics
A.1 Circuit Schematics of Photosensor Board for MICA and MICA2
224

A.2 Circuit Schematics of Photosensor Board for Tmote Sky
225

A.3 Circuit Schematics of the First Generation Wireless Actuation Module
226

A.4 Circuit Schematics of the Second Generation Wireless Actuation Module
227

A.5 Circuit Schematics of the Third Generation Wireless Actuation Module
228

Appendix B
Human Subject Test
B.1 Complete Sensor Placement Test Data
(a)
Sensor placement of subject No. 1
(b)
(a)
Sensor placement of subject No. 2
(b)
229

(a)
Sensor placement of subject No. 3
(b)
(a)
Sensor placement of subject No. 4
(b)
(a)
Sensor placement of subject No. 5
230
(b)

(a)
Sensor placement of subject No. 6
(b)
(a)
Sensor placement of subject No. 7
(b)
(a)
Sensor placement of subject No. 8
231
(b)

(a)
Sensor placement of subject No. 9
(b)
(a)
Sensor placement of subject No. 10
(b)
232

B.2 Comparison of System Performance and User Interface Test Procedure
Testing Procedures for Participants First Exposed to the Research System
(1) Participants start with the initial desktop illuminance of 500 lux under the
research system, use the GUI controller to set the lighting to a comfortable level,
and begin working on the task of their choosing.
(2) At the 5 th , 10 th and 15 th minute the lights are dimmed or brightened by the
investigator. This is to encourage participant override to return the lights to a
comfortable level.
(3) At the 20 th minute the investigator will stop the first set of the experiment and
ask the participant to take a 10-minute rest outside the office. Meanwhile, the
investigator will resume the remaining procedure of the sensor placement
experiment and gather photosensor and light meter measurement as described in
previous section. Before the participant enters the office again for the second set
of 20-minute experiment, the investigator switches the light to commercial
photosensor-dimming system and set the initial light level at 500 lux on the
desktop. Participants will be asked to use the remote controller to override the
initial condition to a comfortable level after settling for the remaining
experiment.
(4) At the 5 th , 10 th and 15 th minute of the second set experiment the lights are again
dimmed or brightened by the investigator using a second remote controller
without notifying the participants.
233

(5) At the 20 th minute the test ends, the participant is debriefed and interviewed, and
their questionnaire is collected.
Testing Procedures for Participants First Exposed to the Commercial System
(1) Participants start with the commercial system with the initial desktop illuminance
of 500 lux, use the remote controller to override the initial condition to a
comfortable level, and begin working on the task of their choosing.
(2) At the 5 th , 10 th and 15 th minute the lights are again dimmed or brightened by the
investigator using a second remote control, without notifying the participants.
(3) At the 20 th minute the investigator will stop the first set of the experiment and
ask the participant to take a 10-minute rest outside the office. Meanwhile, the
investigator will switch the lighting to the mote-based research system, and set
the initial desktop illuminance to 500 lux. Participants will be asked to use the
GUI to override the initial condition to a comfortable level after settling for the
remaining experiment.
(4) At the 5 th , 10 th and 15 th minute of the second set experiment the lights are
dimmed or brightened by the investigator with the GUI at the investigator station.
(5) At the 20th minute the test ends, the participant is debriefed and interviewed, and
their questionnaire is collected.
(6) After the participant leave, the investigator will resume the remaining procedure
of the sensor placement experiment and gather photosensor and light meter
measurement as described in previous section.
234

B.3 Summary of the Responses from the Questionnaires
Sensor Placement
I'd like to be able to choose sensor
locations if working in an office
with automated lighting
Although I made my best effort, I
am still uncomfortable with the
sensor arrangement
Although I made my best effort, to
place the sensors, I couldn't avoid
accidentally toughing or shading
them
Response
Totally agree 7
Partially agree 2
Neither agree nor disagree 1
Partially disagree 0
Totally disagree 0
Totally agree 0
Partially agree 3
Neither agree nor disagree 3
Partially disagree 2
Totally disagree 2
Totally agree 0
Partially agree 4
Neither agree nor disagree 1
Partially disagree 2
Totally disagree 3
235

Comparison of System Performance and User Interface
It was easy to learn how
to use the overriding
controller
I was easily able to set
my desired illuminance
with the overriding
controller
I felt comfortable
working under this
lighting system
The ability of the
system to maintain my
chosen lighting was:
Watt Stopper
commercial system
with handheld remote
overriding controller
Mote-based research
system with GUI
overriding controller
Totally agree 9 9
Partially agree 1 1
Neither agree nor disagree 0 0
Partially disagree 0 0
Totally disagree 0 0
Totally agree 4 8
Partially agree 4 2
Neither agree nor disagree 0 0
Partially disagree 1 0
Totally disagree 1 0
Totally agree 5 8
Partially agree 4 2
Neither agree nor disagree 0 0
Partially disagree 0 0
Totally disagree 1 0
1 (poor) 0 0
2 0 0
3 0 0
4 0 0
5 3 2
6 5 4
7 (excellent) 1 4
Comparing the two
overriding controllers I
prefer:
Handheld remote
overriding controller
GUI controller
No Response
7 2 1
Other than different types of override, the two systems
performed differently
Yes No No Response
6 3 1
236

B.4 Human Subject Test Approval Letter
237

238

B.5 Human Subject Test Protocol Narrative
239

240

241

242

243

244

245

246

247

B.6 Human Subject Test Consent Form
248

249

250

B.7 Human Subject Test Questionnaire
251

252

253

254

B.8 Human Subject Test Interview Script
The following statements will be reminded to the interviewee, “You have the right to
choose not to respond to any of my questions. Please do feel free to refuse to answer.”
1. How do you like the system you were tested?
2. What are the pros and cons you felt or observed about the system during the
test?
3. If you were the developer of the system, how would you like to improve the
system?
4. Was there anything that made you feel disturbing or uncomfortable during the
test?
5. When using the GUI controller, did you use the slide bar or the step up/down
buttons? Which one do you like better?
6. How did you return the light back? According to your memory or chose another
comfortable setting?
If the subjected is observed to have some unanticipated behaviors, questions regarding
those behaviors will be asked. The followings are questions in response to some of the
possible scenarios:
7. (Suppose the subject stood up and stretched during the test.)
Is standing up and stretching your usual way to relax during work? If not, was it
the lighting or the environment that made you feel tired easily?
8. (Suppose the subject tried to override the lighting during the test.)
Was the illuminance provided by the automation system too much or not enough
for you to work with?
9. (Suppose the subject frequently looked up and check the light fixture during the
test.)
Did you feel the light changing all the time during the test? If so, was it very
disturbing because it changed too frequently or too fast?
10. (Suppose the subject seemed to be in a trance most of the time during the test.)
Was it because of the light or the environment that prevented you from
concentrating on your tasks?
11. (Suppose the subject rubbed his eyes frequently during the test.)
Is rubbing your eyes one of your bad habits? If not, did your eyes feel dry or
uncomfortable because of the light? Was it due to the light changing too
frequently or because the illuminance was just not what you used to work with?
255

Appendix C
Cost Analysis
C.1 Average Illuminance Calculation Sheet
GENERAL INFORMATION
Average maintained illuminance for design: 500 lux Lamp data:
Luminaire data:
Type and color: T8 3500K
Manufacturer: Lithonia Lighting Number of luminaire: 4
Catalog number: 2M432A12 Total lumens per luminaire: 10,640
SECTION OF COEFFICIENT OF UTILIZATION
Step 1: Fill in sketch at right.
Step 2: Determine Cavity Ratios
Room Cavity Ratio, RCR = 1.375
Ceiling Cavity Ratio, CCR = 1.375
Floor Cavity Ratio, FCR = 1.375
Step 3: Obtain Effective Ceiling Cavity Reflectance ( CC ). CC = 70%
Step 4: Obtain Effective Floor Cavity Reflectance ( FC ). FC = 19%
Step 5: Obtain Coefficient of Utilization (CU) from Manufacturer’s Data. CU = 0.66
SELECTION OF LIGHT LOSS FACTORS
Unrecoverable
Recoverable
Luminaire ambient temperature: 1.0 Room Surface dirt depreciation (RSDD): 1.0
Voltage to luminaire: 1.0 Lamp lumen depreciation (LLD): 1.0
Ballast factor: 0.88 Lamp burnouts factor (LBO): 1.0
256

Luminaire surface depreciation: 1.0 Luminaire dirt depreciation (LDD): 0.85
Total light loss factor, LLF (product of individual factors above): 0.748
CALCULATIONS
(Average Maintained Illumination Level)
( Maintained Illuminance) Area of Room
Number of Luminaires =
( Lumens per Luminaire) ( CU) ( LLF)
( )
( 500) ( 15 10)
=
= 14.27 15
( 10,640) ( 0.66) ( 0.748) 257

C.2 Example of System Payback Period Calculation
If the 15-person 150m 2 office considered in Chapter 9.5.1 is retrofitted with the
research system, the actuation decisions will change according to the amount of
available daylight, occupancy status, and present occupants’ preferences. The annual
energy consumption is approximated using the following set of assumptions:
a. The average preferred task illuminance is 400 lux (4/5 of the maximum
illuminance) 80% of the time, and only 20% of the time do the occupants
require maximum illuminance for special tasks.
b. The office is roughly divided into three zones according to daylight availability.
The zones are parallel to the south-facing windows as typical practice of
daylight harvesting. 1/3 of the office area (designated as zone 1) receives
significant daylight throughout the day; 1/3 of the office area (designated as
zone 2) is located toward the building core with respect to zone 1, and receives
less daylight than zone 1; there is very low daylight penetration in the remaining
1/3 of the office area (designated as zone 3).
c. Each daylight zone contains five luminaires. The average daily available
daylight in each zone is estimated as in Table C-1.
d. The office is half occupied during lunch hour (12-1pm) and off hours (7-8am
and 6-9pm) and is fully occupied during regular hours.
258

Table C-1 Daylight reception duration in each zone.
Available daylight* % Zone 1 Zone 2 Zone 3
100% 8 hours 0 hour 0 hour
80% 1 hour 4 hours 0 hour
60% 1 hour 3 hours 0 hour
40% 1 hour 2 hours 4 hours
20% 1 hour 2 hours 2 hours
0% 2 hours 3 hours 8 hours
* With respect to the maximum possible illuminance provided by the electric light
alone.
Note: Due to the lack of existing lighting statistics of the room considered, the
above numbers were approximated from the simulation of daylight reception in
such an office located in San Francisco using SPOT [99]. The exact number may
vary in practice.
To take the first two assumptions into account simultaneously when estimating
lighting energy consumption, consider the worst-case scenario where the maximum
illuminance requested in assumption (a) occurs during the evening when no daylight is
available. In order to incorporate the assumption (d) with the feature of the system that
optimizes lighting actuation according to occupancy status as discussed in Chapter 6,
the power consumption during lunch and off hours is considered to be half of that
during regular hours. The electric light, and hence the ballast power, needed for
compensating insufficient daylight in each zone can thus be summarized in Table C-2.
The energy consumption of the system is estimated at 2209.35 kWh as the calculation
shows in Figure C-1. The annual consumption of the retrofitted office is only about
41% of the one with traditional wall switch control. In other words, the energy savings
generated by the intelligent lighting system is 59%.
259

Table C-2 Ballast power drawing duration in each zone
Ballast Power % (W) Zone 1 Zone 2 Zone 3
0% (0W) 8.5 hours 3.5 hours 0 hour
20% (~39W) 1 hour 3 hours 0 hour
40% (~59W) 0.5 hour 2 hours 3.5 hours
60% (~77W) 0.5 hour 1.5 hours 2 hours
80% (~97W) 0 hour 0.5 hour 5 hours
100% (~116W) 1 hours 1 hours 1 hours
Note: The ballast power is only an estimation as no official information is
provided by any manufacturer. The minimum and maximum input power of the
ballast used in this estimation is from the spec sheet cited in [119].
Figure C-1 Calculation of annual energy consumption.
260

At the rate of 9.62 cents per kWh as used in the analysis in Chapter 9.5.1, the
annual electricity cost for the intelligent control is $213. Compared to the original nondimmable
control with $515 annual cost, use of the research system in a 15-occupant
office generates $302 savings. The payback period for implementing the intelligent
lighting system in this particular office is six years.
The energy savings could be much greater, since worst-case and conservative
assumptions were used in the calculation. Recall that all the requests for the maximum
possible illuminance were defined to occur during the night. Moreover, it was assumed
that the office was fully occupied during regular hours; more energy could be saved as
the lighting for absent occupants would be discounted by the system. It should also be
noted that these calculations and estimates do not take into consideration the savings
due to the implementation of other lighting energy management strategies that could be
integrated into the system, such as load shedding.
261