1:114. .. v
2.3...“ 1L
‘1

 

8...
. . 1
... n. I.

. Eyes! I. it.
zwfiohfiycflhn...

...
.61}

 

 

 

 

.153...“
I‘ll

 

 

_ .fiu. :.,.f.).~. LII .u. 3.1.

 

, LIBRARY
3 Michigan State
Ul liVBfSity

U0)

H

 

 

 

This is to certify that the
dissertation entitled

NETWORK—INTEGRATED SENSING AND ENERGY-AWARE
PROTOCOLS IN WIRELESS BODY AREA NETWORKS

presented by

Muhannad Quwaider

has been accepted towards fulfillment
of the requirements for the

Doctoral degree in Electrical Engineering

 

 

9.211% ax. WW

 

Major Professor’s Si nature

05/06 20/0

 

Date

MSU is an Affirmative Action/Equal Opportunity Employer

 

 

A q.--.-I“-'-.-.-.-.-.---.-.-'

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 KzlProleocaPresIClRClDateDueJndd

 

 

NETWORK-INTEGRATED SENSING AND ENERGY-AWARE
PROTOCOLS IN WIRELESS BODY AREA NETWORKS

by

Muhannad Quwaider

A DISSERTATION
Submitted to
Michigan State University
In partial fulfillment of the requirements
For the degree of
DOCTOR OF PHILOSOPHY

Electrical Engineering

2010

- ABSTRACT

NETWORK-INTEGRATED SENSING AND ENERGY-AWARE PROTOCOLS IN
WIRELESS BODY AREA NETWORKS

By

Muhannad Quwaider

The objective of this thesis is to develop an end-to-end framework for network
integrated sensing and energy-aware protocols for supporting applications in
resource-constrained Wireless Body Area Networks (W BAN). A large number of
existing WBAN applications involving physical activity monitoring and body posture
detection use multi-axis accelerometry as the primary sensing modality. While the
accelerometer-based approaches work well for identifying high-activity postures such as
walking and running, they do not work well when it is necessary to differentiate between
low-activity postures such as standing, sitting, lying down, and sometimes with finer
granularity such as sitting upright or reclining. The key contribution of the first part of
this thesis is to develop a novel network-integrated sensing modality, inter-sensor relative
proximity, which is inferred from the measured Received Signal Strength Indicator
(RSSI) of the Radio Frequency (RF) signal between each pair of WBAN sensors. The
concept of RSSI-based proximity is experimentally developed and then integrated within
a Hidden Markov Model (HMM)-based stochastic processing framework for accurately
identifying human body postures in a subj ect-independent manner.

In the second part of the thesis, the issue of energy-aware on-body communication is
addressed by developing a human body posture-aware transmission power control

framework. A closed loop link power assignment framework has been developed in

which the RF power on an on-body network link is dynamically adjusted depending on
the instantaneous postural orientation of a subject individual. It was demonstrated that
such posture-aware mechanisms can outperform the traditional power control algorithms
by leveraging on-body RF attenuation information which heavily depends on postural
configurations.

In the third part of this thesis, an on-body Delay Tolerant Network (DTN) routing
framework has been developed. Ultra-short transmission range is a common constraint
for low-power RF transceivers used for embedded applications with limited energy and
small form-factors. For such ultra-short transmission range, postural body movements
can make the WBANs to be highly prone to topological partitioning, resulting in a body
area Delay Tolerant Network (DTN). Such topological partitioning can often get
aggravated by the unpredictable on-body RF attenuation. The objective of this part of the
thesis was to develop on-body store-and—forward packet routing algorithms, along with
an analytical framework for modeling routing delay in the presence of network
partitioning. The goal is to minimize end-to-end packet delay while minimizing the

end-to-end hop-count, so that the transmission energy drainage is minimized.

Ckmynghtby
Muhannad Quwaider

2010

 

To My Family
ForAll Their

Love and Support

Acknowledgements

I owe my deepest gratitude to my advisor Dr. Subir Biswas for guiding and
supporting me throughout the course of my doctoral study. His ideas, expertise and sound
advice have been a major influence on this thesis. I have learned a great deal under his
able mentorship during this research work. I would also like to thank the other members
of my committee namely Dr. Michael Shanblatt, Dr. Erik Goodman and Dr. Guoliang
Xing for their helpful suggestions, support and advice.

I am greatly indebted to the ECE department and the graduate school for their
support and encouragement in my endeavor. Also, this research study required long and
extensive experiments and simulations and would not have been possible without the
availability of the resources in the NeEWs laboratory and HPCC.

Very special thanks go to my colleagues and fi'iends in the NeEWS laboratory —
Jayanthi Rao, Anthony Plummer, Mahmoud Taghizade, Tao Wu, Fan Yu, Debasmit
Banerjee, Ali Aqel and Qiong Huo with whom I have worked closely during the last few
years and enjoyed sharing research ideas and accomplishments.

Lastly, and most importantly, I thank my mother, my father, my brothers, my wife,
Abeer, and my kids, Bishr, Taqwa and Abdullah, for their understanding, motivation and

support throughout this research work.

TABLE OF CONTENTS

 

 

 

 

 

 

LIST OF TABLES XI
LIST OF FIGURES XII
LIST OF ABBREVIATIONS XVII
CHAPTER 1. BACKGROUND AND THESIS CONTRIBUTION 1
1.1 WIRELESS BODY AREA NETWORKS WBANS ......................................................... I
1.2 APPLICATIONS OF BODY Area NETWORKS .............................................................. 1
1.2.1 Healthcare Applications ................................................................................. 2
1.2.2 Activity Monitoring Applications .................................................................... 4
1.2.3 Military Applications ...................................................................................... 4
1.3 CHALLENGES, ISSUES AND THESIS CONTRIBUTION ................................................. 6
1.3.1 Limited Sensing Modalities ............................................................................ 7
1.3.1.1 Proposed Inter-sensor Relative Proximity .................................................. 7
1.3.2 Energy Management ....................................................................................... 8
1.3.2.1 Transmission Power Control ....................................................................... 8
1.3.2.2 Data Routing with Ultra-low Radio Transmission Range ........................ 10
CHAPTER 2. WBAN LABORATORY PROTOTYPE SYSTEM 12
2.1 WEARABLE SENSOR NODES ................................................................................. 12
2.2 SENSOR NODE HARDWARE ................................................................................... 14
2.3 MICA2DOT RADIO ............................................................................................. 15
2.4 SOFTWARE PLATFORM .......................................................................................... 16
2.5 TRANSMISSION POWER ASSIGNMENT ................................................................... 17
2.6 PROTOTYPE SYSTEM COMPONENTS ...................................................................... 18
CHAPTER 3. NETWORK-INTEGRATED PROXIMITY SENSING FOR
PHYSICAL POSTURE DETECTION 19
3.1 INTRODUCTION AND RELATED WORK ................................................................... 19
3.2 UNI-MODAL ACCELEROMETRY ............................................................................. 19
3.3 INTER-SENSOR RELATIVE PROXIMITY ................................................................... 22
3 .4 TARGET SENSOR MODALITIES .............................................................................. 23
3 .5 COMPUTATION MODES .......................................................................................... 24
3 .6 SENSOR PLACEMENT AND MODALITY USAGE ....................................................... 25
3.7 POSTURE MODELING AND GENERATION ............................................................... 26
3 .8 THRESHOLD BASED IDENTIFICATION ..................................................................... 29
3 .9 CAPTURING STATIONARY BEHAVIOR USING HIDDEN MARKOV MODEL ................ 35
3.9.1 HMM Mapping ............................................................................................. 35
3. 9.2 Posture Detection using HMM ..................................................................... 40
3.10 EXPERIMENTAL RESULTS ...................................................................................... 41
3.10.1 Manual Calibration ...................................................................................... 41

vii

3.10.2 Automatic Observation Calibration ............................................................. 44
3.11 SUMMARY CONCLUSIONS ..................................................................................... 46
3.12 APPENDIX: ITERATIVE HMM WITH AUTOMATIC OBSERVATION CALIBRATION ...... 47

CHAPTER 4. TRANSMISSION POWER ASSIGNMENT WITH POSTURAL

 

POSITION INFERENCE 50
4.1 RELATED WORK ................................................................................................... 50
4.2 DYNAMIC TRANSMISSION POWER ASSIGNMENTS ................................................. 53
4.3 CHARACTERIZATION OF ON-BODY RF LINKS ....................................................... 53

4.3. 1 Experimental Settings ................................................................................... 53
4.3.2 Target link and Posture Sequence ................................................................. 54
4.3.3 Transmission Power Levels .......................................................................... 55
4.3.4 Multi-scale Variations of Link Quality ......................................................... 56
4.3.5 Impacts of On-body Parameters on Link Quality ......................................... 60
4.3.6 Validating Linearity of RSSI for Other Pastures ................................. ' ......... 62
4.4 FIXED POWER ASSIGNMENTS ................................................................................ 63
4.4.1 Relevant Performance Metrics ..................................................................... 63
4.4.2 Fixed Power Assignment Results .................................................................. 64
4.5 DYNAMIC ON-BODY POWER ASSIGNMENTS ......................................................... 67
4.5.1 Closed Loop Control ..................................................................................... 67
4.5.2 Choice of RSSI Thresholds ........................................................................... 69
4.5.3 Target Power Computation by Linear Search .............................................. 70
4.5.4 Computation by Binary Search ..................................................................... 71
4.5.5 Control with Dynamic Postural Position Inference ...................................... 72
4.6 EXPERIMENTAL PERFORMANCE ............................................................................ 75
4.6.1 Medium Access Control ................................................................................ 75
4.6.2 Results and Interpretation ................................... _ ......................................... 7 7
4.6.2.1 Transmit Power and RSSI Dynamics ........................................................ 79
4.6.2.2 Packet Delivery Ratio (PDR) .................................................................... 80
4.6.2.3 Energy Per Packet (EPP) ........................................................................... 82
4.6.2.4 Energy Overhead of Closed Loop Control ................................................ 83
4.6.2.5 Results for Different Subject Individuals .................................................. 84
4.6.2.6 Impacts of RSSI Thresholds ...................................................................... 86
4.6.2.7 Results for Different Postures ................................................................... 86

4.7 SUMMARYAND CONCLUSIONS .............................................................................. 91

CHAPTER 5. POSTURE-PREDICTIVE POWER CONTROL USING

 

LINEAR-QUADRATIC GAUSSIAN CONTROL 93
5.1 LQGI FRAMEWORK .............................................................................................. 94
5.1.1 Problem Formulation ................................................................................... 94
5.1.2 Quantized LQGI ........................................................................................... 97

5.2 PERFORMANCE RESULTS ....................................................................................... 98
5.3 RELEVANT PERFORMANCE METRICS .................................................................... 99
5.4 RESULTS AND INTERPRETATION .......................................................................... 100

5. 4. I Transmit Power and RSSI Dynamics .......................................................... 100
5.4.2 RSSI Target Error (RTE) ............................................................................. 103

viii

5.4.3 Energy Per Packet (EPP) ............................................. I ............................. 104
5.4.4 Energy Overhead of Closed-Loop Control ................................................. 105
5.5 SUMMARY AND CONCLUSIONS ............................................................................ 1 06

CHAPTER 6. DTN ROUTING IN BODY SENSOR NETWORKS WITH

 

DYNAMIC POSTURAL PARTITIONING 108
6.1 INTRODUCTION ................................................................................................... 108
6.1.1 Short RF Transmission Range .................................................................... 108
6.1.2 Routing with Network Partitioning ............................................................. 109
6.1.3 Objective and Contributions ....................................................................... 109
6.2 RELATED WORK ................................................................................................. 110
6.3 CHARACTERIZATION OF ON-BODY NETWORK TOPOLOGY .................................. 1 14
6.3.1 Experimental Settings ................................................................................. 114
6.3.2 Variations of Topology and Network Partitions ......................................... 115
6.4 STORE-AND-FORWARD ROUTING WITH MULTI-SCALE MOVEMENT LOCALITY 119
6. 4. 1 Postural Link Cost with Multi-Scale Locality ............................................ 119
6.4.2 Probabilistic Routing with Postural Link Costs (PRPLC) ......................... 124
6.4.3 Single-Copy Probabilistic Routing (PROPHET) ........................................ 128
6.4.4 Distance Vector Routing with Postural Link Costs (D VRPLC) .................. 129
6.4.5 On-body Store and Flood Routing (OBSFR) .............................................. 132
6.5 EXPERIMENTAL PERFORMANCE .......................................................................... 136
6.5.1 Polling Based Channel Access for Collision Control ................................. 136
6.5.2 Performance Metrics .................................................................................. 1 3 8
6.5.3 T raflic Generation and Data Collection ..................................................... 138
6. 5. 4 Packet Delay (PD) ...................................................................................... 1 3 9
6.5.5 Packet Hop Count (PHC) ........................................................................... 142
6.5. 6 Packet Delivery Ratio ................................................................................. 143

6. 5. 7 Routing Packets from and to Dilfkrent Body Segments .............................. 145

6. 5. 8 Impacts of Postural Stability ...................................................................... I 4 6

6. 5. 9 Impacts of Intra-posture Movements .......................................................... I48
6.5.10 Impacts of Sensor Placements .................................................................... 149
6.6 OFFLINE SIMULATION WITH EXPERIMENTALLY OBTAINED TOPOLOGY ............... 150
6. 6. 1 Delay Benchmark ....................................................................................... 151
6.6.2 Simulated Performance Results .................................................................. 153
6.7 PRPLC WITHOUT 2-HOP ASSUMPTION ............................................................... 158
6. 7.1 Transitive Update of Link Likelihood Factor ............................................. 158

6. 7.2 Performance with Transitive LLF Update .................................................. 159
6.8 SUMMARY AND CONCLUSIONS ............................................................................ 160
CHAPTER 7. MODELING ON-BODY DTN PACKET ROUTING DELAY ...... 162
7.1 INTRODUCTION ................................................................................................... 162
7.2 TOPOLOGY TRACE COLLECTION FOR OFF-BODY ROUTING SIMULATION ............ 164
7.3 MODELING DTN ROUTING PROTOCOLS ............................................................. 165
7. 3. 1 Opportunistic Routing ................................................................................ I 66
7.3.2 Randomized Routing ................................................................................... 168
7.3.3 Utility-based Routing using Link Locality .................................................. 1 73

ix

7.3.4 Probabilistic Routing with Postural Link Cost (PRPLC) ........................... 1 76

 

 

7.3.5 Distance Vector Routing with Postural Link Costs (D VRPLC) .................. 182

7.4 ROUTING DELAY BENCHMARK ........................................................................... 185
7.5 PERFORMANCE EVALUATION .............................................................................. 187
7.5.1 Performance Metrics .................................................................................. 188
7.5.2 T rafi'ic Generation and Data Collection ..................................................... 188

7. 5. 3 Packet Delay (PD) ...................................................................................... 189
7.5.4 Packet Hop Count (PH C) ........................................................................... 190
7.5.5 Packet Delivery Ratio ................................................................................. 191

7. 5. 6 Routing Packets from and to Dzflerent Body Segments .............................. I93

7. 5. 7 Impacts of Postural Stability ...................................................................... 194

7. 5. 8 Impacts of Sensor Placements .................................................................... 196

7.6 EVALUATION OF NODE CRITTCALTTY ................................................................... 197
7.7 SUMMARY AND CONCLUSIONS ............................................................................ 201
CHAPTER 8. SUMMARY AND FUTURE WORK 202
8.1 CONTRIBUTIONS ................................................................................................. 202
8.2 FUTURE WORK ................................................................................................... 205
REFERENCES 207

List of Tables

Table 2.1: Chipcon CClOOO Output Power Settings and Typical Current [75] ................ 15
Table 3.1: Computation Mode and Their Applications ..................................................... 24
Table 3.2: On-Body Sensor Modality and Placement Summary ....................................... 25
Table 3.3: Threshold group values for context identification ........................................... 29
Table 3.4: Observation sub-vectors and Window boundaries for different observation

granularities ....................................................................................................................... 42
Table 4.1: Performance summary of different power assignments ................................... 79
Table 4.2: Power assignment performance for multiple subjects ...................................... 85

Table 4.3: Indoor and outdoor performance summary of different power assignments 89

Table 6.1: Low power and short range RF transceivers .................................................. 109

xi

Figure 1.1:
Figure 1.2:
Figure 1.3:
Figure 2.1:
Figure 2.2:
Figure 2.3:
Figure 2.4:
Figure 3.1:
Figure 3.2:
Figure 3.3:
Figure 3.4:
Figure 3.5:
Figure 3.6:
Figure 3.7:
Figure 3.8:
Figure 3.9:
Figure 4.1:
Figure 4.2:
Figure 4.3:
Figure 4.4:
Figure 4.5:

Figure 4.6:

List of Figures

Body Area Networks Applications ................................................................... 2
Body Area Sensor Network .............................................................................. 3
Pictorial summary of the investigated issues in this thesis ............................... 6
Wearable sensor network and health monitoring prototype ........................... 12
MICA2DOT sensor node [78] ........................................................................ 13
MICA2DOT mote specifications [78] ............................................................ 14
WBAN System Architecture ........................................................................... 17
Accelerometer data during a controlled posture sequence .............................. 20
Frequency domain view of the acceleration reading ...................................... 21
Posture state transition machine ...................................................................... 27
Posture identification using multi-modal thresholds ...................................... 28
Sensor outputs and actual postures ................................................................. 30
Detection accuracy for multiple human subjects ............................................ 33
Posture identification performance ................................................................. 43
Automatic self-calibration of the B matrix ..................................................... 45
Performance of Baum-Welch iterative algorithm ........................................... 48
On-body sensor positions and the used posture sequence .............................. 55
Link quality variations with varying postural positions .................................. 57
Impacts of transmission power at different postures ...................................... 58
Impacts of transmission power on packet loss ................................................ 59
Variation of sensor orientations and on-body obstructions ............................ 6O
Impacts of sensor orientation and on-body obstructions ................................ 61

xii

Figure 4.7: Orientations of a lower-body link with different postures .............................. 62

Figure 4.8: RSSI characteristics power for a lower-body link .......................................... 62
Figure 4.9: PDR performance with fixed power assignment ............................................ 65
Figure 4.10: EPP performance with fixed power assignment ........................................... 66
Figure 4.11: Closed loop power assignment model .......................................................... 67
Figure 4.12: RSSI threshold range for acceptable packet delivery rate ............................ 69
Figure 4.13: Posture-driven DPPI for dynamic power assignments ................................. 72
Figure 4.14: Multiple control schemes with synchronized postures ................................. 76
Figure 4.15: Power level and RSSI dynamics with posture changes ................................ 77
Figure 4.16: PDRs for DPPI and binary search assignments 81
Figure 4.17: Packet drops during posture transitions ........................................................ 82
Figure 4.18: EPP for DPPI and binary search assignments ............................................... 83
Figure 4.19: Cumulative control energy expenditure over time ........................................ 84
Figure 4.20: Energy expenditure with varying RSSI thresholds ....................................... 85
Figure 4.21: Power levels and RSSI dynamics with posture changes, for different power
assignment strategies ......................................................................................................... 87
Figure 4.22: PDR for DPPI and binary search assignments .............................................. 90
Figure 4.23: EPP for DPPI and binary search assignments ............................................... 91
Figure 5 .1: Block diagram of the augmented system ........................................................ 94
Figure 5.2: Block diagram of quantized LQGI Control .................................................... 98
Figure 5.3: On-body sensor positions and the used posture sequence .............................. 99
Figure 5.4: Power level and RSSI dynamics with posture changes ................................ 101
Figure 5.5: RSSI average error and standard deviation ................................................... 103
Figure 5.6: Average EPP for different dynamic search assignments .............................. 105

xiii

Figure 5.7: Cumulative control energy expenditure ........................................................ 106

Figure 6.1: On-body Mounted Sensors ............................................................................ 114
Figure 6.2: Variation of link connectivity with postural mobility ................................... 116
Figure 6.3: Topology and partition properties with posture changes .............................. 117
Figure 6.4: Distributed of number of neighbors .............................................................. 1 18
Figure 6.5: Evolution of multi-scale locality in terms of LLF and HCQ ........................ 123
Figure 6.6: PRPLC routing with LLF capturing multi-scale connection localities ......... 127
Figure 6.7: DVRPLC routing with link cost capturing multi-scale connection localilti3els
Figure 6.8: OBSFR routing logic for packet delay lower bound ..................................... 134
Figure 6.9: Collision-free MAC access Via polling ......................................................... 137
Figure 6.10: Packet delivery delays with static and adaptivew scenarios ....................... 140
Figure 6.1 1: On-body packet delivery delay for different DTN routing protocols ......... 141
Figure 6.12: Distribution of Packet Hop Count (PHC) ................................................... 142
Figure 6.13: Number of transmissions per delivered packet ........................................... 143
Figure 6.14: Packet delivery performance ....................................................................... 144

Figure 6.15: Delivery delay for packets from thigh, waist and arm to right ankle (i.e. node
6) ...................................................................................................................................... 145

Figure 6.16: Delivery delay for packets to arm, thigh and ankle from the upper left arm (i.e.

node 3 in Figure 6.1) ........................................................................................................ 146
Figure 6.17: Impacts of posture duration on packet delay .............................................. 146
Figure 6.18: Impacts of posture duration on packet delivery ratio .................................. 147
Figure 6.19: Packet delivery delay with intra-posture movement ................................... 148
Figure 6.20: Experiments with different sensor placements ........................................... 149

xiv

Figure 6.21: On-body packet delay for: (a) sensor placement P1, and (b) sensor placement

P2 ..................................................................................................................................... 150
Figure 6.22: Experimental topology export for offline simulation ................................. 151
Figure 6.23: Delivery delay from experiments and simulation ....................................... 154
Figure 6.24: Packet hop counts from experiment and simulation ................................... 155
Figure 6.25: Simulated average packet delivery delay .................................................... 156
Figure 6.26: Simulated average of Packet Hop Count .................................................... 157
Figure 6.27: The impacts of transitive updates on delivery delay ................................... 159
Figure 6.28: The impacts of transitive updates on Packet Hop Counts (PHC) ............... 160
Figure 7 .1: On-body Mounted Sensor ............................................................................. 163
Figure 7.2: Topology export for offline and model performance .................................... 164
Figure 7.3: Example connectivity of an on—body link ..................................................... 165
Figure 7.4: Evolution of multi-scale locality in terms of PLC and HCQ ....... . ................. 179
Figure 7.5: Capturing link connectivity locality in PRPLC and UTILITY age-based
routing .............................................................................................................................. 181
Figure 7.6: On-body packet delivery delay for different DTN routing protocols ........... 189
Figure 7.7: Average Packet Hop Count ........................................................................... 191
Figure 7.8: Packet drops observed for different protocols .............................................. 192
Figure 7.9: Delivery delay for packets from thigh, waist and arm to right ankle ............ 193
Figure 7.10: Delivery delay for packets to arm, thigh and ankle from the upper lefi arm (i.e.
node 3 in Figure 7.1) ........................................................................................................ 194
Figure 7.11: Impacts of posture duration on packet delay .............................................. 194
Figure 7.12: Experiments with different sensor placements ........................................... 195

Figure 7.13: On-body packet delay for: (a) sensor placement P1, and (b) sensor placement
P2. .................................................................................................................................... 196

XV

Figure 7.14: Node criticality in terms of: (a) Packet delay, and (b) Packet delay difference
......................................................................................................................................... 198

Figure 7.15: Two node criticality in terms of: (a) Packet delay, and (b) Packet delay
difference ......................................................................................................................... 200

xvi

WBAN

RSSI

DPPI
DTN
FSK
FT

HMM

TPC
TDOA
CDMA
PDR
EPP
LQGI
LQGIQ

RTE
ST
WASP
PLCF
LLF

HCQ

List of Abbreviations

Wireless Body Area Networks
Received Signal Strength Indicator
Radio Frequency

Dynamic Postural Position Inference
Delay Tolerant Network

Frequency Shift Keying

Fourier Transform

Hidden Markov Model

Window Boundary

Transmission Power Control

Time Delay Of Arrival
Code-Division Multiple-Access
Packet Delivery Ratio

Energy Per Packet

Linear Quadratic Gaussian control with an Integrator

Quantized Linear Quadratic Gaussian control with an
Integrator

Target Error

Standard Deviation

Autonomous Spanning tree Protocol
Postural Link Cost Formulation
Link Likelihood Factor

Historical Connectivity Quality

xvii

PROPHET

PRPLC
DVRPLC
LCF
OBSFR
BSDBR
ER

PD

PHC
UTILITY
OPPT
RAND
DD

SIT

STD

Probabilistic Routing Protocol using History of Encounters
and Transitivity

Probabilistic Routing with Postural Link Costs
Distance Vector Routing with Postural Link Costs
Link Cost Factor

On-body Store and Flood Routing

Backward Search for Delay Benchmark Routing
Epidemic Routing

Packet Delay

Packet Hop Count

Utility-age based protocol

Opportunistic protocol

Randomized protocol

Directed Diffusion

Sit straight

Stand

Sit Reclining

Lying Down

Walk

Run

xviii

Chapter 1. Background and Thesis Contribution

1.1 Wireless Body Area Networks WBANS

Wireless Body Area Networks (WBAN) consists of a set of intercommunicating
sensors, either wearable or implanted, which can monitor vital body movement
parameters and collect different body information [1-5]. These devices, communicating
through wireless technologies, can transmit data from the body to a home base station
fi'om where the data can be forwarded to a hospital, clinic, or a service provider in
real-time. The WBAN technology is still in its primitive stage and is being widely
researched. The technology, once accepted and adopted, is expected to be a breakthrough
invention in many healthcare applications, leading to concepts like telemedicine and

mobile health monitoring.

1.2 Applications of Body Area Networks

Initial applications of WBANS are expected to appear primarily in the healthcare
domain, especially for continuous monitoring and logging of vital parameters of patients
suffering from chronic diseases such as diabetes, asthma and heart attacks, as well as in
elder care monitoring. Other emerging applications of this technology include military,
sports and security. Extending the technology to new areas could also assist
communication by seamless exchanges of information between individuals, or between
individual and machines. Figure 1.1 demonstrates a few examples of WBAN

applications.

 

 

Service Provider

 

 

Data Collection

 

Figure 1.1: Body Area Networks Applications

1.2.1 Healthcare Applications

Wearable systems for continuous health monitoring are a key technology in helping
the transition to more proactive and affordable healthcare. They allow individuals to
closely monitor changes in their vital signs and provide feedback to help maintain an
optimal health status. If integrated into a telemedical system, these systems can alert
medical personnel when life-threatening changes occur. In addition, the wearable systems
can be used for health monitoring of patients in ambulatory settings [6]. For example,
they can be used as a part of a diagnostic procedure, optimal maintenance of a chronic
condition, a supervised recovery from an acute event or surgical procedure, to monitor
adherence to treatment guidelines (e.g., regular cardiovascular exercise), or to monitor
effects of drug therapy.

The multiple WBAN sensor nodes shown in Figure 1.2 are capable of sampling,

processing, and communicating one or more vital signs like heart rate, blood pressure,

oxygen saturation, breathing rate, diabetes, body temperature, ECG and activity, or
environmental parameters like location, temperature, humidity, light, movement,
proximity and direction. Typically, these sensors are implanted or placed strategically on
the human body as tiny patches or hidden in users’ clothes allowing ubiquitous health

monitoring in their native environment for extended periods of time.

 

Modalities:

- Blood pressure
- Heart rate

- Breathing rate -.
_ Diabetes ' ‘ '
- Temperatures“
- Humidity ' '

- ECG

 

- Movement
- Proximity
- Direction

 

 

 

 

 

 

 

Out-of-Body
Server

 

 

 

Figure 1.2: Body Area Sensor Network

The Institute of Electrical and Electronics Engineers (IEEE) approved five ISO/IEEE
11073 standards for automatic acquisition of vital signs by clinical information systems
[7-11]. Those five standards are in the domain of information model, the common
nomenclature, the basic communications profiles, and two defining underlying transport
arrangements. All major manufacturers have prototype devices using these five standards

as point-of-care device communication standards, but most of them are reluctant to make

further investment until user demand recovers. Both USA and UK governments have
specified that ISO/IEEE11073 Shall be used for device communication [7],[8].

While the major parts of these standards deal with information representation and
data formatting, they do not propose any specific routing and power control models. The
role of this thesis is to carry out an early analysis of such broad networkng issues in
order to support the standardized data formats as proposed by the IEEE standards.

1.2.2 Activity Monitoring Applications

Human activity has been studied [12-17] for the purposes of understanding the basic
characteristics of human movement and the relationship of physical activity to chronic
diseases such as cardiovascular disease and cancer, as well as for monitoring elderly
patients. The accurate and detailed measurement of physical activity is therefore a crucial
prerequisite to exploring its association with health and disease. Numerous methods have
been used to measure physical activity in the short and long terms [2],[18-23]. They vary
greatly in their applicability in epidemiological research, intervention studies, clinical
practice, and personal assessment.

Human activity monitoring has also been used for studying obesity among children
of all ages [18],[21]. The purpose of these studies was to record the physical activity
levels of children while they attend preschools, to identify the demographic factors that
might be associated with physical activity among those children, and to determine the
extent to which children’s physical activity varies among preschools.

1.2.3 Military Applications

The physical safety and well being of the soldiers in a battlefield is the highest
priority [24-26] of Incident Commanders. Currently, the ability to track and monitor

soldiers relies on visual and verbal communication which can be somewhat limited in

scenarios where the soldiers are deployed inside buildings and enclosed areas that are out
of visual range of the commanders. Also, the need for stealth can often prevent a battling
soldier from sending verbal clues to a commander about his or her physical well being.
Sensor technologies can remotely provide various data about the soldiers including
physiological monitoring and personal alert safety system functionality.

A body area wireless network with multi-modal sensors can monitor the body
movement and other physiological parameters for statistical identification of a soldier’s
body posture, which can then be indicative of the physical conditions and safety alerts of
the soldier in question. Soldier tele-monitoring using WBAN system would include:

1) The wounded soldier’s on-body sensor network to identify the soldier’s body
posture (e. g. lying down on back would indicate a worrisome situation).

2) Assess the soldier’s context with respect to his or her surroundings by sensing
other fellow soldiers and the athletic [24] assets such as tanks, aircrafis and
medical facilities.

3) Delivering information from items (1) and (2) to remotely situated
commanders and to local fellow military personnel.

4) To take an appropriate rescue, medical or evacuation decision.

Since the majority of information gathering and situational assessment processes are
automated, the rescue, evacuation and medical actions can be taken in a swift and safe

manner due to the stealth nature of the outlined process.

 

 

 

 

 

 

 

 

 
    
    
  
  

 

 

    

 

  
  

        
  
 

 

 

 

 

 

 

 

 

CEnergy'iwaie Wireless Body Area ( Appl‘cam’ns \
mumcanon Networks (WBAN) [Limited Sensor ]
Characterization of - gs? Modalities
On-Body RF Links] i it ’ .
RSSI Readingl " - \ taming
/ Body ;.
Transmission Power Movement-{I Network-integrated]
Control . WBAN: Applications, I, Proximity Sensing
(Chapter 4) . Challenges and Archrtecture '
.4 Overview (Chapter 1 & 2)
Fixed 7' l '
\t
I
[ Linear,

Binary, DPPI]/ \ (Chapter 3)

Ultra-Low Postural
Power Sensor Disconnection

Store and Forwarding Modeling .WBAN
Power Control Routin_ Routing

k (Chapter 5) j K (Chapter 6) (Chapter 7) J

Figure 1.3: Pictorial summary of the investigated issues in this thesis

 

 

 

 
 

 

 

 

 

[

Gaussian Control DTN Routing

i
[[Linear-Quadratic

]\ Single Copy J

Posture-Predictive

   

(LQGI)

 
     

 

 

 

 

 

 

 

 

1.3 Challenges, Issues and Thesis Contribution

Deployment of WBAN systems will need to address a number of technical
challenges that lie ahead [27-32]. These include the need for innovative sensor design,
efficient data processing, energy management, and secure data transfer. The goal of this
thesis is to design and investigate a framework for network-integrated sensing and
energy-aware protocols to address the resource constraints inherent in WBANS.
Available energy, processing, and form factor constraints of a body-mounted or
implanted sensor can severely limit its abilities for sensing, and the power and range of

radio transmission for communication purposes. We investigate sensing, processing, and

protocol solutions to address these system constraints. Figure 1.3 shows a Visual
illustration of the scope of the work presented in this thesis. The following sections
elaborate the topics in Figure 1.3 in more details.

1.3.1 Limited Sensing Modalities

A large number of WBAN applications deal with physical activity monitoring and
body posture detection, especially for continuous monitoring of patients suffering from
chronic disease. The number and the accuracy of the detected postures depend on the
number of available sensing modalities and the processing capability. In a number of
papers in the literature [19],[33],[34], multi-axis accelerometers are the only modality
used for the identification of body postures by analyzing the level of accelerations in
different body segments, which are a direct indication of physical activity. These
mechanisms are shown to work [35-3 7] very well for identifying postures such as
walking, jogging, and sprinting. However, for those applications that require context
identification at finer granularities, it is often necessary to differentiate between
low-activity postures such as sitting, lying-down and standing; sometimes with even finer
granularity such as sitting-upright or sitting-reclined. For these non activity-intensive
postures, the traditional accelerometer-based solutions do not work.

1.3.1.1 Proposed Inter-sensor Relative Proximity

To address the above limitation, new sensing modalities are needed in addition to the
standard accelerometry. But adding more modalities in a WBAN sensor can be difficult
due to the ultra-small form factor, energy and processing constraints, especially for
implanted sensors. In this thesis, we have developed a novel network-integrated sensing
modality, inter-sensor relative proximity, which is inferred from the measured Received

Signal Strength Indicator (RSSI) of the Radio Frequency (RF) signal between each pair

of WBAN sensors. By using this new sensing modality, more non-activity-intensive
postures can be detected, without having to add extra components and the associated
energy overhead.

Chapter 3 introduces this network-integrated proximity sensing. An experimental
system is developed for integrating the proximity data with other traditional modalities,
such as acceleration and orientation. This multi-modal data is then processed using an
HMM framework for accurate posture identification.

1.3.2 Energy Management

Since most wirelessly networked devices are operated fiom battery or from
harvesting energy [38-43], one of the major research challenges is the issue of power
limitations [30],[44-48]. Sometimes it is important to guarantee that the device will work
for certain period (c. g. 5 years) before changing the battery [49],[50]. This is particularly
crucial for implanted sensors for which the battery replacement can be an involved
procedure [51-58]. Therefore, better power management schemes to deal with these
issues must be designed. The following issues are dealt with in this thesis.

1.3.2.1 'II'ansmission Power Control

Data transaction across on-body sensors can be point-to-point or multi-point-to-point
depending on specific applications. While real-time on-body detection of an athlete’s
instantaneous posture [19],[59] would require multi-point-to-point data exchange across
various on-body sensors, applications such as monitoring vital signs of a patient will
require all body-mounted and implanted sensors [60],[61] to transmit data to a sink node,
which in turn will relay the information wirelessly to an off-body server.

For both cases, the optimal transmission power required for an on-body wireless link

between two sensors depends on: 1) the physical distance of the link [62], and 2) its

instantaneous channel condition. The distance can vary with mobility driven by human
postures, and the channel condition can change due to unpredictable RF attenuation [63]
caused by many factors including antenna orientation, clothing, and physical stature of
specific subject individuals. We coined the term postural position which defines the
instantaneous state of an on-body wireless link as a result of its distance and the RF
attenuation. An on-body link’s postural position can be highly dynamic due to the reasons
listed above.

AS a result, static pre-defined transmission power is not able to provide continuous
link connectivity while ensuring minimum required transmission power consumption
[62],[64]. Therefore a dynamic link power assignment mechanism for Optimal energy
management is needed.

1.3.2.1.] Proposed Orr-body Transmission Power Assignments

Three different dynamic on-body link transmission power assignment mechanisms
are proposed, namely, Linear, Binary and Dynamic Postural Position Inference (DPPI).
In these approaches, the receiver/control node computes a new desirable transmission
power level for the transmitter node by incrementing or decrementing the current power
level based on the received RSSI values with respect an RSSI threshold range. No
changes are needed when the RSSI falls within the range. This closed-loop control logic
ensures that, for a given postural position, this process eventually achieves a desirable
transmission power level so that the RSSI at the receiver falls within the preset threshold
range. Chapter 4 provides detailed information of transmission power assignment
mechanisms.

The proposed transmission power control mechanism is then enhanced by modeling

human body movement as a stochastic linear system and a quantized Linear Quadratic

Gaussian control with an Integrator (LQGI). The objective is to develop a model based
power control framework in which RF signal strength is predicted and is regulated at a
reference value to enhance the overall energy performance of an on-body wireless sensor.
The predictive LQGI power control will be studied in Chapter 5.

1.3.2.2 Data Routing with Ultra-low Radio Transmission Range

On-body packet routing is needed in the presence of topological partitioning caused
due to ultra-short wireless range and postural body movements. Short transmission range
is a common constraint for low-power RF transceivers designed for embedded
applications with limited energy [62],[64], ofien supplied by harvested operations [65].
Such Situations are particularly pertinent for implantable body sensors. In literature
[65-68] the ultra-low-range transceivers vary in range between Om to 1m, with
corresponding transmission powers varying between 0.7 SmW to 6mW, which are within
a range that can be handled with common harvesting techniques such as piezo-electric
generation from body movements [65].

Low RF transmission ranges also mean that postural body movements can give rise
to frequent partitioning or disconnection in WBAN topologies, resulting in the need for a
body area Delay Tolerant Network (DTN) [69-74]. Such topological partitioning can
ofien get aggravated by the unpredictable RF attenuation caused due to Signal blockage
by clothing material and body segments. Topological partitioning necessitates
store-and-forward routing protocols with performance goals of low end-to-end packet
delay, packet loss, and transmission energy consumption.

1. 3. 2.2.1 Store-and-forward Packet Routing

In this thesis we have developed on-body store-and-forward packet routing

algorithms in the presence of network partitioning. The objective is to minimize

10

end-to-end packet delay by dynamically choosing routes on which the storage/buffering
delays are low. While ensuring low storage delay, the algorithms also attempt to
minimize the end-to-end hop count so that the transmission energy drainage is
minimized, thus leading to long network operating durations.

First, using a prototype WBAN, we have developed a topology characterization
mechanism in order to demonstrate the network partitioning caused due to human
postural mobility. Second, an on-body packet routing fi'amework is developed using a
stochastic link cost formulation, reflecting the body postural trends. Third, the
performance of the proposed protocols is evaluated experimentally and through
simulation using the prototype WBAN. The performance of the routing protocol is then
compared with a probabilistic DTN routing protocol PROPHET [74], and a specialized
flooding based benchmark algorithm that provides routing delay lower bounds. Chapter 6
provides the details of this proposed routing protocol.

Finally, analytical techniques are developed for modeling end-to-end packet delay
for the proposed DTN routing algorithms. DTN routing delay obtained from the
developed model are then compared with results from on-body experiments from the
prototype WBAN and off-body simulation carried out with network topology traces
obtained from the prototype WBAN. Using the model and the topology trace data, a
detailed analysis is then carried out for identifying non-critical links in order to design a
minimal WBAN topology fiom the routing standpoint. Detailed discussion of modeling
techniques and performances in the presence of postural disconnections will be presented

in Chapter 7.

ll

Chapter 2. WBAN Laboratory Prototype System

A laboratory prototype system was developed for carrying out all experimental work

presented in this thesis.

 

     
 

Diabetes
Direction
Pulse

 

Acceleration
Direction

 

 

 

Out—of-Body smk

Server

 

 

 

Figure 2.1: Wearable sensor network and health monitoring prototype

2.1 Wearable Sensor Nodes

The prototype WBAN was constructed by mounting a set of tiny sensor nodes on
different segments of the body (e.g. thigh, arm, ankle, waist, head, etc.), as shown in
Figure 2.1. These wearable sensor nodes consist of a Mica2Dot MOTE [75],[76] radio
node (running the TinyOS operating system) [77], operating with a 915MHz radio using
a Chipcon SmartRF CClOOO radio chip [75], and its sensor card, MT S510, from
Crossbow Inc. [78]. The MicaZDot nodes run from a 570mAH button cell with a total

node weight of approximately 10 grams. The default CSMA MAC protocol is used with a

12

data rate of 38.4kbps. The sensors are worn with elastic bands in such a manner that a
sensor and its antenna orientation with respect to the mounted body segment do not

change with a subject’s posture changes.

 

Atrnel ATMega128

Top—side .. ..

 

 

 

 

Figure 2.2: MICAZDOT sensor node [78]

The body-worn sensors form a mesh or a star topology depending on point-to-point
or multi-point-to-point data requirements in specific applications. Depending on the
topology, a body-mounted sensor sends data to another body-mounted sensor or to a sink
node. The Sink node collects raw data, and sends processed results or events to an
off-body server using a wireless link. This extemal link is created between the on-body
sink node (or it can be any on—body node) and to an off-body base station contacted using
another Mica2Dot radio node. The base station is connected to a Windows PC (compute

server) through a custom-built serial interface, running the RS232 protocol.

13

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Processor/Radio Board MPR500CA Remarks
Processor Performance
Program Flash Memory 128K bytes
Measurement (Serial) Flash 512K bytes >100,000 Measurements
Configuration EEPROM 4K bytes
Serial Communications UART 0-3V transmission levels
Analog to Digital Converter 10 bit ADC 6 channel, 0-3V input
Other Interfaces D10 9 channels
Current Draw 8 mA Active mode
< 15 PA Sleep mode
Mum-Channel Radio
Center Frequency 868/916 MHz ISM bands
Number of Channels 4/50 Programmable, country specific
Data Rate 38.4 Kbaud Manchester encoded
RF Power -20 to +5 dBm Programmable, typical
Receive Sensitivity -98 dBm Typical, analog RSSI at AD Ch.0
Outdoor Range 500 fi 1/4 Wave dipole, line of sight
Current Draw 27 mA Transmit with maximum power
10 mA Receive
< 1 [LA Sleep
Electromechanical
Battery 3V Coin Cell
External Power 2.7 - 3.3 V Connector provided
User Interface 1 LED User programmable
Size (in) 1.0 x 0.25 Excluding battery pack
(mm) 25 x 6 Excluding battery pack
Weight (oz) 0.1 1 Excluding batteries
(grams) 3 Excluding batteries
Expansion Connector 18 pins All major I/O signals

 

Figure 2.3: MICAZDOT mote specifications [78]

2.2 Sensor Node Hardware

The MICAZDOT mote from Crossbow Technology Inc. (see Figure 2.2) is a third
generation tiny wireless mote module which is typically used for low-power wireless
sensor networks. The low-power processor in the Mica2DOT is an Atrnel ATmega 128L
whose 18-pin expansion connector provides various interfacing options including analog

and digital I/O, 12C, SP1 and UART-based communication. The MTSSIO sensor board

14

 

connected to the 18-pin connector can be used for light, temperature, acceleration and

acoustic sensing. The specification sheet for the MICAZDOT is shown in Figure 2.3 [78].

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Pout PA_POW (hex) Current Consumption
(dBm) 915 MHz (mA)
-20 0x02 8.6
-19 0x02 8.8
-18 0x03 9.0
-17 0x03 9.0
-16 0x04 9.1
-15 0x05 9.3
-14 0x05 9.3
-13 0x06 9.5
-12 0x07 9.7
-11 0x08 9.9
-10 0x09 10.1
-9 0x0b 10.4
-8 0x00 10.6
-7 0x0d 10.8
-6 0x0f 1 1 .1
-5 0x40 13.8
-4 0x50 14.5
-3 0x50 14.5
-2 0x60 15.1
—1 0x70 15.8
0 0x80 16.8
1 0x90 17.2
2 OxbO 18.5
3 0xc0 19.2
4 Oxfi) 21 .3
5 Oxff 25.4

 

 

 

 

 

Table 2.1: Chipcon CC1000 Output Power Settings and Typical Current [75]

2.3 MICAZDOT Radio

The radio used by the MICAZDOT is an ISM band RF transceiver designed for
low-power and low-voltage Wireless applications. It uses Chipcon’s CClOOO radio [78],
which employs Frequency Shift Keying (FSK) with Manchester encoding and an

effective data rate of 38.4kbps. The radio on the MICAZDOT is capable of multiple

15

channel operation within the intended band. It can also be software adjusted for a range
of output power levels. The register in the radio that controls the RF power level is
designated PA_POW at address OXOB, and the values and their corresponding RF outputs
are provided on Table 2.1 [75]. It shows the closest programmable value for output

powers in steps of 1 dBm.

2.4 Software Platform

The sensor node software is responsible for data sampling and acquisition, real-time
processing, and WBAN communication. Software running on the sensor nodes and the
network base station is developed using the nesC language under the TinyOS operating
system [77].

TinyOS is an open source component-based event-driven embedded operating
system mainly targeted at wireless sensor networks. It is designed as a set of cooperating
tasks and processes, and it incorporates rapid innovation as well as operating within the
severe memory constraints inherent in sensor networks.

Programming in TinyOS is performed in nesC, a dialect of the C programming
language optimized for memory-limited sensor networks. NeSC has a C-like syntax, but
supports the TinyOS concurrency model, and mechanisms for structuring, naming, and
linking together software components into networked embedded systems. TinyOS
provides built-in interfaces, modules, components, and sensor-board-specific
configurations, which allow programmers to build programs as a set of modules for
program-Specific tasks. Interface to standard hardware inputs, outputs, and sensors are
provided by the components in TinyOS. The application program is statically linked with

TinyOS, and compiled into a small binary file, using a custom GNU tool chain.

16

 

 

 

/ Energy Management in WBAN \
Tier-4: Modeling On-body DTN Routing (Ch-7) \

Ultra-Low f Postural .% ,. Single Copy
, Power Sensor ' Partitioning ' ' DTN Routing)

 

 

 

 

 

 

 

 

 

 

 

 

 

Tier-3 Packet Routing In WBAN (Ch 6)

----------------------------------------

UIII'a-LOW :§ ~ Postural .
Hag; Power Sensor i; . Partitioning E‘

 

 

 

 

 

 

 

 

 

 

 

Tier-2: Transmission Power Control (Ch-4&5)

 

 

Dynamic Power Linear Quadratic Gaussian

k Assignment with an Integrator (LQCy

f Tier-1: Embedded/Wearable Sensor (Ch-3) \

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

' Hardware and Software Mlti-Modal Sensing
Mica2Dot MOTE Accelerortletw
MTSSIO Sensor Card Orientation
TinyOS Operating System ' Network-mtegrated
k ' Relative Proximity /

 

 

 

 

 

Figure 2.4: WBAN System Architecture

2.5 Transmission Power Assignment

The energy required for packet transmissions and receptions, which has been
addressed by transmission power control techniques, is considered as the primary source
of energy drainage in WBAN. The idea is to use the lowest power level that is adequate
to maintain acceptable packet delivery performance. Using a lower power level achieves
reduced energy consumption, reduced interference and limited heat loss. The 22 different
transmission power levels [62],[75] used in our prototype system are shown in Table 2.1
which are used in dynamic transmission power assignments that we will discuss in

Chapter 4 and Chapter 5.

17

2.6 Prototype System Components
The prototype WBAN described in this chapter is consistently used for the
experimental evaluation of all the protocols developed in this thesis. Figure 2.4

demonstrates a layered representation of the various thesis project components on the

prototype.

18

Chapter 3. Network-integrated Proximity Sensing for Physical

Posture Detection

3.1 Introduction and Related Work

In this chapter we have developed a novel network-integrated sensing modality,
inter-sensor relative proximity, which is inferred from the measured Received Signal
Strength Indicator (RSSI) of the Radio Frequency (RF) signal between each pair of
WBAN sensors. By using this new sensing modality, more non-activity—intensive
postures can be detected, without having to add extra components and the associated
energy overhead. An experimental system is developed for integrating the proximity data
with other traditional modalities such as acceleration and orientation. This multi—modal
data is then processed using an HMM fiamework for accurate posture identification.

In a number of papers in the literature [19],[33],[34], multi-axis accelerometers are
the only modality used for the identification of body postures by analyzing the levels of
acceleration in different body segments, which are a direct indication of physical activity.
These mechanisms are shown to work [35-37],[79] very well for identifying postures
such as walking, jogging, and sprinting. However, for applications that require context
identification at finer granularities, it is often necessary to differentiate between
low-activity postures such as sitting, lying down and standing; sometimes with even finer
granularity such as sitting upright or sitting reclined. For these non-activity-intensive

postures, the traditional accelerometer-based solutions do not work.

3.2 Uni-Modal Accelerometry
This section outlines the posture identification process using a traditional uni-modal

approach [19],[33],[34] using only the acceleration information. Controlled experiments

19

using the prototype system illustrated in Chapter 2 and Figure 2.1 are designed in which
human subjects are given predetermined sequences of postures (fi'om the set SIT, STD,
REC, DWN, WLK, and RUN) to follow, and a three-node wearable sensor network is
used for collecting acceleration data from right thigh, upper right arm, and right ankle.
Postures, identified using context detection algorithms, are then temporally correlated

with the actual posture sequence given to the subjects for evaluating the identification

accuracy.

 

 

800 . :

{Thigh

700‘ 7 i

5004:] =

400- ‘ i i A 7
i ii. ; i‘\Ankle f

300- ‘ ' .. =

 
 

 

 

 

Acc. Readings (mg).

 

 

 

SIT STDWLK’STDSTD WLKSTD SIT SIT REC REC SIT STD DWNSTDRUNSTD STD SIT STD

_ SIT: Sit, STD: Stand, REC: Sit-Reclining,
PM” sequence 0"“ “me (5) WLK: Walk, DWN: Lying Down, RUN: Run

 

200

 

 

 

Figure 3.1: Accelerometer data during a controlled posture sequence

Figure 3.1 shows the accelerometers’ readings in milli-g (1 mg is 9.81 mm/sz) from
all three sensor nodes, while a human subject was following a controlled 20-posture
sequence as shown along the horizontal axis of the figure. Each posture Slot in this
experiment lasted for approximately 20 seconds. A sampling rate of 20 Hz was used for
obtaining reading from the accelerometers. The numbers in the figure correspond to the
averages of the accelerations recorded in both the axes of the sensor used.

The figure shows how the acceleration readings increase for the activity-intensive

postures such as WLK and RUN compared to low-activity postures such as SIT and

20

STAND. In fact the readings for SIT, REC, DWN and STD are almost the same due to
the absence of any major physical activity in these postures.

The frequency domain representation of the accelerometer data collected is presented
in Figure 3.2 for all six postures individually. The graph for WLK, for example, is plotted
by applying the Fourier Transform (FT) to the cumulative acceleration data from all the

WLK states as shown in Figure 3.2. The same applies to the other postures as well.

 

 

FFT Amplitude

 

 

 

0 0.02 0.04 0.06 0.08 0.1
Frequency (Normalized by sampling frequency)

 

 

 

Figure 3 .2: Frequency domain view of the acceleration readings

Observe that while the graphs for WLK and RUN demonstrate a noticeable presence
of frequency components in the range 0 to 0.1, the ones for SIT, REC, DWN and STD
are ahnost flat over the entire frequency spectrum. The difference in the peak values for
WLK and RUN indicates the difference of activity levels in those two postures. These
peak values, coupled with suitably chosen thresholds, can be used for identification and
differentiation between the WLK and RUN postures. The results in Figure 3.1 and Figure
3.2 confirm that while unirnodal accelerometry is capable of identifying WLK and RUN,

it is not sufficient for the low-activity postures.

21

To address this limitation, new sensing modalities are needed in addition to the
standard accelerometry. But adding more modalities in a WBAN sensor can be difficult
due to the ultra-small form factor, energy and processing constraints, especially for

implanted sensors.

3.3 Inter-sensor Relative Proximity

In this chapter, we have developed a novel network-integrated sensing modality,
inter-sensor relative proximity, which is inferred from the measured Received Signal
Strength Indicator (RSSI) of the Radio Frequency (RF) signal between each pair of
WBAN sensors. By using this new sensing modality, more non-activity-intensive
postures can be detected, without having to add extra components and the associated
energy overhead. In this chapter, an experimental system is developed for integrating the
proximity data with other traditional modalities including acceleration and orientation.
This multi-modal data is then processed using a Hidden Markov Model (HMM)
framework for assessing the instantaneous body postures. The HMM is leveraged for
dealing with the proximity sensing errors caused by a subject’s clothing, body structure,
irregular RF propagation, and the variability in sensor mounting.

We have developed a WBAN system [24],[60],[6l] that is capable of detecting a
wide set of postures including SIT, SIT-RECLINING, LYING-DOWN, STAND, WALK
and RUN, by leveraging our proposed inter-sensor relative proximity sensing modality
and an HMM processing model. An online video demonstrating the preliminary working

prototype can be found in [80].

22

3.4 Target Sensor Modalities

Three sensor modalities, namely, acceleration, network-integrated proximity sensing
and orientation are used. A two-axis [19],[33] piezoelectric accelerometer in the
Mica2Dot sensor card is used for detecting body movements. Acceleration data is
generated in the units of gravitational acceleration g. While a near-zero acceleration may
mean a very low activity posture such as sitting or lying-down, a high acceleration can
indicate a high activity posture such as jogging or running.

The novel network-integrated proximity sensing or relative proximity between the
sensor nodes is the second sensor modality that is measured in dB using the received
signal strength indication (RSSI) of the radio frequency (RF) signal. Each sensor is set to
periodically send a Hello message with a preset transmission power that is enough to
reach all sensors on the body. Based on those Hello packets, each node creates and
maintains a neighbor table, with information regarding RSSI for all other sensor nodes on
the body. This way each node maintains a measure of the relative proximity with respect
to the other nodes. High RSSI values (high signal strength) indicate that the body parts
are positioned relatively close to each other like during a sitting posture. Similarly,
relatively lower RSSI values indicate that the corresponding body parts are relatively
farther apart (e. g. during a standing posture).

Sensor orientation is the third modality that can indicate the orientation of a body
segment to which a specific sensor node is attached. Orientation information can be used
for identifying low-activity orientation-specific postures such as lying down and
reclining. The two-axis [19],[33] piezoelectric accelerometer in the MTSSIO sensor card

is also used for orientation detection. The constant component of the accelerometer’s

23

output indicates a sensor node’s orientation. Orientation output is extracted by integrating
the acceleration output, and can be assessed for both X and Y directions by the
corresponding accelerometer outputs. Therefore, the orientation indicator shares the same

units as that for the raw accelerometer output.

 

Offline Online

 

Off-body Indoor workout, Indoor patient and
fitness evaluation elderlymonitoring
On-body Outdoor workout, Outdoor patinets,
fitness evaluation soldiers and Sports
monitoring

 

 

 

 

 

 

Table 3.1: Computation Modes and their Applications

3.5 Computation Modes

As Shown in Table 3.1, body context identification can be categorized into four
computation modes: off-body offline, on-body offline, off-body online, and on-body
online. For the off-body case, all sensor data is wirelessly collected to an off-body
processing server (see Chapter 2) which is used for the context identification. In on-body
scenarios, the identification processing is performed at the sensor nodes themselves,
either at a single node or at multiple nodes for improved processing load distribution. The
offline and online processing modes represents whether the identification is done in
real-time or not. For online processing, the amount of available sensor data is generally
less than what is available in the offline scenarios.

From an application standpoint, the on-body processing is more suitable for outdoor
applications since a separate processing server may not usually be available. In indoor
settings, however, such servers may be available and therefore the off-body applications

can be supported. As summarized in Table 3.1, real-time monitoring applications,

24

off-body or on-body, are better supported using the online processing mode. Applications
that require post-collection evaluation are better suited for the offline mode. Results
presented in this chapter correspond to off-body and offline computation mode. The
prototype system described in this chapter performs off-body and offline posture
identification.
3.6 Sensor Placement and Modality Usage

Three sensor nodes are mounted at three body locations, namely, the right thigh, the
upper right arm and the right ankle. The thigh sensor is used for capturing body
acceleration, while all three sensors are used for detecting the relative proximities
between all sensor pairs, and both arm and ankle sensors are used for sensing the
orientations of those body parts. Through extensive experimentation with different
subject individuals it was found that the above sensor placement can provide enough
information diversity for all three sensing modalities for them to be applicable to our

proposed posture identification process.

 

 

 

 

 

Sensor Node Sensor Supported Target
ID Placement Modality Pastures
1 Upper Right Arm Orientation, SIT, DWN,
Proximity STD
2 Lower Right Orientation, SIT, REC,
Ankle Proximity DWN, STD
3 Right Thigh Acceleration, WAK, RUN,
Proximity SIT, STD

 

 

 

 

Table 3.2: On-Body Sensor Modality and Placement Summary

 

A summary of placement, supported modality, and target posture information is
shown in Table 3.2. While all three nodes are programmed to provide RSSI-based

proximity data, the nodes on the arm and ankle are used for generating orientation

25

information, and the sensor node on thigh is used for assessing a subject’s level of bodily
acceleration. The last column identifies which set of physical postures that each specific
sensor node contributes toward measuring. Throughout the rest of the chapter the target
postures will be abbreviated as: SIT (sit straight), STD (stand), REC (sit reclining), DWN
(lying down), WLK (walk), and RUN (run).

Note that while more sensors provide a richer set of data to work with, they also
make the overall sensor-wearing process increasingly cumbersome. Therefore, a key
objective of the system design is to achieve high posture-identification success with as
few sensor nodes as possible. Also, it was found that due to the variability of the RF links
caused primarily by body movements, antenna mis-orientation, and signal blockage by
clothing material, not only does the network topology become unpredictably dynamic,
but the proximity information indicated by the RSSI values can also vary over a very
large range. This has the potential for introducing serious inaccuracies in the posture
identification unless specific measures are taken to suppress the effects of such
measurement errors. A Hidden Markov Model has been used to specifically address these

measurement errors and variability.

3.7 Posture Modeling and Generation

The posture transitions of a human subject are modeled as a Markov Process in
which the subject’s posture transitions are assumed to follow a memoryless process
[81],[82]. The transition probabilities across the postures, as shown in Figure 3.3,
represent a subject’s behavior that is assumed to remain stationary for a certain time
interval. The corresponding transition matrix, termed as A, remains fixed during such an

interval, and can vary across the intervals when there is a broad change in behavior. In

26

the following experiments we generate a sequence of 50 posture states using the

transition probability matrix:
”0.5 0.2 0.1 0.2 0 o 1
0.5 0.5 0 0 0 0
0.2 0 0.5 0.3 0 0
A=iai,j]=
0.3 0 0.1 0.4 0.1 0.1 ,
0.1 0 0.1 0.2 0.4 0.2
__ 0 O 0.1 0.2 0.3 0.4_

 

 

in which states 1 through 6 represent the postures SIT, REC, DWN, STD, WLK and

RUN respectively.

 

 

 

 

 

 

 

Figure 3.3: Posture state transition machine

27

 

 

 

 

is not applicable to

ACfiVityi/ow Moderate High
Accelerometry C Level

@—
B:
i

v
E

IE

 

 

 

  

 

 

 

 

 
 
 
 

 

 

  
 
 

differentiate between ¥
th tur V
ese pos es Accelerometry
is applicable for
_ posture detection
High
RSSI is not
applicable to
differentiate
SIT DWN
Q I lRECl I D between these
postures
Vertical Horizontal
Arm Sensor -
Orientation
is not applicable to
differentiate between
these postures

- SIT: sitting

- REC: sit-reclining
- DWN: lying-down
- STD: standing

- WLK: walking

- RUN: running

Vertical Horizontal

 

 

 

Figure 3.4: Posture identification using multi-modal thresholds

As a part of each experiment, a human subject is handed the resulting posture
sequence and is instructed to follow the sequence with 20 sec being spent in each posture

thus the entire experiment lasts for 1000 see. Note that the transition matrix A is chosen

28

based on long observation of typical behavioral patterns of multiple human subjects in

our laboratory setting.

3.8 Threshold-based Identification

The flowchart in Figure 3.4 depicts a mechanism in which context identification is
accomplished by applying different thresholds for all three sensing modalities. After the
low- and high-activity postures are separated using the degree of acceleration recorded by
the node on the thigh, a proximity threshold (applied in terms of RSSI) is used to
distinguish between STD (stand) and the other remaining postures, namely, SIT, REC,
and DWN. The lying-down (DWN) posture can subsequently be separated using the
orientation information from the node on the arm. Finally, the differentiation between
SIT and REC is performed based on the orientation information from the ankle. Details
about the exact threshold values used for different sensor modalities are presented in

Table 3.3. Results presented in this section correspond to an off-body and offline

 

 

 

 

 

 

 

 

 

computation mode.
Threshold Moderate High Avg. Arm Ankle
Group Activity Activity RSSI Ornt. Ornt.
Level (mg/s) Level (mg/s) (-dBm) (mg) (mgL
T hr] 5 20 80 460 470
Thrz 5 30 70 470 480
Thr3 8 30 80 480 490
Thu 8 30 90 490 500
Thr5 8 30 100 500 510
Tim; 8 30 110 510 520
Tiny 8 30 120 520 530
Thrg 8 30 130 530 540

 

 

 

 

 

 

 

 

Table 3.3: Threshold group values for context identification

Sensor reading for all three modalities and the corresponding actual postures for all

50 posture slots are reported in Figure 3.5. For the sake of brevity, the postures SIT,

29

REC, DWN, STD, WLK and RUN are identified by the letters S, R, D, T, W and U

respectively. The actual state (the posture that the subject is in) during a slot is reported

 

A (a) Thigh accelerometer output

$320 . .. .
‘ Actual States Activityh‘avelof . ; ; l . . _ , '
V240‘?‘;;;;g_: ThighSensorii"'??ff'ili§“‘:‘.;i."

 

    

o . . '1'
TTSSTSSS’I‘W’ITWTSSSSSSSTSSSSSRSRSSRRRRSDTSSTDTUTTSST

 

 

 

Activity Level in
W
O

 

 

 

 

 

 

 

  
  

 

 

 

Time(s)
A (b) Average RSSI values
m320 ‘,I.': ;I‘I;."
'3, :ActualStates .;:;:':‘...;f;“AngSSIVaIues;:.2:';ff
5240‘;::i::‘mt'if:"‘::lA“:‘
0
on
«3
S
> .‘ '.i ~ ..,. .: .
< 111111 I I IIIIIIIIIIIIIII III IIIIIII
TTSSTSSSIWIIWISSSSSSSTSSSSSRSRSSRRRRSDTSSTDTUTTSST
Time(s)
320 (c) Arm accelerometer output
’33 ILL‘Ei-ActualStatesin‘Ji AnnSensorAcc.iT;lii"
E l-I ..I;;'.:l‘::.:I::‘:: -l;
I:
.9
E
s:
g .
TTSSTSSSTWIIWISSSSSSSTSSSSSRSRSSRRRRSDTSSTDTUTTSST

Time (s)

 

(d) Ankle accelerometer output

 

ActualStates AnkleSensOrAcc

 

 

 

0 ll
TTWSSTSSSIWI IWISSSSSSSTSSSSSRSRSSRRRRSDTSSTDTUTTSST
Time(s)

 

 

 

Figure 3.5: Sensor outputs and actual postures

30

by the corresponding letter on the horizontal axis during the slot. With each slot lasting
for 20 seconds, the entire experiment corresponds to 1000 seconds, representing 50
posture slots.

The graph in Figure 3.5:a reports the actual posture states and the corresponding
activity levels, which are computed as the absolute values of the first order derivatives of
the raw accelerometer outputs. The derivative represents the difference between two
successive acceleration samples collected at a 20Hz sampling rate. The computed
derivative numbers are then integrated using a moving average with window size of 5
sampling slots. Finally, those integrated derivative numbers for both X and Y directions
(using the accelerometer outputs for both X and Y axes from the thigh-mounted sensor)
are averaged to produce the activity levels that are plotted in Figure 3.5:a. As expected,
the activity levels are high for the W and U (WALK or RUN) slots, and low for all other
posture slots.

Average proximity information from all three sensor nodes, along with the actual
postures, is reported in Figure 3.5:b. Each node periodically (once in every Hello interval
of 1.5 seconds) computes the average RSSI value based on the radio signal reception
through Hello packets from the other two nodes, and then wirelessly sends that data to the
thigh-mounted sensor. This thigh sensor then computes a master average based on the
averages received from the other two sensors and its own average. This final average (in
-dBm), which is reported in Figure 3.5:b, is then wirelessly transmitted to an off-body
machine for further processing. In these readings, high RSSI dB values indicate high

received radio signal strength and vice versa.

31

The following observations should be made from Figure 3.5:b. First, the average
RSSI has an overall trend to be the lowest for SIT (S) and the highest for STAND (T).
This is consistent since the body parts are generally closely situated during sitting, and
further apart while standing. The average RSSI values for the other two low-activity
postures SIT -RECLINE (R) and LYING-DOWN (D) fall in between those for S and T.
Second, while generally maintaining this trend, there are certain anomalies caused by
several factors including radio signal blockage by the clothing material, unintentional
change of sensor node and antenna orientations, and various other imperfections in sensor
mounting.

Figure 35:0 and Figure 3.5:d show the X—direction orientation indication (as
introduced in Section 3.4) for the sensor nodes attached to the arm and the ankle
respectively. Orientation indication is computed by first averaging the raw accelerometer
output over 20 samples (i.e., 1 second), and then integrating those average values using a
moving average with window size of 5. Both sensors on the arm and the ankle are
mounted such that high X-direction orientation values in Figure 35:0 and Figure 3.5:d
indicate horizontal orientations of the corresponding body segments, and low values
represent relatively vertical orientations. Note that the Y-direction orientation information
is not used in these experiments.

The arm sensor orientation in general can be used to detect the LYING-DOWN (D)
posture, since it is evident in the plot that the arm sensor orientation readings in D
postures are distinctively more horizontal (higher values) compared to those during other
postures. Also, the ankle sensor orientation can be used to detect both LYING-DOWN

(D) and SIT-RECLINING (R), because the orientations of the ankle in these two postures

32

are also distinctively more horizontal (higher values) compared to the other two
low-activity postures SIT (S) and STAND (T). Note that the sensor data patterns, as seen
in Figure 3.5 for all modalities, are consistent with the threshold-based context
identification logic presented in Figure 3.4.

Threshold values of all sensor modalities at different sensors are depicted in Table
3 .3. Each set of threshold combinations is grouped together, and eight such groups are
depicted in the table. The first and the second columns represent the moderate and the
high activity level thresholds to be applied on readings from the thigh-mounted sensor
node for difl‘erentiating between the WLK and RUN postures (see Figure 3.4). The third
column represents RSSI thresholds for the master average RSSI values collected and
computed at the thigh sensor node. The last two columns indicate threshold values to be

applied on the orientation readings from the arm and the ankle-mounted sensor nodes

 

 

     
  

    

 

 

 

 

 

respectively.
90% - g E
a Subject-3: Subject-2
5.; 80% — I , / ~
2 A . ? ’
0 E ‘ I 2
g0 70%- 4' 'i 1
a 3 I ‘
8 ? s b' tl/ ?
g 60%- 2~ ~ “513‘?
50% i ; i i ; i 4 ;
Thrr Thrz Thrs Thr4 Thrs Thre Thr7 Thrs
Threshold Values Combinations

 

Figure 3.6: Detection accuracy for multiple human subjects

Figure 3.6 depicts the threshold based context detection accuracy computed over the

50-state posture sequence generated by the A matrix reported in Section 3.7. Using the

33

thresholds specified in Table 3.3, the comparison algorithm from Section 3.7 has been
applied to the multi-modal sensor data obtained fiom all thee body-mounted sensors for
identifying the instantaneous body posture. The identified posture is then compared with
the subject’s actual posture for computing the success rate as reported in the figure. Such
success rates are presented as percentage matches for different threshold groups and for
different human subjects. Three individuals in these experiments were asked to follow the
same controlled posture sequence as used in Figure 3.5 for several rounds, before the
identification performance was computed.

Observe that in spite of the errors contributed by sensor and antenna mis-orientation
and radio signal blockage by clothing material, this threshold based mechanism can
detect the six postures with up to approximately 84% accuracy. However, since the
identification success rate is heavily sensitive to the threshold values, choosing the right
threshold values is an important design step for this mechanism to work.

A potentially restricting aspect of this threshold-based mechanism is that the optimal
threshold values (threshold groups in this case) are also sensitive to the individual
subjects’ physical and motor aspects during his or her postures. For example, while the
threshold group Thr5 yields the best identification accuracy of 84% match rate for
subject-2, the performance for subject-3 maximizes at 82%, for the threshold group Thr3.
In fact, at Thr5, for subject-3 the system delivers a poor posture identification rate of only
74%. These results allude to a practical limitation of the threshold-based posture
identification in terms of the need for person-specific threshold dimensioning. Other
experiments further indicated that the optimal threshold value can change even for an

individual based on his or her behavioral changes over time. In the next section we have

34

developed a Hidden Markov Model (HMM)-based mechanism for adaptive and

subj ect-independent posture detection.

3.9 Capturing Stationary Behavior Using Hidden Markov Model

The inability of the threshold based mechanism to handle the degraded quality of
sensor data stems from the fact that the identification process does not leverage the
stationary nature of human behavior over certain time intervals. To address this
limitation, we adopt a stochastic posture identification solution that attempts to leverage
the stationary nature of human posture by modeling the posture state machine as a
Hidden Markov Model (HMM) [83-86].

The key concept of the HMM [83] are as follows. A stochastic process is represented
by a discrete time Markov Chain consisting of multiple states which are hidden from an
observer in the sense that an observer cannot directly determine which state the system is
in at any given point in time. However, a number of observable parameters, stochastically
representing the states, are visible to the observer. The idea of HMM formulation is that
if the state transition probability matrix and the observation generation probabilities are
known (or measurable) to the observer, the latter can estimate the current state of the
Markov Chain. Using HMM it is also possible to compute the probability of occurrence
of a specific state sequence [84-87].

3.9.1 HMM Mapping

The posture identification problem with multi-modal sensing framework is mapped

as an HMM formulation as follows.

35

Posture State Space: N postures are modeled as N hidden states with the state space
represented by S = {S1, S2, SN}. In this specific case N = 6, for postures SIT,

SIT-RECLINING, LYING-DOWN, STAND, WALK and RUN.

Observation: At each state, the observation is represented by a vector 0, which is
constructed by combining four sub-vectors 0 = [C U X U R U K], where C represents

the activity level information from the thigh sensor, X represents the master average RSSI
value from all three sensors, and R and K represent the orientation indications from the
arm and ankle sensors respectively. HMM observation vectors are constructed from the
multi-modal sensor data as shown in Figure 3.5. Each sub-vector is created as follows.

The activity level observation at any point in time is represented by the sub-vector C
= {c1, c2 ..... cMC}, in which each cm (m =1, 2, MC) is a binary variable which can be
either ‘0’ or ‘1’. The peak-to-peak activity level range (see Figure 3.5:a) is divided into
MC equal windows, and then depending on which window the current activity level falls
in, the corresponding cm is set to ‘1’, and the rest of the sub-vector elements are set to ’0’.
Note that the value of MC determines the granularity of observation, which in turn, is
expected to influence the quality of the hidden posture state identification process. The
flindow _B_oundary (WB) points for the sub-vector C is represented by WBC. The number
if WB points is one less than the value of MC.

Observation sub-vectors X, R, and K for RSSI values and orientation indications
from the arm and the ankle sensors are constructed using the same mechanism as used
above for the activity level sub-vector C. The corresponding granularity factors (e. g. the
sub-vector size) are indicated as MX, MR and MK respectively, and the window boundary

points are represented by WBX, WBR, and WBK respectively.

36

 

At any time instant t, all four sub-vectors are combined into an overall observation
vector 0,. Also, an overall observation granularity factor M is computed by adding the

individual granularity factors MC, MX, MR and MK. The minimum value of M in our
system was chosen to be 9, with corresponding values of MC, MX, MR and MK to be 3, 2,
2, and 2 respectively. We have experimented with various values of M, ranging from 9
(coarse granularity observation) to 15 (fine granularity observation). For M to be 15, MC,
MX, MR and MK were chosen to be 4, 5, 3, and 3 respectively.

Consider an example in which M is chosen to be 9 with MC, MX, MR and MK as 3, 2,
2, and 2, and the window boundaries WBC, WBX , WBR, and WBK are chosen as {8, 30}
mg/s, 90 dB, 490 mg, and 500 mg respectively. Now with raw sensor outputs
representing activity level of 4 mg/s, RSSI of 70 dB, and arm and ankle orientation
indications of 470 mg and 480 mg, the resulting sub-vectors C, X, R, K will be [1,0,0],
[1,0], [1,0], and [1,0] respectively. Therefore, the overall observation vector 0 will be [1,
0, 0,1, 0,1, 0,1, 0].

As indicated in Figure 3.3, the parameter 0, represents the observation vector at time
slot t, with T as the final observations in an experiment. In all our experiments, the value
of T is 50. In other words, 50 observations, each corresponds a state lasting for 20
seconds, are generated to feed into the HMM estimation system.

Transition Probability Matrix: The posture transition probability matrix is

represented as A = [“111] , where

ai,j:p(q:=Sjlqt—1:Si)a ISi,jSN (3.1)

37

A is an N x N matrix, where N corresponds to the number of postures (states), which

is 6 in our case. The quantity q, denotes the actual posture at time t. The parameter aij

represents the probability that the next posture is j, given the current posture of the
subject is i.

Observation Probability Matrix: As done for the observation vector 0, the
observation probability matrix B is constructed by combining four sub-matrices

as B =[BC UBX UBR UBK] , where BC, BX, BR and BK correspond to activity level,
RSSI and orientation indications from the arm and the ankle sensors respectively.
The elements of sub-matrix BC, whose dimensions are N x MC, are represented by:

bm -_-p(C=[cl = 0,....,cm = l,....cMC = O] |q, =Sj),

ISjSN,ISmSMC, (3'2)

where C represents the activity level observation sub-vector. The parameter bi,"
represents the probability that in posture state j, the element cm in the observation
sub-vector C is ‘1’ and the rest of the elements in C are all zeros. In other words, when a

human-subject is in postures state j (f can be one of six targeted postures in our system),

the quantity bi," indicates the probability that the observed activity level falls in the mth

window of observation within the sub-vector C.

Following the same mechanism as used above for the activity level, observation
probability sub-matrices BX, BR, and BK are constructed for observed RSSI and
orientation indications from the arm and the ankle sensors. The dimensions of those
sub-matrices are N x MX, N x MK, and N x MR. Combing all four sub-matrices, as shown

below, the overall observation probability matrix B with dimension N x M is constructed.

38

B = [BC sub — matrix BX sub — matrix BR sub - matrix BK sub - matrix]

 

 

 

 

 

 

 

where
'ch1 .. b1,cm .. bLCMC I —bl,.x1 " bkfm '° 1’1,qu I
BC: bj,c1 .. bjICm .. bjaCMC ,BX = bj,x1 '° bj,xm °° bj,xMX
_bN,c1 ,, bN,Cm ,, bN,CMC_ _bN,x1 .. bem .. bNJMX_
pblfl.l .. erm .. bum I P b1,.k1 " bl,Icm .. [91$ij -
BR= bin -- be... -- brim and BK = 1’qu -- [’ka .. bLkMK
thfl .. bN,,m .. bNa’MR _ _bN,k1 .. kam .. bNakMK _

 

Initial State Distribution: This is represented by a vector 7r=[7t,.] of length N, so
that: ”i=p(qo=Si)a ISiSN (3-3)

The quantity 7:, represents the probability that the posture Markov chain is

N
initialized at state i. By definition, 2 77: = 1 .
i=1

Based on the above definitions, a system, modeled using HMM, can be fully

specified by the parameters A, B and 1: which are represented together as a tuple:
xl=(A,B,7r) (3.4)
As presented in the next section, we first compute the individual probabilities of the

system being in each possible posture state at a given time. As shown in the derivation,

these probabilities depend on the system’s xi , and the observation sequence {01 02 0 3
. 0T}- After the probabilities are computed, the posture state identification is

39

accomplished by finding the most likely state, which is the one with the highest current
probability.
3.9.2 Posture Detection using HMM

The probability of observing a given sequence 0 = {01.02.---0T} of length T time

steps is represented as P(0|).), and can be evaluated using the forward-backward

procedure [87], as follows:
N
P(0/l)=ZlaT(l)9 (3.5)

where a, (i) is referred to as forward variable, and defined as:
a,(i)EP(01,02,...0,,q, =Si [1). 1S i S N (3.6)
It represents the probability that the partial sequence 01, 02, 0,, until time step t,

has been observed and the current posture state at time t is S,-, given the HMM model It.

a, (l) is a vector of dimension N (which is the total number of possible states). Another
variable ,6, (i), referred to as backward variable, is defined as:
flt(i)EP(0t+1,...,0T |q, =Si,/l). 1S i S N (3_7)

This represents the probability that the partial sequence from time step (t+1) to the

end has been observed and the current posture state at time t is 5,, given the model It.

,6, (l) is also a vector of dimension N. Now another variable 7, (i) is defined such that:

7r(i)=P(qr = Si IOJ.) (3.8)

40

where 7, (1) represents the probability of being in state S; at time t, given an

observation sequence 0, and the model it. Eqn. 3.8 can be expressed in terms of the

forward-backward variables as:
ar(i)flr(i)= a.(i)fl.(i)
“0"” few (2') ‘3”

 

 

7.(i)=

which is a vector of dimension N at time t. Using 7, (l) we can solve for the

individually most likely posture state q, at time t [83], as:

arg max

q’:1sisN

[70(1)]: 1515 T (3.10)

This q, represents the detected posture state at time t.

3.10 Experimental Results

3.10.1 Manual Calibration

In this section we describe the performance of HMM based posture identification and
its performance in comparison with the threshold based approach described in Section

3.8. The same transition probability matrix

 

 

"0.5 0.2 0.1 0.2 o 0"
0.5 0.5 0 o 0 0
A=[a,- .1: 0.2 0 0.5 0.3 0 0
. ’1 0.3 0 0.1 0.4 0.1 0.1.
0.1 o 0.1 0.2 0.4 0.2
_0 0 0.1 0.2 0.3 0.4_

41

as used for the previous experiments, is used for generating a posture sequence to be
followed by the human subjects. Note that for the results in this Section, the A matrix
used for posture sequence generation is also used for the HMM model formulation. In
other words, it is assumed that the A matrix used for HMM is already trained. During an
initial set of known states, the B matrix is first computed, and then the actual posture
identification process was initiated. This initial period is referred to as an observation
calibration phase. As for the Initial State Distribution matrix 1:, we have used [0, 0, 0, 1,
0, 0], for all the experiments results presented here. This means that in all experiments the
subject should start with the posture STAND. These A, B and 1: matrices constitute the
HMM system parameter A.

State identification using/l has been carried out using the HMM technique described
in Section 3.9.2. Figure 3.7 reports the posture identification performance with HMM in
comparison with the threshold based mechanism as introduced in Section 3.8. As done
before, the success rates are measured by comparing the identified postures with the

actual postures from the generated posture sequence using transition probability matrix

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

A.
M MC, MX, MR, MK WBC WBX WBR WBK
9 3,2,2,2 {8,30} 90 490 500
10 3,3,2,2 {8,30} {80,90} 490 500
11 3,4,2,2 {8,30} {70,80,90} 490 500
12 3,5,2,2 {8,30} {70,80,90,100} 490 500
13 3,5,3,2 {8,30} {70,80,90,100} @90,500} 500
14 3,5,3,3 {8,30} {70,80,90,100} (490,500} (490,500}
15 4,5,3,3 {8,30,40} {70,80,90,100} (490,500} (490,500}

 

Table 3.4: Observation sub-vectors and window boundaries for different observation
granularities

The success rate for posture identification using HMM is reported with seven

different observation granularities corresponding to M = 9, 10, ll, 12, 13, 14, and 15. For

42

each such values of M, the corresponding values of MC, MX, MR and W, and their

observation window boundaries are summarized in Table 3.4.

 

 

 

 

Subject-l

 

 

 

$3 fins: 2::

 

$3 Ens; ES:

 

$3 GTE SE:

 

 

$3 ASHE SE:

 

                

 

 

$3 Bus: 22:
$3 87% 22:
$8 8qu 22:
$3 Ne;
$313.
_ H _ _
% % % % % %
0 0 0 0 O 0
m 9 00 7 6 5
582 93583

Subject-2

 

 

 

 

{coo Gun—‘6 ESE

 

$3 Eng 22:

 

$3 GTE 22:

 

 

$3 ANT—.5 2::

 

$3 Eng 22:

 

e3: Sun—>5 SEE

 

$3 ans; :2:

 

KL.» M 2:.

 

 

e\eoc NE...

{eon 15,—.

 

 

 

100%

90%_,_,

£332 owfiaoocom

80% r
70% r
60% ‘
50%

 

Subject-3

 

 

 

$3 GTE 22:

 

$3 Eng 22:

 

 

       

$3 San—.6 22:
$3 STE :2:

$3 STE :2:

 

 

 

 

 

 

 

 

$288—$255
$23522:
$3 2:

$3 "Ea
$3 1.;

_ S _ _
% % % % % %
0 0 O 0 0 0
0 9 oo 7 6 5

£032 owfiaoouom

 

 

 

Figure 3.7: Posture identification performance

43

The first entry for M = 9 (MC=3, MX=2, MR = 2 and MK = 2) indicates that the
three window levels for the C sub-vector are realized with two window boundaries WBC
of 8 mg/s and 30 mg/s. Similarly, two window levels for the X sub-vector are realized
with one window boundary WBX of 90 dB. Observe in the table that with increasing
observation granularities (higher MC, MX, MR and MK) a larger number of window
boundaries are needed to implement higher number of observation window levels.

The following observations are to be made from Figure 3.7. First, the HMM
approach delivers better state match rates (for example 84% to 96% identification success
for human subject-2) compared to the best case performance using the threshold-based
mechanism (84% identification success for the same subject), that is with the threshold
group 4 (8, 30, 70, 490, 510) as shown in Table 3.3. Second, higher observation
granularity (larger M) for HMM provides better posture identification success rate, with
performance saturation occuning beyond the granularity factor around M = 12. Third,
once a sufficiently large observation granularity (e.g. M = 15) is chosen for HMM, unlike
in the threshold-based scheme, no optimal parameter dimensioning is needed. This is a
significant advantage in terms of implementation feasibility. Finally, with similarly large
observation granularities, the HMM continues to provide superior posture identification
performance in a human subj cot-independent manner. This further reinforces the
practicality of the mechanism in not having to dimension any individual-specific
parameter which may cause significant performance variation as observed for the
threshold-based mechanism.

3.10.2 Automatic Observation Calibration

For the results above, the observation probability matrix B has been constructed

during an observation calibration phase before experimenting with each individual human

44

subject. This calibration process (construction of matrix B based on observations)
somewhat compensates for the inconsistencies in the observation values due to variations
in clothing, personal posture specialties and other ambient differences. In fact this
calibration process accounts a great deal for the consistently superior performance of
HMM compared to the threshold-based strategy, as presented in Figure 3.7.

In this section we implement a self-calibration process of the B matrix, so that the
proposed posture identification mechanism can be more practically implemented without
having to manually calibrate the B matrix for each individual subject.

We use the Baum-Welch iterative algorithm [83], for which the key idea is to start
with initial B matrix, and then iteratively adjust it based on the stochastic difference
between the identified (using HMM) posture state sequence and the expected sequence
based on the notion of the state transition matrix A. Details of the Baum-Welch derivation

and the algorithm are included in the Appendix.

 

 

     
 

 

 

 

 

 

100 f l _ = . i
132 f i %
:53 90+B3§ , ;
a E
2 \
0
go 80—«-~r/
E \3
§ Bl
a 70
60 . I . , I ; ;
1 4 7 10 13 16 19 22
Number of Iterations in Baum-Welch

 

Figure 3.8: Automatic self-calibration of the B matrix

45

Figure 3.8 demonstrates the performance of this self-calibration process in terms of
the posture identification accuracy over multiple iterations. Here we used the observation
sequence of human subject-2 of the last experiments, with observation granularity factor
M = 12. Observe that with all three different initial B matrices, the identification accuracy
gradually increases over time with Baum-Welch iterations. For all three cases, the posture
identification process started delivering the best performance within 12 iterations. In a
deployment sense, this means that after wearing the sensors, the subject should continue
with his or her regular behavior for a while for allowing the network to self-calibrate the

HMM B matrix. After that, the identified posture recording should start.

3.11 Summary Conclusions

In this chapter we present a novel network-integrated sensing modality, inter-sensor
relative proximity, which is inferred fi’om the measured Received Signal Strength
Indicator (RSSI) of the Radio Frequency (RF) signal between each pair of WBAN
sensors. The concept of RSSI-based proximity is experimentally developed and then
integrated within a Hidden Markov Model (HMM-based stochastic processing
framework for accurately identifying human body postures in a subject-independent
manner. It was first shown that the traditional uni-modal approach using only the
acceleration information do not work well when it is necessary to differentiate between
low-activity postures, and we addressed that by proposing the relative proximity
sensing modality.

Then, it was demonstrated that although a naive threshold-based mechanism can be
used for reasonable detection performance, the intrinsic errors and unpredictability of the

on-body data collection process require a delicate dimensioning of the used threshold

46

 

values for consistent posture identification performance across various human subjects.
To avoid this, an HMM-based detection process is applied with observation
self-calibration using the Baum-Welch algorithm. It was shown in this chapter that the
HMM method with our novel sensing modalities are able to consistently deliver
significantly better detection performance than the threshold-based mechanism in a more
individual-independent manner.
3.12 Appendix: Iterative Hmm With Automatic Observation Calibration

As preposed in [83],[86], it is possible to calibrate the HMM parameters in ll. such
that the quantity P(0|,l), representing the conditional probability of an observation
sequence (of length T) is maximized. In our specific application of self-calibration as
discussed in Section 3.10.2, it is required to adjust the observation probability matrix B
while keeping the other two parameters A and it in A constant. The Baum-Welch
algorithm [83] is used in our implementation to iteratively obtain an estimate of B that
results in a A which is guaranteed to locally maximize P(0|A).

As defined in Section 3.9.1, the element b,-_,,, in the matrix B represents the probability

that in posture state j, the elements cm, x,,,, rm and km in the observation vector 0 are ‘ l 's’
and the rest of the elements are all zero. The quantity by" can be computed as:

T
27:0)

b S1. of =Vm

jam T
273(1)
t=1

 

(3.11)

where the denominator represents the probability that the system is always in state j with

all possible observations. The nmnerator represents the probability that the system is in

47

 

state j with a specific observation such that the elements vm, where v,,, = {cm x,,,, r,,,, km}

in the observation vector 0 are ‘ l 's’ and the rest of the elements are all zero.
Using Eqn. 3.11 as the iterative step for changing the B matrix, we have implemented

the following algorithm for implementing self-calibration as explained in Section 3.9.2.
1) Collect observations 0 = 0102...0T
2) Initialize A using a starting B matrix with constant A and 7t

3) Given observation sequence 0 = 0102...OT and A, compute:
y,(j), VIStST, ISjSN
4) Compute new B matrix by updating the elements bj,m based on Eqn. 3.11

5) Set new knew using the new B matrix
6) Compute a new quantity MAXLIKELIHOOD as:

MAX LIKELIHOOD = max[P(01...0T Ir), 10(01 ...0T WM]

7) 7»: Knew

8) Go to step-3 and repeat till the quantity MAXLIKELIHOOD converges

 

 

 

 

Log Likelihood

 

 

 

1 3 5 7 91113151719
Number of iterations of the Baum-Welch algorithm

 

 

 

Figure 3.9: Performance of Baum-Welch iterative algorithm

48

The fact that the newly estimated B matrix in Step 4 is computed based on the actual
observation sequence, ensures that the estimation would improve the quantity P(0|/l).
This accounts for the monotonically increasing nature of the MAXLIKELIHOOD, as
evidenced in Figure 3.9, which demonstrates the convergence performance of the
Baum-Welch algorithm in terms of the evolution of the log of the quantity
MAXLIKELIHOOD. Observe that with all three different initial B matrices, the
MAXLIKELIHOOD monotonically increases over the algorithm iterations, and
converges approximately after 12 iterations, which is consistent with what has been

reported in Figure 3.8.

49

Chapter 4. Transmission Power Assignment with Postural Position

Inference

In this chapter we address the issue of energy-aware on—body communication by
developing a posture-aware transmission power control framework. An on-body link’s
postural position can be highly dynamic due to the unpredictable RF attenuation caused
by a slew of factors including subject’s clothing, antenna orientation, and physical stature
of specific subject individuals. In real-time data exchange across various on-body-sensors
with point-to-point or multi-point-to-point data transaction, the optimal transmission
power depends on the physical distance of the link, and its instantaneous channel
condition. As a result, static pre-defined transmission power is not able to provide
continuous link connectivity while ensuring minimum required transmission power
consumption. In this chapter we propose a dynamic link power assignment mechanism

for optimal energy management in on-body links data exchange.

4.1 Related Work

The energy conservation mechanisms in [2],[14],[44] attempt to develop
synchronized sleep scheduling across on-body sensors so that the overall energy
consumption is controlled by the communication duty cycle. These mechanisms do not
control the transmission power of individual on-body links as done in this chapter.
Transmission power control can be executed simultaneously with node sleep scheduling
as a complementary energy-saving strategy.

Several RF Transmission Power Control (TPC) papers [88-90] in the literature use
node localization for computing link distance, which is subsequently used for power

assignment using known models of RF signal attenuation with distance. Performing

50

 

localization for on-body sensors can be complex due to high node mobility triggered by
human posture changes, and energy expensive due to the need for localization
infiastructure such as ultrasound to perform Time Delay of Arrival (TDOA) [91]
computation. Such difficulties to run localization for on-body sensors make localization
oriented TPC protocols unusable for body area networks.

The protocol ATPC in [62] proposes a non-localization—based mechanism in which a
closed loop control is executed based on the received signal strength measurement.
Experimental evaluation of ATPC that has been reported in [62] includes large static
wireless networks deployed in terrains such as parking lots and in-building corridors.
Although ATPC is the closest to the approach adopted in this thesis, the primary hurdle
for mapping ATPC on body area wireless links stems from the fact that the on-body
wireless link characteristics are significantly more dynamic due to human postural
mobility, antenna orientation, clothing, and the physical stature of specific subject
individuals. Additionally, this protocol does not consider the control overhead associated
with closed loop power control mechanism as an algorithmic cost, which can be very
large when applied to on—body links with fi'equently varying link qualities. As a result,
the ATPC protocol, which was designed and tested only for relatively static and
long-distance wireless links, does not apply for highly dynamic on-body links.

Power control mechanisms for links between on-body sensors and off-body static
sinks are explored in [64]. Although the RF link characteristics in this setting are closer to
that of on-body links than those in [62], they do not quite represent the wide variations in
link qualities when both nodes of a link are on-body. Also, the assumption of unlimited

energy availability at the off-body sink node, as adopted in this chapter and in [2],

51

influences the TPC algorithm and energy accounting in ways that are not suitable for
severely energy-constrained and completely on-body links as targeted in this chapter.

The power control mechanisms for IEEE 802.11 MAC protocol in [92—95]
implement closed loop control with RTS and CTS packets carrying the feedback control
information. Full power RTS/CTS and controlled power data transmissions are sent
through separate channels [93],[94], and the transmission power is incremented and
decremented stepwise till a desired power level is achieved. These mechanisms, adapted
specifically for 802.11 networks, are not directly portable to the on-body systems for the
following reason. For a non-802.11 MAC layer, since there are no RTS/CTS packets,
separate feedback control packets will be necessary. Since the cost of implementing
power control is of paramount importance for on-body links, the objective will be to
minimize the usage of such control packets while maximizing the overall energy
efficiency through effective power control. This is not an issue for the mechanisms in
[92-95], since fi'ee RTS/CTS control packets are available in 802.11 on a per data packet
basis.

Transmission power control has also been studied extensively in the context of
cellular networks. In [96-98], closed loop transmission power control for Code-Division
Multiple-Access (CDMA) systems are proposed, where a separate feedback channel with
universal frequency for control data is used. The lack of a separate feedback channel in
an on-body network makes such cellular network power control mechanisms unsuitable.
The lack of a feedback channel poses the same problems imposed due to the lack of per
data packet RTS/CTS control as discussed above. Additionally, the time constant for

power control in the cellular networks are much larger than what is needed in the

52

 

on-body scenarios — causing those closed loop mechanisms in [96-98] to be too slow in

the presence of high postural mobility.

4.2 Dynamic Transmission Power Assignments

The objective of this chapter is to develop a closed loop power assignment
fi'amework that deals with solely on-body links with a wide range of link quality
variations due to the body specific factors as outlined above. Three different dynamic
on-body links transmission power assigrment mechanisms are proposed, namely, Linear,
Binary and Dynamic Postural Position Inference (DPPI). The DPPI proposed mechanism
has been experimentally evaluated and compared with the binary search and the
traditional fixed power increment/decrement strategy (termed as linear search), which
has been used in many closed loop mechanisms in the literature
[64],[96],[92-95],[97],[98]. The proposed mechanisms are distinct in that unlike majority
of the transmission power control mechanisms in literature [44],[14], it does not rely on
the availability of location information for individual sensor nodes, and uses a novel
technique of postural position inference which is specific and can be applied to body area

wireless networks.

4.3 Characterization of On-Body RF Links
This section presents an experimental characterization of on-body wireless links and 3
its quality variations with different transmission power, postural positions, and subject
individuals.
4.3.1 Experimental Settings
We construct a Body Area Network by mounting a set of sensor nodes on different

segments of the body (e.g. thigh, upper arm, ankle, waist etc), as illustrated in the

53

 

prototype system in Chapter 2 and Figure 2.1. In this prototype, the CC1000 chip in
MICA2DOT mote offers a hardware register to specify the transmission power at one of
22 different power levels, corresponding to a range [75] between 21.5mW and 63.5mW.
These power levels are represented as indexes 1 through 22. The sensors are worn with
elastic bands in such a manner that the sensor and its antenna orientation with respect to
the mounted body segnent do not change with frequent posture changes.

The sensors can form a mesh or a star topology depending on point-to-point or
multi-point—to-point data requirements in specific applications. Since the scope of this
chapter is primarily at the link level, the topological configurations are of very little or no
interest. From this point onwards, the links referred to in this chapter belong to a star
topology. All body-mounted sensors in this topology send data to another body-mounted
sink node. This node collects raw data, and sends processed results or events to an
off-body server using a wireless link. This external link is created between the on-body
sink node and to an off-body Mica2Dot radio node connected to a Windows PC (compute
server) through a custom-built serial interface running RS232 protocol, as shown in
Chapter 2 and Figure 2.1.

4.3.2 Target link and Posture Sequence

An on-body wireless link between a sensor node on the upper arm and a sink node at
the waist has been used for studying link quality variations. Since hand is considered to
be the most mobile part of the human body, the arm-mounted sensor has been chosen
specifically to demonstrate the impacts of postural mobility on link quality variations,
and the subsequent need for adaptive transmission power assiglments.

Link quality is characterized by observing the Received Sigral Strength Indication

(RSSI) values at the sink node for a given static transmission power from the

54

arm-mounted sensor node. The variation of RSSI represents the changes in link quality as
a result of body movements, RF attenuation by subject’s clothing and body segments, RF

multi-paths, and other changes in the surroundings.

      
 
 

 

On-body Link x E
WirsleSS/Ifilk‘xli

Out-of-Body Server

Figure 4.1: On-body sensor positions and the used posture sequence

Three different subjects, one female and two males, were asked to follow a sequence
of right hand movements corresponding to four different sitting positions as shown in
Figure 4.1. These hand positions represent natural sitting postures, and also indicate
scenarios with varying degee of RF attenuation due to siglal blockage and sensor node
movements. For each subject, each posture in the sequence lasts for approximately 10
seconds, and the entire sequence is repeated three times resulting in a total experimental
duration of 120 seconds.

4.3.3 Transmission Power Levels

For each posture in the above sequence, link characterization is also carried out with
different transmission powers. We have reported the results with four different power
levels 7, 13, 18 and 22, corresponding to actual transmission powers of 24.8, 34.5, 43,

and 63.5 mW respectively. The rationale for reporting only these power levels is as

55

follows. Level 22, the maximum available power, provides the best packet delivery, but
the worst transmission energy usage. Level 7, the available minimum, delivers the worst
packet delivery, although with very little energy usage. Power level 18 turns out to be the
minimum power level that provides around 100% packet delivery. Finally, level 13 is
reported, because it is somewhere in the middle of power levels 7 and 18, and provides a
reasonable balance between transmission energy consumption and packet delivery.

Sensor on the arm transmits four packets, one at each of the above power levels,
every one second interval. The sink node located at the subject’s waist region measures
the packets’ Received Sigral Strength Indicator (RSSI) and sends the data to the off-body
server (see Chapter 2 and Figure 4.1) for analysis and post-processing. Note that these
bundled transmissions of four packets together ensure that the RS SI corresponding to all
four transmission power levels are captured with the exact same posture along with the
clothing and any other on-body artifacts influencing the RF sigral attenuations.

4.3.4 Multi-scale Variations of Link Quality

The variation of RSSI (in dBm) [75] at the sink node is shown in Figure 4.2. For the
sake of clarity, the RSSI values are shown only for the power levels indices 7, 13 and 22
(24.8mW, 34.5mW and 63.5mW, respectively). These three power levels have been
chosen based on experiments as explained later in this section. The time axes on the
gaphs in Figure 4.2 are divided into postural position slots of width 10 seconds. The
slots are then marked by the posture sequence numbers from 1 through 4 (see Figure 4.1),

which repeats for three cycles.

56

 

 

 

 

 

 

 

 

 

     
   

5

v75 \

: fish
5280 ’r-J

“-85 — . ,.

’4‘."- ..I" .0“; I . '"kno “H'WQ‘RRVUIK/I ‘. a yr LIV”, [ha/"If
'90 Sub-3 ’ PI-7 /' " ' i ‘\ x _
~95 '

10 20 30 40 50 60 70 8O 90 100110120.
Time(s)

 

 

 

 

 

 

Figure 4.2: Link quality variations with varying postural positions
The following observations can be made from Figure 4.2. First, as expected, the link
quality does depend on the transmission power. Second, the RSSI values change quite
significantly with postural positions. For example, the link on Subject-2 with power level

index 13, the maximum sigral strength is -77 dBm in posture 2 of cycle 3, while the

57

minimum sigral strength is -93 dBm in posture l of cycle 3. That represents a
peak—to-peak swing of 16 dBm.

Third, there are distinct differences in the link qualities across the subjects, even with
the same transmission power and at the same posture. For example, using power index
22, the average RSSI for Subject-1’s link is -66 dBm during posture 4 of cycle 3, whereas
for Subject-3, the RSSI during the same posture with the same power level is -74 dBm.

Finally, even for the same subject, with the same transmission power level, the RSSI
can be different for the same postures across different cycles. For example, the RSSI
values for Subject-2 during posture 1 are siglificantly different between cycle-1 and
cycle-3.

To summarize, the experimental results in Figure 4.2 demonstrate that on-body RF
links can vary not only with postures, subject individuals, and transmission powers, but

also with time even when all the above factors are fixed.

 

 

 

 

-60 l l l l l l I l l l I
Position-4\ I . l ;
A'70“ L' I"II!!.
8 , Posrtlon-l . I I ' ' g I 9 ,
e ’ ll ' l - l
:7 '80 7 3 I | I Position-2 ' l
m : I ' | '/ I
Z" w l l ' I I l
I l l I , I
'901 l I l = ' ' I
i l l » - l l i Position-3.
-100 A ’ ‘ i
2 4 6 8 10 l2 l4 l6 18 20 22

 

 

Transmlssion Power Index

 

Figure 4.3: Impacts of transmission power at different postures

Figure 4.3 demonstrates the impacts of transmission power on the RSSI for the four

postural positions shown in Figure 4.1. While a subject remains in each of those

58

positions, the transmission power from the sensor on the upper arm is changed across the
entire available range from level 1 to level 22 (21.5mW to 63.5mW) and the
corresponding RSSI values are noted. At each power level, 20 packets are sent to the sink
which collects RSSI data points for the successfully received packets. The results
obtained demonstrate the spread of RSSI in Figure 4.3 for a given position and at a given
power level. This spread of RSSI indicates that even within a given posture, there are
some intra-posture body movements that affect the on-body RF link qualities.

The results in Figure 4.3 also demonstrate that the minimum power required for a
successful link creation can vary depending on the postural positions. For example, while
in position 4, the link starts delivering packets with power level index of 1, in position 3,
the link starts delivering packets only at a power level index of 8.

The final observation from Figure 4.3 is that for a given postural position, the
average RSSI values can be modeled approximately as a linear function (also reported in
[62],[99-101]) of the transmission power index, although the slope across different
positions can vary. This experimental observation is the key for our proposed power

assigrment mechanism based on postural position inference as presented in Section 4.5.5.

 

 

1.2 . ,
Posrtlon-4

_

g 0.8 — \

Ratio

 

 

.2. \ ‘

'76 0-6 l Position—2 . .

$3 0.4 _ Posrtlon-3

g 0.2 ‘ Position-1

Q 0 I I I I I I f I I I I

 

0 2 4 6 810121416182022
Powerlndex

 

 

 

Figure 4.4: Impacts of transmission power on packet loss

59

Packet losses are recorded from the same experiment corresponding to Figure 4.3,
and are reported in Figure 4.4. This shows that the required transmission power level to
achieve a minimum packet delivery ratio varies across different postural positions. This
implies that for an application with certain amount of packet loss tolerance, the optimal
transmission power can vary depending on the instantaneous posture of a subject.
Therefore, an adaptive transmit power assiglment mechanism is highly desirable for such
on-body links as discussed in this chapter.

4.3.5 Impacts of On-body Parameters on Link Quality

The following experiment was carried out for understanding the impacts of on-body
factors in addition to just the distance between the nodes on a link. Sensor orientation and
on-body postural obstructions for a link between the upper arm and waist sensors have

been targeted as such factors to be understood.

 

Figure 4.5: Variation of sensor orientations and on-body obstructions

As shown in Figure 4.5, four different combinations of the orientation and on-body
obstructions for the link have been experimented with. Orientation of a sensor was
defined by the direction of its antenna alignment, and is indicated by the arrows marked
next to the sensors. The sensor in the picture is too small to clearly show its orientation.

For all four scenarios, while the orientation of the waist sensor was kept constant, the

60

relative orientation was changed by rotating that of the arm sensor. On-body obstruction

of the link has been modulated by having the subject put his hands on his chest.

 

 

 

 

 

-50 l W I I U T n . . . r r
Orlentatlon-l
-60 a. . . Orientation-2 with Obstruction 0 .
A Orientation-1 No Obstruction . . t . l
S _70 _ No Obstruction . 0 ' . . . I
% . . 0 ‘ fl ‘ I |
V . I
l—« . 0 ‘ . 3 3 I I I
U) -80 c 0 A i I I
m t l I I I I
‘34 ' ' l l I l l
90 - I l l I l ' ' I I
- ’ i I | I I I I Orientation-2
, i . I I I I I l with Obstruction
-100
0 2 4 6 8 _10 12 l4 16 18 20 22
Transmrssron Power Index

 

 

 

Figure 4.6: Impacts of sensor orientation and on-body obstructions

The photos in Figure 4.5 demonstrate four different link conditions in terms of the
sensor orientation and obstructions, but for all four, the distance between the sensors was
maintained constant. This control made sure that any link quality variations across the
four scenarios are contributed solely by the changes in orientation and/or obstructions,
and not due to a distance change.

Figure 4.6 presents the RSSI values for different power levels for the experimental
scenarios in Figure 4.5. These results clearly demonstrate that even when the link
distance remains constant, various secondary postural influences such as sensor
orientation and on-body obstructions can affect the link quality as much as the distance
and postural positions do. This conclusion is evident when the amount of RSSI variations

in Figure 4.6 is compared with those in Figure 4.3.

61

 

Figure 4.7: Orientations of a lower-body link with different postures

4.3.6 Validating Linearity of RSSI for Other Postures

To establish the linearity of RSSI vs. transmission power relationship (as observed
for the upper-body link in Figure 4.1) in other parts of the body, we conducted several
experiments on links in different parts of a subject’s body. Figure 4.7 shows such an
experimental setup in which a lower-body link between an ankle sensor and a waist

sensor has been characterized.

 

 

 

 

 

'60 I l l l | T l l l l
. Position-8 (I I I I
-70-i \ x I E i I I I E
a Position-7 g 5; I I I I _ I I .
g 5 x ' I I I I I I
v ~80 _ . x X I E I I I I
a f x A I
a g i 3 I l l . I
I i * i I I ‘ | I Position-5
-90 ~ 1 I . l I I
I I I I I I I Position-6
-100 ‘ L

0 2 4 6 8 10 121416 18 20 22
Transmission Power Index

 

 

 

Figure 4.8: RSSI characteristics power for a lower-body link

62

The RSSI and transmission power characteristics for this link with four postures, as
indicated in Figure 4.7, are shown in Figure 4.8. It is evident that the linearity of RSSI as
observed in Figure 4.3 extends very well for the postures in Figure 4.7. Results from the

links in other parts of body also demonstrated similar linearity relationship.

4.4 Fixed Power Assignments

Each on-body sensor under this design is assigned a static transmit power level for
sending data to the sink node. This power needs to be high enough to achieve a desired
packet delivery rate (see Figure 4.4), while it should be low enough for conserving
energy of the resource-constrained on-body and implanted sensors. Based on the
multi-scale link quality variations as observed in Section 4.3, identifying a power level
that can work in a posture, clothing, and person independent manner can be a challenging
problem.

4.4.1 Relevant Performance Metrics

The performance of transmission power assignment will be evaluated using two
primary metrics, namely, Packet Delivery ratio (PDR) and Energy per Packet (EPP). The
target is to maximize PDR and minimize EPP. PDR is defined as the fraction of packets
sent by an on-body sensor that are delivered to the sink node. The quantity EPP is
computed by dividing the total communication energy, both for transmission and
reception, spent by the on-body sensors (including the sink) by the number of
successfully delivered data packets. Note that the numerator includes communication
energy costs for data as well as any control packets. Although not used for the static
scheme, such control packets are used extensively for the closed loop systems as

described in Section 4.5.1. Note that the EPP metric represents a combined measure of

63

both energy consumption and packet drops. Using EPP, it will be possible to differentiate
between two power assignment strategies for which the total energy consumption may be
the same but one delivers more packets than the other.

A number of secondary metrics, namely, transmission energy, reception energy, and

control overhead, will be also used. The transmission energy cost for a packet is
computed as: E tx = V ' Itx ' S / C , where V, Itx, S, and C represent the supply voltage,

current drawn, packet size, and channel capacity respectively. For the results in this and
in the next section, we have used Vto be 2.5V, and It, to be 9.9mA, 13.8mA and 25.4mA

for power levels 7, 13 and 22 respectively. The data and control packet sizes are 45 and
22 bytes, and the channel rate is 38.4kbps. All these figures correspond to the MicaZMote

hardware [75] including the CC1000 radio chip. The reception costs are computed using
the same formula by replacing the parameter I“, by the receive current I m which is set to

be 8mA [75] for the CC1000 chip.

4.4.2 Fixed Power Assignment Results

The PDR and EPP results are captured for the fixed power assignment at power
levels 7 (24.8mW), 13 (34.5mW) and 22 (63.5mW) with the posture sequence as
described in Section 4.3.2. The experiments were carried out on the same three subject
individuals as in Section 4.3.

Figure 4.9 and Figure 4.10 show the moving average of PDR and EPP within a
sliding window of 10 seconds. Note that like in Figure 4.2, the time axis is divided into
postural position slots of width 10 seconds. The slots are then marked by the posture

sequence numbers from 1 through 4 (see Figure 4.1), which repeats for three cycles.

64

 

l.l

 

1 341,234jlv233;4

2 ' .
0 9- Lied . \..\_LJ
' / / ;

“303 ‘ PI-13 PI-22 ‘x

ch0.7 ~ » \ i

0.6 ‘ Av. for PI-7= 84%
Av. for PI-13=96%

   
  
    
  
 

 

 

 

 

 

 

 

 

 

0'5 ‘ Av. for PI—22=99% Sub-1
0.4 . e r a i .
1.1
1 1 j 2
1 5 ‘x
0.9 — . , , “”~. 5”“ \
n: 0.8 — i ' W“ i
E o 7 1 V i‘“ Y ;_ l
' , f “.‘mm; PI-7
0—6 ‘ Av. for Pl-7=81%
0.5 - Av. for PI-13=93% Sub-2
Av. for PI-22=99%
0.4 I I l l
1.1

 

1 z

0.9 “

M03 ‘ PI-22
a

-7 i E,“
“0.7 ‘ \

0-6 ‘ Av. for PI-7= 79% f Sub-3
0.5 ~ Av. for PI-13=83% Km“? -
Av. for PI-22=97% *
0.4 I l I I I I I I l I l ‘
10 20 30 40 50 60 7O 80 90 100 110 120

Time (3L

 

 
     
  

PI

 
  

 

 

 

 

 

Figure 4.9: PDR performance with fixed power assignment

65

 

 

     

 

 

 

 

 

 

 

 

    

   

 

 

 

1.8
1 2 . 3 4 1 2 . 3 4 1 2 3 4
1.6 4 M - l j l '
E 1-4 ‘ i \ Sub-1
.—. PI—13,Av.: 1.05 mJ/pkt : , PI-22. Av.: 1.54 mJ/pkt
31.2 — . ; ”M; I”...
E 1 ~ '
Lu , XI 0‘ ’ k /
0_8 ‘“ “wu4mtq...«.w'”W»~»>uu«..~~ulM“*”Jl i...” '1’,
J Pl-7, Av.: 0.90 mJ/pkt
0,6 I I I l | I I I I I I
1.8
A 1.6 ‘ M / I /\ ”I
32 1 4 _ .
E - PI—13,Av.: 1.09 mJ/pkt PI—22,Av.:1.57mJ/pkt Sub-2
$ 1 Hf—iu ,9" W?—
0.8 H °"""’""“u«unm;l tuna” . Ewen/f ‘N'IMQH-N_w”'u'l him
\ PI-7, Av.: 0.92 mJ/pkt
0_6 I I I I I I F I I 7 ‘17
1.8
Sub-3
1.6 _W
g 1.4 - PI-13,Av.:1.05 mJ/pkt PI-22,Av.: 1.54 mJ/pkt
H .
51.2 —
Q-r
Q4 1 _
0.8 - .. «4* mum“... ’V
\ PI-7, Av.: 0.90 mJ/pkt
0,6 l I I I I I I I I I I
10 20 30 40 50 _60 70 80 90 100 110 120
Tlme (s)

 

The following observations are to be made. First, as expected, increasing

Figure 4.10: EPP performance with fixed power assignment

transmission power causes larger EPP while providing better communication quality

through higher PDR. Second, since packet drops inflate EPP, every time there is a

66

 

 

 

, HM» an m "Ami
0 8 " We’WW‘rwuu-‘M in“... vies-- ‘m u’

1.8 ‘ 7
l 2 3 4 1 2 : 3 4 1 2 3 4

1.6—W
§I.4~ ’ i \- Sub-1
312 PI—l3,Av.:1.05 mJ/pkt .PI-22,Av.:1.54nu/pkt

' H 1 9 . - We WW":
On
Q 1

 

 

PI—7, Av.: 0.90 mI/pkt
I I

 

 

 

 

 

 

0.6 1 . . I 1 ; 1 I .
1.8
A 1.6 - p... /“‘““\ /\‘1 ‘
5 1-4 * PI-l3 AV'109mJ/pkt PI—22 Av-157mJ/ kt
a , .. . , .. i; p Sub-2
D-u
E 1 — . firm—:-
0 8 )‘H‘*‘{ I (lfiub‘ “/r ‘5‘
' “WI\PI 7, Av.: 0 92 mJ/pkt
0.6 I I I I I T I
1.8

Sub-3

 

LNW
\

1.4 - PI—13,Av.:1.05 mJ/pkt

EPP (mJ/Pkt)
... E;

PI—22, Av.: 1.54 mJ/pkt

 

0.8 H 43"."‘W'5'Nm’nw/fi i 94* ‘ ecu am a k I;
\PI-7 Av.: 0 90 mJ/pkt

um!“

 

 

0.6 I
10 20 30 40 50 .60 70 80 90
T1me (s)

 

100 110120

 

 

Figure 4.10: EPP performance with fixed power assignment

The following observations are to be made. First, as expected, increasing
transmission power causes larger EPP while providing better communication quality

through higher PDR. Second, since packet drops inflate EPP, every time there is a

66

down-spike in the PDR, there is a corresponding up-spike in the EPP results. Third, while
low power levels (e.g. 7 and 13) may be sufficient for 100% PDR for most of the times,
there are occasions during which the packet drops due to body movements can be
unacceptable, and therefore a higher power level, causing larger EPP, needs to be
adopted. Hence a dynamic power assignment is needed for achieving both high PDRs and
low EPPs. As can be seen in the figures, the wide variations of PDR and EPP across

different subjects further reinforce this requirement for dynamic power assignments.

 

 

 

 

 

 

 

 

 

 

  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

r________________'l‘_r_ansmitter Node _ . ______ Receiver Node (Sink/Control)
Data Packet 1 Measure Packet
: Transmission z Receiver RSSI value
at Current
Power Level
i E E I
Data Generator g Current
Th): Power
3 Level
i 5 >~ E :
: 5 '8 i :
5 Update = —<.= E New (RTLMH):
Transmission 5 Power Level 4—— :
| Power Level Computation
I , E Control Packet JNew Power
Receiver "5' """"" "5"" Transmission Level
I ‘ at Full Power

 

 

Figure 4.1 1: Closed loop power assignment model

4.5 Dynamic On-Body Power Assignments

4.5.1 Closed Loop Control

A closed loop mechanism, as shown in Figure 4.11, is adopted for implementing

dynamic on-body power assignments. Based on specific application requirements, data

67

packets are periodically sent from an on-body sensor node to a receiver sink node, which
is also referred to as the control node in the diagram. Each packet is marked with a
monotonically increasing sequence number and the transmit power level information.
Upon receiving a data packet, the control node first decides if the corresponding RSSI
falls within a pre-defined threshold range (between RTL and RT”) in which case the
currently used transmit power level is considered optimal.

If, however, the RSSI falls outside the threshold range, it is perceived that the
transmit power level is too high or too low. In this case, the control node computes a
target transmission power level that should result in an RSSI within the threshold range.
Once computed, this power level is sent back to the transmitter node using a separate
control packet with the maximum possible transmission power. Upon receiving such a
control packet, the transmitter node updates its transmission power and continues packet
transmissions at this new power level until a new control packet, instructing a different
power level, is received. If and when the control node infers a packet loss (by observing
unusual delay in packet reception or gaps in packet sequences) caused by abrupt link
quality deterioration due to postural position changes, the control node assumes that the
current power level being used by the transmitter is insufficient. It sends a control packet
to the transmitter to increase the power level to a value between its current and the
maximum possible value. This process is repeated till the control node starts receiving
packets, and then the RSSI-based control operation as described above resumes. Note that
all the available power levels (1 through 22) are made available during the above iterative

assignment process.

68

4.5.2 Choice of RSSI Thresholds

The RSSI threshold range (RT 1, RT H) is determined based on the packet delivery rate
(PDR) variation as a function of the received RSSI values. To characterize the RSSI vs.
PDR performance, we have conducted an experiment in which while keeping the
transmission power constant, a link’s distance is changed by shifting the receiver’s
position with respect to the transmitter node. At each position, 500 packets are sent, and
the receiver records the RSSI for all the received packets and the number of packets
dropped. From this data, the average RSSI and the corresponding PDR values are

computed and reported in Figure 4.12.

 

 

 

 

 

 

 

 

1.2
1 . RTLL. , RTHL - e-
‘rt: 1* * ' i '
0-8 . : Threshold : d
E 0.6 ' L ”‘33 J. _
a... 1 l‘ ; 'I
0.4 I I -
I I
0.2 _ I . : . ‘ -
O i i i 1 i T' i i 1
-100 -95 -9O -85 -80 -75 -7O -65 -6O
RSSI(dBm)

 

 

 

Figure 4.12: RSSI threshold range for acceptable packet delivery rate

As shown in Figure 4.12, the smallest RSSI that corresponds to near perfect delivery
performance is approximately -88 dBm, which is chosen to be the lower threshold RTL.
The upper threshold RTH is chosen to be -82 dBm for which the dynamic power
assignments, as described in Section 4.6, deliver the best possible energy performance.

More details about the choice of RT" are presented in Section 4.6.2.6.

69

The primary reason for choosing a target RSSI threshold range, as opposed to a
single target RSSI value, is to minimize the frequency of corrective control actions and
its associated energy and capacity overhead. Also, given the 22 discrete power levels for
the Mica2Dot hardware, for a given state of an on-body link, a specific power level at the
transmitter may not be able to produce an exact RSSI value at the receiver.

4.5.3 Target Power Computation by Linear Search

In this approach, the sink/control node computes a new desirable transmission power
level for the transmitter node by linearly incrementing or decrementing the current power
level based on the received RSSI values with respect to the threshold range (RTL, RT”).
No changes are needed when the RSSI falls within the range. The closed-loop control
logic in Figure 4.11 ensures that, for a given postural position, this process eventually
achieves a desirable transmission power level for a link so that the RSSI at the sink falls
within the preset threshold range. The step size for power changes can be chosen based
on the necessary responsiveness.

This linear search-based approach is expected to work well with small incremental
changes in link quality during intra-posture body movements such as moving the hand
slightly while being in any of the positions in Figure 4.1. The linear power level search
during such incremental changes can quickly find a new desirable power level without
having to incur excessive control overhead in terms of control packets from the sink to
the source sensor node. However, for inter-posture body movements, such as when a
subject transitions across the positions in Figure 4.1, this linear approach can be slow and
control expensive due to the large amount of corrective changes needed in the

transmission power level.

70

4.5.4 Computation by Binary Search

Better responsiveness and lighter control for inter-posture body movements can be
achieved via a binary search. In this approach, if the RSSI at sink is lower than the
threshold RTL then the next power level is chosen to be right at the mid-point between
the current and the maximum possible power levels. Similarly, when RSSI is above RTH,
the next level is computed as the mid-point between the current and the minimum
possible power levels. As in the linear case, the closed-loop control logic ensures that the
transmission power eventually converges to a desirable level for the RSSI at sink to be
within the present threshold range.

While providing a lighter control solution for inter-posture movements, the binary
search can be slow and expensive for intra-posture scenarios, in which the transmission
power level has to go back and forth over a wider range even for small incremental
change of the link quality. In other words, for small changes in the link quality, this
search mechanism is likely to overreact and thus takes many steps to converge within the
RSSI threshold range, leading to oscillations and sluggishness. This mechanism suits well
for adults who generally remain relatively still within a posture, but can be expensive for
children for whom generally the intra-posture instabilities are intrinsic.

To summarize, although both linear and binary search mechanisms ensure eventual
convergence to the optimal transmission power level, they are control-expensive and
slow to react for inter-posture and intra-posture body movements respectively. Sluggish
convergence can cost either low PDRs or high EPPs. Transient RSSI levels above RTH
may cause high EPPs while providing perfect PDR. Whereas, RSSI values below RTL

may deliver excellent EPP performance while dropping a large number of packets.

71

Additionally, a slow convergence leads to large number of control packets from the sink
to transmitter, causing the energy per packet numbers to be even higher.

4.5.5 Control with Dynamic Postural Position Inference

The fundamental limitation of the above two mechanisms stems from the fact that no
knowledge about on-body link quality variation is leveraged during the power
assignments. Without such information, the mechanisms amount to blind searches for a
transmission power that can result in a desirable RSSI.

In Dynamic Postural Position Inference (DPPI), the knowledge of on-body link
characteristics is used for inferring a subject’s current postural position for assigning the

best possible power level to a link. The DPPI mechanism is explained in Figure 4.13.

 

Possible New
Posture Position: b 1 ,” Current Posture
I

RSSI

,’ Position: b
B..«'

 

      
   

I
I

G ,I’Possible New
I osture Position: b2

 

 

 

 

I
I

 

 

I
1
1
l
1

iP’ a
Power Level Index

 

 

 

 

Figure 4.13: Posture-driven DPPI for dynamic power assignments
Based on the observed linear relationship between the transmission power and RSSI

in Figure 4.3, we can express:

72

Ri(b)=s(b)°Pi +c(b), (4.1)
where Ri(b) is the RSSI at the sink for a transmission power level P,- (i = I, 2, 22) and

for a body postural position represented as b. The parameters s(b) and C(b) represent the
slopes and intercept of the straight line. The dependency of these parameters on the
current postural position is evident from the characterization graphs in Figure 4.3.

The solid straight line in Figure 4.13 represents a postural position b and a stable
power assignment P,- for which the RSSI value is shown to be within the target threshold

range. Now assume that for the same transmission power level, due a body position

change, the RSSI has changed to a value higher than threshold RT H. In this case, the
dotted straight line above the solid line represents a possible new position b]. The goal is
to infer this new postural position in terms of the slope and the intercept for the
corresponding straight line. Once the equation of the straight line for this new position is

found, a desired transmission power level can be computed by targeting the threshold

mid-point RT to be the desired RSSI.
At the current power level P,», when the postural position changes to b1, the new RSSI

value can be written as:

Ri(bl)=5(bl)°Pi +c(b1). (42)
Upon observing this new RSSI value is higher than the threshold RT H, the sink

(control) node initiate a position inference process by sending a control packet to the

transmitter to send packets at a lower power level Pj. At the sink, this new power level

produces RSSI:

73

Rj(bl)=5(b1)'Pj+c(b1)- (4.3)
Eqns. 4.2 and 4.3 can now be solved for the slope as:
5(b1)=[Ri(b1)"Rj(bl)]/(Pi - Pj),
and for the intercept as: C(b1)=[Pi 'Rj(bl)_Pj 'Ri(b1)l/(Pi — Pj),
These two parameters represent the subject’s current position b 1, which can now be

used for computing a desirable transmission power as:

{RT eig- -R,-<bi>—P,- «twin/(P. —P,-)}
IR.(bt)—R,-<b1)1/<P.- —P,-)}

 

(4.4)

The corresponding operating point is marked as D in Figure 4.13. Upon receiving
this newly computed power level P, the transmitter node continues to use it till a new

value is received from the sink node.

If the new postural position (e.g. b 2) is represented by a straight line that is lower

than that corresponding to the current position (the solid line), then the RSSI will change
to a value lower than threshold RT 1,. The inference and the subsequent power assignment

process still remains the same as what is described above. The only difference is that in

this case, the inference power level Pk needs to be higher than the original level Pi.

Subsequently, the desirable new power level P’ can be computed using Eqn. 4.4 by
substituting the index ‘j’ by ‘k’.

Key novelties of the presented approach includes: 1) adaptive nature with postural
movements and other surrounding changes, 2) inferring postural positions for optimal

assignments, 3) measurement-based as opposed to model-based control, and 4) composite

74

energy accounting via integrating transmission power expenditure with packet delivery

performance.

4.6 Experimental Performance

A subject individual with sensors mounted on his arm and the waist is asked to
follow the posture sequence of Figure 4.1. Each position lasts for 10 seconds, and the
entire sequence is repeated four times, resulting in a total duration of 160 seconds.
Performance of fixed and dynamic power assignments are evaluated using the metrics as
defined in Section 4.5.1. Fixed assignments with four different power levels (power
indexes of 7, 13, 18 and 22), and three dynamic mechanisms, namely, Linear, Binary and
DPPI are compared.

4.6.1 Medium Access Control

An experimental challenge encountered is how to compare all these seven
mechanisms fairly, given that it is almost impossible for a subject to repeat a posture
seven times with the exact same sensor positions along with the clothing and other
on-body artifacts influencing the RF signal attenuations. To address this, we adopted an
experimental technique in which all seven assignment mechanisms are executed
independently with a span of one second cycle (see Figure 4.14), and then the cycle is
repeated continuously for 160 seconds. Considering the minimum time constant of
posture changes is of the order of several tens of seconds, this mechanism ensures that all
seven mechanisms are executed under the exact same posture and other influencing

conditions.

75

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

I Data packet from on-body transmitter to on-body sink _ i
Data packet from on-body sink to out-of-body server T1me : . . .
[In] Control packet from on-body sink to on-body transmitter ’ 5
Linear t-H Binary r-H rpm #4. - . - 2993 293's 30 595’:
L'ZL’. _, ”:1 r; : _
_ ESE-7T , - , .. . ,7 I" . f
\ Y # L Y J L Y J \ I \ , , J g ,__l L— ,_J
150 ms 150 ms 150 ms 100ms 100ms 100ms 100ms 150 ms

 

e 1000 ms Control Cycle 9?

 

I

I
Fixed-7 Fixed-13 Fixed-18 Fixed-22 : Linear

I

I

Figure 4.14: Multiple control schemes with synchronized postures

As shown in Figure 4.14, for every power assignment mechanism, there is a slot
within the one second control cycle. Each such. slot consists of three different types of
packets: 1) data packet from the transmitter sensor to the sink, 2) information packet from
the sink to off-body server for data collection and processing, and 3) control packet from
the sink to the transmitter. This last packet type is not needed for fixed assignments. The
sink node executes the power assignment, measures the RSSI and drop statistics, and
reports them to the off-body server for each of the seven evaluated mechanisms
separately. Appropriate slot duration and guard times between slots have been chosen to
avoid any packet collisions. Note that CSMA, the default MAC protocol implemented in
Mica2Dot’s [78] TinyOS [77] operating system, is used for all the reported experiments.

Also note that the inference power level indexes j and k in the DPPI mechanism in
Section 4.5.5 are chosen to be immediately below or above the index 1' respectively. In

other words, the conditions j 2i —l and k =i +1 are always maintained for the presented

results.

76

4.6.2 Results and Interpretation

Although four different fixed power levels are experimented with, due to space
constraint, we report only the results for power indexes 22 and 13, which produce the
best PDR and EPP respectively among all the fixed assignments.

Figure 4.15: Power level and RSSI dynamics with posture changes

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

24
20
A N
E 16 0
a E
: 12 t3
”3 3
a 8 s3
4
0
24
Packet Loss 20
A Power Index RSSI ><
E 16.3
53 a
r: 12 is
V’ B
32 8 s3
4
0
24
Power Index
-60 — ‘" 20
A _ 16 X
:7 ‘ + 12 is
g -80 q R ‘w I“ WA _ 8 E
“41141.11 1.1 m in 1 ..
-90 — u! I . Ir' / \1 + 4
Thr. ange L‘nea' RSSI
— 0
10 20 30 40 50 60 70 80 90100110120130140150160
Time (s)

 

77

Figure 4.16: Continued

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

-50 1 1 W I W r I f F 1 f 1 1 T r 24

-60 — Power Index _- 20
A K
5 -70 r 1 ‘L 16%
v SI I I f" “11" 1225‘
I-I . 1' ‘I ‘ H I I o
g -80 .-. ....... .I ........... * I ................... ‘ 5 g

.I I III I I i“ .... I with .16", I" 8 g,

-90 - " f .

400 Thr. Range

-50 t I I

-60 _ Power Index
”a?

g -70 ‘
g -80.. r
94
_90 .2 .............
-100
’50 I I I I I I I l I I r 1 1 I 24
Benchmark '

-60 - M I 20
A Power Index - .. , 5
a ._ 16::
Q '70 ‘ \F.“ a
L“ * III I 11 "12%
w -80 - , o
3 LJ r 8 an

-90 _ Thr L3,, RSSK T 4

-100 ° 58 0

 

 

 

10 20 3O 40 50 6O 7O 80 90100110120130140150160

Time (s)

 

78

 

4.6.2.1 Transmit Power and RSSI Dynamics

Figure 4.15 demonstrates the evolution of assigned power levels and the resulting
RSSI for two fixed assignment (power indexes 22 and 13) and three dynamic assignment
(Linear, Binary and DPPI) mechanisms. A best case scenario, termed as the Benchmark,
is also defined for each control cycle (see Figure 4.14). The benchmark results are
constructed by selecting the lowest possible transmit power from all seven implemented
mechanisms, for which a packet was delivered to the sink. This is computed offline and
provides an experimental performance upper bound to compare DPPI’s performance
with. The following observations should be made from Figure 4.15.

First, with fixed high power (e. g. PI=22) assignments, the RSSI at the sink is mostly
above the higher threshold RTH, which is expected to provide excellent packet delivery
rate (PDR), but at the expense of high Energy Per Packet (EPP). With fixed low power
(e. g. Pl=13) assignments, more often the RSSI remains within the desirable threshold
range, causing the EPP performance to be better. But as shown in Figure 4.15, packet
losses during link-obstructive postures can affect the PDR performance. These can be

validated in the time averaged results presented in Table 4.1.

 

 

 

 

 

 

 

Power Av. Av. EPP Tx Energy Rx Energy
Assignment lPDR (mJ/pkt) (mJ) (mJ)

Index-13 83% 1.37 118 39
Index-22 99% 1.59 203 43
Linear 92% 1.32 132 52
Binary 94% 1.09 1 16 49
DPPI 98% 0.95 103 46
Benchmark 99% 0.87 94 43

 

 

 

 

 

 

 

Table 4.1: Performance summary of different power assignments

Second, all three dynamic assignment mechanisms manage to maintain the RSSI

values within the threshold range longer, indicating a better balance between PDR and

79

EPP. Also, given that the subject’s movements were mostly inter-posture as opposed to
intra-posture, the binary search worked better than its linear counterpart. However, for
binary, the oscillation (see Section 4.5.4 for explanation) of the transmission power level
and the subsequent departures of RSSI out of the target threshold range are quite evident
in several instances including those during the time frames 105-205 and 1103-1303.

The DPPI mechanism performs the best by avoiding this oscillation problem via its
postural position inference-based assignment as described in Section 4.5.5. Finally, since
the benchmark corresponds to the minimum transmission power level for a successfully
delivered packet across all seven experimented assignments, it’s RS SI values are always
the lowest and often dwell below the lower threshold R TL.

4.6.2.2 Packet Delivery Ratio (PDR)

Figure 4.17 shows the moving average of PDRs, for the DPPI and binary
assignments, computed using an averaging sliding window of 10 seconds. As shown in
Table 4.1, these two assignments have the best two PDRs among the dynamic
mechanisms. Observe that there are several time durations (e. g. lOs-20s) when the binary
mechanism loses packets, whereas the DPPI is able to deliver them. A closer look to the
RSSI values in Figure 4.15 for that time duration reveals that the binary assignment was
dropping packets due to an oscillation in the assigned power level and the subsequent
RSSI values beyond the desired threshold range. The oscillation started soon after a body
position change by the subject at around 10 seconds time mark. Departures below the

lower RSSI threshold during such oscillations resulted in packet dr0ps.

80

 

 

 

 

 

 

 

 

 

 

 

1.1 f g 3
1 , é ;
0.9 _ R ,‘ U \I-n-JI I U
a: 0.8 - / i I I
DPPI ‘ . ’
E 0.7 — ¥ Binary:
0.6 - i * ‘
0.5 —
0.4 I I I I I I r I I T j I I I I
o 10 20 30 40 so 60 7o 80 9o 100110120130140150160

Time (s)

 

 

 

Figure 4.17 : PDRs for DPPI and binary search-assignments

For DPPI, however, packets are not dropped because the RSSI values in those time
durations are well within the threshold range. This was accomplished by a quick
inference of the new position (i.e. the correct straight line in Figure 4.13) soon after the
subject’s posture change, and then by adjusting the new power values based on the
inferred postural position. As shown in Table 4.1, the resulting PDR is a very high and
close to the best case benchmark performance.

The results in Figure 4.18 show the impacts of posture transitions on the packet
delivery performance. The figure shows the number of consecutively lost packets as a
function of time. It can be observed that fixed power assignment using PI-l3 experiences
higher packet losses during posture transitions than the dynamic assignment schemes.
This is due to the possibility of long periods of link disconnection, resulting from posture
changes (e. g. a transition between postures 3 and 4 in the second, third, and fourth
posture cycles). In contrast, dynamic power assignment schemes experience lower
packet loss in general. DPPI is less affected compared to the binary search mechanism,
mainly because of its ability to generally better estimate the power level required for the

next packet transmission.

81

 

 

\D

I

"U

'T‘

p—A

t»

"" 4:.
I. A

No. of consecutive packets lost

 

 

 

“1L, $111141!

10 20 3O 4O 50 6O 7O 80 90 100 110120130140150
Time(s)

 

 

 

Figure 4.18: Packet drops during posture transitions

4.6.2.3 Energy Per Packet (EPP)

Figure 4.19 shows the moving average of EPPs, for the DPPI and binary
assignments, computed using an averaging sliding window of 10 seconds. It can be seen
that energy performance of DPPI is consistently better than binary search. The energy
advantage of DPPI is contributed by: a) lower transmission power levels (see Figure
4.15), lower packet drops (see Figure 4.17), and smaller closed loop control overhead, as
explained in the next subsection. 1

In addition to the average EPP figures, the last two columns in Table 4.1 report the
cumulative total transmission and reception energy expenditures, for both data and

control packets, during the entire experimental period of 160 seconds.

82

 

 

I"
N

 

Binary

% \ DPPI

.1 III/V

g—n
00
l

 

  

EPP (mJ/Pkt)
I

r
o
I

 

 

0.61 I I 1 I I I I 1 1 rfii,rrr r

0 10 20 3O 40 50 60 70 80 90 100110120130140150160
Time(s)

 

 

 

Figure 4.19: EPP for DPPI and binary search assignments

4.6.2.4 Energy Overhead of Closed Loop Control

The energy costs for the closed loop controlled assignments manifest in the form of
transmission and reception of the control packets from the sink to the transmitter nodes.
The cumulative control packet communication cost over time is shown in Figure 4.20.

Note that this control communication energy for the Linear assignment is the
maximum among the dynamic mechanisms. This is simply because of its slower
convergence and the subsequently frequent control activities in the events of inter-posture
movements. It can be verified from Figure 4.15 that the total count of power assignment
changes for the Linear is significantly more than the other two dynamic mechanisms.

Although the situation is better for binary search, it also has to send quite a few
control packets during the oscillations followed by the postural changes by the subject.
The control packet communication costs for DPPI, however, are smaller than the binary,
thus indicating fewer control packets from the sink to the transmitter. Faster convergence

via postural position inference accounts for this control overhead reduction. This low

83

control energy overhead of DPPI also contributes to its excellent EPP performance as

shown in Figure 4.19.

 

 

A
O
1

 

Cont. Pkt. Energy (ml)

 

 

 

 

H N DJ
O O O O
1 L 1 1

I I I I I I I

O

20 40 60 100 120 140 160

Tim8e0( s)

 

 

 

Figure 4.20: Cumulative control energy expenditure over time

4.6.2.5 Results for Different Subject Individuals

Time averaged PDR and EPP results for three subject individuals, one female and
two males, are summarized in Table 4.2. All subjects go through the same experimental
process as described in Section 4.6.1. Results indicate that the PDR and EPP performance
advantage of DPPI holds for all three subject individuals with different body statures and
clothing outfits. Although it is possible to get a slightly better PDR using high power
fixed assignments (e.g. power level index of 22), the energy cost for that can be very
high.

Observe that the binary assignment does better than the linear assignment for
Subject-1 and Subject-3, but it is slightly worse for Subject-2 in terms of PDR. Our
investigations revealed that Subject-2 was more active within postures, causing the liner
assignment to handle those intra-posture movements better than binary, for which control
oscillations are triggered by such small body movements. The PDDI scheme, however, is

able to handle such cross-subject habitual variations and provide the best performance

84

consistently. Also note that the performance of PDDI is closed to the best case

benchmark performance (both PDR and EPP) for all three subject individuals.

Power
Assignment PDR EPP PDR EPP PDR EPP

1 1
1

 

Table 4.2: Power assignment performance for multiple subjects

To summarize, by using its unique ability of postural position inference, the DPPI is
able to optimally assign on-body link power in a subject-independent manner while
consistently maintaining better balance between packet delivery and the incurred energy
cost. We have conducted similar experiments with different on-body sensor positions and
posture sequences, and found that the above performance trends also hold in a position

and posture independent manner.

 

 

l .3
Control Scheme: Dynamic Postural Position Inference

    

EPP (mJ/Pkt)
_ E: E;

.o
\O
1

 

 

_O
00
1

-87 -86 -85 -83 —82 -81 -80 -78 -77 -76 -75 -74 -72 -71 -70
RSSI Upper Threshold (dBm)

 

 

 

Figure 4.21: Energy expenditure with varying RSSI thresholds

85

4.6.2.6 Impacts of RSSI Thresholds

In order to choose an appropriate RSSI threshold range, the DPPI mechanism was
characterized (see Figure 4.21) with varying values of the upper threshold RT H, while
keeping the lower threshold fixed at -88 dBm. With increasing RT H, the transmission
energy cost is expected to increase because of higher transmission powers. With this, the
packet delivery ratio is also expected to improve. Therefore, the changes in Energy Per
Packet (EPP) numbers will depend on which of the above two effects dominate.

Figure 4.21 demonstrates that initially with increasing RTH, the EPP reduces because
the increase in PDR (and decrease in the number of control packets) dominates the
increase in transmission energy costs. However, beyond approximately -82dBm, that
relative dominating effect reverses, causing the EPP to start increase. This explains why
-82dBm, which provides the best EPP performance, has been chosen as the RT3 (see
Section 4.6.2) for all experiments.

4.6.2.7 Results for Different Pastures

A subject individual with sensors mounted on his arm and the waist in position 1, as
shown in Figure 4.1, is asked to follow the posture sequence of: {Sit (SIT), Lie-Down
(DWN), Stand (STD) and Walk (WLK)}. Each posture lasts for 10 seconds, and the
entire sequence is repeated four times, resulting in a total experiment duration of 160
seconds. The RSSI threshold range was chosen as before to be -88 to -82 dBm. Note that
the power assignment mechanisms discussed in this chapter do not make any assumption
about the posture durations. However, as a general rule, when the ratio of the interval
between posture changes to the response time is low, power control mechanisms are not

very beneficial.

86

Figure 4.22: Power levels and RSSI dynamics with posture changes, for different
power assigmnent strategies

 

 

 

 
   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

'50 srr DWN STD WLK SIT DWN STD WLK SIT DWN STD WLKVSIT pWN STD WLK — 24
-60 - ‘ 20
A K

O
540 — . t 16 g
:7 Power Index ‘ 12 B
3.80 - 5
of. / —— 3 13..
'90 J V - __ 4
Thr. Range PI: 22
-100 0
-50 l I I I I I I T I I I I I I I 24
'60 I RSSI Power Index Pack tL0$ ' __ 20
E I ' . ~— 16 g;
a") ~ in m ‘ a. refs - E
r: : - I “‘ 12 s
m , B
-80
E III“! IFJIM :rNIuI r 8 33
hi" I I : ' II 'I)‘
-90 — f/ U - 'V 4
Thr. R ge , PI: 13
400 10 20 30 40 50 60 7Q 80 90 100 110120 130140 1501600
T1me (s)

 

 

87

Figure 4.23: Continued

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

I 7 I I I I 24
r- 20
I .1
16% I
~ 12.: I
6
........ :L‘ 8 m
11'
, s- 4
0
r r 1 1 1 1 24
' Oscillation
. I . / ’ I " 2°
A . ' M“ “ l6 ><
E E IIII f I I II I ‘I 121;
(0-30 71- 1" {a ,,,,,,, . z . .1 , ........... , _ o
I ‘1‘ ,. I . WWI ........ I. I 1 ............ .8 °‘
-90 - I \1‘ "* * : I , / 4
_50 I I I I T I 1 T I r I I I 24
-60 I Power Index “' 20
A ‘ 16 ><
£401 3
: 12 15
m_ _ B
9 8° ......... ~- 1 8 ‘3
-90 - INT g ~ 1 ‘ ’ 4
\ \ ‘ I
Thr. Range . DPPI , . { RSSI IJ .
'100 10 20 30 40 50 60 70 .80 90 100 110120 130140 150160 0

T1me (s)

 

 

Figure 4.22 demonstrates the evolution of assigned power levels and the resulting

88

RSSI for two fixed assignment (power indexes 22 and 13) and three dynamic
assignments (Linear, Binary and DPPI). Same observations and conclusions can be made
as in Section 4.6.2.1 and Figure 4.15. First, with fixed high power (e.g. PI=22)

assignments, the RSSI at the sink is mostly above the higher threshold RTH, which is

expected to provide excellent packet delivery rate (PDR), but at the expense of high
Energy Per Packet (EPP). With fixed low power (e.g. Pl=13) assignments, more often the
RSSI remains within the desirable threshold range, causing the EPP performance to be
better. But as shown in Figure 4.22, packet losses during link—obstructive postures can
affect the PDR performance. These can be validated in the time averaged results

presented in Table 4.3, which summarizes results for a set of indoor and outdoor

 

 

 

 

 

 

 

experiments.
Indoor Outdoor
Power Av. Av. EPP Tx Rx Av. Av. EPP Tx Rx
Assgn. PDR (ml/pk!) Energy Energy PDR (mJ/pkt) Energy Energy
("11) (M (In-l) (In-I)
Index-l3 82% 1.05 112 37 84% 0.98 109 38
Index-22 96% 1.34 205 41 96% 1.29 198 40
Linear 89% 0.83 126 53 90% 0.74 106 49
Binary 91% 0.74 110 49 91% 0.69 98 46
DPPI 94% 0.66 96 43 95% 0.60 87 42

 

 

 

 

 

 

 

 

 

 

 

Table 4.3: Indoor and outdoor performance summary of different power assignments

Second, all three dynamic assignment mechanisms manage to maintain the RSSI
values within the threshold range longer, indicating a better balance between PDR and
EPP. Also, given that the subject’s movements were mostly coarse-grain inter-posture, as
opposed to fme-grain intra-posture, the binary search worked better than the linear.
However, for binary, the oscillation of the transmission power level and the subsequent
departures of RSSI out of the target threshold range are quite evident in several instances
including those during the time frames 308-408 and 1403-1603, where the subject is in a
high motion posture of Walk. Finally, the DPPI mechanism performs the best by
avoiding this oscillation problem via its postural position inference-based assignment as
described in Section 4.5.5. Table 4.3 summarizes the average of all the performance

indexes for all the assignment mechanisms. While the results show that dynamic

89

mechanisms perform better than static for both outdoor and indoor experiments, it can be

observed that the overall energy performance is slightly better outdoors compared to

indoors. This can be attributed to the fact that dynamic mechanisms use fewer control

packets outdoors due to better channel conditions as there are fewer multipaths.

 

 

 

 

 

  

 

  

 

  

 

 

   

 

 

SIT DWNSTD WLKSIT DWNISTD WLK SIT DWN STDWLK SIT DWN STDWLK
m 0.9 “
D I I .
D-a I _ i f
0.8 ~ : '
/ i \DPPI
0.7 ‘ Binary j ‘ I
06 l r I I I I T I I I T I I I I
O 10 20 3O 40 50 6O 7O 80 90 100110120 130140150 160
Time (s)

 

 

 

Figure 4.24: PDR for DPPI and binary search assignments

Figure 4.24 and Figure 4.25 show the moving average of PDRs and EPPs

respectively, for the DPPI and binary assignments, computed using an averaging sliding

window of 10 seconds. Note that the number of control bits used to provide feedback is

slightly higher (i.e. 4 bits) for DPPI compared to Linear and Binary mechanisms. This

difference is already accounted for in the EPP metric. Also, for protocols with relatively

large preamble size (e.g. 144 bits for 802.11), the additional energy cost due to those

extra bits can be considered insignificant. As shown in Table 4.3, these two assignments

have the best two PDRs among the dynamic mechanisms. Observe that there are several

time durations (e. g. 105—30$) when the binary mechanism loses packets, whereas the

DPPI is able to deliver them. A closer look to the RSSI values in Figure 4.22 for that time

duration reveals that the binary assignment was dropping packets due to an oscillation in

90

the assigned power level and the subsequent RSSI values beyond the desired threshold
range. It can be seen that the energy performance of DPPI in Figure 4.25 is consistently
better than the binary search. The energy advantage of DPPI is contributed by lower
transmission power levels (see Figure 4.15), lower packet drops, and smaller closed loop
control overhead. In addition to the average EPP figures, the last two columns in Table
4.3 report the cumulative total transmission and reception energy expenditures, for both

data and control packets, during the entire experimental period of 160 seconds.

 

 

     

 

 

 

 

 

1'4 SIT DWNSTD WLK srr DWNSTD WLK srr DWNerWLK’srr DWNSTDWLK
1.2 d I 4' : ~ ' I .1
g , : Binary, Avg: 0. 74 mJ/Pkt .
LE 1 _ . DPPI: Avg. 066 mJ/Pkt * =
:03. .~ \ I‘r
a. . ; I.
. » ’\ I
0.6 — , . \ L
0.4 I I i I I I I I I I i i I I I
0 10 20 30 40 50 60 70 80 90 100110120 130140 150160
Time (s)

 

Figure 4.25: EPP for DPPI and binary search assignments

4.7 Summary and Conclusions

A closed loop transmission power assignment framework for on-body wireless links
has been developed in this chapter. Postural positions as it pertains on a given wireless
link are dynamically inferred based on measured RF signal strength and packet drops on
a link by link basis, and then optimal power assignment is done by fitting those
measurement results on a model describing the relationship between the assigned power
and the resulting signal strength. Extensive experimental results are provided to

demonstrate the model building and algorithm performance on a prototype body area

91

network. It has been shown that power assignment with the proposed Dynamic Postural
Position Inference (DPPI) can provide consistently better packet delivery and energy
performance compared to a range of static and other dynamic assignment strategies. It
has also been compared with a performance benchmark established by sampling the best
operating points across a number of static and dynamic power assignment strategies. The
experimental results show that DPPI, using its novel postural position inference strategy,

can deliver subj ect-independent performance that is closed to the experimental

benchmark.

92

Chapter 5. Posture-Predictive Power Control using Linear-Quadratic

Gaussian Control

It was shown in Chapter 4 that the optimal transmission power required for an
on-body wireless link between two sensors depends on the physical distance of the link,
and its instantaneous channel condition. The distance can vary with mobility, driven by
human postures, and the channel condition can change due to unpredictable RF
attenuation [63] caused by a slew of factors including antenna orientation, clothing, and
physical stature of specific subject individuals. As a result, static pre-defined
transmission power is not able to provide continuous link connectivity while ensuring
minimum required transmission power consumption. Therefore dynamic link power
assignment mechanism for optimal on-body energy management is proposed.

In this chapter, the dynamic transmission power control mechanism presented in
Chapter 4 is enhanced by modeling human body movement as a stochastic linear system
and a quantized Linear Quadratic Gaussian control with an Integrator (LQGI). The
objective is to develop a model-based transmission power control framework in which
RF signal strength is predicted and is regulated at a reference value to enhance the overall
energy performance of an on-body wireless sensor.

The LQGI framework has been a popular way to design optimal stochastic control
for various applications [102-107] such as aircraft control, adaptive optics systems and
MEMS camera control. However, to the best of our knowledge, there have been no
attempts to use the LQGI approach to the on-body transmission power assignment as

attempted in this work.

93

LQGI is used for regulating the Radio Signal Strength Indicator (RSSI) of the
receiver node at a fixed reference level for an on-body link. We modeled the RSSI
variation process, which consists of its relationship to the transmission power and the
body configuration process, as a linear stochastic system. In this LQGI approach, a
Kalman filter estimates the state of the system. Simultaneously, a linear quadratic
regulator with an integrator, based on the state estimate fiom the Kalman filter, provides

optimal power control.

5.1 LQGI Framework

5.1.1 Problem Formulation

In this section we present the proposed LQGI-based posture prediction and
transmission power assignment mechanism. To implement LQGI, the RSSI process and
the body posture processes both are modeled as linear, stochastic systems. Both processes

are shown as an augmented system in Figure 5.1.

 

xq (k +1) = Aqxq (k) + qu(k) q(k)

6100 = quq(k) + W(k)

P1 0 Q y<k>

36(k +1) = a3?(k) + bu(k)
70c) = cY(k)+du(k) rm

P2 v(k)

W(k)—-—'

 

 

 

 

 

u(k)

 

 

 

 

 

Figure 5.1: Block diagram of the augmented system

P1 in Figure 5.1 represents the posture change process and P2 denotes the RSSI
process. The discrete-time state-space representation of the augmented system in Figure

5.1 is given by

94

x(k +1) = Ax(k) + Bu(k) + Gw(k)

Y(k) = Cx(k) + du(k) + w(k) + v(k) (5.1)

where A, B, G, C and d are state matrices, and w(k) e 92 , v(k) e 93 are white noises with
the following properties:

E(W(k)) = 0, E(W(i)W(j)) = Q and E(V(k)) = 0, E(V(i)V(J')) = R
respectively, with E (w(i)v( j)) = 0 , where the term E(x(i)) is the mean value of variable x,
and E (x(i)x( j)) is the covariance of variable x.

A discrete-time Kalman filter is designed with the steady-state Kalman filter gain as:
MS =23CT(CXSCT+R)_1 ' (5.2)
where 2 s [102] is the solution to the following algebraic Riccati equation:
2, = AZSAT + GQG T — AZSCT (CZSCT + II)"1 CESAT

The tracking error is defined as 3(k)=J’(k) -y,ef where y ref 6 ER is the

reference RSSI level. An integrator is implemented to eliminate the steady-state tracking
enon

The incremental tracking error integral is given by:
I (k +1) = I (k) + e(k) = I (k) + Cx(k) + du(k) — yref (5.3)

and it is desired to regulate the control action. The augmented system and integrator is

reformulated as a linear quadratic regulator. This introduces a new state equation:

xtrack (k + 1) = Atrackxtrack (k) + Btrack u (k) _ yref

(5.4)
ytrack (k) ___ Ctrack xtrack (k)

95

The delay operator is denoted by q‘1 such that q—lg(k) = g(k — 1) . By multiplying

each side of the xmwk (k) state equation by the difference operator (1 — q 4) we can

obtain:

x(k+1)— x(k)]_ {w(k)} track

—1 track
(1- q )x (k+1)=[1(k+1)_1(k) g(k) —xd (k) (5.5)

Define (1— q " )u(k) = v(k — 1) . We can obtain:

xtrack k nxn OnXl:|[-xd(k_ 1)]+ +|:B-q]l —l k 1_
()= [CW 1 80(4) d(‘1 )u()- ( q >ny (5.6)

=Atrack xtrack(k_ l)+BtraCkv(k— 1)

track

and yd (k-l) is given by:

ytrack(k_ 1): (1_ q—l)ytrack(k)_ (1_ q—1)Ctrackxtrack (k)
___: dthrack track (k- 1)

The overall performance cost function can be written as

”v.2:

 

 

track (k)ll2 +r2"v(k)“2 )

(5.7)
= 205d track (k)TQtrack xg‘ack (k)+ v(k)TRtrack v(k))
k0

T
Qtrack : Ctrack Ctrack and Rtrack

where = r2 > 0. The quantityris a weight

factor on control v(k), i.e, difference between u(k+l) and u(k). The optimal control that

minimizes the overall performance cost function J in Eqn. 5.1 is given by

96

k—l
u(k) = —K"’“ckx”ack (k) = —Kax(k) — Ke Ze(k) (58)
k0

where the optimal control gain matrix is given by

T T
Ktrack = [Ka Ke]= (Rtrack + Btrack PBtrack )—1 Btrack P Atrack where P is

3

the solution to the following algebraic Riccati equation

AtrackTPAtrack _ P- AtrackT P Atrack Ktrack + Qtrack = 0

5.1.2 Quantized LQGI

In this approach, saturation and quantizer blocks are added to the LQGI system, as
shown in Figure 5.2, in order to generate control (or transmission power u) from a finite
set of discrete levels. Figure 5.2 shows quantized LQGI with “S” and “Q” denoting
saturation and quantizer respectively. The absence of both the saturation and quantizer
blocks is just the LQGI system. “KF” in Figure 5.2 denotes the Kalman filter. The
saturation block applies an upper (5 dBm) and lower (-20 dBm) limit on the control input
signal. The quantizer block rounds the control input signal at each period of time to the
nearest discrete value.

As in Eqn. 5.7, r is a weight factor on control v(k), i.e, difference between u(k+l)
and u(k). The quantized LQGI (LQGIQ) approach is used to regulate how frequently the
controller changes the transmission power level, and subsequently, it regulates the
number of control packets sent for a period of time. For lower values of r , there are less
penalty on the differences in control, and more penalty on the tracking error as in Eqn.
5.7, therefore, the power level changes more frequently, and the tracking error decreases,

respectively. On the other hand, for higher values of r , there is more penalty on

97

differences in control efforts, and less penalty on tracking error, hence, the transmission

power level changes less frequently and the tracking error increase.

“2

 

y ref

 

6(k) ,

| H

 

 

 

 

 

“r ya, (k)

 

 

 

y(k (k) +

 

 

£(k)

 

 

Figure 5.2: Block diagram of quantized LQGI Control

5.2 Performance Results

A subject individual with sensors mounted on his arm and the waist was asked to
follow the four-posture sequence as shown in Figure 5.3. The movement between posture
positions is estimated to take 10 sec, and the entire sequence is repeated nine times,
resulting in a total duration of 360 sec. Sensor on the arm transmits two packets every
second, one with the linear search power assignment and the other with the binary search.

The linear and binary search mechanisms were described in Section 4.5.

98

     

‘~~ \
‘~

   

On-body Link

   
  

Wireless Lir‘ik‘x‘}~
On-body Out-of-Body Server
Sensors

 

Figure 5.3: On-body sensor positions and the used posture sequence

Results from these two approaches are collected experimentally from the prototype
WBAN as described in Chapter 2 and Section 4.3.1. Data for the transmission power and
the corresponding RSSI values are also collected during these liner and binary search
control experiments. These data, collected through the off-body server (see Figure 5.3
and Chapter 2), is then used for simulating different LQGI control approaches as
presented in Section 5.1. The simulated LQGI variations include: LQGI with r=2
(LQGI_2), LQGIQ with r=2 (LQG10_2) and LQGIQ with r=700 (LQGIQ700), where
LQGIQ denotes the quantized LQGI approach, as defined in Section 5.1.2. The weight
factor r’s values are chosen to be very small and very large for representing the low and
high energy penalties due to the control packets from the sink to the source sensor. The
five control approaches (two experimental and three simulated) are evaluated and
compared using the same parameter settings as. in Chapter 5.

5.3 Relevant Performance Metrics
The performance of transmission power assignment is evaluated using two primary

metrics, namely, RSSI Target Error (RTE) and Energy per Packet (EPP), as described in

99

Chapter 4. The target is to minimize RTE and EPP. RTE is defined as how far the current
RSSI from the target RSSI. As presented in Section 4.5.2, the RTL and RTH are chosen to
be -88dBm and -82dBm for the linear and binary searches, and ymf for the LQGI
approaches is chosen to be right in between RTL and RTH which is ~85dBm. The quantity
EPP is computed by dividing the total communication energy, both for transmission and
reception, spent by the on-body sensors (including the sink) by the number of
successfully delivered data packets. Note that the numerator includes communication
energy costs for data and control packets. A number of secondary metrics, namely,
transmission energy, reception energy, and control overhead, as described in Chapter 4,

are also measured and evaluated.

5.4 Results and Interpretation

5.4.1 Transmit Power and RSSI Dynamics

Figure 5.4 demonstrates the evolution of assigned power levels and the resulting
RSSI for linear, binary, and the three LQGI-based power assignments. The linear and
binary search mechanisms are described in Section 4.5. Power values of CC1000 from
-20 to 5 dBm correspond to the power levels from I to 22 are used in the figures. The
following observations should be made from Figure 5.4.

First, although for the majority of the time the linear and binary search approaches
manage to maintain the RSSI within the target threshold range, it fails to be in the range
for a significant fiaction of the experimental duration. As explained in Section 4.5.4 and
Figure 4.15, the linear approach suffers fi'om the slow stepwise assignment, while the

binary suffers from the oscillation problem due to overreactions. The resulting departures

100

of RSSI out of the target threshold range are quite evident in several instances including
those during the time frames 405-703, 1203-1503 and 2605-2805.

Figure 5.4: Power level and RSSI dynamics with posture changes

 

 

 

 

 

 

 

 

 

    

 

 

 

 

15
A RSSI\
8 Thr. range -70
a 5 I k
I: II] A
3 r 5
- m
a WWW 1W\ 80“”
G I—l
.° I w
I, a
I-.. I V -90
g Tx Pow/er‘xv
I" Linear
-25 -100
15
E RSSI Tx Power
,3 -70
'3
IS A
E a
g. -80 a
g v
.2 a
.3 a
a-rs - -90
E
‘" Binary
-25 -100
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360
Time (s)

 

 

 

101

Figure 5.5: Continued

 

 

 

 

 

 

 

 

 

  
  

 

 

 

 

 

 

 

 

 

 

 

 

 

15
’é‘ Tx Power + -70
m
3 5 ‘
E a
o -80 m
a. 'u
I: '5 “ I.“
.9 m
3 3
g-IS — -90
.3.
E-‘

-25 -100
A 15
CE T Pow
3 x er -_ -70
2 5 ‘
o E
°‘ L -80 m
.5 -5 3
a g
g - -90 ad
§-15 ~
[—

LQGIQ_2

-25 -100

15
5'3 -70
2’ 5 1
0 A
E I -80 5,
an “o
a -5 ‘ :
.2 m
m (I)
."’ Tx Power of.
g . -~ -90
m -15
g LQGIQ_700

-25 -100

0 20 4O 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360
Time (s)

 

 

Second, it can be seen from the LQGI_2 plot in Figure 5.4, that there is a need for
modifications to the control, due to the transmission power not being constrained to the

transceiver RF power output of —20 to 5 dBm. Therefore, LQGIQ control in Figure 5.2 is

102

implemented. There are two significant changes as a result of the quantized control: 1)
the transmission power in LQGIQZ plot of Figure 5.4 meets the constraints of the
transceiver; 2) the “smooth” signal on LQGI_2 becomes “stair-step” as shown in
LQGIQZ. The penalty r is kept the same in both approaches to show the effects of the
saturation and quantizer blocks on the control.

Finally, LQGIQ_700 plot can be compared with LQGIQ_2 plot to show the effects
of r on LQGIQ control. Notice that when r=2, the controller changes its transmission
power level quite frequently. While LQGIQ700 plot shows when the penalty r is
increased, the controller changes its transmission power level less frequently. The penalty

values of r=2 and r=700 are chosen so that the reader can easily observe this contrast.

    
 
 
      

 

 

 

 

.9
Ur

    

0.0109
\ \

LQGI LQGIQ LQGIQ
2 2 700

O

I
Exp. Results : Sim Results
2.5 g
I
m g m RTE, SD
'31 5 :
F3 I
E”: 1 i
I
I 0.0112
:
I
I
I
I

Linear Binary

Figure 5.6: RSSI average error and standard deviation
5.4.2 RSSI Target Error (RTE)
Figure 5.6 shows statistical analyses of the tracking RSSI error e(k) , where error

average and Standard Deviation (SD) are shown in this figure. Errors for the linear and
binary approaches are computed as follows. If the received RSSI value is not within the

threshold range, the error will be the absolute difference of the RSSI value and the closest

103

threshold value (either RTL or RTH), otherwise the error is considered to be zero (if RSSI
value within the threshold range). With the LQGI approaches, the error is computed as
defined in Section 5.1. Figure 5.6 shows that LQGI_2 and LQGIQ_2 both have the

smallest tracking average error and SD among all approaches. In other words, the RSSI

for these approaches varies the least from the reference trajectory y ref than the others.

This is due to the controller being able to yield the “best” RSSI without regard to
transmission power constraints. Once the constraints are taken into consideration, as seen
in LQGIQ_2, the standard deviation of the tracking error is increased. Notice that a
higher penalty in LQGIQ_700 results a higher average RTE and SD. Linear and binary
approaches, on other hand, both show high RTE and SD, even with having the threshold
range, as described in Section 4.5.

5.4.3 Energy Per Packet (EPP)

Figure 5.7 shows the average EPPs for linear and binary approaches compared with
different LQGI approaches. It can be seen that LQGIQ_2 has the best EPP performance
among all mechanisms. The average EPP for LQGIQ_2 was 0.68 mJ/Pkt. The reasons for
LQGIQZ’S best performance are as follows. First, use of quantizer helps to reduce the
extreme high transmission powers that are used with LQGI_2, as shown in Figure 5.4.
Second, the low penalty on the control by using small r also helps LQGIQZ to reduce

the transmission powers by having more control packets compared with LQGIQ700.

104

0.795

.0
00

0.781

   
  
    

    
      

   

    

:
Exp. Results E Sim Results

I

g 0.75 :
E 0'71 0.70 E
g: 0.7 :
LL] I
' I
g0 0.65 g
< :

0.6

     

LQGI LQGIQ LQGIQ
- 2 2 700

Linear Binary

Figure 5.7: Average EPP for different dynamic search assignments

Finally, linear and binary approaches show slightly higher EPP (0.71 and 0.7 mJ/Pkt,
respectively) compared with LQGIQZ. The slow step search in the linear search and the
power assignment oscillation in the binary search account for this performance loss in
EPP compared to the LQGIQZ scheme. Both linear and binary search approaches,
however, achieve better EPP compared with LQGI_2 and LQGIQ700. This is because
fewer control packets are used in the search-based approaches by having an RSSI
threshold range as opposed to a single reference RSSI. The average EPP of LQGI_2 and
LQGIQ7OO were 0.795 and 0.781 mJ/Plct, respectively.

5.4.4 Energy Overhead of Closed-Loop Control

The energy costs for the closed-loop controlled assignments manifest in the form of
transmission and reception of the control packets from the sink to the transmitter nodes.

The total control packet communication cost is shown in Figure 5.8.

105

104.4

 
 
 
   
   

y—s
O
0

Sim Results

00
O

53.65

59-5
CO

4.64

Total Control Energy (ml).
O\
o

 

O

LQGI LQGIQ LQGIQ
2 2 700

Linear Binary
Figure 5.8: Cumulative control energy expenditure

Note that the control communication energy for the LQGI_2 approach is the

maximum among all mechanisms. Because no qualtizer with low penalty on the control

are applied in this approach to target yr.) , where the total control energy was 104.4

mJ/Pkt. LQGIQ700 has the lowest total control energy of 4.64 mJ/Pkt, because of its
high penalty on the control due to the high r. Linear, binary and LQGIQ_2 , all have total
control energy in between the two extreme values, because the threshold RSSI range and
the quantizer in these approaches worked to reduce the number control packets from the
sink to the source. The total control energy of the linear and binary power assignments

are very closed, and it was 48.43 m] with linear and 49.01 m] with binary.

5.5 Summary and Conclusions

A closed-loop transmission power assignment framework for on-body wireless links
was developed in this chapter. Human body movement has been modeled as a stochastic
linear system and a quantized Linear Quadratic Gaussian control with an Integrator
(LQGI) approach has been utilized for optimal power control. Experimental results from

linear and binary search-based closed-loop design along with simulation results from

106

LQGI approaches are presented and compared. It was shown that power assignment with
quantized LQGI model and small weight factor can provide lower error and energy

performance compared with the search-based closed-loop strategies.

107

Chapter 6. DTN Routing in Body Sensor Networks with Dynamic

Postural Partitioning

In most WBAN applications, ultra-short radio transmission range is a common
constraint for low-power RF transceivers designed for embedded applications with
limited energy and small form-factors. For such ultra-short transmission range, postural
body movements can make the WBANS highly prone to topological partitioning,
resulting in a body area Delay Tolerant Network (DIN). Such topological partitioning
can often get aggravated by the unpredictable on-body RF attenuation. In this chapter, we
have developed on-body store-and-forward packet routing algorithms in the presence of
network partitioning. The objective is to minimize end-to—end packet delay by
dynamically choosing a route on which the storage/buffering delays are low. While
ensuring low storage delay, the developed algorithms will also attempt to minimize the
end-to-end hop-count, so that the transmission energy drainage is minimized, thus

leading to long network operating durations.

6.1 Introduction

6.1.1 Short RF Transmission Range

Short transmission range is a common constraint for low-power RF transceivers
designed for embedded applications with limited energy [62],[64], often supplied by
harvested operations [65]. Such situations are particularly pertinent for implantable body
sensors. Examples of ultra-low range transceivers in the literature include [66] with O-lm,
[65] with 0.2-1m, [67] with 0.2m, and [68] with 0-1m transmission ranges. The
corresponding transmission powers vary between 0.75mW to 6mW, which are within a

range that can be handled with common harvesting techniques such as piezo-electric

108

generation from body movements. Information available in the literature on such low

power RF transceivers is summarized in Table 6.1.

 

 

 

 

 

Reference Tx Tx. Power Rx. Power
Range Consumption Consumption
(meter) (mWatt) (mWatt)
[65] 0.2 - 1 1.5 — 3.5 ~2.5
[66] 0 - 1 2 2
[57] 0.2 0.75 - 3.75 0.75 — 3.75
[68] 0 - 1 6 5.1

 

 

 

 

 

 

Table 6.1: Low-power and short-range RF transceivers

6.1.2 Routing with Network Partitioning

Low RF transmission ranges also mean that postural body movements can give rise
to frequent partitioning or disconnection in WBAN topologies, resulting in a body area
Delay Tolerant Network (DTN) [69-74]. Such topological partitioning can often get
aggravated by the unpredictable RF attenuation caused due to signal blockage by clothing
material and body segments. Although real-time applications such as patient monitoring
may not be supported in the presence of topological partitioning, non-real-time
applications such as athlete’s physiology monitoring can still be supported using on-body
DTN packet routing across disconnected partitions. Performance goals for such protocols
will be to obtain: 1) low end-to-end delay, 2) low packet loss, and 3) low transmission
energy consumption.

6.1.3 Objective and Contributions

The objective of this chapter is to develop on-body store-and-forward packet routing
algorithms in the presence of network partitioning. The objective is to minimize
end-to-end packet delays by dynamically choosing routes on which the storage/buffering

delays caused due to topological disconnections are low. While ensuring low storage

109

delay, the algorithm also attempts to minimize the end-to-end hop-count so that the
transmission energy drainage is minimized, thus leading to long network Operating
durations. Note that in the absence of network congestions in low data-rate WBANS, the
storage delays due to topological disconnections are usually much larger compared to the
congestion delay. That is why the congestion delay is not modeled in this chapter.
Specific contributions of the chapter are as follows. m, we have developed a
prototype body area network for motivating the on-body packet routing problem and

validating the proposed routing algorithm. Second using the prototype network, we have

 

developed detailed topology characterization mechanisms in order to demonstrate the
network partitioning caused due to human postural mobility. M a probabilistic and a
distance vector packet routing framework are developed using a stochastic link cost
formulation that captures multi-scale topological localities in human postural movements.
m the performance of the proposed protocols is experimentally evaluated using the
prototype body sensor network, and is compared with a probabilistic [74], a
utility-age-based [108],[109], and an opportunistic [110] DTN routing protocol from the
literature. My, an off-line simulation model is developed for validating the
experimental performance trends obtained from the prototype WBAN.
6.2 Related Work

Many body area network implementations using on-body sensor communication
have recently been reported in the literature [14],[2],[111],[112]. A number of these
papers focus mainly on the on-body MAC layer issues. The system in [14] uses a slotted
multipoint-to-point architecture in which the data fiom multiple on-body sensors are sent

to a sink node in a collision-flee manner. The transmission slots are synchronized using

110

beacons periodically sent by a pre-designated sink node. The mechanisms reported in [2]
use an on-body adaptation of the standard IEEE 802.15.4/ZigBee-based MAC.

The work reported in [l l 1],[112] investigate on-body MAC-routing cross-layer
issues via distributed transmission coordination in the presence of specific routing
structures. In [111], the authors present an energy-efficient slotted MAC in the presence
of a Wireless Autonomous Spanning tree Protocol (WASP) that is used for on-body
packet routing. The mechanism is cross-layer in that the MAC slot allocation is
customized for the underlying routing tree, thus providing routing-specific energy
economy at the MAC layer. The protocol in [112] adopts a similar tree-based cross-layer
approach, but designed specifically for reducing packet delivery delay over an on-body
spanning tree. This protocol also handles body mobility by adaptively re-constructing and
maintaining the spanning tree used for packet routing.

From the on—body routing standpoint, most of the existing WBAN systems adopt star
or tree topologies on a connected graph; meaning a physically connected end-to—end path
between any pair of on-body sensors is assumed at any given point in time. However,
these models do not apply for the targeted DTN routing paradigm in this chapter, which
handles topology partitioning leading to scenarios in which end-to-end physical
connectivity between node pairs may not be present at times. Such partitioning is caused
mainly due to the ultra-short-range RF transceivers as reported in Section 6.1.1.

The existing research on routing in disconnected networks (i.e. Delay Tolerant
Networks or DTNs) are categorized [69],[70] as: 1) replication-based (multiple copy)
[71],[74],[108], 2) knowledge-based [72],[113], and 3) hybrid of the above two

[73],[108],[113]. The replication approach explores the ways several copies of a packet

111

can be disseminated among several carrier mobile nodes to increase the chance that it
would reach the desired destination. While providing good delay performance, the
primary limitation of these protocols is their energy and capacity overheads due to
excessive packet transmissions. For ultra resource-constrained WBANS, such overheads
are not acceptable. -

The knowledge-based strategies are typically for single copy forwarding and they
make use of information about connectivity dynamics to make efficient forwarding
decisions [72],[1 13]. The hybrid approaches [73],[108],[113] combine the replication and
knowledge-based strategies. The general principle behind these approaches is as follows.
When a node with a packet to be forwarded encounters another node, the forwarding rule
should determine if the packet (or a copy of the packet) should be transferred to that node
or it should continue to be buffered. The rule is based on the estimate whether the
encountered node is more likely than the forwarding node to visit the destination.

For the above mechanisms to work as anything beyond epidemic/viral routing [114],
the nodes need to have certain degree of spatial and temporal locality in their mobility
and meeting patterns. The scheme PROPHET [74], which is an extension of epidemic
routing [115], develops a probabilistic fi'amework for capturing the spatio-temporal
locality present in the node mobility pattern within a dynamically partitioned wireless
network. The authors in [73] define a high-dimensional Euclidean space, called
MobySpace, constructed upon nodes’ mobility patterns. The specific MobySpace
evaluated is based on the locality of movements defined as the frequency of visits of
nodes to each possible location. Node interaction localities can be also captured in the

form a per-link utility as detailed in [108],[116]. The link utility can be formulated as its

112

age [109],[110],[ll6], formation frequency [116], and other historical parameters that can
effectively capture the nodes’ interaction localities.

Two additional routing protocols, namely opportunistic [108],[110],[116] and
randomized [110],[ll6] are also analyzed in the literature for applications in which either
there is no node-interaction locality or such localities can not be evaluated. With
opportrmistic routing, a source node directly delivers its packets to the destination node,
and buffers them till the link with the destination is formed. In randomized routing,
packets are randomly routed following the hot-potato logic [110]. Both these protocols
are hugely outperformed by the locality-based protocols [110],[116] due to their
knowledge about the properties of the links.

The above mechanisms are all applied to networks spanning across local to wide
areas, few extending all the way up to the inter-planetary scale [117]. The objective of
this part of the thesis is to apply the key DTN routing concepts, as identified above, in an
ultra short-range body area environment. The challenge is to develop mechanisms for
capturing the locality of on-body node movements caused by human postural mobility. A
key contribution of our work is to formulate mechanisms that capture multi-scale
topological localities in human postural movements. Unlike the existing utility-based
[108],[109],[116] and probabilistic [74],[115] DTN routing protocols that capture only
short-term node interaction locality, in this work we device mechanisms for capturing
movement localities at both short and long terms. Such multi-scale locality is shown to

improve the packet routing performance in an on-body context.

113

6.3 Characterization of On-Body Network Topology

This section presents an experimental characterization of WBAN network topology
with different postural positions.
6.3.1 Experimental Settings

A WBAN is constructed (see Figure 6.1), as illustrated in the prototype system in
Chapter 2, by mounting seven sensor nodes (attached on two upper-arms, two thighs, two
ankles and one in the waist area). Each wearable node consists of a 915MHz MicaZDot
MOTE (running TinyOS operating system), with Chipcon’s SmartRF CC1000 radio chip
[76], and the sensor card MTSSlO from Crossbow Inc. [78]. The Mica2Dot nodes run
from a 570mAH button cell with a total sensor weight of approximately 10 grams. The

default CSMA MAC protocol is used with a data rate of 38.4kbps.

 

Modalities:
- Blood pressure
- Heart rate

- Breathing rate ............ ._ ,v , .
-Diabetes . -.
- Temperature; g
- Humidity ’ '
- ECG

    
  

 

- Movement
- Proximity
- Direction

Itzzfi'a—

;@ ..... 9

 

 

[:3

l C
o
l - o'
a. . n
'0 | "
I ’.
0 o o 0 ‘ ' 0
o c.
.' I . 0'
o a .
O. I ' V 'l
’ l
o
. o‘ ‘0'. '0 o

a ‘ o. .

o I
. '0 l. .
O . ‘ . 0
I . I

' .
' . ‘ .' o ‘
I 9 -
u . . '
A . . '
A . . -
' o f :
() l n -
t-Of-B 'I a ‘ 0‘ o. a.
o e 0 s 0 o
. 0.. . .
a. o. e. t
O O .o o.

server . _____ 2'51. ,0"

 

 

6; ...... g7

 

 

 

 

 

 

 

Figure 6.1: On-body Mounted Sensors

114

Via software adjustments of the CC1000’s transmission power, the transmission
range is set to be in between 0.3m-0.6m. By doing so, we are able to emulate the
ultra-low transmission range for the embedded transceivers [65-68] s reported in the
literature. Note that the variation of the range is caused due to the variability in antenna
orientation, clothing, and other on-body RF attenuation characteristics.

The sensors form a mesh topology with one or multiple simultaneous network
partitions. The topology and the number of partitions change dynamically based on the
postural positions of the subject individuals. All experiments in this chapter correspond to
multi-point-to-point routing in which data from all other nodes are sent to node-6 (see :
On-body Mounted Sensors), which is designated as the on-body sink node. This node
collects raw data, and sends processed results or events to an off-body server using a
wireless link. This external link is created between the on-body sink node and to an
off-body Mica2Dot radio node connected to a Windows PC through a custom-built serial
interface, running R3232 protocol, as described in Chapter 2.

6.3.2 Variations of Topology and Network Partitions

Experiments were carried out for observing the impacts of postural mobility on
network partitioning. A human subject, fitted with seven sensors, was asked to follow a
pre-determined sequence of postures (SIT, SIT-RECLININ G, LYING-DOWN, STAND,
WALK and RUN), each lasting for 20 sec. The status of three WBAN links (1-3, 4-6, and
5-3) during such an experiment is shown in Figure 6.2. The presence and absence of a
link’s connectivity, as sampled by the nodes on the link, is represented by l and 0

respectively.

115

 

l

 

 

 

 

 

 

 

 

 

 

 

3
.‘3
m
E
.4
0
0
Link: 1-3
SIT REC DWN
l
.. l
§
:73
fi
r—I
0 l
0 20 40
Link: 4-6 Time (S) [0: OFF, 1: ONl
STD WLK RUN
1 I
. ll
§
5
i=5
.4
0 I .
_ 60 80 100 120
Lmk: 5—3 Time (s) [0: OFF, 1: ONI

 

Figure 6.2: Variation of link connectivity with postural mobility

Each node maintains a neighbor table based on Hello messages sent periodically with

low transmission power once in 1.4 see. A time-out period of 2.8 sec. is used for purging

entries from the neighbor table. The link status in Figure 6.2 is constructed by combining

116

the neighbor table information from the nodes relevant for the exhibited links.
Experimentally, the neighbor table information is periodically sent by all seven on-body
nodes to the off-body server (in Figure 6.1) using the fiill transmission power of the
Chipcon’s CC1000 radio.

The following observations can be made from Figure 6.2. First, few links are
connected only during certain postures, which can lead to significant topology variations
and network partitioning across the postures. For instance, link 5-3 (between left front
thigh and upper left arm nodes) shows the effect of distance on connectivity. The link is
connected during most closed postures such as SIT and REC. However, for the stretched
out postures such as LYING-DOWN, STAND and WALK, the link is mostly
disconnected. Similar trends are observed for the other links including link 1-3 and link

4-6, as shown in Figure 6.2.

 

 

 

 

 

 

 

4 6
Avg. Nodes

is W 8
.. , 4a
a No. of i E
'8 2 * Partitions i, ' g,
z LNUU / “5
ob - ’ 2 o'
3: 1 ~ , Z

0 SIT REC DWN STD WLK 0

0 20 4O 80 100 120

TirisieQ (s)
Figure 6.3: Topology and partition properties with posture changes

The second observation is that even within a posture, a link may have intermittent
disconnections (e.g. link 1-3 is disconnected during the SIT posture from ‘0-20’ see.

interval). The reasons for such intra-posture disconnections include minor body

117

movements, RF signal blockage by body segments and clothing material, and also the
relative orientations of the node-pairs forming the link in question.

The topology level impacts of the body posture variation are reported in Figure 6.3.
Observe the wide swing of the node degree (1.5 to 3.8 across the six postures/activities)
which indicates a high level of dynamism in the on-body mesh topology. Also observe
the number of simultaneous network partitions which vary from 1 to 5, indicating
frequent topological partitioning as hypothesized in Section 6.1.2. As expected, the

postures with relatively lower node degree correspond to higher number of network

 

   
   

   

 
 

   

  
  

  
  

  
 
 

 

 

 

 

 

   

   

 

partitions.
40% 1 INode-l
INode-2
30% .1 QNode-3 ‘2
.9 lIIlNode-4 gg : S
3": ' g V : I. \
o INode-6 E ’5 l5 ; \ .
t5: = - I: ’= . \
E '-' IE 5:; §
10% . s :- 65 la \
"—" '-' I: I: \
= L" I: I: \
:l 2 I: ‘3 §
0% _ n... .E E 4.:- ‘ [E x .
0 1 2 3

No. of Neighbors

Figure 6.4: Distributed of number of neighbors

Figure 6.4 demonstrates the distribution of number of neighbors experimentally
obtained for each node in the network for a duration of 1320 see. spanning six postures
(SIT, SIT-RECLINING, LYING-DOWN, STAND, WALK and RUN), each lasting for
20 sec. Figure 6.4 shows that for most of the nodes, the number of neighbors vary
between 2 and 4, and the probability does not exceed 315% in the best case. This

demonstrates as to how the network dynamically partitions based on human postural

118

 

movements. Such topological disconnections necessitate on—body store-and-forward

routing.

6.4 Store-and-Forward Routing With Multi-Scale Movement Locality

This section develops routing protocols that leverage short and long term node
interaction localities in the presence of partitioned and time-varying WBAN topologies as
demonstrated in Section 6.3.

6.4.1 Postural Link Cost with Multi-Scale Locality

As discusses in Section 6.2, a key requirement for store-and-forward routing in
partitioned topologies is to be able to identify any spatio-temporal locality present in the
node movement patterns. In the context of on-body topologies, we propose a novel
paradigm of Postural Link Cost Formulation (PLCF) in which a time-varying cost is
formulated for each WBAN wireless link based on the locality in the connectivity pattern
for the link in multiple time scales. This postural link cost is then applied for executing
store and forwarding protocols, which are an adaptation of epidemic routing [59] that
leverages the localities of link connectivity in Delay Tolerant Networks (DTN). With
posture and activity changes of a subject, the PLCF link costs are automatically adjusted
such that the packets are forwarded to next-hops which are most likely to provide an
end-to-end path with minimum intermediate buffering/storage delays. The link metric we
propose is specifically designed to minimize data delivery delay by reducing the amount

of packet buffering time at intermediate nodes.

Let us define a Link Likelihood Factor (LLF) Bf}- (0 —<- H]- S 1) which represents

the likelihood for the link Lid: (between nodes i and j) to be connected during a discrete

119

 

t
time slot t. If the quantity Pi]- can be designed appropriately to track the locality of link
connectivity then it can be used as the link cost [118] for probabilistic routing.

We propose that the LLF be dynamically updated after the tth time slot as:

Pifj =Piijl + (l —P,.”.' ) - a) if link LU is connected

:1

_1 . . . . (6.1)
Pi]. =P.fj ~a) 1f lznk La,- 15 dzsconnected

t
When the link is connected, P,j increases at a rate determined by the

t
constanta) (0 S a) $1), and the difference between the current value of P. ,- and its

maximum value, which is 1. As a result, if the link remains connected for a long time, the
t
quantity P.1- asymptotically reaches its maximum value of 1. When the link is

t
disconnected, P1,- asyrnptotically reaches zero with a rate determined by the

z
constanta) . To summarize, for a givena) , the LLF PU- responds to the instantaneous

connectivity condition of the link Lid- .

Note that the LLF above captures the locality in short-term link connectivity in a
manner conceptually similar to the age-based utility formulation, as developed in
[109],[l 16]. It is, however, not the same because in the designs in [109],[116], the routing
utility of a link is increased incrementally when the link is formed, and is reduced to zero
as soon as the link is disconnected. This formulation of utility misses out the fact that
even after disconnection, the formation probability of that link may be higher than a
currently-connected link. In other words, those definitions of utility fairly differentiate

across currently connected links, but not across the currently non-connected links. In the

120

formulation of utility in Eqn. 6.1, motivated by the logic used in PROPHET [74], we
track the short-term locality even when a link is not physically connected. This extended
persistency in LLF is expected to improve performance over the existing age-based utility
definitions as used in [109],[116].

The next design step is to dimension the pararneterw for capturing link localities at a
longer time scale. From Eqn. 6.1, the rate of change of the LLF per time slot can be

written as:

§(Pifj)= (1—1’55-0} if link Li,j is connected

t t—l . . . . (6-2)
§(P,-,j) = — P, j - (1 — (0) If lmk Li,j is disconnected

t
Eqn. 6.2 indicates that for a high a) (e.g. 0.9), Pi,j increases fast when the link is

connected, and decreases slowly when the link is not connected. Conversely, for a

t
loww (e.g. 0.1), Pi,j increases slowly when the link is connected, and decreases fast

when the link is not connected. Ideally, it is desirable that for a historically good link (i.e.

1‘
connected frequently on a longer time-scale), Pi, j should increase fast and decrease

slowly, and for a historically bad link, it should increase slowly and decrease fast. This
implies that the parametera) needs to capture the long-term history of the link; hence it

should be link specific and time varying. Based on this observation, we define Historical

Connectivity Quality (HCQ) of an on-body link Lid- at time slot t as:

l
t _ r
wi,j_ zLiJ Twindow

r=t—TW (6'3)

indow

121

r

The quantity L“ is 1, if the link is connected (see Figure 6.2) during the time slot r,
and 0, if it is not connected. The constant T window represents a measurement window (in
number of slots) over which the connectivity quality is averaged. The parameter
[0,-i]- (0 5 (0,21- 5 1) indicates the historical link quality Lid- as a fraction of time the
link was connected during the last T window duration. The parameter T window should be
chosen based on the human postural mobility time constants. Experimentally, we found

the optimal T window values that work well for a large number of subject individuals and

range of postures to be in between 7 sec. and 14 see.

Figure 6.5 shows the evolution of LLF (Pifj) and HCQ (COL-)with time. The top

graph shows an example link activity with the first half indicating a steadily connected
link with a single fiame (1.4 sec.) of disconnection at time frame 10, and the second half

indicates a steadily disconnected link with single frame of connection at time fi'ame 41

sec. The middle graph shows the evolution of (0):j)with a T window set to 7 frames.
l
The bottom graph shows the evolution of Pi}- with constanta) (i.e. 0.9 and 0.1) and

t
link-specific time varying 601., j from Eqn. 6.3, indicating the historical link quality.

When the link is steadily well connected (during the first half), a high constanta) (i.e.

t
0.9) responds well to a momentary disconnection by decreasing Pi} slowly, but

recovering quickly when the link becomes reconnected. A low constanta) (i.e. 0.1)

122

responds poorly in this situation by doing just the opposite - that is a fast decrease and

slow recovery.

0: OFF, 1: ON

1

LiJ

 

 

lllllllllllllllllllllllllllllllll

 

 

 

 

 

Time Frame

Figure 6.5: Evolution of multi-scale locality in terms of LLF and HCQ
Similarly, when the link is steadily disconnected (during the second half), a low

constanta) (i.e. 0.1) responds relatively better than a high constanta) (i.e. 0.9) by

t t
increasing P.1- slowly for a momentary connection, and decreasing ‘Pij quickly after the

123

link becomes disconnected. The lines for two constant a) values clearly show that a
single constant value fora) is not able to handle both good-link and bad-link situations

equally effectively.

As hypothesized, the link-specific and time-varying (1),-{j , on the other hand, is able

to handle both situations well by mimicking the behavior of a) = 0.9 during the
historically good-link situation, and that of a) = 0.1 during the historically bad-link
situation. These results clearly demonstrate the effectiveness of the HCQ and LLF
concepts for designing routing utilities that can capture both short and long term localities
of the on-body link dynamics. With this multi-scale approach, the proposed mechanism
should be able to outperform both age-based (utility) [109],[1 l6] and probabilistic [74]
routing protocols that use only short term locality information.

Note that unlike the entities in Figure 6.2 and Figure 6.3, the LLF and HCQ in Figure
6.5 show the link connectivity localities which depends on the short and long term history
of the link. The localities captures in Eqns. 6.1 and 6.3 are responsible for this
memory-based behavior in Figure 6.5 in contrast with the instantaneous behavior in
Figure 6.2 and Figure 6.3.

6.4.2 Probabilistic Routing with Postural Link Costs (PRPLC)

Combining Eqns. 6.1 and 6.3, an on-body node i can construct and maintain the Link

t
Likelihood Factor P, j , for all j eN, j at: i, where N represents all WBAN nodes (i.e. 7

nodes in Figure 6.1). In other words, a node observes and maintains its likelihood to be in
direct (one-hop) contact with all other nodes in the network at any given point in time.

The routing goal is to reduce the end-to-end packet delivery delay by choosing high

124

likelihood links, thus reducing the intermediate storage delay caused due to packets stuck
at nodes on low-likelihood links.
The routing philosophy is when a node i needs to forward a packet to node d (d is the

sink node), and it meets a node j, the packet is forwarded from node i to node j only if the

l

condition PM < 13er is found true. In other words, a higher link likelihood of node j to

node d indicates that the latter is more likely to meet node (1 than what node i ’s chances
are. That justifies the packet transfer fi'om node i to j with a goal of minimizing the
end-to-end packet routing delay.

This forwarding logic assumes that each on-body node is guaranteed to intermittently
come within up to 2-hop distance from the destination node. In other words, node i is
intermittently able to see other nodes that intermittently come in direct contact with node
d. In our experimental topology this assumption was always found true. In fact for a
WBAN topology, it is generally true that depending on the specific postural patterns, all
nodes intermittently form direct links with all other nodes in the network. This
observation makes the assumption generally applicable for WBANS which usually have a
small network diameter. This distributed routing mechanism for a multipoint-to-point
implementation is summarized in the pseudo-code presented in Figure 6.6. An extension
of link cost formulation without the above assumption is presented in Section 6.7.

Each on-body sensor node needs to execute the algorithm as presented here. Using

the periodic Hello mechanism, as outlined in Section 6.3.2, each node-i gradually

t
develops the P”- values with all other nodes in the network. The node also uses the same

1
Hello messages to send the quantity P, j , its Link Likelihood Factor (LLF) with the

125

 

common destination node-d (e.g. node 6 in Figure 6.1), to all other nodes that are
currently connected to node—i. This way, each node will know the individual LLFs of all
of its direct neighbor nodes to the common destination node-d.

At any given point in time, if there are packets stored in node-i’s buffer (originated at
node-i or at some other node), node-i checks if any of its directly connected neighbors
has a higher LLF to the destination node-d compared to its own LLF to node-d. If
node—i’s LLF is the highest then it continues to keep the packet in its own buffer.
Otherwise it finds the directly connected node with the highest LLF to the destination
node-d, and forwards packets to that node. This ensures that node-i forwards a packet to
the node that is most likely to meet the destination node, thereby reducing the expected
end-to-end delivery delay.

While the pseudo-code in Figure 6.6 shows a multipoint-to-point implementation for
traffic from all on-body nodes to a common destination, the same concept is applicable
for point-to-point routing. In that case, each node will require maintaining the Link
Likelihood Factors (LLFs) to all possible destinations as opposed to only the common
destination, as done in Figure 6.6. Another difference will be that each packet may be
forwarded to different next hops depending on its specific destination. The rest of the

forwarding logic will be similar to what is presented in Figure 6.6.

126

 

 

Routing Logic for node-i to forward packets to sink-d at
time slot-t

while (true){
for (all nodej[jeN,j¢i]){

l
cofd- = 2L2]- Twindow // compute HCQ

r =t‘T window
if ( Lg, j = I) // 1f link Li ,j is connected
-1 —1
Pi,- = Pi; + (1" Fifi )"”i,j

else // if link Li, j is disconnected
—1
Pair =ij wits
. t _
4a,,- — I)
Send Pitd to node-j

}
Find node-k so that P]: d is

maximumfor[keN,k¢i,d,L:.k =1]

for (all buflered packets to be forwarded
to sink node-d){

if (Li. d = I) // node-i has direct link to node-d

Deliver the packet to sink node-d;
else{

'f (Pia >Pifd)
// node-k has better link likelihood with sink-d
Forward the packet to node-k;
else '

Continue buflering the packet in node-i.

 

 

 

Figure 6.6: PRPLC routing with LLF capturing multi-scale connection localities

127

6.4.3 Single-Copy Probabilistic Routing (PROPHET)

A single-copy version of the probabilistic routing protocol PROPHET [74], which
uses only short term link locality information, is implemented in our laboratory prototype
for comparing it with the proposed PRPLC with multi-scale link localities. PROPHET
relies on epidemic algorithms by doing pair-wise exchange of packets between nodes (as
they come in contact with each other) to eventually deliver them to destinations. At a
node A, a probabilistic metric called delivery predictability, is established for each of its
known destinations B. This metric indicates how likely it is for node A to be able to
deliver a message to destination B.

Nodes buffer packets if there is currently no available path to the destination. An
index of these packets called a summary vector is maintained by the nodes, and when two
nodes meet they exchange summary vectors. Updated summary vector information is
then used to decide which packets to request from the other node. In the evaluations in
[74], the forwarding strategy is when two nodes meet, a packet is transferred to the other
node if the delivery predictability of the destination is higher at the other node.

The concept of delivery predictability in PROPHET has some similarities with the

Link Likelihood Factor used in the PRPLC routing, and is updated [74] via three update

equations using update constants Rm: 7: and fl . Since these update constants are not

designed to be link-specific and they do not adapt with historical link qualities, the
protocol PROPHET somewhat corresponds to the constanta) scenarios in PRPLC as
explained in Figure 6.5 and Section 6.4.1. As a result, while the Link Likelihood Factor
with adaptive historical link quality in PRPLC can capture the long-tenn localities in

postural body movements, it is not possible via the delivery predictability updates using

128

constant B,,-,.7. and ,6 , as used in PROPHET. The performance benefits of PRPLC

over the generic design of PROPHET will be presented in Section 6.5.

6.4.4 Distance Vector Routing with Postural Link Costs (DVRPLC)

In DVRPLC, nodes maintain end-to-end cumulative path cost estimates to the
common sink node. As in PRPLC, the primary goal is to reduce the end-to-end packet
delivery delay by choosing a high likelihood end-to-end path, thus reducing the
intermediate storage delay caused due to packets buffered at nodes on low likelihood

links.

Let us define a Link Cost Factor (LCF) Ci}- (0 5 Ci]- 3 Cmax) which represents

the routing cost for the linkL ,- (between nodes i and j) during the discrete time slot t.

We propose that the LCF be dynamically updated after the tth time slot as:

=C: 1 (l— a)? .) if link Lg is connected
C}: =C.’:.maxl+(C C :1)-(1— a); j) if link L, j is disconnected

(6.4)
When the link is connected, Ci,- decreases at a rate determined by (1 —' (0:1) ,
where 60:1- (O s coil. S 1) is the Historical Connectivity Quality, as defined in Eqn. 6.3.
If the link remains connected for a long time, the quantity Ci,- asyrnptotically reaches its
minimum value 0. When the link remains disconnected, Ct ,- increases at a rate

determined by the quantity (1" COL) , and the difference between the current

t
cost C j and its maximum value 1. This formulation ensures that a link’s routing cost

129

 

always reflects the likelihood of the existence of the link while capturing its historical
connectivity trends. Note that the time evolution of LCF in DVRPLC follows a rationale
that is very similar to that of LLF in PRPLC. The main difference is that while the LCF
reduces for connected links, the LLF increases in such situations. Similar difference
exists when a link remains disconnected. To summarize, like in PRPLC the cost in
DVRPC captures both short and long term link localities for minimum delay packet

routing.

t
Let 7w be the end-to-end cumulative cost from node-i to the sink node-d.

According to distance vector routing logic, when a node i needs to forward a packet to

the sink node d, and it meets a node j, the packet is forwarded to node j only if the

t t
condition 713d <4”; is found true. In other words, a lower path cost through node j

indicates that the latter is more likely to forward the packet to node (1 than what node i ’s
chances are. That justifies the packet transfer from node i to j with a goal of minimizing
the end-to-end packet routing delay. This distributed routing mechanism for a
multipoint-to-point implementation is summarized in the pseudo-code presented in
Figure 6.7.

Note that the DVRPLC protocol attempts to minimize end-to-end cumulative routing
costs. The objective is that due to this end-to-end cost minimization, DVRPLC should be
able to outperform (from a delay standpoint) PRPLC which always interprets its LLF

only at the link level and not in an end-to-end cumulative manner.

130

 

Routing Logic for node-i to forward packets to sink-d at time slot-t

while (true){
for (allnodej[jeN,j¢i]){

t
t
a)“. = Z Li}. /Twindow // compute HCQ

r =t ‘ window

if ( Lg, j = I) // if link L," j is connected

C5,,- =Cit,3'1 '(1- mg!)
l _.
Send 713d to node 1

else // if link Lid is disconnected

-l —1
Cir = C5,; + (Cmax ‘ Ci; HI “ 605,1)
if(i = dAND 7:0, > led)
t t t . . .
7i d = C i d // 71-, d IS node-1 to smk—d end-to-end cost
}
Find node-k such that 7]: d is Minimum AND

1 t . . . .
7k,d <7i,d mm” ID EM} it l.d,L§-,j =1]
for (all buffered packets to be forwarded to sink node-d){

if (L: d = I) // node-i has direct link to node-d
Deliver the packet to sink node-d;
else{
if (a valid node-k was found){
Forward the packet to node-k

l
72d = C 1., k + 71:, d // update the end-to-end—cost

}else
Continue buffering the packet in node-i.

 

 

 

Figure 6.7: DVRPLC routing with link cost capturing multi-scale connection
localities

131

To execute the algorithm in Figure 6.7, each on-body sensor node-i uses the periodic

Hello mechanism, as outlined in Section 6.3.2, in order to gradually develop the Cl;

values with all other nodes in the network. It also iteratively updates the quantity 7,1,,
using the computed Ci]. values with respect to all its neighbors. The node then uses the

Hello mechanism to send the quantity 7,1,, , its end-to—end cumulative path cost to the

common destination node-d (e.g. node 6 in Figure 6.1), to all other nodes that are
currently connected to node-i. This way, each node gets updated about the path costs of
all of its direct neighbors’ to the common destination node-d.

6.4.5 On-body Store and Flood Routing (OBSFR)

In order to determine the best case delay performance among the above protocols,
On-body Store and Flood Routing (OBSFR), a modified flooding protocol for partitioned
networks, has been implemented. With flooding, multiple copies of a packet from a
source node can reach to the destination through multiple routes, and the first arrived
copy at the destination indicates the minimum possible end-to-end storage/buffering
delay that can be achieved by PRPLC and DVRPLC protocols.

Although a regular packet flooding mechanism can be applied to this application, a
few additional routing syntaxes are needed in order to avoid packet losses in certain
scenarios that arise specifically due to network partitioning. In addition to a unique
identifier {source_id., seq_no. } , a packet also carries a list of node-ids indicating its path
so far from the source node. When a node-i receives a packet for the first time (detected
from its unique identifier), it buffers the packet till it encounters at least one node that is

not there in the list of node-ids found in the packet. Upon encountering at least one such

132

node, the packet is handed over from i to such nodes using broadcast, and then deleted
from i’s buffer. Like in regular flooding, upon any subsequent reception of the same
packet, node i will ignore it.

This modified flooding protocol is used for reducing the number of packet drops in
the presence of network partitioning. Consider the following situation. Using the
conventional flooding (i.e. not using the list of node-ids) when node i broadcast the
packet to a node j, it is possible that node j had already broadcast forwarded the same
packet and therefore it simply discards it after receiving from node i. If node i and j are
currently forming a network partition that does not contain the packet’s destination, then
the packet is dropped from this partition and will never be forwarded to its final
destination.

However, with the modified flooding that uses the list of node-ids, node i will not
broadcast the packet to node j, since in this situation node j is already in the list of
node-ids in the packet. As a result, node i will buffer the packet till it encounters a node
that is not already traversed trough by the packet in question. This improves the chance
for the packet to be forwarded out of the current partition (formed by nodes i and j),
thereby reducing the overall packet loss probability. This modification can be applied
only to small networks with few nodes, and will not scale for large sensor network with
tens of nodes to be added in the list of node-ids in the packets.

Note however that even with the above flooding modifications there exists a partition
packet saturation situation in which a packet may be lost. Consider the scenario in which
after node-i receives a packet for the first time it gets into a partition with two other nodes

p and q so that all three nodes are fully connected within the partition. Now since nodes p

133

and q do not appear in the list of node-ids in the packet, node i will broadcast forward the
packet to both nodes p and q and delete it from its own buffer. At this point, nodes p and
q will broadcast forward the packet to each other since q does not appear in the list of p’s
copy and p does not appear in the list of q’s copy of the packet. After this round of
forwarding, both p and q will also delete the packet from their own buffers. This will
cause the packet to be dropped from this partition and will never be forwarded to its final
destination. The OBSFR mechanism for a multipoint-to-point implementation is

summarized in the pseudo-code presented in Figure 6.8.

 

Logic for node-i to forward packets to sink-d
while (true){
for (all buflered packets to be forwarded to sink node-d){

for (allnodej[jeN,j¢i]){
if(L§-,j = I and j elist of node_ids in packet ){

// j is a neighbor of i, and the packet did not
// visit node-j before
Broadcast the packet;
Remove it from node-i ’s buffer
Break; // done with this packet forwarding
} else
Continue buffering the packet at node-i;

}
}
}

Logic for node-i after receiving a packet
if ( the packet was not received before)
if (this is not the destination)
Bujfer the packet in for future forwarding;
else
Discard the packet; // it was received before

 

 

 

Figure 6.8: OBSFR routing logic for packet delay lower bound

134

In spite of these occasional packet drops, this On-body Store and Flood Routing
(OBSF R) protocol ensures that the successfully flooded packets to the destination do
represent the minimum possible packet delivery delay that can be achieved by our
proposed PRPLC protocol. However, due to its inherent flooding nature, the transmission
energy cost for OBSFR will be expected to be significantly larger compared to PRPLC,
DVRPLC, PROPHET, and other utility-based mechanisms.

Note that in the absence of network congestions in low data-rate WBANS, the
storage delays due to topological disconnections are usually much larger compared to the
congestion delay. The protocol OBSFR is designed to deliver the delay lower bound in
the absence of any delays caused due to packet congestions.

Although the broadcast-based Epidemic Routing (ER) [115] could have been used
for finding the delay lower bounds, we have chosen to implement OBSFR for the
following reasons. First, unlike in ER which requires summary vector exchange across
neighbor nodes to minimize content transfers, OBSFR uses a node-list in each packet to
accomplish the same goal. This helps OBSF R to significantly reduce wireless traffic and
energy drainage compared to ER, especially when large numbers of packets are buffered.
Second, while in ER a node continues to buffer content even after it is given to another
node epidemically, in OBSF R, the provider node removes the packet from its buffer,
leading to a much smaller buffer budget compared to ER. The flipside of the above two
aspects is that in certain rare situations as explained above as partition packet saturation,
OBSFR may not be able to deliver a packet to its desired destination. But its significant
capacity, buffer, and energy advantages over the Epidemic Routing, especially in the

context of the resource-constrained WBAN sensors, have prompted us to design OBSFR

135

 

which uses mores constrained flooding than ER, and can work efficiently in a small

WBAN with only few nodes.

6.5 Experimental Performance

The same seven sensor laboratory prototype network, as shown in Figure 6.1, is used
for the results presented in this section. Packets originated from all on-body sensors were
routed to the common destination node-6, attached on the right ankle. Most presented
results correspond to packets originated from node-3, representing the longest hop (i.e.
also worst case) packet routing scenario in most of the body postures. Few results have
also been presented for packets originated from node-5 and node-1, indicating the
generality of our proposed routing from other segments of the body as well.

6.5.1 Polling-Based Channel Access for Collision Control

In order to avoid the CSMA MAC collisions inherent to Mic2Dot’s TinyOS
networking stack, we have implemented a higher layer polling-based TDMA access
strategy that is managed by the common sink node (i.e. node-6 in Figure 6.1). The
primary motivation for TDMA over CSMA access strategy is to operate the WBAN in an
energy-efficient manner. The sink node polls the other on-body sensors in a round-robin
fashion. A node forwards its packets (both data and Hello) only when it is polled by the
sink for giving access to the channel. For our seven-node network (see Figure 6.1), a
polling time frame of 1.4 sec is used which is divided into 7 time slots, one for each
on-body node (the sink node also needs a slot for sending Hello packets etc. for link cost
formulation as described in Section 6.3). Although the data packets and the Hello
messages fiom the nodes are transmitted at power adjusted transmission range of

0.3m-0.6m for emulating the low transmission range as outlined in Section 6.3.1, the

136

polling control packets are transmitted by Node-6 at full power so that all on-body nodes
receive such packets for effective polling, leading to collision-free channel access. If a
node misses a polling packet fi'om node-6, it simply misses one transmission opportunity.

As shown in Figure 6.9, each 1.4 sec frame is divided into seven 200 msec time slots.
. Each slot is further divided into three 60 msec sub-slots and a 20 msec guard time
between adjacent slots. The first sub-slot is used for channel access polling packets fiom
the sink node at full power. The second sub-slot is used for data and Hello packet
exchange between on-body nodes at low power, emulating the transceivers discussed in

Section 6.3.1.

 

 

 

ii [[1] Polling packet from sink node E
[\3 Hello/Data packet from regular nodes . I
E Topology information packet T1me. E
i
e-1 Node-2 Node-3 Node-4 Node-5 Node-6 o o o

 

 
    

 

 

 

 

 

 

01.4
,fr...'

 

 

'5 “x. . “x. M ’
. -. . \°' La \
‘Kfl -. "" '1‘: | ':\
. '=._ ‘. a“. "3k. [5‘ -

l I \ l
': 200 ms :' 2‘ 1400 ms Frame Cycle ——9E 200 ms

 

 

 

 

 

 

 

 

 

 

v C
_‘ .

 

‘7

 

I

: I
I I ‘

I ’ ‘ I
I

I

 

 

 

60 ms each 20 ms

Figure 6.9: Collision-free MAC access via polling

The third sub-slot is used for a topology gathering process in which each on-body
node sends its neighbor information to an off-body machine (see Figure 6.1) at full

transmission power. Using time-stamped neighbor information from each on-body node,

137

it is possible to reconstruct the on-body topology evolution as a result of postural body
movements. As presented in Section 6.6, such reconstructed topology information is used
for offline simulation analysis of the proposed packet routing algorithms.

6.5.2 Performance Metrics

The performance of - on-body routing is evaluated using three commonly used
[108],[110],[113],[1 16] primary metrics, namely, end-to-end Packet Delay (PD), Packet
Hop Count (PHC), and Packet Delivery Ratio (PDR). The index PHC is a direct measure
of the communication energy (i.e. for transmission and reception) expenditure of the
routing mechanism. Unlike for routing in conventional un-partitioned networks, the PD
in partitioned on-body networks depends mainly on the storage delay at the intermediate
nodes as a result of network disconnection. The PHC here mainly impacts the number of
transmissions per packet forwarding, indicating the energy expenditure; it does not
impact the packet delay so much. The target is to minimize PHC while also minimizing
the PD by avoiding large packet storage delays by routing a packet through links with
low disconnection probabilities or high Link Likelihood Factor as defined in Section
6.4.1. The PRPLC protocol, as presented in Section 6.4.2, is designed mainly to minimize
the on-body PD values.

6.5.3 Traffic Generation and Data Collection

The source node generates a data packet every 4 polling frames (each frame is 1.4
sec), with a packet size of 46 bytes. Each packet is marked with a monotonically
increasing packet—ID so that by observing the received packet-ID, the packet delivery
ratio (PDR) can be computed. Also, all on-body network nodes are coarse-grain
time-synchronized by the sink node-6 at the beginning of each polling frame. This allows

single-trip packet delay (PD) to be computed from a source node to the sink node-6. On

138

its way to the sink node, a data packet collects the entire route information in the form of
a list of the intermediate node-IDs. This allows the extraction and analysis of route
information including the PHC values.

6.5.4 Packet Delay (PD)

End-to-end packet delivery delays from source node-3 on the left upper arm to the
sink node-6 on right ankle for PRPLC and DVRPLC are reported in Figure 6.10. Three
different versions of PRPLC and DVRPLC was implemented; two with static values of
a) (i.e. 0.9 and 0.1) and one with adaptivew , capturing the long-term locality in terms of
Historical Connectivity Quality of the on-body links. For each of these scenarios, a
separate experiment was run for 1320 see. (i.e. 22 minutes), sending 230 packets, and
spanning 6 different postures (SIT, SIT-RECLININ G, LYING-DOWN, STAND, WALK
and RUN), each lasting for 20 see. Figure 6.10 reports the average of packet delay

computed from each such experiment. All adaptive to results correspond to a connectivity

quality measurement window (i.e. T window ) of seven polling frames or approximately

9.8 sec. Value of T window in this neighborhood has shown to demonstrate the best

performance for multiple subject individuals, indicating a fairly good estimation of the

time constant of the long-term postural movement locality.

139

I—a
O

3000

Average Delay (3)

N

 

      

O

(0) = 0.1) (a) = 0.9) Dynamic (0

Figure 6.10: Packet delivery delays with static and adaptive a) scenarios

The following observations can be made from Figure 6.10. First, with adaptivea) ,
both PRPLC and DVRPLC protocols are able to successfully capture the long-term
locality in postural movements (see Section 6.4.1). As a result, the packet delays for the
adaptive to scenarios (3.6 sec and 3.11 sec for PRPLC and DVRPLC respectively) are
significantly reduced compared to those for the constant a) implementations which
leverage only the short term movement locality as explained in Section 6.4.1. Second,
with all implementations, DVRPLC achieves better packet delay compared with PRPLC,
mainly because of its end-to-end path cost formulation as explained in Section 6.4.4.
Because of its delay superiority, this point onwards we will present results only for the
adaptive to scenarios. Unless stated otherwise, the references to the protocols PRPLC
and DVRPLC will correspond to their adaptive to implementations, reflecting the
multi-scale locality which was not captured in PROPHET [74] and other utility-based

mechanisms [109],[l 16] in the literature.

140

#GOO

N

Average Delay (8)

 

0
PRPLC PROPHET DVRPLC OBSFR UTILITY OPPT.

Figure 6.11: On-body packet delivery delay for different DTN routing protocols

Figure 6.11 reports the experimentally obtained average packet delays for a number
of protocols including the proposed PRPLC, DVRPLC, OBSFR, an utility-age-based
protocol (UTILITY) [116], an opportunistic protocol (OPPT) [110], and the single-copy

probabilistic protocol PROPHET (with its delivery predictability update

constants PM ’7’: and ,3 chosen as 0.75, 0.98, and 0.25 as reported in [74]). Observe

that the flooding protocol OBSFR achieves significantly better packet delay (i.e. the
experimental lower bound) compared to the other protocols mainly due to its
multi-forwarding nature as explained in Section 6.4.5. On the other extreme, the
opportunistic protocol OPPT, in which a source node delivers a packet to the destination
only when it experiences a direct link with the destination, shows the worst delay because

of the very low connection fiequency between the source node 3 and the destination node

6 in Figure 6.1. With static update constants PM, ,7, and fl , PROPHET suffers fiom

the same shortcomings (i.e. leveraging only the short-term locality) of PRPLC with
constanta) , as shown in Figure 6.10. As explained in Section 6.4.1, the age-based utility

approach (UTILITY) also suffers from the same short-term-only locality, which explains

141

its poor performance compared to the packet delays for PRPLC and DVRPLC (i.e. 3.6
sec and 3.11 sec). With adaptive to , these two protocols are able to successfully capture
the multi-scale link localities caused due to postural body movements. It should be noted
that in spite of their superior delay performance with respect to PROPHET and age-based
UTILITY, there is still room for improvement when compared to the experimental delay

lower-bound as demonstrated by the OBSF R flooding protocol.

      

   

 

 

 

   

 

 

$100% . BPRPLC, Avg: 2.15
g [I] PROPHET, Avg: 2.11
8 80% J DVRPLC, Avg: 2.41
5:; 60% _ $2: OBSFR, Avg: 2.33
gin . $33: a UTILITY. Avg: 2.06
.3 40% " : $555: I OPPT, Avg: 1.0
g 20%— he
re 0% _ he: Ell]
2 3
Packet Hop Count (PHC)

Figure 6.12: Distribution of Packet Hop Count (PHC)

6.5.5 Packet Hop Count (PHC)

PHC serves as a measure for routing energy expenditure (i.e. for transmission and
reception) for the on-body sensors. Figure 6.12 demonstrates the distribution of PHC
experimentally obtained for all the protocols implemented on our prototype.

The figure shows that packets in DVRPLC and OBSFR take slightly longer routes
compared to the other protocols. As shown in Figure 6.11, the packets in those two
protocols also experience the lowest packet delay. This means that the DVRPLC and
flooding protocols route packets through better quality links, leading to smaller delays,
even though it requires more number of end-to-end hops. The distribution graph in Figure

6.12 also shows that while majority of the packets from node-3 to node-6 are routed in

142

 

2-hops, certain packets take one, three, or four hops. Also note that certain packets are
delivered directly from node-3 to node-6 (from the upper left arm to the right ankle in
Figure 6.1), especially during the closed postures such as sitting. Since in the
opportunistic routing protocol (OPPT) packets are delivered only when a source node
comes in direct contact of the destination, all packets are delivered with PHC 1.

Figure 6.13 shows the number of data packet transmissions. It is computed as the
ratio of the total number of transmissions and the number of successfully delivered
packets at the sink. This index captures the additional forwarding costs for multiple
packet transmissions in flooding-based protocols such as OBSFR. The large number for
OBSFR explains the impacts of its multi-forwarding compared to all other single-copy
DTN routing protocols. Note that there were no link layer packet retransmissions during
these experiments; any channel errors have resulted in dropped packets, and are captured

by the Packet Delivery Ratio (PDR) reported in the following Section.

& 05

Total N23 of Trans. Per Pkt

 

O

 

PRPLC PR DVRPLC OBSFR UTILITY OPPT.

Figure 6.13: Number of transmissions per delivered packet

6.5.6 Packet Delivery Ratio

Packet losses are observed due to the following two reasons. First, due to postural

mobility, there are transient blackout periods during which a neighbor may appear to be

143

connected in a node’s neighbor table, when in fact it is no longer connected. These
blackout periods are created during a node’s neighbor time-out period, which is in the
vicinity of 2.8 see, as reported in Section 6.3.2. Packet transmissions during such
blackout periods end up in packet losses since no link layer reliability is used. All six
protocols suffer from such packet losses. The second type of losses, applicable only to
flooding, is due to the partition packet saturation, as explained for the OBSFR protocol
in Section 6.4.5.

Figure 6.14 demonstrates that due to its multi-packet forwarding, the flooding-based
OBSFR looses fewer packets compared to all the other protocols, even though the packet
losses due to partition packet saturation are present only for OBSFR. Poor PDR for the
OPPT protocol in this case was caused due to a very unreliable link between the source
and the destination nodes (i.e. nodes 3 and 6 in Figure 6.1) which are physically situated
at two extremes of the subject’s body. Since the OPPT protocol relies on direct
source-destination contact for packet delivery, the source-destination link quality affects

this protocol most.

90%

PDR

80%

 

71%

  

70%

PRPLC PR DVRPLC OBSFR UTILITY OPPT.

Figure 6.14: Packet delivery performance

144

 

 

 

6.5.7 Routing Packets from and to Different Body Segments

Delivery delays for packets from different body segments to the sink node placed on
the right ankle (i.e. node 6 in Figure 6.1) are shown in Figure 6.15. Observe that as the
physical distance between a source and the destination increases, the average packet
delays for all the protocols increase. Relatively though, all the experimented routing
protocols maintain the same trend for the packet delay as observed in Section 6.5.4. The
average packet hop-counts from source nodes 5, 1, and 3 were experimentally logged in

the range of 1.25, 1.5, and 2.3 respectively

 

8 _ E PRPLC /' ' / '
[U PROPHET .

Z DVRPLC
D OBSFR
B UTILITY
I OPPT.

  
 
 
 
 

Ch
4*

 

 

 

Average Delay(s)
A

 

Src = 5 Src =
(Left Thigh) (Waist ) (Upper Left Arm)

Figure 6.15: Delivery delay for packets from thigh, waist and arm to right ankle (i.e. node 6)

Delivery delays for packets from the upper left arm (i.e. node 3) to different body
segments are shown in Figure 6.16. It should be noted that although the absolute delay
values are different, the overall trend in packet delay follows the same pattern for all the

sink nodes that we have experimented with.

145

 

 

 

 

 

 

_ I: PRPLC - ' -
8 III PROPHET \ / 43 28/
ZDVRPLC 12.36s “315 '
’5?
3: 6 2 BOBSFR «.-.-.-
:3 E1 UTILITY :-:-:-:
a I OPPT. ___ ;:::::
d.)
g0
‘63
>
<
0 ‘ Sink=2 Sink=5 ' Sink—_-
(Upper Right Ann) (Left Thigh ) (Right Ankle)

Figure 6.16: Delivery delay for packets to arm, thigh and ankle from the upper left arm
(i.e. node 3 in Figure 6.1)

6.5.8 Impacts of Postural Stability

For all the experiments so far, each individual physical posture was made to last for
20 see. In order to study the impacts of variable postural stability on the routing
performance, the subject was instructed to repeat the same sequence of postures as in

Section 6.3, but with different posture durations ranging from 10 sec. to 40 sec.

    

 

 

 

 

10 - OPPT. UTILITY ~- 60
(u; PROPHET 5 - __ 50 3'
E 3 “ 3‘
0 4.. I—
Q 6 1 40 3
°° t -- 30 Oh
>
< 4 — E
E l —— 20 .
° ‘5

DVRPLC OBSFR O
0 r I l 0
10 20 Posture DuratiOITQS) 40

Figure 6.17: Impacts Of posture duration on packet delay

146

Figure 6.17 shows the impacts of posture duration on average packet delay for all six
on-body routing protocols. Due to its significantly higher values, the packet delays for the
OPPT protocol are plotted as a separate axis in Figure 6.17. Observe that the packet
delays for all the protocols generally increase with higher posture durations. This is
because longer posture duration implies that a connected link remains connected for
longer duration and also a disconnected link remains disconnected longer. As a result, a
packet that is buffered in a node due to network partitioning remains buffered for longer
duration, leading to higher end-tO-end packet delay. In a relative sense, all the
experimented protocols maintain the same performance trend for the packet delay as

Observed in Section 6.5.4, Figure 6.11.

 

 

 

 

 

 

 

 

 

 

95% ‘ ;
90% ‘ \ — M"
85% - /
Q 80% +PROPHET
p" +PRPLC
75% — +038“
+DVRPUC
70% -jl_____ . +m
I - :5- -' OPPT.
65% I l l
10 20 40

30
Posture Duration (3)

Figure 6.18: Impacts Of posture duration on packet delivery ratio

Figure 6.18 shows the impacts Of posture duration on the PDR metric. As explained
in Section 6.5.6, the primary reason for drops is transient blackouts due to postural
mobility. With higher posture durations, the degree Of mobility is less, and therefore the
drops are less frequent. This explains a general increase in packet delivery ratio with

higher posture durations.

147

 

 

29.31 sec/ .

8

6

4

Average Delay (8)

2

 

0
PRPLC PROPHET DVRPLC OBSFR UTILITY OPPT

Figure 6.19: Packet delivery delay with intra-posture movement

6.5.9 Impacts of Intra-posture Movements

All the experiments so far correspond to the inter-posture sequence (SIT,
SIT -RECLINING, LYING-DOWN, STAND, WALK and RUN), as introduced in
Section 6.3.2. Figure 6.19 presents packet delay for similar experiments, but carried out
with an intra-posture sequence comprising Of the positions SIT, SIT-RECLINING,
RECLINING-RIGHT-CROSS, RECLINING-LEFT-CROSS, RECLINING-RAISED-RIGHT,

and RECLINING-RAISED-LEFT, where CROSS and RAISED refer to cross-legged and
leg-raised sub-postures while sitting. The primary objective of these intra-posture
experiments is to study the impacts Of higher granularity postural movements on the
on—bOdy routing protocol performance.

Comparing the performance in Figure 6.19 with those for inter-posture movements in
Figure 6.11, it can be Observed that the relative performance trends across all the six
protocols are still maintained for this high granularity intra-posture case. The absolute
delay values, however, have been slightly reduced due to a better overall connectivity
compared to the inter-posture case. Similar trends were consistently Observed fi'om many

more inter- and intra-posture experiments carried out during this work.

148

 

 

 

 

 

 

Figure 6.20: Experiments with different sensor placements

6.5.10 Impacts of Sensor Placements

Additional on-body sensor placements, as shown in Figure 6.20, were experimented
with for evaluating the validity of the routing results Obtained so far from the sensor
placement shown in Figure 6.1. Different source and sink nodes are used in the two
placement settings P1 and P2 in Figure 6.20. The inter-posture sequence, described in
Section 6.3.2, was followed by the subject and the corresponding packet delay results are
presented in Figure 6.21 for the placement settings P1 and P2 in Figure 6.20. Generally,
the relative performance trends across all the experimented protocols as Observed for the
original sensor placement (in Figure 6.11) remain valid for the new sensor placements P1
and P2 in Figure 6.20. The delay for the placement P1 is larger due to the longer

source-tO-destination distance compared to that in P2.

149

Average Delay (s)

 

 

H PRPLC PROPHET DVRPLC

Average Delay (3)

    

 

PRPLC PROPHET DVRPLC OBSFR UTILITY OPPT

Figure 6.21: On-body packet delay for: (a) sensor placement P1, and (b) sensor
placement P2

6.6 Offline Simulation with Experimentally Obtained Topology

The Objective of this section is to develop an Off-line simulation framework that uses
network topology traces Obtained during the on-line experiments described so far. The
motivations for such simulation are to: l) validate the correctness of the experimental
results, 2) determine a benchmark performance with the minimum possible delay for a
given posture/topology sequence, 3) develop a mechanism for Offline experimentation
with new protocols, before implementing them online which is significantly more

experimentally involving.

150

 

 

 

 

Real-time Experiment 2 3

with PRPchand Real- time Offline
DVRPLC" Topology v dons
Export 5'5 ‘ Simulated
Source 1 "I PRPLC and
I“ 4H5 DVRPLC

i

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6 6 57
Export Dynamic Topology
Comparative Performance
Evaluation

Experimental /\_/
Performance N
—> /\_/
/\_/

 

Figure 6.22: Experimental topology export for offline simulation

In order to keep the results comparable, the simulation is performed on the exact
same topology sequence used for the experiments. During the experiments described in
Sections 6.3 and 6.4, all seven on-body nodes periodically export their neighbor table to
an off-body machine using full transmission power. This external machine then derives
the experimental topology sequence by combining the time-series neighbor table
information. The Offline simulation is performed on this topology sequence and the
results are then compared with those Obtained experimentally. This arrangement is
summarized in Figure 6.22.

6.6.1 Delay Benchmark

In order to determine the best case end-tO-end delay performance, an Offline route
search algorithm, Backward Search for Delay Benchmark Routing (BSDBR) has been
developed. As long as the entire topological sequences for a dynamically partitioned

network are known a priori, the BSDBR algorithm is able to compute the most delay

151

 

Optimal end-to-end path for each packet depending on its source, destination, time Of
origin, and the complete topological sequence information. BSDBR is designed to be an
Offline centralized search algorithm, to be executed in the presence of entire time series

topology information.

Let to be the time instant at which a packet is generated at node-i and routed towards
destination node-d. Given a known topology sequence, the Objective is to find the earliest
time instant after t0 at which the packet can be delivered to the destination. Let t1 be the
earliest time instant (t1 > to) at which destination node-d comes in contact with any other
node-j (j e N, j ¢ d). The minimum possible delivery delay for the packet originated at
time t0 can be written as (t1 - to). This minimum delay is possible only if the necessary
network links are formed across the network during the time interval [t0 t0 t1] so that the
packet could be forwarded multi-hop all the way from the origin node-i to node-j before
time t]. The Obj ective of BSDBR search process is to scan the network topology sequence
in order to find if such link formations are there so that (t1 - to) can represent the

minimum packet delivery delay.
If the search process concludes that the packet cannot be delivered by time t], then

the next feasible time instant t2 is identified and a similar search is conducted to

determine if (t1 - to) can be the minimum delivery delay. The quantity t2 is the earliest

time instant after t] (t2 > t1) at which destination node-d comes in contact with any other

node-j (j EN, j at d). This BSDBR search process is iteratively continued till a valid

minimum delivery delay (tr- to) is found. The time instant tr corresponds to the earliest

152

 

contact time so that the necessary network links are formed across the network during

the time interval [to to tr] so that the packet can be forwarded multi-hop all the way from

the origin node-i to destination node-j by the time tr.

Note that the expected value Of the packet delay lower bound could have been
computed using the Linear Programming formulation as adopted in the full
knowledge-based approach in [113]. Instead, we have chosen to implement BSDBR since
it allows us to determine the minimum packet delay for each individual packet as
Opposed to their average in a statistical sense. Also, the formulation in [113] is more
complex than BSDBR since it incorporates the effects Of message queuing which is not
studied in our implementation.

6.6.2 Simulated Performance Results

The same experimental topology sequence in Section 6.5 is used for the simulation
results presented in this section. Packets are routed from node-3 to node-6 with a data rate
Of one packet every 4 fiames. Figure 6.23 shows the packet by packet delivery latency of
PRPLC and DVRPLC fiom simulation and experiments for a total duration Of 1320 see
(or 22 minutes), involving transmissions Of 230 packets for each protocol experiment.
Figure 6.24 reports the corresponding Packet Hop Count (PHC) from the same

experiment.

153

 

 

 

 

 

   

 

 

 

 

 

100

 

 

 

 

 

 

 

 

 

 

 

 

   

 

16-
312:
E
0
a 3 -
a
ti
fl 4‘
O I I I I I I I T
0 20 40 60 80 100 120 140 160 180 200 220
PacketID
16* lllllllllllllllll
— Sim, Avg: 3.39 sec. “““ t,
312 5 PRPLC - Exp. Avg 3.60 sec. 5
5‘ :
0 I
a 8 - :
3 l
0 A I l
a“: 4‘ g
0 r I I I I i
40 50 6O 70 80 90
PacketID
15 —
§IO~
i
_g ..
g 5
O-t
O I I I
0 20 40 60 80 100 120 140 160 180 200 220
PacketID
15 — """" _ . “““““““““““““““
Srm,Avg:2.89sec. ~-__.
3 —Exp, Avg: 3.11 sec. 5
>5 1
7.310 DVRPLC :
a s
6 I
.34 5 -
8
On
0 I I I f -—'1 I
40 50 60 70 80 90 100
PacketID

Figure 6.23: Delivery delay from experiments and simulation

154

 

 

 

 

 

A4

U

I

3*.

E3

0

o i j

‘1 l

O I

.352 '

O

.2

, ll

D.

II I I V I I I T I
0 20 40 60 80 100 120 140 160 180 200 220

PacketlD

 

 

 

 

 

 

 

 

 

70
PacketID
65
:1:
954
g:
§2
% I
is?
0 1' I l T I l I l I I I I
0 20 40 60 80 100 120 140 160 180 200 220
PacketID
A5 _ _-—- _______________________
g ’ . ~~~~~~~~~~~
8.54 - DVRPLC Sm AVE 2-3
.. —Exp,A .2.41
5
U

 

 

 

40 50 60 70 80 90 100
Packet ID

Figure 6.24: Packet hop counts from experiment and simulation

155

For clarity, delivery latency for the 40th to the 100th packets are zoomed in and shown
in both Figure 6.23 and Figure 6.24. As expected, the simulated packet latencies are
slightly better (approximately 7%) than the experimental values. However, the overall
trends are very similar, indicating that the Operations of the PRPLC and DVRPLC
implementations within the prototype on-body network are very similar to the offline
simulation as arranged in Figure 6.22. The primary reason for the performance loss in the
experiments is the packet drops caused due to the reasons as explained in Section 6.5 and

Figure 6.14 and Figure 6.18. For the Offline simulation, 100% packet delivery helps

 

keeping the packet delivery latency slightly higher.
The PHC in Figure 6.24 shows a very similar trend in which the simulation average

is slightly better than the experimental results.

5.79 sec.

NUJ-PIJIO

Average Delay (8)

H

 

O

PRPLC PROPHET DVRPLC BSDBR UTILITY OPPT.

Figure 6.25: Simulated average packet delivery delay

Figure 6.25 reports the average packet delay of PRPLC and DVRPLC compared with

BSDBR, UTILITY, OPPT, and PROPHET (with the constants 13...,2’, and ,8 chosen

as in Section 6.5.4 and reported in [74]). For all the protocols, the same topology

sequence which was extracted from the on-body experiments is used. Observe that

156

PRPLC and DVRPLC achieve significantly better packet delay compared to that (5.79,
4.05 and 34.8 sec) Of PROPHET, UTILITY and OPPT respectively. This trend agrees
with the experimental results reported in Section 6.5 and Figure 6.11. Also note that the
packet delay of PRPLC and DVRPLC are close to the best case Offered by the benchmark

Obtained using BSDBR.

I IN
()1wa

 

O
p—d

Packet Hop Count (PHC).
o 'u: H

 

PRPLC PROPHET DVRPLC BSDBR UTILITY OPPT.

Figure 6.26: Simulated average of Packet Hop Count
Figure 6.26 shows the corresponding PHC performance. The figure shows that
packets in BSDBR take longer routes compared to the other three protocols. As shown in

Figure 6.25, the packets in BSDBR enjoy the minimum packet delay. This means that

 

BSDBR routes packets through better quality links, leading to smaller delays, even
though it requires more number of hops. These results are very similar tO the
experimental results in Section 6.5.5 and Figure 6.12, where the flooding-based protocol
OBSFR Offers the best delay, but at the expense Of longer routes. Figure 6.26 also shows
that the packets in DVRPLC are delivered using slightly longer routes compared with

PRPLC, PROPHET, UTILITY and OPPT.

157

6.7 PRPLC without 2-Hop Assumption

As explained in Section 6.4.2, the baseline PRPLC in Figure 6.6 assumes that each
on-body node intermittently comes within up to 2-hop contact Of the destination/sink
node. While this assumption is generally expected to be true for WBANS which usually
have a small network diameter, in theory the assumption may get violated for networks
with large diameters. In this section we generalize the link cost logic by using a transitive
component [74].
6.7.1 Transitive Update of Link Likelihood Factor

In the generalized case, a node i may forward a packet to node j, event if node-j has

never directly visited the sink/destination node-d. A consequence Of this generalization is

that the Link Likelihood Factor (LLF) of L13} now can change not only based on the

connectivity status of the link, but it can also increase based on the LLF of the
intermediate nodes that node i may meet over time. This transitive change Of LLF can be

captured by executing the following additional (to Eqn. 6.1) update equation when the

link L5,}, is connected.

t _ t—l H t t t t
BJ—BJ +(1_B,j ) i,k' k,j°a)i,k'wk,j (6-5)
The transitive update equation above indicates that if node-k has a high LLF to

node-j, then every time node-i meets node-k, the LLF Of link LL} goes up by a factor

that depends on the Historical Connectivity Quality (HCQ) Of links L5,]. and Lk,j .

With the above addition of LLF update logic, the same logic in Figure 6.6 can be used for

packet forwarding. Note that in order to support the above transitive update, each node is

158

 

required to report the LLF and HCQ information about all its neighbors in its periodic

Hello messages.

16 1 Packets with performance difference — NO Transitive, Avg: 3.39 sec.
With Transitive, Avg: 3.16 sec.

h—i
N
1

Packet Delay (3)
OO

 

 

 

 

 

 

 

 

 

 

 

 

0 l l I I I ' F 4M. l I

0 20 40 6O 80 100 120 140 160 180 200 220
Packet ID

Figure 6.27: The impacts Of transitive updates on delivery delay

6.7.2 Performance with Transitive LLF Update

Figure 6.27 and Figure 6.28 show delay and Packet Hop Count (PHC) performance
Of PRPLC from simulation experiments conducted with and without transitive LLF
updates. For fair comparison, both sets Of experiments were carried out on the same
topology sequences. The same experimental settings as in Section 6.6 are used.

As shown in Figure 6.27, the inclusion Of transitive LLF update does slightly
improve the packet delivery delay (approximately 6%). The packets for which there is an
improvement are encircled in the figure. These improvements are caused mainly due to
the fact that with transitive update, a node is now able to forward a packet tO the sink
through nodes that may never see the sink directly. This also means that for those packets
that are delivered with lower delay will have higher hop counts. This explains the higher

PHC number in Figure 6.28 for PRPLC with transitive update. For example, in Figure

159

 

 

6.27 packets 7, 8, 18, 38, 82 and 106 are delivered with lower latencies, but in Figure

6.28, higher PHC is reported for delivering those packets.

 

 

 

a 4 _ —NO Transitive, Avg: 2.08

:1: -- With Transitive, Avg: 2.22

EL 3 _

5 _ I;

8 l l . .

g, 2 l‘ v v

a I \

r51 . J \ /

'56 Packets with performance difference

a O I I I I I I T l I I f
0 20 40 60 80 100 120 140 160 180 200 220

Packet ID
Figure 6.28: The impacts Of transitive updates on Packet Hop Counts (PHC)

Note that the delay improvements due to the transitive update mechanism remain
limited due to the small WBAN diameter as explained in Section 6.4.2. For networks
with larger diameter, more number Of intermediate nodes will be available under the
with-transitive version Of the protocol, and the resulting delay improvements in those

cases will be expected to be much higher.

6.8 Summary and Conclusions

Store-and-forward packet routing protocols for wireless body area networks
(WBAN) have been developed in this chapter. The concept Of a stochastic link cost was
introduced for enabling a probabilistic and a distance vector on-body routing protocol in
the presence Of postural mobility Of human body. Performance Of these proposed
protocols were evaluated both experimentally and via simulation, and then compared
with a generic probabilistic routing protocol and a specialized on-body packet flooding

mechanism that provides the routing delay lower-bounds. It was shown that via

160

 

 

successful modeling of the spatio-temporal locality of on-body link disconnection
patterns, the proposed algorithms can provide better routing performance compared to the
existing probabilistic routing protocols in the literature. Ongoing work on this topic
includes developing a distance vector routing algorithm using a similar stochastic link
cost metric, and developing a Kalman Filter-based body movement prediction model for

predictive on-body packet routing.

161

 

 

Chapter 7. Modeling On-body DTN Packet Routing Delay

This chapter presents a stochastic modeling framework for the store-and-forward
packet routing protocols that are presented in Chapter 6 for Wireless Body Area
Networks (WBAN) with postural partitioning. Delay modeling techniques for evaluating
single-copy on-body DTN routing protocols are developed. End-to-end routing delay for
a series of protocols including opportunistic, randomized, and the two proposed
mechanisms in Chapter 6 that capture multi-scale topological localities in human postural
movements have been evaluated. Performance of the analyzed protocols are then
evaluated experimentally and via simulation to compare with the results obtained from
the developed model. Finally, a mechanism for evaluating the topological importance of
individual on-body sensor nodes is developed. It is shown that such information can be
used for selectively reducing the on-body sensor-count without substantially sacrificing

the packet delivery delay.

7 .1 Introduction

The goal of this chapter is to develop analytical modeling mechanisms for computing
packet transfer delay for a series of DTN routing algorithms that can be implemented in
an on-body setting. The dominating delay in DTN routing is contributed by packet
buffering caused due to topological disconnections, as we discussed in Chapter 6. In the
absence of network congestions in low data-rate WBANS, such buffering delays are
usually much larger compared to the congestion delay. That is why the congestion delay
is not modeled in this chapter. Specific contributions of the chapter are as follows. First,
the developed prototype body area network in Chapter 2 is motivated the on-body packet

routing problem. The prototype is conducted on-body routing experiments with the DTN

162

 

 

 

routing protocols that are modeled and evaluated in this chapter. Second, a topology trace
collection mechanism, as discussed in Chapter 6 is developed for wirelessly extracting
network topology, as a function of human postural dynamics, from the on-body sensors
to an off-body server. Third, analytical techniques are developed for modeling the
end-to-end packet delay for a range of DTN routing algorithms, namely, opportunistic
[108],[116],[110] utility-based [108],[110],[109], random [108],[110],[119], PRPLC
[120],[121], and DVRPLC [120],[121] . Fourth, the DTN routing delay obtained from the
developed model are compared with results from on-body experiments from the
prototype WBAN and off-body simulation carried out on network topology traces
obtained from the prototype WBAN. Finally, using the model and the topology trace
data, a detailed analysis is carried out for identifying non-critical nodes in order to design

a minimal WBAN topology from the routing stand point.

 

   
  
 
 
  
   

Modalities:

- Blood pressure
- Heart rate

- Breathing rate .
- Diabetes .
- Temperature?
- Humidity ' ‘

- ECG

 

- Movement
- Proximity
- Direction

 

l:|

Out-of-Body
Server

 

 

 

 

 

 

 

 

 

Figure 7.1: On-body Mounted Sensor

163

 

 

7.2 Topology Trace Collection for Off-Body Routing Simulation

A DTN WBAN is constructed as in Figure 7 .1 (also illustrated in Chapter 6), by
mounting seven sensor nodes (attached on two upper-arms, two thighs, two ankles and
one in the waist area). A remote trace collection mechanism was developed so that real
network topology traces from the prototype WBAN can be wirelessly collected and used

for routing model development and off-line routing simulation experiments.

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

On-body Experiment Exported Dynamic TOPOIOEY
2 3
a, :-
Real-time \é/ 4 Dir-body
Topology . 3 Simulation
Export ; \_
I“ 4“: is ~I°M0d61iw
..-® , .
? 'g Performance
3 i
6 8 é>7
Comparative
Performance Evaluation
/\-/
/\-/
Experimental Delay /\_/
Prototype WBAN Performance of DTN /\_/
Routing Protocols 4'

 

 

 

 

 

 

 

 

Figure 7.2: Topology export for offline and model performance

As depicted in Figure 7.2, during the on-body experiments, the state of each link is
periodically sent to the off-body server at full transmission power. The server collects the
link-state samples (ON or OFF) from all the on-body links and stores them with a
time-stamp from its local clock. All these link-state samples, together, form topology
traces which are then used for delay model development and off-body routing simulations
as presented in Sections 7.3, 7.4 and 7.6. Results from the delay model and off-body

simulations are compared with the routing performance from the on-body experiments

164

 

since all of them use the exact same topology traces, ensuring comparable link state and

network partitioning patterns.

7.3 Modeling DTN Routing Protocols

The objective of this section is to model the delay of: a) a series of existing single
copy DTN routing algorithms applied to on-body settings and, b) two specific routing
algorithms that are specifically developed in Chapter :6 to leverage the locality of WBAN
topology as function of postural body movements.

Definition 1 (Link State): The state of a link between two on-body nodes i and j at the

th . . . . . .
n discrete time slot 18 represented as Li.j (n) , Wthh IS assrgned the value 1 or O to

indicate the state to be connected or disconnected respectively. The time slot here is an
observation time slot which corresponds to the Hello interval period for neighbor/link

discovery. In our prototype implementation, it was set to be 1.4sec.

 

 

 

 

 

 

|i—|—|T‘]fi Time
I— Ti’J
Off—H Ly-20 Ag], 1

Figure 7.3: Example connectivity of an on-body link
Definition 2 {Link Disconnection Probability): The Link Disconnection Probability

(LDP) for the link between node-i and node-j is represented as 13,1- (k). The

quantityPi (k) represents the probability that after an arbitrarily chosen time slot, the

link remains disconnected for k consecutive disconnected time slots. In a sufficiently long

165

 

 

. . . k
topology trace, spanmng T time slots, if n represents the number of occurrences of such

k-slot long disconnections, then the LDP can be expressed as:

,. II; /T, fork 21
321(k): ZL,J(n)/T, fork = o (7.1)
n=1

The casek 21 represents situations for which the arbitrarily chosen slot is a part of

one of the T of periods (except the last slot on the T oflperiod) or the last slot during one of
the T 0,, periods (see Figure 7.3). Similarly, the case k = 0 represents situations for which

the arbitrarily chosen slot is a part of one of the T 0,, periods (except the last slot on the T 0,,

period) or the last slot during one of the T oflperiods. With above definition of if ,- (k) , we

T A
have Zho Pi]- (k) = 1 , and its expected value can be represented as:

ELDM. = :=Ok13, -(k) (7.2)

,1

where ELDiJ- is the Expected Link Delay, representing the average number of

disconnection slots after an arbitrarily chosen slot. In other words, ELDiJ- can be

_ i’j ', ' . . . .
expressed as ELDi, j — Tofl /2 , where T 03er rs the average disconnection duratron
for link i to j.

7.3.1 Opportunistic Routing

In DTN Opportunistic routing (OPPT) [108],[110],[ll6], a source node delivers
packet to the destination node only when the two nodes come into direct contact. This

single copy mechanism offers a simple DTN routing approach for which the delay can be

166

 

 

very large, especially in scenarios with low mobility or infrequent link formation between
the source and destination. As done in [116], the opportunistic routing is modeled and
analyzed in this section for estimating the worst delay performance when complex
algorithms need to be avoided for the resource-constrained on-body sensors.

Since a source node 3 delivers a packet to destination d only when L3"; = 1 and a
packet at node 5 can be generated at any arbitrary time slot, the delivery delay for a
packet is ELDM as developed in Eqn. 7.2. This is true only when the packet generation
rate is low enough so that no more than one packet is generated during a T of period (see
Figure 7 .3). This means that the generated packet can be delivered at the very beginning
of the immediately following T 0,, period without any additional wait.

However, when the packet generation rate is higher so that multiple packets are
generated during a T017 period, the packets need to be delivered one per time slot during
the next T 0,, period. This backlog clearance adds an additional delay component that

needs to be added in addition to the ELDW from Eqn. 7.2. Let B represent the number of

packets generated during the T 017 period. With A being the packet generation rate at the

source node 3, B = I1 ' Tofl' . After the subsequent T 0,, period starts, these B packets are

flushed one packet per time slot, requiring B time slots. During these B slots,

another B .2 packets are generated which are then cleared one per slot.

Combining the backlog clearance delay with the Expected Link Delay (ELD), the

average delivery delay for the packets generated during the T off period can be written

167

 

 

 

B-l
as ELDsd + 2i B . Average delay for the packets generated during the T 0,, period

i=1

B 824 .
can be written as? + 2’. Therefore, the overall average packet delay for on-body

i=1
opportunistic routing can be expressed as:
B-l B 31—1
B- ELDsd +Zi +B-l- 3+ 21'
i=1 i=1

DOPPT = B+B-l , (7-3)

 

where Expected Link Delay (ELD) can be computed in Eqn. 7.2, and

9d _ . o u o
B = A ' To} — 2 ' A ' ELD 3,01 . Note that thlS expressron rs valrd when the system
is stable in the sense that on an average, all packets generated during the T 0,, and T017

periods are able to be delivered during the T 0,, period for the link between nodes 3 and (1.

7.3.2 Randomized Routing

In a randomized routing protocol (RAND), if a node with a data packet does not have
a direct connection with the destination, the node forwards the data packet to a neighbor
chosen at random [110],[l 16]. The packet is subsequently forwarded in the same way, till
it is received at the destination. Unlike for hot-potato routing [110] in large networks, the
delay performance of RAND can often be better than that of opportunistic routing in
small body area networks only with few nodes. Smaller topologies have lesser number of
end-to-end path combinations, leading to quicker delivery. Also, the network partitioning,
as shown in Figure 6.3, helps reducing the path combinations even further. Packet
looping, which is inherent in a randomized routing protocol, can be avoided by recording

a packet’s traversed path in it incrementally so that a forwarding filtering can be

168

 

 

implemented. A packet is never forwarded to a node that is recorded in that packet’s
already traversed path.

Definition 3 (Forwarding Probability): In RAND forwarding, a node-i forwards a packet
uniformly randomly to one of its currently connected neighbors. Therefore, at any time

slot n, the probability of node i forwarding a packet to node j is defined as:

L. .(n)
f _ "
BJ (n)— N j
ELM (n)
j=1

and Pif(n)=0foralljeN,

 

N (7.4)
1' ¢ in at d, ifZL.,,~ (n) ¢ 0 and 1,, (n) = 0

j=1

where N is the number of nodes in the on-body network. Eqn. 7.4 is applicable as long as

N
node i currently has at least one neighbor l'e° ZLiJ (n) at 0 and none of those
j=l

neighbors is the destination node d (i-e- [4,3,1(71): 0). In case when node i has

destination d as a current neighbor, the packet is forwarded to node d with probability ‘1’.
Also, when node i has no current neighbors, it keeps buffering the

- f _ . . f _ . . ,
packet (Le-PL,- (n) — 1 ), resulting mg”,- (n) —- 0, f0r all 1 $1 . Incorporating all these

situations, Eqn. 7.4 can be expanded as:

169

 

 

L. .(n)
Pi£(n)=—Ni—and Pi5(n)=0foralljeN, j¢i,j¢d,

2173101)
j=l

N
if 2L1.) j(n) ¢ 0 and Li", (n) = 0
j=l (7.5)

1351(n)=1, 135(n)=0, 35(n)=0forazlje1v,j¢i,j¢dsz,,d(n)=1

N
135004 and 35(n)=0,fora11jeN,j¢z-,if ZLI-J(n)=0
i=1

Definition 4 (Forwarding Matrix): The forwarding matrix captures the forwarding

probabilities at time slot n across all possible links in the network with N nodes and can

be represented as:

 

1: Plfloz) PIQOI) PLC-(n) fime
2: P2{1(n) P2{2(n) P2201) P21: N(n)
A(n)= i: Pigln) ago) Fifi“) fig/(n) (7.6)
d: 0 O ... 0 ... O
_N; Pilot) P15,2(n) P15, j(n) PIMP“),

 

The Forwarding Matrix A(n) has two notable properties. First, the elements in the (1th

row are all zeros since the destination node d never forwards a packet. The elements in

h . . .
the d1 column, however, are either 1 or zero, depending on node d’s instantaneous

connectivity with the other nodes as expressed above in Eqn. 7.5. Second, the summation

170

 

 

of all elements in a row should be 1. The Forwarding Matrix, which depends on the link

states Li, j (n) , can be created afier the forwarding probabilities are computed using
Eqn. 7.5 based on the observed link states from the collected WBAN topology traces.

. . . th .
Consrder a data packet that rs generated at node 3 during the n time slot, and

delivered to node d at the (n+k)th time slot, resulting in a delay of k slots. The value of k

can vary from O to infinity. Let the probability of the above event (i.e. delivering the

packet with a delay of k slots) be represented as the delivery probability [9:0, 0‘) , which

can be expressed as:

k
pgd (k) = [A(n).A(n +1) ..... A(n + 10]“, = H A(n + i) (7])
i=0 s,d

which is the [s,d] element of the product matrix. Therefore the expected RAND

. th . .
forwarding delay for a packet that was generated at the n time slot can be written as:

T T k
_ n _ °
DRAND “ Zk°ps,d(k) ‘ 2]" HA“ H) (7.8)
k=0 k=0 i =0
s,d

where T is the length (in number of slots) of the experimental topology traces obtained in
Section 7.2. Considering sufficiently long on-body topology traces (i.e. large T), the
maximum value of k in Eqn. 7.8 is set to be T instead of infinity.

To clarify the above forwarding concept further, let us explore the following
example. Consider a 4-node (i.e. N= 4) body sensor network with node-1 as the source

and node-3 as the destination. Example forwarding matrixesA(1) , A(2) and A(3), and the

171

 

 

corresponding network topologies at time slots 1, 2 and 3, obtained from the topology

trace, are given by:

 

 

 

 

00500.50 0 00010 0
1000 0010
A(l)=0000 ’A(2)=0000 ’
L0010_9 9 050.5009 9

'0010‘
A3_001oo o

()"0000

_0010_9 9

 

 

 

Matrix products [A(l)oA(2)] and [A(l)-A(2)-A(3)] are:

'025 0.25 0.5 0‘ o o 0.5 o
[A(1)-A(2)]= O 0 0 1 and [A(1)-A(2)-A(3)]= 0 O 1 O
0 0 0 o 0 0 o 0 .

_ 0 o 0 0- _0 0 0 0_

 

 

 

 

Using the above matrixes, the delay probabilities can be computed

as: [711,3 (0) = 0, [212,3 (1) = 0.5 and [213,3 (2) = 0.5 using Eqn. 7.7. According to A(l),

at time slot 1, node-1 has two neighbors (2 and 4), node-2 has one neighbor (node-l) and

 

node-4 has a direct connection with the destination node-3. Assume that a packet is

generated at source (node-1) at time slot 1. Since node-1 has no direct connection with d

(i.e. P1,]; = 0 ), the packet will be randomly forwarded to either node 2 or 4 with

probability 0.5 each, but the probability of delivering it to the destination node-3 is zero

in the current slot-l (out of all possible infinite number of slots in future). This is

captured by [)1]; (O) = 0 which is the [1,3] element of matrix A(I).

172

At time slot 2, the packet will be forwarded to 3 through 2 with probability 1, that is
if 2 has already received the packet in slot 1. Otherwise (i.e. the packet was forwarded to
node 4 in slot 1), the packet will be forwarded to node-l or node-2 by node-4 at slot 2.

Therefore, the probability of delivering the packet to the destination node-3 in slot-2 (out

of all possible slots) is 0.5. This is captured by [)1]; (1) = 0.5 from the [1,3] element of

the product matrix [A(1)- A(2)].

Since Pl,j3 = {3 = P 41:3 =1 in A(3), the packet is guaranteed to be delivered to

node-3 in slot-3. Since the probability of delivery in slot-1 was zero, and in slot-2 was
0.5, and the delivery is guaranteed in slot-3 (i.e. if it was not delivered in slot-2), the
probability of delivering the packet in slot-3 (out of all possible slots) is 0.5. In other

words, the probability of delivery with a delay of 2 slots (i.e. k=2) is 0.5. This is also

captured as [9113(2) =05 from the [1,3] element of the product

matrix [A(l)-A(2)-A(3)] . Using p113 (0) = 0, p113, (l) = 0.5 and p11} (2) = 0.5 , the
expected delay for random forwarding for this example WBAN topology trace is

0x0+1x0.5+2x0.5=1.5time slots.

7.3.3 Utility-based Routing using Link Locality

In randomized routing, a node does not consider the locality of its connectivity with
other network nodes wile forwarding a packet. In utility-based routing protocols
[74],[110],[116],[108],[109] nodes prefer to forward packets to destination through the
neighbor with the latest encounter with the destination, thus leveraging the link locality in

the form of its age. Each node is assigned a utility value based on the last encounter time

173

 

 

with the destination, and a packet is forwarded to a neighbor with the highest utility
value. Utility represents how useful (fast) this node might be in delivering a data packet

to the destination, and is often implemented using a timer.

Let the utility function U ,3 j (n) represent the utility value of node i with respect to

node j at the n trme slot. Every trme node 1 comes 1n contact wrth node j, the quantity
U i, j (n) is set to a maximum utility value and then for every time slot the node remains

out of contact from the destination, the quantity U ,3 j (n) is decreased based on a

pre-set utility reduction method [116],[109] as a function of elapsed time. The update rule

for U131“) can be written as:

U , i L- - n+1 =1
Ui, (n+1): max f "1( ) (7.9)

where Umax is the maximum utility value to the destination each node can has. These
utility values are exchanged between neighbors within the periodic Hello messages.

With the above definition of utility, at the nth time slot node-i will forward a packet

(destined to node-d) to node-j only if
Ui,d (n) < U jad (n) and U 13d (n) 2 Ukfl (”XV/C E Viol) , where V1,: (72) is the set of

all neighbors of node-i during the nth time slot.

Note that the above forwarding logic assumes that each on-body node is guaranteed
to intermittently come within up to 2-hop contact from the destination node. In other

words, a source node is intermittently able to see other nodes that intermittently come in

174

 

 

direct contact with the destination node. In our experimental topology this assumption
was always found true [121]. In fact for a WBAN topology, it is generally true that
depending on the specific postural patterns, all nodes intermittently form direct links with
all other nodes in the network. This observation makes the assumption generally

applicable for WBANS which usually have a small network diameter [116],[109].
The packet routing delay in utility-based forwarding (UTILITY) can be computed
using the same logic as in random forwarding (RAND) except that the forwarding

probabilities Bf, (n) in Eqn. 7.5 need to be reformulated for UTILITY. The forwarding

probability in this case can be expressed as:

13.502) =1and 13.5.01) = 0, Vj ¢ i e NifUi,d(n) 2 Uj,d(n),Vj awn)
13.5. (n) = 1 and 113.5(1)): 0 if Ui,d(n) < Uj’d(n) and Uj’d(n) 2 Uk,d(n),

Vk El/ll-(n) and Vr¢j EN
fig(n)=1and g{}(n)=0,Vj¢d eNifd «am-(n)

(7.10)

where N represents the set of all on-body nodes and where V401) is the set of all

neighbors of node-i during the nth time slot. The top line of Eqn. 7.10 represents a

srtuatron 1n WhJCh e1ther node-z does not have any nerghbor dunng the n trrne slot, or 1ts

own utility to the destination node-d is higher than those of all its current neighbors.
Either way, the node buffers the packet with probability 1. The middle line of the
equation codes the utility-based forwarding rule as stated after Eqn. 7.10. The bottom line
represents the situation in which the destination node-d is a direct neighbor of node-i,

causing a direct delivery.

175

 

 

Once the forwarding probabilities are computed applying Eqn. 7.10 on the on-body

topology traces collected in Section 7.2, the forwarding matrix A(n) and the delivery

n
probabilities ps,d (k) are computed using the same rules presented in Eqns. 7.6 and

7.7. Finally, the delivery delay is computed using Eqn. 7.8 as

T k
DUTILITY = 2k- HAM“)

k=0 i=0 s,d

7.3.4 Probabilistic Routing with Postural Link Cost (PRPLC)

Routing using PRPLC utilizes a Postural Link Cost (PLC) [120] which captures
WBAN link localities in multiple time scales. PRPLC routing protocol is discussed in
Chapter 6 and Section 6.4.2. We redefrne this protocol in this chapter to construct the
overall analytical model of this protocol. For on-body packet forwarding, the PLC is used
exactly the same way as for the UTILITY routing; that is by replacing the utility values
by the PLCs. The routing mechanism of this routing protocol is summarized in the
pseudo-code presented in Figure 6.6. With posture and activity changes of a human
subject, the PLC link costs are automatically adjusted such that the packets are forwarded

to next-hops which are most likely to provide an end-to-end path with minimum

intermediate buffering/storage delays. PLC is defined as ,6”- (n) , (05,3),1- (”)51) ,

which is similar to the Link Likelihood Factor (LLF) that is defined in Section 6.4.2. It

represents the probability of frnding L, I") =1. The update equations for PDC are
.1

formulated as in Eqn. 6.1 [121],[120],[122]:

176

 

 

 

fli,j(n)=fl,~,j(n)+(l—fli’jm_1)).w iflink LU (n):1
fli,j(n)= .3“- (n —1)-a) iflink L13} (,1) = 0 (7-11)

According to Eqn. 7.11, when the link is connected, the Postural Link Cost (PLC)

fli, j (71) increases at a rate determined by the constant a) (0 S a) S 1) , and the difference
between the current value of ,6,’ J- (n) and its maximum value, which is 1. As a result, if
the link remains connected for a long time, the quantity 161', j (n) asymptotically reaches
its maximum value of 1. When the link is disconnected, 161', j (n) asymptotically reaches
zero with a rate determined by the constant a). To summarize, for a givena) , 16,-, j(n)

responds to the instantaneous connectivity condition of the link Li, j .

With time invarianta) , the PLC update rules in Eqn. 7.11 captures the locality in
short-term link connectivity in a manner conceptually similar to the age-based utility
formulation, as developed in [109],[116]. It is, however, not the same because in the
designs in [109],[116], the routing utility of a link is increased incrementally when the
link is formed, and is reduced to zero as soon as the link is disconnected. This
formulation of utility misses out the fact that even after disconnection, the formation
probability of that link may be higher than a currently-connected link. In other words,
those definitions of utility fairly differentiate across currently connected links, but not
across the currently non-connected links. In the formulation of PLC in Eqn. 7.11,
motivated by the logic used in PROPHET [74], we track the short-term locality even

when a link is not physically connected. This extended persistency in PLC is expected to

177

 

 

 

improve performance over the existing age-based utility definitions as used in
[109],[116].
The next design step is to dimension the parametera) for capturing link localities at
a longer time scale. From Eqn. 7.11, the rate of change of the PLC per time slot can be
written as:
mid-(on (1-fli,,-(n-1))-w iflink L.,,- (n) =1

(7.12)
§(,Bi,j(n))=—,Bi,j(n-1)-(1—a)) iflink Lid- (n) = 0

Eqn. 7.12 indicates that for a high a) (e.g. 0.9), ,Bw-(n) increases fast when the
link is connected, and decreases slowly when the link is not connected. Conversely, for a
lowa) (e.g. 0.1), ,6,”- (n) increases slowly when the link is connected, and decreases
fast when the link is not connected. Ideally, it is desirable that for a historically good link
(i.e. connected frequently on a longer time-scale), 4.1“) should increase fast and

decrease slowly, and for a historically bad link, it should increase slowly and decrease
fast. This implies that the parameter a) needs to capture the long-term history of the link;

hence it should be link specific and time varying. Based on this observation, we redefine
the Historical Connectivity Quality (HCQ) of an on-body link Li, j at time slot n as:
n
0),” (n) = Z Li,j (r) T window (7.13)
r =""T window
The constant T window represents a measurement window (in number of slots) over

which the connectivity quality is averaged. The factor

178

 

 

 

 

(01-d- (n), (0 S (Oz-,1- (n) S 1) indicates the historical link quality as a fraction of time the

link was connected during the last T window slots. The parameter T window should be
chosen based on the human postural mobility time constants. Experimentally, we found
the optimal T window values that work well for a large number of subject individuals and

range of postures to be in between 7 sec. and 14 sec.

 

 

 

 

 

 

1 ::::::::A .1 ON
"a 1.31 Li,j(n)
0’ 7 HCQ :3
U 0.6 A
I 0.4 —
0.2 ~ _
0 fi vececcettceeeceeY} -; OFF

 

 

 

 

 

 

 

0 l 0 20 30 40 50
Time Slot

Figure 7.4: Evolution of multi-scale locality in terms of PLC and HCQ

Figure 7.4 shows the evolution of PLC )6”- (n) and HCQ (1),-J (n) with time. The

top graph shows an example link activity (indicated by L,- (n)) with the first half

indicating a steadily connected link with a single time slot (1.4 sec.) of disconnection at
time slot 10, and the second half indicates a steadily disconnected link with single slot of

connection at the 41St slot. The top graph also shows the evolution of

179

 

 

 

 

50,” (n) according to Eqn. 7.13 with a T window set to 7 time slots. The bottom graph

shows the evolution of '61}! (n) with constant a) (i.e. 0.9 and 0.1) and link-specific

time varying (0),,- (n) from Eqn. 7.13, indicating the historical link quality. When the

link is steadily well connected (during the first half), a high constant a) (i.e. 0.9)

responds well to a momentary disconnection by decreasing ,6- (n) slowly, but
1,]

recovering quickly when the link becomes reconnected. A low constant a) (i.e. 0.1)
responds poorly in this situation by doing just the opposite - that is a fast decrease and
slow recovery.

Similarly, when the link is steadily disconnected (during the second half), a low

constant a) (i.e. 0.1) responds relatively better than a high constant a) (i.e. 0.9) by

increasing fli,j(") slowly for a momentary connection, and decreasing

fig, 11”) quickly after the link becomes disconnected. The lines for two constant a)

values clearly show that a single constant value for a) is not able to handle both

good-link and bad-link situations equally effectively.

As hypothesized, the link-specific and time-varying ,5}, j (n) , on the other hand, is

able to handle both situations well by mimicking the behavior of a) = 0.9 during the
historically good-link situation, and that of a) = 0.1 during the historically bad-link
situation. These results clearly demonstrate the effectiveness of the HCQ and PLC
concepts for designing routing utilities that can capture both short and long term localities

of the on-body link dynamics. With this multi-scale approach, the proposed mechanism

180

 

 

should be able to outperform both age-based (UTILITY) [109],[116] and probabilistic

[74] routing protocols that use only short term locality information.

 

UTILITY

 

 

Utility
function
U(0

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

   

PLC
Computation
Model ch

Routing Decision

 

 

 

M . Short-term
locality

 

 

 

 

I Capturing short- and long-term localities

 

 

Figure 7.5: Capturing link connectivity locality in PRPLC and UTILITY age-based
routing.

Note that unlike the entities in Figure 6.2 and Figure 6.3 in Chapter 6, the PIC and
HCQ in Figure 7.4 show the link connectivity localities which depends on the short and
long term history of the link. The localities captures in Eqns. 7.11 and 7.13 are
responsible for this memory-based behavior in Figure 7.4 in contrast to the instantaneous
link behavior in Figure 6.2 and Figure 6.3. Figure 7.5 summarizes the structural
difference between PRPLC [122] and the UTILITY [109],[116] age-based protocol from

the link locality capture standpoint. As shown in the figure, while UTILITY extracts only

181

 

 

short term locality from the link on-off dynamics, PRPLC extracts an additional long
term locality by observing the Historical Connectivity Quality (HCQ) as presented in
Eqn. 7.13.

The forwarding rule in PRPLC is identical to what stated for UTILITY-based

forwarding in Section 7.3.3 with the utility function U i, j (n) replaced by the postural

link cost flu (’1). Consequently, the forwarding probabilities Pg]- (n), the forwarding

matrix A(n), and the delivery probabilities P: d (k) can be computed using Eqns. 7.6,

7.7 and 7.10 respectively, and finally, the end-to-end packet delay can be computed using

T k
Eqn. 7.8 as DPRPLC = 2k. HAW”)

k=0 i=0 S,d

7.3.5 Distance Vector Routing with Postural Link Costs (DVRPLC)

In DVRPLC, nodes maintain end-to-end cumulative path cost estimates to a common

sink node, as defined in Chapter 6 and Section 6.4.4. Where the Link Cost Factor (LCF)

is defined as Cid- (n), O S CiJ (n) S Cmax which represents the routing cost for the

link LU (between nodes i and j) during the discrete time slot 12. The update equations

for LCF are formulated as [121],[120],[122]:

Ci,j(n)=Ci,J-(n—l)-(1—a)i,j(n)) iflink Li,j (n) =1

, (7.14)
Cid. (n)=Ci,j (n — 1) +(Cmax —C,.,J. (n —1))-(1—(ol-’j(n)) iflink Lid- = 0

182

 

 

 

When the link is connected, CM (71) decreases at a rate determined

by (1‘60“- (n)) , where wiJ (n) , (0 S a),.,j(n)51) is the Historical Connectivity
Quality, as defined in Eqn. 7.13. If the link remains connected for a long time, the

quantity C i, j (n) asymptotically reaches its minimum value 0. When the link remains
disconnected, C ,3 j (n) increases at a rate determined by the quantity (1 “60,3,- (n)) , and

the difference between the current cost C13; (n) and its maximum value 1. This

formulation ensures that a link’s routing cost always reflects the likelihood of the
existence of the link while capturing its historical connectivity trends. Note that the time
evolution of LCF in DVRPLC follows a rationale that is very similar to that of PLC in
PRPLC. The main difference is that while the LCF reduces for connected links, the PLC
increases in such situations. Similar difference exists when a link remains disconnected.
To summarize, like in PRPLC,Ithe cost in DVRPC captures both short and long term link

localities for minimum delay packet routing.

Let 7,4101) be the minimum end-to-end cumulative cost from node-i to the sink

node-d. According to distance vector routing logic, when a node i needs to forward a

packet to the sink node d, and it meets a node j, the packet is forwarded to node j only if

the condition 713d (71) < 7i.d (n) is found true. In other words, a lower path cost through

node j indicates that the latter is more likely to forward the packet to node d than what
node i’s chances are. That justifies the packet transfer from node i to j with a goal of
minimizing the end-to-end packet routing delay. The forwarding mechanism of this

protocol is summarized in the pseudo-code presented in Figure 6.7.

183

 

 

Note that the DVRPLC protocol attempts to minimize end-to-end cumulative routing
costs. The objective is that due to this end-to-end cost minimization, DVRPLC should be
able to outperform (from a delay standpoint) PRPLC which always interprets its PLC
only at the link level and not in an end-to-end cumulative manner.

To execute DVRPLC, each on-body sensor node-i uses the periodic Hello

mechanism, in order to gradually develop the C ,3 J- (n) values with all other nodes in the
network. It also iteratively updates the quantity 7,3,1 (’1) using the computed CL} (11)
values with respect to all its neighbors. The node then uses the Hello mechanism to send

the quantity 7M (n) , its end-to-end cumulative path cost to the common destination

node-d (e. g. node 6 in Figure 7.1), to all other nodes that are currently connected to

node—i. This way, each node gets updated about the path costs of all of its direct

neighbors’ to the common destination node-d. The update equation for 7,3,, (11) :
r.,.(n)=min(r.,.(n), r.,.(n)+C.,.(n>) (7.15)
where node-k has the minimum 719d (’1) among all the current neighbors of node-i.

Because of its end-to-end nature, the forwarding rule in DVRPLC is based on the

end-to-end cost 7,3,1 (n) as opposed to based on local parameters U i, J- (n) or ,3,, j (n) as

used in UTILITY and PRPLC both of which do not rely on end-to-end cost. The distance
vector forwarding rule for a packet from node-i to destination node-d can be formalized

as follows. If node-i is a direct neighbor of node-d, forward the packet. Otherwise find

node-k such that node—k has the minimum 71,401) among all the current neighbors of

184

 

 

node-i. Then forward the packet to node-k only if 7”, (n) < 7,3,, (n); otherwise,

continue buffering the node as node-i. With this forwarding rule, the forwarding

probabilities Bf,- (n) can be expressed as:

Pif;,(n)=1and 115(n)=0,vj° ¢i,d e NifLLd(n)=1
13.5,(n)=1 and 13f;(n)=o,v)° ¢i,k eNij’Li,d(n)¢1and

7k,d(")<7i,d(") andyk,d(n) S 7r,d,Vr e 911,-(n),r ¢ i,d,k e N (7.16)
fl§(n)=land 133(n)=0,Vjad,ieNthi’d(n)¢1and

7k,d(") 2 7i,d(") andykfl (n) S yr,d,Vr e wi(n),r at i,d,k e N

where N represents the set of all on—body nodes and where W11") is the set of all

neighbors of node-i during the nth time slot. The forwarding matrix A(n), and the delivery

probabilities p: d (k) can be computed using Eqns. 7.6 and 7.7 respectively, and finally,

the end-to-end packet delay can be computed using Eqn. 7.8 as

r k
DDVRPLC = Zk{HA(n+i)] .
.gd

k=0 i=0

7 .4 Routing Delay Benchmark

In order to determine the best case end-to-end delay performance, an offline route
search algorithm, Backward Search for Delay Benchmark Routing (BSDBR) has been
used, which is same delay benchmark routing used in Chapter 6. As long as the entire
topological sequences for a dynamically partitioned network are known a priori, the
BSDBR algorithm is able to compute the most delay optimal end-to-end path for each

packet depending on its source, destination, time of origin, and the complete topological

185

 

 

sequence information. BSDBR is designed to be an offline centralized search algorithm,

to be executed in the presence of entire time series topology information.

Let t0 be the time instant at which a packet is generated at node-i and routed towards
destination node-d. Given a known topology sequence, the objective is to find the earliest
time instant afler to at which the packet can be delivered to the destination. Let t] be the
earliest time instant (t1 > to) at which destination node-d comes in contact with any other
node-j (j e N, j 7* d ). The minimum possible delivery delay for the packet originated at
time t0 can be written as (t1 - to). This minimum delay is possible only if the necessary
network links are formed across the network during the time interval [to to t1] so that the
packet could be forwarded multi-hop all the way from the origin node-i to node-j before
time t]. The objective of BSDBR search process is to scan the network topology sequence
in order to find if such link formations are there so that (t1 - to) can represent the

minimum packet delivery delay.
If the search process concludes that the packet cannot be delivered by time t’, then

the next feasible time instant t2 is identified and a similar search is conducted to

determine if (t1 - to) can be the minimum delivery delay. The quantity t2 is the earliest

time instant afier tl ((2 > t!) at which destination node-d comes in contact with any other

node-j (j EN, j at d). This BSDBR search process is iteratively continued till a valid
minimum delivery delay (tr- to) is found. The time instant tr corresponds to the earliest

contact time so that the necessary network links are formed across the network during

186

 

 

,4

 

 

the time interval [to to tr] so that the packet can be forwarded multi-hop all the way from

the origin node-i to destination node-j by the time tr.

Note that the expected value of the packet delay lower bound could have been
computed using the Linear Programming formulation as adopted in the full
knowledge-based approach in [113]. Instead, we have chosen to implement BSDBR since
it allows us to determine the minimum packet delay for each individual packet as
opposed to their average in a statistical sense, as we discussed in Chapter 6. Also, the
formulation in [113] is more complex than BSDBR since it incorporates the effects of

message queuing which is not studied in our implementation.

7.5 Performance Evaluation

The same seven-sensor laboratory prototype network, as shown in Figure 7.1, was
used for the on-body experimental evaluation of all the analyzed routing protocols.
Packets originated from all sensors were routed to the common destination node-6,
attached on the right ankle. Unless stated otherwise, results correspond to packets
originated fiom node-3, representing the longest hop (i.e. also worst case) packet routing
scenario in most of the body postures. Results are also presented from off-body
simulation experiments, carried out on network topology traces collected during the
actual on-body experiments so that the simulation results can be compared with the
experimental data for the exact same topology traces. Those traces are also used to create
the forwarding matrix in Eqn. 7.6 for computing the analytical packet delay numbers for
all the analyzed routing protocols.

In order to avoid the CSMA MAC collisions inherent to Mic2Dot’s TinyOS

networking stack, we have implemented a higher layer polling access strategy managed

187

 

 

 

by one on-body node similar to the polling mechanism that is described in Section 6.5.1.
This polling node polls the other six sensor nodes in a round-robin fashion, so that a
regular node forwards packets (both data and Hello) only when it is polled by the polling
node and given access to the channel. A polling time frame of 1.4 sec. is used which is
divided into 7 time slots, one for each of the seven on-body nodes. Note that the polling
node itself also needs to send Hello packets etc. for link cost formulation as described in
Section 6.3. Although the data packets and the Hello messages from the nodes are
transmitted at software power adjusted transmission range of 0.3m-0.6m, the polling
packets are transmitted by the polling node at full power so that all on—body nodes
receive such packets. If a node misses a polling packet in a frame due to channel error, it
misses transmission opportunity only in that frame.
7.5.1 Performance Metrics

The primary performance index is the end-to-end Packet Delay (PD), which is
modeled in this chapter and is attempted to be explicitly minimized by the UTILITY,
PRPLC and DVRPLC protocols as presented in Sections 7.3.3, 7.3.4 and 7.3.5. Unlike in
conventional un-partitioned networks, the PD in partitioned on-body networks depends
mainly on the storage delay at the intermediate nodes. Two secondary metrics, namely,
Packet Hop Count (PHC), and Packet Delivery Ratio (PDR) are also recorded for a more
complete understanding. The index PHC captures the number of transmissions per packet
forwarding, indicating the energy expenditure; it does not indicate the packet delay since
the PD in this context depends more on the buffering/ storage than the hop-count.
7.5.2 Traffic Generation and Data Collection

A chosen source node is programmed to generate data packets at the rate of 1 packet

every 4 discrete time slots (each slot is 1.4 see), with a packet size of 46 bytes. As in

188

 

'1"-

 

 

 

Chapter 6, all on-body network nodes are slot-level time-synchronized by the sink node
(i.e. node-6 in Figure 7.1) using periodic synchronization packet broadcast at a high
transmission power [120]. By stamping a packet with the transmission time slot-id by the
source node and subtracting it fiom the reception time slot-id at the sink node, it is
possible to compute the single—trip packet delay (PD) at the sink node. On its way to the
sink node, a data packet collects the entire route information in the form of a list of the
intermediate node-IDs. This allows the extraction and analysis of route information

including the PHC values.

 

41.22\ -.—-/42.56 43.2

12 /

 

 

 

 

u—n
O

 

Average Delay (3)
4} 0) 00

N

   

O

       

OPPT. RAND. UTILITY PRPLC DVRPLC BSDBR

Figure 7.6: On-body packet delivery delay for different DTN routing protocols

7.5.3 Packet Delay (PD)

End-to-end packet delivery delays for a packet from the source node-3 on left upper
arm to the sink node-6 on right ankle for all routing protocols analyzed in Section 7.3 are
reported in Figure 7.6. For each of these protocols, a separate experiment was run for
1320 sec. (i.e. 22 minutes), sending 230 packets, and spanning 6 different body postures
and activities (SIT, SIT-RECLINING, LYING-DOWN, STAND, WALK and RUN),

each lasting for 20 sec. Figure 7.6 reports the average of packet delay computed from the

189

 

 

analytical model, on-body experiment, and off-body simulation using network topology
traces collected during the on-body experiments. The figure also shows the delay
lower-bound obtained by applying the BSDBR benchmark algorithm (presented in
Section 7.4) on the topology traced collected from on-body experiments.

The following observations are be made in Figure 7.6. First, the experimental,
simulation, and model-generated analytical results closely matched across all protocols.
Second, as a general trend the delay performance improves with the amount of
knowledge leveraged on topological locality. Both PRPLC and DVRPLC achieve
significantly better delay compared to the other protocols and very close to BSDBR

benchmark delay, because they are able to capture multi-scale topological localities in

human postural movements using the cost parameters 161.1 (n) and C,j (n) , as explained in

Sections 7.3.4 and 7.3.5. The age-based approach UTILITY uses only the short-term
locality, which explains its larger delay compared to PRPLC and DVRPLC, but smaller
delay than OPPT and RAND, both of which do not leverage any topological locality
information and responds based solely on instantaneous link conditions. Randomized
forwarding provides slightly better delay since in a typically small WBAN, there are only
few possible end-to-end path combinations, leading to quicker delivery than the
Opportunistic mode in which a delivery is possible only when the source directly meets
the destination.

7.5.4 Packet Hop Count (PHC)

Figure 7 .7 shows the average PHC which serves as an indirect measure for
communication energy expenditure (i.e. for transmission and reception) for the on-body

sensors. The large number for RAND explains the impacts of random forwarding

190

 

 

 

compared to all other protocols. The protocols DVRPLC and BSDBR take slightly longer
routes compared to the other protocols, although those two offer better packet delays.
This means that they route packets through better quality links, leading to smaller delays,
even though it requires more number of end-to-end hops. Since with the opportunistic
routing (OPPT) packets are delivered only when a source comes in direct contact of the

destination, all packets are delivered with PHC l.

4.15

Packet Hop Count (PHC)

 

       

OPPT. RAND. UTILITY PRPLC DVRPLC BSDBR

Figure 7.7: Average Packet Hop Count

7.5.5 Packet Delivery Ratio

Since no link layer packet retransmissions are deployed, the system is not able to
recover from the packet drops observed due to the following reason. Due to postural
mobility, there are transient blackout periods during which a neighbor may appear to be
connected in a node’s neighbor table, when in fact it is no longer connected. These
blackout periods are created during a node’s neighbor time-out period, which was chosen
to be two polling frames or 2.8 sec, as reported in Section 6.3.2. Packet transmissions

during such blackout periods end up in packet drops since no link layer reliability is used.

191

All five evaluated protocols suffer from such packet losses, which are captured in the

Packet Delivery Ratio (PDR) as reported below.

90% 89%

ratio

85%

ry
OO
‘3
e\

Packet Delive

75%

71%

               

0
70 /° OPPT. RAND. UTILITY PRPLC DVRPLC

Figure 7.8: Packet drops observed for different protocols

In Figure 7.8, the poor PDR for the OPPT protocol is caused due to a very unreliable
link between the source and the destination nodes (i.e. nodes 3 and 6 in Figure 7.1) which
are physically situated at two extremes of the subject’s body. Since the OPPT protocol
relies on direct source-destination contact for packet delivery, the source-destination link
quality affects this protocol most. For RAND, since the hop count is large (see Figure
7 .7), it is more likely for a packet to encounter the transient blackout period during its
end-to-end trajectory, thus leading to higher drops and low PDs. Lower h0p-counts (see
Figure 7.7) for the locality-based protocols, namely, UTILITY, PRPML, and DVRPLC,

suffers from fewer drops due to the relatively lower occurrences of the transient

 

blackouts. Note that the concept of such drops does not apply for the Benchmark case,

and that is why there is no entry for BSDBR in Figure 7.8.

192

7.5.6 Routing Packets from and to Different Body Segments

Delivery delay computed using the developed model and from on-body experiment
for packets from different body segments to the sink node placed on the right ankle (i.e.
node 6 in Figure 7.1) are shown in Figure 7.9. As shown in the figure, the experimental
and model-generated analytical results closely match across all protocols. Observe that as
the physical distance between a source and the destination increases, the average packet
delivery delays for all the protocols increase. Relatively though, all the experimented
routing protocols maintain the same trend for the packet delay as observed in Section
7.5.3. The average packet hop-counts from source nodes 5, 1, and 3 were experimentally

logged in the range of 1.25, 1.5, and 2.3 respectively.

 

I OPPT. Online 43.2\.. /41 .22

 

        

Ir.
y};

 

       
    
 
  
 
 

  
      
  

  

   
      
  
       
  
    

  
 

   

 
  
  
  
  

  
 
       

a DVRPLC Online

1 5 _ ca OPPT. Model sass »
a RAND. Online 3% 1:31;:
’t'o‘ _ In RAND. Model 53351 it?
v 12 as has
>. : 11111.er Online ,, 2’s:
2 are :\
o 9 _ 121 UTILITY Model gas; :9:
e E1 PRPLC Online get) $135
mi!» ,\’§\‘
5,? [ll PRPLC Model 4%
g
< ,.

E DVRPLC Model

 

 

”gem:- _.' ,'

 

 

       

  
  

 

3 . 3:3:1 :-:w._

'-'-': ‘-— W‘s“ .-..:’4= ::::

'.-.'I’= iii; '.'..':"= /::::

_ . 5:23:53; 5 iii? :-:'§$i§ flags
Src= 5(LeftThigh) =1(Waist) (Up perLeftArm)

Figure 7.9: Delivery delay for packets from thigh, waist and arm to right ankle

Delivery delay for packets fiom the upper left arm (i.e. node 3) to different body
segments are shown in Figure 7.10. It should be noted that although the absolute delay
values are different, the overall trend in packet delay follows the same pattern for all the

sink nodes that we have experimented with.

193

 

 

   
 
   
 
  
 
 

I OPPT. Online
15 _ OPPT. Model
B RAND. Online
1!] RAND. Model
E UTILITY Online
E UTILITY Model
E PRPLC Online

 
  
  
 
      
 
 

    
            
     

’.
y

"’I‘w'
111‘ 1

    

 

  

Average Delay (5).
\O

  

           

 

 

    

u PRPLC Model fig
6 - :1 DVRPLC Online iiޤ
1:11 DVRPLC Model K§
_ :3??? if?
Sink = 2 (Upper Right Arm) Sink = 5 (Left Thigh ) Sink = 6 (Right Ankle)

Figure 7.10: Delivery delay for packets to arm, thigh and ankle from the upper left arm
(i.e. node 3 in Figure 7.1)

7.5.7 Impacts of Postural Stability

For all the experiments so far, each individual physical posture was made to last for
20 sec. In order to study the impacts of variable postural stability on the routing
performance, the subject was instructed to repeat the same sequence of postures as in

Section 6.3, but with different posture durations ranging from 10 sec. to 40 sec.

   
   

 

 

 

 

18 s OPK
»; RAND. -~ 50,5).
5* 14 ~ g)
S 40:13
no .

1 1 an
2 0 E U‘l‘lLlTi _, 302
E 1—3
o .
4.3 6 — __ “r
O s #gmkfl’fl 20%

. m—«~—~
2 ‘" \DVRPDC 10
10 20 Posture Duration (3) 30 40

 

+ DVRPLC Online + DVRPIJC Model + PRPLC Online + PRPIII Model “'9" UTILITY Online
7—9" UTILITY Model —B- RAND. Online ‘43— RAND. Model + OPPT. Online ‘79— OPPT. Model

 

 

Figure 7.11: Impacts of posture duration on packet delay

194

 

Figure 7.11 shows the impacts of posture duration on average packet delay for all
five protocols. Due to its significantly higher values, the packet delays for the OPPT
protocol are plotted as a separate axis in Figure 7 .1 1. Observe that the packet delays for
all the protocols generally increase with higher posture durations. This is because longer
posture duration implies that a connected link remains connected for longer duration and
also a disconnected link remains disconnected longer. As a result, a packet that is
buffered in a node due to network partitioning remains buffered for longer duration,
leading to higher end-to-end delay. In a relative sense, all the experimented protocols
maintain the same performance trend for packet delay as observed in Section 7.5.3,

Figure 7.6.

 

 

 

 

 

Figure 7.12: Experiments with different sensor placements

195

 

 

 

(a)

 
 
 

 

 

 

  
  
 

 

 

 

 

 

  

 

16
E g Online
g 12 IModel
Q
go 8
’12
< 4
0 OPPT. RAND. UTILITY PRPLC DVRPLC
26.44\ : /24.91
16 b
3 ( ) EOnline
g,» 12 IModel
g 8
0
3’
’9
< 4
0

OPPT. RAND. UTILITY PRPLC DVRPLC

Figure 7.13: On-body packet delay for: (a) sensor placement P 1 , and (b) sensor placement P2.
7.5.8 Impacts of Sensor Placements

Additional on-body sensor placements, as shown in Figure 7.12, were experimented
with for evaluating the validity of the routing results obtained so far fi'om the sensor

placement shown in Figure 7.1. Different source and sink nodes are used in the two

 

placement settings P1 and P2 in Figure 7.12. The inter-posture sequence, described in
Section 6.3.2, was followed by a subject and the corresponding packet delay results are
presented in Figure 7.13 for the placement settings P1 and P2 in Figure 7.12. Generally,
the relative performance trends across all the experimented protocols as observed for the

original sensor placement (in Figure 7.6) remain valid for the new sensor placements P1

196

and P2 in Figure 7.12. The delay for placement P1 is larger due to the longer

source-to-destination distance compared to that in P2.

7.6 Evaluation of Node Criticality

The objective of this section is to evaluate the topological criticality of the individual
WBAN sensor nodes. Once identified, such criticality information can be used for
selectively reducing the on-body sensor-count without substantially sacrificing the packet
delivery delay. The mechanism for such analysis is to first remove all links attached to a
given sensor S from the collected topology trace (see Section 7.2), and then run off-line
routing simulation experiments to evaluate the new delay on this reduced network.
Comparison between the delays from the original trace and this reduced trace would
indicate the topological criticality of the target node S from a routing standpoint.

The above mechanism is applied to the 7-node WBAN with node-3 as the source and
node-6 as the destination as shown in Figure 7.1. Figure 7.14 shows the resulting
end-to-end packet delay characteristics when a specific node (chosen from the set 1, 2, 4,
5, and 7) is selectively removed from the network under different routing protocols.
Figure 7.14:3 shows the new delay after a node is removed, and Figure 7.14:b shows the
difference between the new delay and delay obtained from the complete topology without
any node removed. The latter indicates the topological criticality of the removed node
from a routing standpoint. A positive low difference in Figure 7.14:b would indicate that
the removed node is not particularly critical for the corresponding routing protocol.
Conversely, a positive high difference would indicate that the removed node is critical. A
negative difference actually means that the routing performance has improved after the

node is removed. Meaning, the corresponding routing protocol was non-optimally

197

 

 

 

choosing the node after removing which the routing protocol actually. found a better
route, leading to lower delay. Results for all the analyzed protocols except opportunistic
routing (OPPT) are presented in Figure 7.14. Since OPPT relies on direct
source-destination contact for packet delivery, removal of any intermediate node from the

topology does not impact the delivery delay, which is why it is not included.

16 GRAND. QUTILITY EPRPLC .DVRPLC [DBSDBR

 

1—1
h

 

u—ot—i
ON

Average Delay (5)

ONhONOO

All nodes n-l remvd. n-2 remvd. n-4 remvd. n-5 remvd. n-7 remvd.

 

|IRAND.EIU'111r1Y EPRPLC IDVRPLC IrBSDBfl

 

(b)

t" 5° 9°
.1: 4> A
l l 1

Delay Difference (s).

.O
b
l

 

 

 

-0.6 ‘

All nodes n-l remvd. n-2 remvd. n-4 remvd. n-S remvd. n-7 remvd.

Figure 7.14: Node criticality in terms of: (a) Packet delay, and (b) Packet delay difference

Observe in Figure 7.14:b that removing a node from the network creates different

amount of delay difference depending on the specific routing protocol. Note that for the

198

age-based UTILITY-based routing the delay difference is negative, although small, when
any of the nodes 1 or 2 is removed. This indicates that UTILITY was non-optimally
choosing those two nodes and by removing any of them is forcing the protocol to choose
a better route, resulting in negative difference. This inaccuracy is introduced by the
short-term-only connectivity locality in the protocol. On the other hand, by leveraging
both short and long term connectivity locality, as introduced in Sections 7.3.4 and 7.3.5,
the protocols PRPLC and DVRPLC completely eliminates the negative differences,
meaning they always choose an optimal route through nodes removing which actually
worsens the delay. Similarly, in Figure 7.14:b the BSDBR algorithm with benchmark
delay also demonstrates optimal routes indicated by no negative differences. Finally,
observe that the node dependencies of the RAND is generally more than all other
protocols mainly because with randomized forwarding any node in WBAN can be in a
packet’s path as described in Section 7.3.2.

From the above specific analysis and the results in Figure 7.14:b, the following
conclusions can be made. First, if PRPLC or DVRPLC (these two provides the best
end-to-end delay as shown in Figure 7.6) is deployed, one node from the set 1, 2, and 7
can be easily removed without sacrificing packet delay. Second, generally, nodes 4 and 5
are topologically more critical than all other nodes (in the context of source 3 and
destination 6) primary because of their physical proximity to the destination and physical
placement with respect to possible routes from node-3 to node-6. Note that the analysis in
his section shows what happens if only one node is removed from the topology. In order
to find the impacts of removing multiple nodes, similar analysis is needed by removing

different combinations of the nodes and then by measuring the delay differences. Also,

199

 

 

removal of a node will be feasible only if the primary purpose of the node is routing and

 

   

  
    
 
   
   

 

 

 

not sensing.

20 ES RAND. (a)
A m 1117le
£316 a PRPLC
E DVRPLC
a 12 m Bmark
D
g0 8
13'
>
<1 4

        

     

    

          

     

 

 

       

 

 

 

 

 

 

0 All 1.2 1.4 1.5 1.7 2.4 2.5 2.7 4.5 4.7 5.7
rs RAND.
10 15 UTILITY (b)
r; a PRPLC
E 7 DVRPLC
8 ll] Bmark
‘5 4
a
Q 1
32‘
a-Z
-5

All 1.2 1.4 1.5 1.7 2.4 2.5 2.7 4.5 4.7 5.7

Figure 7.15: Two node criticality in terms of: (a) Packet delay, and (b) Packet delay
difference

Figure 7.15 shows the resulting end-to—end packet delay characteristics when two

 

nodes (chosen from the set 1, 2, 4, 5, and 7) are selectively removed at a time. Figure
7.15:a shows the new delay and Figure 7.15:b shows the difference between the delays
new delay and delay obtained from the complete topology without any two nodes
removed. As in Figure 7.14:b, Figure 7.15:b indicates the topological criticality of the
removed two nodes fiom a routing standpoint. For example, with PRPLC or DVRPLC
routing in place, removal of any of the node pairs {1, 2}, {1, 4}, {1, 7}, {2, 4}, {2, 7} at a
time will have insignificant impacts on the routing delay. With a target of topology size

reduction by two nodes, a specific pair can be chosen depending on the sensing criticality

200

 

of the individual nodes from an application perspective. Similar analysis can be
performed with other routing protocols. Also note that similar criticality analysis can be

performed for combination of higher number of nodes.

7.7 Summary and Conclusions

This chapter develops a delay modeling fiamework for store-and-forward packet
routing in Wireless Body Area Networks (WBAN). Using a prototype WBAN for
experimentally characterizing and capturing on-body topology traces, an analytical delay
modeling technique was developed for evaluating single-copy DTN routing protocols.
End-to-end routing delay for a series of protocols including opportunistic, randomized,
and two other mechanisms that capture multi-scale topological localities in human
postural movements have been evaluated. Performance was evaluated experimentally, via
simulation, and using the developed models. It was shown that via multi-scale modeling
of the spatio-temporal locality of on-body link disconnection patterns, it is possible to
attain better delay performance compared to opportunistic, randomized and utility-based
DTN routing protocols in the literature. Finally, a mechanism for evaluating the
topological importance of individual on-body sensor nodes is developed. It is shown that
such information can be used for selectively reducing the on-body sensor-count without
substantially sacrificing the packet delivery delay. Future work on this topic includes
developing a Kalman Filter-based body movement prediction model for predictive

on—body packet routing with lower delay objectives.

201

 

 

Chapter 8. Summary and Future Work

8.1 Contributions

In this thesis, a framework for network integrated sensing and energy-aware protocols
to address the resource constraints inherent to Wireless Body Area Networks (WBANS)
were investigated. Available energy, processing, and form factor constraints of a
body-mounted or -implanted sensor can severely limit its abilities for sensing, and the
power and range of radio transmission for communication purposes. In this thesis we
have investigated sensing, processing, and protocol solutions to address these system
constraints.

Chapter 3 has developed a novel network-integrated sensing modality, namely,
inter-sensor relative proximity, which is inferred from the measured Received Signal
Strength Indicator (RSSI) of the Radio Frequency (RF) signal between each pair of
WBAN sensors. By using this new sensing modality, non activity-intensive postures can
be detected without having to add extra components and the associated energy overheads.
An experimental system was developed for integrating the proximity data with other
traditional modalities, such as, acceleration and orientation. This multi-modal data was
then processed using a stochastic processing framework Hidden Markov Model (HMM)
for accurate posture identification in a subj ect-independent manner.

Data transaction across on-body sensors can be point-to-point or multi-point-to-point
depending on specific applications. The optimal transmission power required for an
on-body wireless link between two sensors depends on the physical distance of the link
and its instantaneous channel condition. The distance can vary with mobility driven by

human postures, and the channel condition can change due to unpredictable RF

202

 

 

 

attenuation caused by a slew of factors including antenna orientation, clothing, and
physical stature of specific subject individuals. An on-body link’s postural position can
be highly dynamic due to the reasons listed above. As a result, static pre-defined
transmission power is not able to provide continuous link connectivity while ensuring
minimum required transmission power consumption. Therefore, a dynamic link power
assignment mechanism for optimal energy management is needed.

In Chapter 4, three different on-body link transmission power assignment
mechanisms, namely, Linear, Binary and Dynamic Postural Position Inference (DPPI),
were proposed. In these approaches, the receiver/ control node computes a new desirable
transmission power level for the transmitter node by incrementing or decrementing the
current power level based on the received RSSI values with respect an RSSI threshold
range. No changes are needed when the RSSI falls within the range. This closed-loop
control logic ensures that, for a given postural position, this process eventually achieves a
desirable transmission power level so that the RSSI at the receiver falls within the preset
threshold range.

The proposed transmission power control mechanism was then enhanced in Chapter
5 by modeling human body movement as a stochastic linear system and a quantized
Linear Quadratic Gaussian control with an Integrator (LQGI). The objective was to
develop a model-based power control framework in which RF signal strength was
predicted and regulated at a reference value to enhance the overall energy performance of
an on-body wireless sensor.

Chapter 6 and Chapter 7 of this thesis focused on on-body packet routing in the

presence of topological partitioning caused due to ultra-short wireless range and postural

203

 

 

 

body movements. Short transmission range is a common constraint for low-power RF
transceivers designed for embedded applications with limited energy, often supplied by
harvested operations. Such situations are particularly pertinent for implantable body
sensors. Low RF transmission ranges also mean that postural body movements can give
rise to frequent partitioning or disconnection in WBAN topologies, resulting in a body
area Delay Tolerant Network (DTN). Such topological partitioning can often get
aggravated by the unpredictable RF attenuation caused due to signal blockage by clothing
material and body segments. Topological partitioning necessitates store-and-forward
routing protocols with performance goals of low end-to-end packet delay, packet loss,
and transmission energy consumption.

Chapter 6 explored and developed on-body store-and-forward packet routing
algorithms in the presence of network partitioning. The objective was to minimize
end-to-end packet delay by dynamically choosing routes on which the storage/buffering
delays are low. While ensuring low storage delay, the algorithms also attempt to
minimize the end-to-end hop-count so that the transmission energy drainage is
minimized, thus leading to long network operating durations.

Using a prototype WBAN, a topology characterization mechanism has been
developed in Section 6.3 in order to demonstrate the network partitioning caused due to
human postural mobility. Then, an on-body packet routing framework was developed in
Section 6.4 using a stochastic link cost formulation, reflecting the body postural trends.
The performance of the proposed protocols was evaluated experimentally and through

simulation using the prototype WBAN. The performance of the routing protocols were

204

 

‘1

 

then compared with the existing single copy DTN routing protocols, and a specialized
flooding-based benchmark algorithm that provides routing delay lower-bounds.

In Chapter 7, analytical techniques for modeling end-to-end packet delay for the
proposed single copy DTN routing algorithms were developed. DTN routing delay
obtained from the developed model were then compared with results from on-body
experiments from the prototype WBAN and off-body simulation carried out with network
topology traces obtained from the prototype WBAN. Using the model and the topology
trace data, a detailed analysis was then carried out in Section 7.6 for identifying
non-critical links in order to design a minimal WBAN topology from the routing stand
point.

8.2 Future Work

In Chapter 6 and Chapter 7, it was assumed that the sensor nodes have infinite and
continuous source of energy. But in reality, this may not be true for the implanted
sensors. Instead, the energy for the implanted WBAN sensors are likely to come from
harvested energy sources such as piezoelectric, magnetic, and thermo-electric generators
[3 8-43]. While the energy-harvesting sources can vary widely
[123],[124],[58],[125-127], energy harvesting using vibration of piezoelectric sensors
[38],[40-43],[51],[53-56],[127-136] are particularly relevant for the WBANS. Since the
piezoelectric energy harvesting depends on movements [38],[42], the amount of energy
harvested at a specific on-body sensor node will depend on the movement pattern of the
body part that the node is attached to. As a result, the specific energy generation profile at

the WBAN nodes does depend on the postural body movement patterns over time.

205

 

This dynamic energy generation profile provides an additional layer of constraints on
the routing problem that was discussed in Chapter 6. To deal with this new constraint, an
energy-aware [137-147],[75] store-and-forward packet routing protocol is necessary in
the presence of network partitioning and harvested energy [88],[138],[139],[l48-156].
The goal will be to minimize end-to-end packet delay, packet loss, and transmission
energy.

The objective is to minimize end-to-end packet delays by dynamically choosing
routes on which the storage/buffering delays are low. While ensuring low storage delay,
the algorithm should also attempt to minimize the variance of the delay across all packets
generated at different nodes, because low-harvested nodes may have tendency of
buffering packets for too long. In addition, the algorithm should also work to:

o Minimize the average per node down time.

o Minimize the variance of downtime across all nodes.

0 Minimize average packet loss and the variance across all sources.

0 Minimize the end-to-end hop-count so that the transmission energy drainage is

minimized, thus leading to long network operating durations.

We will develop energy-aware on-body routing mechanisms for incorporating the

above objectives.

206

 

[1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

References

E. Jovanov, A. Milenkovic, C. Otto, P. De Groen, B. Johnson, S. Warren, and G
Taibi, “A WBAN System for Ambulatory Monitoring of Physical Activity and
Health Status: Applications and Challenges,” Engineering in Medicine and Biology
Society, 2005. IEEE-EMBS 2005. 27th Annual International Conference of the,
2005, pp. 3810-3813.

E. Jovanov, A. Milenkovic, C. Otto, and P.C.D. Groen, “A Wireless Body Area
Network of Intelligent Motion Sensors for Computer Assisted Physical
Rehabilitation,” Journal NeuroEng. and Rehab, vol. 2, Mar. 2005, p. 6.

R. Bartalesi, F. Lorussi, M. Tesconi, A. Tognetti, G Zupone, and D. De Rossi,
“Wearable kinesthetic system for capturing and classifying upper limb gesture,”
Eurohaptics Conference, 2005 and Symposium on Haptic Interfaces for Virtual

Environment and Teleoperator Systems, 2005. World Haptics 2005. First Joint,
2005, pp. 535-536.

Yun Liang, Abhik Roychoudhury, and Tulika Mitra, “Timing Analysis of Body
Area Network Applications,” Dec. 2008.

D. Simic, A. Jordan, Rui Tao, N. Gungl, J. Simic, M. Lang, Luong Van Ngo, and V.
Brankovic, “Impulse UWB Radio System Architecture for Body Area Networks,”
Mobile and Wireless Communications Summit, 2007. 16th 1ST, 2007, pp. 1-5.

R. Istepanian, E. J ovanov, and Y. Zhang, “Guest Editorial Introduction to the
Special Section on M-Health: Beyond Seamless Mobility and Global Wireless
Health-Care Connectivity,” Information Technology in Biomedicine, IEEE
Transactions on, vol. 8, 2004, pp. 405-414.

M. Martinez-Espronceda, L. Serrano, I. Martinez, J. Escayola, S. Led, J. Trigo, and
J. Garcia, “Implementing ISO/IEEE 11073: proposal of two different strategic
approaches,” Conference Proceedings: Annual International Conference of the
IEEE Engineering in Medicine and Biology Society. IEEE Engineering in
Medicine and Biology Society. Conference, vol. 2008, 2008, pp. 1805-1808.

L. Schmitt, T. Falck, F. Wartena, and D. Simons, “Novel ISO/IEEE 11073
Standards for Personal Telehealth Systems Interoperability,” Proceedings of the
2007 Joint Workshop on High Confidence Medical Devices, Software, and Systems
and Medical Device Plug-and-Play Interoperability, IEEE Computer Society, 2007,
pp. 146-148.

J. Yao and S. Warren, “Applying the ISO/IEEE 11073 Standards to Wearable
Home Health Monitoring Systems,” Journal of Clinical Monitoring and
Computing, vol. 19, 2006, pp. 427-436.

207

 

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

J. Yao and S. Warren, “Applying the ISO/IEEE 11073 Standards to Wearable
Home Health Monitoring Systems,” Journal of Clinical Monitoring and
Computing, vol. 19, 2006, pp. 427-436.

ISO - International Organization for Standardization and ISO - International
Organization for Standardization, “ISO - International Organization for
Standardization.”

Shu-Di Bao, Yuan-ng Zhang, and Lian—Feng Shen, “Physiological Signal Based
Entity Authentication for Body Area Sensor Networks and Mobile Healthcare
Systems,” Engineering in Medicine and Biology Society, 2005. IEEE-EMBS 2005.
27th Annual International Conference of the, 2005, pp. 2455-2458.

B. Lo, S. Thiemjarus, R. King, and G. Yang, “Body Sensor Network — A Wireless
Sensor Platform for Pervasive Healthcare Monitoring,” Proceddings of the 3rd
International Conference on Pervasive Computing (PERVASIVE 2005), May. 2005,
pp. 77-80.

C. Otto, A. Milenkovic, C. Sanders, and E. Jovanov, “System Architecture of a
Wireless Body Area Sensor Network for Ubiquitous Health Monitoring,” Journal
of Mobile Multimedia, vol. 1, 2006, pp. 307-326.

M. Moh, B. Culpepper, Lan Dung, Teng-Sheng Moh, T. Harnada, and Ching-Fong
Su, “On data gathering protocols for in-body biomedical sensor networks,” Global
Telecommunications Conference, 2005. GLOBECOM '05. IEEE, 2005, pp. 6
pp=2996.

A. Milenkovic, C. Otto, and E. Jovanov, “Wireless Sensor Networks for Personal
Health Monitoring: Issues and an Implementation,” Special issue: Wireless Sensor

Networks: Performance, Reliability, Security, and Beyond, 2006.

A. Milenkovic, C. Otto, and E. Jovanov, “Wireless sensor networks for personal

health monitoring: Issues and an implementation,” Computer Communications,
vol. 29, 2006, pp. 2521—2533.

R.R. Pate, K.A. Pfeiffer, S.G Trost, P. Ziegler, and M. Dowda, “Physical Activity
Among Children Attending Preschools,” Pediatrics, vol. 114, Nov. 2004, pp.
1258-1263.

K. Chen and DJ. Bassett, “The Technology of Accelerometry-based Activity
Monitors: Current and Future,” Med Sci Sports Exerc, vol. 37, 2005, pp.
490-500.

J. Brusey, R. Rednic, E.I. Gaura, J. Kemp, and N. Poole, “Postural activity
monitoring for increasing safety in bomb disposal missions,” Measurement Science
and Technologi, vol. 20, 2009, p. 075204.

M.J.C. Almeida, K.L. Mclver, K.A. Pfeiffer, M. Dowda, and RR. Pate, “Validation

208

 

[221

[231

[24]

[25]

[26]

[27]

[28]

[29]

[30]

[31]

[32]

[33]

and Calibration of the Actical Accelerometer in Preschool Children.”

S.G Trost, K.L. Mclver, and RR. Pate, “Conducting accelerometer-based activity
assessments in field-based research,” Medicine and Science in Sports and Exercise,
vol. 37, Nov. 2005, pp. S531-543.

G.J. Welk, “Principles of design and analyses for the calibration of
accelerometry-based activity monitors,” Medicine and Science in Sports and
Exercise, vol. 37, Nov. 2005, pp. 8501-511.

S. Biswas and M. Quwaider, “Remote Monitoring of Soldier Safety through Body
Posture Identification using Wearable Sensor Networks,” SPIE Defense and
Security Symposium, Multisensor; Multisource Information Fusion: Architectures,
Algorithms, and Applications, 2008, pp. 1-14.

K. Venkatasubramanian, A. Banerjee, and S. Gupta, “Plethysmogram—based secure
inter-sensor communication in Body Area Networks,” Military Communications
Conference, 2008. MILCOM 2008. IEEE, 2008, pp. 1-7.

E. Jovanov, D. Raskovic, J. Price, J. Chapman, A. Moore, and A. Krishnamurthy,
“Patient monitoring using personal area networks of wireless intelligent sensors,”
Biomedical Sciences Instrumentation, vol. 37, 2001, pp. 373-378.

K. Yazdandoost and R. Kohno, “UWB antenna for wireless body area network,”
Microwave Conference, 2006. APMC 2006. Asia-Pacific, 2006, pp. 1647-1652.

C. Cordeiro and M. Patel, “Body area networking standardization: present and
future directions,” Proceedings of the ICST 2nd international conference on Body
area networks, Florence, Italy: ICST (Institute for Computer Sciences,
Social-Informatics and Telecommunications Engineering), 2007, pp. 1-2.

K. Lorincz, D.J. Malan, T.R.F. Fulford-Jones, A. Nawoj, A. Clavel, V. Shnayder, G
Mainland, M. Welsh, and S. Moulton, “Sensor networks for emergency response:
Challenges and opportunities,” IEEE Pervasive Computing, vol. 3, 2004, pp.
16-23.

GZ. Yang and M. Yacoub, Body sensor networks, Springer-Verlag New York Inc,
2006.

B. Lo and GZ. Yang, “Key technical challenges and current implementations of
body sensor networks,” IEE Proceedings of the 2nd International Workshop on
Body Sensor Networks (BSN 2005), 2005, pp. 1-5.

B. Gyselinclcx, J. Penders, and R. Vullers, “Potential and challenges of body area

networks for cardiac monitoring,” Journal of Electrocardiology, vol. 40, 2007,
pp. 165-168.

A. Ylisaukko-oja, E. Vildjiounaite, and J. Mantyjarvi, “Five-point acceleration

209

 

 

[34]

[351

[36]

[371

[38]

[39]

[40]

[41]

[421

[43]

[441

sensing wireless body area network - design and practical experiences,” Wearable
Computers, 2004. IS WC 2004. Eighth International Symposium on, 2004, pp.
184-185.

E. Farella, A. Pieracci, L. Benini, and A. Acquaviva, “A Wireless Body Area
Sensor Network for Posture Detection,” Computers and Communications, 2006.
ISCC '06. Proceedings. 11th IEEE Symposium on, 2006, pp. 454-459.

Seon-Woo Lee and K. Mase, “Activity and location recognition using wearable
sensors,” Pervasive Computing, IEEE, vol. 1, 2002, pp. 24-32.

A. Krause, D. Siewiorek, A. Smailagic, and J. Farringdon, “Unsupervised,
dynamic identification of physiological and activity context in wearable
computing,” Wearable Computers, 2003. Proceedings. Seventh IEEE International
Symposium on, 2003, pp. 88-97.

N. Ravi, N. Dandekar, P. Mysore, and ML. Littman, “Activity recognition from
accelerometer data,” PROCEEDINGS OF THE NATIONAL CONFERENCE ON
ARTIFICIAL INTELLIGENCE, Menlo Park, CA; Cambridge, MA; London; AAAI
Press; MIT Press; 1999, 2005, p. 1541.

E. Minazara, D. Vasic, and F. Costa, “Piezoelectric Generator Harvesting Bike
Vibrations Energy to Supply Portable Devices.”

M.P. Rothney, M. Neumann, A. Beziat, and KY Chen, “An artificial neural
network model of energy expenditure using nonintegrated acceleration signals,” J
Appl Physiol, vol. 103, Oct. 2007, pp. 1419-1427.

L. Mateu and F. Moll, “System-Level Simulation of a Self-Powered Sensor with
Piezoelectric Energy Harvesting,” Sensor Technologies and Applications, 2007.
SensorComm 2007. International Conference on, 2007, pp. 399-404.

S. Priya, R. Taneja, R. Myers, and R. Islam, “Piezoelectric Energy Harvesting
using Bulk Transducers,” Piezoelectric and Acoustic Materials for Transducer
Applications, 2008, pp. 373-388.

E. Minazara, D. Vasic, F. Costa, and G Poulin, “Piezoelectric diaphragm for
vibration energy harvesting,” Ultrasonics, vol. 44, Dec. 2006, pp. e699-e703.

Q. Zheng and Y. Xu, “Asymmetric air-spaced cantilevers for vibration energy
harvesting,” Smart Materials and Structures, vol. 17, 2008, p. 055009.

Sungdae Choi, Seong-Jun Song, Kyomin Sohn, Hyejung Kim, Jooyoung Kim, J.
Yoo, and Hoi-Jun Yoo, “A Low-power Star-topology Body Area Network
Controller for Periodic Data Monitoring Around and Inside the Human Body,”
Wearable Computers, 2006 10th IEEE International Symposium on, ISWC, 2006,
pp. 139-140.

210

 

[45]

[46]

[47]

[48]

[49]

[50]

[51]

[52]

[53]

[54]

[551

M. Hanson, H. Powell, A. Barth, K. Ringgenberg, B. Calhoun, J. Aylor, and J. Lach,
“Body Area Sensor Networks: Challenges and Opportunities,” Computer, vol. 42,
2009, pp. 58-65.

Y. Lin, B. Veeravalli, and S. Viswanathan, “Critical-Path based Low-Energy
Scheduling Algorithms for Body Area Network Systems,” RT CSA '07:
Proceedings of the 13th IEEE International Conference on Embedded and

Real-Time Computing Systems and Applications, IEEE Computer Society, 2007, pp.
308, 301.

M. Walsh, M. Hayes, and J. Nelson, “Robust performance for an energy sensitive
wireless body area network — an anti-windup approach,” International Journal of
Control, vol. 82, 2009, p. 59.

I. Lamprinos, A. Prentza, E. Sakka, and D. Koutsouris, “Energy-efficient MAC
Protocol for Patient Personal Area Networks,” Engineering in Medicine and
Biology Society, 2005. IEEE-EMBS 2005. 2 7th Annual International Conference of
the, 2005, pp. 3799-3802.

T. Norgall, “Body Area Network — a Key Infiastructure Element for
Patient-Centric Health Services,” ISO TC215 / WG7 / IEEE 1073 Meeting Berlin,
2005, pp. 9-11.

I. Culpepper, Lan Dung, and M. Moh, “Hybrid indirect transmissions (HIT) for
data gathering in wireless micro sensor networks with biomedical applications,”
Computer Communications, 2003. CC W 2003. Proceedings. 2003 IEEE 18th
Annual Workshop on, 2003, pp. 124-133.

J. Ajitsaria, S.Y. Choe, D. Shen, and DJ. Kim, “Modeling and analysis of a
bimorph piezoelectric cantilever beam for voltage generation,” Smart Materials
and Structures, vol. 16, 2007, pp. 447-454.

V.R. Challa, M.G. Prasad, Y. Shi, and ET. Fisher, “A vibration energy harvesting
device with bidirectional resonance frequency tunability,” Smart Materials and
Structures, vol. 17, 2008, p. 015035.

H. Fang, J. Liu, Z. Xu, L. Dong, L. Wang, D. Chen, B. Cai, and Y. Liu,
“Fabrication and performance of MEMS-based piezoelectric power generator for
vibration energy harvesting,” Microelectron. J., vol. 37, 2006, pp. 1280-1284.

M. Ferrari, D. Marioli, and A. Taroni, “Modeling, Fabrication and Performance
Measurements of a Piezoelectric Energy Converter for Power Harvesting in
Autonomous Microsystems,” Instrumentation and Measurement, IEEE
Transactions on, vol. 55, 2006, pp. 2096-2101.

S. Jiang, X. Li, S. Guo, Y. Hu, J. Yang, and Q. Jiang, “Performance of a
piezoelectric bimorph for scavenging vibration energy,” Smart Materials and
Structures, vol. 14, 2005, pp. 774, 769.

211

 

 

[56]

[57]

[581

[59]

[60]

[61]

[62]

[63]

[64]

[65]

[66]

S. Roundy and P. Wright, “A piezoelectric vibration based generator for wireless
electronics,” Smart Materials and Structures, vol. 13, 2004, pp. 1142, 1131.

D. Shen, J. Park, J. Ajitsaria, S. Choe, H.C.W. III, and D. Kim, “The design,
fabrication and evaluation of a MEMS PZT cantilever with an integrated Si proof
mass for vibration energy harvesting,” Journal of Micromechanics and
Microengineering, vol. 18, 2008, p. 055017.

S.P. Beeby, M.J. Tudor, and NM. White, “Energy harvesting vibration sources for
microsystems applications,” Measurement science and technology, vol. 17, 2006,
p. 175.

“WebPage of Department of Health and Human Services,”
wwwgrantsZ.nih.gov/grants/guide/pa-files/PA-07-354.htrnl.

M. Quwaider and S. Biswas, “Physical Context Detection using Multi-modal
Sensing using Wearable Wireless Networks,” Journal of Communication Software
and Systems (JCOMSS ’08), Special Issue on Medical Applications for WSN, vol.
4, 2008, pp. 191-202.

M. Quwaider and S. Biswas, “Body posture identification using hidden Markov
model with a wearable sensor network,” Proceedings of the ICST 3rd international
conference on Body area networks, Tempe, Arizona: ICST (Institute for
Computer Sciences, Social-Informatics and Telecommunications Engineering),
2008, pp. 1-8.

S. Lin, J. Zhang, G. Zhou, L. Gu, J. Stankovic, and T. He, “ATPC: adaptive
transmission power control for wireless sensor networks,” SenSys '06: Proceedings

of the 4th international conference on Embedded networked sensor systems, ACM
Press, 2006, pp. 236, 223.

G Zhou, T. He, S. Krishnamurthy, and J. Stankovic, “Impact of radio irregularity
on wireless sensor networks,” MobiSys '04: Proceedings of the 2nd international
conference on Mobile systems, applications, and services, ACM Press, 2004, pp.

138, 125.

S. Xiao, V. Sivaraman, and A. Burdett, “Adapting radio transmit power in wireless
body area sensor networks,” Proceedings of the ICST 3rd international conference
on Body area networks, Tempe, Arizona: ICST (Institute for Computer Sciences,
Social-Informatics and Telecommunications Engineering), 2008, pp. 18

S. Mikarni, T. Matsuno, M. Miyama, M. Yoshimoto, and H. Ono, “A
Wireless-Interface SoC Powered by Energy Harvesting for Short-range Data
Communication,” Asian Solid-State Circuits Conference, 2005, 2005, pp. 241-244.

D. Sagan, “RF Integrated Circuits for Medical Applications: Meeting the

Challenge of Ultra Low Power Communication,” Ultra-Low-Power
Communications Division, Zarlink Semiconductor, 2005.

212

 

[67]

[68]

[69]

[70]

[71]

[72]

[73]

[74]

[75]

[76]
[77]
[781
[79]

[80]

[81]

E. Strommer, M. Hillukkala, and A. Ylisaukkooja, “Ultra-low Power Sensors with
Near Field Communication for Mobile Applications,” International Conference on
Wireless Network, 1C WN’07, 2007, pp. 1-12.

T. Falck, H. Baldus, J. Espina, and K. Klabunde, “Plug 'n Play Simplicity for
Wireless Medical Body Sensors,” Pervasive Health Conference and Workshops,
2006, 2006, pp. 1-5.

E. Jones and P. Ward, “Routing strategies for delay-tolerant networks,” ACM
Computer Communication (CCR), 2006.

Hui Guo, Jiang Li, Yi Qian, and Ye Tian, “A practical routing strategy in Delay
Tolerant Networks using multiple pigeons,” Military Communications Conference,
2008. MILCOM 2008. IEEE, 2008, pp. l-7.

J. Leguay_afi1 n2, T. Friedman, and V. Conan, “Evaluating MobySpace-based
routing strategies in delay-tolerant networks: Research Articles,” Wirel. Commun.
Mob. Comput., vol. 7, 2007, pp. 1171-1182.

V. Conan, J. Leguay, and T. Friedman, “Fixed point opportunistic routing in delay
tolerant networks,” Selected Areas in Communications, IEEE Journal on, vol. 26,
2008, pp. 773-782.

J. Leguay, T. Friedman, and V. Conan, “Evaluating Mobility Pattern Space Routing
for DTNs,” INF OCOM 2006. 25th IEEE International Conference on Computer
Communications. Proceedings, 2006, pp. 1-10.

A. Lindgren, A. Doria, and O. Schelén, “Probabilistic routing in intermittently
connected networks,” SIGMOBILE Mob. Comput. Commun. Rev., vol. 7, 2003,
pp. 19-27.

“Link to Cipcon SmartRF CC1000 Datasheet,”
http://www. chipcon. com/fi les/C C 1 000_Data_Sheet_2_I .pdf.

“Link to chipcon.com/,” http://www. chipcon.com/.
“Webpage of TinyOS Technology,” http://www. tinyos.net/.
“Crossbow Technology, Inc.,” http://wwwxbow. com.

L. Bao and S. Intille, “Activity Recognition from User-Annotated Acceleration
Data,” Pervasive 2004, Apr. 2004, pp. 17, 1.

“Link to posture detection videoz,”
http://www. egr: msu. edu/~sbiswas/Research/Wearable. mpg.

L. Liporace, “Maximum likelihood estimation for multivariate observations of
Markov sources,” Information Theory, IEEE Transactions on, vol. 28, 1982, pp.

213

 

[82]

[83]

[84]

[85]

[86]

[87]

[88]

[89]

[90]

[91]

[92]

[93]

729-734.

B.H. Juang, “Maximum-likelihood estimation for mixture multivariate stochastic
observation of Markov chains,” AT & T TECH. J., vol. 64, 1985, pp. 1235-1250.

L. Rabiner, “A tutorial on hidden Markov models and selected applications in
speech recognition,” Proceedings of the IEEE, vol. 77, 1989, pp. 257-286.

L. Wang, M. Mehrabi, and EL Kannatey-Asibu, “Hidden Markov Model-based
Tool Wear Monitoring in Turning,” Manufacturing Science and Engineering, vol.
124, 2002, pp. 651-658.

J. Allanach, Haiying Tu, S. Singh, P. Willett, and K. Pattipati, “Detecting, tracking,
and counteracting terrorist networks via hidden Markov models,” Aerospace
Conference, 2004. Proceedings. 2004 IEEE, 2004, p. 3257 Vol.5.

V.N. James and J. Clark, “Automated Visual Surveillance Using Hidden Markov
Models,” International Conference on Vision Interface, 2002, pp. 88-93.

0. Brdiczka, J. Maisonnasse, and P. Reignier, “Automatic detection of interaction

groups,” Proceedings of the 7th international conference on Multimodal interfaces,
Torento, Italy: ACM, 2005, pp. 32-36.

T. Lim and G Mohan, “Energy aware geographical routing and topology control to
improve network lifetime in wireless sensor networks,” Broadband Networks,
2005. BroadNets 2005. 2nd International Conference on, 2005, pp. 771-773 Vol. 2.

Guoliang Xing, Chenyang Lu, and R. Pless, “Localized and Configurable
Topology Control in Lossy Wireless Sensor Networ ,” Computer
Communications and Networks, 2007. ICCCN 2007. Proceedings of 16th
International Conference on, 2007, pp. 75-80.

J. Liu and B. Li, “Distributed topology control in wireless sensor networks with ’
asymmetric links,” Global Telecommunications Conference, 2003. GLOBECOM
'03. IEEE, 2003, pp. 1257-1262 vol.3.

U. Klee, T. Gehrig, and J. McDonough, “Kalman filters for time delay of
arrival-based source localization,” EURASIP J. Appl. Signal Process, vol. 2006,
2006, pp. 167-167.

P. Nar and E. Cayirci, “PCSMAC: a power controlled sensor-MAC protocol for
wireless sensor networks,” Wireless Sensor Networks, 2005. Proceeedings of the
Second European Workshop on, 2005, pp. 81-92.

Daji Qiao, Sunghyun Choi, Amit Jain, and K. Shin, “Adaptive transmit power

control in IEEE 802.11a wireless LANS,” Vehicular Technology Conference, 2003.
VTC 2003-Spring. The 57th IEEE Semiannual, 2003, pp. 433-437 VOL].

214

 

[94] S.A. Srikanth, S. Agarwal, S.V. Krishnamurthy, R.H. Katz, and SK. Dao,
“Distributed Power Control in Ad-hoc Wrreless Networks,” IEEE International
Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC),
vol. 2, 2001, pp. 59--66.

[95] E. Lopez-Aguilera and J. Casademont, “A transmit power control proposal for
IEEE 802.11 cellular networks,” 6th International Workshop on on Applications
and Services in Wireless Networks, New York: Institute of Electrical and
Electronics Engineers, 2006, pp. 235-242.

[96] H. Saghaei and A. Neyestanak, “Variable Step Closed-Loop Power Control in
Cellular Wireless CDMA Systems under Multipath Fading,” Communications,
Computers and Signal Processing, 2007. PacRim 2007. IEEE Pacific Rim
Conference on, 2007, p. 157-160.

[97] M. Rintamaki, H. Koivo, and I. Hartirno, “Adaptive closed-loop power control
algorithms for CDMA cellular communication systems,” Vehicular Technology,
IEEE Transactions on, vol. 53, 2004, pp. 1756-1768.

[98] Rintamaki, “Power control in CDMA cellular communication, systems,” Wiley
Encyclopedia of Telecommunications, J. G Proakis, 2002.

[99] K. Srinivasan, P. Dutta, A. Tavakoli, and P. Levis, “Understanding the causes of
packet delivery success and failure in dense wireless sensor networks,” SenSys '06:
Proceedings of the 4th international conference on Embedded networked sensor

systems, Boulder, Colorado, USA: ACM, 2006, pp. 420, 419.

[100] N. Dankwa, “An Evaluation of Transmit Power Levels for Node Localization on
the Mica2 Sensor Node,” signal, vol. 5, 2004, p. 03.

[101] J. Kim, S. Chang, and Y. Kwon, “ODTPC: On-demand Transmission Power
Control for Wireless Sensor Networks,” Information Networking, 2008. [COIN
2008. International Conference on, 2008, pp. 5, l.

[102] J. Choi, R Morales, Y. Zhao, and A. Majurndar, “Design and Control of a Thermal
Stabilizing System for a MEMS Optomechanical Uncooled Infrared Imaging
Camera,” Elsevier Sensors and Actuators, vol. 203, 2003, pp. 132-142.

[103] M. Athans, “The role and use of the stochastic linear-quadratic-Gaussian problem
in control system design,” Automatic Control, IEEE Transactions on, vol. 16,
1971, pp. 529-552.

[104] K. Zhou, J. Doyle, and K. Glover, Robust and Optimal Control, Prentice Hall
Upper Saddle River, NJ, 1996.

[105] A. Maddi, A. Guessoum, and D. Berkani, “Using Linear Quadratic Gaussian
Optimal Control for Lateral Motion of Aircraft,” World Academy of Science,
Engineering And Technology, vol. 37, 2009.

215

 

[106] E. Maneri and W. Gawronski, “Linear Quadratic Gaussian Controller Design
Using a Graphical User Interface: Application to the Beam-Waveguide Antennas,”
Oct. 1999.

[107] DP. Looze, “Linear—quadratic-Gaussian control for adaptive optics systems using a
hybrid model,” Journal of the Optical Society of America A, vol. 26, Jan. 2009,

pp. 1-9.

[108] T. Spyropoulos, K. Psounis, and CS. Raghavendra, “Efficient routing in
intermittently connected mobile networks: The multiple-copy case,” IEEE/A CM
Transactions on Networking (TON), vol. 16, 2008, pp. 77—90.

[109] H. Dubois-ferriere, M. Grossglauser, and M. Vetterli, “Age matters: Efficient route
discovery in mobile ad hoc networks using encounter ages,” 2003.

[110] J. Leguay, T. Friedman, and V. Conan, “DTN routing in a mobility pattern space,”
Proceedings of the 2005 ACM SIGCOMM workshop on Delay-tolerant networking,
Philadelphia, Pennsylvania, USA: ACM, 2005, pp. 276-283.

[111] B. Braem, B. Latre, I. Moerman, C. Blondia, and P. Demeester, “The Wireless
Autonomous Spanning tree Protocol for Multihop Wireless Body Area Networ ,”
Mobile and Ubiquitous Systems - Workshops, 2006. 3rd Annual International
Conference on, 2006, pp. 1-8.

[112] B. Latre, B. Braem, I. Moerman, C. Blondia, E. Reusens, W. Joseph, and P.
Demeester, “A Low-delay Protocol for Multihop Wireless Body Area Networks,”
Mobile and Ubiquitous Systems: Networking & Services, 2007. MobiQuitous 2007.
Fourth Annual International Conference on, 2007, pp. 1-8.

[113] S. Jain, K. Fall, and R. Patra, “Routing in a delay tolerant network,” Proceedings
of the 2004 conference on Applications, technologies, architectures, and protocols
for computer communications, 2004, pp. 145—158.

[114] A. Jindal and K. Psounis, “Performance analysis of epidemic routing under
contention,” Proceedings of the 2006 international conference on Wireless
communications and mobile computing, Vancouver, British Columbia, Canada:
ACM, 2006, pp. 539-544.

[115] A. Vahdat and D. Becker, “Epidemic Routing for Partially Connected Ad Hoc
Networks,” Technical Report CS-200006, Duke University, 2000.

[116] T. Spyropoulos, K. Psounis, and C. Raghavendra, “Efficient Routing in
Intermittently Connected Mobile Networks: The Single-Copy Case,” Networking,
IEEE/ACM Transactions on, vol. 16, 2008, pp. 63-76.

[117] S. Burleigh, A. Hooke, L. Torgerson, K. Fall, V. Cerf, B. Durst, K. Scott, and H.
Weiss, “Delay-tolerant networking: an approach to interplanetary Internet,”
Communications Magazine, IEEE, vol. 41, 2003, pp. 128-136.

216

[118] D. Thakore and S. Biswas, “Routing with persistent link modeling in intermittently
connected wireless networks,” Military Communications Conference, 2005.
MILCOM 2005. IEEE, 2005, pp. 461-467 Vol. 1.

[119] M. Shin, S. Hong, and l. Rhee, “DTN routing strategies using optimal search
patterns,” Proceedings of the third ACM workshop on Challenged networks, San
Francisco, California, USA: ACM, 2008, pp. 27-32.

[120] M. Quwaider and S. Biswas, “Disruption Tolerant Network Routing in Body
Sensor Networks with Dynamic Postural Partitioning,” In Press, To Appear;
Elsevier Ad Hoc Networks, 2010.

[121] M. Quwaider and S. Biswas, “Probabilistic routing in on-body sensor networks
with postural disconnections,” Proceedings of the 7th ACM international
symposium on Mobility management and wireless access, Tenerife, Canary

Islands, Spain: ACM, 2009, pp. 149-158.

[122] M. Quwaider and S. Biswas, “On-body Packet Routing Algorithms for Body
Sensor Networks,” NE T COM '09. First International Conference on Networks and
Communications, Dec. 2009, pp. 171-177.

[123] SP. Beeby, M.J. Tudor, and NM. White, “Energy harvesting vibration sources for
microsystems applications,” Dec. 2006.

[124] G Park, T. Rosing, M.D. Todd, C.R. Farrar, and W. Hodgkiss, “Energy harvesting
for structural health monitoring sensor networks,” Journal of Infiastructure
Systems, vol. 14, 2008, p. 64.

[125] NE. Bharatula, S. Ossevoort, M. Stager, and G Troster, “Towards wearable
autonomous microsystems,” Pervasive computing: second international

conference, Pervasive 2004, Linz/Vienna, Austria, April 18-23, 2004: proceedings,
2004,p.225.

[126] A. Hande, T. Polk, W. Walker, and D. Bhatia, “Self-Powered Wireless Sensor
Networks for Remote Patient Monitoring in Hospitals,” Sensors, vol. 6, 2006, pp.
1102—1117.

[127] NS. Shenck and J.A. Paradiso, “Energy scavenging with shoe-mounted
piezoelectrics,” IEEE MICRO, 2001, pp. 30—42.

[128] GK. Ottrnan, H.F. Hofinann, A.C. Bhatt, and GA. Lesieutre, “Adaptive
piezoelectric energy harvesting circuit for wireless remote power supply,” IEEE
Transactions on Power Electronics, vol. 17, 2002, pp. 669—676.

[129] B. Yang, C. Lee, W. Xiang, J. Xie, J .H. He, R.K. Kotlanka, S.P. Low, and H. Feng,
“Electromagnetic energy harvesting from vibrations of multiple fiequencies,”
Journal of Micromechanics and Microengineering, vol. 19, 2009, p. 035001.

217

 

[130] HA. Sodano, D.J. Inman, and G Park, “Comparison of piezoelectric energy
harvesting devices for recharging batteries,” Journal of Intelligent Material
Systems and Structures, vol. 16, 2005, p. 799.

[131] GA. Lesieutre, GK. Ottman, and HF. Hofmann, “Damping as a result of
piezoelectric energy harvesting,” Journal of Sound and Vibration, vol. 269, 2004,
pp. 991—1001.

[132] GK. Ottman, H.F. Hofmann, and GA. Lesieutre, “Optimized piezoelectric energy
harvesting circuit using step-down converter in discontinuous conduction mode,”
IEEE Transactions on Power Electronics, vol. 18, 2003, pp. 696—703.

[133] E. Lefeuvre, A. Bade], C. Richard, and D. Guyomar, “Piezoelectric energy
harvesting device optimization by synchronous electric charge extraction,” Journal
of Intelligent Material Systems and Structures, vol. 16, 2005, p. 865.

[134] M.J. Ramsay and W.W. Clark, “Piezoelectric energy harvesting for bio-MEMS
applications,” Proceedings of SPIE, 2001, p. 429.

[135] S. Kim, W.W. Clark, and Q.M. Wang, “Piezoelectric energy harvesting with a
clamped circular plate: analysis,” Journal of intelligent material systems and

structures, vol. 16, 2005, p. 847.

[136] L. Mateu and F. Moll, “System-Level Simulation of a Self-Powered Sensor with
Piezoelectric Energy Harvesting,” Sensor Technologies and Applications, 2007.
SensorComm 2007. International Conference on, 2007, pp. 399—404.

[137] K. Akkaya and M. Younis, “An energy-aware QoS routing protocol for wireless
sensor networks,” Distributed Computing Systems Workshops, 2003. Proceedings.
23rd International Conference on, 2003, pp. 710-715.

[138] B. Guo, Z. Li, and D. Li, “Energy-Aware Improved Directed Diffusion Algorithm
for Area Event Monitoring in Wireless Sensor Network,” Lecture notes in
computer science, Third international conference, ICDCIT, 2006, pp. 20-23.

[139] A. Pourkabirian and A. Haghighat, “Energy-aware, delay-constrained routing in
wireless sensor networks through genetic algorithm,” Software,
Telecommunications and Computer Networks, 2007. SoftCOM 2007. 15th
International Conference on, 2007, pp. 1-5.

[140] AB. Mohanoor, S. Radhakrishnan, and V. Sarangan, “Online energy aware routing
in wireless networks,” Ad Hoc Netw., vol. 7, 2009, pp. 918-931.

[141] AB. Mohanoor, S. Radhakrishnan, and V. Sarangan, “On energy aware routing in
wireless networks,” Broadband Communications, Networks and Systems, 200 7.
BROADNET S 2007. Fourth International Conference on, 2007, pp. 690-697.

[142] Zhi Ang Eu, Hwee-Pink Tan, and W. Seah, “Routing and Relay Node Placement in

218

 

Wireless Sensor Networks Powered by Ambient Energy Harvesting,” Wireless
Communications and Networking Conference, 2009. WCNC 2009. IEEE, 2009, pp.
1-6.

[143] B. Wenning, D. Pesch, A. Timm-Giel, and C. Gorg, “Environmental Monitoring
Aware Routing in Wireless Sensor Networks,” Wireless and Mobile Networking,
2008, pp. 5-16.

[144] M. Rahimi, H. Shah, G Sukhatme, J. Heideman, and D. Estrin, “Studying the
feasibility of energy harvesting in a mobile sensor network,” Robotics and
Automation, 2003. Proceedings. ICRA '03. IEEE International Conference on,
2003, pp. 19-24 vol.l.

[145] A. Kansal, J. Hsu, S. Zahedi, and MB. Srivastava, “Power management in energy
harvesting sensor networks,” ACM Trans. Embed. Comput. Syst., vol. 6, 2007, p.
32.

[146] A. Mahapatra, K. Anand, and DP. Agrawal, “QoS and energy aware routing for
real-time traffic in wireless sensor networks,” Computer Communications, vol.
29, Feb. 2006, pp. 437-445.

[147] H. Kwon, D. Noh, J. Kim, J. Lee, D. Lee, and H. Shin, “Low-Latency Routing for
Energy-Harvesting Sensor Networks,” Ubiquitous Intelligence and Computing,
2007, pp. 422-433.

[148] G Schaefer, F. Ingelrest, and M. Vetterli, “Potentials of Opportunistic Routing in
Energy-Constrained Wireless Sensor Networks,” E WSN '09: Proceedings of the
6th European Conference on Wireless Sensor Networks, Cork, Ireland:
Springer-Verlag, 2009, pp. 133, 118.

[149] M. Jakobsen, “Energy Harvesting Aware Routing and Scheduling in Wireless
Sensor Networks,” Master Thesis, Department of Informatics and Mathematical
Modeling, Technical University of Denmark DTU Informatics, 2008.

[150] Yongchang Yu and Gang Wei, “Energy Aware Routing Algorithm for Target
Region Monitoring in Wireless Sensor Network,” Wireless Communications,
Networking and Mobile Computing, 2007. WzCom 2007. International Conference
on, 2007, pp. 2617-2620.

[151] R. Shah and J. Rabaey, “Energy aware routing for low energy ad hoc sensor
networ ,” Wireless Communications and Networking Conference, 2002.
WCNC2002. 2002 IEEE, 2002, pp. 350-355 vol].

[152] K. Zeng, K. Ren, W. Lou, and RI. Moran, “Energy-aware geographic routing in
lossy wireless sensor networks with environmental energy supply,” Proceedings of
the 3rd international conference on Quality of service in heterogeneous
wired/wireless networks, Waterloo, Ontario, Canada: ACM, 2006, p. 8.

219

 

 

[153] Nilanjan Banerjee, Jacob Sorber, Mark D. Corner, Sami Rollins, and Deepak
Ganesan, “Triage: Balancing Energy and Quality of Service in a Microserver,”
2007.

[154] N. Banerjee, M. Corner, and B. Levine, “An Energy-Efficient Architecture for
DTN Throwboxes,” INF OCOM 2007. 26th IEEE International Conference on
Computer Communications. IEEE, 2007, pp. 776-784.

[155] Liming He, “Delay-Minimum Energy-Aware Routing Protocol (DERP) for
Wireless Sensor Networks,” Jul. 2007.

[156] C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva, “Directed
difl‘usion for wireless sensor networking,” Networking, IEEE/A CM Transactions
on, vol. 11, 2003, pp. 2-16.

220

ll
lllll
ll”
Ill”-
I"
Ill
“
m l
u
E “l
Ill
I.|..
N H
A m
nlU ”ll
H I

31293 03063 6017