Medical Embedded Systems

University of California
Los Angeles
Medical Embedded Systems
A dissertation submitted in partial satisfaction
of the requirements for the degree
Doctor of Philosophy in Computer Science
by
Roozbeh Jafari
2006
c Copyright by
Roozbeh Jafari
2006
The dissertation of Roozbeh Jafari is approved.
Benjamin M. Wu
Glenn Reinman
Milos D. Ercegovac
Majid Sarrafzadeh, Committee Chair
University of California, Los Angeles
2006
ii
To my parents, Zoya and Rahim
for teaching me to love learning
iii
Table of Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Recongurable Fabric Vest . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Driver Application . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Design of Medical Jacket . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.1 Interconnection Topology . . . . . . . . . . . . . . . . . . 10
2.3.2 Computational Units/Controllers . . . . . . . . . . . . . . 10
2.3.3 Switches . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.4 Sensing Components . . . . . . . . . . . . . . . . . . . . . 12
2.3.5 Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Ischemia Detection and Drug Delivery Method (IDDD) . . . . . . 14
2.4.1 ECG Interpretation . . . . . . . . . . . . . . . . . . . . . . 14
2.4.2 Ischemia Detection and Drug Delivery . . . . . . . . . . . 15
2.4.3 Implementation of Algorithm on Processing Units . . . . . 17
2.5 Reconguration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5.1 Network Topology in Medical Jacket . . . . . . . . . . . . 20
2.5.2 Processing, Synchronization and Reconguration in Medi-
cal Jacket . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
iv
2.6 Experimental Analysis . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3 An Ecient Placement and Routing Technique for Fault-tolerant
Wearable Computers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.2 Driver Application . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3.1 Interconnection Topology . . . . . . . . . . . . . . . . . . 37
3.3.2 Switch Model . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3.3 Circuit Switching vs. Packet Switching . . . . . . . . . . . 39
3.3.4 Power Lifetime . . . . . . . . . . . . . . . . . . . . . . . . 39
3.3.5 Power Consumption Model . . . . . . . . . . . . . . . . . . 40
3.3.6 Fault Lifetime . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.7 Fault Model . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.8 Integer Linear Programming Formulation . . . . . . . . . . 48
3.3.9 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.4 Problem Denition . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.5 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5.1 Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5.2 Placement . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
v
3.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 Adaptive Electrocardiogram Feature Extraction on Distributed
Embedded Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3 Automated Feature Set Detection . . . . . . . . . . . . . . . . . . 74
4.4 Software Proling . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Target Architecture Model . . . . . . . . . . . . . . . . . . . . . . 80
4.6 Dynamic Reconguration . . . . . . . . . . . . . . . . . . . . . . . 81
4.7 Feature Set Partitioning . . . . . . . . . . . . . . . . . . . . . . . 82
4.8 Simulation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5 Customizable Medical Monitoring Platform . . . . . . . . . . . . 97
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.4 Design of CustoMed . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.4.1 Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.4.2 Med Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.4.3 Pocket PC . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.4.4 Server . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
vi
5.5 Driver Application . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.6 Sensor Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6 CMAS: Clinical Movement Assessment System for Neuromotor
Disorders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.2 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3 Hardware design . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.3.1 Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.3.2 Handgrip device . . . . . . . . . . . . . . . . . . . . . . . . 121
6.3.3 Med node . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.3.4 Pocket PC . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.4 Software design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.4.1 Med node . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.4.2 Pocket PC . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.5 Test sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.6 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . 127
7 On Minimal Energy Skew Routing in Lossy Wireless Sensor Net-
works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
vii
7.3 Preliminaries and Models . . . . . . . . . . . . . . . . . . . . . . . 133
7.4 Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.5 Minimal-Skew Routing . . . . . . . . . . . . . . . . . . . . . . . . 136
7.5.1 Problem Formulation and -Approximation Algorithm . . . 136
7.5.2 Discussion on Lossy Channels . . . . . . . . . . . . . . . . 145
7.5.3 Dissimilar Initial Energy Levels . . . . . . . . . . . . . . . 146
7.5.4 Time Complexity Analysis . . . . . . . . . . . . . . . . . . 147
7.5.5 Discussion on Multicommodity . . . . . . . . . . . . . . . 148
7.6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
viii
List of Figures
2.1 System Architecture of the Medical Jacket . . . . . . . . . . . . . 6
2.2 Fabric with Mesh of Wires . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Switch Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 A Dot-Mote along with a Switch . . . . . . . . . . . . . . . . . . 12
2.5 Electrode for Transdermal Iontophoresis System . . . . . . . . . . 14
2.6 A typical ECG Signal . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 Normal and Abnormal ECG Signals . . . . . . . . . . . . . . . . . 15
2.8 Ischemia Detection Operations . . . . . . . . . . . . . . . . . . . . 17
2.9 Network Topology . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.10 Task Multiplexing in Medical Jacket Network . . . . . . . . . . . 29
2.11 Medical Jacket Prototype . . . . . . . . . . . . . . . . . . . . . . 30
2.12 Power Consumption of Medical Jacket . . . . . . . . . . . . . . . 30
3.1 Interconnection topology . . . . . . . . . . . . . . . . . . . . . . . 38
3.2 Tear model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Representation of tears in dual graph . . . . . . . . . . . . . . . . 44
3.4 Region subdivisions for every edge in dual graph . . . . . . . . . . 45
3.5 Probability of edge failure in the grid for all possible distinct tears 48
3.6 ILP variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.7 Fault/power lifetime ratio for 7x7 mesh size . . . . . . . . . . . . 62
3.8 Lifetime analysis (
f
= 1/24 h
1
) . . . . . . . . . . . . . . . . . . 63
f
= 1/30 h
1
) . . . . . . . . . . . . . . . . . . 64
ix
f
= 1/36 h
1
) . . . . . . . . . . . . . . . . . . 65
3.11 Lifetime analysis on xed mesh size varying communication distance(
f
=
1/24 h
1
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
f
=
1/30 h
1
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
f
=
1/36 h
1
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.14 Performance comparison of our technique with a greedy method
(
f
= 1/24 h
1
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.15 Performance comparison of our technique with a greedy method
(
f
= 1/36 h
1
) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.1 ECG analysis schematic . . . . . . . . . . . . . . . . . . . . . . . 88
4.2 Automatic ECG segmentation performed on ltered signal . . . . 89
4.3 The sampling intervals of ECG morphologies 1 and 2. Morphol-
ogy 1 consists of samples extracted from QRS onset and oset.
Morphology 2 consists of samples from S-wave oset and T-wave
oset. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 The sampling intervals of ECG morphologies 3 and 4. Morphology
3 consists of samples between 50 ms before and after the ducial
point (FP). Morphology 4 consists of samples between 150 ms and
500ms after the ducial point. . . . . . . . . . . . . . . . . . . . . 90
4.5 Graph generated from proling analysis containing only major
blocks required for feature extraction. . . . . . . . . . . . . . . . . 91
x
4.6 Delays corresponding to ECG feature detection modules extracted
in proling phase . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.1 System Components . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.2 CustoMed GUI on a Personal Computer . . . . . . . . . . . . . . 101
5.3 Application Development Cycle . . . . . . . . . . . . . . . . . . . 102
5.4 A Med Node along with Sensors . . . . . . . . . . . . . . . . . . . 108
5.5 Data Monitoring on a Remote Server . . . . . . . . . . . . . . . . 109
5.6 System Hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.7 Med Nodes Placement on Body . . . . . . . . . . . . . . . . . . . 112
5.8 CustoMed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.9 Base Station Med Node and a Pocket PC . . . . . . . . . . . . . . 117
5.10 Sensor Readings during Indoor Slow Walk . . . . . . . . . . . . . 118
6.1 Handgrip device . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.2 Connection setup window . . . . . . . . . . . . . . . . . . . . . . 125
6.3 Calibration window . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.4 CMAS task window . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.5 Average and individual forces during Task 1 . . . . . . . . . . . . 128
7.1 Minimum Skew Denition . . . . . . . . . . . . . . . . . . . . . . 135
7.2 Node partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.3 Original network and the resulting network after partitioning . . . 139
7.4 Flow exchange to obtain minimal-skew solution . . . . . . . . . . 143
7.5 Edge splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
xi
7.6 Graph topology (n=100, grid distribution) . . . . . . . . . . . . . 150
7.7 Normalized energy consumption rate in sorted order (n=100, grid
distribution) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
7.8 Graph topology (n=100, random distribution) . . . . . . . . . . . 152
7.9 Normalized energy consumption rate in sorted order (n=100, ran-
dom distribution) . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
7.10 min-skew/min-cost lifetime for various number of splits - k (n=100,
grid distribution) . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
grid distribution) . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
rand distribution) . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
rand distribution) . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
7.14 Average delay of min-skew/min-cost for various number of splits -
k (n=100, grid distribution) . . . . . . . . . . . . . . . . . . . . . 156
k (n=100, grid distribution) . . . . . . . . . . . . . . . . . . . . . 156
k (n=100, rand distribution) . . . . . . . . . . . . . . . . . . . . . 157
k (n=100, rand distribution) . . . . . . . . . . . . . . . . . . . . . 157
xii
List of Tables
2.1 Fault Detection/Reconguration Time . . . . . . . . . . . . . . . 26
2.2 Power Consupmtion/Reliability of the Medical Jacket . . . . . . . 27
2.3 Computational Capability of Dot-Motes . . . . . . . . . . . . . . 27
2.4 Number of Unknown Beat Queries Sent to Pokcet PC . . . . . . . 31
3.1 Power/fault lifetime comparison in 5x5 grid . . . . . . . . . . . . 61
3.2 Power/fault lifetime comparison in 7x7 grid . . . . . . . . . . . . 61
4.1 Features categorized by groups . . . . . . . . . . . . . . . . . . . . 77
4.2 Benchmark statistics . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.3 Number of queries exchanged among processing units : sampling
rate = 360 sample/sec . . . . . . . . . . . . . . . . . . . . . . . . 94
rate = 200 sample/sec . . . . . . . . . . . . . . . . . . . . . . . . 95
rate = 100 sample/sec . . . . . . . . . . . . . . . . . . . . . . . . 96
5.1 CustoMed System Specication . . . . . . . . . . . . . . . . . . . 112
5.2 Wireless Channel Characteristics during Indoor Slow Walk . . . . 114
5.3 Wireless Channel Characteristics during Outdoor Normal Walk . 114
6.1 Itemized force error and frequency dierence . . . . . . . . . . . . 129
xiii
Acknowledgments
First and foremost, I would like to thank my advisor, Professor Majid Sar-
rafzadeh. Without his invaluable support and guidance, my thesis work would
not have been possible. I am very grateful for his patience, motivation, enthu-
siasm, and immense knowledge that, taken together, make him a phenomenal
mentor.
I would like to thank other great mentors I had the opportunity to work with:
Professors V. Reggie Edgerton, Miodrag Potkonjak and Devin L. Jindrich. I
would also like to extend my appreciation to my thesis committee members, Pro-
fessors Milos Ercegovac, Glenn Reinman and Benjamin Wu.
I would like to thank present and past members of Embedded and Recongurable
lab for their friendship and constructive discussions we had. I had the opportu-
nity to supervise and guide many undergraduate and graduate students. I wish
to thank them for giving me the chance to discover my passion for teaching and
academic research. I must thank the ones who played a decisive role: Hyduke
Noshadi, Azad Zahoory and Andre Encarnacao.
I would like to thank all my wonderful friends in Los Angeles for the great and
joyous moments we shared together. It would be a long list to mention all of the
friends that I am indebted to. I gratefully thank all of them.
I wish to thank my relatives in the United States: my aunt Rafat, my uncles
Karim, Farhad and Abbas, and my cousins Sepehr, Mehrdad, Maryam, Farnaz
and Ali for their support and love.
Most of all, my deepest gratitude goes to my family, my Mom Zoya, my Dad
Rahim, my sister Ramak and my brother Ramin for the unconditional love and
support that they have given me. One of my most wonderful childhood memories
is when my mom used to read books to me and my sister with such great patience.
xiv
I will never forget the moment my Dad brought in my rst computer. Ramak
and Ramin are the most amazing siblings one could ever have. Last but certainly
not least a very special thanks to my Grandma for her love and encouragement.
Thank you!
xv
Vita
1977 Born, Tehran, IRAN.
2000 B.S. (Electrical Engineering), Sharif University of Technology,
Tehran, IRAN.
2002 M.S. (Electrical Engineering), State University of New York at
Bualo, NY.
2004 M.S. (Computer Science), University of California, Los Angeles,
CA.
Publications
Optimal Register Sharing for CFG Synthesis in SSA Form Philip Brisk, Foad
Dabiri, Roozbeh Jafari, Majid Sarrafzadeh IEEE Transactions on CAD (TCAD),
vol 25 no. 5, pp 772-779, May 2006.
Light-weight Embedded Systems Foad Dabiri, Roozbeh Jafari, Ani Nahapetan,
Majid Sarrafzadeh In Computer Engineering Handbook, Edited by Vojin Oklob-
dzija, Taylor & Francis/CRC Press, to appear.
A Unied Theory of Timing Budget Management Soheil Ghiasi, Elaheh Bo-
zorgzadeh, Po-kuan Huang, Roozbeh Jafari, Majid Sarrafzadeh accepted for pub-
lication in IEEE Transactions on Computer Aided Design (TCAD), to appear.
xvi
Probabilistic Delay Budgeting for Soft Realtime Applications Soheil Ghiasi, Po-
kuan Huang, Roozbeh Jafari accepted for publication in IEEE Transactions on
VLSI (TVLSI), to appear.
Recongurable Fabric Vest for Fatal Heart Disease Prevention Roozbeh Jafari,
Foad Dabiri, Philip Brisk, Majid Sarrafzadeh accepted for publication in Journal
of Embedded Computing (JEC), to appear.
-Optimal Minimal-Skew Battery Lifetime Routing in Distributed Embedded
Systems Roozbeh Jafari, Foad Dabiri, Majid Sarrafzadeh Journal of Low Power
Electronics (JOLPE), vol 1, no. 2, pp 97-107, September 2005.
Micro-Sequencer Approach Speeds Reconguration Roozbeh Jafari, Henry Fan,
Majid Sarrafzadeh Computers O-The Shelf (COTS) Journal, vol 5, no. 6, pp
49-55, June 2003.
An Ecient Placement and Routing Technique for Fault-tolerant Distributed
Embedded Computing Roozbeh Jafari, Majid Sarrafzadeh accepted for publica-
tion in ACM Transactions on Embedded Computing Systems (TECS), to appear.
Adaptive Electrocardiogram Feature Extraction on Distributed Embedded Sys-
tems Roozbeh Jafari, Hyduke Noshadi, Soheil Ghiasi, Majid Sarrafzadeh IEEE
Transactions on Parallel and Distributed Systems special issue on High Perfor-
mance Computational Biology (TPDS), vol. 17, no. 8, pp 1-11, August 2006.
xvii
Abstract of the Dissertation
Medical Embedded Systems
by
Roozbeh Jafari
Doctor of Philosophy in Computer Science
University of California, Los Angeles, 2006
Professor Majid Sarrafzadeh, Chair
Light-weight embedded systems are now gaining more popularity due to recent
technological advances in fabrication of more powerful tiny processors. In par-
ticular, the eld of pervasive computing and medical monitoring is emerging due
to the development of tiny and low-prole computing nodes, mated with sensors
and actuators, in close proximity to the human. In this thesis, I will demonstrate
some of my Ph.D. work that has been focused on design and analysis of hardware
and software for medical embedded systems. In particular, I will discuss two of
our platforms: Recongurable Fabric Vest and CustoMed (Customizable Medical
Monitoring System). I will further describe some of the fundamental problems I
have been investigating:
An Ecient Placement and Routing Technique for Fault-tolerant Wearable
Computers
Adaptive Electrocardiogram Feature Extraction on Distributed Embedded
Systems
-Optimal Minimal-Skew Battery Lifetime Routing in Distributed Embed-
ded Systems
xviii
CHAPTER 1
Introduction
In recent years, exciting technological advances have been made in development
of exible electronics. These technologies oer the opportunity to weave com-
putation, communication and storage into the fabric of the every clothing that
we wear, therefore, creating intelligent fabric. Chapter 2 presents a medical vest
which has sensors for physiological readings and software-controlled, electrically-
actuated trans-dermal drug delivery elements. Furthermore, computational el-
ements are embedded in the vest for collecting data from sensors, processing
them and driving actuation elements. Since this vest will be used for medical,
life-critical applications, the single most critical requirement of such a vest is an
extremely high level of robustness and fault tolerance. Meantime, the key tech-
nological constraint for these mobile systems is their power consumption. Our
target application for our medical vest is the detection of possibly fatal heart
problems, specically unstable angina pectoris or ischemia. We illustrate the de-
sign stages of our medical vest as well as the technical details of both software
and network reconguration schemes (to enhance the robustness and the perfor-
mance of our system). We also discuss the details of ischemia detection algorithm
employed in our vest. Moreover, we evaluate the robustness of our system with
existence of various faults. Finally we measure the performance of our algorithm
as well the power consumption of several congurations of our vest.
Chapter 3 presents an ecient technique for placement and routing of sen-
1
sors/actuators and processing units in a grid network. The driver application
that we present is a medical jacket which requires an extremely high level of ro-
bustness and fault tolerance. The power consumption of such jacket is another
key technological constraint. Our proposed interconnection network is a mesh
of wires. A jacket made of fabric and wires would be susceptible to accidental
damage via tears. By modeling the tears, we evaluate the probability of having
failures on every segment of wires in our mesh interconnection network. Then
we study two problems of placement and routing in the sensor networks such
that the fault tolerance is maximized while the power consumption is minimized.
We develop ecient integer linear programming (ILP) formulations to address
these problems and perform both placement and routing simultaneously. This
ensures that the solution is a lower bound for both problems. We evaluate the
eectiveness of our proposed techniques on a variety of benchmarks.
Tiny embedded systems have not been an ideal outt for high performance
computing due to their constrained resources. Limitations in processing power,
battery life, communication bandwidth and memory constrain the applicability
of existing complex medical analysis algorithms such as the Electrocardiogram
(ECG) analysis. Among various limitations, battery lifetime has been a ma-
jor key technological constraint. In Chapter 4, the issue of partitioning such a
complex algorithm while the energy consumption due to wireless transmission
is minimized is addressed. ECG analysis algorithms normally consists of pre-
processing, pattern recognition and classication. Considering the orientation
of the ECG leads, we devise a technique to perform preprocessing and pattern
recognition locally in small embedded systems attached to the leads. The fea-
tures detected in pattern recognition phase are considered for the classication.
Ideally, if the features detected for each heartbeat reside in a single processing
node, the transmission will be unnecessary. Otherwise, to perform classication,
2
the features must be gathered on a local node and thus, the communication is
inevitable. We perform such a feature grouping by modeling the problem as a
hypergraph and applying partitioning schemes which yields a signicant power
saving in wireless communications. Furthermore, we utilize dynamic recongura-
tion by software module migration. This technique with respect to partitioning
enhances the overall power saving in such systems. Moreover, it adaptively alters
the system conguration in various environments and on dierent patients. We
evaluate the eectiveness of our proposed techniques on MIT/BIH benchmarks
and on average achieve 70% energy saving.
Ubiquitous computing denes a new era in which a large number of hetero-
geneous systems collaborate to support the execution of medical applications
and health care monitoring. In Chapter 5, we propose a system for rapidly con-
structing low-cost prototypes for medical monitoring by leveraging wireless sensor
nodes. Our platform is a new architecture called CustoMed that will reduce the
customization and reconguration time for medical systems. This architecture is
a network enabled system that supports various wearable sensors. Moreover, it
contains on-board general computing capabilities for executing individually tai-
lored event detection, alerts, and network communication with various medical
informatics services. The customization of such system with a large number of
med nodes is extremely fast in a few minutes even by non-engineering sta. In
this paper, we present the design stages of such system along with experimental
analysis that evaluates its performance.
We are developing a clinical movement assessment system (CMAS) to quantify
motor eort following neuromotor injury. The system will allow for easy and
versatile data collection by non-technical sta and patients from a variety of
devices by using recent advances in technology for real-time reconguration. We
3
will demonstrate the eectiveness of our system by developing a patient-friendly
device in Chapter 6. A prototype of the device has been implemented and used
to measure nger forces in exion and extension. The eciency of the CMAS
devices will be demonstrated by making quantitative assessments of neuromotor
eorts following stroke in a patient population. At present, our experimental
results on two subjects show the eectiveness of our device.
For several sensor networks applications, it is critical to extend the lifetime of
each individual sensor node in order to remain operational for the longest time
possible. Therefore, the power consumption rate should be evenly distributed
over all the nodes in the system. Traditional routing algorithms attempt to min-
imize the total power consumption of the system, but they do not attempt to
evenly distribute the load over all the nodes in the network. In Chapter 7, we
present an ecient routing algorithms that minimizes the energy skew among
nodes in a network with lossy links. We propose an -optimal polynomial time
centralized multi-hop routing technique that maximizes the lifetime a system
of distributed power sources, considering the quality of the wireless links and
the vagaries of the radio communication channel. Our technique aims to evenly
distribute the power consumption rate which yields in a minimal-skew solution.
We theoretically prove that our technique is ecient. Finally, we illustrate the
quality of the solutions provided by our algorithms on a set of benchmarks that
consider the quality of the wireless channels based on models using real RF tran-
scievers. We show that our solution provides signicant increase in the lifetime
of the network (up to ve times) at the cost of a slight increase in the latency of
the end-to-end paths (up to 10%).
4
CHAPTER 2
Recongurable Fabric Vest
2.1 Introduction
2.1.1 Motivation
Computation, storage, and communication are now woven into the fabrics of
our society with much of the progress being due to the relentless march of the
silicon-based electronics technology as predicted by Moores Law. The emerging
technology of exible electronics, where electronics components such as transis-
tors and wires are built on a thin exible material, oers a similar opportunity
to weave computation, storage, and communication into the fabric of the very
clothing that we wear. The implications of seamlessly integrating a large number
of communicating computation and storage resources, mated with sensors and
actuators, in close proximity to the human body are quite exciting. For example,
one can imagine biomedical applications where biometric and ambient sensors
are woven into the garment of a patient or a person in a hazardous environment
to trigger or modulate the delivery of a drug. Realizing this vision is not just
a matter of developing innovative materials for exible electronics, along with
accompanying sensors and actuators. The characteristics of the exible electron-
ics technology and the novel applications enabled by it require innovation at the
system-level technology level. The natural applications of these systems have
5
environmental dynamics, physical coupling, resource constraints, infrastructure
support, and robustness requirements that are distinct from those faced by tra-
ditional systems. This combination requires one to go beyond thinking of these
systems as traditional systems in a dierent exible form factor. Instead, a re-
thinking of the architecture and the design methodology for all layers of these
systems is required.
2.1.2 Driver Application
Figure 2.1: System Architecture of the Medical Jacket
Although the goal of the our research is to investigate system architecture
and design methodology concepts for electronic textiles, successful systems re-
6
search of this nature requires that one not only develop these concepts, but also
validate them in context of real technology constraints and application require-
ments. Therefore, we use pervasive patient monitoring and sensor-driven person-
alized trans-dermal drug delivery as our driver application. One possibility of
leveraging electronic textile technology in the context of such an application is to
create a exible garment (i.e., vest) that the patient can wear, which has sensing,
computation, communication, and actuation elements embedded in it. We are
developing such a prototype vest called medical jacket. Ideally, such a person-
alized drug delivery vest should have sensors on both the interior (to measure
physiological readings) and on the exterior (to measure environmental readings
such as the presence of toxins in the surroundings), and a software-controlled,
electrically-actuated trans-dermal drug delivery system. It should allow low-
latency, ne-grained adaptation of the drug dosage based on continual physiolog-
ical measurements in the case of patients, and based on both environmental and
physiological measurements in the case of people operating in hazardous envi-
ronments. More generally, our application driver is representative of biomedical
applications where information technology is integrated into fabrics and textiles.
Figure 2.1 shows the overall system architecture of the medical jacket. As is
evident from the Figure 2.1, the vest consists of four main subsystems: control
(or computation), communication, sensing, and actuation.
One of target applications of our medical jacket is the detection of possibly fatal
heart problems, specically unstable angina pectoris. Angina Pectoris is a fatal
medical condition with more than 7 million suerers in the U.S. alone. In the
unstable angina pectoris form of the condition, fatal attacks happen in an unpre-
dictable manner, even when the patient is at rest. Using the architecture that
has been discussed in the previous sections, we can achieve constant monitoring
and life-saving drug delivery in the emergency situation even when the patient
7
is away from any conventional emergency medical help. With the medical jacket
the patient can have a personalized tuning for the system by his/her physician
and can continue with the normal life activities with the vest. Drug delivery
algorithms running on the vest, nely tuned by the patients physician, can be
used prevent any fatal results.
Angina pectoris is characterized as an acute chest pain or discomfort due to
coronary heart disease and is considered as a symptom of myocardial ischemia.
Electrocardiogram analysis is the standard used for the diagnosis of ischemia.
However, ECG waves are highly patient dependent and the analysis is generally
performed by the clinician manually (through observation of the ECG waves).
2.2 Related Work
Several wearable technologies exist to continually monitor patients vital signs,
utilizing low cost, well-established disposable sensors such as blood oxygen nger
clips and electrocardiogram electrodes.
The Smart Shirt from Sensatex [sen] is a wearable health monitoring device that
integrates a number of sensory devices onto the Wearable Motherboard from
Georgia Tech [PMJ02]. The Wearable Motherboard is woven into an undershirt
in the Smart Shirt design. Their interconnect is a exible data bus that can sup-
port a wide array of sensory devices. These sensors can then communicate via the
data bus to a monitoring device located at the base of the shirt. The monitoring
device is integrated into a single processing unit that also contains a transceiver.
The SmartShirt design features plastic optical ber that can be used to detect
punctures - however, they do not have any means of dealing with these punctures
other than reporting them via the transceiver. In contrast, our design features a
8
recongurable interconnect - rather than a large data bus - which can dynami-
cally adjust to punctures or tears. Moreover, we provide further fault tolerance
through distributed control buttons. In the event that their single processing
unit (and transceiver) is damaged or the control lines leading to the device are
torn, the Smart Shirt is virtually inoperable. In our design, we distribute control
to multiple processing elements and can accommodate multiple communication
buttons. This ensures that a tear or puncture to our vest will not result in total
system failure. This fault tolerance is essential in the demanding and/or haz-
ardous environments targeted by this research. Fireghters, policemen, soldiers,
astronauts, athletes, and others working in hazardous environments need robust
material that can sustain damage and yet still reliably provide service. Finally,
our design supports actuation, such as a drug delivery system that could pro-
vide immediate medical attention to individuals in environments that would be
dicult to bring medical personnel - either due to their remoteness (i.e. in the
case of an astronaut or mountain climber) or the danger involved (i.e. in the case
of a potential biohazard or re). By automating the treatment, as well as the
detection, our envisioned capability will provide more fault tolerant means of safe-
guarding the life of the wearer of garments such as the medical jacket described
earlier. Several other technologies have been introduced by MIT called MIThril
[DMP03], e-Textile from Carnegie Mellon University [MMK02] and Wearable
e-Textile from Virginia Tech [MJE03]. None supports the concept of recongu-
ration due to faults or tears as the medical jacket does.
9
2.3 Design of Medical Jacket
2.3.1 Interconnection Topology
The interconnection medium for our proposed system is a mesh of wire segments.
The mesh interconnection topology is a wire-frame that has a regular structure,
each vertex being connected to exactly four other vertices.
Mesh networks have several signicant advantages. Each node has a dedicated
communication link with every other node on the network and also has access
to the full bandwidth available for that link. Nodes on buses must share the
bandwidth available on the bus medium. Besides, in a mesh, multiple paths exist
between devices. This brings a great robustness against faults. If a direct path
between two nodes goes down, messages can be rerouted through other paths.
Moreover, it has considerable scalability and can be easily manufactured. The
manufacturing issues become more signicant because in our system wires are
integrated into the fabric and with current fabric manufacturing technology, this
can be easily achieved. Furthermore, the mesh interconnection is highly regular
which assist us in routing and placement of sensors and computational units. A
picture of our fabric with mesh interconnection is shown in Figure 2.2.
2.3.2 Computational Units/Controllers
The dot motes developed at the University of California, Berkeley oer a tiny,
low cost computation platform for embedded applications. It comprises a pro-
grammable ATmega128 microcontroller from Atmel. The ATmega128 [atm] is
a low-power CMOS 8-bit RISC microcontroller. The ATmega128 achieves a
throughput of up to 1 MIPS per MHz. The ATmega128 also provides 128K
10
Figure 2.2: Fabric with Mesh of Wires
bytes of In-System Programmable Flash, 4K bytes of EEPROM, 4K bytes of
SRAM, and several peripherals including a real time counter, four timers, two
UARTs (Universal Asynchronous Receiver and Transmitter), an ADC, and a byte
oriented two-wire serial interface also called the Inter Integrated Circuit (I
2
C)
interface.
2.3.3 Switches
The dot motes are connected to UART lines through programmable switches to
create the distributed computation fabric drawn in Figure 2.3. The switches used
for this purpose are ADG714 ICs from Analog Devices. These are CMOS, octal
SPST (single-pole, single-throw) switches controlled via a 3-wire serial interface.
All three terminals on each side is connected to a horizontal or a vertical wire
segment. Six type of connection may be established with this switch. Two con-
11
Figure 2.3: Switch Model
nections are straight connections whereas the other are bent connections.
Figure 2.4: A Dot-Mote along with a Switch
2.3.4 Sensing Components
Myocardial Ischemia is basically caused by lack of oxygen and nutrients to the
contractile cells. Frequently it may lead to myocardial infraction. This causes
severe consequence of heart failures or arrhythmia that may yield death. An elec-
trocardiogram (ECG / EKG) is an electrical recording device of the heart signals
and is used in the investigation of heart diseases including Ischemia. The ECG
monitoring system that we use utilizes ten electrodes. The electrodes (or leads)
12
are attached to the patients arms, legs, and chest. The electrodes detect the elec-
trical impulses generated by the heart, and transmit them to the ECG machine.
From the ECG tracing, the heart rate, the heart rhythm, whether there has been
a prior heart attack, whether there may be coronary artery disease and whether
the heart muscle has become abnormally thickened can be determined. Digitized
samples are sent to the dot-motes for Ischemia detection process through UART
protocol. The ECG that we utilize is manufactured by Midmark Diagnostics
Group [mid].
2.3.5 Actuators
Nitroglycerin is used to prevent Ischemia (Angina). It works by relaxing the
blood vessels to the heart, so the blood ow and oxygen supply to the heart
is increased. Various routes of administration are proposed such as sublingual
tablets, extended-release capsule, skin patches (transdermal), spray and oint-
ment. A rich area of research over the past 10 to 15 years has been focused on
developing transdermal technologies that utilize mechanical energy to increase the
drug ux across the skin by either altering the skin barrier (primarily the stratum
corneum) or increasing the energy of the drug molecules. These so-called active
transdermal technologies include iontophoresis (which uses low voltage electrical
current to drive charged drugs through the skin). Ionotophosis is dened as the
introduction, by means of a direct electrical current, of ions of soluble salts into
the tissues of the body for therapeutic purposes [SM94]. It is a technique used
to enhance the absorption of drugs across biological tissues, such as the skin.
Transdermal iontophoresis is an excellent candidate for our application since it
is controlled by electrical signals. Such a drug delivery unit is shown in Figure 2.5.
13
Figure 2.5: Electrode for Transdermal Iontophoresis System
Nitroglycerin may lose its eectiveness over time, so physicians generally
schedule nitrate-free breaks to prevent tolerance. A valid concern exists that
nitrate-free periods might increase the risk for angina and adverse heart events.
With on-line monitoring performed in our medical jacket, we avoid excessive
amount of drug administration and eliminate its risks.
2.4 Ischemia Detection and Drug Delivery Method (IDDD)
2.4.1 ECG Interpretation
As the heart undergoes depolarization and repolarization, the electrical currents
that are generated spread not only within the heart, but also throughout the
body. This electrical activity generated by the heart can be measured by an ar-
ray of ECG electrodes placed on the body surface. A typical ECG signal is shown
in Figure 2.6. P, QRS and T Segments which are analyzed in the algorithm are
shown in the Figure.
14
Figure 2.6: A typical ECG Signal
2.4.2 Ischemia Detection and Drug Delivery
Figure 2.7: Normal and Abnormal ECG Signals
Our ischemia detection algorithm incorporates a collection of initial tuning
sessions to extract the normal heart pattern of the patient. These initial ECGs
are selected to include a range that span dierent activity levels and various days
of the initial tuning period (rest, exercise, stress...etc). For diagnosis, common
ECG patterns of the heart signal are used such as: heart rate, PR interval, QRS
duration, QT interval duration, ST interval ...etc. In addition, a collection of ab-
normal ECG patterns associated with the ischemic conditions are stored. Normal
15
ECG patterns of the patient along with the standard indications and abnormal
patterns of ischemia are incorporated as a basis for the rest of the automated
analysis that is performed by the recongurable vest as shown in Figure 2.7. At
the end of the initial tuning period, the drug delivery decision is based upon tun-
ing information provided by the patients cardiologist. Tuning of the algorithm
will determine the criticality of the wave according to the deviation from the
normal values of the patient and the matching with abnormal ischemic patterns
according to the guideline provided by the clinician through tuning.
Nitroglycerin delivery through a patch is used as an initial response to heart at-
tack. One of the main reasons that nitroglycerin is used in this system is the fact
that it does not generate any risk to the patient if the condition is not actually
a heart attack.
For the initial version of the algorithm we focused on ST region analysis as one
of the main indications of ischemia. Ischemia has common indications such as:
lack or inversion of T region in the ECG signal, deviation in the elevation and
slope of the ST interval.
In the ischemic ECG, the ST region of the curve has values abnormally lower
or non-existing compared to patients average heart signal shown in Figure 2.7.
Similarly an abnormal heart signal detected with unusual QRS complex with ex-
tremely low 1mV R voltage on the specic probe of the ECG as shown in Figure
2.7.
Our ischemia detection algorithm is capable of detecting ischemic heart attacks
that show ST region indications such as: non-existing ST region, elevation, slope
abnormalities according to the preset normal and abnormal ranges. Furthermore,
the algorithm is capable of detecting other heart signal abnormalities such as QRS
complex width, abnormal interval ranges and voltages for P,R,S,T,Q waves, etc.
As a crucial part of its nature the algorithm is fast and runs on real-time data.
16
2.4.3 Implementation of Algorithm on Processing Units
Figure 2.8: Ischemia Detection Operations
As shown in Figure 2.8, at the rst step of the algorithm, signals from ECG
are passed through a bandpass lter to reduce the noise and remove the bias volt-
age. The lter is an FIR window band-pass lter with cuto frequencies adjusted
at 6Hz and 400Hz. Rest of the algorithm can be partitioned into two sections,
ECG signal characterization and ischemia pattern matching.
ECG Signal Characterization:
After the signal is ltered, QRS complex is detected and consequently heart rate
is observed. Each time a local peak is detected, it is classied as QRS complex,
T-Wave or noise. So, at the end of each period new values of QRS complex, ST
interval and T-Wave amplitude are updated. Basic rules to nd a QRS and ST
17
segments are as follows:
Peaks that precede or follow larger peaks by less than 200 ms (50 samples)
are ignored.
If a peak has occurred after 360 ms (90 samples) of the previous peak and
the maximum derivative of the signal is at least half of the previous peak
(i.e. QRS complex), identify the peak as T-Wave.
If the peak is larger than the threshold, call it QRS complex.
Update the threshold as the mean of the last 8 QRS peaks detected multi-
plied by constant factor TH (TH = 0.9). A good estimate on QT duration
can be achieved from the following equations [PMV01]:
QT =
0.384RR + 99 RR < 600ms

0.156RR + 236 600ms RR 1000ms
0.166RR + 277 RR > 1000ms
RR is the heart beat period. At the beginning, the threshold is estimated to
be the mean of the two maximum peaks in a 2 second interval.
Ischemia Pattern matching:
When a beat is detected, it is characterized by a number of features such as
width, amplitude, and R-to-R interval. The beat is also compared to previously
detected beat templates. If the beat closely matches a previously classied beat
template, its information is retrieved from its matching type. Otherwise, the beat
is classied based on its beat features as described below:
18
Ischemia may change ECG signals in dierent ways and mostly aect ST-T com-
plex. In our algorithm, four major patterns in ECG are considered as ischemic
patterns [GLS98]:
T-Wave Enlargement: If the T-Wave Amplitude is increased by more than
1mV.
ST Level increase: If ST segment deviation is greater than 1mV.
QRS end change: An upper shift in the QRS end (S-wave) may be a sign
of ischemia.
QRST deformation: Myocardial ischemia may cause a deformation in ECG
signals which results in increase of the integral of the waveform from S-
wave to T-wave. At each period this integral is evaluated and compared to
a threshold.
If a beat matches neither of the above criterion nor any of the beat tem-
plates already stored in the memory, it is considered as an UNKNOWN beat.
In this case, the mote inquire about the type of the unknown beat from other
processing units in the network and once the information is retrieved, the mote
replaces the least recently accessed beat template with the new beat in its mem-
ory. This process increases the communication overhead of the system between
various processing units and hence increases the power consumption. However,
it is inevitable since the accuracy of the algorithm is extremely important and
we may not tolerate to miss unknown beats.
Once an ischemic behavior is detected in one period of ECG signals, we keep
track of ischemic patterns for at least 30 seconds [ZR03]. If we detect a con-
19
tinuous 30 second ischemic behavior, this would be considered as occurrence of
ischemia and drug administration will start.
There exists various tools on PhysioToolkit web-site [phyb] for bio-signal pro-
cessing, however, our processing units (dot-motes) with their very constrained
resources as well as the special communication requirements of electronic devices
in our medical vest forced us to reproduce the algorithm and implement the soft-
ware for our embedded system.
2.5 Reconguration
2.5.1 Network Topology in Medical Jacket
As discussed earlier, the interconnection medium of the medical jacket is a mesh
of wire segments. We employ four dot-motes as computational units placed at the
corners of a square in our interconnection network as shown in Figure 2.9. Every
dot-mote is accompanied with a switch which is described in Section 2.3.3. In
this prototype, we use a mesh of size 4x4. The center square of wire segments are
associated with inter-mote communication and will be referred as inner square
in this chapter. The outer square of wire segments (mesh of size 4x4) is respon-
sible for carrying ECG and drug delivery signals. The switches, controlled by
dot-motes, connect ECG, drug delivery or inter-mote communication lines to the
motes. There exist two pairs of wires in inner square to improve the reliability
of inter-mote communication. In case if a pair of wires fails, switches can swap
to the other pair and recongure the circuit.
20
Figure 2.9: Network Topology
2.5.2 Processing, Synchronization and Reconguration in Medical
Jacket
The ischemia detection algorithm is previously elaborated in Section 2.4.3. The
dot-motes perform the processing, however, since we have eight channels of ECG,
the computational capability of one dot-mote does not suce to process the data
from all ECG channels. Therefore, the algorithm is distributed over four dote-
motes. Each mote accomplishes interpretation of ECG signals of two channels.
The synchronization ensures that every ECG channel is being processed by at
least a dot-mote. Therefore, synchronization among dot-motes is inevitable. Fur-
thermore, in case of fault occurrence, the network is repaired in reconguration
21
stage. Once ischemia is detected, appropriate signals are sent to the drug delivery
units. As for drug delivery units, due to some limitations, we do not utilize actual
drug delivery units. Instead we exploit a Pocket PC for monitoring purposes. We
emphasis that the pocket PC does not carry out any computations for ischemia
detection algorithm.
The only communication protocol that is available on dot-motes is UART (I
2
C
still is not fully functional on mica2dots), hence all processing, synchronization
and reconguration tasks are carried out through the UART protocol. Therefore,
tasks are executed sequentially as shown in Figure 2.10.
According to [ZR03], when ischemia happens, the ECG signals stay deformed
for at least 30 seconds. Physicians recommend that the appropriate drug is
administered in about one minute once ischemia occurs. In the worst case, if
a fault occurs at the same time when ischemia occurs, the network may require
reconguration in order to accomplish successful drug administration. If ischemia
is detected suddenly after synchronization/reconguration phase, the circuit will
be recongured in the next synchronization/reconguration phase which takes
5T and drug is administered in the rst upcoming drug delivery phase which
is 5T + 4T away from the time ischemia is detected. The time limit for drug
administration is 60 seconds, therefore:
30 + 9T 60 (2.1)
We chose T = 2.5sec. Decreasing the time T would decrease the response time
to the faults, however, it will increase the overhead of reconguration/synchronization
and may aect the accuracy of processing and ischemia detection algorithm
since the data sent by the ECG device during drug delivery or synchroniza-
22
tion/reconguration process is lost. No data ow control mechanism is designed
by the manufacturer of the ECG device. The ECG device transmits its data at
115200 bps.
The motes communicate with each other using UART. The interconnection re-
conguration algorithm needs to maintain a broadcast medium in the face of link
failures. The general idea of the algorithm is that the nodes collect connectivity
information by sending ping messages to each other periodically at synchroniza-
tion/reconguration phase. When a node notices that it has not received a re-
sponse from one or more nodes within some timeout period, the node attempts
to recongure its local switch to regain connectivity. Because a node has a map
of the interconnect topology (based on the results of sending ping messages) it
can determine how the network has been partitioned. In our prototype, there are
only four nodes. Therefore, the only partitions possible are horizontal, vertical,
or a corner partitions. In the case of a horizontal (vertical) partition there are
two groups of two nodes. Because a nodes horizontal and vertical connectivity
are independent, a dot-mote is able to change its horizontal connectivity without
aecting its vertical connectivity (and vice versa. When a node detects a hor-
izontal (vertical) partition, it only changes its vertical (horizontal) connectivity
bus. In the case of a corner partition, a single node is partitioned from the other
three. The single node must change both its horizontal and vertical data bus. In
the three node partition, only the two adjacent nodes can aect the connectiv-
ity. A broadcast algorithm is designed to check the connectivity of the network
periodically. In case of a partition, a token is generated to circulate among the
partitioned nodes. After some predened time, when no response is received
from one or more motes, the mote which has the token sends a halt message to
all the motes on the UART bus and will attempt to x the topology based on
the algorithm described before. Upon receiving the halt message, all the other
23
motes initiate to x the topology as well. When xing topology is nished, they
all go into the halt mote. The master mote wakes all the motes up by broadcast-
ing a wake up message once it observes that the network is xed and connected
again. In our implementation, each node is connected to two horizontal and two
vertical buses. Therefore, in the worst case, three faults can be sustained while
maintaining connectivity.
2.6 Experimental Analysis
This section presents various experimental analysis performed on our medical
jacket. Initially we present a picture of our prototype in Figure 2.11. As shown
in the picture, a PC sends ECG signals (extracted from ECG database) to the
medical jacket and a pocket PC is employed as a drug delivery unit (for moni-
toring purposes). The Pocket PC also assists the dot-motes with beat template
recognition by storing the new beat templates in its memory. It, however, does
not carry any computations. The sole processing units in the system responsi-
ble for computations are dot-motes. Our medical jacket is designed such that
it could be attached to the ECG device introduced in Section 2.3.4, however,
for experimental analysis, we used already recorded benchmarks. All the ex-
periments are carried out with real ECG signals. We used ECG signals from
MIT-BIH Arrhythmia Database. The MIT-BIH Arrhythmia Database contains
48 half-hour excerpts of two-channel ambulatory ECG recordings, obtained from
47 subjects studied by the BIH Arrhythmia Laboratory between 1975 and 1979.
The recordings were digitized at 360 samples per second per channel with 11-bit
resolution over a 10 mV range. We used 25 of 48 complete records freely available
from PhysioNet [phya]. We also downsampled all the benchmarks to 250 samples
24
per second to facilitate the compatibility with our ECG device. Each MIT-BIT
record has the recordings of two channel, therefore, in order to generate ECG
signals for 8 channels, each record is replicated and redundantly broadcasted.
The signals extracted from the database are sent to the medical jacket by a
PC through UART. The data format of PC is exactly the same as ECG device.
Hence, our medical jacket is fully functional when we replace the PC with the
ECG device.
In the rst set of experiments, we measured the latency of fault detection and
circuit reconguration for various types of faults as shown in Table 2.1. Wire
faults in general are referred to wire disconnection in the network. V/H wire
faults refer to vertical/horizontal partitions of the broadcast medium. Corner
wire faults refer to corner partitions. ECG and drug delivery faults refer to the
disconnection of wires responsible for carrying ECG and drug delivery signals.
Finally, mote fault corresponds to the loss of dot-motes (due to either physical
failures or power outage).
The next set of experiments are performed on power consumption versus reli-
ability trade-os of the medical jacket. In general, when we decrease the number
of dot-motes on the jacket, the power consumption is reduced while the reliability
of the system also degrades. Reducing the number of dote-motes, however, has
another impact on the system. Since dot-motes are closely collaborating with
each other, reducing them would decrease the amount of memory available to
the system for storing new unknown beat templates. Therefore, there will be a
larger number of unknown beat template queries sent to the Pocket PC. This
would increase the power consumption of the system due to the fact that the
25
Fault Fault Detection
Type Detection + Reconguration
Time (msec) Time (msec)
V/H Wire
Fault 58 63
Corner Wire
Fault 59 67
ECG Wire
Fault 8 9
Drug Delivery
Wire Fault 4 5
Mote Fault 92 N/A
Table 2.1: Fault Detection/Reconguration Time
Pocket PC has a greater power consumption rate than dot-motes. A summary
on the maximum number of faults which can be tolerated while the system is
still functional is presented in Table 2.2. This table also presents the average
power consumption of the system performing Ischemia Detection on MIT-BIT
benchmarks. As shown in the table, the power consumption increases when the
number of dot-motes decreases to one from two. This is mainly due to the heavy
utilization of the Pocket PC for storing the new templates. The Pocket PC that
we employ is iPAQ H5550. During the experiments all unused modules including
the LCD and frontlight were turned o. Also the iPAQ is programmed such that
it automatically goes to the sleep mode after being idle for 30 seconds. Dot-motes
can wake up the Pocket PC when needed. Figure 2.12 illustrates the power con-
sumption of the system for each record of MIT-BIH benchmarks.
26
# of Power # of Failed Worst Case #
Dot-Motes Consumption Motes of Failed Wires
(mW) Tolerated Tolerated
4 155.39 3 3
2 137.20 1 1
1 216.95 0 1
Table 2.2: Power Consupmtion/Reliability of the Medical Jacket
Moreover, in Table 2.3 we present a brief summary on the processing time
required for ischemia detection on each channel of ECG device on average. As
shown, the processing time increases largely with complex ECG signals of the
MIT-BIT benchmarks.
Signal Sample(/Channel) Sample
Complexity Processing Rate
Time (usec) (Hz)
Average 250 250
Highest 420 250
Table 2.3: Computational Capability of Dot-Motes
Finally Table 2.4 illustrates the number of requests sent to the Pocket PC
to inquire about the unknown new beat templates generated for each MIT-BIT
record in various congurations of our medical jacket. As mentioned before, dot-
motes are very constrained in terms of their resources. Their limited memory
27
size allows us to store a limited number of templates. Moreover, once the num-
ber of dot-motes decreases due to the faults, since the remaining motes must
process more ECG channels, the amount of memory assigned for storing beat
template decreases. Therefore, the system experiences more new unknown tem-
plates and they must inquire from the Pocket PC (which has a large memory).
Second column in Table 2.4 refers to the number of request messages sent to the
Pocket PC when the medical jacket employs four motes. The third and the fth
columns represents the case where a fault has occurred and the medical jacket
is functioning with two or one mote respectively. The forth column illustrates
in percentage how the number of request messages are increased in the 2-mote
conguration compared to the 4-mote conguration. The sixth column shows the
same comparison between 1-mote and 2-mote congurations.
2.7 Conclusion
Our project promises to make advances in sensors, materials, and system archi-
tecture for intelligent garments. In the future we plan to expand our prototype,
creating larger topologies, and updating our link reconguration algorithm for use
in larger topologies. In addition, we plan to work on a medical jacket prototype
where stand-by mode would be supported on the dot-motes. This can reduce the
power consumption of our jacket drastically. Lastly, we plan to connected the
ECG sensors directly to the motes to increase the reliability of the system.
28
Figure 2.10: Task Multiplexing in Medical Jacket Network
29
Figure 2.11: Medical Jacket Prototype
Figure 2.12: Power Consumption of Medical Jacket
30
Overhead Overhead
Record # 4 Motes 2 Motes Percentage 1 Mote Percentage
Compared Compared
to 4 Motes to 2 Motes
100 0 0 0.0 1 N/A
101 0 5 N/A 7 40.0
102 9 46 411.1 295 541.3
103 0 0 0.0 1 N/A
104 36 68 88.9 341 401.5
105 115 133 15.7 161 21.1
106 15 40 166.7 305 662.5
107 2 8 300.0 33 312.5
118 0 2 N/A 5 150.0
119 0 0 0.0 7 N/A
200 14 41 192.9 160 290.2
201 5 11 120.0 51 363.6
202 6 11 83.3 17 54.5
203 223 367 64.6 623 69.8
205 0 5 N/A 22 340.0
207 170 228 34.1 347 52.2
208 22 35 59.1 321 817.1
209 4 10 150.0 14 40.0
210 22 37 68.2 98 164.9
212 0 3 N/A 6 100.0
213 4 10 150.0 45 350.0
214 12 33 175.0 120 263.6
215 3 7 133.3 29 314.3
217 22 103 368.2 364 253.4
219 0 3 N/A 10 233.3
Average Overhead Percentage 129.0 265.3
Table 2.4: Number of Unknown Beat Queries Sent to Pokcet PC
31
CHAPTER 3
An Ecient Placement and Routing Technique
for Fault-tolerant Wearable Computers
3.1 Introduction
3.1.1 Motivation
The past few years have seen exciting advances in the development of perva-
sive computing technologies. Computation, storage, and communication are now
more or less woven into the fabrics of our society with much of the progress being
due to the relentless march of Silicon-based electronics technology as predicted
by Moores Law. The emerging eld of exible electronics, where electronic com-
ponents such as transistors and wires are built on a thin exible material, oers a
similar opportunity to weave computation, storage, and communication into the
fabric of the very clothing that we wear, thereby creating an intelligent fabric (also
called electronic textiles or e-textiles) [MM02]. The implications of seamlessly in-
tegrating a large number of communicating computation and storage resources,
mated with sensors and actuators, in close proximity to the human body are
quite exciting; for example, one can imagine biomedical applications where bio-
metric and ambient sensors are woven into the garment of a patient to trigger
and modulate the delivery of a drug. Realizing such novel applications is not just
a matter of developing innovative materials for exible electronics, along with
32
accompanying sensors and actuators. The characteristics of the exible electron-
ics technology and the requirements of the applications enabled by it necessitate
radical innovation in system-level design. Electronic components built of exi-
ble materials have characteristics that are very dierent from that of silicon and
PCB-based electronics. Fault tolerance become more eminent in this architecture
due to criticality of the applications running on this architecture. Further, the
operating scenarios of these systems involve environmental dynamics, physical
coupling, resource constraints, infrastructure support, and robustness require-
ments that are distinct from those faced by traditional systems. This unique
combination requires one to go beyond thinking of these systems as traditional
electronic systems in a dierent form factor. Instead, a re-thinking and complete
overhaul of the system architecture and the design methodology for all layers of
these systems is required.
3.1.2 Driver Application
Although the goal of our project is to investigate system architecture and design
methodology concepts for electronic textiles, successful systems research of this
nature requires that one not only develop these concepts, but also validate them
in context of real technology constraints and application requirements. Therefore,
we use pervasive patient monitoring and sensor-driven personalized trans-dermal
drug delivery as our driver application. One possibility of leveraging electronic
textile technology in the context of such an application is to create a exible
garment (i.e., vest) that the patient can wear, which has sensing, computation,
communication, and actuation elements embedded in it. We are developing such
a prototype vest called medical jacket. Ideally, such a personalized drug delivery
33
vest should have sensors on both the interior (to measure physiological readings)
and on the exterior (to measure environmental readings such as the presence of
toxins in the surroundings), and a software-controlled, electrically-actuated trans-
dermal drug delivery system. It should allow low-latency, ne-grained adaptation
of the drug dosage based on continual physiological measurements in the case of
patients, and based on both environmental and physiological measurements in
the case of people operating in hazardous environments. More generally, our
application driver is representative of biomedical applications where information
technology is integrated into fabrics and textiles. Figure 2.1 shows the overall
system architecture of the medical jacket. As is evident from the Figure 2.1, the
vest consists of four main subsystems: control (or computation), communication,
sensing, and actuation.
3.2 Related Work
Placement and routing problems on a grid have been studied extensively in the
eld of VLSI CAD. There has been considerable research eort to solve these
problems. Like most other VLSI layout problems, it is believed that placement
and routing cannot be solved optimally in a reasonable amount time [She93]
[SW96]. Therefore, heuristic algorithms are used to obtain near-optimal solu-
tions.
Placement algorithms can be generally classied into two main categories: con-
structive method and iterative improvement [SW96]. Constructive methods gen-
erate a placement layout in a global sense. This category includes the min-cut
algorithm [DK85] , the quadratic placement algorithm [Hal70] and the force-
directed algorithm [QB79]. Iterative improvement algorithms starts from an ini-
34
tial solution and tries to improve the solution quality by making changes to the
current solution. Simulated annealing [SS95] and genetic algorithm [SM90] be-
long to the iterative improvement method. Regardless of the basic algorithm be-
ing used, most recent placement tools [KSJ91] [CCK99] [WYS00] [CCK00] adopt
hierarchical or multilevel paradigm by incorporating well-established graph parti-
tioning [KAK97a] or clustering techniques, in order to handle the high complexity
of current designs.
The most basic approach to the routing problem is maze-running algorithm based
on the breadth rst search [Lee61]. Heuristic Steiner tree algorithms are used to
nd a minimum cost path for a multi-terminal net. To route multiple nets, var-
ious approaches can be used such as sequential greedy algorithm, hierarchical
approach [BP83], multicommodity ow-based algorithm [SK87], or integer linear
programming [LTK92].
The main dierence between our approach and other existing techniques is that
we perform both placement and routing simultaneously. The solution that we
nd is also a lower bound for both problems. Our technique easily accommo-
dates our requirements in terms of the instance size and the number of sensors,
actuators and processing units to be placed and routed.
Message routing in large interconnection networks has received a great deal of at-
tention in recent years in the eld of parallel computing. To decrease the amount
of time in transmitting data, second generation multicomputers adopt wormhole
routing mechanism, e.g., Ametek 2010, nCUBE-2, Intel Paragon, J-machine, and
iWARP [NM93]. Adaptive wormhole routing algorithms have been proposed for
mesh, torus, and hypercube topologies. With wormhole routing, a message is
divided into a sequence of xed-size units of data, called its. The header it of a
message contains all the information needed to decide the selection of next chan-
nel on the route. As the header it advances, the remaining its follow it in a
35
pipeline fashion. When the header it reaches a node that has no suitable output
channel available, all of the its in the message are blocked until the channel is
freed. This form of routing technique makes the message transmission time almost
independent of the distance between two nodes if the network is contention-free.
A good survey paper on wormhole routing techniques can be found in [NM93].
The main dierence between our approach and wormhole routing is that we em-
ploy the concept of circuit switching for communication while wormhole routing
utilized packet switching technique. In our model, since we have continuous data
collection from sensors, we employ the concept of circuit switching which pro-
vides dedicated communication line between nodes. Barrenechea et al. studied
optimal routing algorithms for regular sensor networks, namely, square and torus
grid sensor networks, in both, the static case (no node failures) and the dynamic
case (node failures) [BBV04]. Their model, however, utilizes packet switching
approach.
Multihop routing in wireless networks with respect to power optimization and
fault-tolerance have been studied extensively. Generally speaking, a node in a
wireless network independently explores its surrounding region and establishes
connections with other neighboring nodes that are within its transmission and
reception range. Our model signicantly diers since the interconnect network is
not via wireless media. The wired interconnect networks reduces the environmen-
tal eects on the reliability of the network and greatly reduces the communication
power consumption.
Furthermore, energy-aware routing for e-textile platforms was studied in [KM05].
This paper derives an analytical upper bound for the achievable number of jobs
completed over all possible routing strategies. Their routing technique achieves
about fty percent of this analytical upper bound. [SMM03] addresses the mod-
eling of computation, communication and failure in e-textile and investigate the
36
performance of code migration and remote execution. Despite their interesting
approach, our main focus is to recongure/congure the architecture optimally
based on routing/routing an placement. Furthermore, we model the tears and
evaluate the probability of having failures on every segments of wires in our mesh
interconnection network.
3.3 Preliminaries
3.3.1 Interconnection Topology
The interconnection medium for our proposed system is a mesh of wire segments.
The mesh interconnection topology is a wire-frame that has a regular structure,
each vertex being connected to exactly four other vertices.
Mesh networks have several signicant advantages. Each node has a dedicated
communication link with every other node on the network and also has access
to the full bandwidth available for that link. Nodes on buses must share the
bandwidth available on the bus medium. Besides, in a mesh, multiple paths exist
between devices. This brings a great robustness against faults. If a direct path
between two nodes goes down, messages can be rerouted through other paths.
Moreover, it has considerable scalability and can be easily manufactured. The
manufacturing issues become more signicant because in our system wires are
integrated into the fabric and this can be easily achieved with the current fabric
manufacturing technology. Furthermore, the mesh interconnection is highly reg-
ular which assists us in routing and placement of sensors and processing units.
A picture of our fabric with mesh interconnection is shown in Figure 2.2.
In the remainder of this chapter, we refer to mesh interconnection topology by
terminology grid. The interconnection topology is modeled as a grid graph
37
illustrated in Figure 3.1. Each edge in the grid represent a wire or a communica-
tion link and each node represents a processing unit and a switch. Most nodes are
connected to four edges. Switches are responsible for either connecting the pro-
cessing units to any of the edges or establish the connections between edges. In
this chapter, we focus on routing and placement of sensors and processing units.
We emphasize on nding disjoint paths between every pair of communicating
nodes because our communication medium is a mesh of wires and each wire can-
not handle more than one communication at a time. In addition, we chose the
wired interconnection medium due to high available bandwidth of wires and the
requirements of our application (Electrocardiogram signals transmit at 115.2Kbps
in our application).
Figure 3.1: Interconnection topology
3.3.2 Switch Model
Our proposed switch model is shown in Figure 2.3. Our switches are capable of
connecting any terminals to any other terminals.
38
3.3.3 Circuit Switching vs. Packet Switching
Circuit switching has been considered for our proposed technique due to many
advantages it oers to our applications. Our scheme has been designed for ap-
plications with high bandwidth requirements. Therefore, dedicated paths for
communication enhances the overall performance of such systems. Moreover,
due to low delay requirements of most time-sensitive medical applications, the
quality of service can be guaranteed through the circuit switching approach. The
circuit switching scheme demand circuit setup time which may assess an overhead
on the system. Yet, once setup, its overhead on the network becomes negligible
over time. In packet switching scheme, however, the path of each packet is deter-
mined independently. Each packet may travel by a dierent path. Considering
our objective in nding the lower-bound on the system lifetime and the com-
plexity of our formulation, it may not be feasible to impose such overhead for
every packet being transmitted. Finally circuit switching facilitates composition
of larger packets and reduces the overhead of packet headers.
3.3.4 Power Lifetime
The mobility of our system requires the use of constrained power sources (i.e.
batteries). Therefore, in the design of such system, power optimization is one of
the major objectives. The amount of power resources available to the system de-
termines the life time of the system which will be referred to as power lifetime (l
p
)
in this chapter. As shown in the following equations, E indicates the remaining
energy of the system power source while P
w
represents the power consumption
rate of each node in the grid which includes the power consumption of both the
39
processing unit and the switch.
l
p
= System power lifetime
E = Energy remaining in the system power source
P
w
= power consumption of each active node
V is a set of all the nodes of the grid while R
v
consists of the nodes chosen
and utilized for routing. Moreover, the size of set R
v
is denoted by n.
V = {v
i
|v
i
is a node in the grid}
R
v
= {v
i
|v
i
V v
i
used in routing}
n = |R
v
|
Finally, we dene the power lifetime of the system as follows:
l
p
=
E
P
w
n
=
1
p
n
where
p
=
P
w
E
(3.1)
We also dene
p
which will be used in the ILP problem formulation of rout-
ing and placement.
3.3.5 Power Consumption Model
In our proposed architecture, it is crucial to minimize the power consumption
since the system is battery-powered. The power consumption due to processing
is rather constant as long as the processing units remain operational. We, how-
ever, can reduce the power consumption of inter-node communications. When
a communication between a pair of nodes is established, a number of switches
is utilized. The power consumption of the communication is proportional to the
40
number of switches along the path considering that they must remain operational
throughout the communication. The other important factor which is generally
considered is the delay. However, in our system, the communications are estab-
lished using circuit switching scheme. Therefore, the wire length of each path is
rather small and delay concerns become a non-issue.
3.3.6 Fault Lifetime
As discussed in Section 3.3.1, our system is made of fabric which is susceptible to
accidental damage via tears and punctures. Such faults may yield loss of a con-
nection and therefore, may result in total system failure. Fireghters, policemen,
soldiers, astronauts, athletes working in hazardous environments may experience
more faults than others. We dene
f
as the rate of fault occurrence per unit
time. We dene E and R
e
as set of edges in the grid and set of edges used in
routing, respectively:
E = {e
i
|e
i
is an edge in the grid}
R
e
= {e
i
|e
i
E e
i
used in routing}
The set of edges not used in routing is dened by R
e
:
R
e
= E R
e
= {e
i
|e
i
E e
i
NOT used in routing}
When a fault occurs, probability of loosing a particular edge e
i
is represented
by P(e
i
).
P(e
i
) = Pr{e
i
fails|a fault has occurred} (3.2)
41
Assuming that the routing is performed (i.e. the edges for inter-communication
between nodes are selected), when a fault happens, the probability of having sys-
tem failure due to link failures is depicted by P(e
R
).
P(e
R
) = Pr{an edge in R
e
fails|a fault has occurred}
=
e
i
R
e
P(e
i
) =
Obviously, failures in e
i
R
e
does not aect the system functionality since
e
i
R
e
is not utilized for routing. The probability of not having system failure
when a fault occurs is shown with P(e
R
).
P(e
R
) = 1 P(e
R
) = 1
The expected fault lifetime of the system now can be dened in the following
equation. The rst term in the series indicates the probability of having system
failure after the occurrence of the rst fault. The second term denotes the prob-
ability of system failure with two faults while the rst fault refrains the system
intact. The same arguments hold for all the remaining terms.
l
f
= Expected lifetime of the system
=

f
+
2(1)
f
+
3(1)
2
f
+
=

i=1
i(1 )
i1
The expected fault lifetime is shown by Equation 3.3 where
f
and are the
fault occurrence rate and probability of having system failure respectively.
l
f
=

f
(1 (1 ))
2
=
1
(3.3)
42
3.3.7 Fault Model
In this section we model tears or cuts that may occur to the fabric and yield
loss of a connection in the network. For simplicity, we consider only monotone
tears. In practice, this assumption is not far from reality because he results
from tear and puncture analysis experiments on textile conrm that the nature
of tears is usually monotone (WARP and WEFT tears) [WF04]. Various me-
chanical instruments are being used for these experiments such as Elmendorf
(falling pendulum) [Bon94] and COMPUTE (COMbined PUncture and TEar)
[Pri01] on cotton fabric. In addition, several standards have been implemented
for tear analysis on textile such as PE-P-04640:1976 Fabric measurement meth-
ods. Woven and knitted fabrics. Determination of tear strength and PE-EN
ISO 13937-3:2002:1-4 Textiles - Tear properties of fabrics - Parts 1-4.
Unfortunately most reports published in this area only describe statistics about
the strength/length of tears and not about their physical distribution throughout
the fabric. Therefore, the following theoretical tear analysis was carried out with
the assumptions based on the outcome of practical tear analysis experiments.
In Figure 3.2, a few tears are shown. Each tear can be modeled as a path
Figure 3.2: Tear model
43
Figure 3.3: Representation of tears in dual graph
between two nodes in the dual graph of the grid as shown in Figure 3.3. The
dual graph refers to the graph in which each region of the grid is presented with a
node. Whenever two regions in the grid have an edge in common, there exists an
edge between corresponding nodes in the dual graph. According to this model,
the total number of dierent tears that may happen in the fabric is the total
number of distinct paths in the dual graph. Therefore, the number of distinct
tear occurrence on a wire segment in the grid is the same as the number of paths
going through the corresponding edge in the dual graph. Consequently, we must
count the number of paths which passes through each edge in the dual graph. To
show how this is performed, initially, we present the following lemmas.
Lemma 3.3.1. The number of distinct paths from node v
11
to v
nm
on a grid is:
P(v
11
, v
nm
) =
n 1
r 1
(3.4)
44
Proof: Length of each path from v
11
to v
nm
is m+n 1 and each path can
be represented with a string of length m + n 2 containing of right (R) and up
(U) moves. Each string has exactly n 1 (U)s. The total number of distinct
strings is shown in (3.4) which is equivalent to the number of paths.
Now, consider an edge w which connects nodes v
ij
to v
ij+1
as shown in Figure
Figure 3.4: Region subdivisions for every edge in dual graph
3.4. Since we consider only monotone tears, the total number of paths which
passes through w, T(w), can be evaluated by (3.5):
T(w) = R(I, v
ij
) R(IV, v
ij+1
)
+ R(III, v
ij
) R(II, v
ij+1
)
(3.5)
R(X, v
ij
) is the total number of paths connecting every node in rectangle X
to node v
ij
. Possible rectangles for X are shown in Figure 3.4. We have:
R(I, v
i,j
) =
i
s=1
j
t=1
P(v
st
, v
ij
)
45
To nd a closed form for R(I, v
i,j
) we observe that:
i
s=1
P(v
s1
, v
ij
) =
i
s=1
j 1 +i s
j 1
j +i 1
j
This implies that the total number of paths originating from every node on
the left border of rectangle I to v
ij
is equal to P(v
11
, v
ij+1
).
Finally we conclude that:
R(I, v
ij
) =
j
t=1
i +t 1
t
i +j
j
1 (3.6)
Similar equations hold for R(IV, v
ij+1
), R(III, v
ij
), R(II, v
ij+1
). Note that
we only considered horizontal edges. The same argument is valid for vertical
edges. We previously showed that each tear in the grid is represented by a path
in the corresponding dual graph. Therefore, assuming that a tear occurs, the
probability of loosing an edge e
i
in the grid is proportional to the number of
distinct paths in the dual graph crossing that edge. In other words, it would be
proportional to the total number of paths that passes through w
i
(i.e. T(w
i
)),
which corresponds to e
i
in the dual graph. To nd the probability of loosing an
edge, we calculate total possible tears in the network which is equivalent to all
possible paths between any two nodes (v
ij
, v
pq
) in the dual graph. Therefore, the
total number of possible tears in the network is:
T
N
= 2
n+1
i=1
m+1
j=1
i
p=1
j
q=1
j q +i p
j q
T
straight
(3.7)
The nested summation is taken over all pairs of (v
ij
, v
pq
) and is multiplied by
two to take into account both descending and ascending paths. T
straight
is the
total number of straight paths including paths with length zero (single nodes).
46
Since straight paths are being considered as both ascending and descending, they
will be counted twice in the rst term of the above equation. Therefore, T
straight
must be deducted from the nested summation in Equation 3.7. T
straight
can be
simply written as:
T
straight
= (n + 1)
m+ 1
2
+ (m+ 1)
n + 1
2
+ (n + 1)(m+ 1) (3.8)
The latter term in Equation 3.8 is basically the number of paths in the graph
with length zero (the number of nodes in the graph). Therefore, the total number
of possible tears in the network can be expressed as:
T
N
= 2
n+1
i=1
m+1
j=1
(
i +j
j
1)[(n+1)
m+ 1
2
+(m+1)
n + 1
2
+(n+1)(m+1)]
(3.9)
Hence, probability of having failure on edge e
k
is calculated as follows:
Pr(e
k
) =
T(w
k
)
T
N
(3.10)
In Equation 3.10, each w
i
in dual graph corresponds to e
i
in the grid. Figure
3.5 illustrates the probability of failure on each edge in the grid when a tear occurs.
In the model we illustrated, our analysis show that some wires are more sus-
ceptible to failures. If the wires already utilized for communication fail, the
system will fail. Furthermore, since the lifetime of the system is determined by
the shortest lifetime of all wire segments selected for routing, we prefer to employ
47
Figure 3.5: Probability of edge failure in the grid for all possible distinct tears
wires which are less susceptible to faults.
3.3.8 Integer Linear Programming Formulation
ILP is used to specify problems where a function has to be maximized or min-
imized and the variables are constrained by inequality and equality constraints
and/or integral restrictions on either a subset or all the variables. The objective
function and the inequality or equality constraints are all linear. We dene the
mixed integer programming problem (MIP) as:
max{cx +hx : Ax +Gy b, x Z
n
+
, y R
n
+
}
48
where Z
n
+
is the set of non-negative integral n-dimensional vectors, R
n
+
is
the set of non-negative integral p-dimensional vectors, and x = {x
1
, ..., x
n
} and
y = {y
1
, ..., y
p
} are the variables in the problem. We restrict our attention to a
special case of the MIP where all variables may have only integer values. The
integer linear programming (ILP) is dened as
max{cx : Ax b, x Z
n
+
}
More specically, we restrict the integer values of x to be either 0 or 1. There-
fore, 0-1 ILP where x is redened as x B
n
where B
n
is a set of n-dimensional
binary vector.
3.3.9 Terminology
In the remainder of this work, we will specify the ILP problem in the following
way.
y = MAX(cx) (3.11)
Ax b (3.12)
Equation 3.11 species the objective function for the problem where we dene
x = x
1
, ..., x
n
as 0-1 variables. The constraints of the problem can be formulated
in matrix form as shown in Equation 3.12. In the remainder of the chapter, we
will use uppercase letter representation to specify constant matrices or values,
49
and lowercase letters to represent variables.
3.4 Problem Denition
We introduced the terminologies power lifetime and fault lifetime as well as the
tear analysis technique. The lifetime of our system is highly dependent on the
remaining energy in the power source and the frequency of fault occurrence.
Sometimes the system may fail due to lack of energy and in some cases, faults
may parallelize the system. To increase the power lifetime, once we perform
placement and routing, we seek to minimize the total length of routing. To im-
prove the fault lifetime, we attempt to use the edges which are less susceptible
to faults. The overall lifetime of the system is dependent both on power and
fault lifetime. Therefore, to maximize the overall lifetime of the system, we shall
maximize the minimum of power and fault lifetimes.
MAX(MIN(l
p
, l
f
)) (3.13)
However, the objective function in an ILP formulation may not contain both
MAX and MIN as shown in (3.13). Therefore, we redene the objective func-
tions as illustrated in (3.14) and (3.15) and have the solver optimize both objec-
tives individually in two dierent runs.
MAX(l
p
)
with constraint l
p
< l
f
(3.14)
50
MAX(l
f
)
with constraint l
f
< l
p
(3.15)
The maximum value of (3.14) and (3.15) is the value of objective function
(3.13). In Section 3.5 we specify in details how the objective functions are con-
structed.
3.5 Problem Formulation
3.5.1 Routing
In both placement and routing problems, we assume that the interconnection
network is a grid of size M N. The vertices correspond to processing units
along with switches while the edges corresponds to the links between them. Each
vertex and edge in the grid is assigned a variable. These variables specify if an
edge and/or vertex is selected for routing and/or placement. Further details on
how these variables are used will be explained later. However, the accordance
between the indices of variables and the position of vertices and/or edges are as
follows:
Variable x
pij
represent if node with coordinations i and j in the grid is chosen
for the pth communication pair. As shown in Figure 3.6, variables y
hij
and y
vij
represents horizontal and vertical edges corresponding to node ij in the grid, re-
spectively.
51
Figure 3.6: ILP variables
Problem: Routing to Minimize the Power Consumption
Instance: Graph G(V, E) s.t. G is a grid, collection of disjoint vertex pairs
(s
1
, t
1
), (s
2
, t
2
), ..., (s
l
, t
k
)
Question: Find k mutually vertex-disjoint paths in G, connecting s
i
to
t
i
for each i, 1 i k s.t. the total length of paths is minimum
In the ILP formulation we dene variables x
pij
where p = 1, ..., k. x
p
ij rep-
resents whether vertex v
ij
is selected as a part of the path between (s
p
, t
p
). We
also dene q
pij
as follows:
52
x
pij
=
1, if vetex v
ij
is selected for
communicating pair (s
p
, t
p
)
0, otherwise.
(3.16)
y
(h/v)pij
=
1, if edge y
(h/v)pij
is selected for
communicating pair (s
p
, t
p
)
0, otherwise.
(3.17)
q
pij
=
1, if vertex v
ij
= s
p
v
ij
= t
p
0, otherwise.
(3.18)
Using the above variables, we dene the constraints (3.19), (3.20) and (3.21)
for the ILP formulation. (3.19) ensures that a feasible path is found between s
p
and t
p
while (3.20) denotes that an each edge in the grid may not be used for
more than one path. (3.21) enforces that the paths are vertex disjoint. To ensure
that a path is feasible, we enforce the condition that exactly two neighboring
edges of a vertex must be selected if the vertex is chosen for routing. In the case
where the vertex is not chosen, none of the neighboring edges may be selected.
This condition is slightly dierent for the source and sink vertices of the path
since exactly one neighboring edge must be selected for routing. This condition
is expressed in (3.19).
Finally the objective function, represented in (3.22), attempts to minimize the
power consumption or the number of nodes employed for routing. Subsequently,
this would maximize the power lifetime.
53
2x
pij
y
phij1
y
pvi1j
y
phij
y
pvij
q
pi
= 0
1 p k
1 i M
1 j N
(3.19)
k
p=1
y
p(h/v)ij
1 (3.20)
k
p=1
x
pij
1 (3.21)
MIN(
k
p=1
m
i=1
n
j=1
x
pij
) (3.22)
Problem: Routing to Maximize the Fault Tolerance
Instance: Graph G(V, E) s.t. G is a grid, fault occurrence weights on edges of
G, collection of disjoint vertex pairs (s
1
, t
1
), (s
2
, t
2
), ..., (s
l
, t
k
)
Question: Find k mutually vertex-disjoint paths in G, connecting
s
i
to t
i
for each i, 1 i k s.t. (3.23) is minimum.
Tear analysis described in Section 3.3.7 generates probability of having fail-
ures in every edge of the grid once a tear occurs. This information helps us to
improve the fault tolerance of the system. As opposed to the previous problem,
the objective function is to pick edges for routing such that the probability of
fault occurrence in the selected edges is minimized. Therefore, the constraints
(3.19), (3.20) and (3.21) still hold while the objective function (3.22) changes
according to the new requirements.
54
MIN(
k
p=1
m
i=1
n
j=1
(P(e
hij
)y
hpij
+P(e
vij
)y
vpij
)) (3.23)
Problem: Routing to Maximize the Fault Tolerance and Minimize the Power
Consumption
1
, t
1
), (s
2
, t
2
), ..., (s
l
, t
k
)
i
to
t
i
for each i, 1 i k s.t. (3.24) and/or (3.25) is minimum.
The rst problem that we discussed earlier in Section 3.5.1 addresses the is-
sue of power optimization. It attempts to minimize the total number of nodes
selected for routing. Since selecting each node in the grid requires activation of
a switch and a processing unit, therefore the objective function minimizes the
power consumption of the network. The previous problem, however, denotes
how we can improve the fault tolerance of our interconnection medium through
using edges which are less exposed to faults. Now we present a new problem
formulation which addresses both power optimization and fault tolerance issues
simultaneously. The constraints (3.19), (3.20) and (3.21) are still valid, however,
the objective functions (3.24) and (3.25) determine the lower bound for the life-
time of our system as described in Section 3.4.
55
MIN(
k
p=1
m
i=1
n
j=1
x
pij
)
with respect to
k
p=1
m
i=1
n
j=1
x
pij
>
k
p=1
m
i=1
n
j=1
(P(e
hij
)y
hpij
+P(e
vij
)y
vpij
)
(3.24)
MIN(
k
p=1
m
i=1
n
j=1
(P(e
hij
)y
hpij
+P(e
vij
)y
vpij
))
with respect to
k
p=1
m
i=1
n
j=1
(P(e
hij
)y
hpij
+P(e
vij
)y
vpij
) >
k
p=1
m
i=1
n
j=1
x
pij
(3.25)
The next problem that we present is an extension to the previous problem
where each edge in the grid carries more than one wire. We dene the channel
capacity for each edge in the grid and assume that each edge contains c wires.
Thus, at most c disjoint paths may use one edge simultaneously.
Consumption
1
, t
1
), (s
2
, t
2
), ..., (s
l
, t
k
)
i
to t
i
for each i, 1 i k s.t. each edge in E is not asso-
ciated with more than c paths and (3.24) and/or (3.25) is minimum.
With our new assumption, constraint (3.19) is still eective while (3.26) and
56
(3.27) are enforced as well. The objective function remains the same as (3.24)
and (3.25).
k
p=1
y
p(h/v)ij
c (3.26)
k
p=1
x
pij
c (3.27)
The last ILP formulation for routing is for the case where a processing unit
is connected to multiple sensors. This case happens where multiple sensors com-
municate with a single processing unit simultaneously and eciently.
Consumption
G, collection of disjoint tuples (S
1
, t
1
), (S
2
, t
2
), ..., (S
k
, t
k
)
Question: Find k mutually vertex-disjoint paths in G, connecting all s
i

S
i
to t
i
for each i, 1 i k s.t. each edge in E is not as-
sociated with more than c paths and (3.24) and/or (3.25) is minimum.
Constraints (3.20), (3.21) and (3.26) still hold. Only constraint (3.19) is mod-
ied and redened in (3.28). |S
p
| is the cardinality (i.e. the number of members)
of S
p
.
57
|S
p
|x
pij
y
phij1
y
pvi1j
y
phij
y
pvij
q
pi
= 0
1 p k
1 i M
1 j N
(3.28)
3.5.2 Placement
Traditional placement and routing problems in VLSI CAD are considered as two
separate problems. In this section, we present an ILP formulation where we solve
both placement and routing of our sensor network simultaneously. In most cases,
the medical application itself derives the location of sensors. However, in some
applications, the exibility of placing sensors within a specic range may be be-
stowed. This can be easily accommodated by our proposed ILP technique and
the optimal location of sensor to minimize our objective function will be found.
For instance, a temperature sensor may be required to be placed on chest, but
its exact location is not given. In this case, our technique attempts to optimize
its location too.
58
Problem: Placement and Routing to Maximize the Fault Tolerance and Mini-
mize the Power Consumption
G, collection of disjoint vertex set (S
1
, S
1
), (S
2
, S
2
), ..., (S
k
, S
k
) s.t. s
i
S
i
V
and t
i
S
i
V , S
i
and S
i
are set of nodes which can potentially be source or
sink of a path.
i
to
t
i
for every i, 1 i k s.t. (3.24) and/or (3.25) is minimum.
In this problem, the nodes which are candidates to be either source or sink
of the paths are given in sets S
i
s and S
i
s. For each tuple (S
i
, S
i
), however,
exactly one node from S
i
and one node from S
i
will be selected. In the case
where the position of a node is xed (i.e. the node must be placed at a specic
position), the set S
i
or S
i
may have only one member. To solve this problem,
our ILP formulation is modied and we dene two sets of new variables for each
communication pair.
s
pij
S
p
s
pij
= 1
1 p k
1 i M
1 j N
(3.29)
59
pij
S
p
s
pij
= 1
1 p k
1 i M
1 j N
(3.30)
2x
pij
y
phij1
y
pvi1j
y
phij
y
pvij
s
pij
s
pij
= 0
1 p k
1 i M
1 j N
(3.31)
The constraints (3.29) and (3.30) enforce that only one node from sets S
i
or S
i
is selected as a source or a sink of a path. Constraint (3.31) is modied
from (3.19) to ensure that a feasible path for every communication pair is found.
Constraint (3.20), (3.21) and objective functions (3.24) and (3.25) still hold for
this problem.
We can apply all the variations of routing problems described in Section 3.5.1 to
the placement and routing formulations.
In this Section, we present the simulation results on fault/power lifetime opti-
mization on various benchmarks. We used the CPLEX solver on a 400MHz Sun
Ultra-10 machine and allowed the maximal runtime to be 30 minutes. Various
benchmarks are generated for the experimental analysis. The size of benchmarks
varies from 5 5 to 11 11. The number of communication pairs are altered
60
as well. The communication pairs are placed across the grid randomly for each
instance size. The power consumption of the processing units is assumed to be
P
w
= 24mW (Berkeley mica2dots - power supply voltage

= 3V ). We also as-
sume that the system is equipped with ten 1.5V , 2000mAh AA batteries. Hence,
p
= 0.0008h
1
. Fault occurrence rate,
f
, picks the values of 1/24h
1
, 1/30h
1
and 1/36h
1
.
f
= 1/24h
1
implies that on average, a random tear occurs in
every 24 hours.
Fault Lifetime (hours) Power Lifetime (hours)
f
= 1/24 h
1
51.35 N/A
f
= 1/30 h
1
56.25 55.30
f
= 1/36 h
1
N/A 64.38
Table 3.1: Power/fault lifetime comparison in 5x5 grid
Fault Lifetime (hours) Power Lifetime (hours)
f
= 1/24 h
1
31.74 N/A
f
= 1/30 h
1
33.71 33.15
f
= 1/36 h
1
N/A 40.11
Table 3.2: Power/fault lifetime comparison in 7x7 grid
Table 3.1 illustrates the simulation results for a 5 5 grid with 8 communi-
cation pairs. The rst and the second columns present the values of objective
function (3.25) and (3.24), respectively. As shown in the Table 3.1, for small
values of
f
(i.e. = 1/24), the routing must be performed with most emphasis on
power lifetime optimization while for larger
f
s, (i.e. = 1/36), the fault lifetime
61
Figure 3.7: Fault/power lifetime ratio for 7x7 mesh size
is optimized to achieve the best results. For
f
= 1/30h
1
, both fault and power
objective functions yields approximately similar lifetimes. Table 3.2 exhibits sim-
ilar simulation results for a 7 7 grid. Figure 3.7 represents fault/power lifetime
ratio for a 7 7 grid. In this set of experiments, the location of source and
sink nodes in each communication pair was selected randomly across the grid
such that the rectilinear distance between source and sink nodes is not greater
than l
max
where 1 l
max
7. Both the number of communication pairs and
l
max
are varied. Each plot in Figure 3.7 corresponds to a specic l
max
. For each
data-point, 20 benchmarks were generated and the average ratio of all 20 bench-
marks is shown in the diagram. Overall, 1540 benchmarks were generated for
f
= 1/30h
1
. Ratio greater than one indicates that the system is less sensitive
to faults and most probably dies due to the power outage while ratio less than 1
implies that most likely the system becomes paralyzed due to the faults. Accord-
ing to Figure 3.7, the fault/power lifetime ratio drops to less than one for large
62
number of communication pairs distributed randomly with l
max
3. Hence,
more eorts must be spent on fault lifetime optimization of such benchmarks
while in less congested networks, power lifetime gains more importance. This is
due to the fact that in congested networks we are forced to use edges that are
highly susceptible to faults and therefore fault lifetime decreases rapidly while
the power lifetime is not decreasing at the same order.
Figure 3.8: Lifetime analysis (
f
= 1/24 h
1
)
Figures (3.8), (3.9) and (3.10) present lower bounds for both power and fault
lifetime of systems with various congurations. Each data-point again reects the
average of lifetime acquired from 20 benchmarks. In each benchmark, the source
and the sink of communication pairs are placed with uniform distribution across
the grid where l
max
= instance size. As shown, once the fault occurrence rate,
f
, increases, the overall lifetime of the system decreases and more eorts must
be spent on fault tolerant routing. As illustrated in all diagrams, the solver was
unable to nd a feasible solution for any 30 communication pairs placed randomly
on a 5 5 grid within 30 minutes.
63
f
= 1/30 h
1
)
In Figures (3.11), (3.12) and (3.13), the lower bounds for both power and fault
lifetime are represented where the size of the grid was set to 7 7 and l
max
as
well as the number communication pairs are varied accordingly. Each data-point
is the average of 20 benchmarks. The other observation is that when the number
of communication pairs grows, if the network becomes relatively congested, both
power and fault lifetime become relatively constant and similar. The reason is
that most resources of the system is already utilized and the routing algorithm
do not have the exibility of choosing the best solution with respect to either
power optimized or fault tolerant routing techniques.
Our objective may not be achieved by a greedy method since it attempts
to maximize the minimum of two objective functions. However, to exhibits the
benets of our proposed formulation over a greedy approach, we considered a
method where a communication pair is selected randomly and its shortest path
based on the power objective function is calculated. This process is repeated
until shortest paths between all pairs are determined. If a feasible solution is
64
f
= 1/36 h
1
)
not found, another sequence of communication pairs is selected until a solution
is found. The system lifetime in such network corresponds to the minimum of
fault and power lifetimes. The same greedy approach is repeated while the fault
objective functions is considered. The greedy fault lifetime is compared against
the greedy power lifetime and the maximum is picked and reported in Figures
3.14 and 3.15. These set of experiments are performed on a 7 7 grid while the
fault rate are assumed to be
f
= 1/24h
1
and
f
= 1/36h
1
for Figures 3.14
and 3.15, respectively. Each data-point is the average of 20 7 random bench-
marks where 1 l
max
7. In some benchmarks, the greedy approach was not
successful in nding a feasible solution. In such cases, the benchmark did not
contribute to the data-point. On average, the lifetime of the system designed by
our proposed technique for
f
= 1/24h
1
is 79% higher than the greedy method.
The improvement has been 103% for
f
= 1/36h
1
. It illustrates the eectiveness
of our proposed technique when the fault rate increases.
65
Figure 3.11: Lifetime analysis on xed mesh size varying communication
distance(
f
= 1/24 h
1
)
3.7 Conclusion
Our proposed ILP technique has the two following properties:
It considers placement and routing problems at the same time targeting our
performance objective. Generally, placement and routing are being treated
as two independent design steps and each is quite complex. Therefore, opti-
mizing the nal solution or nding near optimal answer become even more
challenging. Yet, our proposed formulation attempts to perform both two
design steps simultaneously while acquiring a lower bound on the solution.
Power consumption and reliability are two major performance measures
66
distance(
f
= 1/30 h
1
)
in embedded computing systems. Given that the aforementioned objective
metrics have dierent nature, optimizing both is impossible. However, both
aect the lifetime of the system independently. By applying the concept
of lifetime and dening fault lifetime and power lifetime, these ob-
jectives are merged into a single measure. The resulting ILP algorithm
guarantees to optimize system lifetime taking into consideration both ob-
jectives. Usually in such scenarios, the problem is treated as a constrained
optimization in which one objective is bounded (becomes a constraint) while
the other is optimized. Our proposed ILP scheme uses the same concept
in two dierent phases which provably results in optimizing both objectives.
67
distance(
f
= 1/36 h
1
)
Our proposed ILP technique is an oine task in our system design procedure. In
other words, it does not attempt to recongure the system after the design stage.
The oine solution is used to optimally place and route sensors and actuators
in grid based networks. In this chapter, we are not considering any recovery
scheme from faults. The main objective is to avoid faults by placing and rout-
ing the sensors using locations less prune to tears. Specically, with respect to
our target architecture and the nature of faults in corresponding applications,
a tear can potentially cut many communication links at the same time whereas
in fault tolerant routing algorithms, recovery from single link loss (or a limited
number of link loss) is being considered.
Maximizing system lifetime in more general network topologies may be consid-
ered for future works. Although a grid network is a common topology for many
wearable embedded networked systems, removing this limitation can potentially
68
Figure 3.14: Performance comparison of our technique with a greedy method
(
f
= 1/24 h
1
)
lead to a more general methodology for fault prune networks. In addition, exploit-
ing network architectures in which recovery from faults can be done eciently is
unquestionably an interesting and a challenging problem that may be studied in
future.
69
Figure 3.15: Performance comparison of our technique with a greedy method
(
f
= 1/36 h
1
)
70
CHAPTER 4
Adaptive Electrocardiogram Feature Extraction
on Distributed Embedded Systems
4.1 Introduction
The electrocardiogram (ECG) is the record of variation of bioelectric potential
with respect to time as the human heart beats. Due to its ease of use and non-
invasiveness, ECG plays an important role in patient monitoring and diagnosis.
Multichannel electrocardiogram (ECG) data provide cardiologists with essential
information to diagnose heart disease in a patient. Our primary objective is to
address feasibility verication of implementing an ambulatory ECG analysis al-
gorithm with real-time diagnosis functions for wearable computers. ECG analysis
algorithms have always been very dicult tasks in the realization of computer
aided ECG diagnosis. Implementation of such algorithms becomes even harder for
small and mobile embedded systems that should meet the given latency require-
ments while minimizing overall energy dissipation for the system. Distributed
embedded systems are successfully deployed in various wearable computers. Dis-
tributed architectures have been developed for cooperative detection, scalable
data transport, and other capabilities and services. However, the complexity
of algorithms running on these systems has introduced a new set of challenges
associated with resource constrained devices and their energy concerns. These
71
obstacles may dramatically reduce the eectiveness of embedded distributed al-
gorithms. Thus, a new distributed, embedded, computing attribute, dynamically
recongurable, must be developed and provided to such systems. In these sys-
tems, reconguration capability in particular, may be of great advantage. This
capability can adaptively alter the system conguration to accommodate the ob-
jectives and meet the constraints for highly dynamic systems.
There have been exciting advances in the development of pervasive computing
technologies in the past few years. Computation, storage, and communication
are now more or less woven into the fabrics of our society with much of the
progress being due to the relentless march of Silicon-based electronics technology
as predicted by Moores Law. The emerging eld of exible electronics, where
electronic components such as transistors and wires are built on a thin exible
material, oers a similar opportunity to weave computation, storage, and com-
munication into the fabric of the very clothing that we wear, thereby creating an
intelligent fabric (also called electronic textiles or e-textiles) [MM02]. The impli-
cations of seamlessly integrating a large number of communicating computation
and storage resources, mated with sensors and actuators, in close proximity to
the human body are quite exciting; for example, one can imagine biomedical ap-
plications where biometric and ambient sensors are woven into the garment of
a patient to trigger and modulate the delivery of a drug. Realizing such novel
applications is not just a matter of developing innovative materials for exible
electronics, along with accompanying sensors and actuators. The characteristics
of the exible electronics technology and the requirements of the applications
enabled by it necessitate radical innovation in system-level design. Electronic
components built of exible materials have characteristics that are very dierent
from that of silicon and PCB-based electronics. Further, the operating scenar-
ios of these systems involve environmental dynamics, physical coupling, resource
72
constraints, infrastructure support, and robustness requirements that are distinct
from those faced by traditional systems. This unique combination requires one
to go beyond thinking of these systems as traditional electronic systems in a
dierent form factor. Instead, re-thinking and complete overhaul of the system
architecture and the design methodology for all layers of these systems are re-
quired.
4.2 Related Work
Georgia Tech [PMJ02]. The Wearable Motherboard is woven into an under-
shirt in the Smart Shirt design. Their interconnect is a exible data bus that
can support a wide array of sensory devices. These sensors can communicate
via the data bus to a monitoring device located at the base of the shirt. The
monitoring device is integrated into a single processing unit that also contains
a transceiver. The SmartShirt design features plastic optical ber that can be
used to detect punctures - however, they do not have any means of dealing with
these punctures other than reporting them via the transceiver. In the event
that their single processing unit (and transceiver) is damaged or the control lines
leading to the device are torn, the Smart Shirt is virtually inoperable. In our
design, we distribute control to multiple processing elements and can accommo-
date multiple communication buttons. This ensures that a fault in our system
73
will not result in total system failure. The fault tolerance is essential in de-
manding and/or hazardous environments targeted by this research. Fireghters,
policemen, soldiers, astronauts, athletes, and others working in hazardous en-
vironments need robust material that can sustain damage and yet still reliably
provide service. Several other technologies have been introduced by MIT called
MIThril [DMP03], e-Textile from Carnegie Mellon University [MMK02], Wear-
able e-Textile from Virginia Tech [MJE03], CustoMed and RFab-Vest from UCLA
[JEZ05][JDB05]. Furthermore, the Lifeguard project being conducted at Stan-
ford University is a physiological monitoring system comprised of physiological
sensors (ECG/Respiration electrodes, Pulse Oximeter, Blood Pressure Monitor,
Temperature probe), a wearable device with built-in accelerometers (CPOD),
and a base station (Pocket PC). The CPOD acquires and logs the physiolog-
ical parameters measured by the sensors [lif]. The Assisted Cognition Project
conducted at the University of Washingtons Department of Computer Science
explored the use of AI systems to support and enhance the independence and
quality of life of Alzheimers patients. Assisted Cognition systems use ubiquitous
computing and articial intelligence technology to replace some of the mem-
ory and problem-solving abilities that have been lost by an Alzheimers patient
[KEF03]. Nevertheless, none of the above projects/systems supports the concept
of scalability and adapting complex processing algorithms.
4.3 Automated Feature Set Detection
Given the goal of classifying objects based on their attributes, the functionality
of an automated pattern recognition system can be divided into two basic tasks:
the description task generates attributes of an object using feature extraction
techniques, and the classication task assigns a group label to the object based
74
on the attributes with a classier.
There are two dierent approaches for implementing a pattern recognition sys-
tem: statistical and structural. Each approach utilizes dierent schemes within
the description and classication tasks which incorporates a pattern recognition
system. Statistical pattern recognition [DHS00][JDM00] concludes from statisti-
cal decision theory to discriminate among data from dierent groups based upon
quantitative features of the data. The quantitative nature of statistical pattern
recognition, however, makes it dicult to discriminate among groups based on
the morphological (i.e., shape-based or structural) subpatterns and their interre-
lationships embedded within the data. This limitation provided the impetus for
development of structural approaches to pattern recognition.
Structural pattern recognition [Gra95][Pav77] relies on syntactic grammars to
discriminate among data from dierent groups based upon the morphological in-
terrelationships (or interconnections) present within the data. Structural pattern
recognition systems are eective for image data as well as time-series data.
We have investigated an accurate ECG processing algorithm based on structural
pattern recognition (as depicted in Figure 4.1) mapped onto our processing units
(dot-motes) [xbo]. The algorithm consists of three stages: preprocessing, pattern
recognition and classication. We perform preprocessing and pattern recognition
locally, i.e. within close proximity to the ECG leads. The preprocessing includes
ltering while the pattern recognition includes heartbeat detection (through the
QRS complex detection), segmentation as well as feature extraction. Once the
features are extracted, they will be processed for classication.
The ltering is performed by nite impulse response (FIR) lters with cut-o
frequencies of 5-150 Hz for sampling rate of 360 samples/sec. The heartbeat de-
tection is implemented with a QRS detector based on the algorithm of Pan and
Tompkins [PT85] with some improvements that employs slope information. The
75
scheme proposed by Laguna et al [LMG97] is used to extract the ducial points.
All oset and onset points are detected based on the location and convexity of
R point. We detect each point onset by locating the largest isoelectric region
before the point. Then we search for the inection point followed by largest
negative slope for convex R-wave or largest positive slope for concave R-wave.
We also detect the point oset by searching for signicant up slope following the
end of the last down slope for P, T and S osets in particular. Consequently,
features relating to heartbeat intervals and ECG morphology are calculated for
each heartbeat. The list of features is included in Table 4.1 and are based on
[COR04] and [CB04] with minor additions. In addition, a sample ltered ECG
signal which was automatically segmented by our tool is depicted in Figure 4.2.
We extract a total of 23 features from the ECG signals, and each derives from
one of the groups below:
RR Interval Features: We extract four features based on RR Intervals. The
RR interval is the interval between two successive heartbeat ducial points, ob-
tained from the maximum of the R-wave. The pre-RR interval is the RR-interval
between a detected heartbeat and the previous one. The post-RR interval is the
interval between a given heartbeat and next detected one. The average-RR inter-
val is the average of all detected RR intervals, and the local average-RR interval
is the average of the ten most recent RR-intervals.
Heartbeat Interval Features: We extract ve features related to heartbeat
intervals. QRS duration is the time between QRS oset and QRS onset. T-wave
duration is the time between T-wave onset and T-wave oset. The PR, ST, and
QT duration are additions to the automated classication system. ST duration
is the time between S-wave oset and T-wave onset. The PR duration is the
time between P-wave onset and R. The QT duration is the time between Q-wave
76
Group label Features
RR Interval Pre-RR interval
Post-RR interval
Average-RR interval
Local average-RR
Heartbeat Interval QRS duration
T-wave duration
PR duration
ST duration
QT duration
Geometric Baseline (DC level)
Number of samples in 70%-100% of R-amplitude
Maximum positive peak
Maximum negative peak
Slope velocity between Q-onset and R
Slope velocity between R and S
Morphology 1A ECG morphology (10 samples for sampling rate: 360 s/sec)
between Q and S
Morphology 1B Normalized ECG morphology (10 samples for sampling rate: 360 s/sec)
between Q and S
between S and T-wave oset
between S and T-wave oset
between R-50 ms and R+100 ms
between R-50 ms and R+100 ms
between S + 150 ms and S + 500ms
between S + 150 ms and S + 500ms
Table 4.1: Features categorized by groups
77
onset and T-wave oset. All of these features are obtained by rst determining
the start and end point of each interval, and then subtracting the end point from
the start point.
Geometric Points: We calculate the signal DC shift level by taking the aver-
age base line of the previous ve successive detected heartbeats. The maximal
positive and minimal negative peaks are detected by computing the voltage dif-
ference between each sample in the heartbeat and DC shift level. In addition, we
extract the number of samples in 70-100% range of absolute peak value. Finally
we compute the slope velocity of Q-onset - R as well as R - S segments.
ECG Morphology Features: We extract eight features based on ECG mor-
phologies arranged into four groups. Two groups consist of samples from heart-
beat segments and two groups consist of samples from xed intervals. Within
each group, one feature consists of samples from the original ECG signal, while
the other feature is extracted from the normalized ECG signal. The normaliza-
tion is done through scaling down the amplitude of samples by standard deviation
of the same heartbeat.
We extract samples from heartbeat segments in ECG morphology 1 and 2 (see
Figure 4.3). In morphology 1, ten samples between QRS onset and oset are
extracted, and in morphology 2, nine samples between S-wave oset and T-wave
oset are obtained. The number of samples collected is also contingent upon the
sampling rate and scales with various sampling rates (the aforementioned num-
bers are for the original sampling rate of 360 samples per second).
We extract samples from a xed interval in ECG morphology 3 and 4 (see Fig-
ure 4.4). In morphology 3, ten samples between R50ms before and R+100ms
are extracted, and in morphology 4, eight samples between between R + 150ms
and R + 500ms are acquired.
78
For all ECG morphologies, the elements that fall in between two samples are
estimated using linear polarization. We have repeated such feature extraction
for three input sampling rates of 360, 200 and 100 samples per second. 360 sam-
ples/seconds is the original sampling rate for the MIT/BIH [MM90] benchmarks
and the sampling rates of 200 and 100 samples/second was acquired by down-
sampling the input.
Despite our objective is to minimize the communication among processing nodes
before the classication phase, this study does not investigate the problem of clas-
sication. Therefore, we did not implement a classier for our platform. However,
any classier suitable for constrained embedded systems may be deployed.
4.4 Software Proling
To measure the execution delay of our heartbeat detection and feature extraction
system, we used Avrora [TLP05], a microcontroller simulator framework devel-
oped at UCLA. Avrora is a precise and exible simulator that preserves all timing
and behavior of the instrumented program, while allowing user-dened proling
of application information. With Avrora, users can easily prole application spe-
cic information such as branch frequency, maximum stack size, and memory
access by adding custom program monitors.
For our experiments, we implemented a program monitor on Avrora that gener-
ates the control ow graph (CFG) while measuring the execution frequency and
delay of each basic block. Since the CFG of our system is very large, only the
major processes are shown in Figure 4.5. The CFG is dynamically generated
based upon our compiled and assembled ECG program while Avrora simulates
the program execution. Unlike static analysis, parts of a program that are not
79
executed during the simulation will not be accounted for. Also the generated
graph will accurately reect compiler optimizations. Hardware interrupts, which
occur intermittently during execution, are accounted for as well.
Delay analysis for each function is performed during CFG generation for prac-
ticality since basic block information may be too detailed. Function delay is
measured as the duration when execution enters a function to when execution
exits. Calls to other functions are accounted for, while interrupts are not. Since
execution delay may be inconsistent due to functions containing dierent execu-
tion paths, the average function delay is gathered from each execution instance.
However for our purposes, the functions that extract features from heartbeat
signals all consist of a single execution path. Therefore we lost no precision in
our analysis. The delays of feature detection modules for sampling rate of 360
sample/second are illustrated in Figure 4.6.
4.5 Target Architecture Model
Networked sensor nodes containing constrained, often battery-powered embedded
computers can densely sample phenomena that were previously dicult or costly
to observe. Sensor nodes can be placed anywhere on patients body [JEZ05]
[JDB05]. Due to the mobility of such systems, wireless sensor networks are ex-
pected to be both autonomous and long-lived, surviving environmental hardships
while conserving energy as much as possible.
It is well-known that the amount of energy consumed for a single wireless
communication of one bit can be many orders of magnitude greater than the
energy required for a single local computation [SHC04]. Thus, we focus on the
energy used for wireless communication. In our model, since all nodes are placed
80
within close proximity of each other, we assume they communicate directly and
multi-hop communication is not required. Therefore, the total energy consumed
for in-network communication is:
(n) = b(n) e(n) (4.1)
where b(n) is the number of packets transmitted and e(n) is the average
amount of energy required to transmit one packet. In our design, we consider
Collision Free Model (CFM) which simplies the programming by abstracting
out all of the details of low level channel contention and packet collision from
the algorithm designers. By abstracting reliable communication as an atomic
operation, programming based on CFM bears resemblance to existing algorithm
design in parallel and distributed computation. CFM does not really capture
the impact of packet collision that distinguishes wireless communication from
wired communication, which makes performance analysis under CFM not very
accurate. However, for the sake of simplicity, we consider CFM in our design.
4.6 Dynamic Reconguration
Sensor nodes are composed of embedded systems as well as general-purpose sys-
tems, introducing a tension between resource and energy constraints and the
layers of indirection required to support true general-purpose operating systems.
TinyOS [LMG04], the state-of-the-art sensor operating system, tends to priori-
tize embedded system constraints over general-purpose OS functionality. TinyOS
consists of a collection of software components written in the NesC language
[GLB03], ranging from low-level parts of the network stack to application-level
routing logic. However, our target operating system is SOS, a new operating sys-
81
tem for mote-class sensor nodes that takes a more dynamic point on the design
spectrum [HKS05]. SOS consists of dynamically-loaded modules and a common
kernel, which implements messaging, dynamic memory, and module loading and
unloading, among other services. Dynamic recongurability is one of our primary
assumptions. In the domain of embedded computing, recongurability is the abil-
ity to modify the software on individual nodes of a network after the network has
been deployed and initialized. This provides the ability to incrementally update
the sensor network after it is deployed, add new software modules to a sensor
node after deployment, and remove unused software modules when they are no
longer needed. The growing tensions between large hard to update networks
and complex applications with incremental patches has made recongurability
an issue that can no longer be ignored. SOS supports a mechanism that enables
over the air reprogramming of the sensor nodes. Using this methods, software
modules may be modied, added or removed.
4.7 Feature Set Partitioning
A hypergraph is a generalization of a graph, where the set of edges is replaced
by a set of hyperedges. A hyperedge extends the notion of an edge by allowing
more than two vertices to be connected by a hyperedge. Formally, a hypergraph
H = (V, E
h
) is dened as a set of vertices V and a set of hyperedges E
h
, where
each hyperedge is a subset of the vertex set V [DD96], and the size a hyperedge
is the cardinality of this subset. Let w
i
denote the weight of vertex v
i
V . A
K-way vertex partition = {V
1
, V
2
, . . . , V
k
} of H is said to be balanced with an
overall load imbalance tolerance 1 if each V
i
satises the following equation:
82
W
k
W
avg
(1 +), for k = 1, 2, . . . , K (4.2)
where,
W
k
=
v
i
V
k
w
i
(4.3)
W
avg
= (
v
i
V
w
i
)/K (4.4)
In a partition of H, a hyperedge that has at least one vertex in a partition is
said to connect that partition. Connectivity set
j
of a hyperedge e
j
is dened
as the set of partitions connected by e
j
. Connectivity
j
= |
j
| of a hyperedge
e
j
denotes the number of partitions connected by e
j
. A hyperedge h
j
is said
to be cut (external) if it connects more than one partitions (i.e.
j
> 1), and
uncut (internal) otherwise (i.e.
j
= 1). Therefore, the denition of cut-size is
as follows:
cutsize() =
e
j
E
h
(
j
1) (4.5)
Hence, the cutsize is equal to the number of cut nets. The hypergraph parti-
tioning is dened as dividing it into two or more parts such that the cutsize is min-
imized, while a given balance criterion among the partition weights is achieved.
The hypergraph partitioning problem is known to be NP-hard [GJ79b].
During the software partitioning, it is quite important to be able to divide the
system specication into clusters so that the inter-cluster (inter-mote) connec-
tions are minimized. Hypergraphs can be used to naturally represent feature
extraction algorithms. The vertices of the hypergraph are modeled as features,
their weights represent the computational time required for features detection,
83
and the hyperedges resemble the number of times a set of features is triggered
simultaneously. Partitioning the graph such that the cut-size is minimized while
the partitions are balanced can reduce the communication that is required among
various processing units for classication phase. The vision is that all features
selected must be classied at a local node, thus, in the events where selected fea-
tures reside on distributed nodes, inter-node communication is inevitable. A high
quality hypergraph partitioning algorithm greatly aects the feasibility, quality,
and the cost of the resulting system.
We employed a hypergraph partitioning algorithm that is based on the multi-
level paradigm. In the multilevel paradigm, a sequence of successively coarser
hypergraphs is constructed. A bisection of the smallest hypergraph is computed
and used to obtain a bisection of the original hypergraph by successively pro-
jecting and rening the bisection to the next level ner hypergraph. We have
used hMETIS, a program for partitioning hypergraphs implemented for PCs
[KAK97b]. Same algorithm can be easily ported on a mobile computer such
as a Pocket PC to facilitate dynamic reconguration. The vision is that the
hypergraph information is collected real-time from the processing nodes of the
wearable computer. Subsequently, the algorithm running on the motes are re-
congured. The number of partitions is determined as described below:
The preprocessing tasks as well as pattern recognition must be completed before
the next heartbeat arrives. Let the heartbeat be N beats per minute (bpm).
Therefore, the heartbeat rate period can be obtained from:
T
heartbeat
= 60/N (4.6)
Let the time required for preprocessing and pattern recognition be t
pre
and
t
recog
respectively.
84
t
pre
+t
recog
< (T
heartbeat
) where < 1 (4.7)
The factor is selected to be 0.9 to ensure a margin that prevents from
overloading the processing units. Therefore, the maximum CPU time that may
be assigned to pattern recognition is (T
heartbeat
) t
pre
where t
pre
is xed and
can be computed from proling stage. As described earlier, the weight on vertices
represents the required computational time for each feature. In addition, W
k
is
already outlined in Equation 4.4. Therefore, the following objective should be
accommodated:
Minimize K
s.t.
W
k
< t
recog
k = 1..K
(4.8)
To determine the value of K, we consider the total time required for pattern
recognition on all features , T
recog
, (extracted from proling analysis). It is trivial
that the lowerbound on K can be obtained from the following Equation:
K = T
recog
/t
recog
(4.9)
Once partitioning is performed based on the value of K, the solution may be
imbalanced and violates the constraint described in Equation 4.8. In this case,
K must be incremented and the features are re-partitioned until a feasible solution
is determined.
85
4.8 Simulation Analysis
This section presents various simulation analysis performed to exhibit the eec-
tiveness of our technique. All experiments were carried out with ECG signals
from MIT-BIH Arrhythmia database. The MIT-BIH Arrhythmia database con-
tains 48 half-hour excerpts of two-channel ambulatory ECG recordings, obtained
from 47 subjects studied by the BIH Arrhythmia Laboratory between 1975 and
1979. The recordings were digitized at 360 samples per second per channel with
11-bit resolution over a 10 mV range. We used all 48 complete records freely
available from PhysioNet [phya]. We also repeated the experiments by downsam-
pling all the benchmarks to 200 and 100 samples per second. As illustrated in
Table 4.2, each MIT-BIH record has the recordings of two channels. Yet, we only
used the rst channel. The second channel was not used for the sake of simplic-
ity. Originally, in MIT/BIH benchmarks, the electrodes placed on the chest were
selected due to their small noise level.
We performed proling analysis on the algorithm described in Section 4.3
using Avrora to compute the computational delay of feature detection modules.
The ECG algorithm was ported both for dot-motes (SOS) and PCs. The algo-
rithm for PC was written in C language. The simulation for feature and hyper-
graph extraction was done on PC due to a number of software instability that
we encountered in SOS. As for hypergraph partitioning, we utilized hMETIS.
The MIT/BIH benchmarks were used with three sampling rates as illustrated in
Tables 4.3, 4.4 and 4.5. The original sampling rate was 360 samples/sec while
200 and 100 samples/sec were acquired by downsampling the data. In Tables
4.3, 4.4 and 4.5, two scenarios for conguration was considered. In one scenario,
features were adaptively assigned to processing units based on hypergraph parti-
86
tioning. In the other scenario, the optimized conguration was determined using
hypergraph partitioning on benchmark 100 and remained xed throughout our
experiments (xed conguration). The number of partitions were obtained from
Equation 4.9 for each benchmark. Table 4.3 gures the number of queries ex-
changed in both scenarios. Considering that the experiments were carried out
through simulations, we were unable to measure the wireless power consumption.
However, given the number of features we examined - twenty three, each query
may be incorporated in a wireless packet of dot-motes (30 bytes). Therefore, tak-
ing into account Equation 4.1, the wireless power consumption is proportional
to the number of queries exchanged. On average, the communication energy
consumption was reduced approximately by 70% in all set of experiments. The
wireless communication overhead for partitioning was negligible due to the small
size, sparsity and slowly changing nature of our hypergraphs. The recongura-
tion was performed only once for each benchmark. Therefore, its eect on the
performance of the system was negligible.
4.9 Conclusion
We proposed a technique for software partitioning in tiny embedded systems.
Our target application was an ECG analysis algorithm which is generally clas-
sied as a complex medical application. We addressed the problem of mapping
such application onto resource constrained embedded systems while extending
the lifetime of the system. This was achieved by reducing the energy consump-
tion due to wireless communications. We demonstrated the eectiveness of our
technique on various ECG excerpts form MIT/BIH benchmarks. On average, the
energy consumption rate was reduced by 70%.
87
Figure 4.1: ECG analysis schematic
88
Figure 4.2: Automatic ECG segmentation performed on ltered signal
Figure 4.3: The sampling intervals of ECG morphologies 1 and 2. Morphology 1
consists of samples extracted from QRS onset and oset. Morphology 2 consists
of samples from S-wave oset and T-wave oset.
89
Figure 4.4: The sampling intervals of ECG morphologies 3 and 4. Morphology
3 consists of samples between 50 ms before and after the ducial point (FP).
Morphology 4 consists of samples between 150 ms and 500ms after the ducial
point.
90
Figure 4.5: Graph generated from proling analysis containing only major blocks
required for feature extraction.
91
Figure 4.6: Delays corresponding to ECG feature detection modules extracted in
proling phase
92
Record # Channel used Patient Medication
100 MLII male, age 69 Aldomet, Inderal
101 MLII female, age 75 Diapres
102 V5 female, age 84 Digoxin
103 MLII male, age not recorded Diapres, Xyloprim
104 V5 female, age 66 Digoxin, Pronestyl
105 MLII female, age 73 Digoxin, Nitropaste, Pronestyl
106 MLII female, age 24 Inderal
107 MLII male, age 63 Digoxin
108 MLII female, age 87 Digoxin, Quinaglute
109 MLII male, age 64 Quinidine
111 MLII female, age 47 Digoxin, Lasix
112 MLII male, age 54 Digoxin, Pronestyl
113 MLII female, age 24 None
114 V5 female, age 72 Digoxin
116 MLII male, age 68 None
118 MLII male, age 69 Digoxin, Norpace
119 MLII female, age 51 Pronestyl
121 MLII female, age 83 Digoxin, Isordil, Nitropaste
122 MLII male, age 51 Digoxin, Lasix, Pronestyl
123 MLII female, age 63 Digoxin, Inderal
124 MLII male, age 77 Digoxin, Isordil, Quinidine
200 MLII male, age 64 Digoxin, Quinidine
201 MLII male, age 68 Digoxin, Hydrochlorthiazide, Inderal, KCl
202 MLII male, age 68 Digoxin, Hydrochlorthiazide, Inderal, KCl
203 MLII male, age 43 Coumadin, Digoxin, Heparin, Hygroton, Lasix
205 MLII male, age 59 Digoxin, Quinaglute
207 MLII female, age 89 Digoxin, Quinaglute
209 MLII male, age 62 Aldomet, Hydrodiuril, Inderal
213 MLII male, age 61 Digoxin
214 MLII male, age 53 Digoxin, Dilantin
217 MLII male, age 65 Digoxin, Lasix, Quinidine
219 MLII male, age not recorded Digoxin
220 MLII female, age 87 Digoxin
221 MLII male, age 83 Hydrochlorthiazide, Lasix
222 MLII female, age 84 Digoxin, Quinidine
228 MLII female, age 80 Digoxin, Norpace
230 MLII male, age 32 Dilantin
232 MLII female, age 76 Aldomet, Inderal
233 MLII male, age 57 Dilantin
Table 4.2: Benchmark statistics
93
# of queries # of queries Transmission power
Record # exchanged with adaptive exchanged with xed saving with adaptive
partitioning conguration partitioning (%)
100 10 10 0.00
101 76 154 50.65
102 121 157 22.93
103 53 96 44.79
104 37 281 86.83
105 187 258 27.52
106 35 186 81.18
107 64 162 60.49
108 112 302 62.91
109 58 490 88.16
111 144 378 61.90
112 63 218 71.10
113 63 100 37.00
114 4 57 92.98
115 19 70 72.86
116 52 457 88.62
117 53 78 32.05
118 26 169 84.62
119 13 204 93.63
121 86 179 51.96
122 32 160 80.00
123 5 21 76.19
124 49 183 73.22
200 17 48 64.58
201 2 217 99.08
202 41 71 42.25
203 19 233 91.85
205 83 170 51.18
207 232 547 57.59
208 20 292 93.15
209 225 272 17.28
210 143 183 21.86
212 7 119 94.12
213 63 552 88.59
214 145 177 18.08
215 181 603 69.98
217 117 237 50.63
219 35 160 78.13
220 74 155 52.26
221 23 244 90.57
222 6 170 96.47
223 144 328 56.10
228 54 468 88.46
230 15 27 44.44
231 4 16 75.00
232 13 189 93.12
233 47 328 85.67
234 73 192 61.98
Average 65.52 216.00 69.67
Table 4.3: Number of queries exchanged among processing units : sampling rate
= 360 sample/sec
94
100 5 5 0.00
101 58 84 30.95
102 173 276 37.32
103 58 170 65.88
104 49 188 73.94
105 169 252 32.94
106 41 343 88.05
107 68 229 70.31
108 101 314 67.83
109 62 478 87.03
111 194 347 44.09
112 56 114 50.88
113 26 156 83.33
114 17 141 87.94
115 48 265 81.89
116 154 361 57.34
117 51 69 26.09
118 63 273 76.92
119 2 237 99.16
121 88 255 65.49
122 12 127 90.55
123 15 105 85.71
124 61 223 72.65
200 1 222 99.55
201 206 296 30.41
202 35 194 81.96
203 132 495 73.33
205 78 222 64.86
207 201 497 59.56
208 32 319 89.97
209 244 387 36.95
210 163 339 51.92
212 37 312 88.14
213 123 207 40.58
214 153 187 18.18
215 136 180 24.44
217 25 160 84.38
219 68 286 76.22
220 42 160 73.75
221 85 443 80.81
222 12 183 93.44
223 117 528 77.84
228 62 433 85.68
230 16 157 89.81
231 34 177 80.79
232 46 217 78.80
233 125 529 76.37
234 72 271 73.43
Average 79.50 258.60 69.26
= 200 sample/sec
95
100 11 11 0.00
101 71 152 53.29
102 101 267 62.17
103 45 86 47.67
104 15 186 91.94
105 196 594 67.00
106 48 296 83.78
107 78 181 56.91
108 47 188 75.00
109 30 414 92.75
111 94 354 73.45
112 31 242 87.19
113 38 149 74.50
114 43 120 64.17
115 21 70 70.00
116 15 284 94.72
117 25 103 75.73
118 28 185 84.86
119 26 169 84.62
121 70 257 72.76
122 37 407 90.91
123 13 136 90.44
124 60 147 59.18
200 14 181 92.27
201 128 250 48.80
202 45 231 80.52
203 104 272 61.76
205 85 109 22.02
207 311 557 44.17
208 25 253 90.12
209 227 623 63.56
210 168 291 42.27
212 107 365 70.68
213 51 598 91.47
214 113 183 38.25
215 300 532 43.61
217 6 174 96.55
219 29 374 92.25
220 58 95 38.95
221 12 242 95.04
222 11 98 88.78
223 154 331 53.47
228 217 494 56.07
230 71 287 75.26
231 14 185 92.43
232 24 301 92.03
233 139 425 67.29
234 61 178 65.73
Average 75.35 263.06 71.36
= 100 sample/sec
96
CHAPTER 5
Customizable Medical Monitoring Platform
5.1 Introduction
Computation, storage, and communication are now woven into the fabrics of
our society with much of the progress being due to the relentless march of the
silicon-based electronics technology as predicted by Moores Law. The emerging
technology of exible electronics, where electronics components such as transis-
tors and wires are built on a thin exible material, oers a similar opportunity
to weave computation, storage, and communication into the fabric of the very
clothing that we wear. The implications of seamlessly integrating a large num-
ber of communicating computation and storage resources, mated with sensors,
in close proximity to the human body are quite exciting. For example, one can
imagine biomedical applications where biometric and ambient sensors are woven
into the garment of a patient or a person in a hazardous environment to trigger or
modulate the delivery of a drug. The natural applications of these systems have
environmental dynamics, physical coupling, resource constraints, infrastructure
support, and robustness requirements that are distinct from those faced by tra-
ditional systems. This combination requires one to go beyond thinking of these
systems as traditional systems in a dierent exible form factor. Instead, a re-
thinking of the architecture and the design methodology for all layers of these
systems is required. An innovative system such as our proposed architecture for
97
biomedical monitoring requires that the spatially distributed components on its
surface be able to communicate and coordinate to achieve the desired sensing
and possibly actuation function. Sensing will take advantage of the spatial dis-
persion over the human body surface to improve detection eciency, eectiveness
and robustness. Such collaboration requires exible communication among the
elements. While the resource constraints imply that the communication must
be done in a highly ecient manner, the distributed control nature of the ap-
plications requires that the communication be highly predictable (latency) and
reliable. The ultimate goal of this research is to develop the communication, co-
ordination and control primitives with which a wide variety of applications will
be enabled, and to demonstrate their eectiveness compared to other approaches,
using a biomedical application as a motivator.
Ubiquitous and pervasive computing environments are typically researched and
developed by means of rapid prototyping. A common approach is to use o-
the-shelf networked embedded systems or to design and implement specialized
hardware, which is then post hoc attached to an existing object or into the envi-
ronment. Moreover, ubiquitous computing environments for health care are ex-
tremely dynamic systems with ever changing requirements and constantly rened
specications. Therefore, a prototype-based development cycle is most appropri-
ate to cope with new demands.
While utilizing ubiquitous components for biomedical applications is an exciting
approach, medical researchers face considerable hurdles if they wish to create
even simple application for health care monitoring. Most lack the necessary tech-
nical training. Even technical sta who are willing to learn will nd themselves
spending most of their time building and debugging circuit boards, rmware and
low-level protocols rather than building their medical monitoring systems. One
interesting approach to creating rapidly recongurable medical system is to build
98
a software-based GUI on a PC that interfaces with the individual components of
the system. The GUI is responsible for obtaining the application requirements
from the medical personnel and program various components of the system to
meet the applications needs. Such architecture allows early evaluation of ser-
vice ideas, prot opportunities, user aspects, and technological feasibility of the
applications. Besides, it generates an evolving pool of innovation prototypes for
future services and products. Moreover, it creates services that can be produc-
tied and forwarded to product development. One major concern in design of
such prototype is that it provides justication and prove for services. It also has
to work and allow reliable test settings.
5.2 System Description
Our proposed CustoMed is a fascinating and critical class of distributed embed-
ded systems. The main attribute of CustoMed is its rapid prototyping and fast
and easy customizability capability based on patients needs. Firstly, the cus-
tomization may be at the device level. The choice of the device, the placement of
the device on the body, and the interaction level of the device with the environ-
ment will be tailored to the individual and his/her needs. Secondly, the software
downloaded onto the devices can be customizable. Depending on the gender,
age, medical condition, and other variables, the software downloaded onto the
devices diers. The idea of customization has not been emphasized before, but is
an important concern, if the system is made to be robust enough to handle many
dierent needs, as well as unexpected needs that may arise.
Non engineering-sta can easily custom-built CustoMed for patients. The sys-
tem will be quickly assembled from basic parts and congured for use. The vision
is that in the doctors oce in about ve minutes, the appropriate devices and
99
Figure 5.1: System Components
the correct number of them will be assembled and axed to the patient. Other
system take months or even years to built, and hence lack the adaptability the
system we propose. The physician will also pick from a wide range of code to
download onto the devices. We developed such a tool which enables physicians/
developers to pick a specic variation of a code for a particular application and
download it to the system components of CustoMed. This tools ensures the
abstract design of medical monitoring application and software programming in-
dependence. Furthermore, it works with the environment made available to the
patient. For example, in case of an emergency, the sensors can alert the security
system in the house. In less urgent case, an email can be sent across the Internet
or a home appliance can be turned o or on.
Figure 5.2 shows the Graphical User Interface of our system on a PC. This
GUI supports scripting such that the specications of the application may be eas-
100
Figure 5.2: CustoMed GUI on a Personal Computer
ily loaded onto the components. Some parameters that may be obtained from the
personnel includes the type of medical sensors employed, the frequency of data
sampling and information regarding sensor calibration. Students and researchers
who were not involved in the design of the supporting infrastructure built easily
some of these applications.
Using the scripting capabilities, we implemented a distributed bootstrap mecha-
nism that automated the initialization process. We dened conguration les
with information about the required system services, initial parameters, and
communication schemes. The bootstrap mechanism implements the following
functionality: automatically starts the services in the specied components and
registers the services in the base component. The bootstrap process made the
101
system portable. Every ubiquitous computing environment has its own bootstrap
conguration le.
Figure 5.3 illustrates the development cycle of the applications. The rapid pro-
totyping and fast customization capabilities enable the developers to easily eval-
uate inter-application behavior and test its performance on various subjects. By
observing this, developers can incrementally improve the reliability and the e-
ciency of the system very fast.
Figure 5.3: Application Development Cycle
The main component of our system is called med nodes which incorpo-
rate sensing, processing and communications (both wired and wireless). The
processors of med nodes support variety of analog and digital sensors and are
programmable to process the collected data from sensors. Figure 5.1 illustrates
our CustoMed architecture connected to the Internet through a Pocket PC.
The list of supported services by our system which facilitates rapid prototyp-
ing and customization are as follows:
102
Distributed med node management: All med nodes directly communicate
with a base station med node which is connected to a Pocket PC. Since
they all share the same communication medium, a token cycling algorithm
is implemented on the base station. Possession of the token grants the right
to transmit. If a node receiving the token has no information to send, it
passes the token to the next end station. Each station can hold the token
for a maximum period of time which is specied by the developers.
Med node presence detection: Base station med node detects active med
nodes based on their recent activities and accordingly adapt the token cy-
cling algorithm.
Distributed med node database: The developer can choose whether sensor
readings must be stored locally and some specic trigger points will be
transmitted to the Pocket PC or all the sensor readings will be directed
to the Pocket PC. Moreover, the architecture supports inter med nodes
query processing. This feature specically assist with data correction in
med nodes considering high correlations among data of sensor readings.
Notication service: to disseminate information about the state of the ubiq-
uitous computing environment (i.e. radio channel characteristics). This
service is based on events and event channels that decouple the information
senders from the information receivers.
5.3 Related Work
103
Georgia Tech [PMJ02]. The Wearable Motherboard is woven into an undershirt
in the Smart Shirt design. Their interconnect is a exible data bus that can
support a wide array of sensory devices. These sensors can then communicate
via the data bus to a monitoring device located at the base of the shirt. The
monitoring device is integrated into a single processing unit that also contains a
transceiver. The SmartShirt design features plastic optical ber that can be used
to detect punctures - however, they do not have any means of dealing with these
punctures other than reporting them via the transceiver. In the event that their
single processing unit (and transceiver) is damaged or the control lines leading
to the device are torn, the Smart Shirt is virtually inoperable. In our design, we
distribute control to multiple processing elements and can accommodate multiple
communication buttons. This ensures that a fault in our system will not result
in total system failure. This fault tolerance is essential in the demanding and/or
hazardous environments targeted by this research. Fireghters, policemen, sol-
diers, astronauts, athletes, and others working in hazardous environments need
robust material that can sustain damage and yet still reliably provide service.
Moreover, our design supports an alert system which could provide immediate
medical attention to individuals in emergency medical situations. Several other
technologies have been introduced by MIT called MIThril [DMP03], e-Textile
from Carnegie Mellon University [MMK02] and Wearable e-Textile from Virginia
Tech [MJE03]. Furthermore, the Lifeguard project being conducted at Stanford
University is a physiological monitoring system comprised of physiological sensors
(ECG/Respiration electrodes, Pulse Oximeter, Blood Pressure Monitor, Temper-
ature probe), a wearable device with built-in accelerometers (CPOD), and a base
station (Pocket PC). The CPOD acquires and logs the physiological parame-
104
ters measured by the sensors [lif]. The Assisted Cognition Project conducted
at the University of Washingtons Department of Computer Science is exploring
the use of AI systems to support and enhance the independence and quality
of life of Alzheimers patients. Assisted Cognition systems use ubiquitous com-
puting and articial intelligence technology to replace some of the memory and
problem-solving abilities that have been lost by an Alzheimers patient [KEF03].
Nevertheless, none of the above projects/systems supports the concept of being
custom-built in minutes and reconguration in the extent that CustoMed does.
5.4 Design of CustoMed
CustoMed is composed of the following devices:
5.4.1 Sensors
We employ various types of sensors for continual physiological measurements as
well as environmental measurements to identify wearers physiological conditions
and disorders and the case where people operating in hazardous environments.
We have available to us pressure sensors, galvanic skin response sensors, ex
sensors, piezoelectric lm sensors, temperature sensors and accelerometers.
Flex Sensors: The ex sensor changes resistance when bent. It will only
change resistance in one direction. They can be used to measure the angular
motion of various parts of body such as knees and neck.
Pressure Sensors: These sensors are ideal for measuring forces without dis-
turbing the dynamics of a test. They can be used to measure both static
and dynamic forces. They are thin enough to enable non-intrusive mea-
105
surement. The resistance of the sensor decreases as force is applied. Mea-
surement of pressure under the foot is of particular interest to the clinician
and researcher, particularly in the elds of diabetes, orthopaedics, sports
science and rheumatology. The pressure sensors could be placed in shoes
to perform such measurements.
Piezoelectric Film Sensors: Piezoelectric thin lm sensors generate analog
voltage signals in response to applied dynamic forces. They can be used to
monitor if a patient being attacked or abused by other patients.
Galvanic Skin Response (GSR) Sensors: The galvanic skin response (GSR)
also referred to as the electrodermal response (EDR), measures electrical
skin conductance from the ngers or palms that is associated with sweat
gland activity. It is commonly used in psychophysiology experiments to
infer emotional state and cortical arousal. The GSR is commonly used in
Biofeedback experiments.
Temperature Sensors: The polymer based bulk heterojunction solar cells
have demonstrated the ability to act as temperature sensor. These high per-
formance photodetectors combined with organic-thin lm transistors can be
used to realize sensors that can be fabricated on exible substrates includ-
ing fabric. Such devices possess a combination of electrical performance
with superior properties such as light weight, low cost of fabrication, and
exibility [CXY04].
Accelerometers: Such sensors can be used for orientation and body move-
106
ment tracking. The idea is to track the orientation of each segment of the
body by a sensor to determine the body posture for medical monitoring.
5.4.2 Med Nodes
The most important component of our system is the med nodes. Med nodes are
stand-alone components equipped with processing units and batteries. They sup-
port various types of sensors for physiological reading from human body. These
blocks enable the system to be exible, however their basic structure remains
xed almost all the time and thus the reconguration time is no longer a severe
limitation. Furthermore, customization of such system with a large number of
med nodes is extremely fast. The med nodes can possess several parameters such
that they can be complex enough to suit a range of applications. They can as well
be more basic that can perform multiple operations. Furthermore, they support
variety of analog and digital sensors such as ex sensors, piezoelectric sensors,
pressure sensors, etc. Also the block is left software programmable which can be
customized for various applications and sensors. On-chip memory blocks are also
available for data storage. The processors of med nodes are dot-motes developed
at University of California, Berkeley and manufactured by Crossbow [xbo]. A
med node along with a few sensors is shown in Figure 5.4.
5.4.3 Pocket PC
A pocket PC is responsible for collecting data from med nodes and classifying
them. It dispatches the critical events detected by med nodes or the pocket
PC, itself, to the Internet. Moreover, it coordinates and controls the overall
functionality of the system. Pocket PC is also equipped with Global Positioning
System (GPS) that reports the location of the patient to the Hospital in case of
107
Figure 5.4: A Med Node along with Sensors
emergency.
5.4.4 Server
Pocket PC transmits the real-time data to an external server through the Internet.
The server provides with various monitoring and data logging capabilities. A
screen-shot of the server program is shown in Figure 5.5.
Figure 5.6 depicts the system hierarchy of CustoMed along with the commu-
nication scheme employed in various levels.
5.5 Driver Application
The design and development of the system framework is meant to be versatile
enough to be applied to many dierent medical applications. The requirements
of the system, specically with the ability to quickly and easily make a custom-
made system per patient, were deeply aected by this vision. The following is
a sample application that have inspired the overall system requirements and the
108
Figure 5.5: Data Monitoring on a Remote Server
initial idea.
Post-knee surgery tracking of patients:
Structuring of imaging data from the musculoskeletal system will be accom-
plished utilizing novel MR pulse sequences to support canonical imaging
values from tissue independent of acquisition techniques. Two med nodes
are placed on knees and two on the ankles. The med nodes on the knees are
equipped with ex sensors while med nodes on the ankles utilizes pressure
sensors. Med nodes on the knees are responsible for tracking the angular
motion of the knees and transmitting the collected data to the pocket PC.
Med nodes on the ankles, however, measure the forces and pressures under
the foot as well as the load distribution. In addition, we place another med
node on the neck to measure the angular motion of the back/neck.
Used to aid Alzheimers patients:
109
Figure 5.6: System Hierarchy
The rise of Alzheimers disease is one of the greatest health crises facing
the industrialized world. Today, approximately four million Americans suf-
fer from Alzheimers disease; by 2050, the number is expected to rise to
15 million people. CustoMed can make huge dierences in the quality of
life of Alzheimers patients. It would be appropriate of the walking well
Alzheimers patients, who are patients that are fully mobile. Galvanic skin
response sensors, which detect arousal and/or agitation by measuring skin
conductance, can be placed on the patients body. Perhaps they can be
placed in their socks or in the nap of theirs neck, two places, where there is
large amount of perspiration. When agitation is sensed, proper verbal cues
can be issues to the patient, to reorient them or calm them down. Also, in
case of emergency CustoMed can email physicians and/or family members.
CustoMed can extend the time span when patients are independent, and
hence improves the quality of their life and reduce the cost incurred. Cus-
110
toMed can prove to be very useful in assisted living homes, also, where the
ratio of the sta to patients is about one to ten. In such situations, being
hit or abuse has been a problem. Pressure sensors placed on the patients
can determine a hit, but even more importantly the detection of agitation,
before any harm has been done can be used to protect the patients and the
sta. Perhaps, during times of agitation the stas pagers can be cued.
5.6 Sensor Calibration
Sensor calibration is the process of forcing a sensor to conform to a given in-
put/output mapping. This is often done by adjusting the sensor internally but
can equivalently be done externally by passing the sensors output through a cali-
bration function that maps the actual sensor response to the desired response. We
observe each sensors behavior in a controlled environment and map its observed
output to the desired output by a mapping function. The developer download
such function to the appropriate med node during the implementation.
In this section, we illustrate system specications of CustoMed along with the
experimental analysis that we performed to evaluate the performance of the sys-
tem. The processing units in med nodes are mica2dot motes developed at the
University of California, Berkeley. The code on mica2dots is implemented with
NesC under TinyOS [HSW00]. The software on Pocket PC is written with Em-
bedded Visual C++ while the tools on PC are written with Visual C++. A
summary on the system specications of the CustoMed is presented in Table 5.1.
111
All the experiments were carried with med nodes transceivers operating at
Max Sampling Rate 200Hz
Bits per Sample 10bit
Max Number of Sensors per Med Node 5
Power Consumption of each Med Node 54 mW
(Wireless ON, CPU active)
Battery Lifetime of each Med Node 34 hours
Table 5.1: CustoMed System Specication
433MHz. All med nodes were equipped with one 3V Duracell DL2450 batteries
except the base station med node placed at location 5 (with three batteries) since
it is connected to the Pocket PC and must remain on continuously.
Figure 5.7: Med Nodes Placement on Body
We considered various systems with dierent combinations of med nodes/sensors
on the body as depicted in Figure 5.7. Base station med node 5 was responsi-
ble for collecting data from other med nodes and transmitting it to the Pocket
112
PC through UART (Universal Asynchronous Receiver/Transmitter). Overall, 8
congurations of med nodes were implemented and was tested on 3 subjects in
5 dierent environments. The loss rate in wireless communication on average
was measured less than 4% which exhibits the reliability of the system. As for
wearability, all test subjects indicated that the system is comfortable. Due to the
small size of the med nodes and its compact design, the system did not greatly
interfere with wearers daily activities. Figure 5.8 shows a system conguration
of CustoMed with four med nodes (two with piezoelectric sensors, one with ex
and one with pressure sensor). Figure 5.9 depicts the base station med node
connected to the Pocket PC.
In regards to the channel characterization, we performed various experiments
with CustoMed. The antenna utilized was a standard mica2dots 17.3cm solid
copper antenna. One major consideration on med nodes is their wearability fac-
tor. Therefore, we made the antenna circular since we could easily t it in the case
of med nodes. We conducted our experiments with transmission power setting in
med nodes set to 0dBm. The transmission power setting can be easily changed
in dot-motes through ChipCon CC1000 chip. In these experiments, we trans-
mitted TinyOS [HSW00] default packets and measured the reliability of channel
between every communicating pair. Tables 5.2 and 5.3 illustrate the loss rate in
terms of percentage along with packet round-trip time delay (RTT) in terms of
milliseconds between every pair of med nodes for two dierent walking speeds as
well as indoor and outdoor environments. As shown, the channel characteristics
(delay and packet round-trip delay) is highly dynamic due to the mobility and
the close proximity of the med nodes to the body.
113
RTT(ms)/Loss Rate(%) 1 2 3 4 5
1 - 62.1 / 16.9 82.7 / 12.4 62.5 / 10.73 59.9 / 5.6
2 62.1 / 16.9 - 63.6 / 4.0 53.0 / 2.1 74.6 / 0
3 82.7 / 12.4 63.6 / 4.0 - 72.9 / 4.9 135.6 / 3.6
4 62.5 / 10.7 53.0 / 2.1 72.9 / 4.9 - 94.4 / 0
5 59.9 / 5.6 74.6 / 0 135.6 / 3.6 94.4 / 0 -
Table 5.2: Wireless Channel Characteristics during Indoor Slow Walk
RTT(ms)/Loss Rate(%) 1 2 3 4 5
1 - 52.9 / 11.5 75.7 / 14.4 62.2 / 9.3 53.4 / 10.1
2 52.9 / 11.5 - 54.3 / 15.2 65.7 / 7.7 86.3 / 6.3
3 75.7 / 14.4 54.3 / 15.2 - 75.8 / 2.6 106.8 / 2.6
4 62.2 / 9.3 65.7 / 7.7 75.8 / 2.6 - 99.3 / 0
5 53.4 / 10.1 86.3 / 6.3 106.8 / 2.6 99.3 / 0 -
Table 5.3: Wireless Channel Characteristics during Outdoor Normal Walk
5.8 Conclusions
In this paper, we proposed a system architecture which enables rapid prototyp-
ing and quick customization. We tested our proposed architecture under various
conditions and with dierent applications.
Dynamic power management techniques such as shutdown have emerged as pow-
erful methods for power-aware computing systems including signal-processing
systems. New system-wide power-aware operation is critically important in the
system envisioned in this paper. This requires distributed versions of power man-
agement techniques, which consider communication and sensing aspects as well
as computation. Privacy and security is another major concern. Such technology
become less acceptable if the privacy of individuals is compromised. The dangers
do not stop when the data is transmitted. Indeed, CustoMed may be invasive,
delivering drugs and life-saving debrillator shocks to the patients as needed. Of
course, for both passive and active CustoMeds, security is a big concern and a
technological obstacle to large-scale deployment. The data must be safely and
wirelessly transmitted among the med nodes and to the hospital. This involves
encryption, error-correction and authentication protocols, where all of these must
114
be suciently light-weight in the sense of computational resources, battery needs
and transmission cost to make it realizable for the CustoMed. Furthermore, the
CustoMed should be able to execute light-weight wireless identication scheme
with the hospital/base station. Such an identication scheme should be robust in
the sense that no adversary can pretend to be a hospital base-station to a Cus-
toMed and at the same time an adversary should not be able to masquerade
as a legitimate CustoMed and send erroneous or out-dated data.
115
Figure 5.8: CustoMed
116
Figure 5.9: Base Station Med Node and a Pocket PC
117
Figure 5.10: Sensor Readings during Indoor Slow Walk
118
CHAPTER 6
CMAS: Clinical Movement Assessment System
for Neuromotor Disorders
6.1 Introduction
Emerging technologies are generating opportunities for improving medicine and
public health. We are currently focusing on using recent advances in wireless
networks and real-time reconguration to improve diagnosis and treatment of
neuromotor impairments.
Strokes, spinal cord injuries, traumatic orthopedic injuries, and other disor-
ders commonly lead to a loss of motor functions of the upper limbs. Currently,
clinical assessments of motor functions are routinely scored subjectively or using
ordinal scales [1-3]. New treatment and rehabilitation strategies are creating a
need for more accurate and quantitative measures of motor function that can be
made readily and accurately [4, 5]. Such Standardized, quantitative functional
measurements would contribute to understanding neuromotor injuries. Moreover,
quantitative measurements of motor function are the rst step in developing ap-
propriate population databases to enable rapid, patient-specic assessments of
motor function.
The architecture we propose is a mobile and low prole system that may be
used in home settings and can potentially assist stroke survivors with several
119
therapy exercises.
6.2 Related work
Many studies have characterized individual nger function and coordination in
both unimpaired and impaired populations [6-12]. Quantitative measurements
of maximum force, time course of force onset and oset, and nger coordination
yield valuable information about neuromotor function following injury, and could
be very helpful in describing the characteristics of use-dependent plasticity during
recovery. Although most current eorts are primarily limited to a research setting,
there is growing interest in using quantitative measurements of function as clinical
tools [13-17].
6.3 Hardware design
Our CMAS is based on a system architecture called CustoMed (Customizable
Medical Monitoring device) developed at UCLA [JEZ05]. CustoMed is a plat-
form for health monitoring using wireless sensor networks (WSN). This archi-
tecture is a network enabled system that supports various wearable sensors and
contains on-board general computing capabilities for executing individually tai-
lored event detection, alerts, and network communication with various medical
informatics services. Our platform is composed of sensors, med nodes and a
pocket PC. The most important component of our system, the med node, is a
stand-alone component consisting of a processing unit and a battery which can
support various types of sensors for physiological reading from human body. The
customization of such system with a large number of med nodes is extremely
fast even by non-engineering sta. In addition to support of various wearable
120
sensors, the software downloaded onto the devices can be customizable over-the-
air. Depending on sex, age, medical condition, and other variables, the software
downloaded onto the devices diers. The idea of customization has not been em-
phasized before, but is an important concern, if the system is made to be robust
enough to handle many dierent needs, as well as unexpected needs that may
arise.
6.3.1 Sensors
CustoMed can interface with various types of sensors, such as pressure, galvanic
skin response, ex, piezo-electric lm and temperature sensors. These sensors
can provide continual physiological measurements as well as environmental mea-
surements, even when people operating in hazardous environments. In our rst
CMAS prototype, we utilize pressure sensors to measure the force applied by
individual hand digits to a handgrip. Pressure sensors are ideal for measuring
forces without disturbing the dynamics of a test. They can be used to measure
both static and dynamic forces. They are thin enough to enable non-intrusive
measurement. The resistance of the sensor decreases as force is applied.
6.3.2 Handgrip device
The CMAS handgrip device is capable of measuring forces generated by indi-
vidual ngers using pressure sensors. The resistance against squeeze force in our
handgrip device is adjustable. A picture of the handgrip device along with sensors
and a med-node is shown in Figure 6.1.
121
Figure 6.1: Handgrip device
6.3.3 Med node
Med Nodes enable the system to be exible. Although their basic structure can
remain xed most of the time, customization of the system with a large number
of med nodes is extremely fast. The med nodes can be complex enough to
suit a range of applications, such as tracking knee motion after knee surgery to
aiding Alzheimer patients residing in assisted living homes by detecting arousal
and/or agitation (by measuring skin conductance). Furthermore, they support a
variety of analog and digital sensors. The block is software programmable and
can be customized for various applications and sensors. On-chip memory blocks
are also available for data storage. The processors of med nodes are dot-motes
developed at the University of California, Berkeley. A med node along with a
handgrip and pressure sensors is shown in Figure 6.1.
122
6.3.4 Pocket PC
A pocket PC is responsible for collecting data from med nodes and classifying
them. It dispatches the critical events detected by med nodes or the pocket PC,
itself, to the Internet. Moreover, it coordinates and controls the overall function-
ality of the system. It can as well perform resource management to accommodate
several objectives such as optimizing the power or enhancing the fault-tolerance.
In many medical monitoring applications, pocket PCs can facilitate interactions
with patients with their large displays while the mobility is not sacriced.
6.4 Software design
6.4.1 Med node
The software for med nodes has been developed in C and under SOS, a new
operating system for mote-class sensor nodes that takes a more dynamic point
on the design spectrum [19]. SOS consists of dynamically-loaded modules and
a common kernel, which implements messaging, dynamic memory, and module
loading and unloading, among other services. Dynamic recongurability is one of
our primary objectives. In the domain of embedded computing, recongurability
is the ability to modify the software on individual nodes of a network after the
system has been deployed and initialized. This provides the ability to incremen-
tally update and add new software modules to med nodes after deployment, and
remove unused software modules when they are no longer needed. This in par-
ticular is benecial when the medical sta ought to remotely update the software
that runs on med nodes. The ability to recongure the software on med nodes
may potentially enhance the power consumption of the system (and the lifetime)
by limiting the wireless transmission to only the most valuable data. In tiny
123
embedded systems that we utilize Berkeley dot-motes the most power hungry
component is the radio module. In addition, the local processing in med-nodes
may be dynamically altered based on individuals need.
6.4.2 Pocket PC
The software on Pocket PC provides both serial connection with a gateway med
node and Internet connection with an external server. The existing radio com-
munication protocol on med-nodes is incompatible with the wireless receiver on
Pocket PCs. Therefore, we utilize a base-station med node that collects informa-
tion from all med-nodes and sends to the pocket PC through the serial connection.
The base station med-node also facilitates the software reconguration (in SOS
the new software modules that is installed must be dispatched from a base sta-
tion). The Internet connection with an external server may be utilized for remote
monitoring. As shown in Figure 6.2, in the rst window of CMAS terminal on
pocket PC, the user species the proper port conguration for both serial and
Internet connections. When the external server is not available, the user may
choose to bypass the Internet connectivity.
For the particular handgrip device we utilize, four pressure sensors are em-
ployed. Hence, it is imperative that patients hand digits are placed properly
on individual sensors. In addition, due to the variable physical capabilities of
patients, the applied force must be normalized. Hence, the screen shot shown in
Figure 6.3 is designed to assess the maximum voluntary contraction (MVC) and
ensure that the ngers are properly placed on the sensors.
Following the calibration process, the main window of CMAS terminal is
displayed as shown in Figure 6.4. The top window features a guide or a target
cursor that moves on a sinusoidal wave (Fig. 6.4, black dot). The patients
124
Figure 6.2: Connection setup window
response (i.e. the average forces applied to the pressure sensors) is displayed
in the same window (Fig. 6.4, blue dot bounded by horizontal lines). The
objective is to have the patient follow the guide by applying appropriate force to
the handgrip device. The force may be scaled to 80%, 60%, 40% or 20% of the
MVC force. Moreover, forces applied by individual ngers are illustrated on the
CMAS terminal in four separate windows. Finally, the patient can choose from
several test sets which will be described in the next section.
6.5 Test sets
Four test sets have been developed for the CMAS:
Test 1: The guide or the target travels on a sinusoidal waveform between
125
Figure 6.3: Calibration window
zero and 100% of the patients maximum strength. The period of the sine
wave six seconds.
Test 2: The objective in test 2 is to evaluate the fatigability in patients.
The target moves between 20% and 40% of the MVC with the period of
200ms. The patient is expected to follow the indicator by rapidly squeezing
the handgrip device.
Test 3: This test is developed for evaluating the ability to nely modulate
force production. The guide moves between 15% and 30% of the MVC on
a sinusoidal wave with period of six seconds.
Test 4: This test is the high strength ne grain motor control. This test is
similar to test 3 except that the guide moves between 45% and 65% of the
126
Figure 6.4: CMAS task window
MVC.
6.6 Experimental results
Preliminary experiments conducted on unimpaired subjects (N=2) showed that
subjects were able to nely modulate nger forces to achieve a desired average
force (Figure 6.5). The variance in errors between average force and target forces
were 6.4 1.0%, 0.9 5%, 38 0.3% and 0.3 0.3% of the guide variance
for Task 1, 2, 3 and 4, respectively for both subjects over 6 trials. Subjects
precisely matched the guide frequency (frequency dierences < 0.005 Hz for all
trials). Crosscorrelating the guide signal with the average force signal revealed
that average forces from subjects lagged the guide forces by 80 100 ms and 300
127
0 ms for tasks 2 and 3. For tasks 1 and 4, the time mean time lags of -50 ms
and -21 ms fell well within the variance 170 and 470 ms, respectively.
Figure 6.5: Average and individual forces during Task 1
Whereas subjects were able to precisely track the guide signals, they did not
accomplish this by generating comparable forces with each nger (Figure 6.5).
Large forces produced by middle ngers compensated for small forces generated
by index and little ngers, reecting an eective force synergy that maintained
target average forces. The variance in errors between each individual force and
the target forces is shown for every task and subject in Table 6.1. In addition,
in frequency domain, the dierence of the peak in the power spectrum between
each individual force and the target forces is shown. Please note that each data
point in Table 6.1 is the average taken over 3 trials.
128
T
a
s
k
S
u
b
j
e
c
t
V
a
r
i
a
n
c
e
i
n
e
r
r
o
r
s
w
i
t
h
r
e
s
p
e
c
t
t
o
t
h
e
t
a
r
g
e
t
f
o
r
c
e
(
%
)
F
r
e
q
u
e
n
c
y
d
i
e
r
e
n
c
e
(
H
z
)
A
v
e
r
a
g
e
L
i
t
t
l
e
n
g
e
r
R
i
n
g
n
g
e
r
M
i
d
d
l
e
n
g
e
r
I
n
d
e
x
n
g
e
r
A
v
e
r
a
g
e
L
i
t
t
l
e
n
g
e
r
R
i
n
g
n
g
e
r
M
i
d
d
l
e
n
g
e
r
I
n
d
e
x
n
g
e
r
1
1
5
.
5
6
8
8
.
1
8
4
6
.
0
0
2
3
.
5
1
9
0
.
8
0
0
.
0
0
2
-
0
.
0
8
5
0
.
0
0
0
0
.
0
0
2
0
.
0
0
1
1
2
7
.
3
4
7
7
.
4
0
3
0
.
5
8
5
7
.
4
8
9
9
.
2
6
0
.
0
0
2
0
.
0
0
3
0
.
0
0
2
0
.
0
0
3
-
0
.
0
5
2
2
1
0
.
7
8
8
9
.
4
0
1
0
.
8
2
6
6
.
6
3
9
9
.
7
8
0
.
0
0
0
-
0
.
0
2
3
-
0
.
0
0
8
-
0
.
0
0
2
-
0
.
0
4
4
2
2
1
.
0
6
7
2
.
5
1
2
1
.
2
2
6
7
.
6
1
1
0
0
.
0
0
0
.
0
0
0
-
0
.
0
2
5
0
.
0
0
0
-
0
.
0
1
1
-
0
.
0
4
9
3
1
6
1
.
1
5
9
3
.
5
5
4
4
.
3
9
5
4
.
8
4
9
8
.
3
4
0
.
3
6
0
-
0
.
3
6
8
0
.
3
5
6
0
.
3
6
1
-
0
.
0
3
7
3
2
1
5
.
8
9
7
8
.
4
4
3
3
.
1
8
5
5
.
2
3
1
0
0
.
0
0
0
.
0
0
3
-
0
.
2
0
4
0
.
0
0
3
0
.
0
0
2
-
0
.
6
8
3
4
1
0
.
4
4
8
8
.
7
4
3
2
.
9
5
4
0
.
5
6
8
6
.
1
1
-
0
.
0
0
1
-
0
.
0
2
2
-
0
.
0
3
1
-
0
.
0
2
2
-
0
.
0
2
9
4
2
0
.
1
5
7
7
.
2
9
2
9
.
6
4
4
0
.
4
4
9
8
.
4
8
0
.
0
0
1
-
0
.
0
3
5
-
0
.
0
1
5
-
0
.
0
1
0
-
0
.
0
2
3
T
a
b
l
e
6
.
1
:
I
t
e
m
i
z
e
d
f
o
r
c
e
e
r
r
o
r
a
n
d
f
r
e
q
u
e
n
c
y
d
i
e
r
e
n
c
e
129
CHAPTER 7
On Minimal Energy Skew Routing in Lossy
Wireless Sensor Networks
7.1 Introduction
Battery-powered networked embedded systems are being widely used in various
application domains. Practical limitations do not allow replacement of dead bat-
teries after system deployment. Hence, reduction of energy consumption, as a
mechanism to prolong the battery lifetime, has gained signicant attention in
sensor networks community. Previous research eorts have shown that commu-
nication dissipates signicant amount of energy [SHC04]. Consequently, design
choices and protocols that aect the amount of required transmission have a great
impact on system lifetime. For example, minimum-hop routing algorithm reduces
the number of transmissions required to deliver a message at the destination, and
improves system energy dissipation [IGE03].
On the other hand, many applications heavily rely on robust and non-stop
operation of sensing and computing nodes (or simply nodes) to perform their in-
tended functionality. In such classes of applications, failure of a node causes the
entire system to fail, and hence, the lifetime of the system is determined by the
lifetime of the rst node that fails. Consequently, even distribution of energy dis-
sipation becomes an important design concern in addition to total system energy
130
dissipation. Intuitively, even distribution of (minimum-skew) energy dissipation
maximizes the lifetime of the nodes with shortest lifetime, and prolongs the ap-
plication lifetime. The existing energy-aware routing algorithms are generally
oblivious to fair distribution of energy consumption among nodes, and exhibit
poor results.
In this paper, we present an ecient routing algorithms that minimizes the
energy skew among nodes in a mesh network. Our algorithm is a centralized
scheme that is globally optimal. This algorithm eciently nds the best routing
strategy for a given trac pattern in any network topology, including mesh net-
works. Experimental results on our benchmarks show an increase of up to ve
times in system lifetime when comparing with optimal shortest path routing with
a slight penalty in the overall latency of the end-to-end paths.
Before we continue, we would like to clarify some assumptions used in the
remaining of our work. First, our solution is applicable when the quality of the
links used for routing does not dramatically change over time. This assumption
is conrmed for good quality links by [CWK05b]. Second, our solution works
only in the case of unicast routing algorithms. The use of broadcast or multicast
routing mechanisms has been left for future work.
7.2 Related Work
Most of the previous routing protocols ([JM96][MG96][PC97] [BMS01][PR99]) for
wireless ad-hoc networks concentrate on nding and maintaining routes in the
face of changing topology caused by mobility or other environmental changes.
Typical protocols use shortest path algorithms based on hop count, geographic
distance, or transmission power. The rst two are important in minimizing delay
131
and maximizing throughput. The third objective is peculiar to wireless ad-hoc
networks, and is important because typically the nodes involved have a limited
power supply, and radio communication consumes a large fraction of this supply.
To address this issue, several power-aware routing protocols have been developed
([SWR98] [RM98] [SL01] [RH00] [WLB01]). In most of these approaches, the
aim is to minimize the energy consumed per packet in order to deliver it to the
destination. The typical approach is to use a distributed shortest path algorithm
in which the edge costs are related to the power required to transmit a packet
between the two nodes involved. The problem with this technique is that nodes
on the minimum-energy path are quickly drained of power, aecting the network
connectivity when they fail. While some of the most sophisticated routing al-
gorithms associate a cost with routing through a node with low power reserves
([SWR98] [SL01]), they present at best heuristic solutions.
Researchers have explored the fundamental limits of energy-ecient collab-
orative data-gathering by deriving upper bounds on the lifetime of increasingly
sophisticated sensor networks [BC02]. But they do not devise any ecient algo-
rithm for routing.
Another method proposed to extend the sensor network operational time con-
sist of organizing the sensors into a maximal number of disjoint set covers that
are activated successively. Only the sensors from the current active set are re-
sponsible for monitoring all targets and for transmitting the collected data, while
nodes from all other sets are in a low-energy sleep mode [CD05]. The proposed
method, however, is a heuristic. Furthermore, a shortest cost path routing algo-
rithm is studied which uses link costs that reect both the communication energy
consumption rates and the residual energy levels at the two end nodes [CT04].
This approach also formulates the technique as a linear programming problem.
132
Nevertheless, an -optimal polynomial time routing algorithm that maximizes the
lifetime of the system with respect to the distributed energy sources has not been
devised in any of the prior studies. Furthermore, to the best of our knowledge,
no probabilistic distributed routing scheme that attempts to minimize the skew
in energy have been proposed.
Authors in [SL04] present a distributed method for lifetime maximization.
Their method starts out with assuming a certain lifetime and attempts to search
for the optimal lifetime with a bounded error (by starting out with an arbitrary
lifetime) within a certain number of rounds. Our method however, proposes a
simpler approach (despite being centralized) that calculates the optimal lifetime
in polynomial time. More importantly, the main advantage of our method is
that we take into account the lossy channels. Authors in [JDS05] also propose a
similar approach assuming the communication links are lossless.
7.3 Preliminaries and Models
The problem at hand is to design an eective routing mechanism to maximize
the application lifetime, which is given by the lifetime of the rst node that fails.
If nodes have identical initial energy resources, the objective can be rephrased as
minimizing the maximum energy consumption rate of the nodes.
We model the network using an unconstrained connected directed graph G
s
=
(V
s
, E
s
), where V
s
denotes set of nodes v
1
, v
2
, . . . , v
n
, and E
s
represents the set of
directed links l
ij
. Note that directed graphs can be utilized to model any network
topology. We will show that our algorithm is optimal for any directed graph.
We represent the energy consumed by node v
i
to transmit or receive an in-
formation unit (also called packet or message in this paper) by t
i
and r
i
, respec-
133
tively. These two parameters account for the energy consumption incurred by
communication and computation (or possibly other sources of energy consump-
tion) required to transmit or receive a packet. We assume that node v
i
has an
energy level E
i
, which is initially equal to E
i,0
((i, j) V
s
E
i,0
> 0). Note that
E
i
decreases as packets are received and/or sent by node v
i
.
Let f
ij
be the number of packets transmitted from node i to its immediate
neighbor node j over link l
ij
in a routing scenario. We assume that network
has some periodicity property in data transmission, in which case, f
ij
can be
interpreted as the number of packets transmitted over l
ij
in unit of time or its
transmission rate. The lifetime of node v
i
is dened as follows:
T
i
=
E
i,0
(t
i
.
j; l
ij
E
s
f
ij
) + (r
i
.
j; l
ji
E
s
f
ji
)
(7.1)
Where
j; l
ij
E
s
f
ij
is the total number of packets transmitted from node v
i
,
and similarly,
j; l
ji
E
s
f
ji
is the total number of packets received by v
i
.
7.4 Denition
In the following formulation we attempt to minimize the skew in energy con-
sumption (due to wireless communication) in the network. The skew is dened
as follows:
There exists exponential number of paths connecting source to destination nodes.
For every path, we form the following denition. There exists a node in every path
that has the maximum energy consumption (ideally if each path is isolated from
the rest of the network, the energy consumption rate must be identical throughout
the path, however, in reality, nodes may have incoming/outgoing edges from/to
the rest of the network. Therefore, the energy consumption rate may not be
134
necessarily uniformly distributed). In every path P
i
connecting a source to a
destination node, we identify the node with maximum energy consumption rate
as p
imax
. We dene the skew of energy consumption as the dierence between
p
imax
and p
jmax
of paths i and j. In the following Sections, we will illustrate
how our formulation minimizes the dierence of maximum energy consumption
in every two paths. Please note that despite the number of paths are exponential,
the upper bound on the number of nodes with maximum energy consumption is
still n (where n is total number of nodes). The denition of minimum skew is
rephrased in Equation 7.2.
P
i
&P
j
connecting source(s) to destination(s)
Minimize|p
imax
p
jmax
| (7.2)
Figure 7.1: Minimum Skew Denition
135
7.5 Minimal-Skew Routing
7.5.1 Problem Formulation and -Approximation Algorithm
Given a sensor network G
s
and a specic trac pattern, the objective is to route
the packets so that after completion of packet routing, the minimum remain-
ing battery energy among all of the nodes is maximized. This objective would
maximize the lifetime of the rst failing node, and subsequently, maximizes the
application lifetime. We assume that there is a specic node t in G
s
, which serves
as the gateway or base station or destination node, and all of transmitted pack-
ets have to be delivered to t. Specically, the given trac pattern is composed
of a set of (node, quantity) pairs, which species the source and the number of
packets that need to be sent to t. Our formulation, however, accommodates a
more extensive objective which covers the aforementioned objective as described
in Equation 7.2.
Furthermore, we temporarily assume that all nodes have identical initial en-
ergy levels. In Subsection 7.5.3, we will extend our results to show that our
model and technique is extensible to cases where nodes have dissimilar initial
energy levels.
Each packet transmitted from a source node to t can be viewed as a unit
ow in the network G
s
. More precisely, x
ij
units of ow going through the link
l
ij
represent x
ij
packets transmitted from v
i
to v
j
. This trac decreases the
energy resources at v
i
and v
j
by x
ij
t
i
and x
ij
r
j
, respectively. The given
trac pattern can be interpreted as a ow supply vector, FS, which specied the
amount of ow supply at nodes. All of packets have to be received at node t, and
hence, Node t is the only node with negative ow supply (positive ow demand).
The supply for node t is set such that the total supply over all the nodes in the
136
network is zero.
The problem of packet routing is equivalent to nding a feasible network ow
in G
s
. Intuitively, the objective of maximizing the minimum remaining energy of
the nodes is equivalent to minimizing the maximum ow passing through nodes.
Note that the remaining energy at intermediate (not a source or destination)
node v
i
with x
i
units of ow passing through it is E
i
x
i
.(r
i
+ t
i
). In the
remainder of this section, we transform G
s
to a new network G
t
in which, min-
cost ow solution generates the optimal solution. We proceed to describe the
transformation procedure, along with mathematical properties of our technique
and proofs of correctness.
We construct the network G
t
= (V
t
, E
t
) from the graph G
s
according to the
following rules. Each node in graph G
s
is split into k + 2 nodes where k is
a tunable parameter that controls the accuracy of the solution. We refer to the
resulting set of nodes as a partition. In each partition, two nodes serve as receiver
and transmitter (or input and output) and the rest of the nodes are called splits.
Figure 7.2 illustrates an example partition in which, nodes v
(k+2)(i1)+1
and v
(k+2)i
are the receiver and the transmitter of the partition, respectively. Figure 7.3
shows an example network G
s
and the resulting network G
t
after transformation.
Each intermediate node v
i
in G
s
consumes r
i
+ t
i
units of energy to relay a
packet. Therefore, v
i
can relay at most u
i
= E
i
/(r
i
+ t
i
) packets before running
out of battery. u
i
forms an upper bound on the number of packets that can pass
through node v
i
. In order to embed this constraints into the problem, we assign
the ow upper bound of u
i
= u
i
/k to each of the split nodes. The upper bounds
guarantee that a partition cannot pass more ow after relaying u
i
units of ow,
or equivalently, a node cannot route packets after its battery is dead. The ow
upper bound for all of the receiver and transmitter nodes is innity. Note that the
137
Figure 7.2: Node partitioning
upper bound u
i
= u
i
/k on the split nodes in partition i, implicitly sets the upper
bound of u
i
for the receiver and transmitter nodes of the partition. The receiver,
transmitter and split nodes are assigned a special sequence of costs for passing the
unit ow. The cost associated with receiver and transmitter nodes of a partition
is zero. The costs associated with split nodes of a partition are increasing from
left to right (Figure 7.2). This directs the min-cost ow solution to utilize the split
nodes from left to right, when passing ow through the partition. Specically,
we assign the following costs to the nodes in partition i:
138
Figure 7.3: Original network and the resulting network after partitioning
c
(k+2)(i1)+1
= 0
c
(k+2)(i1)+2
= 1
c
(k+2)(i1)+3
= n +
c
(k+2)(i1)+4
= n(n + 1 +) +
.
.
.
c
(k+2)i1
= n
k
j=2
c
(k+2)(i1)+j
+
c
(k+2)i
= 0
(7.3)
The cost on each split node is enforced such that it would be greater than the
cumulative cost of the split nodes with smaller indices over all of the partitions.
Intuitively, our cost assignment technique enforces the optimal min-cost ow
solution to utilize the split nodes with smaller indices, before trying to utilize a
particular split node. This simple yet eective idea exhibits the main property of
139
our technique by which, we minimize the maximum energy consumption of the
nodes. For simplicity, we dene cost rank where cost rank R
i
has the following
property:
R
i
>
j<i
n R
j
(7.4)
Consequently, Equations 7.3 may be reworded as follows:
c
(k+2)(i1)+1
= 0
c
(k+2)(i1)+2
= R
1
c
(k+2)(i1)+3
= R
2
c
(k+2)(i1)+4
= R
3
.
.
.
c
(k+2)i1
= R
k
c
(k+2)i
= 0
(7.5)
The loss in edges may also be accommodated easily. For the sake of simplic-
ity, we add the extensions for lossy edges in Section 7.5.2 The aforementioned
procedure provides a well-dened set of steps to create G
t
from G
s
. A packet
routed from a sensor node s, to the base station t, in G
s
, corresponds to a unit
of s > t ow in G
t
. Similarly, our partition construction scheme along with
the upper bound settings on split nodes guarantee that any ow solution in G
t
represents a feasible routing solution in G
s
. Note that the given trac pattern for
G
s
can be readily translated into a ow supply vector (FS) for G
t
. FS species
the amount of ow supply or demand for all of the nodes.
We now prove that our cost assignment strategy implies that the min-cost
ow solution in G
t
corresponds to a routing scheme in G
s
that minimizes the
maximum consumed energy at the nodes (or equivalently, maximizes the mini-
mum remaining energy over all of the nodes). Let x
ij
represent the amount of
ow on edge (i, j). Let y
i
represent the amount of ow going through split node
k in partition i. Similarly, let c
i
denote the associated cost of unit ow passing
through that node in G
t
. The min-cost ow problem for graph G
t
with the given
140
supply and demand vector, FS, can be written as:
Minimize
i
c
i,k
y
i,k
(7.6)
Subject to:
{j:(i,j)A}
x
ij

{j:(j,i)A}
x
ji
= b(i) (7.7)
0 x
ij
u
ij
; (i, j) V
t
(7.8)
0 y
i
u
i
; (i) V
t
(7.9)
Equation (7.7) is the ow conservation condition at each node and Equations
(7.8) and (7.9) are capacity constraints for the arcs and the nodes respectively.
We assume that the lower bounds on arc ows as well as the lower bounds on
node ows are zero. Moreover, the number of nodes in network G
s
is denoted by
n.
The ow that passes through the receiver or transmitter nodes of a partition
represents the total ow passing through that partition, or equivalently, it deter-
mines the energy consumption at the sensor node corresponding to the partition.
The more the ow, the shorter the lifetime of the sensor node. Therefore, the
objective is to minimize the maximum amount of ow passing through partitions.
The ow passing through any partition has to pass through its split nodes,
and subsequently, min-cost ow solutions utilize the splits with lower costs before
higher cost splits. We denote the number of split nodes in partition i that carry
a non-zero ow by
i
.
i
=
k+1
j=2
y
(k+2)(i1)+j
u
(7.10)
141
Theorem 7.5.1. The objective function in Equation 7.6 minimizes the maximum
ow in the nodes of network G
s
= (V
s
, E
s
) with maximum error where u
i
=
u
i
/k = 1/k. This is equivalent to maximizing the minimum battery lifetime of
the nodes in network G
s
(or its transformed network G
t
) and therefore analogous
to maximizing the lifetime of the system.
Proof. Proof is formed by contradiction. Assume our technique generates solu-
tion L where the ow entering each partition is represented by f
i
. Let f
max
=
(max(f
i
)i). Assume there exist another solution L
with maximum ow denoted

by f
max
where f
max
< f
max
+.
Thus, (7.10) conveys that:
max
<
max
Considering the cost on splits in Figure 7.2 and Equation (7.3), we conclude
that the overall cost of ow for solution L is greater than the overall cost of
ow for solution L
. This contradicts the optimality of min-cost ow technique

since the solution found (L) does not have the minimum cost. Therefore, by
contradiction, the solution L
cannot exist and our formulation minimizes the

maximum ow (or maximizes the minimum lifetime).
Due to the ow conservation condition in min-cost ow technique, it is trivial
that the ow in any nodes may not be reduced individually to minimize the
objective function.
The intuition behind our proposed technique is that the cost assignment on
the splits forces the network to rout a ow from the kth split of node v
i
, if it
cannot be routed through any number of other nodes whose (k 1)th splits is
empty.
142
Theorem 7.5.2. The solution L , generated by our technique, minimizes the
dierence of the maximum node ows throughout every two disjoint paths con-
necting source to a destination node (with tolerance of = 1/k - minimal-skew).
In other words, these two nodes further must have the minimum lifetime of the
nodes along each path and are regarded as the bottleneck nodes (i.e. such nodes
have the maximum energy consumption rate along their corresponding paths).
Proof. This proof is presented by contradiction as well. Assume there exists a
feasible solution L
which was transformed from L and the ows of two partitions

i and j was altered such that their dierence is reduced. We denote the new ows
in solution L by f
i
and f
j
. Without loss of generality, we assume f
i
< f
j
.
Figure 7.4: Flow exchange to obtain minimal-skew solution
We investigate both possible scenarios. One scenario occurs when the trans-
formation involves changing the ow in only one partition (increasing f
i
or de-
creasing f
j
) According to the ow conservation theorem, this scenario is not
feasible. The other scenario, as shown in Figure 7.4, takes place when the ow
of partition i is increased and the ow of partition j is decreased (by a value of
143
greater than ) such that:
|f
j
f
i
| < |f
j
f
i
| +
and
f
j
+f
i
= f
j
+f
i
+
(7.11)
From Equation 9, it can be easily shown that:
f
i
< f
i
< f
j
< f
j
(7.12)
Therefore, the number of splits that are utilized for passing the ow follows
the same convention:
i
<
i
<
j
<
j
(7.13)
As discussed before, in each partition, each split has a cost that alone is greater
than n times the cumulative cost of all the precedent splits with smaller indices
(n = number of nodes in G
s
(original graph)). Since the solutions L and L
are
similar except in partitions i and j, therefore, the overall cost of solution L is
greater than the cost of L
(
j
>
j
). This contradicts the optimality of min-cost
ow algorithm as the solution L must have the minimum cost. Hence, solution
L cannot exist.
Theorem 7.5.3. The solution of minimal-skew routing is unique in the sense
that the lifetime of nodes in the network in descending order is unique.
Proof. Assume F and F
are the vectors containing the ow of all nodes in de-

scending order for two optimal solutions, L and L
. Obviously F[1] = F
[1],
otherwise the two total costs would be dierent (that contradicts the optimal-
ity of the solution). Inductively, this argument holds for every index. Assume
F[i] = F
[i], for i = 1 . . . k. If F[k + 1] < F
[k + 1] then because of the special

cost assignment of the splits, the cost of F
[k + 1] itself would be larger than

the total cost of F[i]s, i = k + 1 . . . n. Therefore, the total cost of solution L
is
144
greater than the cost of L. This contradicts the optimality of solution L
. This
completes the proof.
7.5.2 Discussion on Lossy Channels
In general, accommodating lossy communication channels in minimum cost ow
technique is not quite an easy task due to ow conservation conditions. In the
case where the existence of lossy communication links are inevitable, the packets
received at the receiver is less than the number of packets transmitted. This
scenario corresponds to ow attenuation in our minimum cost formulation. To
address the data/ow loss in communication links, ideally we prefer to remove
certain portion of the ow traveling through an edge and transfer it to a aux-
iliary node which is considered as a dummy sink node. Other gateways and
sink nodes are also connected to the aforementioned node. This can be accom-
plished by adding intermediate nodes on every interconnection edge. Moreover,
the intermediate nodes have to be connected through lossy edges with cost zero
a dummy sink node. The upper bound on lossy edges corresponds to the
loss rate of every interconnection edge. The limitation we face, however, is that
the initial ow in every edge tends to be directed through the lossy edge until
the lossy edge reaches its capacity. This causes a potential problem where the
interconnection edges are not utilized with their full capacity. Therefore, this will
incur in an increase in loss rate. To address this issue, we propose the following
transformation:
Each edge is divided into k split edges e
ik
, ..., e
i(k+1)1
and k lossy edges
l
ik
, ..., l
i(k+1)1
connecting intermediate nodes to a dummy node as shown in Fig-
ure 7.5. The costs on split edges correspond to R
1
, R
3
, ...R
2k
while the cost on
lossy edges correspond to R
2
, R
4
, ...R
2k+1
. This will ensure that initially edge e
ik
145
Figure 7.5: Edge splitting
is utilized and subsequently lossy edge l
ik
. Despite, we attempt to characterize
the probabilistic nature of loss with a deterministic approach, over a long run, the
behavior of our model become analogous to steady behavior of real systems. The
above cost distribution will ensure that edges e
ik
, l
ik
, e
ik+1
, ..., e
i(k+1)1
, l
i(k+1)1
are utilized accordingly. Changing the cost ranks associated with the edges en-
ables us to accommodate other combinations of edge utilization. All Theorems
and Lemmas proven in the this Section still holds for graphs with lossy edges
(with minor modications). The proofs are omitted due to lack of space.
7.5.3 Dissimilar Initial Energy Levels
Throughout our formulation, we assumed that the initial energy levels in all
nodes are similar. Dissimilar initial energy levels, however, can be simply accom-
modated by modifying the capacity of the splits (upper-bounds on the splits).
Throughout our formulation, we assumed that the initial energy levels in all
nodes are similar. Dissimilar initial energy levels, however, can be simply accom-
146
modated by modifying the capacity of the splits (upper-bounds on the splits).
The main property of our technique is that we assign various costs to dierent
levels of energy stored in a battery. When a battery is fully charged, it can be
used more easily than the case it has half of the full charge. Therefore, in the case
where some nodes in the network do not have full energy level, the already used
portion of their battery corresponds to the splits with lesser cost. Hence, those
splits could be assigned upper bound of zero in the min-cost ow formulation as
if they have been already used.
7.5.4 Time Complexity Analysis
The formulation used for the proposed min-cost ow problem in Equations (2)
through (7.9) is an LP formulation which can be solved with standard LP solvers.
Throughout our experiments, we used Matlab as an LP solver. The use of LP-
solvers enabled us the ability to have non-integral capacity in the formulation and
non-integral ow in the solution. In order to provide combinatorial algorithms
to solve such problem, integrality constraints must be enforced. Fortunately
all the problem parameters such as supply/demands and capacities assigned to
splits can be scaled by a factor of k (the number of splits) and meet the integrality
constraints. Therefore, the time complexity of our technique is after scaling would
be O((mlog(nk))(m+nk log(nk))) where m is the number of edges in the graph;
n and k are the number of nodes and splits in each partition, respectively. The
time complexity can be easily derived from the time complexity of the min-cost
ow algorithms and the size of our constructed network (O(nk)). The original
time-complexity of min-cost ow is reported in [AMO93].
147
7.5.5 Discussion on Multicommodity
In scenarios where several sources and destinations are involved and the com-
munication pairs exchange dierent type of messages, routing problems can
be modeled as multicommodity ow. Out technique can not directly address
this class of problems [GJ79a]. The problem itself is known to be NP-complete.
However, we believe that our methodology can be applied in conjunction with
the known heuristics for multicommodity ow problems and generate reasonable
results. We have not considered this class of networks but we plan to study it in
near future.
7.6 Experimental Results
We generated various benchmarks based on random graphs with relatively large
number of nodes to illustrate the eectiveness of our centralized technique. For
simplicity, we assumed that the energy level in all benchmarks is uniform. Our
benchmark sets resemble real-world networks. In most network applications, it
might be unlikely that a large portion of sensor nodes are placed within close
proximity of each other. Instead, they are placed on a grid with certain random
properties. It can be envisioned as a locally random globally regular networks.
Such scenario can be imagined with the following example: Certain number of
sensor nodes is required to be placed in a building. Each room has a specic
number of sensor nodes which is constant, yet, the position of the nodes is random
within each room. We call this set as random networks with grid distribution.
We generate such benchmarks by dividing the area into unit-size tiles. A tile that
does not have a sensor node is selected randomly. We place a node in the tile
with uniform distribution. This procedure is repeated until all tiles are covered.
148
If more sensor nodes are required to be inserted, the same course is recurred until
all sensor nodes are placed. One hundred sensors are placed within areas of size
160 40, 160 60, 160 80, 200 40, 200 60, 200 80 and 200 100. We
refer to this set of benchmarks as grid networks. In all networks, three source
nodes are placed on far left side of the square area while the destination nodes/
gateways are placed on the right side of the square. This particularly assist us to
place the source/destination not within the close proximity of each other. The
connectivity between nodes is determined by statistical models developed using
real RF transcievers under dierent conditions and scenarios [CWK05a]. Only
communication links with reliability greater than 80% are considered. In the
next set of experiments, the locations of the nodes are generated conforming to
a random uniform distribution over areas various size. The We refer to this class
of random networks as random networks.
Firstly, we illustrated the eectiveness of our technique on two benchmarks.
One benchmark belongs to our class of grid networks while the other resembles
random network. Each network consists of 100 nodes scattered over an area of
200 square meters.
The graph topology of grid and random networks are illustrated in Fig-
ures 7.6 and 7.8 respectively.
Figure 7.7 depict the normalized energy consumption rate of the nodes in the
system sorted in ascending order for various numbers of splits (k) in the grid
network. The eectiveness of our algorithm on how the energy consumption rate
becomes evenly distributed is clearly demonstrated when k is increased. Figure
7.7 illustrates the same graph for the random network shown in Figure 7.8.
In the next set of experiments, the size on benchmarks is varied. For every
particular size, twenty benchmarks are generated with three source and three
149
0 20 40 60 80 100 120 140 160 180 200
0
10
20
30
40
50
60
70
80
90
100
Graph Topology grid200x100
(meter)
(
m
e
t
e
r
)
Figure 7.6: Graph topology (n=100, grid distribution)
destination nodes. To compare our scheme against other routing algorithms, we
consider a shortest path routing algorithm based on minimum cost ow. Figures
7.10, 7.11 , 7.12 and 7.13 demonstrate the lifetime resulted from our scheme
compared to min-cost technique. In all diagrams, each data-point corresponds
to the average taken over twenty benchmarks. Throughout all diagrams, as the
number of splits increases, the lifetime improves due to forcing the network to
utilize vertex disjoint paths more eectively. Overall, the average lifetime of the
systems (for k = 4) is increased by a factor of 4.38 compared to min-cost routing
approach.
Figures 7.14, 7.15, 7.16 and 7.17 exhibits the delay trade-os of our routing
algorithm with respect to the min-cost shortest path for various benchmarks.
Overall, the average delay of our scheme (for k = 4) is 10% greater than
150
1
1.5
2
2.5
3
3.5
4
0
20
40
60
80
100
0
0.2
0.4
0.6
0.8
1
k (number of splits)
grid200x100
node #
N
o
r
m
a
l
i
z
e
d

e
n
e
r
g
y

c
o
n
s
u
m
p
t
i
o
n

r
a
t
e
Figure 7.7: Normalized energy consumption rate in sorted order (n=100, grid
distribution)
the min-cost shortest path. The average delay is not vastly increased due to
existence of multiple vertex disjoint paths between sources and destinations in
our benchmarks. The average delay also increases slightly as the k (number of
splits) is increased.
In general, highly connected networks such as grids provide a large number of
parallel paths between nodes which is of our interest and enhances the exibility
of data routing.
151
0 20 40 60 80 100 120 140 160 180 200
0
10
20
30
40
50
60
70
80
90
100
Graph Topology rand200x100
(meter)
(
m
e
t
e
r
)
Figure 7.8: Graph topology (n=100, random distribution)
7.7 Conclusion
We proposed a polynomial time -optimal technique for multi-hop routing in
wireless networks with distributed battery sources. Our technique maximizes the
lifetime of the system. Furthermore, it evenly distributes the energy consumption
rate which yields in a minimal-skew solution for node utilization. We theoreti-
cally proved that our technique is ecient and has polynomial time complexity.
Furthermore, our technique accommodates routing through lossy links while the
optimality is not sacriced. Our investigation on various benchmarks revealed the
quality of the solutions generated by our methodology even with a small number
of splits.
152
1
1.5
2
2.5
3
3.5
4
0
20
40
60
80
100
0
0.2
0.4
0.6
0.8
1
k (number of splits)
rand200x100
Node #
N
o
r
m
a
l
i
z
e
d

e
n
e
r
g
y

c
o
n
s
u
m
p
t
i
o
n

r
a
t
e
Figure 7.9: Normalized energy consumption rate in sorted order (n=100, random
distribution)
153
Figure 7.10: min-skew/min-cost lifetime for various number of splits - k (n=100,
grid distribution)
grid distribution)
154
rand distribution)
rand distribution)
155
Figure 7.14: Average delay of min-skew/min-cost for various number of splits -
k (n=100, grid distribution)
k (n=100, grid distribution)
156
k (n=100, rand distribution)
k (n=100, rand distribution)
157
References
[AMO93] Ravindra K. Ahuja, Thomas L. Magnanti, and James B. Orlin. Net-
work ows: theory, algorithms, and applications. Prentice-Hall, Inc.,
1993.
[atm] ATmega128L Manual. http://www.atmel.com.
[BBV04] Guillermo Barrenechea, Baltasar Beferull-Lozano, and Martin Vet-
terli. Lattice sensor networks: capacity limits, optimal routing and
robustness to failures. In IPSN04: Proceedings of the third interna-
tional symposium on Information processing in sensor networks, pp.
186195. ACM Press, 2004.
[BC02] M. Bhardwaj and A. Chandrakasan. Bounding the lifetime of sensor
networks via optimal role assignments., 2002.
[BMS01] Prosenjit Bose, Pat Morin, Ivan Stojmenović, and Jorge Urru-
tia. Routing with guaranteed delivery in ad hoc wireless networks.
Wirel. Netw., 7(6):609616, 2001.
[Bon94] M. Bona. Textile Quality. Texilia, 1994.
[BP83] M. Burstein and R. Pelavin. Hierarchical Wire Routing. IEEE
Transactions on Computer-Aided Design of Integrated Circuits, pp.
223234, 1983.
[CB04] I Christov and G Bortolan. Ranking of pattern recognition param-
eters for premature ventricular contractions classication by neural
networks. Physiological Measurement, 25(5):12811290, 2004.
[CCK99] A.E. Caldwell, H. Choi, A.B. Kahng, S. Mantik, M. Potkonjak,
G. Qu, and J.L. Wong. Eective Iterative Techniques for Finger-
printing Design IP. In Design Automation Conference, pp. 843848.
ACM/IEEE, 1999.
[CCK00] T. F. Chan, J. Cong, T. Kong, and J. R. Shinnerl. Multilevel Op-
timization for Large-Scale Circuit Placement. pp. 171176. IEEE,
2000.
[CD05] M. Cardei and D.-Z. Du. Improving Wireless Sensor Network Life-
time through Power Aware Organization. ACM Wireless Networks,
11(3), May 2005.
158
[COR04] Philip de Chazal, Maria ODwyer, and Richard B. Reilly. Automatic
Classication of Heartbeats Using ECG Morphology and Heartbeat
Interval Features. IEEE TRANSACTIONS ON BIOMEDICAL EN-
GINEERING, 51(7):11961206, 2004.
[CT04] Jae-Hwan Chang and Leandros Tassiulas. Maximum lifetime
routing in wireless sensor networks. IEEE/ACM Trans. Netw.,
12(4):609619, 2004.
[CWK05a] Alberto Cerpa, Jennifer L. Wong, Loaune Kuang, Miodrag Potkon-
jak, and Deborah Estrin. Statistical Model of Lossy Links in
Wireless Sensor Networks. In Proceedings of the ACM/IEEE
Fourth International Conference on Information Processing in Sen-
sor Networks (IPSN05), Los Angeles, CA, USA, April 2527 2005.
ACM/IEEE.
[CWK05b] Alberto Cerpa, Jennifer L. Wong, Loaune Kuang, Miodrag Potkon-
jak, and Deborah Estrin. Temporal Properties of Low Power
Wireless Links: Modeling and Implications on Multi-Hop Routing.
In Proceedings of the Sixth ACM/IEEE MOBIHOC05), Urbana-
Champaign, IL, USA, May 2528 2005. ACM/IEEE.
[CXY04] Fang-Chung Chen, Qianfei Xu, and Yang Yang. Enhanced e-
ciency of plastic photovoltaic devices by blending with ionic solid
electrolytes. Applied Physics Letter., p. 3181, 2004.
[DD96] Shantanu Dutt and Wenyong Deng. A probability-based approach
to VLSI circuit partitioning. In DAC 96: Proceedings of the 33rd
annual conference on Design automation, pp. 100105, New York,
NY, USA, 1996. ACM Press.
[DHS00] Richard O. Duda, Peter E. Hart, and David G. Stork. Pattern Clas-
sication. John Wiley and Sons, Inc., second edition, 2000.
[DK85] A. E. Dunlop and B. W. Kernighan. A Procedure for Placement
of Standard Cell VLSI Circuits. IEEE Transactions on Computer-
Aided Design of Integrated Circuits, 4(1):9298, January 1985.
[DMP03] R. DeVaul, J. Gips M. Sung, and A. Pentland. MIThril 2003: ap-
plications and architecture. In Wearable Computers, Seventh IEEE
International Symposium on,, pp. 411. IEEE, 2003.
[GJ79a] M. R. Garey and D. S. Johnson. Computers and intractability: a
guide to the theory of NP-completeness. W. H. Freeman, 1979.
159
[GJ79b] Michael R. Garey and David S. Johnson. Computers and Instractabil-
ity: A Guide to the Theory of NPCompleteness. W.H Freeman, San
Francisco, CA, 1979.
[The Bible. An indispensible guide to NPcompleteness and related
issues. Updated by Johnsons semiregular column in the Journal of
Algorithms.]
[GLB03] David Gay, Philip Levis, Robert von Behren, Matt Welsh, Eric
Brewer, and David Culler. The nesC language: A holistic approach
to networked embedded systems. In PLDI 03: Proceedings of the
ACM SIGPLAN 2003 conference on Programming language design
and implementation, pp. 111, New York, NY, USA, 2003. ACM
Press.
[GLS98] J. Garcia, P. Lander, L. Sornmz, S. Olmos, G. Wagner, , and P. La-
guna. Comparative Study of Local and Karhunen-Loeve-Based ST-
T Indexes in Recordings from Human Subject with Induced Myocar-
dial Ischemia. Computer and Biomedical Research, 31(CO981481),
1998.
[Gra95] Amara Graps. An Introduction to Wavelets. IEEE Computational
Sciences and Engineering, 2(2):5061, 1995.
[Hal70] K.M. Hall. An r-dimensional Quadratic Placement Algorithm.
Management Science, pp. 219229, 1970.
[HKS05] Chih-Chieh Han, Ram Kumar, Roy Shea, Eddie Kohler, and Mani
Srivastava. A dynamic operating system for sensor nodes. In Mo-
biSys 05: Proceedings of the 3rd international conference on Mo-
bile systems, applications, and services, pp. 163176, New York, NY,
USA, 2005. ACM Press.
[HSW00] Jason Hill, Robert Szewczyk, Alec Woo, Seth Hollar, David E. Culler,
and Kristofer S. J. Pister. System Architecture Directions for Net-
worked Sensors. In Architectural Support for Programming Lan-
guages and Operating Systems, pp. 93104, 2000.
[IGE03] Chalermek Intanagonwiwat, Ramesh Govindan, Deborah Estrin,
John Heidemann, and Fabio Silva. Directed diusion for wireless
sensor networking. IEEE/ACM Trans. Netw., 11(1):216, 2003.
[JDB05] Roozbeh Jafari, Foad Dabiri, Philip Brisk, and Majid Sarrafzadeh.
Adaptive and fault tolerant medical vest for life-critical medical
monitoring. In SAC 05: Proceedings of the 2005 ACM symposium
160
on Applied computing, pp. 272279, New York, NY, USA, 2005. ACM
Press.
[JDM00] Anil K. Jain, Robert P. W. Duin, and Jianchang Mao. Statistical
Pattern Recognition: A Review. IEEE Trans. Pattern Anal. Mach.
Intell., 22(1):437, 2000.
[JDS05] Roozbeh Jafari, Foad Dabiri, and Majid Sarrafzadeh. -Optimal
Minimal-Skew Battery Lifetime Routing in Distributed Embedded
Systems. Journal of Low Power Electronics, 1(2):97107, 2005.
[JEZ05] Roozbeh Jafari, Andre Encarnacao, Azad Zahoory, Foad Dabiri, Hy-
duke Noshadi, and Majid Sarrafzadeh. Wireless Sensor Networks for
Health Monitoring. In MobiQuitous 05: Proceedings of the Second
Annual International Conference on Mobile and Ubiquitous Systems,
2005.
[JM96] David B Johnson and David A Maltz. Dynamic Source Routing in
Ad Hoc Wireless Networks. In Imielinski and Korth, editors, Mobile
Computing, volume 353. Kluwer Academic Publishers, 1996.
[KAK97a] G. Karypis, R. Aggarwal, V. Kumar, and S. Shekhar. Multilevel
Hypergraph Partitioning: Application in VLSI Domain. pp. 526
529. IEEE/ACM, 1997.
[KAK97b] George Karypis, Rajat Aggarwal, Vipin Kumar, and Shashi Shekhar.
Multilevel hypergraph partitioning: application in VLSI domain.
In DAC 97: Proceedings of the 34th annual conference on Design
automation, pp. 526529, New York, NY, USA, 1997. ACM Press.
[KEF03] Henry Kautz, Oren Etzioni, Dieter Fox, and Dan Weld. Founda-
tions of Assisted Cognition Systems. Technical report, University of
Washington, Computer Science Department, Technical Report, 2003.
[KM05] Jung-Chun Kao and Radu Marculescu. Energy-Aware Routing for
E-Textile Applications. In DATE 05: Proceedings of the conference
on Design, Automation and Test in Europe, pp. 184189, Washing-
ton, DC, USA, 2005. IEEE Computer Society.
[KSJ91] J. M. Kleinhans, G. Sigl, F. M. Johannes, and K. J. Antreich.
GORDIAN: VLSI Placement by Quadratic Programming and Slic-
ing Optimization. IEEE Transactions on Computer-Aided Design
of Integrated Circuits, 10(3):365365, 1991.
161
[Lee61] J.F. Lee. An Algorithm for Path Connection and Its Application.
IRE Transactions on Electronic Computer, pp. 346365, 1961.
[lif] Lifeguard Monitoring System. http://lifeguard.stanford.edu.
[LMG97] Pablo Laguna, Roger G. Mark, Ary Goldberger, and George B.
Moody. A Database for Evaluation of Algorithms for Measurement
of QT and Other Waveform Intervals in the ECG., 1997.
[LMG04] Philip Levis, Sam Madden, David Gay, Joseph Polastre, Robert
Szewczyk, Alec Woo, Eric Brewer, and David Culler. The Emer-
gence of Networking Abstractions and Techniques in TinyOS. In
First Symposium on networked system design and implementation
(NSDI04), pp. 114, San Francisco, California, USA, 2004.
[LTK92] N. Lek, R.W. Thaik, and S.M. Kang. A New Global Router Using
Zero-One Integer Linear Programming Techniques for Sea-of-Gates
and Custom Logic Arrays. IEEE Transactions on Computer-Aided
Design of Integrated Circuits, 1992.
[MG96] Shree Murthy and J. J. Garcia-Luna-Aceves. An ecient routing
protocol for wireless networks. Mob. Netw. Appl., 1(2):183197,
1996.
[mid] MIDMARK Diagnostics Group. http://midmarkdiagnostics.com.
[MJE03] T. Martin, M. Jones, J. Edmison, and R. Shenoy. Towards a design
framework for wearable electronic textiles. In Wearable Computers,
Seventh IEEE International Symposium on,, pp. 190 199. IEEE,
2003.
[MM90] G. B. Moody and R. G. Mark. The MIT-BIH Arrhythmia Database
on CD-ROM and software for use with it. In Computers in Cardiol-
ogy 1990, pp. 185188, Los Alamitos, 1990. IEEE Computer Society
Press.
[MM02] D. Meoli and T. May-Plumlee. Interactive Electornic Textile Devel-
opment: A Review of Technologies. Journal of Textile and Apparel,
Technology and Management, 2(2), 2002.
[MMK02] D. Marculescu, R. Marculescu, and P.K. Khosla. Challenges and
opportunities in electronic textiles modeling and optimization. In
Design Automation Conference, 2002. Proceedings. 39th, pp. 175
180. ACM/IEEE, 2002.
162
[NM93] Lionel M. Ni and Philip K. McKinley. A Survey of Wormhole Rout-
ing Techniques in Direct Networks. IEEE Computer, 26(2):6276,
1993.
[Pav77] T. Pavlidis. Structural Pattern Recognition, volume 1 of Springer
Series in Electrophysics. Springer-Verlag, 1977.
[PC97] Vincent D. Park and M. Scott Corson. A Highly Adaptive Dis-
tributed Routing Algorithm for Mobile Wireless Networks. In IN-
FOCOM 97: Proceedings of the INFOCOM 97. Sixteenth Annual
Joint Conference of the IEEE Computer and Communications Soci-
eties. Driving the Information Revolution, p. 1405, Washington, DC,
USA, 1997. IEEE Computer Society.
[phya] PhysioBank - physiologic signal archives for biomedical research.
http://www.physionet.org/physiobank/.
[phyb] PhysioToolkit - open source software for biomedical science and en-
gineering. http://www.physionet.org/physiotools/.
[PMJ02] Sungmee Park, K. Mackenzie, and S. Jayaraman. The wearable
motherboard: a framework for personalized mobile information pro-
cessing (PMIP). In Design Automation Conference, 2002. Proceed-
ings. 39th, pp. 170174. ACM/IEEE, 2002.
[PMV01] S. Papadimitriou, S. Mavroudi, L. Vladutu, , and A. Bezerianos.
Ischemia Detection with a Self-Organizing Map Supplemented by
Supervised Learning. IEEE Trans. on Neural Networks, 12(3), 2001.
[PR99] Charles E. Perkins and Elizabeth M. Royer. Ad-hoc On-Demand
Distance Vector Routing. In WMCSA 99: Proceedings of the Sec-
ond IEEE Workshop on Mobile Computer Systems and Applications,
p. 90, Washington, DC, USA, 1999. IEEE Computer Society.
[Pri01] Anthony Primentas. Puncture and Tear of Woven Fabrics. Journal
of Textile and Apparel, Technology and Management, 1(4), 2001.
[PT85] J. Pan and W. J. Tompkins. A real-time QRS detection algorithm.
IEEE Trans. Biomed. Eng, 32(3):230236, 1985.
[QB79] N.R. Quinn and M.A. Breuer. A Force Directed Component Place-
ment Procedure for Printed Circuit Boards. IEEE Transactions on
Circuits and Systems, pp. 377388, 1979.
163
[RH00] Ram Ramanathan and Regina Hain. Topology Control of Multihop
Wireless Networks Using Transmit Power Adjustment. In INFO-
COM (2), pp. 404413, 2000.
[RM98] V. Rodoplu and T. Meng. Minimum energy mobile wireless net-
works., 1998.
[sen] Sensatex. http://www.sensatex.com.
[SHC04] Victor Shnayder, Mark Hempstead, Bor rong Chen, Geo Werner
Allen, and Matt Welsh. Simulating the power consumption of large-
scale sensor network applications. In SenSys 04: Proceedings of the
2nd international conference on Embedded networked sensor systems,
pp. 188200, New York, NY, USA, 2004. ACM Press.
[She93] N. A. Sherwani. Algorithms For VLSI Physical Design Automation.
Kluwer Academic Publishers, 1993.
[SK87] E. Shargowitz and J. Keel. A Global Router Based on Multicom-
modity Flow Model. The VLSI Journal, pp. 316, 1987.
[SL01] Ivan Stojmenovic and Xu Lin. Power-Aware Localized Routing in
Wireless Networks. IEEE Transactions on Parallel and Distributed
Systems, 12(11):11221133, 2001.
[SL04] Arvind Sankar and Zhen Liu. Maximum Lifetime Routing in Wire-
less Ad-hoc Networks. In INFOCOM, 2004.
[SM90] K. Shahookar and P. Mazumder. A Genetic Approach to Stan-
dard Cell Placement Using Meta-genetic Parameter Optimization.
IEEE Transactions on Computer-Aided Design of Integrated Circuits,
9(5):500511, 1990.
[SM94] P. Singh and HI. Maibach. Iontophoresis in drug delivery: ba-
sic principles and applications. Crit Rev Ther Drug Carrier Syst.,
11:161213, 1994.
[SMM03] Phillip Stanley-Marbell, Diana Marculescu, Radu Marculescu, and
Pradeep K. Khosla. Modeling, Analysis, and Self-Management of
Electronic Textiles. IEEE Trans. Comput., 52(8):9961010, 2003.
[SS95] W. J. Sun and C. Sechen. Ecient and Eective Placement for
Very Large Circuits. IEEE Transactions on Computer-Aided De-
sign of Integrated Circuits, 14(3):349359, March 1995.
164
[SW96] M. Sarrafzadeh and C.K. Wong. An Intorduction to VLSI Physical
Design. McGraw-Hill, 1996.
[SWR98] Suresh Singh, Mike Woo, and C. S. Raghavendra. Power-aware rout-
ing in mobile ad hoc networks. In MobiCom 98: Proceedings of the
4th annual ACM/IEEE international conference on Mobile comput-
ing and networking, pp. 181190, New York, NY, USA, 1998. ACM
Press.
[TLP05] Ben L. Titzer, Daniel Lee, and Jens Palsberg. Avrora: Scalable Sen-
sor Network Simulation with Precise Timing. In IPSN05, Fourth
International Conference on Information Processing in Sensor Net-
works, 2005.
[WF04] B. Witkowska and I. Frydrych. A Comparative Analysis of Tear
Strength Methods. Fibers and Textiles in Eastern Europe, 12(2),
June 2004.
[WLB01] Roger Wattenhofer, Li Li, Paramvir Bahl, and Yi-Min Wang. Dis-
tributed Topology Control for Wireless Multihop Ad-hoc Networks.
In INFOCOM, pp. 13881397, 2001.
[WYS00] M. Wang, X. Yang, and M. Sarrafzadeh. Dragon2000: Fast
Standard-cell Placement for Large Circuits. pp. 260263, 2000.
[xbo] Crossbow Technology Inc. http://www.xbow.com.
[ZR03] MW Zimmerman, MT Johnson RJ Povinelli, , and KM Ropella.
A Reconstructed Phase Space Approach for Distinguishing Ischemic
from Non-Ischemic ST Changes using Holter ECG Data. Computers
in Cardiology, 30, 2003.
165

Medical Embedded Systems

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Medical Embedded Systems

Uploaded by

Copyright:

Available Formats

University of California

0.384RR + 99 RR < 600ms

with maximum ow denoted

. This contradicts the optimality of min-cost ow technique

cannot exist and our formulation minimizes the

which was transformed from L and the ows of two partitions

are the vectors containing the ow of all nodes in de-

[i], for i = 1 . . . k. If F[k + 1] < F

[k + 1] then because of the special

[k + 1] itself would be larger than

You might also like