NSEL Report Series

Report No. NSEL-001

November 2007

Structural Health Monitoring

Using Smart Sensors

Tomonori Nagayama
and
Billie F. Spencer, Jr.

NEWMARK STRUCTURAL ENGINEERING LABORATORY

Department of Civil and Environmental Engineering

University of Illinois at Urbana-Champaign
 UILU-ENG-2007-1801

ISSN: 1940-9826

© The Newmark Structural Engineering Laboratory

The Newmark Structural Engineering Laboratory (NSEL) of the Department of Civil and
Environmental Engineering at the University of Illinois at Urbana-Champaign has a long history
of excellence in research and education that has contributed greatly to the state-of-the-art in civil
engineering. Completed in 1967 and extended in 1971, the structural testing area of the
laboratory has a versatile strong-floor/wall and a three-story clear height that can be used to carry
out a wide range of tests of building materials, models, and structural systems. The laboratory is
named for Dr. Nathan M. Newmark, an internationally known educator and engineer, who was
the Head of the Department of Civil Engineering at the University of Illinois [1956-73] and the
Chair of the Digital Computing Laboratory [1947-57]. He developed simple, yet powerful and
widely used, methods for analyzing complex structures and assemblages subjected to a variety of
static, dynamic, blast, and earthquake loadings. Dr. Newmark received numerous honors and
awards for his achievements, including the prestigious National Medal of Science awarded in
1968 by President Lyndon B. Johnson. He was also one of the founding members of the
National Academy of Engineering.

Contact:
Prof. B.F. Spencer, Jr.
Director, Newmark Structural Engineering Laboratory
2213 NCEL, MC-250
205 North Mathews Ave.
Urbana, IL 61801
Telephone (217) 333-8630
E-mail: [email protected]

This technical report is based on the first author's doctoral dissertation under the same title
which was completed in April 2007. The second author served as the dissertation advisor for
this work.

Financial support for this research was provided in part by the National Science Foundation
(NSF) under NSF grants CMS 03-01140 and CMS 06-00433 (Dr. S. C. Liu, Program Manager).
The first author was supported by a Vodafone-U.S. Foundation Graduate Fellowship. This
support is gratefully acknowledged.

The cover photographs are used with permission. The Trans-Alaska Pipeline photograph was
provided by Terra Galleria Photography (http://www.terragalleria.com/).
 ABSTRACT

Industrialized nations have a huge investment in the pervasive civil infrastructure on

which our lives rely. To properly manage this infrastructure, its condition or serviceability
should be reliably assessed. For condition or serviceability assessment, Structural Health
Monitoring (SHM) has been considered to provide information on the current state of
structures by measuring structural vibration responses and other physical phenomena and
conditions. Civil infrastructure is typically large-scale, exhibiting a wide variety of
complex behavior; estimation of a structure's state is a challenging task. While SHM has
been and still is intensively researched, further efforts are required to provide efficient and
effective management of civil infrastructure.
Smart sensors, with their on-board computational and communication capabilities,
offer new opportunities for SHM. Without the need for power or communication cables,
installation cost can be brought down drastically. Smart sensors will help to make
monitoring of structures with a dense array of sensors economically practical. Densely
installed smart sensors are expected to be rich information sources for SHM.
Efforts toward realization of SHM systems using smart sensors, however, have not
resulted in full-fledged applications. All efforts to date have encountered difficulties
originating from limited resources on smart sensors (e.g., small memory size, small
communication throughput, limited speed of the CPU, etc.). To realize an SHM system
employing smart sensors, this system needs to be designed considering both the
characteristics of the smart sensor and the structures to be monitored.
This research addresses issues in smart sensor usages in SHM applications and
realizes, for the first time, a scalable and extensible SHM system using smart sensors. The
architecture of the proposed SHM is first presented. The Intel Imote2 equipped with an
accelerometer sensor board is chosen as the smart sensor platform to demonstrate the
efficacy of this architecture. Middleware services such as model-based data aggregation,
reliable communication, and synchronized sensing are developed. SHM Algorithms
identified as promising for the usage on smart sensor systems are extended to improve
practicability and implemented on Imote2s. Careful attention has been paid to integrating
these software components so that the system possesses identified desirable features.
The damage detection capability and autonomous operation of the developed system
are then experimentally verified. The SHM system consisting of ten Imote2s are installed
on a scale-model truss. The SHM system monitors the truss in a distributed manner to
localize simulated damage.
In summary, this report proposes and realizes a scalable and autonomous SHM system
using smart sensors. The system is experimentally verified to be effective for damage
detection. The autonomous nature of the system is also demonstrated. Successful
completion of this research paves the way toward full-fledged SHM systems employing a
dense array of smart sensors. The software developed under this research effort is open-
source and is available at: http://shm.cs.uiuc.edu/.
 Contents

Page

CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Monitoring of civil infrastructure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 SHM using smart sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Overview of report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
CHAPTER 2 BACKGROUND . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Smart sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Smart sensor’s essential features . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Smart sensors to date . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Middleware services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Structural Health Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Attempts toward SHM using smart sensors . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.1 Research attempts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4.2 Difficulties in using smart sensors for SHM applications . . . . . . . 23
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
CHAPTER 3 SHM ARCHITECTURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1 Desirable characteristics of an SHM system employing smart sensors . . . . . 28
3.2 SHM system architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Network system architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.2 Smart sensor platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.3 Middleware services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.4 Damage detection algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
CHAPTER 4 SENSOR BOARD CUSTOMIZATION . . . . . . . . . . . . . . . . . . . . . . . 41
4.1 Strain sensor board development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2 AA filter board development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3 Experimental verification of strain sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
CHAPTER 5 MIDDLEWARE SERVICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1 Data aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.1.1 Estimate on data amount in SHM applications . . . . . . . . . . . . . . . . 51
5.1.2 Model-based data aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.1.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Reliable communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2.1 The effects of data loss on SHM applications . . . . . . . . . . . . . . . . 57
5.2.2 Packet loss estimation in RF communication . . . . . . . . . . . . . . . . . 61
5.2.3 Reliable communication protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 62
 5.3 Synchronized sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3.1 Time synchronization effect on SHM applications . . . . . . . . . . . . 74
5.3.2 Estimation on time synchronization error . . . . . . . . . . . . . . . . . . . . 77
5.3.3 Issues toward synchronized sensing . . . . . . . . . . . . . . . . . . . . . . . . 79
5.3.4 Realization of synchronized sensing . . . . . . . . . . . . . . . . . . . . . . . . 82
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
CHAPTER 6 ALGORITHMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.1 Natural Excitation Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.2 Eigensystem Realization Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.3 Damage Locating Vector method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.4 Distributed Computing Strategy for SHM . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.5 Stochastic Damage Locating Vector method . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.6 Extension of DCS for SHM with the SDLV method . . . . . . . . . . . . . . . . . . 101
6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
CHAPTER 7 REALIZATION OF DCS FOR SHM . . . . . . . . . . . . . . . . . . . . . . . 103
7.1 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.1.1 Fast Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.1.2 Singular value decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.1.3 Eigensolver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.1.4 Complex matrix inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.1.5 Sort . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.2 DCS implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.2.1 Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.2.2 NExT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.2.3 ERA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.2.4 DLV methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
7.2.5 DCS logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.2.6 Final implementation on the Imote2 . . . . . . . . . . . . . . . . . . . . . . . 126
7.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
CHAPTER 8 EXPERIMENTAL VERIFICATION . . . . . . . . . . . . . . . . . . . . . . . 138
8.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.2 NExT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
8.3 ERA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.4 DLV methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.5 DCS logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.6 Calculation and communication time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
8.7 Battery life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.8 Damage detection results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
CHAPTER 9 CONCLUSIONS AND FUTURE STUDIES . . . . . . . . . . . . . . . . . . 159
9.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
9.2 Future studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.2.1 Sensing capability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.2.2 Damage detection capability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
 9.2.3 Power harvesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
9.2.4 Power management and scheduling . . . . . . . . . . . . . . . . . . . . . . . 164
9.2.5 Monitoring of occasional events such as earthquakes . . . . . . . . . 164
9.2.6 Multihop . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
9.2.7 Communication range adjustment . . . . . . . . . . . . . . . . . . . . . . . . 165
9.2.8 Reliability of the system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
9.2.9 Environmental hardening . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
9.2.10 Multiple purpose usage of smart sensors . . . . . . . . . . . . . . . . . . 165
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
 Chapter 1

INTRODUCTION

1.1 Monitoring of civil infrastructure

Our lives rely heavily on the pervasive civil infrastructure in which industrialized
nations have huge investments. Malfunctioning of civil infrastructure has caused
tremendous economic loss and claimed numerous human lives. Civil infrastructure is,
thus, critical to keep our economy running, while the infrastructure itself is an important
asset to be managed.
To properly manage civil infrastructure, its condition, or serviceability, must be
assessed. Many variables can be monitored and used for the assessment. For instance,
Intelligent Transportation Systems make use of traffic surveillance information to
efficiently manage the transportation system. Tunnels are monitored for traffic accidents
and air quality. The Urgent Earthquake Detection and Alarm System (Nakamura, 2004)
detects primary seismic waves and stops trains before severe secondary waves approach.
Measurement and proper data processing are expected to give a reasonable assessment of
serviceability that can then be improved based on the assessment.
The physical state of a structural system, for example, applied load, vibration level,
and existence of structural damage, is among the factors that determine serviceability.
Sensing physical quantities in detail offers the potential to better estimate structural
conditions. For river bank protection, for instance, water level may be monitored and the
associated load estimated. Precipitation rate and groundwater level are important
indicators to predict slope failure. Strain and temperature measurements can be utilized to
monitor concrete gravity or arch dams. Engineers, owners, and users can make better
decisions based on the measured information.
Structural condition assessment is, however, not always straightforward as in the case
of the Structural Health Monitoring (SHM) of buildings, bridges, and towers. The
structural condition is oftentimes sought in terms of structural characteristics, i.e., mass,
damping, stiffness matrices, damage existence, and/or applied load to the system. These
structures are large and consist of many members, which makes such structural condition
assessment difficult and/or prohibitively expensive. One approach in SHM to alleviate this
difficulty is based on vibration measurement. Though structural characteristics and
applied load are difficult to assess directly, dynamic behavior, which is a function of the
structural characteristics and applied load, can be measured. The structural characteristics
and applied load information lurk in the dynamic behavior. Structural soundness is
expected to be estimated by inverse analyses of the dynamic behavior.
Because buildings, bridges, and towers are typically large and complex, information
from just a few sensors is inadequate to accurately assess the structural condition. The
dynamic behavior of these structures is complex in both spatial and time scale. Moreover,

1
damage/deterioration is intrinsically a local phenomenon. Therefore, to comprehend such
dynamic behavior, the motion of structures needs to be monitored by densely located
sensors at a sampling frequency sufficiently high to capture salient dynamic
characteristics.

1.2 SHM using smart sensors

When many sensors are implemented, wireless communication appears to be
attractive. The high cost associated with the installation of wired sensors (Celebi, 2002;
Farrar, 2001) can be greatly reduced by employing wireless sensors. Wireless sensors
often convert analog signals to digital signals prior to radio frequency (RF) transmission,
while many wired systems collect analog signals at one or several base stations where the
signals conversion takes place. The digital conversion on the wireless sensor node
eliminates possible signal degradation during analog signal communication through long
cables. Wireless sensor systems are, thus, promising as data acquisition systems with a
large number of sensors installed on sizable structures.
Being “smart”, i.e., having data processing capability in the sensors, is an essential
feature that further increases the potential of wireless sensors. Smart sensors can locally
process measured data and transmit only the important information through wireless
communication. As a network, wireless sensors extend the capability. For instance,
sensors that are malfunctioning in the network can be detected, and other sensors can
rebuild sensor topology without this dead node. As another instance, location mapping can
be done automatically by a localization service (Doherty et al., 2001; Kwon et al., 2005a;
Kwon et al., 2005b), which helps civil engineers determine and confirm the location of
large numbers of sensors on complex structures.
Smart sensors, however, have limited resources, prohibiting direct application of
traditional structural monitoring strategies. For example, the communication speed is too
slow to centrally collect all of the measured information. Clocks on sensor nodes are not
always synchronized. Some communication packets may be lost. Storage and memory
space is limited. Processor speed is slower than that of a PC. Smart sensors do not
necessarily offer a real-time system; programmers may not be able to assign appropriate
priority to given tasks. Moreover, battery power imposes limitations on many aspects of
smart sensors. Any task consuming large amounts of power becomes impractical on a
battery-operated smart sensor node. Smart sensor systems need to overcome these
limitations using deliberate system design, as seen in some of the time synchronization
and reliable communication research efforts (Ganeriwal et al., 2003; Maroti et al. 2004;
Mechitov et al., 2004).
From the perspective of SHM, being smart makes it feasible to monitor structural
response densely both in time and space. The amount of data generated from a monitored
structure can be enormous due to the large number of sensors and high sampling
frequency. For example, the Tsing Ma and Kap Shui Mun Bridges in Hong Kong produce
63 MB of data every hour (Wong, 2004). Being smart is expected to allow significant data
compression at the node level by extracting only the information necessary for the task at

2
hand, thus reducing the amount of data to be stored or transferred through wireless
communication.
Furthermore, being smart gives the possibility of autonomous structural health
monitoring, with reduced user interaction. Smart sensors communicate with each other
through the RF link to share measurement data. The data can be utilized to judge structural
soundness. Once the smart sensor network detects structural damage, the network informs
users about the damage or necessary repair. Microprocessors on the smart sensors make it
possible to perform this procedure autonomously.
Many smart sensor prototypes have been developed and several attempts to use smart
sensors for SHM are reported so far. There are, however, many problems to be solved.
Ruiz-Sandoval (2004) addressed some sensor hardware problems from a civil engineering
viewpoint. As for algorithms, most attempts at SHM with smart sensors only substitute the
wired link with wireless communication and apply traditional damage detection
algorithms at the base station. An SHM system with such algorithms assuming central
data collection does not scale to a large number of smart sensors. Researchers have also
proposed approaches where simple data processing is performed on each smart sensor
node without interaction with other nodes; these attempts do not employ spatial
information, and, therefore, have room for improvement in terms of damage detection
capability. Information from a dense array of smart sensors should be processed in a
coordinated manner, rather than independently. SHM algorithms for distributed and
coordinated data processing making use of the smart sensor's distributed computing and
sensing resources have only recently appeared.
Gao (2005) recently proposed a new distributed computing strategy (DCS) for SHM
envisioning smart sensor usage. DCS is intended to use the smart sensor's data processing
capability in a coordinated way to achieve SHM. Computer analysis and experimental
validation on a simulated wireless network showed DCS for SHM is promising.
Implementation of DCS on smart sensors by addressing implementation issues and
experimental validation of the proposed scheme are imperative.

1.3 Overview of report

This research focuses on the realization of a vibration-based SHM framework
employing smart sensors that can detect and localize damage. In general terms, damage
can be defined as changes introduced into a system that adversely affect the current or
future performance of that system (Doebling et al., 1998). The effect of damage on a
structure can be classified as linear or nonlinear. A linear damage situation is defined as
the case when the structure remains linear-elastic after damage, while damage introducing
nonlinear behavior is defined as nonlinear damage (Doebling et al., 1996). The type of
damage considered in this research is linear damage. Metal corrosion, concrete spalling/
scour, and yielding of beam-column joints are typical linear damage examples of interest
to civil engineers. Linear damage changes structural characteristics such as mass,
damping, and stiffness. Subsequent changes in modal parameters are identified and
utilized in damage detection. A framework for such a vibration-based SHM employing
smart sensors is proposed herein. Among the important features to be obtained are

3
scalability to a large number of smart sensors and autonomous operation, as well as
effective damage detection capability.
Chapter 2 provides the background of this research. Smart sensors, middleware
services, and SHM are briefly reviewed. Research efforts employing smart sensors for
SHM applications are then summarized and difficulties in these attempts are addressed. In
the subsequent chapters, an SHM framework is realized on a smart sensor network that
resolves the majority of these difficulties.
Chapter 3 describes the SHM architecture developed in this research. A homogeneous
hardware configuration is selected, while smart sensor nodes are functionally
differentiated into several categories. Smart sensors, middleware services, and damage
detection algorithms used in such networks are briefly explained.
In Chapter 4, sensor boards for one of the smart sensor platforms, the Mica2, are
developed to demonstrate sensor board customizability according to SHM requirements.
Because strain sensors for the Mica2 were not available, a strain sensor board is developed
as well as an Anti-Aliasing (AA) filter board. Scale-model experiments show that these
sensor boards can facilitate accurate measurement of structural responses.
Middleware services realized as part of this research for SHM applications are
discussed in Chapter 5. Middleware services include reliable communication, model-
based data aggregation, and synchronized sensing. These middleware services can be used
in a wide variety of civil engineering applications.
Chapter 6 discusses algorithms to be implemented on smart sensors. The DCS
algorithm has the potential to realize densely deployed smart sensor networks for SHM
because of its distributed and coordinated data processing. Algorithmic components of
DCS are briefly reviewed. The damage detection algorithm in DCS is then extended by
replacing the mass perturbation Damage Locating Vector (DLV) method with the
Stochastic Damage Locating Vector (SDLV) method. The SDLV method is shown to
simplify damage detection and reduce total power consumption. This chapter provides the
algorithmic basis for the subsequent chapters.
In Chapter 7, the DCS algorithm is implemented on the Imote2 smart sensor platform
using the middleware services and algorithms. First, numerical functions are ported to the
Imote2. Second, the capabilities of the generic sensor board are examined. Third, each of
the DCS algorithms is implemented on smart sensors, and its validity is numerically
investigated.
Chapter 8 describes experimental verification of the developed framework. Smart
sensor nodes are placed on a scale-model, three-dimensional truss. One of the bar
elements of the truss is replaced with a more slender element to simulate linear damage to
the truss. The smart sensor system measures acceleration responses of the model and
localizes damage. Calculation and communication time, the battery life, and damage
detection capability are discussed based on findings from the experiments.
Chapter 9 summarizes the research detailed in this report and presents possible
directions for future research on SHM using smart sensors.

4
 Chapter 2

BACKGROUND

This chapter presents the background for this research. Realization of an SHM
framework using smart sensors requires interdisciplinary studies. The pursuit of a single
component of the system does not necessarily realize the framework. Technical
background and research efforts on three subjects closely associated with SHM using
smart sensors, i.e., smart sensors, middleware services, and SHM algorithms, are
overviewed. Research efforts directed toward SHM using smart sensors are then
reviewed, and difficulties encountered in these attempts are summarized.

2.1 Smart sensor

2.1.1 Smart sensor’s essential features

Smart sensor technology has been under rapid development in recent years. A smart
sensor usually has five essential features: 1. on-board microprocessor, 2. sensing
capability, 3. wireless communication, 4. battery-powered, and 5. low cost. This section
describes each of these features in detail.

1. On-board microprocessor

The essential difference between a smart sensor and a standard integrated sensor is its
intelligence, i.e., the on-board microprocessor. Programs can be embedded in the
microprocessor, which allows smart sensors to save data locally, perform desired
computations, make “if-then” decisions, scan necessary information, send results quickly,
schedule multiple tasks, coordinate with surrounding sensors, etc. The on-board
microprocessor can also control the time and duration that the sensor will be fully awake
in order to efficiently manage power consumption. The smart sensors can arrange
autonomous networks to achieve multiple tasks, such as SHM, power saving, multihop
communication, self-configuration and self-healing of the network, dynamic routing, etc.
There are several options for embedding intelligence in a hardware design (Texas
Instruments, 2007). Among the choices are: a digital signal processor (DSP), a Field
Programmable Gate-Array (FPGA), an Application-Specific Integrated Circuit (ASIC),
and a general purpose processor (GPP). Compared to the generally designed
microprocessor, a DSP is specifically designed for rapid signal processing, such as real-
time processing of audio signals on cell phones, often using an optimized instruction set
(ISA). A multiply-accumulate (MAC) operation, which is suitable for matrix operation,
such as convolution for filtering, dot product, or even polynomial evaluation, is commonly
implemented on DSP chips. An FPGA has the capability of being reconfigurable within a
system and offers greater raw performance per specific operation because of its dedicated

5
logic circuit. However, FPGAs are more expensive and typically have higher power
dissipation than DSPs and GPPs. ASICs can be tailored to perform specific functions
extremely well, and can be made quite power efficient. However, since ASICS are not
field-programmable, their functionality is not easily changed or updated. Also, without
mass production, the initial setup costs are prohibitively high. Smart sensor applications
require multiple tasks, such as wireless communication, data logging, data acquisition,
and diverse signal processing. Among the above-mentioned options, GPPs are best suited
for performing a broad array of tasks. When tasks on smart sensors include intensive data
processing, a DSP or an FPGA can be used in combination with a GPP to improve data
processing performance. Once smart sensor applications are well-understood, the
necessary functions may be refined and ASIC may be implemented to perform specific
tasks. Smart sensors on the market targeting a broad range of applications, in most cases,
utilize only a GPP. If deemed advantageous, smart sensor systems developed on a GPP
can then be improved by utilizing special purpose microprocessors such as DSP, FPGA,
and ASIC.

2. Sensing capability

A smart sensor is able to convert the physical state of an object or environment such
as temperature, light, sound, and/or motion into electrical or other types of signals that
may be further processed. A single smart sensor node may have several sensors measuring
different physical quantities. Micro-Electro-Mechanical System (MEMS) devices, which
are the integration of mechanical elements, sensors, actuators, and electronics on a
common silicon substrate through microfabrication technology, are often employed for
sensors because of their small size, inexpensive cost (when mass produced), and low
power consumption. Data acquisition parameters, such as sampling frequency and data
length, can be controlled by the on-board processor. The on-board microprocessor can
also access and process the acquired data. The sensing capability provides the interface
between the smart sensor's on-board microprocessor and real-world physical phenomena.
One of the promising technologies in sensors is the use of quasi-digital sensors, such
as frequency, duty-cycle, and pulse number output sensors, as proposed by Kirianaki et al.
(2000) and Yurish (2005). Low noise sensitivity, wide dynamic range, and a simple
interface which is directly connected to microprocessors are advantages of the quasi-
digital sensors. The use of quasi-digital sensors is increasing. For example, the Intel Mote
developed by Intel Corporation (Kling, 2004) utilizes a duty cycle output accelerometer.

3. Wireless communication

Smart sensors communicate with each other through a wireless link. While RF
communication is most widely used, smart sensors with different transmission media,
such as acoustic, laser, and infrared transmission, have also been studied (Hollar, 2000).
Acoustic communication utilizes a sounder and a microphone. The microphone requires
little power, while the sounder power is comparable to low power RF devices; these
features are very attractive from a power saving perspective. Some disadvantages in
acoustic communication include the following: the surrounding background noise limits

6
the communication quality, and the atmosphere attenuates the signal power over large
distances. Moreover, its slow wave propagation speed becomes a serious problem for
timing-critical applications. Laser communication can be classified as either passive or
active communication, according to whether the laser signal is generated through a
secondary source or not. Passive devices receive signals through a laser, and send back
data depending on the received command by modulating mirrors located on the passive
device. Without the necessity of generating a laser signal, the passive device consumes a
small amount of power. Active laser communication points an on-board laser transmitter
toward a receiver and transmits data. Long communication ranges, up to 21.4 km (Hollar,
2000), have been reported. Because of the short wavelength, i.e., 100 to 350 nm for lasers
as compared to 0.01 to 30 m for RF, and the resultant small divergence, information can
be transmitted at lower power than RF. However, searching for the correct receiver
direction can take a significant amount of time, and the point-to-point communication is
fixed unless the transmitter is redirected through a time- and power-consuming receiver
search. Infrared communication uses infrared light with wave lengths ranging from 700
nm to 1 mm. IrDA, an infrared wireless communication technology, consumes
significantly less power than RF transceivers (Miller, 2001). The IrDA, however, uses a
relatively narrow focused beam which usually requires that the transmitter and the
receiver be carefully aligned with each other.
Most of the smart sensors to date employ RF-based wireless communication. As
opposed to the laser and infrared communication, RF communication is omni-directional,
and does not need direct line-of-sight. In the United States, the Federal Communications
Commission (FCC) has allocated the three RF bands around 900 MHz, 2.4 GHz, and 5.0
GHz (0.33 m, 0.125 m, and 0.06 m, respectively, in terms of wavelength) as unlicensed
industrial, scientific, and medical (ISM) bands. Many of today's wireless technologies
operate on these ISM bands. Small RF chips, which require low power and have few
external components, are available from many suppliers. For example, the CC1000 from
Chipcon, Inc. (2007) consumes only 8.6-25.4 mA for transmission and 9.6 mA for
reception at 868 MHz with a 3 V supply voltage. However, RF chip communication speed
is typically limited; the throughput of the CC1000 is only 38.4 kbps. Omni-directional
communication with moderate power consumption suits a dense array of smart sensors.
RF communication in the ISM band has often been adapted for applications needing small
amounts of data transfer such as temperature monitoring.
The RF link also has a limited communication range. As the distance between
transmitter and receiver gets longer, more data is lost. Communication between two nodes
can be achieved by either home runs, where two nodes communicate with each other
directly, or hopping, where nodes between two communication endpoints work as relay
nodes.
For communication in a network of a large number of smart sensors, hopping
communication has advantages in terms of power. Power requirements for RF
communication are proportional to the distance raised to some power. Depending on
interference effects, the power requirement could be proportional to the fifth power of
distance (Zhao & Guibas, 2004). Also, maximum transmission power for the ISM bands is
restricted to be less than 1 W by FCC, limiting the communication range. For these

7
reasons, smart sensors spatially distributed over large areas, as is the case for SHM of
bridges and buildings, need to communicate with each other through hopping. However,
this simple analysis on power consumption does not include all the communication related
tasks affecting the total power consumption in a system. For example, waiting for possible
incoming messages in a listening mode also consumes power. Additionally, latency and
robustness concerns accompany a large number of relay nodes. Optimal design for
communication range is sought with these considerations.
Such RF communication links result in the following characteristics of smart sensors:
(a) Wireless communication is not as fast as wired communication; (b) Data loss is
inherent to exchange of communication packets; and (c) The network route is not
physically fixed, which enables self-configuration and self-healing of networks, and
dynamic routing with the help of the on-board microprocessor.
Several standards have provided RF communication specifications. For example, the
Bluetooth standard provides the radio link, baseband link, and the link manager protocol
(the Open Systems Interconnection reference model; Zimmermann, 1980) for low-power
Wireless Personal-Area Networking (WPAN). The expected application is short-range,
low-power voice and data communication, such as Personal Digital Assistant (PDA),
mobile phones, and laptops. Bluetooth is based on proximity networking at 2.4 GHz, with
an expected communication range of around 100 m at an output power of 100 mW,
although 1 and 2.5 mW specifications are also available. Data throughput is 723.1 Kbps.
IEEE 802.11 is a Wireless Local Area Networking (WLAN) specification. The
specification includes robust, Ethernet-style data networking components, such as a
Transmission Control Protocol/Internet Protocol (TCP/IP) stack. At the expense of power
consumption, this technology possesses communication ranges larger than 100 m and a
throughput of up to 54 Mbps. For example, an 802.11b compatible transceiver chip,
MAX2820 from Maxim Integrated Products, Inc. (2007), consumes more than 200 mW,
while other external components also draw current. IEEE 802.11 technology is well-suited
for devices with stable power supplies and high data rate requirements. The IEEE
802.15.4 standard defines the physical and Media Access Control (MAC) layer protocols
for WPAN with low data rates but needing a very long battery life. The data rate is 250
Kbps. ZigBee is an industry consortium to promote this standard. The CC2420 from
Chipcon, Inc. implements the IEEE 802.15.4 communication standard, consuming only
8.5-17.4 mA for transmission and 19.7 mA for reception.
These RF communication standards bring significant benefit to smart sensor
applications, though some applications may need customization. By using the standards,
users can avoid the necessity to design all the layers of the OSI reference model
(Zimmermann, 1980) from scratch. Users can choose a standard which best fits each
application. Also the standards allow independently developed systems to communicate
with each other easily. The available standards, however, do not necessarily offer the best
RF solution to a specific smart sensor application. Users may want to customize the
standards if available standards do not offer a suitable solution at the expense of losing
compatibility with other systems.

8
4. Battery powered

Smart sensors are frequently powered by local batteries, particularly when other
power sources are not easily reached. Even when power is available, such as in a building,
tethering a large number of smart sensors to power sources is expensive and reduces the
merits of wireless communication. There have been several attempts to harvest energy at
sensor nodes locally, for example, from mechanical vibration (Rahimi et al., 2003).
Vibration energy for civil infrastructure is, however, typically contained in the low
frequency range, where electric energy is difficult to harvest. Currently available smart
sensors rely on the local battery power supply, which has finite capacity and finite life.
Because replacing or recharging batteries in smart sensors installed on a structure may not
be a trivial task, power saving is a major concern. This finite energy source problem
imposes a strict constraint.

5. Low cost

Smart sensors primarily composed of MEMS and other integrated circuits (IC) have
the potential to be produced at a low cost as well as being small in size. This feature of
smart sensors, combined with inexpensive installation cost due to wireless
communication, enables numerous smart sensors to be densely distributed over civil
infrastructure, thus offering the potential to capture the structure's state in detail, and
drawing us closer to realizing the dream of ubiquitous sensing.

2.1.2 Smart sensors to date

Some of the first efforts in developing smart sensors for application to civil
engineering structures were presented by Straser and Kiremidjian (1996, 1998) and
Kiremidjian et al. (1997). In their work, a wireless sensor unit was constructed consisting
of a microprocessor, radio modem, data storage, and batteries. To reduce the battery
consumption, the smart sensor could be either in a waiting mode or an operational mode.
Since these first efforts, numerous researchers have developed smart sensing platforms.
Lynch and Loh (2006) cited over 150 papers on wireless sensor networks for SHM
conducted at over 50 research institutes worldwide. These platforms can be grouped into
two primary categories: proprietary and nonproprietary
Many proprietary platforms have been developed by individual research groups,
including the work of Straser and Kiremidjian. Other examples include Mason et al.
(1995), Bult et al. (1996), Agre et al. (1999), Aoki et al. (2003), Basheer et al. (2003),
Kawahara et al. (2003), Kottapalli et al. (2003), Shinozuka (2003), Wang et al. (2003),
Casciati et al. (2004), Sazonov et al. (2004), Farrar et al. (2005), and Lynch (2006), all of
whom designed their smart sensor hardware using commercial-off-the-shelf (COTS)
components. Each of these papers cited shortcomings in previously developed smart
sensors as motivation for the development of their own hardware. While the individual
researchers advanced the state-of-the-art in smart sensor technology, progress was slow
due to the lack of coordination and leveraging of efforts.

9
Figure 2.1. EYES project sensor prototype.

Coordinated research efforts, such as was done in the EYES project on self-
organizing and collaborative energy efficient sensor networks (EYES, 2006; Law et al.,
2002) out of Europe, leveraged individual contributions. The EYES project aimed to
develop architecture and technology that will enable the creation of a new generation of
self-organizing and collaborative sensors. These sensors will be capable of effectively
networking together, to provide a flexible platform to support a large variety of mobile
sensor network applications. EYES will make use of the effort invested in the DataGrid
project (DataGrid Project, 2006), which is to build the next generation of computing
infrastructure, providing intensive computation and analysis of shared large-scale
databases. This EYES project includes more than 12 WorkPackages (WP) that deal with
middleware, applications, and management. The architecture of EYES has two levels. The
first level deals with the sensors and the network, i.e., internal sensor architecture,
distributed wireless access, routing protocols, reliable end-to-end transport,
synchronization and localization of nodes. The second level provides distributed services
to the application, deals with information collection, lookup, discovery and security.
Figure 2.1 shows a sensor prototype from the EYES project. Though this coordinated
effort provides a more effective framework than the individual approaches, the work is
still proprietary in nature, limiting the technical information available to the public.
Commercialized smart sensors also offer the possibility of leveraging contributions
among a large number of users. The availability of wireless sensor hardware, especially to
application researchers who do not specialize in hardware design, has prompted smart
sensor applications in various areas (e.g., machinery monitoring, building HVAC control,
etc.). Companies, such as Dust Networks (Dust Networks, 2007), Microstrain
(Microstrain, Inc. 2007), Millennial Net (Millennial Net, 2007), Sensametrics
(Sensametrics, 2007), and Sensicast Systems (Sensicast Systems, 2007) have designed
their own hardware, middleware, and application software, and provide associated
support. These sensors tend to emulate wired sensors where raw data from sensor nodes is
transmitted to the base station, which is not scalable to a large number of smart sensors
due to the limited bandwidth. The lack of scalability becomes substantial, especially when
data is acquired at a high sampling ratio, as is the case for civil engineering applications.

10
The specialized and proprietary nature of the commercial smart sensors has inhibited free
hardware/software development for civil engineering applications.
The first available open hardware/software research platform, which allows users to
customize hardware/software for a particular application, was the Berkeley Mote. Under
substantial support of the U.S. Defense Advanced Research Projects Agency (DARPA),
researchers at the University of California at Berkeley have developed the open platform.
The main objective is to create massively distributed sensor networks, which consist of
hundreds or thousands of sensor nodes; these nodes have been termed as Smart Dust or
Mote. The goal is to have fully autonomous sensor nodes that are a cubic millimeter in
size.
A family of Berkeley Mote hardware has been developed since the first Mote, termed
COTS Dust (Hollar, 2000), arrived, incorporating communications, processing, sensors,
and batteries into a package about a cubic inch in size. The generation of the Mote
following COTS Dust was the Rene, developed in the summer of 2000. The third
generation of Mote, the Mica, was released in 2001 (Hill & Culler, 2002), having
improved memory capacity and using a faster microprocessor (Atmega128L, 4 MHz).
Subsequent improvements to the Mica platform resulted in the Mica2, Mica2Dot, and
MicaZ. Sensors for the Micas are designed on stackable sensor boards separated from the
main radio/processor board, so that users can switch between several sensor boards
according to their application. The next generation of the Berkeley Mote is the Telos
platform, which seeks to achieve three major goals: ultra-low power operation compared
with previous Motes, ease of use, and robust hardware and software implementation
(Polastre et al., 2005). The Berkeley Mote hardware schematic is publicly available;
researchers can follow and customize the design, and expand its sensing modality with
their own sensor boards. Moreover, these sensors have been made commercially available
through Crossbow Technology, Inc. (2007). The hardware specification of representative
Berkeley Motes is summarized in Table 2.1. Note that the limited radio bandwidth, small
RAM size, and low resolution Analog-to-Digital Converter (ADC) of these Motes, as well
as the modest microprocessor power can be a restraining factor for intensive sensing/
calculation applications like SHM.

11
 Table 2.1. Smart Sensor Specifications
(The power consumption is estimated based on associated components’ data sheet.)

MicaZ Mica2 Mica2dot Telos Imote2

Microprocessor ATmega128L ATmega128L ATmega128L TIMSP430 XScalePXA271
Bus size (bits) 8 8 8 16 32
Clock speed (MHz) 7.373 7.373 4 6.717 13-416
Program flash (bytes) 128 k 128 k 128 k 48 k 32 M
EEPROM (bytes) 4k 4k 4k
256 k+ 32 M
RAM (bytes) 4k 4k 4k 10 k
external
44 @ 13 MHz
Active power (mW) 24 @ 3 V 24 @ 3 V 15 @ 3 V 10 @3 V
570 @ 416 MHz

12
Sleep power (PW) 75 75 75 8 100
ADC resolution (bits) 10 10 10 12 N/A
ADC channels 8 8 6 8 N/A
32 M
Nonvolatile storage (bytes) 512 k 512 k 512 k 1,024 k (Identical to the
program flash)
Memory write power (mW) 45 45 45 60 120
Memory read power (mW) 12 12 12 12 108
 Table 2.1. Smart Sensor Specifications (Continued)
(The power consumption is estimated based on associated components’ data sheet.)

MicaZ Mica2 Mica2dot Telos Imote2

Radio Chip CC2420 CC1000 CC1000 CC2420 CC2420
Radio frequency (MHz) 2400-2483.5 315/433/915 315/433/915 2400-2483.5 2400-2483.5
Data rate (kbps) 250 38.4 38.4 250 250
Outdoor range (m) 100 300/300/150 300/300/150 100 100
Transmission power at 1 mW
52 31 31 52 52
power (mW)
Reception power (mW) 59 22 22 59 59
Typical battery option 2 x AA 2 x AA CR2354 2 x AA 3 x AAA

13
Typical capacity (mAh) 2 x 1500-2000 2 x 1500-2000 560 2 x 1500-2000 3 x 700
Typical supply voltage 3 3 3 3 4.5
Size (mm) 58 x 32 x 7 58 x 32 x 7 25 x 6 65 x 31 x 6 48 x 36 x 7
Light Light
Temperature Temperature
Humidity Humidity
Barometric Barometric Light
Light Light
pressure pressure Temperature
Available Sensor Temperature Temperature
Accelerometer Accelerometer Humidity
Accelerometer Humidity
GPS GPS Accelerometer
Microphone Microphone
Sounder Sounder
Magnetometer Magnetometer
 While a number of sensor boards for the Berkeley Mote have been designed for
generic applications and are commercially available (see Table 2.1), available sensors are
not optimized for use in civil infrastructure applications. Acceleration and strain are
among the most important physical quantities to judge the health of a structure. While
acceleration measurements are essential to obtain global responses of a structure,
structural strains provide an important indicator of local structural behavior. Studer and
Peters (2004) demonstrated that multiscale sensing yields better results than single-scale
measurements for damage identification. Available sensor boards have accelerometers,
though their applicability to civil infrastructure is limited. Ruiz-Sandoval (2004)
examined an acceleration sensor board, the MTS310CS, from the civil engineering
perspective; he found the accelerometer's performance is deficient, particularly in the low
frequency range, and developed a new sensor board, 'Tadeo,' with high sensitivity and low
noise level. The sensor boards of the Berkeley Mote can be conveniently customized
according to the applications.
The Berkeley Mote uses a open-source operating system, TinyOS, which is available
online (http://www.tinyos.net). TinyOS is a component-based operating system designed
for sensor network applications on resource-constrained hardware platforms, such as the
Berkeley Mote. More specifically, it is designed to support concurrency intensive
operations such as are required by networked sensors with minimal hardware
requirements; TinyOS fits in about 4 kB of memory space. TinyOS is, however, not a real-
time operating system. TinyOS has only two levels of priority, each of which corresponds
to hardware interrupt and tasks. Users cannot assign priority arbitrarily to specific tasks or
assume highly precise scheduling. This limitation needs to be well considered in designing
a system using TinyOS.
A wide community uses TinyOS to develop and test various algorithms and
protocols. The operating system has been ported to over a dozen platforms employing
numerous sensor boards. The autonomous characteristics of the smart sensor can be
realized by developing programs under TinyOS and then running these programs, along
with the operating system, on the on-board microprocessor. Similar software for the
Berkeley Mote and other platforms includes MANTIS (Bhatti et al., 2005) and EmStar
(Girod et al., 2004). Because of the open nature, numerous researchers have contributed to
TinyOS enhancement; users can take advantage of the wealth of previous studies.
Open-interface smart sensor platforms developed by companies, such as Ember
(2007) and Intel (2007), have similar potential. Users have access to the programming
interface and can program the sensor nodes as needed. Ember nodes must be programmed
on their own OS, but users can implement necessary functionalities through an open
interface. Ember proposed its own algorithm to construct the mesh-network. Though the
ready-to-use mesh networking is advantageous for some applications, the patented nature
of networking may result in an inflexible network system design. Intel recently developed
the Intel Imote2, which runs TinyOS or Linux (Adler et al., 2005). The Imote2 provides
enhanced computation and communication capabilities that allow resource-demanding
sensor network applications, such as SHM of civil infrastructure, to be supported, while
low-power operation and small physical size are still among the objectives. The wealth of
studies using TinyOS can be implemented directly to the Imote2; this commercial open-

14
interface platform will further advance the state-of-the-art in smart sensor technology,
especially for SHM application.

2.2 Middleware services

Prior to implementing SHM algorithms on a smart sensor network, many issues
intrinsic to the wireless sensor must be addressed. These issues include time
synchronization, routing, and reliable communication. These issues are not unique to civil
engineering applications, but are common for many smart sensor applications, including
environmental and habitat monitoring and sniper detection (Simon et al., 2004). However,
services to address these issues are not in general provided by operating systems.
Middleware services usually address these issues. Wireless sensor network middleware
services tend to be more application specific, as compared to those for PCs and servers,
because the limited resources of smart sensor networks (e.g., power, memory, and
bandwidth) need to be optimally utilized according to the specific applications (Romer et
al., 2002). Requirements for each middleware service need to be provided clearly; users
choose or develop middleware services which suit their requirements. Middleware
services relevant to this research are reviewed herein.
Because smart sensor nodes have their own clock, these nodes do not share global
time. This lack of a global clock is problematic for SHM applications. For example, if
modal analysis is conducted without accurate time synchronization, the identified mode
shape phase will be inaccurate, possibly falsely indicating structural damage. To address
this issue, several time synchronization techniques have been proposed so far. Reference
Broadcast Synchronization (RBS; Elson et al., 2002), Flooding Time Synchronization
Protocol (FTSP; Maroti et al., 2004) and Timing-sync Protocol for Sensor Networks
(TPSN; Ganeriwal et al., 2003) are among the well-known synchronization methods.
Lynch et al. (2005) implemented the RBS and reported a maximum time delay of 0.1
seconds. Mechitov et al. (2004) implemented FTSP as a part of wireless data acquisition
system for SHM. Their system can maintain better than 1ms synchronization accuracy for
a long period of time. When clock rate differs from node to node, clock drift is not
negligible. Synchronization needs to be applied periodically, or the difference in clock
rates needs to be estimated and compensated. Also packet transfer associated with time
synchronization is subject to packet loss. Reliability of synchronization in a lossy
communication environment is also an issue. Though accuracy and reliability of time
synchronization depends on the method and the hardware/software architecture,
synchronization better than 1 ms has been claimed by many researchers (Elson et al.,
2002; Maroti et al., 2004; Mechitov et al., 2004) and is becoming a reasonable
assumption.
Smart sensors can build several network topologies, including (a) the star topology
where one coordinator node directly communicates to other nodes, (b) the mesh topology
where multiple routing nodes communicate with each other to make a redundant network,
and (c) the cluster tree topology where routing nodes are organized in a tree to connect all
the nodes, usually without redundant connections. This classification is consistent with the
ZigBee standard. The star topology is not geometrically scalable because all of the nodes

15
need to be within the coordinator's communication range. Ember (2005) employed the
mesh network to monitor water treatment plants, using their ZigBee chipset. The first
generation Intel Mote (Kling, 2003; Kling et al., 2005) employed the cluster tree topology
utilizing Bluetooth technology. Mechitov et al. (2004) employed the tree topology with
Mica2s. By refreshing the topology periodically, the sensor networks are resistant to
communication link failures. Most of the current civil engineering applications using
smart sensors utilize the star topology. For mesh or cluster tree topologies, the routing
path needs to be optimized. To collect data efficiently and reliably at the base station,
Mechitov et al. (2004) set the path length as the primary criterion for establishing the tree
structure, with link quality being the secondary criterion.
Packet loss intrinsic to wireless communication needs to be addressed. Some data
need to be delivered reliably while others need to be delivered with better than a certain
probability. Although packet loss is inevitable, the loss of data can be avoided or reduced.
Sending a packet multiple times increases the possibility of successful delivery.
Furthermore, a packet can be repeatedly sent until acknowledgment is received. This
approach guarantees successful communication unless two nodes are completely out of
the communication range of each other or one of the nodes fails. Reliable communication
with acknowledgment messages was implemented on the Mica2 platform (Mechitov et al.,
2004). Also, standards such as Bluetooth and IEEE 802.11, provide their own
specifications to deal with communication errors. Successful delivery can be guaranteed
or increased at the expense of extra packets to exchange, larger header size, and/or extra
computation.
Because RAM is one of the power-consuming components on a smart sensor, the size
of RAM is usually small, sometimes necessitating virtual memory or other solutions.
Applications such as SHM deal with large response data records and need a large amount
of memory. For example, a modal analysis method, Eigensystem Realization Algorithm
(ERA; Juang and Pappa, 1985) employs singular value decomposition of two Hankel
matrices, whose size could be much larger than the RAM size; a 50 x 50 rather small
Hankel matrix consisting of 8-byte, double precision matrix elements needs 20 kB of
RAM, while hundreds by hundreds Hankel matrices are not uncommon. Mica2's 4 kB of
RAM memory is apparently insufficient for this SHM application. Kwon et al. (2005b)
proposed ActorNet which employs virtual memory on the Mica2. In addition to the virtual
memory, large physical memory can be a solution. For example, the Imote2 possesses 32
MB of RAM at the expense of larger power consumption.
Because of efficient power usage considerations, TinyOS employs nonblocking I/O,
which makes programming code complicated. The nonblocking I/O system does not wait
for a return value after it calls a function. A called task is posted for execution, and the
main thread keeps running. Once the task is executed, the completion is signaled. Kwon et
al. (2005b) proposed converting nonblocking I/O to blocking I/O to make the program
development more efficient. This nonblocking I/O is not a critical problem to be solved,
but this middleware service will increase the productivity of the inexperienced developers.
Though many middleware services have been proposed, studied, and implemented,
their performance has not been carefully examined from the SHM perspective. Moreover,
the middleware requirements for SHM are unclear. Based on SHM application

16
requirements, appropriate middleware services will need to be adopted, modified, or
developed.

2.3 Structural Health Monitoring

Structural Health Monitoring (SHM) strategies measure structural response and aim
to effectively detect, locate, and assess damage produced by severe loading events and by
progressive environmental deterioration. Structural response reflects the structural
condition as well as the excitation force. By analyzing the response data, SHM strategies
are expected to reveal structural condition, such as the damage existence. SHM has seen
intense research efforts in mechanical, aerospace, and maritime, as well as civil
engineering applications. In particular, condition-based machinery maintenance utilizing
vibration monitoring has seen many applications (Al-Najjar, 2000; Collacott, 1977). Once
damage is detected, detailed Non-Destructive Testing (NDT), repair, and/or suspension of
service follow. Well-known NDT techniques include visual inspection, eddy current
testing, acoustic emission, ultrasonic testing, and radiographic inspection (Gros, 1997).
SHM leads to well-organized maintenance of structures.
Efficient and effective maintenance is among motivations to introduce SHM
strategies for various applications. The Electric Power Research Institute (EPRI) has
successfully demonstrated the use of diagnostic technology at its Eddystone Station fossil-
fuel power plant outside of Philadelphia. Since the 1987 installation of a dual-purpose
vibration monitor and rotor-crack detector for bearing wear at the plant, EPRI says it has
saved $250,000 in costs for teardown of machinery. In military applications, SHM is
expected to reduce maintenance and manpower requirements by approximately 20 to 40
percent, increase combat sorties by 25 percent, and reduce the complexity of the logistics
trail by 50 percent, compared to current military strike aircraft (Becker et al., 1998). SHM
in these fields has been shown to be a practical and an attractive solution for maintenance.
From a civil engineering perspective, SHM is expected to provide an efficient and
effective tool for management of infrastructure. Civil infrastructure is a valuable asset,
which keeps the economy and people's life running. For example, a long-span bridge such
as the Akashi-Kaikyo Bridge in Japan costs billions of dollars, while the Golden Gate
Bridge in California has about 40 million crossings per year (Golden Gate Bridge,
Highway and Transportation District, 2005). These structures, however, deteriorate with
age. Many of the 1,100 major long-span bridges in the U.S. are over 50 years old and
several notable ones are over 100 years old. More than 800 of the long-span bridges in the
National Bridge Inventory are classified as fracture-critical (Pines & Aktan, 2002). The
structural integrity and serviceability of these structures are not necessarily apparent.
Because failure of such civil infrastructure can have significant negative impacts on
society at large, structures need to be managed based on the understanding of their current
structural conditions. The current practice, visual inspection, is expensive both in time and
cost. For example, the biennial visual inspection of a major bridge such as the Brooklyn
Bridge in New York is reported to last for over 3 months at a cost of $1 million (Dubin &
Yanev, 2001; Pines & Aktan, 2002); note that the reliability of such visual inspection is

17
being examined (Moore et al., 2001). SHM potentially offers better and more efficient
understanding of structural conditions.
In addition to structural deterioration due to aging, SHM can be employed to estimate
structural damage due to severe loading events, such as earthquakes, hurricanes, or
tornados, facilitating more timely recovery from such disasters. It is imperative that
emergency facilities and evacuation routes, including bridges and highways, be assessed
for safety. Traditional detailed assessments can be significantly expensive and time-
consuming, as was seen after the 1994 Northridge earthquake with the numerous buildings
that needed to have their moment-resisting connections inspected. Additionally, structures
that are internally, but not obviously, damaged in an earthquake may be in great danger of
collapse during aftershocks; structural integrity assessment can help to identify such
structures to enable evacuation of building occupants and contents prior to aftershocks.
SHM enhances safety and reliability of civil infrastructure after such disastrous events.
Though the necessity is clear, the features of civil infrastructure pose difficulties to
SHM. These features include large scale, a myriad of elements, one-of-a-kind designs,
low natural frequencies, structural redundancy, nontrivial test excitation, challenging
input force estimation, long and continuous service time, and variable environmental
conditions. For example, changes in resonant frequencies of redundant structures, which
often imply structural damage, can be insignificant as compared to frequency shifts due to
changes in ambient conditions, such as temperature and support stiffness (Aktan et al.,
1994; Chowdhury, 1990; Farrar et al., 1994; Salawu, 1997; Tang & Leu, 1989). Many
researchers have been working to overcome these difficulties.
Doebling et al. (1996) reviewed research on vibration-based damage identification
and health monitoring. The cited papers were categorized based on the method, such as
frequency change, mode shape change, mode shape curvature, flexibility, matrix update,
nonlinear methods, and neural network-based methods. Sohn et al. (2003) reviewed
papers on SHM published between 1996 and 2001. A summary of statistical approaches
for damage detection was also included. Other review papers include Salawu and
Williams (1995), Salawu (1997), Doebling et al. (1998), Doebling and Farrar (1999), and
Farrar et al. (2003). Major algorithms for SHM are briefly described below.
Most of the SHM methods mentioned above employ modal analysis to obtain modal
parameters such as natural frequencies, damping ratios, and mode shapes. A number of
modal analysis methods have been proposed. Peak-picking is a simple frequency domain
modal analysis method. Improvements by incorporating the coherence function and using
frequency domain decomposition (Brincker et al., 2001) have also been reported.
Maximum likelihood identification, which includes curve-fitting, estimates modal
parameters by minimizing an error norm in the frequency domain. Time domain modal
analysis methods include the complex-exponential method (Maia & Silva, 1997); Ibrahim
time domain method (Ibrahim & Mikulcik, 1977); ERA (Juang & Pappa, 1985); and
stochastic subspace identification (Hermans & Auweraer, 1999). Identified modal
parameters are further analyzed for damage detection.
Natural frequency changes reflect structural conditions, offering clues to estimate the
structural conditions. By analyzing the natural frequency change, the cause of structural

18
damage is expected to be revealed. For example, the tension in the stay cables of cable-
stayed bridges can be estimated by natural frequency measurement (Gardner-Morse &
Huston, 1993). The frequency changes, however, may have low sensitivity to damage as
mentioned earlier. Begg et al. (1976) pointed out that cracks have little influence on global
modes and suggested that local high-frequency bending modes of the individual members
provide a better indication of cracking. Srinivasan and Kot (1992) conducted experiments
on a shell structure and found that the resonant frequencies of the shell structure were
insensitive to damage. Also, frequency shifts depend on each mode. Salawu (1997)
mentioned that the stress induced by modal deformation is minimal at modal nodes, where
modal displacement is zero; a small change in a particular modal frequency could mean
that the defect may be close to the modal node. Hearn and Testa (1991) developed a
damage detection method that examines percentage changes in natural frequencies. To
relate observed frequency changes or percentage changes to structural damage, these
methods require theoretical structural models as well as a damage model or sensitivity
analysis; these model estimations and the analyses are not straightforward (Salawu, 1997).
Fluctuation in damping values is much larger than that in frequencies (Williams &
Salawu, 1997) and can potentially be a damage indicator. High damping would suggest
more energy dissipation mechanisms, indicating the possibility of cracks in the structure
(Morgan & Osterle, 1985). Changes in damping values up to 80 percent were reported by
Agardh (1991). Williams and Salawu (1997), however, pointed out that damping
properties are the most difficult to model analytically and can only be realistically
obtained through vibration tests. Nashif et al. (1985) and Sun and Lu (1995) give
comprehensive discussions. Relatively large estimation and modeling errors are against
the usage of damping for SHM.
Mode shape information has also been investigated. West (1984) compared mode
shapes of a space structure, using Modal Assurance Criterion (MAC), before and after
exposure to loading. The mode shapes were partitioned and changes in the local MAC
values along the mode shapes were used to locate the damage. Fox (1992) stated that
mode shape changes are insensitive to damage and “Node line MAC,” a MAC variant
based on measurement points close to a node point for a particular mode, is a more
sensitive indicator of damage. Pandey et al. (1991) demonstrates that absolute changes in
mode shape curvature, calculated through difference analysis of displacement mode
shapes, can be a good indicator of damage. Chance et al. (1994) and Nwosu et al. (1995)
used measured strains instead of the calculated curvature, reducing the numerical error
associated with difference analysis. Salawu and Williams (1994), however, pointed out
that the selection of which modes are used in the analysis is an important factor, and
curvature changes do not typically give a good indication of damage using experimental
data. Mode shape phase has also been reported to be a possible damage indicator. Mode
shapes of proportionally damped dynamic systems are theoretically purely real, i.e., they
are aligned within a plane, resulting in either 0 or S radian phase. Structural damage,
especially damage with friction type mechanisms such as loose bolt connections, often
introduces nonproportional damping, which results in complex mode shapes, i.e., the
phases will differ from 0 and S These phase changes can potentially indicate structural
damage, although structural damage is not the only cause to change the phase of mode
shapes (Nagayama et al., 2005). While some mode shape related indicators reflect

19
structural damage, the analysis to quantitatively relate the mode shape information to
structural damage is yet to be completed.
The dynamically measured flexibility matrix, calculated from the mass-normalized
mode shapes and modal frequencies, has been employed to determine structural damage.
By interpreting the physical meaning of the flexibility matrix, Toksoy and Aktan (1994)
observed that anomalies in the flexibility matrix can indicate damage even without a
baseline data set. Gao et al. (2005) reported that the flexibility matrix is estimated with
moderate accuracy using only a few lower modes, as opposed to the stiffness matrix (the
inverse of the flexibility matrix), which needs almost all of the modes. Though some
researchers insist that higher modes better indicate the damage due to their insensitivity to
support conditions and high sensitivity to local damage (Alampalli et al., 1992; Biswas et
al., 1990; Chowdhury, 1990; Lieven & Waters, 1994), estimation accuracy of these modes
is arguable. If adequate for the purpose, lower modes with more accurate estimation
would better suit SHM. The ability of the flexibility matrix to be estimated with only a few
lower modes is, therefore, advantageous. He and Ewins (1986), Lin (1990), and Zhang
and Aktan (1995) further studied flexibility-based SHM strategies. Deficient points of
these techniques include that the error due to the unmeasured modes and flexibility change
due to damage cannot be clearly separated, and that structural constraints such as
symmetry and connectivity have not been fully incorporated, though Doebling (1995)
conducted research to tackle this problem.
Matrix update methods offer another class of SHM methods. Structural characteristics
such as mass, stiffness, and damping are represented in matrix form and updated to be
consistent with the observation of the structure under consideration. As Doebling et al.
(1996) described, the objective function is numerically optimized to update the matrices
under constraints such as matrix sparsity, connectivity, symmetry, and matrix positivity.
Doebling et al. (1996) categorized matrix update methods as closed-form optimal matrix
update, sensitivity-based update, eigenstructure assignment, and hybrid matrix update.
Zimmerman and Kaouk (1994) introduced minimum rank perturbation theory (MRPT) to
the optimal matrix update problem; this approach has been published extensively (Kaouk
& Zimmerman, 1994, 1995; Zimmerman & Simmermacher, 1995; Zimmerman et al.,
1995). The limitation of these methods is that the rank of the perturbation is always equal
to the number of modes used in the computation of the modal force error. Also, the
number of degrees-of-freedom (DOFs) and type of FEM models can affect the final
results.
Wavelet transforms and analytical signals have been among popular data analysis
tools. Wavelet transforms are often used to extract features from complicated data.
Staszewski (1998) summarized the application of wavelet analysis for SHM. While
wavelet transforms provide flexible and powerful data analysis tools that can be used in
combination with other methods, the transforms do not possess a clear physical meaning,
as opposed to the Fourier transform. Feldman and Braun (1995) used an analytical signal,
which is a combination of the actual signal and its Hilbert transform, to get an
instantaneous estimate of the modal parameters. The analytical signal may not work when
more than one modal component exists in the signal. Features of nonstationary processes
are often analyzed with these tools.

20
 Nonlinearity of structural behavior due to structural damage has also been studied.
Crack opening and closing was investigated by Actis and Dimarogonas (1989) and
Krawczuk and Ostachowicz (1992). Lin and Ewins (1990) contemplated response level
dependencies of structural properties utilizing modal properties measured at different
response level. Sohn et al. (2007) proposed damage diagnostics based on the concept of
time reversal acoustics and consecutive outlier analysis to detect nonlinearity between a
pair of sensors.
Sohn et al. (2003) emphasized the importance of statistical models to enhance the
SHM process. Statistical model development can be classified as supervised learning and
unsupervised learning. Sohn et al. (2003) discussed supervised learning methods such as
neural networks (Bishop, 1994; Nakamura et al., 1998; Zhao et al., 1998), genetic
algorithms, support vector machines (Vapnik, 1998), outlier detection, and hypothesis
testing. These supervised learning algorithms are first trained by structures of known
properties or FEM models, and then applied to the structures to be tested. Therefore, this
algorithm may require significant training with the undamaged structure as well as
structures damaged in one or more of the failure modes. While this training may be
possible for structures produced in larger lots (e.g., aircraft), the training is nontrivial for
unique structures (e.g., buildings and bridges). On the other hand, unsupervised learning
may detect abnormalities, or damage presence, but locating damage and assessing the
severity need further theoretical development. Sohn et al. (2002) proposed to combine
AR-ARX model, nonlinear principal component analysis, and statistical analysis to
distinguish environmental and structural changes. Computer simulation and experimental
results proved that this method can detect damage. By placing sensors at each DOF and
applying this procedure for each sensor, Sohn et al. (2002) successfully located damage of
an 8 DOF experimental model under environmental change.
Bernal (2002) proposed a flexibility-matrix-based damage localization method, the
Damage Locating Vector (DLV) method. The DLV method has advantages in that
structural responses do not need to be measured at all the DOFs, though small numbers of
sensors result in limited damage detection capability. A set of load vectors, designated as
DLVs, were computed from the change in the flexibility matrix. The flexibility matrix can
be dynamically estimated. When the DLVs are applied as static forces on the undamaged
structure, the stress field in the structure bypasses the damage areas. This unique
characteristic of the DLVs can be employed to localize damage in the structure.
Although many methods have been proposed, none have proven to be sufficient for
full-scale application. Many of the SHM algorithms mentioned above have been shown to
detect damage well when a large enough number of modes are accurately measured at all
the DOFs. Even the DLV method, which does not require response measurements at all
the DOFs, performs poorly when used with a small number of sensors. The more sensors
used, the more information about structures SHM is expected to give. However, structures
are usually large and have numerous DOFs; accurate and thorough measurements have
been impractical. Sensors have limited accuracy and the associated installation cost,
including cabling, has been prohibitively expensive. For example, Lynch and Loh (2006)
cited Farrar (2001), which reported that the cost of installing over 350 sensing channels on
the Tsing Ma Bridge in Hong Kong was more than $8 million. Celebi (2002) estimated

21
that sensor installation cost including labor, cabling, and recording systems is about
$4,000 per sensing channel. The emergence of smart sensors with wireless
communication capabilities, as described in the following section, offers the possibility of
SHM with dense measurements.

2.4 Attempts toward SHM using smart sensors

2.4.1 Research attempts

Several SHM applications with smart sensors have been reported using scale models.
Demonstrated tasks include data acquisition with a single wireless node, synchronized
data acquisition with multiple nodes, on-board data processing, etc. Tanner et al. (2002,
2003) embedded data processing on a smart sensor unit. A Mica node was programmed to
measure acceleration responses of a beam on both sides of its bolted joint and then
calculate the correlation coefficient of the responses to detect a loose bolt. Lynch et al.
(2002) implemented Fast Fourier Transform (FFT) to reveal the five-story building
model's response in the frequency domain. Nitta et al. (2005) implemented an AR model
on the Mica2 and experimentally verified its validity on a three-story building model.
These studies have briefly demonstrated the applicability of smart sensor systems to SHM
applications using simplified models.
Full-scale buildings and bridges have also been the subject of smart sensor research.
Straser and Kiremidjian (1998) and Lynch et al. (2003) measured structural responses of
the Alamosa Canyon Bridge to validate their smart sensors' performance. Galbreath et al.
(2003) monitored a highway bridge on the LaPlatte River in Vermont, using Microstrain's
wireless strain sensor unit (Microstrain, Inc., 2007). Aoki et al. (2003) measured the
acceleration response of a light pole on the Tokyo Rainbow Bridge in Japan. The data was
transmitted to a data repository using a WLAN. Chung et al. (2004) installed a DuraNode
sensor unit on a pedestrian bridge at the University of California, Irvine. Wirelessly
collected data was then analyzed on a PC to give the first three vibration modes. Ou et al.
(2005) installed eight Mica nodes in the Di Wang Tower in China. Lynch et al. (2005)
installed 15 smart sensor units on the Geumdang Bridge in Korea to measure the forced
vibration response. FFT was applied to measurement signals on smart sensor nodes
independently, and the Fourier transform results were sent back to the base station. In the
geotechnical research field, Chen et al. (2005) proposed to use wireless sensor for MEMS-
based vertical seismic array, called Terra-Scope. These research attempts have
demonstrated the capability of smart sensors to measure acceleration for full-scale civil
infrastructure, though the quality of data was not necessarily examined.
Through the laboratory and full-scale structural applications, benefits in terms of the
sensor installation time were reported. Straser and Kiremidjian (1998) reported that
installation of the wireless system on the Alamosa Canyon Bridge took 30 minutes, which
was five times faster than the cable-based system. Lynch et al. (2003) implemented smart
sensor units on the same bridge, and the installation time was half the time to install the
cable-based system.

22
 Many researchers have also been working on another approach to SHM with wireless
communication technology. The data is acquired using a traditional wired acquisition
system and then sent back to a remote data depository using a cell phone, WLAN, or other
wireless communication (Karbhari et al., 2003; Mufti, 2003;Oshima et al., 2000).
However, this framework does not possess the distributed on-board microprocessors, nor
does it eliminate the cabling cost for sensor installation. Other approaches should be taken
if densely instrumented measurement is an objective.

2.4.2 Difficulties in using smart sensors for SHM applications

Though many researchers have demonstrated the use of smart sensors for SHM
applications, none of them have resulted in a full-fledged SHM system. There are
difficulties civil engineers commonly encounter when SHM applications are implemented
on smart sensors. For example, many of the demonstrated systems are not scalable to a
large number of smart sensors. Some smart sensors cannot acquire reliable measurement
data due to deficiencies in the sensor, time synchronization errors, data loss, etc. Major
difficulties are summarized herein from an SHM application perspective.

1. Sensor hardware

Though several types of sensors have been implemented on smart sensors (see
Table 2.1 on page 12), these sensors do not cover all types of sensors that civil engineers
typically need. For example, only a few smart sensors have employed a strain sensor,
while strain is one of important physical quantities from which to judge structural
condition. Velocity or displacement sensors have not seen smart sensor application. When
civil engineers do not find appropriate sensors on smart sensor platforms, sensor boards
need to be customized; open source platforms make it easier to customize sensor boards.
While accelerometers are some of the most commonly employed sensors, their
applicability to civil engineering applications is not apparent. Users cannot simply assume
MEMS accelerometer’s characteristics to be similar to those of conventional
accelerometers. Acceleration of shake tables or the response of structure model has been
recorded with smart sensors to examine the performance of their accelerometers (Arici &
Mosalam, 2003; Casciati et al., 2003; Hou et al., 2005; Kurata et al., 2004; Lynch et al.,
2002; Ruiz-Sandoval, 2004; Straser & Kiremidjian, 1998).
Limited sensitivity and high noise floor are sometimes problems with MEMS
accelerometers. While acceleration noise density of one of the conventional piezoelectric
accelerometers, PCB 393B04 (PCB Piezoelectronics, Inc., 2007), is 0.04Pg/ Hz , that of
MEMS accelerometer used on many smart sensor prototypes, ADXL201 (Analog
Devices, Inc., 2007) is 200 Pg/ Hz . The accelerometer on the Imote2 sensor board,
LIS3L02DQ (STMicroelectronics, 2007), has a noise density of 50 Pg/ Hz . This noise
floor is considered low enough for scale-model dynamic testing using an exciter or shake
table; however, applicability to ambient vibration measurement of buildings needs further
investigation. Ruiz-Sandoval (2004) and Ruiz-Sandoval et al. (2006) employed an
accelerometer with a high sensitivity and low noise level, SD1221 (Silicon Designs, Inc.,
2007). This MEMS type accelerometer has the noise density of 5 Pg/ Hz . Ruiz-Sandoval

23
(2004) indicated that this accelerometer is suitable for SHM applications. Users need to be
aware of the sensing characteristics of smart sensors.
Characteristics of accelerometers need to be well-examined, especially in the low
frequency range, because major vibration modes of civil infrastructure appear in this
frequency range. Natural frequencies of tall buildings, towers, or long bridges can be as
low as 0.1 Hz. In terms of vibration amplitude, acceleration in the low frequency range is
small, underlining the importance of high resolution and sensitivity of a sensing system.
Ruiz-Sandoval (2004) and Ruiz-Sandoval et al. (2006) calibrated their sensor board with
special emphasis in the low frequency range. Though many smart sensors with
accelerometers have been proposed, only a limited number of acceleration sensor boards
can measure low frequency vibration accurately.
In addition to the sensor itself, the ADC, AA filter, and supply voltage regulator also
influence the quality of measurement signals. A low resolution ADC degrades measured
signals by introducing large quantization errors. The ADC on the Mica2, for example, has
only a 10-bit resolution, limiting the dynamic range of sensors. Appropriate lowpass
filters are essential to obtain digital signals free of aliasing. The sensor’s supply voltage
needs to be regulated so that current drawn by the microprocessor, radio device, or flash
memory does not destabilize current flow to sensing components. These components need
to be carefully designed. Otherwise, the structural information submerged in the
measurement signals may not be extracted. Because all of these issues affect signal
quality, smart sensor users cannot simply assume the sensing characteristics of a sensor
node are the same as those of a sensor component.
Even a dense array of smart sensors is not a rich information source for SHM if
physical quantities needed cannot be measured precisely by each smart sensor.
Development of sensor boards for SHM applications is still an important research issue as
well as calibration of these sensor boards.

2. Data aggregation

Data aggregation for SHM applications often encounters the following three issues:
(a) data size is too large, (b) data may be lost during wireless communication, and (c)
communication range is limited. Each of these problems is summarized below.
Smart sensors in general are not designed to collect a large amount of data, while
SHM applications benefit from data acquired from numerous sensors with high sampling
frequencies. Early applications of smart sensors, such as habitat monitoring, handled only
a small amount of data on an infrequent basis. On the other hand, SHM applications
typically acquires tens of thousands data points, each of which is represented as two- or
four-byte data. Sampling frequencies higher than 100 Hz and total sampling time longer
than a minute are quite common. In view of the need to handle a large amount of data,
SHM applications with smart sensors can be categorized into two groups, neither of which
has fully exploited the smart sensor’s capability.
In the first group, the smart sensors are employed in the same manner as traditional
wired sensors, with all data being collected for processing at a centralized location (see

24
 Data processing

Data acquisition

Figure 2.2. Centralized data acquisition approach.

Data acquisition & processing

Processing

Figure 2.3. Independent data processing approach.

Figure 2.2). Centralized SHM algorithms are then applied to this data. This approach
allows for application of a wealth of traditional SHM algorithms reviewed in 2.3
Structural Health Monitoring. As the number of smart sensors increases, however, the
measurement data to be centrally collected exceeds the network bandwidth, whether
homerun or hopping communication is adopted. The lack of scalability is a serious
deficiency of this approach. One approach toward a scalable solution is to have a tiered
network. Chintalapudi et al. (2006) utilized lower tier nodes and powerful upper tier
nodes. Assuming upper tier nodes have sufficient power, power consumption at lower tier
nodes is moderated. The tiered network approach is applicable only when installation of
powerful nodes and power supply to these nodes are practical.
The second group, on the other hand, assumes that each smart sensor measures and
processes data independently without sharing information among the neighboring nodes
as illustrated in Figure 2.3 (Lynch et al., 2005; Nitta et al., 2005; Sohn et al., 2002).
Because only the data processing outputs are sent back to the base station, the required
communication can be quite small.
Consequently, this approach is scalable to a large number of smart sensors. However,
the independent sensor node approach does not utilize available information from
neighboring nodes; all spatial information is neglected. For example, information
regarding mode shapes cannot be obtained and used in this approach. Information from

25
different type of sensors connected to separate nodes cannot be combined. The inability of
this approach to incorporate spatial information limits its effectiveness.
Gao (2005) proposed a Distributed Computing Strategy (DCS) for SHM that offers a
scalable approach that can incorporate spatial information. This DCS approach, based on
the DLV method (Bernal, 2002), does not need to centrally collect and analyze the
measurement data. Instead, the DCS shares data among the neighboring nodes to utilize
spatial information. Due to this local data sharing with limited numbers of neighboring
nodes, the total amount of data to be transmitted throughout the network is kept modest.
Therefore, this SHM strategy is scalable to a large number of sensors densely deployed
over large structures. While the DCS does not require measurements at all the DOFs, the
method’s performance improves with the number of DOFs measured. Computer analysis
and experimental validation on a simulated wireless network showed the DCS is a
promising SHM scheme. This strategy, however, has not yet been implemented on smart
sensors and experimentally verified.
Data loss during wireless communication is also problematic for SHM applications.
Wireless communication suffers from packet loss unless lost packets are resent. Kurata et
al. (2004) reported data loss during shake table experiments. Many civil engineering
applications using smart sensors do not address this data loss problem. Some of them
simply ignore lost data while others coincidentally receive all the packets during
experiments. However, SHM methods developed so far assume that data acquired at
sensors are available for data processing at the base station. The influence of lost data on
structural analysis has not been clearly investigated. Mechitov et al. (2004) employed a
reliable communication protocol to address this problem. Because acknowledgment
packets are sent frequently, the communication speed is slower than communication
without acknowledgment. Reliable communication services suitable to transfer large
amount of data is expected to advance SHM applications employing smart sensors.
The communication range of smart sensors is usually shorter than the size of civil
infrastructure. Centrally collecting data or sending commands to smart sensor nodes on
structures involves multihop communication. A routing path usually needs to be
determined prior to multihop communication (Mechitov et al., 2004). If a communication
path between two arbitrarily selected nodes is required, a very large table to store routing
paths needs to be constructed on each sensor node. Application-specific knowledge on
communication, such as collecting data to a sink node and dissemination, potentially
simplifies routing. Once paths are found, multihop communication can be started. Routing
path, packet structure, overhead information, etc. need to be carefully designed.

3. Time synchronization

Dynamic analyses of structures assume that data is synchronized, which is not the
default case with smart sensor networks. As described in section 2.2, several
synchronization protocols have been proposed. Some of them have synchronization
accuracy as good as tens of microseconds. However, the requirement of SHM on time
synchronization accuracy is not studied. Based on the fact that natural frequencies of
structures used in analyses are usually below 10 Hz, civil engineers may erroneously think

26
that time synchronization error smaller than a millisecond is acceptable. SHM strategies
need to be examined against time synchronization error prior to their use.

4. Limited computational capability

Some SHM applications utilize numerical operations that require large computational
power and memory allocations. Methods requiring manipulation of large matrices, such as
ERA, fall into this category. Because memory space on smart sensors is limited, these
methods are impossible to implement on smart sensors, or the performance is limited
(Chintalapudi et al., 2006; Nagayama et al., 2004).

5. Power

Power consumption is an important issue to civil infrastructure monitoring. As

compared to other smart sensor applications (e.g., habitat monitoring), monitoring of
structures may have easier access to power sources. Buildings are equipped with power
outlets, and many long bridges have power that is used for light poles, etc. However,
wiring power to numerous sensors takes a significant amount of time and increases
installation cost, negating one of the important advantages of smart sensor. Also, sensor
installation location is not always close to a power source on a structure. Available power
may need to be transformed to an appropriate voltage and frequency. Some structures are
without power. Therefore, power source cannot be assumed to be available. Smart sensors
employing batteries are beneficial and oftentimes the only solution.
Once smart sensors are installed on civil infrastructure, their batteries may not be
easily changed. Some nodes may be installed at places with low accessibility.
Maintenance during which batteries on smart sensors can be replaced is not frequently
scheduled. Power consumption for SHM applications is, however, in general larger than
other wireless sensor applications, shortening battery life. So far, no battery-powered
smart sensor systems for SHM has been implemented on a permanent basis. Much
research effort is being devoted to resolve this issue; one promising approach is power
harvesting.

2.5 Summary
Recent technological advance in smart sensor technology offers new opportunities in
SHM for civil infrastructure. A dense array of smart sensors is expected to be a rich source
of information for SHM. Attempts toward SHM using smart sensors, however, always
encounter difficulties. Although there are reports on smart sensor usage in structural
vibration measurement, none of them has resulted in a full-fledged SHM system. Issues in
scalability, sensing capability, synchronization accuracy, etc. must still be addressed. This
research realizes a framework for SHM using smart sensors by addressing these issues.
The next chapter presents the SHM architecture employed in this report to address these
difficulties.

27
 Chapter 3

SHM ARCHITECTURE

This chapter discusses the architecture for an SHM framework that addresses many of
the difficulties described in the previous chapter. Prior to architecture design, desirable
characteristics of an SHM framework employing smart sensors are summarized. The
smart sensor network system, smart sensor platform, middleware services, and damage
detection algorithms to be employed in this report are then discussed in detail (Spencer &
Nagayama, 2006).

3.1 Desirable characteristics of an SHM system

employing smart sensors
Desirable characteristics for an SHM framework using smart sensors must first be
established. Many researchers have been working on an SHM employing smart sensors by
a variety of approaches. Some researchers utilize only the wireless communication
capability, while others emphasize the use of the embedded microprocessor. Assumptions
on the type of power source also vary widely. Desirable characteristics for an SHM
strategy implemented on a network of smart sensors are specified herein and serve as
guidelines for this research.

1. Scalable resource aware system

Scalability

A large number of sensors densely distributed over a region of interest are considered
necessary to understand the complete state of a structure. Because damage to a structure is
a local phenomenon, signals from sensors near damage are expected to contain more
information on damage than those remotely located. The spatial variation of measurands
can be accurately assessed only when sensors are distributed in a sufficiently dense
manner. Dense instrumentation on structures such as the 3.9 km-long Akashi-Kaikyo
Bridge and the 443 m-tall Sears Tower need a great number of smart sensor nodes. The
inexpensive nature of the smart sensors potentially makes a large number of smart sensors
economically feasible, while the wireless communication capability dramatically reduces
installation costs. A network of smart sensors, however, is not scalable to a large number
of sensors unless the limited resources in the network are respected and well-managed.
For example, smart sensor networks that are intended simply to collect all of the measured
data to a single base station suffer from insufficient bandwidth and are not scalable. Both
the hardware and software should allow dense arrays of smart sensors.

28
Power awareness

Power available for smart sensor networks is typically limited, with most smart
sensors being battery-operated. Power consumption needs to be managed so that the smart
sensor's lifetime is prolonged. In particular, radio communication and flash memory
access need to be planned well, as they are power-demanding tasks (see Table 2.1 on
page 12). Within the network, nodes with more available power should be assigned more
computation and communication tasks to extend the lifetime of the overall network.
Effective power management should result in longer lifetimes for both individual nodes
and the network.

Bandwidth awareness

The use of RF communication should be well-managed in order to limit network

congestion. Excess communication increases data collision, and may even surpass the
network's communication capacity. The higher the traffic load on a network, the more
likely collisions will take place and the less reliable the communication will be. As a
network, nodes with more available communication bandwidth should be assigned more
relay tasks. To increase reliability and efficiency of communication, as well as to save
power, data transferred needs to be kept moderate and bandwidth use should be optimized.

Memory awareness

Flash memory and RAM are among the limited resources on smart sensors.
Moreover, accessing flash memory is power-intensive; read and write accesses should be
limited. Variables needing frequent accesses should be kept on RAM instead of flash
memory. RAM on a smart sensor, on the other hand, is a component which constantly
consumes nonnegligible power; consequently, the current RAM size on smart sensors is
relatively small and expected to increase slowly in the future. For example, Mica2 has
only 4 kB of RAM, which allows storage of only about 1,000 data points in a single-
precision floating point format. When an application requires processing of large data,
such as Discrete Fourier Transform (DFT) of a long time history, Virtual Memory (VM) is
one approach to address the problem of small RAM at the expense of flash memory usage
and associated power consumption. Memory usage needs to be well-managed.

Computational power awareness

As compared to a microprocessor on a PC, the one on a smart sensor has limited

speed and functionality. Such a microprocessor is in charge of multiple tasks, for example,
data acquisition, flash memory access, RF control, numerical calculation, and task
scheduling. Only a single task, or a few tasks if the microprocessor is designed
specifically, can be executed at a time. The tasks on a microprocessor need to be well-
organized with priorities, so that timing critical tasks, such as data acquisition, is not
disrupted or delayed by other less-time-critical tasks.

29
Modal operation

Several modal operations for the microprocessor, such as sleep, watchdog, and
awake, facilitate a resource-efficient operation. For example, following a preprogrammed
schedule, nodes enter the watchdog mode to save power. When no significant event to
observe is expected, the nodes go into the sleep mode. The smart sensors should be able to
transition from one operation mode to another, depending on the tasks to be achieved.

2. Autonomous distributed embedded computing

Model-based data aggregation

The data measured at nodes should be processed locally so that a reduced amount of
data needs to be sent throughout smart sensor networks. Data size reduction without data
degradation should be sought. Algebraic operations such as averaging, numerical filtering,
and resampling, are simple examples. More complex mathematical manipulation such as
frequency analysis may better compress information. Data aggregation based on
knowledge or insight on a structural system is expected to further condense measured data
without compromising the structural information contained. Ideally, only necessary and
sufficient information for the task is transmitted throughout networks.

Collaborative distributed data processing

Distributed data processing eliminates the problem of having a single point of failure
and balances the power consumption among nodes. Also, distributed processing offers
efficient computation, which can be much faster than centralized processing. Distributed
processors, accompanied by cache memory and RAM, contribute to fast computation.
Although resources on each node are limited, sensor networks as a whole possess
appealing computational capabilities.

Autonomous initial configuration and maintenance

The network topology is desired to be dynamically and autonomously configurable.

When smart sensors are physically installed, the sensor nodes need to construct a network
and configure the topology. Because node loss is likely to take place, network
configuration should be adjusted autonomously so that loss of a single node does not take
the system down. Autonomous reconfiguration also allows balancing power consumption
among sensor nodes by switching relay nodes on multihop communication paths.
Individual nodes should be reconfigurable through the network as well. For the initial
set up of a smart sensor network, a large number of nodes need to be programmed as well
as physically mounted. In the long term, the nodes may need to be reprogrammed to
implement different tasks or to reschedule the tasks. Smart sensors are desired to be
programmable through the network, because manual reprogramming of thousands of
nodes is too time consuming and error prone.

30
3. Fault tolerant system

Message packet loss tolerance

RF communication inherently has data loss. Resending lost data packets can increase
transmission reliability at the price of increased communication demand and power
consumption. Algorithms tolerant of data loss are appealing from both power and
bandwidth perspectives.

Node failure tolerance

During the life of a smart sensor node, functionalities of the node may become
impaired. Power depletion stops all of the functionalities of a smart sensor. Wireless link
disconnection impairs only the communication capability. The network should be tolerant
of these node losses by reconfiguration of networks, while application algorithms
allowing the loss of sensing signals from a few nodes are desirable.

Byzantine error tolerance

Even when data is acquired, the data can be faulty and misleading. For example,
when a node with an accelerometer comes unglued, the node acquires faulty acceleration
data from the sensor. As another example, RF interference could induce errors in received
data. Low battery power may also result in random errors in the radio, CPU, and/or sensor.
A system to address this Byzantine error problem is desirable.

4. Desirable algorithmic characteristics specific to SHM for civil infrastructure

Multiscale information

SHM techniques are needed that can employ measured information at multiple scales.
Different types of sensors measure different physical quantities, each of which has its own
sensitivity to certain structural conditions. For example, acceleration and strain are among
the most important physical quantities to judge the health of a structure. While
acceleration measurements are essentially global responses of a structure, structural strain
provides an important indicator of local structural behavior. By using such multiscale
sensor information, structural condition is expected to be assessed more accurately.

Collaboration in local sensor communities

Smart sensors, densely distributed over structures, are expected to communicate with
each other, at least in a local sensor community, to make efficient and effective use of
measured data. Closely located nodes are anticipated to give highly correlated
measurements, which might be fused without significant loss of information, although the
definition of closeness depends on each measurand. Data processing of a spatially
sampled measurand may reveal important information, in a similar way to data processing
of a measurand sampled in the time domain. For example, estimation of a mode shape and

31
its spatial gradient, which is claimed to change according to the health of a structure
(Farrar et al., 1994; Nagayama et al., 2005; Pandey et al., 1991; West, 1984), needs spatial
information. Independent data processing without data sharing among nodes cannot utilize
such spatial information. On the other hand, collaboration of all of the smart sensors in the
whole network is neither practical nor desirable due to substantial communication and
power requirements. Collaboration in local communities, which are spatially large and
dense enough to capture local structural condition, is thus essential for SHM using a dense
array of smart sensors.

Redundancy

Because measurement error, inaccurate modeling, numerical error, etc. introduce

uncertainty in SHM results, SHM algorithms on smart sensor networks should possess
redundancy. For example, more than one set of smart sensors monitor a given element and
make their own judgments on the damage existence; the judgments are then compared
with each other to examine their consistency. Redundancy is an effective means to deal
with the uncertainty, which is problematic for most SHM algorithms.

Multiple functions

Smart sensor networks should preferably achieve multiple tasks, for example,
continuous SHM of a structure, baseline measurement for SHM, monitoring of rare events
such as earthquakes, and sensor calibration. Inclusion of other tasks such as traffic
monitoring and local weather monitoring further enhances the value of smart sensor
systems and makes introduction of smart sensor systems more attractive from a cost/
benefit perspective.

5. Desirable platform characteristics specific to SHM of civil infrastructure

Appropriate sensor availability

Appropriate sensors need to be available for smart sensors. Acceleration and strain
are among the most important measurands for SHM, while velocity, displacement,
temperature, humidity, wind velocity, and wind direction as well are sometimes measured
in full-scale structural monitoring (Ko & Ni, 2005; Wong, 2004). Sensors for these
measurands are needed on smart sensor platforms. One of the most commonly adopted
sensors on smart sensor prototypes is an accelerometer. The applicability of
accelerometers on smart sensors to civil engineering applications is, however, not
necessarily clear. Strain sensor adoption to smart sensors is not common (Arms et al.,
2004; Lynch & Loh, 2006; Nagayama et al., 2004). Sensors being capable of accurately
capturing structural behavior and environmental conditions need to be developed or
customized under the constraint of limited resources of smart sensors, especially battery
power.

32
Sufficient RAM, flash memory and RF bandwidth

Because SHM for civil infrastructure measures dynamic vibrations of large structures
at high sampling frequencies, the amount of data to be saved, processed, and transmitted is
large even after local data processing. The sampling frequency of dynamic vibration
measurements is often higher than 100 Hz. RAM size, flash memory size, and radio
throughput are desirably large, though power consumption still needs to remain moderate.

Environmentally hardened

Because smart sensors placed on structures are often subjected to severe

environmental conditions (e.g., wind, rain, sunshine, extreme temperature, and vibration),
the sensors need to be rugged. Packaging is one solution. Without ruggedness, severe
environmental conditions may result in faulty data or node loss.

Open source OS and interface

The operating system (OS) and the interface should be open to the public. Tasks are
ideally optimized to make efficient use of the limited resources. Services offered by the
OS may not give the best solution for a specific task. Algorithms and middleware services
to be implemented may work more efficiently if the OS is customized. Certain algorithms
or middleware services may be implementable only if the OS is customized. In such cases,
users should have access to the source of the OS. Also, open source OS and interfaces
allow user communities to participate in improving OS and software. In particular, at this
early stage of smart sensor network research, contributions from broad communities are
important to improve the technology. Open source OS and interfaces will advance smart
sensor network technology and its application to SHM.

3.2 SHM system architecture

An SHM system architecture to address many of the difficulties explained in the
previous chapter and the desirable characteristics is proposed herein. Network system
architecture, smart sensor platform, middleware services, and algorithms to be employed
in this report are described.

3.2.1 Network system architecture

A homogeneous configuration with a single type of node is employed instead of the

tiered system approaches employing powerful upper-level nodes and less powerful lower-
level nodes. Note that even though only one type of node is deployed over a structure, a
PC may be needed in this architecture as an interface to users. Usage of only one type of
node allows for development of a simple and robust sensor network system, as explained
later in this section.

33
 Manager node Manager node
Cluster head node Cluster head node
Leaf node Leaf node
Base station Base station

(a) (b)
Figure 3.1. SHM system architecture: Different roles are assigned to nodes in (a) and (b).

In terms of functionality, smart sensor nodes in the proposed system are differentiated
as follows: the base station, the manager node, cluster heads, and leaf nodes. All of the
sensors deployed on a structure, in principle, work as leaf nodes. Leaf nodes receive
commands from the other nodes and perform preprogrammed tasks such as sensing, data
processing, and acknowledgment. A node in a local sensor community is assigned as a
cluster head and coordinates most of the communication and data processing in the
community. In addition to tasks inside the community, the cluster head communicates
with the cluster heads of the neighboring communities to exchange information. One of
the cluster heads also functions as the manager node. When intracluster RF
communication spans multiple clusters, the manager deals with time sharing among
clusters to avoid RF interference. The manager also exchanges packets with the leaf nodes
to manage operations in which all of the leaf nodes participate; sensing that is triggered by
the manager sensor is an example. The base station node is the gateway between smart
sensor networks and the PC. The PC with a user interface sends commands and
parameters to smart sensor networks via the base station. The PC also receives data and
calculation results from the base station. While the base station can communicate with any
node in direct communication range, most communication involving the base station is
routed through the manager or cluster heads with the exception being transmission of a
large amount of data or calculation results from leaf nodes to the PC for debugging
purposes. Thus, smart sensors nodes are functionally differentiated into four categories
(see Figure 3.1).
This system architecture can be compared with tiered networks to clarify its
characteristics. A tiered network assumes ample hardware resources at the upper level
nodes. By reducing the constraints on hardware (i.e., power source, RAM space, flash
memory space, radio bandwidth, etc.), a tiered network aims to implement more
functionality on the network easily. However, such an assumption is not necessarily the
case. When smart sensors are installed on a long suspension bridge, there needs to be

34
many powerful upper-tier nodes connected to power sources. Preparing power sources for
these upper-tier nodes spatially distributed over a long structure is expensive, negating one
of the appealing features of smart sensors. The proposed system architecture using a
homogeneous configuration does not assume such ample hardware resource availability.
Another advantage of a homogenous configuration is its robustness to node failure.
All of the smart sensors distributed over civil infrastructure in this architecture have the
same hardware configuration; thus, their roles can be reassigned in the case when some
nodes stop working due to battery exhaustion, OS failure, etc. Cluster heads and manager
sensors can be chosen from the survivors. Robustness to node failure is an important
feature for smart sensor networks, because node failure is an easily anticipated event.
Reassignment of roles can also be performed based on available hardware resources.
The roles of smart sensor nodes can be switched as shown in Figure 3.1. When the battery
voltage of a cluster head becomes lower than a threshold, another node in the
neighborhood with ample battery power can take over the role of the cluster head. If
necessary, only a part of the cluster head’s tasks can be reassigned to neighboring nodes.
In this manner, resources distributed over a large number of smart sensors can be
efficiently utilized.

3.2.2 Smart sensor platform

The Imote2 is chosen as the smart sensor platform for an SHM system. The Imote2
has much larger RAM space, larger flash memory size, and a faster on-board processor as
compared to other smart sensor platforms as shown on Table 2.1 on page 12. This
hardware specification is considered as being able to provide the necessary data
processing required by DCS for SHM. The base station Imote2 can communicate through
serial ports with the PC connected to a USB cable and the programming board.
The sensor board employed in this research has an LIS3L02DQ triaxial digital
accelerometer (STMicroelectronics, 2007). The accelerometer employs a sigma-delta
ADC (Stewart & Pfann, 1998) to obtain the digital signals. The ADCs produce bitstreams
representing the acceleration signals. Sinc3 filters are then applied to the bitstreams to
reconstruct the signal at the specified sampling rate. These reconstruction filters have a
low cutoff frequency so that the output signals are decimated without the signal being
aliased. The user can select a decimation factor from 8, 32, 64, and 128. The cutoff
frequency of the reconstruction filter and the sampling rate of the signal is fixed by the
decimation factor. The Imote2 processor board receives these digital signals through the
I2C or SPI ports on its basic I/O connector.
TinyOS is employed as the operating systems on smart sensors. This operating
system fits in small memory footprint and is suited for smart sensors with limited
resources. TinyOS has a large user community and many successful smart sensor
applications. However, from a civil engineering perspective, TinyOS may pose limitations
on the SHM system’s functionalities. TinyOS does not support real-time operations. In
other words, the operating system has only two types of threads of execution: tasks and
hardware event handlers, leaving users with only a little control to assign priority to

35
commands; execution timing cannot be arbitrary controlled. This feature of TinyOS needs
to be well-considered when designing a system

3.2.3 Middleware services

Middleware services are developed to provide functionalities that TinyOS does not
offer but that are needed by SHM applications. The middleware services considered
herein are model-based data aggregation, reliable communication, and synchronized
sensing. Model-based data aggregation utilizes application-specific knowledge to
efficiently collect information from measured data in a network. Model-based data
aggregation can save communication resources and contributes to scalability of a smart
sensor network. The reliable communication service enables communication without loss
of information by sending packets repeatedly. This service addresses the problem of lossy
communication. Synchronized sensing utilizes a time synchronization service and obtains
synchronized signals from a network of smart sensors. Synchronized clocks on smart
sensors in a network does not mean measured signals are synchronized, because smart
sensors cannot necessarily control sensing task timing precisely based on their clocks.
This service contributes to accurate measurement in terms of sensing timing. These
middleware services are realized on the smart sensor platform.

3.2.4 Damage detection algorithm

The damage detection algorithm employed in this work is an extension of the

Distributed Computing Strategy (DCS) for SHM proposed by Gao (2005). Smart sensors
form local sensor communities. Local sensor communities measure acceleration responses
of a structure and perform modal analysis. Locally determined modal parameters are then
utilized to construct a portion of the flexibility matrix of the structure. Changes in the
flexibility matrix before and after damage are then analyzed with Singular Value
Decomposition (SVD) to estimate the Damage Locating Vectors (DLVs). DLVs are
applied to the numerical model of the structure as input static force, and elements with
small stress are identified as potentially damaged elements. This damage detection is
performed in each local sensor community. Cluster heads communicate with each other in
order to exchange information about damaged elements. The details of this approach are
explained later in Chapter 6. By distributing and coordinating data processing, the DCS
for SHM and the proposed extension offers a solution for SHM system employing a dense
array of smart sensors.

3.3 Summary
This chapter describes desirable characteristics for smart sensor SHM systems and the
SHM architecture employed in this report. The difficulties, desirable characteristics, and
approaches in this report are summarized in Tables 3.1 and 3.2. The next chapter
demonstrates customizability of sensor boards, which leads to availability of appropriate
sensors.

36
Table 3.1. Difficulties in SHM Using Smart Sensors and Approaches in This Report

Difficulty Approach
• The Imote2 sensor board, which has a
triaxial accelerometer, is employed.
The performance of the accelerometer
Sensor hardware
is compared with reference sensors.
• Sensor board customizability is
demonstrated.
• Model-based data aggregation is
proposed.
• DCS for SHM allows distributed data
Data aggregation
processing, eliminating the need for
centrally collecting a large amount of
data.
• Synchronized sensing middleware
Time synchronization
service is proposed.
• Computational tasks are distributed by
employing model-based data
Limited computational capability aggregation and DCS for SHM.
• The Imote2 has a relatively fast
microprocessor as well as large RAM.
• Power consumption is kept moderate
by employing DCS for SHM and
model-based data aggregation.
Power • CPU clock frequency is switched,
depending on tasks to perform.
• Power harvesting is not addressed in
this research.

37
Table 3.2. Desirable Characteristics and Approaches in This Report

Desirable characteristic Approach

Scalable resource aware system
• DCS for SHM and model-based data
Scalability
aggregation offers scalability.
• RF and flash memory usage are kept
moderate through DCS for SHM and
model-based data aggregation, as well
as implementation ingenuity.
Power awareness
• The homogeneous hardware
configuration of the network potentially
offers reassignment of roles based on
available power.
• RF communication is managed by
manager node and cluster heads.
• Reliable communication protocols,
Bandwidth awareness
which wait for longer intervals in a
congested communication network, are
developed.
• SHM system is designed so that
requests for memory usage do not
Memory awareness
exceed the currently available memory
size of the Imote2, that is 256 kB.
• Computational tasks are distributed by
employing DCS for SHM and model-
based data aggregation.
Computational power awareness
• The frequency scalable processor on the
Imote2 is utilized to make efficient use
of computational power.
• Smart sensors get ready to enter sleep
mode.
Modal operation
• TinyOS that supports modal operation
is employed.
Autonomous distributed
embedded computing
• Model-based data aggregation is
Model-based data aggregation
proposed and implemented.
• DCS for SHM and model-based data
Collaborative distributed data
aggregation realize collaborative,
processing
distributed data processing.

38
Table 3.2. Desirable Characteristics and Approaches in This Report (Continued)

Desirable characteristic Approach

Autonomous initial configuration
• Not realized in this report.
and maintenance
Fault tolerant system
• Reliable communication protocols are
Message packet loss tolerance
proposed.
• The homogeneous hardware
Node failure tolerance configuration of the network potentially
addresses the node failure.
Byzantine error tolerance • Not realized in this report.
Desirable algorithmic
characteristics specific to SHM
for civil infrastructure
Multiscale information • Not realized in this report.
Collaboration in local sensor • DCS for SHM and model-based data
communities aggregation offers this collaboration.
• DCS for SHM gives redundancy
Redundancy
through overlapping clusters.
• Baseline measurement and continuous
SHM of a structure are implemented on
smart sensors.
Multiple functions
• Rare event monitoring, traffic
monitoring, local weather monitoring,
etc. are not realized.

39
Table 3.2. Desirable Characteristics and Approaches in This Report (Continued)

Desirable characteristic Approach

Desirable platform
characteristics specific to SHM
for civil infrastructure
• The Imote2 sensor board, which has a
triaxial accelerometer, is employed.
The performance of the accelerometer
Appropriate sensor availability
is compared with reference sensors.
• Sensor board customizability is
demonstrated.
• The Imote2 has relatively large RAM
and flash memory.
Sufficient RAM, flash memory, • The Imote2 has a ZigBee-compliant RF
and RF bandwidth component. The throughput is larger
than RF components on other smart
smart sensor platforms such as Mica2.
Environmentally hardened • Not realized in this report.
Open-source OS and interface • TinyOS is employed.

40
 Chapter 4

SENSOR BOARD CUSTOMIZATION

The lack of appropriate sensors and sensing components has hindered progress for
SHM applications with smart sensors. In this chapter, the customizability of sensor boards
is demonstrated. In particular, a strain sensor board and an AA-filter board will be
designed, constructed, and tested. Here, the Berkeley Mote, instead of the Imote2, is
chosen as the smart sensor platform because of its availability during the early stages of
this research. The customization procedure for the Imote2 is the same as that for the
Berkeley Mote. Also, the Berkeley Mote is designed for generic applications, while the
Imote2 has been developed for more data intensive applications. Development of a
Berkeley Mote sensor board demonstrates that a general-purpose sensor board can be
customized for SHM applications.
The Berkeley Mote was designed for generic applications; available sensors have not
been optimized for use in civil infrastructure applications. While acceleration
measurements are essential to obtain global responses of a structure, structural strains
provide an important indicator of local structural behavior. Studer and Peters (2004)
demonstrated that multiscale sensing yields better results than single-scale measurements
for damage identification. However, commercially available sensor boards only have
accelerometers for dynamic structural response measurement, and their applicability to
civil infrastructure is limited (Ruiz-Sandoval et al., 2006).
Although a great variety of strain sensors exist, strain sensor boards have yet to be
developed for the Berkeley Mote platform. The search for, selection of, and testing of an
appropriate strain sensor tailored to the Mote platform are presented herein. An AA filter
board to properly condition the signal from the strain sensor board to the Mica2’s ADC is
also designed, constructed, and tested. Experimental results demonstrate that the strain
sensor board has good resolution, and that the output is comparable with conventional
wired strain sensors.
A limitation in the Mica2’s hardware was identified during the experimental
verification - a systemic fluctuation was observed in the acquired data during reading/
writing of data by the Mica2’s microprocessor. This problem is addressed through careful
design of the analog circuit and subsequent digital signal processing.

4.1 Strain sensor board development

Many different types of strain sensors exist that use diverse methods for
measurement, such as mechanical, optical, or electrical means. Mechanical strain sensors,
for example a slide caliper, are simple but generally have poor resolution. A device with
levers to amplify strain to readable values can be devised, but it is prohibitively large.
Optical sensors are typically costly and too delicate to be densely deployed on civil
infrastructure. Electrical sensors, which include the piezoelectric sensor, semiconductor

41
 Figure 4.1. Foil strain gage.

strain gage, and the widely used foil strain gage (see Figure 4.1), have high resolution and
are small, sturdy, and inexpensive. Thus, they are good candidates for developing a
wireless strain sensor. Several desirable characteristics for a smart wireless strain sensors
for civil infrastructure applications are presented in the following paragraphs along with a
discussion of the associated impact on the design.
First, low-frequency responses (e.g., <1 Hz) typically found in tall or long civil
infrastructure need to be measured; therefore, sensors with DC capability are preferable.
Polyvinylidene fluoride (FVDF) film sensors, one type of piezoelectric sensor, are
appealing as they are rugged, inexpensive, and have low power requirements. However,
their sensitivity in the low frequency range is poor; researchers are working to resolve this
difficulty (Satpathi et al., 1999), but it is still problematic. As for semiconductor strain
gages, though large sensitivity to strain is an advantage, its sensitivity to variations in
temperature and tendency to drift are nontrivial disadvantages. In contrast, foil strain
gages (see Figure 4.1) have a wide frequency range and possess a DC capability. The gage
is a metallic resistor whose resistance varies almost linearly with its strain. A gage bonded
to an object deforms with the object's surface and yields resistance change. A circuit with
a Wheatstone bridge and amplifier converts the resistance change to a readable voltage
change. For this project, foil strain gages were chosen.
Second, low power consumption is an important characteristic as the wireless sensor
network usually operates on local battery power. For example, one prototype of the
Berkeley Mote can operate for up to a year on a single battery while in the power-down
mode. The power-down mode shuts off everything except for a watchdog and the
asynchronous interrupt logic necessary for wake up. However, it operates for only 30
hours at peak load (Hill et al., 2000). The power consumption in a strain sensor's
Wheatstone bridge is inversely proportional to the resistance. Thus, a high resistance
strain gage (e.g., 4,500 : is chosen as opposed to the widely used 120/350 : gages, in an
effort to moderate power consumption.
The measurement range selected is 1 to 2,000 PH. The lower limit is set based on the
resolution of a commercial wireless strain sensor product, the SG-Link Wireless Strain
Gauge System (MicroStrain, Inc., 2007). The upper limit is set based on the yielding strain
of steel. To achieve a wide measurement range despite the Mica2's 10-bit ADC restriction,
a variable gain amplifier was implemented.
Finally, a target noise level is sought that is equal to the target resolution, 1PH.
Significant high frequency noise was found to be present in the strain sensor board. The
low-pass filter described in the next section was employed to remove the high-frequency

42
 5.0V
LTC1682-5

100K:
:
-RG +RG

ga
.5
4

ge
-IN +VS Output

:
4
+IN Output

.5

5K
K

4.
:
-V REF
AD623

Figure 4.2. Wireless strain sensor circuit schematic.

Figure 4.3. Strain sensor board.

noise. This filter also reduces the problem of aliasing. To have a larger signal-to-noise
ratio and to mitigate problems with supply power fluctuation in the Mica2, a voltage
doubler/regulator, LTC1682-5 (Linear Technology, 2007), was added, which provides a
constant 5 V excitation for the strain bridge. Note that an amplifier with low noise, the
AD623 (Analog Devices, Inc., 2007), was selected for this circuit (see Figure 4.2).
The strain sensor board was fabricated as shown in Figure 4.3 (Screaming Circuit,
2007; PCBExpress, 2007). The terminal block in the top right corner of the strain sensor
board is wired to the strain gage’s terminals. Using appropriate wiring and the switch
settings on the board, the strain sensor board can be configured both as a quarter-bridge
circuit and a half-bridge circuit. Switches on the strain sensor board allow any of the three
different gages (i.e., 120/350/4,500 : strain gages) to be employed.

4.2 AA filter board development

Because the Berkeley Mote platform does not have antialiasing (AA) filters on the
processor/radio unit, a new AA filter board needs to be developed to address the aliasing
problem. When a continuous signal is sampled at a sampling frequency, f s , any signal

43
 As
dB
fn fc f sb fN
passband transition stopband
Frequency
Figure 4.4. AA filter design parameters.

component whose frequency is higher than the Nyquist frequency, f N = f s e  , is folded

back to the frequency range from 0 to f N , i.e., the signal is aliased. As a result of this
sampling phenomenon, signals in the frequency range above f N are superimposed onto the
original signal components in the baseband frequency range after the sampling. Once the
signal is contaminated with aliasing components, it cannot be corrected. Therefore, the
high-frequency components above the Nyquist frequency need to be eliminated prior to
the sampling process.
Ideally, a Linear Time Invariant (LTI) analog circuit, whose gain is unity over the
passband range with linear phase response and then attenuates quickly to zero is desirable.
In practice, LTI circuits close to the ideal circuit are employed as AA filters (see Figure
4.4). To completely eliminate aliasing in a digital signal, the following relation needs to
hold
f n  f c  f sb  f N (4.1)

where f n is the highest frequency of signal components to be analyzed; f c is filter’s

passband cutoff frequency; and f sb is the stopband cutoff frequency. The stopband
attenuation, A s , is determined so that any signal in the stopband is smaller than the
resolution of the ADC connected to the filter. Other requirements for AA filters include
small-gain variance in the passband; and linear phase over the passband, which keeps the
signals undistorted in the time domain.
There are many variations in LTI circuits for AA filters. Though an arbitrary LTI
system, which satisfies the above mentioned requirement, works as an AA filter, several
filter types have been proposed and used for their specific characteristics. The Butterworth
filter, which is also called the “maximally flat magnitude” filter, has a frequency response
that is as flat as mathematically possible in the passband. The Bessel filter has the
maximally linear phase response. The elliptic filter has an equiripple magnitude response
in both the passband and stopband, minimizing the maximum error in both bands. The

44
order of the filter design, or number of poles, is also a filter parameter. High-order filters,
in general, have more freedom in filter design, giving better magnitude characteristics at
the expense of design difficulty (Skolnick & Levine, 1997) and larger time lag; the slope
of a filter’s transfer function phase is the time lag.
If an 8-pole, elliptic-type AA filter along with 16-bit ADC is employed as in many

applications in civil engineering, the attenuation, A s is about -96 dB ( NQI  = ),
and f sb is 1.28 times f c . Therefore, the sampling frequency is set about 2.56 times the
highest frequency signal of interest. The combination of another type of AA filter and
ADC with different resolution gives different values for the attenuation and the stopband
cutoff frequency.
A 4-pole, Butterworth low-pass filter with a cutoff frequency of 50 Hz was selected
as the AA filter for the Berkeley Mote platform. The filter order and cutoff frequency were
set in order to investigate concurrently the ‘Tadeo’ acceleration sensor board developed
by Ruiz-Sandoval (2004) and the strain sensor board characteristics over a wide frequency
range (i.e., 0-50 Hz), and to limit the associated time lag. A small time lag is important for
applications with real-time response requirements, such as structural control and
triggering smart sensor tasks based on the measurement; high filter order and a low cutoff
frequency result in large deterministic time lags. From the AA perspective, this 4-pole
filter cannot completely eliminate aliasing for Mica2’s 10-bit ADC when sampled at 100
or 250 Hz. f sb for a 10-bit ADC is around 230 Hz. The noise is, however, unlikely to have
an amplitude as large as the ADC’s input range. Small noise components below f sb and all
of the components above f sb are completely eliminated by this filter.
There are three major designs to implement the AA filter as an electrical circuit: (a)
passive filter, (b) active filter, and (c) switched capacitor filter designs (National
Semiconductor, 1991). A passive filter is made up of passive components without
amplifiers, i.e., resistors, capacitors, and inductors. Though having a small noise level,
being free from power supply, and having the capability to deal with a large voltage and
current are advantages of this filter design, inflexibility in the design of the filter is
detrimental. Active filter designs employ operational amplifiers, eliminating the need for
the hard-to-handle inductors, and have more flexibility in design. The noise level is higher
than that of the passive design, but it is still moderate. Switched-capacitor filters are
clocked, sampled-data systems widely available in monolithic form. The cutoff frequency
is typically set by an external clock frequency. This filter has design flexibility, a variable
cutoff frequency, and insensitivity to temperature change. Poor DC precision and high
noise level, however, need to be addressed by additional external circuits, according to
application requirements. Frequency components higher than half the clock frequency
cannot be eliminated by this filter; they need to be eliminated with passive or active filters.
In this research, an active design was employed because of its design flexibility, moderate
noise level, and good DC precision.
A 4-pole Butterworth filter was realized by the Sallen-Key active filter design (Sallen
& Key, 1955), as shown in Figure 4.5. Rail-to-rail input/output amplifiers, MAX4132
(Maxim Integrated Products, Inc., 2007), give the filter a rail-to-rail input/output property,
which results in the efficient use of the input/output voltage range. The cutoff frequency

45
 8.32PF 3V 3V
3.45PF

Input 1K: 1K: 1K: 1K:

Output

1.22PF 2.94PF

Figure 4.5. Sallen-Key fourth-order active low-pass filter.

Figure 4.6. Four-pole AA filter board.

0 Design
Magnitude (dB)

Measured
-10

-20

-30
0 20 40 60 80 100 120
Frequency (Hz)
Figure 4.7. Four-pole AA filter board transfer function.

was set as 50 Hz by appropriate choice of resistors and capacitors. Note that filters with
other cutoff frequencies can be easily constructed by changing resistors and capacitors.
The final printed circuit board (see Figure 4.6) was manufactured at the Electronics
Services Shop (ESS, 2007) based on the cascaded two second-order filter design. The
filter's transfer function was verified as shown in Figure 4.7. The difference between the
design value and the measured value of the transfer function is within the expected range
of error due to finite precision of the capacitors on the board.
Moreover, another AA filter, an 8-pole Butterworth low-pass filter with a cutoff
frequency of 25 Hz, was designed and fabricated (PCBExpress, 2007; Screaming Circuit,
2007). This filter board has a stopband cutoff frequency, f sb , of about 60 Hz. Aliasing is
completely eliminated when sampled at 120 Hz or at a higher frequency, at the expense of

46
Figure 4.8. Eight-pole AA filter board.

0 Design
Magnitude (dB)

Measured
-20

-40

-60
0 10 20 30 40 50 60 70
Frequency (Hz)

Figure 4.9. Eight-pole AA filter board transfer function.

a larger deterministic time lag. This filter is well-suited for measurements in which no
signal of interest is above 25 Hz and the time lag of the filter, 34 ms, is acceptable. (see
Figures 4.8 and 4.9)

4.3 Experimental verification of strain sensor

The strain sensor/AA filter boards must be calibrated before use. These boards are
stacked on the Mica2, and shunt calibration is conducted to determine the sensitivity of the
strain sensor. Shunt calibration simulates a resistance change in the strain gage by
shunting the Wheatstone bridge with a known resistor. The output can then be calibrated.
The data acquired by the Mica2 is first stored on on-board flash memory and then
wirelessly transmitted to the base station (Mechitov et al., 2004). For convenience of
calibration, the strain sensor board is equipped with four switches to shunt the bridge with
different resistors that are wired in parallel with the strain gage. The 4-pole AA filter
board is employed for this calibration as well as the experiments described subsequently.
The sensitivity and offset of the strain sensor board is thus calibrated.
The sensor noise level must also be characterized. Based on the RMS of the measured
signal, the resolution of the strain gage is estimated to be 3.0 PH, which is slightly larger
than the target noise level. By using a more precise amplifier, as well as electromagnetic
shielding, further reduction in the noise level of the analog circuit is considered possible.
In addition, the flash memory writing process is a noise source. The writing process needs
a relatively large current and destabilizes the on-board power supply. Analog-to-digital

47
 80 w/o Digital Filter
w/ Digital Filter
60

Strain (v)
40
20
0
2 1 0 1 2
Time (sec)

Figure 4.10. Shunt calibration.

conversion was noticeably affected by this process, introducing peak noise corresponding
to flash memory access. This phenomenon needs to be suppressed.
Rather than seeking a hardware solution to reduce the noise level, the on-board
microprocessor was exploited to yield a better resolution of the strain sensor. A frequency
domain analysis indicates that the noise of the signal consists mainly of a 50 Hz
component. The natural frequencies of the experimental structure model, described in
detail in the following paragraph, are estimated to be smaller than 5 Hz. Consequently, a
12-pole Butterworth digital filter with a cutoff frequency of 25 Hz was employed to
eliminate the 50 Hz noise. Because the Mica2 does not possess sufficient computational
capability to apply the digital filter on-the-fly, the acceleration record is first stored in
flash memory and subsequently digitally processed. Flash memory access was
programmed to periodically take place at a frequency higher than the digital filter cutoff
frequency so that noise resulting from flash memory access is also suppressed by the
digital filter. In addition to simply applying a digital filter to the measurement data,
downsampling is embedded to efficiently utilize available memory. Because the filtered
signal does not have high-frequency components, the signal can be resampled at a lower
sampling frequency without aliasing. The digital filter is applied to the original data
sampled at 250 Hz, and then every fourth sample is stored on the flash memory. The
resultant sampling frequency is, therefore, 62.5 Hz. The shunt calibration procedure is
repeated with this digital filter (see Figure 4.10). The noise level of this strain sensor was,
thus, reduced to 0.2 PH.
The accuracy of the strain sensor/AA filter boards is experimentally verified using a
three-story scale model building (see Figure 4.11). A strain gage was installed on the first-
story wall of the structural model and connected to the strain sensor board on the Mica2.
The acquired data was transmitted to the base station using wireless communication. A
reference strain gage was also attached on this structure, and the gage was wired to a
conventional strain measurement system. The outputs of the wireless strain sensor and the
reference strain sensor were compared for the case where the structure was responding
under free vibration. Because the strain sensor board has an adjustable amplifier, in this
experiment the gain was adjusted to provide a large signal-to-noise ratio. Through these
analog circuit considerations and digital signal processing, the test on the three-story scale

48
 Accelerometer

Strain gage

Strain sensor board Base station

Figure 4.11. Experimental setup (a three-story building model and smart sensors).

600
Strain sensor board
Reference
400

200
Strain (PH )

-200

-400
10 11 12 13 14 15 16
Time (sec)

Figure 4.12. Strain sensor readings.

model showed that the wireless strain sensor has a good resolution and that the output is
comparable to conventional wired strain sensors (see Figure 4.12).

4.4 Summary
The customizability of sensor boards was demonstrated using the Berkeley Mote
platform. While strain is one of important physical quantities for SHM applications, strain
measurement using smart sensors has been rare. No strain sensor board was available for
the Berkeley Mote. Strain sensor and AA filter boards were developed and their

49
performance was experimentally validated. Sensor board customization procedures for
other smart sensor platforms are considered similar to those demonstrated in this chapter.
As demonstrated, physical quantities needed in SHM applications can be measured
by customized sensor boards. These physical quantities are processed and transferred by
middleware services and algorithms. In the next chapter, middlewere services will be
developed.

50
 Chapter 5

MIDDLEWARE SERVICES

In this chapter, middleware services for smart sensors are studied and realized.
Among the middleware services are data aggregation, reliable communication, and
synchronized sensing (Nagayama et al., 2007). These middleware services are
implemented on the Imote2 running TinyOS, while ideas behind these services are
generally applicable to other smart sensor platforms.

5.1 Data aggregation

The amount of data transferred in SHM applications is considerable. Long vibration
records are acquired from densely distributed smart sensors. If they are collected at a
single sink node using multihop communication, communication time easily exceeds the
time necessary for any other smart sensor task. The amount of data utilized in SHM
applications is first estimated. Distributed estimation of correlation functions is then
proposed as model-based data aggregation (Nagayama et al., 2006b, 2007; Spencer &
Nagayama, 2006). This data aggregation method is scalable to networks of large numbers
of smart sensors. Implementation issues are then discussed.

5.1.1 Estimate on data amount in SHM applications

To capture the vibration response of civil infrastructure, sensors need to acquire data
at an appropriate sampling rate for a sufficient period of time at various locations.
Analysis methods are among the factors in determining these parameters. While there are
many analysis methods for SHM applications, most output-only-measurement approaches
for civil infrastructure utilize Power Spectral Density (PSD) or Cross Spectral Density
(CSD) estimation. The amount of data involved in spectral density estimation for civil
infrastructure, thus, can be estimated.
The number of data points used in spectral density estimation is determined by the
number of FFT data points and the number of averages. Spectral density estimation is
performed by averaging the outcomes of FFT analyses. Though any power of two
theoretically works as the number of FFT data points, 1,024 or 2,048 points are most often
used and give reasonable results. The error in the spectral density estimation is inversely
proportional to the number of averages. A larger number of averages is, therefore,
advantageous. The drawback of having a large number of averages is the associated large
amount of data and long measurement time. If the measurement time is too long,
environmental conditions such as wind velocity and temperature may change during
measurement; averaging signals measured when a structure is behaving in a transient
manner is not desirable. The number of averages in spectral density estimation practically

51
ranges from 10 to 20. If 50 percent overlap is employed in the spectral density estimation
with 2,048 FFT data points, 20 time number of averages needs 21,504 data points.
There are other factors to be considered when the number of data points is
determined. One of them is the sampling frequency. The sampling frequency typically
ranges from 100 to 256 Hz for civil infrastructure. These sampling frequencies are high
enough considering the dominant modes of bridges, buildings, and towers. The
frequencies of dominants modes are usually lower than 10 Hz. The reasons why the
sampling frequency is much higher than the minimum frequency to capture vibrations
below 10 Hz include that higher modes may also be influential to the response, that a
periodic wave consists of multiple frequency components, that transient signal may have
high frequency components, and that ground motion, whose peak values are often utilized
in earthquake engineering, usually has higher-frequency components. A sampling
frequency of 256 Hz is a reasonable assumption for SHM system with smart sensors.
When a 0.1 Hz component of a signal sampled at 256 Hz is analyzed with an FFT of
length 1,024, only less than a half of the natural period fits in one window; the resolution
of FFT is too coarse. In such cases, the sampling frequency is lowered or the number of
FFT points is increased. Increasing the number of FFT points results in an increase in the
total number of data points.
The data type also affects the amount of memory required to store a measured signal.
The most often utilized are 16- and 32-bit integers, and double precision data. The sizes of
these data types are 2, 4, and 8 bytes, respectively. The effective number of bits of
resolution in the measured data determines the requirements on the data type. For
example, an acceleration signal with a resolution of 20 effective bits needs to use either a
32-bit integer or double precision data. Considering that many ADCs on smart sensors has
resolution poorer than 16-bit, 16-bit integers are most likely able to represent measured
sensor data.
The combination of 21,504 data points and 16-bit integer yields 43 kB of data per
sensing channel. Though this number will vary with changes in the number of data points
or the data type, 43 kB is used from here on as a typical amount of data generated per
sensing channel.
The number of sensing channels also depends on the applications. For example, one
of the bridges in Hong Kong famous for its densely instrumented sensors, Ting Kau
Bridge, has 67 channels to measure acceleration and 132 channels to measure strain.
These channels are normally sampled at 25.6 Hz. As such, hundreds of sensor nodes are
considered herein. One sensor node may have multiple sensing channels; for example,
acceleration measurement in three directions needs three channels. In total, this research
has a target of approximately a thousand channels.
A thousand sensing channels, each of which generates 43 kB of data, produce 43 MB
of data per measurement. Centrally collecting such a large amount of data is not practical.
Mechitov et al. (2004) reported that communication at the sink node is the bottle neck and
that collection of 480 kB of data from 16 Mica2s required more than 6,000 seconds. If
smart sensors with limited hardware resources, especially battery and RF components, are

52
used for SHM applications, transfer and processing of this large amount of data need to be
well-planned.

5.1.2 Model-based data aggregation

The amount of data involved in SHM applications normally exceeds practical

communication capabilities of smart sensor networks if all of the measurement data needs
to be collected centrally; data aggregation has been an important issue to be addressed
before an SHM system employing smart sensors is realized. One approach to overcome
this data aggregation problem has been an independent data processing strategy (i.e.,
without internode communication). This approach, however, cannot fully exploit
information available in the sensor network. Distribution of data processing and
coordination among smart sensors play a central role in addressing many smart sensor
implementation issues, including data aggregation. This section will demonstrate that
distribution and coordination can be well-planned so that the data aggregation problem is
addressed without sacrificing performance of the SHM algorithms.
The auto-correlation and cross-correlation functions are the inverse Fourier transform
of the PSD and CSD functions, respectively; their estimation is the beginning step of
many output-only SHM algorithms. CSD estimation requires data from two sensor nodes.
Measured data needs to be transmitted from one node to the other before data processing
takes place. Associated data communication can be prohibitively large without careful
consideration of the implementation. Distributed estimation of correlation functions is
proposed in this section.
Correlation functions are, in practice, estimated from finite length records. PSD and
CSD functions are estimated first through the following relation (Bendat & Piersol, 2000):
nd
ˆ 
G xy  Z  = --------- ¦ X i  Z Y i  Z  (5.1)
nd T
i=

ˆ
where G xy  Z  is an estimate of CSD G xy  Z  between two stationary Gaussian random
processes, x  t  and y  t  . X  Z  and Y  Z  are the Fourier transforms of x  t  and y  t  ; *
denotes the complex conjugate. T is time length of sample records, x i  t  and y i  t  . When
n d =  , the estimate has large random error. The random error is reduced by computing
an ensemble average from n d different or partially overlapped records. The normalized
ˆ
RMS error H  G xy  Z   of the spectral density function estimation is given as

ˆ Z  
H G xy = ------------------- (5.2)
J xy n d


 G xy  Z 
J xy  Z  = ------------------------------------ (5.3)
G xx  Z G yy  Z 

53
 Node 1
E ª¬ x1  t  xi  t  W  º¼
x1
i =1,2,…,ns

Node 2
x2
Node 3
x3
Node 4
x4
Node 5
x5
... Node ns
xns

Figure 5.1. Centralized correlation function estimation.


J xy  Z  is the coherence function between x  t  and y  t  , indicating the linear
dependence between the two signals. Through the averaging process, the estimation error
is reduced. Averaging 10 to 20 times is common practice. The estimated spectral densities
are then converted to correlation functions via the inverse Fourier transform.
An implementation of correlation function estimation for a small community of
sensors in a centralized data collection scheme is shown in Figure 5.1, where node 1
works as a reference sensor. Assuming n s nodes, including the reference node, are
measuring structural responses, each node acquires data and sends it to the reference node.
The reference node calculates the spectral density as in Eq. (5.1). This procedure is
repeated n d times and averaged. After averaging, the inverse Fourier transform is taken to
calculate the correlation function. All calculations take place at the reference node. When
the spectral densities are estimated from discrete time history records of length N , the
total data to be transmitted over the network using this approach is N u n d u  n s –   .
In the next scheme, data flow for correlation function estimation is examined and data
transfer is reorganized to take advantage of computational capability on each smart sensor
node. After the first measurement, the reference node broadcasts the time record to all of
the nodes. On receiving the record, each node calculates the spectral density between its
own data and the received record. This spectral density estimate is locally stored. The
nodes repeat this procedure n d times. After each measurement, the stored value is updated
by taking a weighted average between the stored value and the current estimate. In this
way, Eq. (5.1) is calculated on each node. Finally the inverse Fourier transform is taken of
the spectral density estimate locally. The resultant correlation function is sent back to the
reference node. Because subsequent modal analysis such as ERA uses, at most, half of the
correlation function data length, N e  data points are sent back to the reference node from
each node. The total data to be transmitted in this scheme is, therefore,
N u n d + N e  u  n s –   (see Figure 5.2).
As the number of nodes increases, the advantage of the second scheme, in terms of
communication requirements, becomes significant. The second approach requires data
transfer of O  N ˜  n d + n s   , while the first one needs to transmit to the reference sensor
node data of the size of O  N ˜ n d ˜ n s  . The distributed implementation leverages
knowledge regarding the application to reduce communication requirements as well as to
utilize the CPU and memory in a smart sensor network efficiently.
The data communication analysis above assumes that all the nodes are in single-hop
range of the reference node. This assumption is not necessarily the case for a general SHM

54
 Node 1
E ª¬ x1  t  x1  t  W  º¼
x1

Node 2
x2
Node 3
x3
Node 4
x4
Node 5
x5
... Node ns
xns
E ª¬ x1  t  x2  t  W  º¼ E ª¬ x1  t  x3  t  W  º¼ E ª¬ x1  t  x4  t  W  º¼ E ª¬ x1  t  x5  t  W  º¼ E ª¬ x1  t  xns  t  W  º¼

Figure 5.2. Distributed correlation function estimation.

application. However, Gao (2005) proposed a DCS for SHM which supports this idea.
Neighboring smart sensors in a single-hop communication range make local sensor
communities and perform SHM in the communities. In such applications, the assumption
of nodes being within single-hop range of a reference node is reasonable.

5.1.3 Implementation

In this section, the data aggregation strategy introduced previously is validated on a

network of Imote2s. To understand the performance of the algorithms, consider the truss
model shown in Figure 5.3. Vertical acceleration responses under random input are
measured at nodes 9, 11, 13, 15, 17, 19, 21, 23, and 25. The responses are then injected
into a network of nine Imote2s. The Imote2 corresponding to node 9 works as the
reference sensor. This node broadcasts its own data to the other nodes and collects
correlation function estimates. The length, N , of the injected data is 2,048 and the number
of FFT points is 512, resulting in seven averages with a 50 percent overlap. The sampling
rate is 500 Hz.
The acceleration response data and correlation function estimates are stored and
transferred as a 16-bit integer data type, considering that accelerometer can reasonably be
assumed to have a 12-bit resolution. In this way, memory space and bandwidth are
efficiently utilized, while double precision data is used during the spectral density
estimation and the inverse FFT calculations.
Simulation data is then reliably transferred to the destination nodes using a reliable
communication protocol described in the next section. In this way, outputs of this
correlation function estimate are compared with those calculated on a PC without being
mixed with the effect of data loss.
Correlation functions are estimated on the Imote2s and reported back to the reference
node. Figure 5.4 shows the estimated correlation function between nodes 9 and 11. This
estimate matches the corresponding correlation function estimated on a PC assuming the

            
6 10 14 18 22 26 30 34 38 42 46 50
3 7 11 15 19 23 31 35 39 43 47 51
27
2 5 9 13 17 21 25 29 33 37 41 45 49 53
1 4 8 12 16 20 24 28 32 36 40 44 48 52
              

Figure 5.3. Numerical model of a 2D truss structure.

55
 Correlation function estimate (m2/s4)
4

í2

í4
0 0.1 0.2 0.3 0.4 0.5
Time (sec)
Figure 5.4. A correlation function estimate.

associated quantization error. Because the difference between these two signals is smaller
– 
than  , only the estimate on the Imote2 is shown on Figure 5.4.
Data transferred in a network is reduced as explained in section 5.1.2. In this example,
data transmission is reduced by a factor of 5, as compared to centralized correlation
function estimation implementation. This reduction factor can be larger depending on the
number of sensors and the associated averaging. This reduction shows the advantage of
the distributed correlation function estimation.
Further consideration is necessary to accurately assess the efficacy of the distributed
implementation. Power consumption of smart sensor networks is not simply proportional
to the amount of data transmitted. Acknowledgment messages are also involved. The
radio listening mode consumes power, even when no data is received. However, the size
of the measured data is usually much larger than the size of the other messages to be sent
and should be considered the primary factor in determining power consumption. Small
data transfer requirements realized by the proposed model-based data aggregation
algorithm will lead to reduced power consumption.

5.2 Reliable communication

RF communication is not reliable unless lost packets are specifically addressed.
Packets may not be transmitted properly. When the distance between nodes is too long,
packets may not reach the destination. Multiple nodes trying to send packets at the same
time can cause packet collisions. If packets carrying commands are lost, destination nodes
fail to perform certain tasks. The sender is unsure whether the destination nodes have
received commands. If packets carrying measurement data are lost, destination nodes
cannot fully reconstruct the sender's data. Therefore, packet loss may cause a system to be
in an unknown state and may degrade measurement signals. SHM applications employing
smart sensors must address packet loss.

56
 When smart sensor applications involve more and more complicated internode data
processing and are assigned more and more tasks by commands sent through packets,
these commands needs to be delivered reliably. Reliable communication of short
messages is clearly a significant help in developing SHM systems with complicated
internode data processing.
The need to transfer a large amount of data reliably is not necessarily clear. In many
SHM research efforts, the data loss problem is not addressed. Loss of a few data points has
often been considered acceptable. However, the rationale behind accepting a small packet
loss rate is not clear. In section 5.2.1, the effects of data loss on SHM applications are
assessed.
The packet loss rate of the Imote2 is then experimentally examined. The packet loss
rate varies from experiment to experiment. An experiment with nodes close to each other
is expected to have less packet loss than an experiment with sparsely distributed nodes.
Packet loss rate is estimated under several conditions in section 5.2.2
Subsequently, reliable communication protocols suitable for sending large amounts of
data, as well as protocols to send single packets, are proposed.

5.2.1 The effects of data loss on SHM applications

In a wireless network, some packets are inevitably lost during communication unless
the communication protocol is specifically designed for reliable communication.
Conventional statistical, modal, and structural analyses of structural response data,
however, assume that no loss of data takes place. Some researchers have been working to
develop reliable communication without data loss, while others just ignore the data loss
effects on their analyses. Modal analysis and damage localization has not yet been
examined from the perspective of data loss. The impact of data loss on SHM applications
is investigated herein.
The Distributed Computing Strategy (DCS) for SHM described in section 6.4 is used
as a benchmark application. The correlation function and impulse response function
estimation, as well as modal analysis, employed as a part of DCS are widely used to
analyze ambient vibration data of civil infrastructure. The outcome of this data loss
analysis is applicable to many vibration-based SHM applications. The damage detection
method adapted in DCS is the DLV method, and the effect of data loss on this damage
detection method is also investigated. Understanding the effect of data loss may provide
insight into how to accommodate communication with data loss that is less demanding on
resource limited smart sensors than the communication without data loss (Nagayama et
al., 2007).
A computer simulation study is conducted for a truss model (see Figure 5.3),
assuming various data loss levels to investigate the data loss effect. Smart sensors are
assumed to be placed at the 13 nodes on the lower chord to measure the vertical
acceleration. The vertical input excitation at node 11 is measured. The sampling frequency
is set at 380 Hz so that the Nyquist frequency is above the fourth natural frequency of the
structure. After the data is acquired, a certain percentage of the data is randomly dropped

57
Figure 5.5. Effect of data loss and observation noise: acceleration power spectral density.

to simulate data loss. Data loss is assumed to take place only at this step, where the largest
amount of data is transferred among the sensor nodes.
Data is processed following each step of DCS. The impulse response functions
associated with each measurement points are estimated. Then the impulse response
functions are fed into the ERA routine for modal analysis. The natural frequency, damping
ratio, and mode shapes are among the major outputs. These modal parameters yield
flexibility matrix estimations. Then, one of the truss elements of the structure, element 9,
is replaced with a damaged element, which has a 40 percent cross-section reduction. Data
acquisition, impulse response estimation, modal analysis, and flexibility matrix estimation
are repeated on this data. From the estimated flexibility matrices for an undamaged and
damaged model, the DLVs are calculated. The accumulated stress, small values of which
indicate possible damaged elements, is then calculated. This simulation is conducted
assuming several data loss levels.
For the results of the data loss analysis to be better interpreted, the outcome is
compared to that of computer simulation including observation noise (but without data
loss). A band-limited white noise is superimposed onto each of observed signals. RMS
noise level is specified as a percentage of the RMS of physical responses.
The effect of data loss is first examined by comparing the PSD and coherence
functions of the measurement data. Figure 5.5 shows representative PSD functions
calculated with and without noise/data loss. The PSD’s peaks remain nearly unchanged
when the data loss or the noise is introduced; however, zeros are blurred. Although further
investigation is necessary for quantitative judgment, these results indicate that data loss of
0.5 percent and 5 percent observation noise have similar impact on PSD estimation.
Coherence functions indicate the degree of linearity between two variables. In this
computer simulation, the input excitation is applied only at node 11. Because the response
of the truss structure is linear, the coherence function between the two measurement
signals is expected to be unity over the entire frequency range. When no data loss and no
observation noise are considered, the coherence function is indeed close to one, as shown

58
 (a) (b)
Figure 5.6. Effect of data loss and observation noise: coherence functions between (a)
nodes 11 and 19; and (b) nodes 19 and 25.

in Figure 5.6. The isolated downswings correspond to zeroes of the system; extremely
small responses of the structure at these frequencies at the zeros result in numerical error,
giving small coherence function.
The loss of 0.5 percent of data yields a much smaller coherence function, except at the
system’s poles (see Figure 5.6). Because of the random nature of data loss occurrence and
the excitation, the coherence function varies from simulation to simulation. Twenty time
average of the coherence functions are shown in this figure. The coherence function’s
discrepancy from unity depends on the two measurement nodes between which the
function is calculated. All of the investigated coherence function plots, however, support
that the loss of 0.5 percent of data affects the coherence function in a similar way as 5 to
10 percent measurement noise addition.
Though the data loss clearly impacts the PSD estimation and coherence function
negatively, the consequence in subsequent modal analyses is unclear. The modal
properties are mainly represented by frequency components near the natural frequency.
The PSD and coherence function are not affected by data loss and noise as much near the
poles. The ERA modal analysis method is then applied to the impulse response functions
to estimate the modal properties. Impulse response functions are the inverse Fourier
transform of the transfer functions from the input force to measured outputs, which is
calculated from CSD and PSDs.
Figure 5.7 shows how the estimates of the first natural frequency and the damping
ratio vary as a function of the data loss level. For comparison, these modal parameters are
also estimated from the simulation model subjected to measurement noise (see Figure
5.8). Again, a loss of 0.5 percent of data is approximately equivalent to 5 to 10 percent
measurement noise.
The first four modes are identified and these modal properties are used to estimate the
flexibility matrix F  . The flexibility matrix estimate Fˆ is compared with F  and the
estimation error is shown in Figure 5.9 in terms of the Frobenius norm. F represents the

59
Frobenius norm. Because only a limited number of modes are utilized in the flexibility
matrix estimation, the Frobenius norms do not approach zero at 0 percent data loss and 0

26.026 0.66
1st Natural Frequency (Hz)
26.024 0.65

1st Damping Ratio (%)

26.022
0.64
26.02
0.63
26.018
0.62
26.016
mean 0.61 mean
26.014 mean ± RMS
mean ± RMS
26.012 0.6
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Data loss (%) Data loss (%)
(a) (b)
Figure 5.7. Data loss dependency of (a) the 1st natural frequency; and (b) damping ratio.

26.026 0.66
1st Natural Frequency (Hz)

26.024 0.65
1st Damping Ratio (%)

26.022
0.64
26.02
0.63
26.018
0.62
26.016
26.014 mean 0.61 mean
mean ± RMS mean ± RMS
26.012 0.6
0 2 4 6 8 10 0 2 4 6 8 10
Noise Level (%) Noise Level (%)
(a) (b)
Figure 5.8. Noise level dependency of (a) the1st natural frequency; and (b) damping ratio

0.11 0.2
Normalized accumulated stress

mean
mean ± RMS
0.1
0.15
@ 0||F/||F0||F

0.09
0.1
||F-F

0.08

0.07 0.05
mean
mean ± RMS
0.06 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Data loss (%) Data loss (%)
(a) (b)
Figure 5.9. Data loss dependency of (a) flexibility matrix estimation error; and (b)
normalized accumulated stress in the damaged element.

60
percent noise. The flexibility matrix is estimated both before and after damage is
introduced, and then the DLV method is applied to these matrices. The accumulated stress
calculated in the DLV method is an indication of damage. The stress in a damaged
member should be zero. The accumulated stress in the damaged element is plotted on
Figure 5.9, together with the flexibility matrix estimation error. The stress is indeed zero
for 0 percent data loss. As the amount of data loss increases, the accumulated stress
becomes larger. A stress value around 0.1 is considered small, indicating that the element
is damaged. The estimation error for these quantities due to noise is shown in Figure 5.10
for comparison. From these figures, a loss of 0.5 percent data is considered to be
equivalent to 5 to 10 percent of measurement noise.
From this analysis, data loss is found to introduce error into the modal analysis; the
error affects the subsequent analysis. As data loss increase, the quality of data degrades.
Because of the statistical approach (i.e., averaging) associated with the PSD and the CSD
estimation, a certain amount of data loss can be accommodated. A loss of 0.5 percent data
might be acceptable, considering that a corresponding 5 to 10 percent observation noise is
unexceptional in the monitoring of civil infrastructure.

5.2.2 Packet loss estimation in RF communication

The packet loss rate is experimentally estimated in this section. Together with the
results from the previous section, experimental data are examined to see whether the
packet loss rate is acceptable for SHM applications.

1. Experimental setup

Eight Imote2s, one programming board, and a PC were configured to perform this
experiment. One of the Imote2s is programmed as the base station. Another Imote2 works
as the sender. The other six Imote2s are configured to be receivers. A java program
running on the PC reads parameters from an input file and sends commands to the base
station node. The base station receives information about the number of receivers, node

0.11 0.2
Normalized accumulated stress

mean
mean ± RMS
0.1
0.15
@ 0||F/||F0||F

0.09
0.1
||F-F

0.08

0.05
0.07
mean
mean ± RMS
0.06 0
0 2 4 6 8 10 0 2 4 6 8 10
Noise Level (%) Noise Level (%)
(a) (b)
Figure 5.10. Noise level dependency of (a) flexibility matrix estimation error; and (b)
normalized accumulated stress in the damaged element.

61
 90

80 Node1
Node2
70 Node3
Node4
60 Node5
Node6
50

40

30

20

10

0
1 2 3 4 5 6 7 8 9 10
Trial

Figure 5.11. Packet loss rate estimation.

IDs of the sender and receivers, the number of packets to be sent, and the number of
repetitions. The base station forwards this information to the sender. On reception, the
sender starts broadcasting packets. In this experiment, 100 packets are broadcast. After the
sender completes transmission of the 100 packets, the sender tells the receiver that
transmission is complete and queries each receiver regarding how many packets were
received. This procedure is repeated 10 times.
This experiment is conducted on the lawn in front of Newmark Civil Engineering
Laboratory at the University of Illinois at Urbana-Champaign. The sender and receiver
nodes are held at the height of about 1 m. The distance between the sender and the receiver
is 3.3 m. Figure 5.11 summarizes the results. In this experiment, the packet loss is usually
smaller than 20 percent. There are many rounds of data transfer without any lost packets,
while the maximum packet loss rate reached 86 percent. The packet loss rate does not
show a clear trend among the six nodes. One node without any packet loss in one round of
data transfer may suffer from severe data loss in the subsequent round of communication.
While a loss of no or a few packets is anticipated to take place often in communication,
such a low packet loss rate cannot always be assumed. Smart sensor communication
sometimes experiences a loss of a large number of packets.
A packet loss of 20 percent or 86 percent is apparently much larger than the data loss
level discussed in the previous section and is not acceptable for SHM applications. A
reliable communication protocol suitable for transfer of a large amount of data, as well as
a protocol for short messages, is proposed in the next section.

5.2.3 Reliable communication protocol

Reliable communication protocols for long data records and short messages are
developed in this section. Communication packets range from one-bit acknowledgments
to lengthy acceleration time histories. A communication protocol suitable for long data

62
records is not necessarily efficient for single packet transfer. Communication protocol for
long data records and single packets are, therefore, designed separately.
In the SHM architecture employed in this research, most communication takes place
inside a smart sensor cluster as unicast or multicast. Unicast is communication between
one sender and one receiver, while multicast is communication from one sender to
multiple receivers. Both unicast and multicast protocols are, therefore, developed for long
data records and for short message communication. Communication protocols developed
in this chapter suffice for most communication requirements in DCS for SHM with the
exception being communication involved in time synchronization.

1. Communication protocol for long data records

Communication protocols suitable to transfer long data records reliably are

developed. The length of data assumed in this section is the acceleration record length
used in DCS. The first step of DCS estimates the CSD using FFT. The estimation is
frequently based on 10 to 20 times the number of data points used in the FFT. Assuming
the number of FFT points is 1,024, the target length of data to be transmitted is set as
11,264.
Though repeated data transmission without acknowledgment can statistically improve
the reliability of communication, such protocols cannot guarantee communication success
rate deterministically. If the packet loss rate is expected to be approximately constant over
long time, repeated sending practically results in no data loss; the number of repetitions
can be dynamically adjusted based on the packet loss rate of the last communication.
When many packets are lost statistically, the number of repetitions can be increased.
However, in smart sensor networks, burst packet loss may take place. RF transmission
devices nearby, for example, may cause the large number of communication packets to be
dropped. The fluctuation in packet loss rate is considered large, as shown in Figure 5.11.
Repeated transmission is not a sufficient solution to the packet loss problem.
Acknowledgment packets potentially guarantee reliable communication between
nodes. However, a poorly-designed communication protocol involving acknowledgment
messages can be notably inefficient. Many acknowledgment messages may be required,
waiting times may be long, and the same packets may need to be sent many times. With
careful consideration of efficiency, reliable communication protocols are realized in this
section using acknowledgment messages. These protocols are similar to error control
methods for data transmission, which are briefly reviewed and then modified for use in the
proposed reliable communication protocol for SHM applications.
Automatic Repeat reQuest (ARQ) is an error control method which repeats sending
packets based on the request from the receiver. On reception of packets without error, the
receiver replies with a positive acknowledgment (ACK). If an error is detected, the
receiver sends a negative acknowledgment (NACK) and requests retransmission. There
are several ARQ protocols, some of which are described in the following paragraphs.
In the stop-and-wait protocol, the sender waits for an acknowledgment after the
transmission of each packet. If an ACK is received, the sender transmits the next packet.

63
On reception of a NACK, the sender resends the last packet. Simplicity is a main
advantage of this method. The IEEE 802.11 wireless LAN adapts the stop-and-wait
protocol. This protocol is, however, not efficient in the sense that the sender needs to wait
for acknowledgment of each packet.
In the go-back-N protocol, the sender transmits more than one packet without waiting
for acknowledgment. The receiver keeps receiving packets replying ACK until an error is
detected. If an error is found in the N-th packet, the N-th packet and all subsequent packets
are discarded. The receiver sends back a NACK with the number corresponding to the lost
packet. The sender resends packets from the N-th packet. Discarding the received packets
after the N-th packets lowers the efficiency of this protocol.
Selective-repeat protocol addresses the inefficiency in the go-back-N protocol. The
sender continues sending unsent packets. Once the receiver detects an error in a packet, a
NACK is sent back to the sender. Packets received after detection of an error are stored in
a reception buffer. Only the packet with an error is disregarded. On reception of a NACK,
the sender temporary stops sending unsent packets and resends the lost packet. Then the
sender continues to send the remaining packets.
These concepts of ARQ can be utilized to guarantee reliable communication among
smart sensors. The reliable communication protocol developed by Mechitov et al. (2004)
sends a set of packets and then waits for acknowledgment. If the receiver does not receive
ACK, the same set of packets is sent again. Upon reception of ACK, the sender moves on
to the next set of packets. Even when one of packets sent in a group is lost, all of the
packets are sent again, reducing the efficiency of the protocol.
In this section, concepts from the ARQ protocol are modified for reliable wireless
communication. The RF component on the Imote2 is in either a listening mode or a
transmission mode. During transmission, the Imote2 cannot receive packets. However, in
selective-repeat protocol, transmission and reception are deeply interwoven. The receiver
may send an acknowledgment while the sender is in transmission mode. Packet loss and
retransmission are expected to be more frequent if the ARQ protocol is implemented
directly. Scheduling interwoven transmission and reception may result in long waiting
times. In the proposed protocols, the sender transmits all of the packets before this node
expects an acknowledgment packet. The receiver stores all of the received data in a buffer.
Once the receiver gets a message indicating the end of data transmission, the receiver
replies to the sender, indicating which packets are missing. Only missing packets are sent
again. In this way, the number of acknowledgments and retransmissions can be greatly
reduced.
This protocol is designed to send either 64-bit double precision data, 32-bit integers,
or 16-bit integers. Many ADCs on traditional data acquisition systems have a resolution
less than 16 bits, supporting the need for transfer of 16-bit integer format data. Some
ADCs have a resolution better than 16 bits, necessitating data transfer in 32-bit integer
format. Once an acceleration record is processed, the outcome may need more bits.
Onboard data processing such as FFTs and SVDs is usually performed using double
precision calculations. Even when the effective number of bits is smaller than 32,
debugging of onboard data processing greatly benefits from transfer of double precision

64
 Sender Receiver

Send parameters
• Sender sends parameters to the receiver.
(e.g. data length, sensitivity, offset, data
type, message ID, etc.)
• Sender keeps sending the parameters until Acknowledge
acknowledgement packet is received. • Receiver saves received parameters and
return acknowledgement packet
Data
• Sender sends series of data packets Receive data
• Receiver saves received data and keeps
track of missing packets.
End of data
• Sender notifies the receiver that all the data
is sent. This message is repeatedly sent
Acknowledgement/Request missing
until acknowledgement is received.
packets
• Receiver acknowledges. If receiver missed
no
Missing packets? packets, request sender resend the missing
packets.
yes
Send missing data Receive data
• Sender reply to the request. • Receiver saves received data and keeps
track of missing packets.

Done
• Sender tells the receiver that the data Acknowledge
transfer is complete. Sender keeps sending • Receiver acknowledges.
packets until an acknowledgement packet is
received.

Figure 5.12. Simple block diagram of the reliable unicast protocol for long data records.

data; data processing results on Imote2s can be directly compared with those on a PC,
which are most likely in double precision format. Transfer of 64-bit double precision data
is supported based on such needs.

Unicast

In the unicast protocol, the sender seeks to reliably deliver long data records to the
receiver utilizing acknowledgment messages. Figure 5.12 shows a simple flowchart of this
protocol.The details of the protocol are explained in the following paragraphs as well as in
the block diagram in Figure 5.13.
On start-up, the communication protocol requests that the main application assign a
buffer space to keep data to be transferred. This buffer needs to be large enough to hold
the entire data record. When 11,264 data points in 16-bit integer format are transferred,
this buffer size is about 22.5 kB. The main application returns the pointer to this buffer.
The need for this large buffer size is a drawback of this approach. However, with
knowledge of the application, the main application can utilize this memory space for other
purposes when communication is not expected to take place. This buffer is allocated in the
main application and only the pointer is given to the communication program.
When the application is ready to send data, the sender informs the receiver of the
necessary parameters (e.g., data length, data type, message ID, destination node ID, source
node ID, etc.). If the data is in integer format, the sensitivity and offset of the data are also
conveyed to the receiver. A packet containing these parameters is sent to the receiver.

65
 Sender Receiver
SendLData.send()
Pack parameters
(e.g., Destination, message
ID, data length, sensitivity, ReceiveNoticeMsg.receive()
etc. ) Receiver stores received
parameters.
Afterward, message ID and
SendNoticeSubTask()
sender’s address will be
used to rejects packets for
SendNoticeMsg.sendDone() other communication
call TimerNotice.start() request.

TimerNotice.fired() SendAckMsg()
if acknowledged,
post SendDataTask() ReceiveNoticeMsg.receive()
SendNoticeSubTask()
else set flag as acknowledged
post SendNoticeSubTask()

SendDataTask()
Pack a packet with data
TimerData.start
ReceiveDataMsg.receive
TimerData.fired Receiver stores received
data
SendDataMsg.sendDone
if all the data is sent
post SendNoticeEnd
else
post SendDataTask

SendNoticeEnd()
Pack a packet. Notification
that sender has sent all the
data
ReceiveNoticeMsg.receive()
Receiver checks for
SendNoticeSubTask
missing packets and report
to the sender.
SendNoticeMsg.sendDone()
call TimerNotice.start()

TimerNotice.fired()
if acknowledged,
post
Comment3CheckUnicast()
else SendAckMsg()
post ReceiveNoticeMsg.receive()
SendNoticeSubTask() set flag as acknowledged SendNoticeSubTask()
check request for resending
Comment3CheckUnicast()
if there are missing packets
post SendDataTask
else
pack a packet with ‘Done’
message
post SendNoticeSubTask

ReceiveNoticeMsg.receive()
SendNoticeSubTask Acknowledge the sender
Signal received()
SendNoticeMsg.sendDone()
call TimerNotice.start()

TimerNotice.fired() SendAckMsg()
if acknowledged,
signal sendDone() ReceiveNoticeMsg.receive()
SendNoticeSubTask()
else set flag as acknowledged
post
SendNoticeSubTask()

Figure 5.13. Detailed block diagram of the reliable unicast protocol for long data records.

66
After this transmission, the sender periodically checks whether an acknowledgment from
the receiver has arrived. If not, this packet is sent again. The message ID of the sender is
incremented when the sender engages in the new round of reliable communication. Once
two nodes engage in a round of reliable communication, packets with different message
IDs, destination addresses, or source addresses are disregarded.
To appropriately interpret the received packet, the Imote2 uses a 1-byte comment on
each packet. For example, a packet from the sender with the comment ‘1’ is interpreted as
the first packet with parameters such as data length, data type, etc. On reception of a
packet, tasks predetermined for the comment are executed.
Also, the current state of the sender and the receivers are internally maintained using
another 1-byte variable on each node. For example, the sender’s current state is ‘1’ after
sending the packet with the comment ‘1’. This state changes to ‘2’ when the sender finds
that an acknowledgment for this packet is received. The tasks to be executed in events
such as packet reception and timer firing are determined based on the current state and the
comment on the most recently received packet.
Upon reception of a packet with the comment ‘1’ from the sender, the receiver checks
whether it is ready to receive data. Unless this node has already agreed to receive data
from another node or send data to another node, this node replies to the sender with an
acknowledgment that it will receive the data from the sender.
When the sender notices that the acknowledgment has arrived, the periodical check
for acknowledgment stops and the entire data record is sent to the receiver. One packet
contains either three double precision data points, six 32-bit integers, or twelve 16-bit
integers. Each packet also contains a 2-byte node ID for the sender and a 2-byte packet
number, which the receiver uses to reconstruct the original data in the buffer. To keep
track of received packets, the receiver has an array, which is initialized to zeroes on
reception of a packet with the comment ‘1’. One bit of the array is assigned to each data
packet. Upon reception of each packet, the corresponding bit is changed to one. By
examining this array and the total number of packets calculated from the data type and the
total length of data, the receiver knows which packets are missing.
When the sender finishes transmitting all of the data, the end of the data is signaled to
the receiver by sending a packet with the comment ‘2’. The sender periodically checks the
response from the receiver. If nothing is received, this packet is sent again.
Knowing that all of the data were transmitted from the sender, the receiver checks the
array to keep track of packet reception. If all of the packets are received, an
acknowledgment is returned to the sender. The sender tells the receiver that the data
transfer is complete by returning a packet with the comment ‘4’ to the receiver. This
message is also resent until acknowledgment is returned from the receiver. When the
receiver first acknowledges the packet with the comment ‘4’, the receiver signals to the
main application on the receiver that a set of data has been successfully received, thus
completing the receiving activity on this node. Once the sender receives this
acknowledgment, the main application on the sender is notified that the data has been sent
successfully, thus completing the sending activity on this node.

67
 If there are missing packets, the receiver responds to the end-of-data notice by
sending a packet with the missing packets’ IDs. The first eight missing packets’ IDs are
sent to the sender. The sender packs these missing packets and sends them to the receiver.
At the end of transmission, the packet with the comment ‘2’ is sent again. If there are
missing packets, the same procedure is repeated. If all of the data is received, an
acknowledgment is returned, and the communication activity is completed as described in
the previous paragraph.
One of the difficulties with this protocol is to judge when communication ends. The
sender can clearly know the end of communication when the acknowledgment message
with the comment ‘4’ is received. However, the receiver cannot tell the end of
communication in a strict sense. The receiver cannot know whether the acknowledgment
with the comment ‘4’ reached the sender or not. If the receiver does not receive any packet
after the transmission of the acknowledgment packet with the comment ‘4’, there are two
possibilities: The acknowledgment packet reached the sender and this round of
communication was successfully completed; or the acknowledgment packet did not reach
the sender and the sender is periodically transmitting packets with comment ‘4’, none of
which reaches the receiver. There is no way to judge between the two possibilities.
This inability to judge the end of communication may cause a serious problem. If
only the receiver is disengaged from the communication, then the sender will keep
sending the last packet to the receiver; the receiver never replies because the receiver is
already disengaged.
This problem is addressed by storing message and source IDs of the last 15 rounds of
reliable communication and separating the last acknowledgment process from the rest of
the process. When the receiver gets the packet with the comment ‘4’ for the first time, this
node sends an acknowledgment back and is disengaged from this communication activity.
The message and source IDs are stored on this node. If this node later receives the same
packet, the structure storing the message and source IDs is searched. Once the receiver
finds the message and source IDs of the received packet in the structure, an
acknowledgment is sent back to the sender. This acknowledgment with the comment ‘4’
which is returned to the sender is sent using a message type assigned only for this purpose.
Even when the receiver is engaged in the next round of communication, this process of
acknowledgment can be done without affecting the subsequent or ongoing communication
activities or internal variables.

Multicast

In the multicast communication protocol, the sender reliably transmits long data
records to multiple receivers utilizing acknowledgment messages. Figure 5.14 shows a
simple flowchart of this protocol. The details of this protocol are explained in the
following paragraphs as well as in the block diagram in Figure 5.15.
The primary difference between the multicast and unicast protocols is the first step.
Before the sender sends the required parameters, such as data length and data type, the
node IDs of the receivers are broadcast, as well as the source and message IDs. The

68
 Sender Receiver
Inquiry
• Inquire of receivers whether they are ready. Acknowledge
• Message ID is sent to the receiver • Receiver saves received parameters and
return acknowledgement packet
Send parameters
• Sender sends parameters to the receiver.
(e.g. data length, sensitivity, data type, etc.) Acknowledge
• Sender keeps sending the parameters until • Receiver saves received parameters and
acknowledgement packet is received. return acknowledgement packet

Data Receive data

• Sender sends series of data packets • Receiver saves received data and keeps
track of missing packets.
End of data
• Sender notifies the receiver that all the data
is sent. This message is repeatedly sent
Acknowledgement/Request missing
until acknowledgement is received.
packets
• Receiver acknowledges. If receiver missed
no
Missing packets? packets, request sender resend the missing
packets.
yes
Send missing data Receive data
• Sender reply to the request. • Receiver saves received data and keeps
track of missing packets.

Done
• Sender tells the receiver that the data
transfer is complete. Sender keeps sending Acknowledge
packets until an acknowledgement packet is • Receiver acknowledges.
received.

Figure 5.14. Simple block diagram of the reliable multicast protocol for long data

comment field of this packet is ‘f’ in hexadecimal. A maximum of eight receiver node IDs
are sent in a single packet. When data is sent to more than eight nodes, multiple packets
containing the receiver node IDs are broadcast. As is the case for unicast communication,
the sender periodically sends these packets until the receivers reply.
Upon reception of this initial packet, a node check is conducted to determine whether
the received node IDs contain its own node ID. If the node is not one of the destination
nodes, this node simply disregards the packet. If the node is one of destination nodes, this
node sends an acknowledgment to the sender and commits itself to this round of
communication. Note that if multiple receivers reply to the sender at the same time, packet
collision may take place; timing for replies to the sender needs to be scheduled
appropriately. The order of node IDs in the packets with the comment ‘f’ is used to
schedule the timing for replies. For example, the node corresponding to the eighth node ID
in the packet with the ‘f’ comment waits 105 (= 15 ms x 7) ms before the acknowledgment
is sent back to the sender. The order of and the waiting time for the reply are stored on
each receiver; all of the following acknowledgment messages in this multicast protocol
use this schedule. Optimization of the waiting interval, 15 ms, may result in faster
communication, but such optimization is not pursued in this research.
Once the sender receives an acknowledgment from all of the receivers, the parameters
such as data length, sensitivity, offset, and data type are broadcast. Although the broadcast
transmission reaches all of the nodes in the neighborhood, only the engaged nodes process
the received packet. The source ID, the destination ID (i.e., ID ‘0xffff’ which corresponds
to broadcast address in TinyOS), and the message ID are used on the receiver to
distinguish packets to be processed from other packets.

69
 Sender Receiver
SendLData.bcast()
Pack parameters
(destination, messageID)
post BCBeginTask

BCBeginTask
Inquire of receivers whether
they are ready to receive
data.
if all the receivers ReceiveNoticeMsg.receive()
acknowledge, Receiver stores received
post BeginTask(); parameters.
Afterward, message ID and
sender’s address will be
SendNoticeSubTask()
used to rejects packets for
other communication
SendNoticeMsg.sendDone() request.
post BCBeginTask()

TimerNotice.fired() SendAckMsg()
if acknowledged,
post SendDataTask() ReceiveNoticeMsg.receive()
SendNoticeSubTask()
else set flag as acknowledged
post SendNoticeSubTask()

BeginTask
Pack parameters
(e.g., data length,
sensitivity, etc. )
ReceiveNoticeMsg.receive()
SendNoticeSubTask() Receiver stores received
parameters.
SendNoticeMsg.sendDone()
call TimerNotice.start()

TimerNotice.fired() SendAckMsg()
if acknowledged,
post SendDataTask() ReceiveNoticeMsg.receive()
SendNoticeSubTask()
else set flag as acknowledged
post SendNoticeSubTask()

SendDataTask()
Pack a packet with data
TimerData.start
ReceiveDataMsg.receive
TimerData.fired Receiver stores received
data
SendDataMsg.sendDone
if all the data is sent
post SendNoticeEnd
else
post SendDataTask

SendNoticeEnd()
Pack a packet. Notification
that sender has sent all the
data
ReceiveNoticeMsg.receive()
SendNoticeSubTask Receiver checks for
missing packets and report
to the sender.
SendNoticeMsg.sendDone()
call TimerNotice.start()

TimerNotice.fired()
if acknowledged,
post
Comment3Check()
else SendAckMsg()
post ReceiveNoticeMsg.receive()
SendNoticeSubTask() set flag as acknowledged SendNoticeSubTask()
check request for resending
Comment3Check()
if there are missing packets
post SendDataTask
else
pack a packet with ‘Done’
message
post SendNoticeSubTask
ReceiveNoticeMsg.receive()
SendNoticeSubTask Acknowledge the sender
Signal received()
SendNoticeMsg.sendDone()
call TimerNotice.start()

TimerNotice.fired()
if acknowledged, SendAckMsg()
signal sendDone()
else ReceiveNoticeMsg.receive()
SendNoticeSubTask()
post set flag as acknowledged
SendNoticeSubTask()

Figure 5.15. Detailed block diagram of the reliable multicast protocol for long data records

70
 Another difference between unicast and multicast is that the sender needs to check the
acknowledgment messages from all of the receivers. The sender keeps track of
acknowledgments from the receivers using an array. One bit of the array corresponds to
one receiver. Before the sender transmits a packet needing acknowledgement from the
receiver, this array is initialized to zeroes. Upon reception of an acknowledgment from a
receiver, the bit corresponding to that receiver is changed to one. While the size of this
array is set so that up to 30 nodes can be receivers currently, parameters can be easily
adjusted to accommodate a larger number of receivers at the expense of memory space
required for the array.
Another point to consider for this multicast communication protocol is the way that
requests for retransmission are handled. After sending the end-of-data notice to the
receivers, the sender waits for a reply from the respective receivers. Within the assigned
time slot, each receiver reports to the sender; a packet containing missing packet numbers
is sent back. After the time assigned for all the nodes to reply has passed, the sender
broadcasts the requested packets. At the end of broadcast of these requested packets, a
packet with the comment ‘2’ is sent, asking for packets still missing.
The end of one round is difficult to judge, as explained in the unicast reliable
communication protocol section. A similar solution is implemented for multicast reliable
communication. In addition to the message and source IDs, this multicast protocol needs
to store the time to wait for before replying with an acknowledgment message for the last
15 rounds of communications.

2. Communication protocol for short messages

Communication protocols suitable to transfer a single packet reliably are also needed.
SHM applications involve sending and receiving many commands, each of which fits in
one packet. These commands need to be delivered reliably. If a packet containing a
command to start sensing is lost, the destination node does not start sensing. If loss of a
packet is not detected, the sender assumes the destination nodes has performed tasks based
on the command. The current state of a smart sensor node is difficult to estimate without
reliable communication. Ensuring the performance of SHM systems without reliable
command delivery is extremely complex if not impossible. The reliable communication
protocol developed for long data records is, however, not efficient for single packet
transfer. Unicast and multicast reliable communication protocols suitable for single-
packet messages are developed in this section.
This protocol is again similar to the ARQ protocol. Because only one packet is sent,
an acknowledgment is returned to the sender for each packet, as is the case for the stop-
and-wait protocol. The protocol is designed to send commands and parameters in 8- and
16-bit integers. One packet can hold nine 16-bit integers and one 8-bit integer in addition
to the parameters necessary for reliable communication.

71
 Sender Receiver
Send a packet Acknowledge
• Sender sends a packet to the receiver. • Receiver saves received parameters and
• Sender keeps sending the packet until return acknowledgement packet
acknowledgement packet is received.

Figure 5.16. Simple block diagram of the reliable unicast protocol for short messages.

Sender Receiver
SendSMsg.send()
Pack parameters and data
(e.g., Destination, message
ID, data, etc. )
ReceiveShortMsg.receive()
Receiver stores received
SendShortSubTask() parameters and data.
Acknowledge the sender
SendShortMsg.sendDone() Signal received()
call TimerNotice.start()

TimerShort.fired() SendAckMsg()
if acknowledged,
signal SendDone() ReceiveNoticeMsg.receive()
SendShortSubTask()
else set flag as acknowledged
post SendShortSubTask()

Figure 5.17. Detailed block diagram of the reliable unicast protocol for short messages.

Sender Receiver
Inquiry Acknowledge
• Inquire of receivers whether they are ready • Receiver saves received parameters and
return acknowledgement packet

Send a packet
• Sender sends a packet to the receiver. Acknowledge
• Sender keeps sending the packet until • Receiver saves received parameters and
acknowledgement packet is received. return acknowledgement packet

Figure 5.18. Simple block diagram of the reliable multicast protocol for short messages.

Unicast

The sender transmits a packet, including the destination ID, source ID, message ID,
parameter, and commands. Until this packet is acknowledged, the sender continues to
transmit this packet. The receiver extracts information from the packet and sends an
acknowledgment back to the sender (see Figures 5.16 and 5.17)
The end of one round is difficult to judge, as explained in the reliable unicast protocol
for long data records. The same solution applies to short messages. The source and
message IDs for the last 15 rounds of communication are stored so that acknowledgment
can be sent even after the receiver is disengaged from this round of communication.

Multicast

The sender first transmits a packet including the receivers’ node IDs, as was the case
for the multicast communication protocol for long data records. Once all of the receivers
return an acknowledgment to the sender, the sender transmits a packet containing
parameters and commands. The receivers also acknowledge this packet. To judge the end

72
 Sender Receiver
SendSMsg.bcast()
Pack parameters
(e.g., Destination, message
ID)
post BCBeginTask

BCBeginTask
Inquire of receivers whether
they are ready to receive
data. ReceiveNoticeMsg.receive()
if all the receivers If the receiver is among the
acknowledge, destination nodes,
post BeginTask(); receiver stores received
parameters.
Afterward, message ID and
SendNoticeSubTask()
sender’s address will be
used to rejects packets for
SendNoticeMsg.sendDone() other communication
post BCBeginTask() request.

TimerNotice.fired() SendAckMsg()
if acknowledged,
post SendDataTask() ReceiveNoticeMsg.receive()
SendNoticeSubTask()
else set flag as acknowledged
post SendNoticeSubTask()

BeginTask()
Pack parameters and data
(e.g., Destination, message
ID, data, etc. )
ReceiveShortMsg.receive()
Receiver stores received
SendShortSubTask() parameters and data.
Acknowledge the sender
SendShortMsg.sendDone() Signal received()
call TimerNotice.start()

TimerShort.fired() SendAckMsg()
if acknowledged,
signal SendDone() ReceiveNoticeMsg.receive()
SendShortSubTask()
else set flag as acknowledged
post SendShortSubTask()

Figure 5.19. Detailed block diagram of the reliable multicast protocol for short messages.

of communication, the receiver keeps the source ID, message ID, and waiting time of the
last 15 rounds of communication (see Figures 5.18 and 5.19).

5.3 Synchronized sensing

Time synchronization error in a smart sensor network can cause inaccuracy in SHM
applications. Time synchronization is a middleware service common to smart sensor
applications and has been widely investigated. Each smart sensor has its own local clock,
which is not synchronized initially with other sensor nodes. By communicating with
surrounding nodes, smart sensors can assess relative differences among their local clocks.
For example, Mica2 motes employing TPSN are reported to synchronize with each other
to an accuracy of 50 Psec; different algorithms and hardware resources may result in
different precision. Whereas time synchronization protocols have been intensively
studied, requirements for synchronization from an application perspective have not been
clearly addressed. The effect of time synchronization on SHM applications is first studied.
As stated earlier, spectral density and correlation function estimation are oftentimes

73
utilized in SHM applications followed by modal analysis. Nagayama et al. (2007) has
discussed the effects of synchronization errors on these estimates. Time synchronization
accuracy realized on the Imote2, the smart sensor platform employed in this research, is
then estimated and evaluated for the SHM applications. Because time synchronization
among smart sensors does not necessarily offer synchronized measurement signals, issues
critical to synchronized sensing are then investigated. Finally, synchronized sensing is
realized utilizing resampling (Nagayama et al., 2006a; Spencer & Nagayama, 2006).

5.3.1 Time synchronization effect on SHM applications

Consider the signal x  t  from a smart sensor in the local clock coordinate t . This
signal can be written in terms of the reference (or global) clock t as
x  t  = x    + D t – Gt  (5.4)

where Gt is the initial time synchronization error and D is the clock drift rate. In the fre-
quency domain, this relationship is expressed as
 Gt Z
X  Z  = ------------- GZR § – iZ -------------· X § -------------· (5.5)
+D ©  + D¹ ©  + D¹

where X  Z  and X  Z  are the Fourier transform of x  t  and x  t  , respectively. In the

following derivation, D is assumed to be zero.
The effect of time synchronization errors on modal parameters, such as the natural
frequency, damping ratio, and mode shape, are examined. Though only output
measurements can be obtained during most of the civil infrastructure monitoring, for
completeness, the effect of time synchronization error is investigated both for the input-
output measurement case and the output-only measurement case.
When both input force and output structural responses are measured, transfer
functions from input to output signals are first estimated and then modal analysis follows.
Therefore, the effect of time synchronization error on transfer function estimates is
studied.
Consider a structure for which the equation of motion is written as
Mx··  t  + Cx·  t  + Kx  t  = f  t  (5.6)

where M , C , and K are mass, damping, and stiffness matrices, respectively. x  t  , x·  t  ,

x··  t  , and f  t  are the displacement, velocity, acceleration, and input force vectors,
respectively. By taking the Fourier transform of Eq. (5.6), the displacement vector can be
written as

X  Z  =  – MZ  + CiZ + K  – F  Z 
(5.7)
= T FX  Z F  Z 

74
where X  Z  and F  Z  are the Fourier transforms of x  t  and f  t  , respectively.
T FX  Z  is the transfer function from the input force to the displacement. The i-th element
of x  t  , x i  t  with time synchronization error of Gt xi is modeled as x i  t-Gt xi  . By
transforming to the frequency domain, the measured displacement vector of size n, x  t 
with time delay Gt i  i=1,2,...,n is written in frequency domain as

GZR  – i ZGt 1   } 
 GZR  – i ZGt 2  } 
XZ = XZ (5.8)
}
# }
# }
% }#
  } GZR  – i ZGt n 

A similar relation holds for the measured force f  t  . Therefore, the transfer function
from the input force to the displacement with synchronization error, T FX  Z  is expressed
as in Eq. (5.9),

GZR  – i ZGt x   } 
 GZR  – i ZGt x   } 
T FX  Z  = (5.9)
#
} #
} %
} }#
  } GZR  – i ZGt xn 

GZR  iZGt f   } 
 GZR  iZGt f   } 
T FX  Z 
#
} #
} %
} } #
  } GZR  iZGt fm 

where Gt fi is the time synchronization error for the i-th force measurement. Eq. (5.9)
shows that the time synchronization error does not affect the denominator of the transfer
function; natural frequencies and damping ratios determined only by the denominator are
immune to time synchronization error. The numerator, however, is multiplied by complex
values of unit magnitude, resulting in phase errors in the mode shapes.
When only output structural responses are measured, correlation functions between
the output measurements can be used to determine modal parameters, as explained in
section 6.1. Therefore, the effect of time synchronization on correlation functions and the
associated modal parameters is investigated.
Consider a cross correlation function R xi x ref  W  between the stationary random
response x i  t  at location i and reference signal x ref  t  . R xi xref  W  satisfies the equation of
motion for free vibration given by
·· ·
MR xx ref  W  + CR xx ref  W  + KR xx ref  W  =  W t  (5.10)

75
where R xx ref is a vector consisting of R x i xref  W  . When the sensor node at location i has a
time synchronization error of Gt xi relative to the reference node, the correlation function,
R xi xref  W  can be written as

R xi xref  W  = E > x i  t – Gt xi + W x ref  t  @ = R x i xref  t – Gt xi  (5.11)

where E> @ is the expectation operator. If Gt xi is positive, the correlation function for the
interval (0, Gt xi ) does not have the same characteristics as the succeeding signal. This
portion of the cross correlation function will correspond to negative damping; therefore,
the beginning portion of the signal needs to be removed. When Gt xi is unknown, a
segment corresponding to the maximum possible time synchronization error, Gt x max , is
truncated from the correlation function.
The correlation functions, each having independent time synchronization errors, do
not satisfy the equation of motion, Eq. (5.10). The correlation functions after the
truncation, R xi xref'  W  , however, can be decomposed into modal components as follows:

R xi x ref'  W  = )'/A

)' = M  M  } M 2n

I  GZR  – O  Gt x1  I  GZR  – O  Gt x1  } I 2n GZR  – O 2n Gt x1 

I  GZR  – O  Gt x  I  GZR  – O  Gt x  } I 2n2 GZR  – O 2n Gt x 
=
# # % #} (5.12)
I m GZR  – O  Gt xm  I m GZR  – O  Gt xm  } I 2nm GZR  – O 2n Gt x m 

/ = diag( GZR  O  W  GZR  O  W  } GZR  O 2n W  )


Oi = hi Zi + j Zi  – hi

A = diag( a  a  } a n )

where I ij is j-th element of i-th mode shape, h i and Z i are i-th modal damping ratio and
modal natural frequency, respectively, and a i is a factor accounting for the relative
contribution of the i-th mode in the correlation function matrix. Modal analysis techniques
such as ERA can be used to identify these modal parameters. As Eq. (5.12) shows, the
natural frequencies and damping ratios remain the same. The observed mode shapes )'
are, however, different from the original mode shapes; changes in mode shape amplitude
are negligible due to small h i and t xi ; however, phase shifts in the mode shapes can be
substantial. Mode shape phases can indicate structural damage and are important modal
characteristics from an SHM perspective. The requirements on time synchronization
accuracy for modal analysis need to be assessed mainly from the viewpoint of the mode
shape phase.

76
 Though the effects of time synchronization error on transfer functions and correlation
functions are shown, the estimates calculated from a finite length of measured data are not
exactly the same as these functions. This issue is not theoretically pursued herein.
However, numerical simulation conducted to date using finite length simulation data
records supports the relations given in Eqs (5.9) and (5.12). Thus, requirements on time
synchronization errors are concluded to be dictated primarily by their effect on mode
shape phase.

5.3.2 Estimation on time synchronization error

The accuracy of the time synchronization protocol implemented on the Imote2 is

evaluated. The Flooding Time Synchronization Protocol (FTSP) is implemented on the
Imote2 under the advice of Kirill Mechitov at the University of Illinois at Urbana-
Champaign and Intel Corporation. This protocol is first briefly reviewed.
FTSP utilizes time stamps on the sender and receivers. A beacon node broadcasts a
packet to the other nodes. At the end of the packet, a time stamp, t send , is appended just
before the packet is transmitted from the RF device on the beacon. Upon reception of the
packet, the receivers stamp the time, t receive , from their own local clocks. The delivery
time, t delivery , between these two events includes interrupt handling time, encoding time,
and decoding time. t delivery is usually not small enough to be ignored; the variance of
t delivery over time is usually small. t delivery can be estimated in several ways. An
oscilloscope connected to both nodes and on-board clocks can keep track of the
communication time stamp. In this research, t delivery is first assumed to be zero and then
adjusted so that Imote2s placed on the same shake table give synchronized acceleration
data. The phases of the transfer functions among Imote2 signals should be constant with
zero phase over wide frequency range if these nodes are synchronized. t delivery is
determined to give a constant phase of zero. The details of this adjustment are described in
section 5.3.4. The offset between the local clock on the receiver and the reference clock on
the sender is determined as t receive – t send – t delivery . This offset is subtracted from the
local time when global time stamps are requested afterward.
To evaluate time synchronization error, a group of nine Imote2s are programmed as
follows. The beacon node transmits a beacon signal every 4 seconds. The other eight
nodes estimate the global time using the beacon packet. Time synchronization is, thus,
performed. Two seconds after the beacon signal, the beacon node sends the second packet,
requesting replies. The receivers get time stamps on reception of this packet and convert
them to a global time stamp using the offset estimated 2 seconds before. These time
stamps are subject to two error sources: First, time synchronization error, and second,
delay in time stamping upon reception of the second packet. The receivers take turns to
report back these time stamps. This procedure is repeated more than 300 times. These time
stamps from the eight nodes are compared. Figure 5.20 shows the difference in the
respective global time stamps using one of the eight nodes as a reference. The time
synchronization error seems to be less than 10 Ps for most of the time. Scattered peaks
may indicate large synchronization error. Note that the time synchronization error is one

77
 80

60

40

20

Time (Psec)
0

-20

-40

-60

-80

-100
50 100 150 200 250 300
Repetition
Figure 5.20. Time synchronization error estimation.

of the two above-mentioned factors explaining the error in the time stamps. This figure
indicates that an upper-bound estimate of time synchronization error is 80 Ps.
The time synchronization error estimated above is considered small for SHM
applications. The delay of 10 Psec corresponds to a 0.072-degree phase error for a mode at
20 Hz. Even at 100 Hz, the corresponding phase error is only 0.36 degree.
While a global clock estimates 2 seconds after a beacon signal is found to be accurate,
local clocks may drift over time. Large clock drift necessitates frequent time
synchronization to maintain certain accuracy. The clock drift in the Imote2 is estimated
next.
The same approach is utilized to estimate clock drift. Upon reception of the second
packet, which requests replies, the receivers return to the sender their offsets to estimate
the global time, instead of global time stamps. If clocks on nodes are ticking at exactly the
same rate, the offsets should be constant over a long time. This experiment, however, did
not show constant offsets. Figure 5.21 shows the offsets of nine receiver nodes. One of
them stopped responding around 45 seconds, exhibiting a short line on the figure. This
figure shows that the drift is quite constant in time. The maximum clock drift among this
set of Imote2 nodes is estimated to be around 50 Ps per second. Note that this estimate
from the nine nodes is not the upper limit of clock drift because of the small sample size.
This drift is small but not negligible if measurement takes a long time. For example, after
200 second measurement, the time synchronization error may become as large as 10 ms.
One solution to address this clock drift problem is frequent time synchronization.
Time synchronization is performed often before the effect of clock drift accumulates and
results in large time synchronization error. However, frequent time synchronization is not
always feasible. When other tasks are running, such as sensing, time synchronization may
not perform well. Time synchronization requires precise time stamping as explained
earlier. Moreover, sensing requires precise timing and needs higher priority in execution.
Scheduling more than one high priority tasks is challenging, especially for an operating
system such as TinyOS which has no support for real-time control. If the time

78
 8000

6000

4000

Drift (Psec)
2000

-2000

-4000
20 40 60 80 100 120 140 160 180
Time (s ec)
Figure 5.21. Drift estimation.

synchronization interval is shorter than the sensing duration, another solution needs to be
sought to avoid scheduling both time synchronization and sensing tasks.
Alternatively, a clock’s drift rate can be compensated directly. The slopes of the lines
in Figure 5.21 are nearly constant and provide an estimate of the clock drift rate to be
compensated. If time synchronization offset values can be observed for a certain amount
of time, the slope can be estimated using a least-squares approach.

5.3.3 Issues toward synchronized sensing

Accurate synchronization of local clocks on Imote2s does not guarantee that

measured signals are synchronized. Measurement timing cannot necessarily be controlled
based on the global time.
Sensing on Imote2s is performed in the following way. A sensing application posts a
task to prepare for sensing. Parameters such as the sampling frequency and the total
number of data points are passed to the driver. Once the sensor driver is ready, sensing
starts. The sensing task continues running until the predetermined amount of data is
acquired. During this sensing, the acquired data points are first stored in a buffer. Every
time the buffer is filled, the driver passes the data to the sensing application. This block of
data is supplied with a local time stamp marked when the last data point of the block is
acquired. The clock used for the time stamp runs at 3.25 MHz. If time synchronization is
performed prior to sensing, the offset between the global and local times can be utilized to
convert the local time stamp to a global time stamp where needed. The data and time
stamps passed to a sensing application are copied to arrays, and the buffer is returned to
the driver to be used for the next block of data. The size of the buffer was set to keep 110
data points. When 1100 data points are acquired, for example, the buffer is filled ten
times; everytime the buffer is filled, its contents are passed to the application.

79
 Difficulties encountered in realizing synchronized sensing using Imote2s are
summarized in the following paragraphs.

1. Failure in sensing

Imote2s do not always succeed in sensing. When a node fails to start sensing, no
acceleration record is stored on this node. Processing data without recognizing such
failures results in large errors in the final outcome. Failures in sensing need to be detected
so that data from the associated nodes is not processed together with the other data.

2. Uncertainty in start-up time

Starting up sensing tasks at all of the Imote2 nodes simultaneously is challenging.

Even when the commands to start sensing are queued at exactly the same time, the
execution timing of the commands is different on each node. Thus, measured signals are
not necessarily synchronized to each other.
TinyOS has only two types of threads of execution: tasks and hardware event
handlers, leaving users little control to assign priority to commands; if sensing is queued
as a task, this task is executed after all the tasks in the front of the queue are completed. As
such, waiting time cannot be predicted. If the command is invoked in a hardware interrupt
as an asynchronous command, this command is executed immediately unless other
hardware interrupts interfere. However, invocation of commands as a hardware interrupt
from a clock firing at very high frequency is not practical; firing the timer at a frequency
corresponding to the synchronization accuracy, tens of microseconds, is impossible.
In addition, the warm-up time for sensing devices after the invocation of the start
command is not constant. Even if the commands are invoked at the same time, sensing
will not start simultaneously.

3. Difference in sampling rate among sensor nodes

The sampling frequency of the accelerometer on the available Imote2 sensor boards
has nonnegligible random error. According to the data sheet of the accelerometer, the
sampling frequency may differ from the nominal value by at most 10 percent
(STMicroelectronics, 2007). Such variation was observed when 14 Imote2 sensor boards

80
 4
set value: 560 Hz

C ount
2

0
520 530 540 550 560 570 580 590 600 610 620
Sampling frequency (Hz )
Figure 5.22. Sampling frequencies of 14 sensor boards.
(Frequency calibration was conducted by Jennifer A. Rice.)

were calibrated on a shake table (see Figure 5.22). Differences in the sampling frequencies
among the sensor nodes result in inaccurate estimation of structural properties unless
appropriate post processing is performed. If signals from sensors with nonuniform
sampling frequency are used for modal analysis, one physical mode may be identified as
several modes spread around the true natural frequency (Nagayama et al., 2006a).

4. Fluctuation in sampling frequency over time

A nonconstant sampling rate was observed with the Imote2 sensor boards, which if
not addressed, results in a seriously degraded acceleration measurement. The Imote2
receives the digital acceleration signal once a block of data is available from the
accelerometer. The block size is set at 110 data points. The Imote2 puts a time stamp on
the data once a block is available. By comparing differences between two consecutive
time stamps, the sampling frequency of the accelerometer is estimated. The difference
between the time stamps, shown in Figure 5.23, provides an indication of the variation in

81
 5
x10
1. 8715
meas urement 1
meas urement 2
1. 871

T ime (Psec)
1. 8705

1. 87

1. 8695

1. 869
0 50 100 150 200 250
# of data blocks
Figure 5.23. Variation in the sampling frequency over time.

the sampling frequency. The difference in two consecutive timestamps fluctuates by about
0.1 percent. Though imperfect time stamping on Imote2 is a possible source of the
apparent nonconstant sampling rate, fluctuation with a nonzero average values suggests
that the variable sampling frequency as a credible cause of the phenomenon. With this
fluctuation, measurement signals may suffer from a large synchronization error.

5.3.4 Realization of synchronized sensing

The observed problems discussed previously are addressed using the smart sensor’s
computational capabilities. Failure during sensing is detected and sensing is repeated until
success is achieved. The other three issues are dealt with concurrently by resampling the
measured time histories based on time stamps at the end of each block of data. These two
approaches realize synchronized sensing and are discussed in this section.

1. Detection of sensing failure and repetition of sensing

To detect sensing failure, a timer on the Imote2 is utilized. This timer is set when
sensing is initiated; the timer is scheduled to be fired after the planned total measurement
time has passed. If sensing is successfully completed, this timer is stopped at the end of
sensing. Therefore, the timer never fires unless sensing fails during measurement. If the
timer fires, a flag is used to mark sensing failure. After a predetermined time has passed,
the manager sensor inquires of the other nodes and itself whether sensing has failed. If a
flag indicating sensing failure is found, the memory space storing acquired data is
reinitialized, and a task to restart sensing is posted at all of the nodes.

82
 L M

Figure 5.24. The basic idea of resampling.

2. Resampling-based approach

Resampling based on the global time stamps addresses three of the problems
previously cited: (a) uncertainty in start-up time, (b) difference in sampling rate among
sensor nodes, and (c) time fluctuation in the sampling frequency. The basics of resampling
and polyphase implementation of resampling are first reviewed. The resampling approach
is then modified to achieve a sampling rate conversion by a noninteger factor. Finally, this
proposed resampling method is combined with time stamps of measured data to address
concurrently the three issues.

Resampling

Resampling by a rational factor is performed by a combination of interpolation,

filtering, and decimation. Consider the case in which the ratio of the sampling rate before
and after resampling is expressed as an rational factor, L e M . The signal is first upsampled
by a factor of L. Then the signal is downsampled by a factor of M. Before downsampling,
a lowpass filter is applied to eliminate aliasing and unnecessary high-frequency and
aliasing components (see Figure 5.24).
During upsampling, the original signal x[n] is interpolated by L -1 zeroes as in Eq.
(5.13),

­ x > n/L @, n =  r L r L,...

y> n@ = ® (5.13)
¯ , otherwise

where y > n @ is the upsampled signal. In the frequency domain, insertion of zeroes creates
scaled mirror images of the original signal in the frequency range between the new and old
Nyquist frequencies. A discrete-time lowpass filter with a cutoff frequency, S e L , is
applied so that all of these images except for the one corresponding to the original signal
are eliminated. To scale properly, the gain of the filter is set to be L .
The signal is then downsampled by a factor of M as in Eq. (5.14).
z > n @ = y > nM @ (5.14)

83
Before this decimation is applied, all of the frequency components above the new Nyquist
frequency need to be eliminated. A discrete-time, lowpass filter with a cutoff frequency
S e M and gain 1 is applied.
The lowpass filter applied in the downsampling can be combined with the one in the
upsampling process. The cutoff frequency of the filter is set to be the smaller value of
S e L and S e M . The gain is L . This filtering process is reviewed more in more detail in
the subsequent paragraphs.
Numerical filters of various types can be represented as in the following equation,
Nb Na

y>n@ = ¦ b > k @x > n-k @ – ¦ a > k @y > n-k @ (5.15)

k=0 k=1

where x > n @ is the input to the filter, y > n @ is the output from the filter, a > k @ and b > k @ are
filter coefficients for outputs and inputs, respectively. N a and N b represent the numbers
of filter coefficients, a > k @ and b > k @ . These filters can be classified as either Finite
Impulse Response (FIR) filters or Infinite Impulse Response (IIR) filters. An FIR filter has
nonzero coefficients corresponding only to the inputs. In other words, N a is zero for an
FIR filter. An IIR filter has one or more filter coefficients corresponding to the outputs.
FIR and IIR filters have their own advantages over the other. FIR filters are always
stable no matter what coefficients are chosen. Another advantage of an FIR filter is its
linear phase characteristic. The delay introduced by an FIR filter is constant in frequency.
On the other hand, IIR filters with a given performance can be designed using fewer
coefficients, and, thus, are usually less computationally expensive.
One of the possible error sources of this resampling process is imperfect filtering. A
perfect filter, which has a unity gain in the passband and a zero in the stopband, needs an
infinite number of filter coefficients. With a finite number of filter coefficients, passband
and stopband ripples cannot be zero. A filter design with 0.1 to 2 percent ripple is
frequently used. A filter needs to be designed considering these filter characteristics.
Figure 5.25 shows signals before and after filtering. A signal analytically defined as a
combination of sinusoidal waves is sampled at three slightly different sampling
frequencies. Two of the signals are then resampled at the sampling frequency of the other
signal. As can be seen from Figure 5.25, after resampling, the three signals are almost
identical. These signals are, however, not exactly the same due to the imperfect filtering.
Though this signal distortion during filtering is preferably suppressed, this resampling
process is not the only cause of such distortion. AA filtering and digital filtering also use
imperfect filters. The filter in the resampling process needs to be designed so that the filter
does not severely degrade signals as compared with other filters.
The resampling process is considered to be extremely challenging if the upsampling
factor, L , is large. This issue is explained herein with examples. When a signal sampled at
100 Hz is resampled at 150 Hz, the rational factor, L e M , is 3/2. The original signal is
upsampled by a factor of 3. A lowpass filter with a cutoff frequency of S e  can be easily
designed with a reasonable number of filter coefficients. Note that the cutoff frequency of

84
 3 3

2 2

1 1

Samples
Samples

0 0

-1 -1

-2 -2

-3 -3
27.9 27.95 28 27.9 27.95 28
Time (sec) Time (sec)
(a) (b)
Figure 5.25. Signals (a) before and (b) after resampling.

the filter is expressed in radian normalized to the sampling frequency of an upsampled

signal. This filter is applied to the upsampled signal, which is three times longer than the
original one, and then downsampled by a factor of 2. When a signal sampled at 101 Hz is
resampled at 150 Hz, on the other hand, the rational factor, L e M , is 150/101. Upsampling
by a factor of 101 greatly increases the data size. A lowpass filter with a cutoff frequency
of S e  requires a large number of filter coefficients. If such a filter is applied to the
upsampled signal, the upsampled signal is 101 times longer than the original one. Direct
implementation of such resampling on resource-limited smart sensors is intractable.

Polyphase implementation

An FIR filter can be computationally much more inexpensive as a filter for

resampling if the polyphase implementation is employed. The polyphase implementation
leverages knowledge that upsampling involves inserting many zeroes and that an FIR
filter does not need to calculate the output at all of the upsampled data points. This
implementation of resampling is explained in this section mainly in the time domain
because the extension of the method to achieve synchronized sensing is pursued using a
time domain analysis. Oppenheim et al. (1999) provided a detailed description of the
polyphase implementation in the Z-domain.
Upsampling and lowpass filtering with an FIR filter can be written in the following
manner:
N-1

y> n@ = ¦ h > k @v > n-k @

k=0 (5.16)
­ x > m @, l = Lm, m =0, r , r , }
v> l@ = ®
¯ , otherwise

85
where x > n @ is the original signal, v > l @ is the upsampled signal, and y > n @ is the filtered
signal. h > k @ is a vector of the filter coefficients for the lowpass FIR filter. The length of
the vector of filter coefficients is N . L is, again, an interpolation factor. By combining the
two relationships in Eq (5.16) and the following,
Lm = n-k (5.17)

upsampling and filtering can be written as

n/L

y> n@ = ¦ h > n-Lm @x > m @ . (5.18)

m= n-N+ 1 e L

where and represent the ceiling and floor functions, respectively. The number of
algebraic operations is reduced in this equation using the knowledge that v > l @ is zero at
many points.
Downsampling by a factor of M is formulated as
z > j @ = y > jM @
jM/L
(5.19)
= ¦ h > jM-Lm @x > m @ j=0, r  r  }
m= jM-N+ 1 e L

where z > j @ is the downsampled signal. The outputs do not need to be calculated at all of
the data points of the upsampled signal, as can be seen in Eq. (5.19). The output needs to
be calculated only at every M-th data point of the upsampled signal. If an IIR filter was
employed, the filter would need outputs at all of the data points of the upsampled signal.
This polyphase implementation, thus, reduces the number of numerical operations
involved in resampling. However, implementation is still challenging if the upsampling
factor, L , is large.

Resampling with a noninteger downsampling factor

FIR filter design becomes extremely challenging when sampling frequencies need to
be precisely converted. For example, resampling of a signal from 100.01 to 150 Hz
requires a lowpass filter with a cutoff frequency of S e  . The filter needs tens of
thousands of filter coefficients. Design of such a filter is computationally challenging.
Also a large number of filter coefficients may not fit in the available memory on smart
sensors. This problem is addressed mainly by introducing linear interpolation.
First, the resampling is generalized by introducing an initial delay, l i . Eq. (5.19) is
rewritten with l i as follows:

86
 z > j @ = y > jM + l i @
 jM+l i /L
(5.20)
= ¦ h > jM+l i -Lm @x > m @ j=0, r  r  }
m=  jM + l i – N +   e L

By introducing l i , the beginning point of the downsampled signal can be finely adjusted.
If the time difference at the start of sensing is taken into account by l i , the synchronization
accuracy of signals is not limited by the original sampling period.
Resampling is then combined with linear interpolation to achieve the necessary
accuracy. The integer, M is replaced by a real number, M r . The upsampling rate, L a ,
must remain an integer. M r and L a are not uniquely determined. These values are chosen
so that L a is not too large. A large value of L a requires a high-order lowpass filter, as is
the case for the normal polyphase implementation of resampling. Using these upsampling
and downsampling factors, resampling is performed. Upsampling is same as before.
However, the downsampling process shown in Eq. (5.20) cannot be directly applied,
because of the noninteger downsampling factor. Output data points to be calculated do not
necessarily correspond to points on the upsampled signal. Output data points often fall
between upsampled data points. Linear interpolation is used to calculate output values as
follows:
z > j @ = y > pl @ ˜  pu – pj  + y > pu @ ˜  pj – pl 
pl e La

= ¦ h > p l – L a m @x > m @ ˜  p u – p j 
m=  p l – N +   e L a
pu e La

h > p u – L a m @x > m @ ˜  p j – p l  (5.21)

+ ¦
m=  p u – N +   e L a
p j = jM r + l i
pl = jM r + l i
pu = pl + 

In this way, resampling of an arbitrary noninteger rational factor can be achieved. When
L a is not too small, the approximation of linear interpolation gives reasonable results. A
value of L a ranging from 20 to 150 is employed in algorithmic testing and shown to give
reasonable results.

Implementation on Imote2s

The proposed resampling approach is employed to address issues toward

synchronized sensing. This approach is first overviewed.

87
 Timer firing every T3 110 data points from sensing driver

T3 T3 T3 T3 t1 t2 t3 t4 t5 t6 t7 t8 tn
110 110 110 110 110 110 110 110 … 110

Data copied to an array

Start a Start storing Check for
sensing task data sensing
failure

t1 t1+T1 t1+T2 t1+T4

Time
Figure 5.26. Time stamps and waiting times used for synchronized sensing.

The time stamp to indicate the beginning of the entire sensing process, t1 , and the
predetermined waiting times, T1 , T2 , and T3 , are utilized to implement this resampling
approach as well as the time stamp at the end of each block of data, t i  i=1,2, ..., n  . T1 is
the waiting time before the sensing task is posted. A timer component that fires every T3
checks whether the waiting time T1 has passed. T2 is the waiting time before the Imote2
starts storing the measured data. The time stamps, t i , are utilized later to compensate for
misalignment of the starting time, as well as for individual differences in sampling
frequency and the time varying sampling frequency. Figure 5.26 illustrates the use of
these time stamps and waiting times. This figure also shows the waiting time T4 for
sensing failure detection. If sensing has not finished at t1+T4 , the sensing is restarted as
previously stated. The details of this approach are explained in the subsequent paragraphs.
When smart sensors are ready to begin sensing, the manager sensor gets a global time
stamp, t1, and multicasts it to the other nodes. The predetermined waiting times, T1 and
T2, are also multicast. All of the nodes then start a timer component, which is set to be
fired every T3 . When the timer is fired, the global time is checked. If the global time is
larger than t1+T1, a task to start sensing is posted and the periodic timer stops firing.
When a block of sensed data is available, an event is signaled. In this event, the global
time is checked again. If the global time is greater than t1+T2, then the block of data is
stored in memory. Otherwise, the block of data is discarded. From the time stamp of the
last data point of the current block, t current , that of the last block, t last , and t1+T2 , the
initial misalignment is estimated. When the number of data points in a block is N data , the
misalignment in time of the first data point, t ini , is estimated as follows:
t current – t last
t ini = t1+T2 – t last – --------------------------- (5.22)
N data

where t ini is later used in resampling as l i in Eq. (5.21) for the first block of data to
compensate for misalignment of the starting time (see Figure 5.27).
Timestamps and resampling also compensate for difference and fluctuation in the
sampling frequency (see Figure 5.28). The sampling frequency of the current data block,
fs current ,is estimated from t current and t last .

88
 The 1st block

Original signal

Resampled signal
tlast t1+T2 tcurrent
t1+T2 - tlast
tcurrent  tlast
N
li

Figure 5.27. Resampling of the first block of data.

Resampling for
the i-th blcok

xprev[n] xcurrent[n] xnext[n]

x’[n] Original signal

tcurrent Resampled signal

t
tlast, resamplelast

 tcurrent  tlast  N  1
tlast-tlast, resample
N
1
fs
li

Figure 5.28. Resampling of the i-th block of data (i>1).

t current – t last
fs current = --------------------------- (5.23)
N data

The sampling rate conversion by a factor L a e M r is applied to the block of data. The factor
is determined by
L a = fs e d s
(5.24)
M r = fs current e d s

where d s is the common denominator introduced to have the integer, L a , in an

appropriate range. fs is the sampling rate after the rate conversion. l i in Eq. (5.21) for the
first block of data is t ini . For the subsequent blocks, l i is determined by the time of the last
resampled signal of the previous block, t last,resample . Because Eq. (5.21) requires x > m @ in

89
blocks before or after the current block, data in these blocks are also utilized. To be more
specific, when resampling is applied to the current block of data, data from the block
before and after the current block are used as part of the input signal for the resampling,
x' > n @ , i.e.,

­ x prev > n @  d n<N data

°
x' > n @ = ® x current > n-N data @ N data d n  N data (5.25)
°
¯ x next > n-2N data @ N data d n  N data

where x prev , x current , and x next are data in the previous, current, and next blocks,
respectively. The sampling frequency of x' > n @ is assumed to be same as that of
x current > n @ . l i for x' > n @ is then calculated as

l i =  t current – t last  ˜  N data –   e N data –  t last – t last, resample  + ---- (5.26)
fs

Using these parameters, Eq. (5.21) is applied to each data block.

This resampling-based approach, however, cannot be applied on-the-fly by the
Imote2s. The resampling is applied after all of the Imote2s have acquired signals.
This approach can also address differences in sampling frequency among nodes and
fluctuation of the frequency over time. Due to the denominator, d s in Eq. (5.24), the
upsampling factor is kept moderate, eliminating the need for a lowpass filter with an
extremely large number of filter coefficients.
While this algorithm can be implemented on Imote2s to achieve synchronized
sensing, numerical operations are nontrivial. Eq. (5.21) still needs a large number of
multiplications, additions, etc.; the number of coefficients is still large, frequently greater
than a thousand. One thousand filter coefficients, for example, occupy 8 kB of memory
space. These issues are addressed by employing integer operations when applicable.
As is apparent from Eq. (5.15), scaling FIR filter coefficients results in filter outputs
scaled by the same factor. FIR filter coefficients are multiplied by a large constant, K , so
that these coefficients can be well approximated by 16-bit integers.
h' > n @ = h > n @ ˜ K +  (5.27)

These 16-bit integers representing h' > n @ are stored on the Imote2s instead of 64-bit double
precision data, saving considerable memory. The scaled output, y > p l @ and y > p u @ in Eq.
(5.28) are estimated only by integer operations; h' > n @ and x > n @ are both integer type
variables.

90
 z > j @ =  y > pl @ ˜  pu – pj  + y > pu @ ˜  pj – pl   e K
pl e La
§
= ¨ ¦ h' > p l – L a m @x > m @ ˜  p u – p j 
¨
© m=  pl – N +   e La

pu e La
·
(5.28)
+ ¦ h' > p u – L a m @x > m @ ˜  p j – p l  ¸ e K
¸
m=  p u – N +   e L a ¹
p j = jM r + l i
pl = jM r + l i
pu = pl + 

Then linear interpolation is performed by casting all associated integers to double

precision data. The final outcome is adjusted to account for the scaling factor of the filter
coefficients. In this way, implementation of the proposed resampling approach becomes
less numerically challenging.

5.4 Summary
Middleware services for smart sensors were developed in this chapter. Model-based
data aggregation, including distribution and coordination, provided scalability to a large
number of smart sensors without sacrificing performance of the SHM algorithms. The
data loss problem, which was shown to have adverse effects on an SHM algorithm, was
addressed by developing reliable communication protocols. To realize synchronized
sensing, a resampling-based approach was proposed. These middleware services are
implemented on the Imote2 running TinyOS. In Chapter 7, these middleware services will
be used in realization of an SHM system.

91
 Chapter 6

ALGORITHMS

In this chapter, algorithms for SHM to be implemented on smart sensors are

discussed. The Distributed Computing Strategy (DCS) for SHM proposed by Gao (2005)
has the potential to realize densely deployed networks of smart sensors for SHM, because
of its local data sharing and processing. The algorithmic components of DCS for SHM,
i.e., Natural Excitation Technique (NExT), Eigensystem Realization Algorithm (ERA),
the Damage Locating Vector (DLV) method, and DCS, are briefly reviewed. Though most
of the data processing is performed locally, the initialization phase of the strategy to
estimate mode shape normalization constants involves more cumbersome processes; the
initialization requires either input force measurement or output measurement before and
after a known mass perturbation. Recent algorithmic developments of a stochastic DLV
(SDLV) method by Bernal (2006) allows estimation of DLVs without input force
measurement or mass perturbation. DCS is extended with this stochastic DLV method to
allow for less demanding initialization.

6.1 Natural Excitation Technique

To understand the NExT (James et al., 1992, 1993) considered the equation of motion
in Eq. (6.1) under the assumption of the stationary random process.
Mx··  t  + Cx·  t  + Kx  t  = f  t  (6.1)

where M , C , and K are the n x n mass, damping, and stiffness matrices, respectively;
x  t  is a n x 1 displacement vector; f  t  is a m x 1 force vector; x·  t  and x··  t  are the
velocity and the acceleration vectors, respectively. By multiplying the displacement at the
reference sensor and taking the expected value, Eq. (6.1) is transformed as follows:
ME > x··  t+W x ref  t  @ + CE > x·  t+W x ref  t  @ + KE > x  t+W x ref  t+W  @ (6.2)
= E > f  t+W x ref  t  @

Because the input force and response at the reference sensor location are uncorrelated for
W t  , the right-hand side of Eq. (6.2) is zero. The expectation between the two signals is
the correlation function. Therefore, by denoting E > x  t + W y  t  @ as the correlation
function R xy  W  , Eq. (6.2) is rewritten as
MR x·· xref  W  + CR x· x ref  W  + KR xx ref  W  =  W t  (6.3)

When A  t  and B  t  are weakly stationary processes, the following relation holds
(Bendat & Piersol, 2000).

92
 R ·  W  = R· AB  W  (6.4)
AB

A similar relation holds for higher derivatives,

m
R m W = R AB  W  (6.5)
A B

where the superscript, m, denotes the m-th derivative. Consequently, Eq.(6.3) can be
rewritten as

MR·· xxref  W  + CR· xx ref  W  + KR xx ref  W  =  W t  (6.6)

Thus, correlation functions for the stationary responses are shown to satisfy the equation
of motion for free vibration. This fact can be directly used for the subsequent modal
analysis.

6.2 Eigensystem Realization Algorithm

ERA (Juang & Pappa, 1985) identifies modal parameters from free vibration
responses. When measurement at p sensors are available in a measurement vector y > n @ ,
the Markov parameters Y > n @ from m sets of measurement are constructed as follows:

Y > n @ = y >  @ y >  @  }  y > m @ (6.7)

A generalized Hankel matrix, H rs > k –  @ , is formed as a r u s block matrix:

Y>k@ Y > k+j 1 @ } Y > k+j s-1 @

Y > h 1 + k @ Y > h 1 + k+j 1 @ } Y > h 1 + k+j s-1 @
H rs > k –  @ = (6.8)
# # % #
Y > h r-1 + k @ Y > h r-1 + k+j 1 @ } Y > h r-1 + k+j s-1 @

A SVD of this Hankel matrix yields

T
H rs >  @ = PDQ (6.9)

where the superscript T denotes the matrix transpose. Components in these matrices
corresponding to small singular values are considered noise and replaced by zeroes. Juang
– 1/2 T – 1/2 1/2 T T 1/2
and Pappa (1985) derived that the triple, ^ D P H rs >  @QD  D Q E m E p PD `
is a minimum realization of the measured system. A matrix to extract the first m columns,
E m is defined using an identity matrix of order m and a null matrix of size m.
T
E m = > I m  m  m }  m @ (6.10)

93
 The eigenproblem of the system matrix is then solved:
– 1/2 T – 1/2
D P H rs > k @QD \ i = z i \ i (6.11)

The natural frequencies f i and damping ratios ] i are then calculated from the eigenvalues,
z i as
O i = NP z i e  k't 
f i = O i e  S  (6.12)
] i = Re(O i) e O i

where NP z i is the principal value of natural logarithm of z i . 't is the sampling period.
Re  takes the real part of a complex number. The mode shape, I i , corresponding to z i
is calculated as
T 1/2
I i = E p PD \i . (6.13)

The initial modal amplitudes, 4 , which can be utilized to calculate modal amplitude
coherence, an indicator to quantitatively distinguish the system and noise modes, is
estimated as follows:
-1 1/2 T
4 = \ D Q Em (6.14)

6.3 Damage Locating Vector method

One of the flexibility-based SHM approaches is the Damage Locating Vector (DLV)
method, which is briefly described here. The DLV method, first developed by Bernal
(2002), is based on the determination of a special set of vectors, the so-called damage
locating vectors (DLVs). These DLVs have the property that when they are applied to the
structure as static forces at the sensor locations, no stress is induced in the damaged
elements. This unique characteristic can be employed to localize structural damage (Gao,
2005).
For a linear structure, the flexibility matrix, F , at the sensor locations is constructed
from measured data. Gao (2005) explained two approaches to estimate the flexibility
matrix. Both of them reconstruct the flexibility matrix using mode shapes and
normalization constants. One approach estimates the normalization constants assuming
that the input force can be measured, while the other utilizes output-only measurements
before and after a known mass perturbation.
Formulation of the flexibility matrix based on input force measurements is
summarized by the following equation involving complex conjugate pairs of arbitrary
normalized mode shapes, < = M M .

94
 T
F = – <D g < (6.15)

D g = diag § d 1 d 2 } d j } · is the matrix of modal normalization constant, d j . If all of

© ¹
the modes are observed and available for reconstruction, F can be estimated. When only a
part of structural modes is available, Eq. (6.19) gives an approximation of F. Gao (2005)
showed that F can be approximated well from a limited number of modes.
The normalization constants can be estimated through the following relation:
– p +   T T –
dj = Oj I j Bq f > diag  < m,j q m  @ (6.16)

where mode shapes at sensor locations < m and eigenvectors I satisfies the following
relationship:
< m = CI (6.17)

AI = I/ (6.18)
T – T
I j is the j-th row of the inverse of eigenvector matrix I , and < m,j is the j-th row of the
T
transposed mode shape matrix < m . / is a diagonal matrix of eigenvalues. When there is
more than one colocated sensor and actuator pair, multiple estimates of d j will be
obtained. Bernal and Gunes (2004) suggested that d j corresponding to the component in
T
vector < m,j q m with the largest magnitude might be used. Using the normalization
constants, the flexibility matrix is estimated as in Eq. (6.15); only a part of the flexibility
matrix corresponding to sensor locations is practically estimated by replacing the mode
shape matrix < with the matrix of mode shapes at sensor location, < m .
When the measurement of input force is impractical, output-only measurements
before and after a known mass perturbation are utilized to estimate normalization
constants for flexibility matrix reconstruction. The estimated mode shapes and
eigenvalues, as well as normalization constants, reconstruct F as follows:
– T
F = )/ )
(6.19)
– T
=  ID /  ID 

where ) is the matrix of mass normalized mode shapes, and I is the matrix of arbitrarily
normalized mode shapes. D is a diagonal matrix of mass normalization constants, D i .
Note that estimation of normalization constants from a known mass perturbation is
theoretically derived for a proportionally damped system; mode shapes should be real. An
approximate flexibility matrix is estimated from a limited number of modes.
In the mass perturbation approach (Bernal, 2004), the mode shapes and eigenvalues
are determined before and after a known mass perturbation is introduced to the structure.
The mass matrix of the modified structure, M 1 , is expressed as
M 1 = M 0 + 'M (6.20)

95
where M 0 is the mass matrix of the original structure, and 'M represents the mass
perturbation. D i is estimated from modal parameters and 'M as follows:

2 O  i – O  i q ii
D i = ------------------------ ˜ -------------------------
- (6.21)
O  i T
I 'MI
T
0,i ,i

T T – T
q i = q 1i q 2i } q ni =  I 0 I 0  I 0 I 1,i (6.22)

The subscripts 0 and 1 indicate quantities before and after the mass perturbation,
respectively, and n is the number of observed modes. While Eq. (6.22) uses mode shapes
at all DOFs, the usage of mode shapes at sensor location is shown to give reasonable
results (Gao, 2005). I in Eq. (6.19) is in practice replaced with mode shapes at sensor
locations, yielding a corresponding flexibility matrix.
Estimation of the flexibility matrix using either input force measurement or known
mass perturbation is performed before and after damage to localize the damage. Estimated
flexibility matrices of the undamaged and damaged structures are denoted as F u and F d ,
respectively. All of the linearly independent load vectors L are collected, which satisfy the
following relationship:
F d L = F u L or F ' L =  F d – F u L =  (6.23)

F ' is change in the flexibility matrix. This equation implies that the load vectors L
produce the same displacements at the sensor locations before and after damage. From the
definition of the DLVs, these vectors also satisfy Eq. (6.23); that is, because the DLVs
induce no stress in the damaged elements, damage in those elements does not affect the
displacements at the sensor locations. Therefore, DLVs are indeed the vectors in L . To
calculate L , SVD is employed. The SVD of the matrix F ' leads to

T
T 6  V
F ' = USV = U  U  (6.24)
  VT


Recall from the characteristics of SVD that

T
V
V V = I (6.25)
T
V

where I is identity matrix. Eq.(6.24) can then be rewritten as

F' V F' V = U S  (6.26)

96
From Eq. (6.26), one obtains
F' V =  (6.27)

Eqs. (6.23) and (6.27) indicate that L = V  , i.e., DLVs can be obtained from the
SVD of the difference matrix F ' . In Eq. (6.24), because of noise and computational
errors, the singular values corresponding to V  are not exactly zero. To select DLVs from
SVD results of the matrix F ' , Bernal (2002) proposed an index svn defined as

si ci
svn i = --------------------------
- (6.28)

max  s k c k 
k

where s i is the i-th singular value of the matrix F ' ; c i is the constant that is used to
normalize the maximum stress in the structural element, which is induced by the state load
c i V i , to have a value of one; and V i is the right singular vector of F ' .
Each of the DLVs is then applied to an undamaged analytical model of the structure,
and the stress in each structural element is calculated. If an element has zero normalized
accumulative stress, this element is a candidate damaged element. The normalized
accumulative stress for the j-th element is defined as
Vj
V j = ----------------------- (6.29)
max  V k 
k

where

§ n V ij ·
V j = abs ¨ ¦ -----------------------
-¸ (6.30)
¨ i ¸
© i= max  V k ¹
k

In Eq. (6.30), V ij is the stress in the j-th element induced by the i-th DLV; V j is the
cumulative stress in the j-th element. In practice, the normalized accumulative stress
induced by DLVs in the damaged elements may not be exactly zero due to noise and
uncertainties. A reasonable threshold should be chosen to select damaged elements.

6.4 Distributed Computing Strategy for SHM

Gao (2005) proposed the Distributed Computing Strategy (DCS) for SHM employing
the algorithms reviewed in previous sections of this chapter. In this approach, a hierarchy
of local sensor communities is organized, so that the data does not need to be centrally
collected or analyzed. Because these algorithms are applicable to data only from nodes in
a local sensor community, the analysis can be performed within a local sensor community
independent of the other nodes; data from a given community is processed within the

97
same community, eliminating the need to globally transfer a large amount of data through
a long communication path.
While most of the derivation in NExT, ERA, and the DLV method can be directly
applicable to data from a local sensor community, the mass normalization constants, d j or
D i , for the DLV method may need data at DOFs outside the community. To estimate d j in
a local sensor community, the input force and the output response at that location need to
be available. D i needs mode shapes to be estimated for DOFs corresponding to the
location of the mass perturbation, as shown in Eq. (6.21). In addition, q ii in Eq. (6.22)
may contain a large error if only a limited number of sensors in a local area is used. Gao
(2005) suggested that estimation of mass normalization constants be performed from
globally obtained data during initialization of the SHM system. Normalization constants
obtained during initialization are then converted to normalization constants for each
sensor community through the following equations:

i § < j ref  i · 
dj = d j ˜ ¨ ------------------
-¸ (6.31)
© < ij ref ¹

i I j ref  i 
D j = D j ˜ ----------------
i
- (6.32)
I j ref

i i
d j and D j are the j -th mode mass normalization constants for the sensor community, i .
< j ref  i  and I j ref  i  are the j -th global mode shape’s components corresponding to the
i i
reference sensor node of the i-th sensor community. < j ref and I j ref are components of j-
th mode shape determined in the i-th sensor community. These components correspond to
the reference sensor DOF of the community. In this way, mass normalization constants for
each sensor community can be estimated. Thus, monitoring is realized within local sensor
communities.
DCS also describes how local sensor communities collaborate with each other to
judge structural damage, including redundancy to increase fault tolerance. Local sensor
communities are formed with overlap, so that more than one sensor community monitors
the same element. This overlap provides the means to deal with false-positive and false-
negative damage detection. Information about localized damaged elements is exchanged
among local sensor communities. If damage localization in one sensor community
contradicts that in another community, measurement and analysis are repeated. If a
damaged element is detected and no inconsistency is reported, the base station is informed
of the damage (see Figure 6.1).

6.5 Stochastic Damage Locating Vector method

Recent algorithmic development of the Stochastic DLV (SDLV) method by Bernal
(2006) has the potential to improve the DCS for SHM by eliminating the need to estimate
normalization constants. The initialization phase of DCS involving either input force

98
 measured data (outputs only)

data collection

auto and cross-correlation functions

NExT & ERA

modal parameters
i i i
(Y dor l ,dy ) d

i i
Y dor y normalized
d to
have a unit magnitude

data aggregration undamaged normalization

i
for community i constants Dg or a i(eq. 5.1)

potentially damaged
flexibility matrix Fd

FD = Fd - Fu

singular value decomposition of FD

damage locating vectors (DLVs)

obtain normalized cumulative

stress under DLVs
no send an "ok" signal
damage in community i?
back to central station
yes

no community i reports
part of adjacent communities? damage information
yes
decision making
identified as damaged candidates yes report damage
involving adjacent information
in all of these communities?
communities
no
retake data, and reconduct data
aggregation and decision making

Figure 6.1. DCS logic (Gao, 2005).

99
measurement or mass perturbation can be greatly simplified with the SDLV method. The
SDLV method is briefly reviewed in this section.
Bernal (2006) stated that the null space of the change in the flexibility matrix will be
T T
contained in the null of 'Q . The matrix 'Q is the transpose of the change in Q which
is defined as follows:
– p +  
Q = – CA H p† L (6.33)

–p
H p = CA (6.34)
–p
CA

L = I (6.35)


where I denotes identity matrix,  denotes a zero matrix. p = 0, 1, or 2, depending on

whether the measured outputs are displacement, velocity, or acceleration. The system
matrix A and observation matrix C are determined through modal analysis such as ERA.
T
The null vectors of 'Q are treated as DLVs.
Bernal (2006) also proposed a way to determine the number of DLVs and to combine
information from multiple DLVs. When there are q DLVs and the stress corresponding to
each vector is V j , a normalized stress index (nsi) is defined as follows:
Vj
nsi j = --------------- (6.36)
V j max

Weighting can be incorporated to yield a weighted stress index (WSI) as

q

WSI = ¦ Jj nsij (6.37)

j=1

where J j is a weighting parameter. Potentially damaged elements ( PD ) are chosen as

PD = ^ elements|WSI d tol ` (6.38)

Bernal (2006) took J j =  and tol = 0.1 ˜  WSI  max . The number of DLVs, q , can be
estimated as

­ si ½
q =  ˜ ® # of J values d  J = ------------
-¾ (6.39)
¯ sj ¿
max

100
where s j are the singular values of 'Q . Bernal (2006) demonstrated that the proposed
WSI successfully indicates damage in a numerical truss structure.

6.6 Extension of DCS for SHM with the SDLV method

The SDLV method simplifies DCS for SHM. The SDLV method does not require
normalization constant estimation, eliminating the need for initialization with input force
measurement or mass perturbation. The derivation does not assume response
measurement at locations distributed over the entire structures. Structural responses are
measured only within local sensor communities before and after damage, and the SDLV
method is applied subsequently.
There are several ways to construct the system matrix A and the observation matrix
– 1/2 T – 1/2 1/2 T T 1/2
C . The triple ^ D P H rs >  @QD  D Q E m E p PD ` , determined in ERA is a
minimum realization of the system; the first and third terms can be used as A and C .
However, in practice, the triple may contain noise modes. Another way to formulate A
and C matrices is to use modal parameters, i.e.,

A = diag § O  O  } O n O  O  } O n · (6.40)
© ¹

C = < 1 < 2 } < n < 1 < 2 } < n (6.41)

By selecting modes to be used in Eqs. (6.40) and (6.41), only specified modes contribute
to the matrices. C can also be formulated as

C = <1 <2 } <n   }  (6.42)

assuming the observed variables are the real part of the outputs. Following this
T
formulation, 'Q becomes a complex matrix yielding complex DLVs and stress field.
The absolute value of the complex stress induced by complex DLVs is used as V j .
Numerical simulation showed the C matrix formulation in Eq. (6.42) gives more accurate
damage detection results than that in Eq. (6.41). One explanation is that the imaginary
parts of the C matrix in Eq. (6.41) are cancelled out in Eq. (6.33), losing certain
information; the C matrix from the other formation maintains the information from the
imaginary part. In the subsequent chapters, the formulation in Eq. (6.42) is utilized.

6.7 Summary
Algorithms to be implemented on smart sensors are briefly reviewed in this section.
DCS for SHM is proposed as an algorithm enabling distributed SHM employing densely
deployed smart sensors. DCS is extended with the SDLV method to eliminate the need for
input force measurement or mass perturbation during the initialization. In the next chapter,

101
these algorithms are implemented on the Imote2 platform using the middleware services
developed in the previous chapter.

102
 Chapter 7

REALIZATION OF DCS FOR SHM

The Distributed Computing Strategy (DCS) for SHM reviewed in Chapter 6 is

developed for the Imote2 sensor platform in this chapter. First, a number of requisite
functions, such as FFT and SVD, are developed and their accuracy limitations are
evaluated. DCS is then implemented and validated on a component by component basis.

7.1 Functions
DCS involves numerical operations that are not provided by a standard math library.
NExT estimates correlation functions using FFT and the inverse FFT. ERA applies SVD
to a matrix of real numbers, as in Eq. (6.9). Eq. (6.11) needs a complex eigensolver. If the
initial modal amplitudes need to be estimated to distinguish system and noise modes, the
inverse of a complex matrix is essential, as shown in Eq. (6.12). Sorting is also needed in
ERA, as modes are usually sorted by their natural frequencies. DLV methods involve
SVD and sorting. Furthermore, the SDLV method may require SVD on a complex matrix,
depending on formulation of an observation matrix, C . These functions need to be
implemented on the Imote2 to realize structural health monitoring applications.
Numerical Recipes in C (Press et al., 1992) and the CLAPACK User’s Guide
(Anderson et al., 1999) explain a variety of numerical functions. Necessary functions are
coded based on these references. Because the code and libraries need to be cross-compiled
for the Xscale processor on the Imote2, instead of the x86 processor on a PC, the codes in
Numerical Recipes in C or in CLAPACK cannot be directly used. Relevant files are
extracted and modified for use with Imote2 applications.

7.1.1 Fast Fourier Transform

Fast Fourier Transform (FFT) is essential for implementation of the NExT

algorithms. NExT uses the auto- and cross-correlation functions for the measured systems
as input. First, time histories are converted to the frequency domain using the FFT and
averaged to produce the power- and cross-spectral densities. The auto- and cross-
correlation functions are then obtained by applying the inverse FFT. The Cooley and
Turkey (1965) algorithm, using the Danielson and Lanczos Lemma (Danielson &
Lanczos, 1942), enables FFT in O  N NQI N  operations. Double-precision FFT available
as a part of Numerical Recipes in C (Press et al., 1992), is modified to be used on the
Imote2.
The numerical accuracy of the FFT implementation on Imote2 is examined by
comparing FFT results from the Imote2 with those calculated by MATLAB on a PC. A
band-limited white noise is generated as a signal to be analyzed, and the FFT is applied to
this signal. As seen in Figure 7.1, the FFT implementation on the Imote2 and on the PC

103
 0
10

Fourier amplitude |F|

-2
10

-4
10
0 500 1000 1500 2000
-12
10
error ratio |F-F0|/|F0|
Fourier amplitude

-14
10

-16
10
0 500 1000 1500 2000
discrete sample point
Figure 7.1. FFT accuracy check (Number of FFT data points: 4,096).

give numerically identical FFT results, considering that double precision data has about 15
effective significant digits.
The FFT on double precision data needs n B memory space to store the data to be
analyzed; the data is then replaced with the FFT result. The Imote2’s 256 kB of on-board
memory can hold about 32,000 double data points. If FFT is applied to a longer record,
larger RAM or virtual memory is necessary. Note that the Imote2 has 32 MB of RAM,
which can hold a large amount of data. At the time of this research, however, only 256 kB
of memory is accessible.

7.1.2 Singular value decomposition

Singular Value Decomposition (SVD) is a part of both the ERA and DLV methods.
While SVD in the ERA and mass perturbation DLV method are applied to real valued
matrices, the SDLV method may involve SVD of a complex matrix, depending on how
the observation matrix, C, is reconstructed through modal analysis. A double precision
SVD function for real matrices, as well as one for complex matrices needs to be available.
The source code for SVD of a real matrix is found in Numerical Recipes in C and
CLAPACK, while CLAPACK also provides source codes for SVD of a complex matrix.
First, the matrix is transformed to a bi-diagonal form using the Householder reduction.
The bidiagonal matrix is then diagonalized to obtain the singular values. The source code
from Numerical Recipes in C and CLAPACK are modified for implementation on the
Imote2.
The accuracy of the implementation is examined. The singular values of the matrix
are chosen as accuracy indicators; singular values of various magnitudes are inspected. An
arbitrary matrix is first constructed and decomposed by SVD on MATLAB. The singular
values are then replaced with a decreasing geometric series; singular values of this matrix

104
 Singular value
0 0

Singular value
10 Imote2 Sv 10 Imote2 Sv
Matlab Sv0 Matlab Sv0

-20 -20
10 10
0 20 40 60 80 100 0 20 40 60 80 100
0 0
10 10

Singular value
Singular value

error ratio
error ratio

|Sv - Sv0|/|Sv0|

|Sv - Sv0|/|Sv0|

-20 -20
10 10
0 20 40 60 80 100 0 20 40 60 80 100
Order of singular values Order of singular values
(a) (b)
Figure 7.2. SVD accuracy check on 100 x 100 matrix (Numerical recipes functions): (a)
smaller singular values; and (b) larger singular values.
Singular value

Singular value
0
10 Imote2 Sv
10 Imote2 Sv
Matlab Sv0 Matlab Sv0

-20 -20
10 10
0 20 40 60 80 100 0 20 40 60 80 100
0
0 10
Singular value

10
Singular value

error ratio
error ratio

|Sv - Sv0|/|Sv0|
|Sv - Sv0|/|Sv0|

-20 -20
10 10
0 20 40 60 80 100 0 20 40 60 80 100
Order of singular values Order of singular values
(a) (b)
Figure 7.3. SVD accuracy check on 100 x 100 matrix (CLAPACK functions): (a) smaller
singular values; and (b) larger singular values.

are consequently distributed equally in log scale, allowing a check of the algorithms for
singular values of various magnitudes. A matrix is reconstructed from the calculated
singular values and the two unitary matrices obtained in the SVD. This SVD and
reconstruction are performed on both the Imote2 and a PC. Figures 7.2 and 7.3 show
singular values estimated on the two platforms and the ratio of the numerical difference
between the two to the singular values on a PC. The accuracy of the SVD on the Imote2
and on a PC is numerically the same. As another accuracy indicator, the 2-norm of the
difference between the original matrix and the matrix reconstructed from the SVD results
– 
are calculated as shown in Table 7.1. These values are as small as  . The SVD
implementation on the Imote2 is considered to be accurate to within the precision of the
data type.
Limitations on the matrix size are also examined. Because smart sensors have limited
memory, numerical operations on large matrices are not feasible. An n u n double

precision matrix occupies n B of memory. Therefore the two unitary matrices of SVD

105
Table 7.1. SVD Accuracy Check on Matrices of Various Size

|USV-A|2/|A|2
matrix size
Numerical Recipes CLAPACK
3x3 1.01 x e-15 1.01 x e-15
30x30 6.39 x e-16 4.96 x e-15
50x50 1.09 x e-15 4.0337 x e-16
100x100 9.72 x e -16 4.1957 x e-15


and the diagonal matrix of singular values require n + n B of memory. Even when
the entire 256 kB of RAM on the Imote2 is used for storing these matrices, the maximum
size of the matrix to be analyzed is  u  .

7.1.3 Eigensolver

ERA requires an eigensolver to estimate modal parameters. The eigensolver is

employed to the estimated system matrix, A. Modal frequencies and damping ratios are
estimated from its eigenvalues, while mode shape estimation requires the eigenvectors.
Therefore, both eigenvalues and eigenvectors need to be calculated. Moreover, the A
matrix is not necessarily symmetric. A complex eigensolver capable of calculating both
eigenvalues and eigenvectors is needed.
CLAPACK contains a double precision complex eigensolver. The algorithm reduces
a matrix to its upper Hessenberg form. The Hessenberg matrix is then reduced to its Schur
form. The eigenvalues are obtained as the diagonal elements of the Schur form matrix.
The eigenvectors of this matrix is then calculated and transformed back considering the
Hessenberg and Schur transforms. The source code is modified for use on the Imote2.
The accuracy of the implementation is examined. A matrix is constructed in the same
way as in the SVD function accuracy test. The eigensolver is applied to this matrix on both
the Imote2 and the PC. Figure 7.4 shows eigenvalues estimated on the two platforms and
the normalized difference between the two eigenvalues. Eigenvalues on the Imote2 and on
the PC are considered to be numerically equal. As another accuracy indicator, the Modal
Assurance Criterion (MAC) for eigenvectors calculated on the Imote2 and on the PC is
investigated. The MAC between two vectors I  and I  are defined as follows:

MAC = I  ˜ I  e  I  ˜ I  I  ˜ I   (7.1)

where * denotes complex conjugate. Identical vectors give a MAC value of 1.0, even if
one of them is multiplied by a constant. Uncorrelated vectors give a MAC value of zero.
The deviation of the MAC values from one is plotted on Figure 7.5. These values are as
– 
small as  when the corresponding eigenvalues are sufficiently large. Although the
eigenvectors corresponding to eigenvalues smaller than the precision of the double data
type are different from those on the PC, the eigensolver implementation on Imote2 is
considered to be accurate with the precision of the data type.

106
 5
0
Imote2 O 10
10

Eigenvalue
Imote2 O

Eigenvalue

magnitude
Matlab O0

magnitude
Matlab O0

-20 -5
10 10
0 10 20 30 40 50 0 10 20 30 40 50
0 -10
10 10

Eigenvalue
Eigenvalue

|OO0|/|O0|

error ratio
error ratio

|OO0|/|O0|

-20 -20
10 10
0 10 20 30 40 50 0 10 20 30 40 50
Order of eigenvalues Order of eigenvalues
(a) (b)
Figure 7.4. Complex eigensolver accuracy check on 50 x 50 matrix (eigenvalues): (a)
smaller eigenvalues; and (b) larger eigenvalues.

-15
x10
1 2

0.8
1
0.6
1-MAC

1-MAC

0
0.4

0.2 -1

0 -2
0 10 20 30 40 50 0 10 20 30 40 50
Order of eigenvalues Order of eigenvalues
(a) (b)
Figure 7.5. Complex eigensolver accuracy check on 50 x 50 matrix (eigenvector): (a)
smaller eigenvalues; and (b) larger eigenvalues.

The matrix size for which an eigenanalysis can be performed is limited by available
memory. The size of a complex double precision matrix is twice as large as that of real
double precision matrix. Double precision eigenvalues and eigenvectors of an n u n

complex matrix occupies n + n B of memory. The eigensolver uses another n u n

matrix internally, resulting in a need for n + n B memory in total; the size of
memory needed to keep other small internal variables are not counted. When n = 89, the
RAM space occupied by these matrices exceeds the size of the 256 kB of on-board RAM
on the Imote2.

107
7.1.4 Complex matrix inverse

A complex matrix inverse function is developed based on the Gauss-Jordan matrix

inverse method described in Numerical Recipes in C (Press et al., 1992). The accuracy of
the algorithm is examined by calculating the matrix product of the inverse and original
matrices. The identity matrix was obtained as the output.
The complex matrix inverse algorithm for an n u n complex matrix needs
 
n + n B of memory. n B is first used to store the original matrix and then
replaced with the matrix inverse. n B is internally used. The necessary memory space,

n + n , exceeds 256 kB when n =  .

7.1.5 Sort

A quick sort algorithm is developed based on Numerical Recipes in C (Press et al.,

1992). One application example requiring two vectors be sorted simultaneously is the
ordering of natural frequencies in ascending order and the corresponding reordering of
multiple modal parameters. This function is implemented using floating point operations.
The memory size required to run this function is approximately n B, where n is the
length of the vector to be sorted; the size of floating type variable is 4 B. Sorting of more
than 30,000 data points can be achieved with 256 kB memory. This function is not
considered to be expensive in terms of memory usage.

7.2 DCS implementation

Components of DCS (i.e. Sensing, NExT, ERA, DLV methods, and DCS logic) are
implemented on the Imote2s. The outputs of each step of the DCS algorithm on Imote2s is
compared with reference values calculated on the PC using MATLAB. Sensing is
examined by comparing signals from the Imote2s with those from conventional reference
accelerometers. To check the validity of the various DCS components, acceleration data is
injected from the PC, instead of acceleration measured on the Imote2, and used for the
numerical evaluation. The Imote2 is shown to perform DCS calculations as designed.

7.2.1 Sensing

The middleware service developed in section 5.3 is implemented on the Imote2, and
the sensing capability investigated. The sensing capability is first calibrated against
reference sensors. In anticipation of the damage detection experiment with Imote2s in the
next chapter, the Imote2s and reference sensors are placed on the three-dimensional truss
structure, and the measured acceleration signals are examined. Synchronization of the
measured signals is considered by sensing acceleration responses of the truss and
comparing the phase characteristics of the signals.

108
Figure 7.6. Three-dimensional, 5.6 m-long truss structure.

            
6 10 14 18 22 26 30 34 38 42 46 50
3 7 11 15 19 23 31 35 39 43 47 51
27
2 5 9 13 17 21 25 29 33 37 41 45 49 53
1 4 8 12 16 20 24 28 32 36 40 44 48 52
              

Figure 7.7. Node and element IDs of the truss.

Figure 7.8. Pin and roller ends (Gao, 2005).

The structure on which the calibration is conducted is the 5.6m long, three-
dimensional truss (see Figures 7.6 and 7.7) located at the Smart Structures Technology
Laboratory (SSTL) of the University of Illinois at Urbana-Champaign (http://
cee.uiuc.edu/sstl/). Originally designed at the Structural Dynamics and Control/
Earthquake Engineering Lab in the University of Notre Dame (Clayton & Spencer, 2001),
the truss has 14 bays, each of which is 0.4 m in length. The truss sits on two rigid supports.
One end of the truss is pinned to the support, and the other is roller-supported (see Figure
7.8). The pinned end can rotate freely with all three translations restricted. The roller end
can move in the longitudinal direction.
The truss members are steel tubes with an inner diameter of 10.9 mm and an outer
diameter of 17.1 mm. The joints of the elements are specially designed so that the truss

109
Figure 7.9. Details of the joint (Gao, 2005).

Figure 7.10. Magnetic shaker (Gao, 2005).

member can be easily removed or replaced to simulate damage without dissembling the
entire structure. A detailed picture of the joint is shown in Figure 7.9. As can be seen, a
truss member can be installed/removed by tightening/loosening the collars at the two ends
of a member.
The truss is excited vertically by a Ling Dynamic Systems permanent magnetic V408
shaker (see Figure 7.10) that can generate a maximum force of 20 lbs. with a dynamic
performance ranging from 5 to 9,000 Hz. A band-limited white noise is sent from the
computer to the shaker to excite the truss structure up to 100 Hz. The shaker is connected
to the bottom of the outer panel using a small steel rod. A PCB piezotronics load cell
(model 208B02) is installed between the steel rod and the bottom of the joint to measure
the input to the structure. This load cell has a sensitivity of 50 mV/lb, a frequency range of
0.001 to 36,000 Hz, and a measurement range of 100 lbs. in both compression and tension.
In this report, input force measurement is not used for damage detection. However, the
load cell is installed to confirm that the input force is approximately a band-limited white
noise.
Two Imote2s are firmly attached on a node of the truss by hot glue to measure
structural responses at this point in three directions (see Figure 7.11). Two Imote2s are
installed next to each other; corresponding signals from the two Imote2s are expected to
be identical if the truss node behaves as a rigid body without rotational motion. Even
though rotational behavior of the truss node is not zero, translational motion is expected to
be of much larger magnitude. The two signals should, therefore, show good agreement.
Differences in the two signals are attributed primarily to observation noise specific to each

110
Figure 7.11. The Imote2s on a truss node.

Figure 7.12. Accelerometer: model 353B33 and magnetic base.

of the Imote2 nodes, while noise common to both Imote2 nodes needs to be detected in
comparison with reference sensors.
PCB high-sensitivity piezotronics accelerometers (see Figure 7.12.) are used as
reference accelerometers. These accelerometers have a sensitivity of approximately 100
mV/g, a frequency range of 1 to 4,000 Hz, and a measurement range of ±50g. These
accelerometers are mounted on the structure through a magnetic base. These magnetic
bases facilitate easy relocation of the accelerometers between tests.
Four PCB accelerometers are attached on the same structural node as the one that the
Imote2s are attached to. Two of them measure acceleration in the longitudinal direction,
while the others measure acceleration in the transverse direction. To investigate vertical
acceleration measurements, the two accelerometers used in longitudinal acceleration
measurement are relocated so that these sensors measure vertical acceleration. Thus, at
least two reference accelerometers measure acceleration in one direction; differences in
signals from these two sensors indicate noise components that are not common to both of
them.
A four-channel 20-42 Siglab spectrum analyzer (Spectral Dynamics, Inc., 2007)
provides the signal to the shaker and measures the inputs or reference sensor signals.

111
Figure 7.13. Amplifier (Gao, 2005).

Eight-pole elliptical AA filters are employed for both the input and output measurements
with a cutoff frequency of 100 Hz. The sampling rate is 256 Hz. Other equipment used
during the testing includes:
• Amplifier: A Sony STR-D315 amplifier (see Figure 7.13) uses the input signal from
the Siglab spectrum analyzer to drive the magnetic shaker.
• Signal conditioner: PCB four-channel signal conditioners (model 441B104) have
been used in this experiment.
Using the synchronized sensing middleware service, the Imote2 is programmed to
capture acceleration responses of the truss. The LIS3L02DQ accelerometer
(STMicroelectronics, 2007) on the Imote2 sensor board applies an AA filter and yields
digital outputs. The cutoff frequency of the AA filter is fixed with respect to the sampling
frequency (see Table 7.2). Because the structural modes of the truss whose natural
frequencies of interest are as high as 100 Hz need to be analyzed, the sampling frequency
is set as 560 Hz. After measurement, signals are resampled at 280 Hz, with the resulting
usable bandwidth of the signal being a little over 100 Hz.

Table 7.2. Cutoff and Sampling Frequencies of LIS3L02DQ Accelerometer

Decimation factor Cutoff frequency (Hz) Sampling frequency (Hz)

128 70 280
64 140 560
32 280 1120
8 1120 4480

The digital lowpass filter employed in the resampling approach is designed using
Matlab (The Mathworks, Inc., 2007). The passband cutoff frequency needs to be higher
than the frequency range of interest, i.e. 100 Hz, while the stopband cutoff frequency is
usually lower than the Nyquist frequency of the resampled signals. The filter design with
these two cutoff frequencies being much lower than the sampling frequency of the
upsampled signal and close to each other requires a large number of filter coefficients. To
alleviate this filter design difficulty, the stopband cutoff frequency is increased. Instead of

112
 1

Correlation Coefficient
0

-1
-4.8 -4. 7 -4. 6 -4.5 -4.4
Time Shift (sec)

Figure 7.14. Correlation coefficients plotted against time shift.

keeping the stopband cutoff frequency lower than the Nyquist frequency and completely
eliminating the aliasing components, the cutoff frequency is set so that the aliasing
components are above 100 Hz. For example, when sampling frequency is converted from
560 to 280 Hz with the passband cutoff frequency of 100 Hz, the stopband cutoff
frequency is set to 180 Hz instead of 140 Hz. Note that the actual sampling frequency of
the sensor board may have 10 percent variation and is not exactly 560 Hz. The stopband
cutoff frequency normalized to the upsampled sampling rate is set to be smaller than 180/
560L, allowing the variation in the denominator. L is the upsampling factor. Likewise, the
normalized passband cutoff frequency is set to be larger than 100/560L. In this way, a
filter allowing aliased components outside of the frequency range of interest is designed
with fewer filter coefficients than a filter completely eliminating aliased components.
The length of the measured data is subject to the RAM size limitation. The sensing
function on the Imote2 first stores data in the on-board RAM; the size of RAM limits the
sensing duration. For example, a three-axis acceleration measurement using a 16-bit data
type for each axis and having 11,264 data points occupies about 67 kB of RAM. Because
sensing is first performed at a higher sampling frequency and then downsampled, the
RAM originally occupied by the measured acceleration data is about 134 kB. 256 kB of
RAM on the Imote2 does not allow continuous measurement of acceleration for an
indefinite amount of time. Once the memory space becomes full, sensing is stopped, the
stored data is copied to Flash memory, and sensing can be started again.
Acceleration signals from the Siglab spectrum analyzer and Imote2s are compared
with each other to examine the sensing capability of the Imote2. Because the sampling
frequency of the reference signals is different from that of Imote2 signals, which is set at
280 Hz, the reference signals are resampled to 280 Hz. The resampling algorithm
described in section 5.3 is utilized. Also, the Siglab system is not synchronized with
Imote2s, making direct comparison of the two time domain signals difficult. Signals from
the two data acquisition systems are synchronized using the resampling algorithm
discussed previously. To estimate the offset, the reference sensor signals are shifted on a
time axis by a specified offset and the correlation coefficient between the two signals is
calculated. Figure 7.14 provides a plot of the correlation coefficient against the time shift.
The offset between the two sets of clocks is determined as the time shift giving the

113
 0.08 0.2

0.04
Acceleration (g) 0.1

Acceleration (g)
0
0

-0.04
-0.1
Imote2 1 Imote2 1
Imote2 2 Imote2 2
-0.08 Reference 1 Reference 1
Reference 2 Reference 2
-0.2
10 10.02 10.04 10.06 10.08 10.1 10 10.02 10.04 10.06 10.08 10.1
Time (sec) Time (sec)

(a) longitudinal (b) transverse

0.4
Imote2 1
Imote2 2
Reference 1
Reference 2
0.2
Acceleration (g)

-0.2

10 10.02 10.04 10.06 10.08 10.1

Time (sec)
(c) vertical
Figure 7.15. Acceleration record from Imote2s and reference sensors.

maximum absolute correlation coefficient. Note that the time shift is not restricted to
multiples of the sampling period. The resampling algorithm with linear interpolation
allows noninteger multiple of the sampling period as a time shift. Synchronized signals are
then compared in time and frequency domains.
The measured signals from the Imote2s and PCB accelerometers show reasonable
agreement in the time domain. Figure 7.15 shows a comparison of the four time histories.
Two of them are from Imote2s and the other two are from the PCB accelerometers. The
vertical acceleration signals from Imote2s and PCB accelerometers are almost identical,
indicating high sensing performance of Imote2s. Transverse acceleration, however, is not
as good as the vertical acceleration. Signals from PCB accelerometers and Imote2s have
an observable difference, while signals from the same systems are close to each other.
Longitudinal acceleration shows noticeable differences among the four signals. Here, the
two signals from the PCB accelerometers show differences while the Imote2 signals
appear to be the same. The noise on Imote2 signals specific to each node is considered
small.

114
 0
10
Imote2 1
Imote2 2 0
Reference 1 10

Power spectral densities (g2/Hz)

Power spectral densities (g2/Hz)

Reference 2
-2
10
-2
10

-4
10
-4
10

Imote2 1
10
-6 Imote2 2
-6
10 Reference 1
Reference 2
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Frequency (Hz) Frequency (Hz)
(a) longitudinal (b) transverse

0
10
Power spectral densities (g2/Hz)

-2
10

-4
10

Imote2 1
-6 Imote2 2
10 Reference 1
Reference 2
0 20 40 60 80 100 120 140
Frequency (Hz)
(c) vertical
Figure 7.16. Acceleration power spectral densities of Imote2s and reference sensors.

The same observation is made when the signals are compared in terms of power
spectral density (see Figure. 7.16). Note that signals above 100 Hz cannot be directly
comparable because the cutoff frequencies of filters involved are set around 100 Hz. As
with the time domain comparison, the power spectral density of the longitudinal
accelerations shows differences between the two Imote2 signals. The observable
difference in the signals is not constant over frequency. Another finding from the plot for
the longitudinal acceleration is that Imote2s gives smaller signal magnitude above 80 Hz
than PCB accelerometers. In general, the Imote2 signals are close to each other.
Transfer functions between the Imote2 accelerometers, between reference
accelerometers, and between the Imote2 and reference accelerometers are then estimated
(see Figure 7.17). The transfer function magnitude plotted against frequency reveals
differences in the sensitivities of two signals as a function of frequency. As can be seen in
Figure 7.17, the transfer function magnitude for the vertical acceleration is nearly flat
below 100 Hz. The compared sets of signals are considered close to each other confirming
findings from Figures 7.15 and 7.16. The fluctuation found around 30 Hz, as well as
fluctuation in low-frequency range, is considered to be due to the signals being extremely
small, as shown in Figure 7.16. Transfer functions for the transverse acceleration confirm

115
 3 3
Imote2_2/Imote2_1 Imote2_2/Imote2_1
Imote2_1/Siglab_1 Imote2_1/Siglab_1
Siglab2/Siglab1 Siglab2/Siglab1

Magnitude of transfer function

Magnitude of transfer function

2 2

1 1

0 0
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Frequency (Hz) Frequency (Hz)
(a) longitudinal (b) transverse
3
Imote2_2/Imote2_1
Imote2_1/Siglab_1
Siglab2/Siglab1
Magnitude of transfer function

0
0 20 40 60 80 100 120 140
Frequency (Hz)
(c) vertical
Figure 7.17. Magnitude of transfer function among Imote2s and reference sensors.

that signals from the same data acquisition system are close to each other, but signals from
the Imote2 and the Siglab spectrum analyzer have noticeable differences. The transfer
functions for the longitudinal acceleration fluctuate considerably, indicating that the four
signals have discrepancies. In terms of the transfer functions, the longitudinal acceleration
measurements seem to be noisy, though the problem is not only specific to Imote2s but
also for the PCB accelerometers.
Imote2 and PCB accelerometer signals are also compared in terms of coherence (see
Figure 7.18). While the transfer function magnitude can detect frequency-dependent
amplification factors in two signals, the coherence function indicates the degree of
linearity between two signals. Signals that are linearly related result in a coherence
function that is one over all frequencies. Small values in the coherence function, on the
other hand, suggest nonlinearity or noisy measurements in the corresponding frequency
range. Vertical and transverse accelerations show coherence functions close to one. The
gradual drop of the coherence function between an Imote2 and PCB accelerometer signals
may indicate imperfect time synchronization between the two signals. Coherence
functions for the longitudinal acceleration are small in some frequency ranges. The

116
 1 1

0.8 0.8

Coherence function
Coherence function

0.6 0.6

0.4 0.4

0.2 0.2
Imote2- Imote2 Imote2- Imote2
Imote2- Siglab Imote2- Siglab
Siglab - Siglab
Siglab - Siglab
0 0
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Frequency (Hz) Frequency (Hz)
(a) longitudinal (b) transverse
1

0.8
Coherence function

0.6

0.4

0.2
Imote2- Imote2
Imote2- Siglab
Siglab - Siglab
0
0 20 40 60 80 100 120 140
Frequency (Hz)
(c) vertical
Figure 7.18. Coherence functions among Imote2s and reference sensors.

coherence function between the Imote2s is smaller than that between PCB accelerometers
around 50 Hz and above 80 Hz.
To investigate the sensitivity of the other axes of the Imote2 accelerometer, the
orientation of the Imote2 is changed. The vertical acceleration again seems to be accurate,
while the longitudinal acceleration shows noticeable differences among the signals. The
transverse acceleration measurements are not as accurate as the vertical acceleration
measurements. Similar to the previous experiment, the vertical acceleration is accurate
and the longitudinal measurement seems noisy. Therefore, the accuracy of the measured
accelerations does not appear to depend on the specific channel used on the Imote2.
There are several possible explanations for why the vertical acceleration is accurate,
while longitudinal acceleration appears noisy. First, consider the magnitude of the
accelerations. Because the truss’s motion is relatively small in the longitudinal direction
(see Figures 7.15 and 7.16), the signal-to-noise ratio (SNR) is small for longitudinal
measurements. However, similar experiments with smaller excitation also show that the
vertical acceleration is accurate even when signal’s strength is small; small SNR alone
cannot explain why the longitudinal measurements are noisy. Another possible

117
 200 3

Phase of spectral densities (degree)

Phase of spectral densities (degree)
2
100
1

0 0

-1
-100
-2

-200 -3
0 20 40 60 80 100 0 20 40 60 80 100
Frequency (Hz) Frequency (Hz)

(a) zoom-out (b) zoom-in

Figure 7.19. Phase of cross spectral densities.

explanation concerns crosstalk. The crosstalk specification for the Imote2 accelerometer is
a maximum of 2 percent, while that of the PCB accelerometer is 5 percent. Because
vertical acceleration is much larger than longitudinal acceleration, measured acceleration
signals for the longitudinal direction may contain substantial crosstalk. If the actual value
of the crosstalk varies from sensor to sensor within the maximum bounds, the measured
longitudinal acceleration may differ from sensor to sensor; thus, each sensor may have a
different contribution from the acceleration in the perpendicular axes. Further
investigation may confirm this conjecture.
Time synchronization of measured signals is now examined. Six Imote2s are placed
on structural nodes to measure acceleration response. Phase of cross-spectral densities
among these signals is considered to indicate the accuracy of time synchronization.
Synchronized signals are expected to have a phase of zero, while synchronization error
results in linear phase. The larger the error is, the steeper the slope of phase is. Figure 7.19
shows the phase of the cross-spectral densities between the reference node signal and
signals from the other five nodes. The phase is almost flat over the frequency range below
100 Hz. Theoretically, the phase is one degree at 100 Hz if the signals have a
synchronization error of 28 Ps. The phase indicates that the synchronization of the
measured signals is approximately 30 Ps.

7.2.2 NExT

Natural Excitation Technique (NExT) is implemented on the Imote2 using the model-
based data aggregation strategy described in section 5.1.2. The validity of this
implementation is examined by analyzing the same set of data both on a group of Imote2s
and on the PC.
The data to be analyzed is the acceleration response of the three-dimensional truss
structure. The data analyzed by Gao (2005) is converted to integers by shifting, scaling,
and rounding. Assuming the usage of a 16-bit ADC with approximately 12 effective bits,
the conversion maps the maximum data value to about 1,500 and the minimum value to

118
 -3
x10
8
Imote2
Matlab

Correlation function (g2)

4

-4

0 1 2
Time (sec)
Figure 7.20. Correlation function estimates.

about -1,500. Note that the full range of a 12-bit signed integer is from -2,048 to 2,047.
The integer data is then injected to the group of Imote2s using the reliable communication
protocol for 16-bit integer data type described in section 5.2.3.
The acceleration response data is first split into several blocks and then injected. A
block of 11,264 data points is injected to a node at a time, and this injection is repeated
three times to each node. Eventually each node receives 33,792 data points per sensing
channel. Dividing time histories into several blocks is incorporated to simulate the Imote2
sensing process, which must be performed repeatedly, as explained in section 7.2.1, to
collect long records of data.
NExT is performed on the data injected to the Imote2s corresponding to nodes 2 to 7
(see Figure 7.7). Node 4 is designated as the cluster head, organizing distributed
correlation function estimation. The number of data points used in the FFT calculations is
1,024, resulting in 21 time averages in estimation of the spectral densities for the 11,264
point data records and 63 time averages for the 33,792 point data records. Then the
spectral densities are converted to correlation functions by the inverse FFT. Figure 7.20
shows the cross-correlation function between the vertical acceleration responses at nodes
3 and 4 (see Figure 7.7) calculated on the Imote2 as well as that calculated on the PC. The
estimate from the PC is overlaid by the estimate from the Imote2. Note that only the first
256 data points, or one second, of the correlation function estimate on the Imote2 is
reported to the base station and the cluster head to reduce communication requirements.
The subsequent modal analysis using ERA usually utilizes only a portion of the
correlation function, eliminating the need to transfer the entire correlation function. Figure
7.21 shows the difference between a correlation function estimated on the Imote2 and the
corresponding correlation function estimated on the PC. From these figures, NExT on the
Imote2 is considered to be numerically equal to that on the PC with the precision of a
double data type.

119
 -17
x 10
1.5

Difference in correlation function (g2)

1

0.5

-0. 5

-1

-1. 5
0 0. 5 1
Time (sec)

Figure 7.21. Difference in correlation function estimates.

7.2.3 ERA

The ERA method is implemented on the Imote2 using the previously developed
eigensolver and SVD functions. To study the validity of the implementation, the outputs
from ERA on the Imote2 are compared with those calculated on the PC.
While the algorithm is well-established as described in section 6.2, the algorithm
implementation on Imote2 is subject to memory limitations, necessitating cautious
consideration. One of the numerical operations requiring a large amount of memory is the
SVD of the Hankel matrix as shown in Eq. (6.9) . The number of columns and rows is
often in the hundreds. A Hankel matrix with more than a thousand rows is not uncommon.
The Imote2’s 256 kB of on-board RAM, however, is filled to capacity by a double
precision matrix of size  u  . SVD of a matrix outputs two unitary matrices and a
vector containing the singular values, further limiting the size of the matrix to be analyzed
with ERA. In practice, some of the memory is already used by other variables, other
programs, and the OS. Therefore, the size of the Hankel matrix on the Imote2 is limited by
the available memory.
A  u  Hankel matrix, occupying about 19 kB of RAM, is employed for
implementation of ERA on the Imote2. In general, a small Hankel matrix results in the
inclusion of more noise modes in the modal identification. To avoid missing physical
modes in modal identification, the number of nonzero singular values in Eq. (6.9) is
assumed to be larger than the expected number of physical modes. Envisioning
identification of the truss vibration modes in the frequency range lower than 100 Hz, the
ERA implementation assumes 28 singular values as nonzero. Among 14 identified pairs of
modes, those with a small modal amplitude coherence (Juang & Pappa, 1985) or with a
large damping ratio are considered to be noise and are eliminated. Even after this noise
mode elimination, modes that do not show clear peaks on the cross-spectral density plots
may still be present in the identified sets of modes. The Imote2 first estimates the peaks of
the cross-spectral density amplitudes as local maximums, and picks the ERA output

120
modes that have natural frequencies close to these peaks as the physical modes. ERA is,
thus, implemented on the Imote2 having limited memory.
The accuracy of the developed ERA routine is now numerically examined. ERA is
applied to the correlation functions estimated in section 7.2.2. The Imote2 corresponding
to node 4 performs ERA and identifies the natural frequencies, damping ratios, mode
shapes, modal amplitude coherences, and initial modal amplitudes. Natural frequencies,
damping ratios, and mode shapes are compared with those calculated on the PC and are
summarized in Table 7.3. The error in the frequency and damping ratio estimates is
calculated as the difference between the estimates on the Imote2 and the PC. The
estimation error in the mode shapes are investigated in terms of the Modal Assurance
Criterion (MAC) defined in Eq. (7.1). As is seen in Table 7.3, the modal identification
results on the Imote2 and those on the PC are identical with the precision of double data
type.

Table 7.3. Identification of Natural Frequencies, Damping Ratios and Mode Shapes

Natural Frequency Damping Ratio Mode Shape

difference difference 1-MAC
mode f (Hz) –  ] (%) –  – 
(  ) (  ) (  )
1 20.1526 5.3078 3.5523 2.3794 2.2204
2 41.2039 -0.4263 0.3439 0.1252 0
3 61.9044 -0.1137 0.3812 -0.0944 0
4 66.9313 -0.0142 0.5763 -0.0648 -2.2204
5 70.4128 -0.5826 0.7909 -0.9630 2.2204
6 93.8481 -0.0426 0.3118 0.7626 -2.2204

7.2.4 DLV methods

Implementation of the DLV methods on the Imote2 is described in this section. These
implementations are validated through comparison between the results on the Imote2 and
the PC.
Deliberate consideration is given to the implementation of the DLV methods to
accommodate the limited hardware resources on the Imote2. One of the numerical
operations requiring a large amount of memory and CPU time is stress analysis under the
DLV loading. If an FEM model of the entire structure needs to be stored on a smart sensor
node, the model may exceed available memory. Numerical operations involved in the
analysis of the FEM model, i.e., calculation of static displacements and stresses under the
DLV loading, are not trivial. Instead, the linear nature of the analysis is used to simplify
the process. A matrix to convert input force to stress is calculated in advance and injected
to the cluster heads; rather than to run the entire structural analysis, the cluster heads
simply need to compute the product of the matrix and DLVs. Furthermore, the conversion
matrix needs to keep only the submatrix corresponding to the structural node and elements
in the local sensor community corresponding to the cluster head. The submatrix converts

121
input forces applied at nodes in the local sensor community to stress in elements in the
same community. The structural analysis is, thus, performed as the product of the modest
size matrix and DLVs
The DLV method implementation on the Imote2 is now numerically examined. The
DLV method based on a known mass perturbation is employed in this numerical
validation; a similar procedure was employed to implement and validate the SDLV
method.
Prior to monitoring local sensor communities, mass normalization constants are
estimated. These constants are also estimated on the Imote2. The structural responses of
the truss before and after the addition of the known mass at node 11 are injected to 10
Imote2s corresponding to nodes 7, 9, 11, 13, 15, 17, 19, 21, 23, and 25. The Imote2 at
node 21 works as the cluster head and applies NExT and ERA before and after the mass
perturbation. Using the identified modal parameters, the cluster head estimates the mass
normalization constants using Eq. (6.21). The estimated mass normalization constants are
listed in Table 7.4 together with the difference between the constants estimated on the
Imote2 and the PC. The mass normalization constant estimation on the Imote2 is
numerically identical to that on the PC.
Monitoring of the local sensor community consisting of six Imote2s at nodes 2, 3, 4,
5, 6, and 7 is now examined. The injection of data, correlation function estimates by
NExT, and modal identification by ERA are repeated for the truss acceleration response
data measured before and after element 9 is replaced with a thinner element. Identified
modal frequencies and mode shapes before and after damage, as well as the estimated
mass normalization constants, are used to construct flexibility matrices. As intermediate
results, the singular values in Eq. (6.24) estimated on the Imote2 and on the PC are
compared in Table 7.5. The DLVs calculated on the Imote2 are also compared with those
estimated on the PC. The singular values and DLVs are considered numerically identical.
The DLVs are then applied to the truss model to estimate the normalized accumulated
stress. As shown in Table 7.6, the stress in element 9 is smaller than the threshold value of
0.3. Because the normalized accumulated stress is only compared with the threshold
value, the stress does not need to have many effective bits. Therefore, the stress estimation
on the Imote2 is performed in a single-precision float data type having approximately
seven effective digits. The normalized accumulated stress estimated on the Imote2 and on
the PC is identical with respect to the precision of the float data type.

7.2.5 DCS logic

The DCS logic to find inconsistency in damage localization results is implemented on

the Imote2. Sensor communities close to each other exchange information about damaged
elements. Because the topology of sensor communities is formulated to have overlaps in
monitored elements, at least two sensor communities monitor each element of the
structure. If some communities find contradicting damage detection results, these
communities repeat sensing, data processing, and the DCS logic. Also, if sensor
communities detect damaged elements in a consistent manner, these communities report

122
Table 7.4. Mass Normalization Constant Estimation

Difference in Mass
Mass Normalization Normalization Constant
Mode Frequency
Constant – 
(  )
1 19.9078 0.011098 0.3530
2 41.1663 0.007703 1.0122
3 62.6000 0.008077 0.0298
4 67.3655 0.006587 -0.2212
5 71.4202 0.003260 -1.7906
6 94.2266 0.011514 0.7700

Table 7.5. Summary of SVD in the DLV Method

Singular Values DLVs

Values –  – 
Difference (  ) 1-MAC (  )
–  0.4224 0
2.8380 u10
–  0.5740 -2.2204
2.6216 u10
–  4.4483 -2.2204
1.9715 u10
–  0.0432 -4.4409
4.2358 u10
–  2.9774 -2.2204
1.3588 u10
–  -1.7915 -2.2204
3.0088 u10
–  1.8335 0
5.4073 u10
–  1.2435 0
6.8856 u10
– 4.0434 2.2204
1.2346 u10
– -3.6429 2.2204
1.6543 u10
– 4.1185 0
2.9298 u10
– 3.4197 0
1.1808 u10

the damaged elements to the base station. The details of the DCS logic are described
herein with reference to Figure 7.22.
When the manager sensor, which also works as the cluster head of a sensor
community, completes application of the DLV method, the manager initializes itself to
apply the DCS logic and sends a command to the next cluster head. For explanation
purposes, the cluster heads are assumed to be arranged as shown in Figure 7.22. Cluster
123
 CH1 CH2 CH3 …. CHn-1 CHn
Initialize Initialize Initialize Initialize Initialize

Manager sensor No

yes damaged element?

Inconsistent?
More inquiry?

No
yes
Flag|=
RETAKE
Move to
next

No No

yes damaged element? yes

Inconsistent? Inconsistent?
More inquiry

No
yes yes
Flag|= Flag|=
RETAKE RETAKE

Move to
next
No

damaged element? yes

Inconsistent?
More inquiry

No
yes
Flag|=
RETAKE

Report Report Report Report Report

to the BS to the BS to the BS to the BS to the BS

Inquire about Back to the Back to the Back to the Back to the
RETAKE Manager Manager Manager Manager

Back to the Back to the Back to the Back to the

Manager Manager Manager Manager

No No No
Manager
Retake? Retake? Retake? Retake?
again?

yes yes yes yes yes

Start sensing New manager New manager New manager New manager
as manager

Start sensing Start sensing Start sensing Start sensing

as manager as manager as manager as manager

Figure 7.22. DCS logic.

124
Table 7.6. Normalized Accumulated Stress Estimation

Normalized Accumulated –
Element ID Difference (  )
Stress
3 0.6377 -2.1416
4 0.6662 0.3378
5 0.9751 4.9952
6 0.4506 -2.5601
7 0.3470 1.9430
8 0.6078 -5.3461
9 0.1368 0.2829
10 1.0000 0
11 0.6226 0.2349

configuration. However, cluster heads do not need to physically line up; rather they need
to line up logically. In Figure 7.22, the manager sensor, CH1, notifies the cluster head to
the right, CH2. The receiver cluster head also initializes itself and notifies the next cluster
head. The initialization and notification are repeated until no cluster head to the right is
found.
The right-most cluster head then starts the procedure of finding inconsistent damage
detection results. This cluster head first checks whether damage is detected in its sensor
community. If not detected, the operation moves to the left cluster head. If detected, the
cluster head inquires with the neighboring cluster heads whether the damage detection
results are consistent. If inconsistent, a flag on the cluster heads involved in the transaction
are marked as “RETAKE”. The search for inconsistency then moves to the left cluster
head, where the procedure is repeated. This inconsistency search and transition to the left
node are repeated until the process reaches the left-most cluster head.
When the left-most cluster head finishes the inconsistency search, the process of
reporting to the base station begins. The left-most cluster head reports damage detection
results, i.e. damaged element ID and any inconsistencies, to the base station. After this
cluster head finishes reporting, the operation moves to the right cluster head. This
procedure is repeated until all of the cluster heads report their results. Then, the operation
moves back to the manager sensor. While all of the nodes are designed to report DCS
logic results to the base station for debugging purpose, the code can be easily changed so
that only cluster heads with damaged elements report to the base station.
The manager sensor then inquires of the other cluster heads whether local sensor
communities organized by these cluster heads need to retake data. Only local sensor
communities with inconsistent damage detection results participate in the next round of
DCS. If the manager sensor in the previous round of DCS does not participate in the next
round, one of the cluster heads takes over the manager sensor’s role.

125
 1
2 Node ID 102 RETAKE 1
3
4 INCONSISTENT
5
6 Inconsistent Element ID: 8
7
8 ID 102 # of Damaged Element: 0
9
10
11 Node ID 91 RETAKE 1
12
13 INCONSISTENT
14
15 Inconsistent Element ID: 8
16
17 ID 91 # of Damaged Element: 1
18
19 Damaged Element ID: 8

Figure 7.23. Report on damage elements.

This DCS logic is implemented on Imote2s and tested assuming acceleration

measurement of the three-dimensional truss structure. Two Imote2s corresponding to
cluster heads at nodes 4 and 6 receive simulated damaged element information from the
PC. These two Imote2s then apply the DCS logic. Figure 7.23 shows the report from these
two cluster heads after application of the DCS logic. In this case, the cluster head at node
4 is assumed not to have detected damage, while the cluster head at node 6 is supposed to
have detected damage at element 8, as indicated in lines 8, 17, and 19. These two nodes
find inconsistency and decide to retake data, as indicated by the RETAKE flag on lines 2
and 11. Node IDs on lines 2 and 11 are sensor node IDs unique to each Imote2 and
different from the truss’s structural node IDs. The DCS logic implementation on the
Imote2 is, thus, demonstrated.

7.2.6 Final implementation on the Imote2

With the developed components of DCS for SHM, the entire strategy is realized on
Imote2s. Simple flow charts of the process are shown on Figures 7.24 and 7.25. DCS for
SHM involves at least four sets of measurements: (a) measurement over the whole
structure without mass perturbation for mass normalization constant estimation, (b)
measurement over the whole structure with mass perturbation for mass normalization
constant estimation, (c) measurement in a local sensor communities to obtain the base line
flexibility matrix, F u , and (d) measurement in a local sensor communities for monitoring.
All of these measurements require similar procedures, with the differences being:
participating sensor members, the parameters injected at the beginning, and the numerical
calculation applied to the ERA results.
Extension of DCS for SHM using the SDLV, as explained in section 6.6, simplifies
these steps. The SDLV does not require estimation of mass normalization constants,
eliminating the need for the first two sets of measurement over the entire structure. The
third and fourth sets of measurements are the same as the original DCS for SHM, with the
exception being numerical operations in the DLV method.

126
 Initialization: without mass perturbation Initialization: with mass perturbation
Inject parameters Inject parameters
• NodeID • NodeID
• Measurement direction • Measurement direction
• # of samples • # of samples
• NExT/ERA parameters • NExT/ERA parameters
• Modal parameters from the previous
step
Time synchronization
• Send beacon signal
• Estimate local clock offset
Time synchronization
• Estimate clock drift • Send beacon signal
• Estimate local clock offset
• Estimate clock drift
Sensing
• Update local clock offset Sensing
• Start sensing • Update local clock offset
• Copy acquired data to global • Start sensing
variables • Copy acquired data to global
• Acquire timestamp of each block of variables
data • Acquire timestamp of each block of
• resample and synchronize the data
measured data • resample and synchronize the
measured data

Report to the base station

Report to the base station
Correlation function estimation
• broadcast data Correlation function estimation
• cross-spectral density estimation • broadcast data
• ifft • cross-spectral density estimation
• averaging • ifft
• averaging

yes
Need more yes
averaging ? Need more
averaging ?
no
no
Report to the cluster heads and base
Report to the cluster heads and base
station
station
ERA ERA/Mass normalization constant
estimate
Report to the base station
Report to the base station

Figure 7.24. The block diagram of mass normalization constant estimation.

At the beginning, parameters such as node IDs of the leaf nodes and measurement
directions are injected to the manager sensor and cluster head nodes. Also, these nodes are
notified which of the four sets of measurements shown in Figures 7.24 and 7.25 needs to
be performed. Information from the base station or users is transferred only at this stage. It
is considered possible to inject these parameters only once and store them on nonvolatile
memory. In this way, networks of smart sensors can work as autonomous stand-alone
systems.
At the end of parameter injection, time synchronization begins. To increase the
possibility of successful time synchronization, and to estimate clock drift, time
synchronization is repeated multiple times. Time synchronization is also performed later
between measurements.

127
 Initialization: within local sensor communities Monitoring
Inject parameters Inject parameters
• NodeID • NodeID
• Measurement direction • Measurement direction
• # of samples • # of samples
• NExT/ERA parameters • NExT/ERA parameters
• Modal parameters from the previous • Modal parameters from the previous
step step
• mass normalization constants

Time synchronization
• Send beacon signal Time synchronization
• Estimate local clock offset • Send beacon signal
• Estimate clock drift • Estimate local clock offset
• Estimate clock drift

Sensing
• Update local clock offset Sensing
• Start sensing • Update local clock offset
• Copy acquired data to global • Start sensing
variables • Copy acquired data to global
• Acquire timestamp of each block of variables
data • Acquire timestamp of each block of
• resample and synchronize the data
measured data • resample and synchronize the
measured data

Report to the base station

Report to the base station
Correlation function estimation
• broadcast data Correlation function estimation
• cross-spectral density estimation • broadcast data
• ifft • cross-spectral density estimation
• averaging • ifft
• averaging

yes
Need more yes
averaging ? Need more
averaging ?
no
no
Report to the cluster heads and base
station Report to the cluster heads and base
station
ERA
ERA/DLV
Report to the base station
Report to the base station

Figure 7.25. The block diagram of monitoring in a sensor community.

Synchronized sensing is performed in the manner described in section 5.3. The

number of sensing samples is set to be 11,264. The raw data is sent back to the base station
for debugging purposes. This reporting allows numerical operations on Imote2s to be
reproduced on the PC; the validity of data processing on the Imote2 can also be examined.
Centrally collecting all of the data takes a substantial amount of time and can be
eliminated once this DCS system is fully developed.
Correlation functions are then estimated in a distributed manner. The model-based
data aggregation approach explained in section 5.1 is employed. If the number of averages
in correlation function estimation, n d in Eq. (5.1), is smaller than the predetermined value,
sensing, reporting to the base station, and correlation function estimation are repeated (see
Figures 7.24 and 7.25).

128
 The estimated correlation functions are reported back to the cluster heads and base
station. While cluster heads need correlation function estimates from their leaf nodes, the
base station does not need the estimates except for the debugging purposes. Reporting to
the base station can be omitted once the system is fully developed.
Cluster heads then apply ERA to the correlation function estimates. Natural
frequencies and mode shapes are calculated. Based on these modal parameters, mass
normalization constants are estimated, or damage localization is performed. During
monitoring in the local sensor communities, the DCS logic is applied to the outcome of the
DLV procedure. The results are then sent back to the base station. Some of these results
are injected to cluster heads at the beginning of the next set of measurement. For example,
the estimated mass normalization constants are injected to the cluster heads at the
beginning of the monitoring process by the local sensor communities. This reporting is
also not necessary, once the system is fully developed. The outcome of one set of
measurements can be kept on RAM or nonvolatile memory for the next set of
measurements, eliminating the need for the parameter injection.
In the implementation of DCS for SHM, the frequency scalable feature of the Imote2
is utilized. The microprocessor on the Imote2 can operate at multiple frequencies. Because
the driver on TinyOS supports only the 13 and 100 MHz modes at the time, the
operational frequency is switched between only these two values. Numerical calculations
in NExT, ERA, and DLV, as well as sensing, are all performed at 100 MHz. Other tasks
such as communication are performed at 13 MHz.
All of the tasks in DCS for SHM need to be performed in the proper order by relevant
nodes. A 1-Byte variable containing instruction for the next task is utilized in organizing
all of the tasks. All of the communication packets have 1-byte variable to describe the
instructions. At the end of transmission and reception, this instruction is processed in a
task named “ProcessInstruction”. “ProcessInstruction” then looks for the block of codes
following the “switch” statement corresponding to the instruction value. Operations
described in the block of codes are executed. At the end of the block, new values can be
assigned to the instruction byte to execute the next task, or a communication packet
carrying new values for the instruction byte can be sent out. Components of DCS for SHM
explained in this chapter are synthesized in this manner. Table 7.7 summarizes the
instructions, associated operations, and information accompanying the communication
packet. The implementation on the Imote2 is summarized from Figures 7.26 to 7.30.

7.3 Summary
This chapter described the implementation of DCS for SHM on a network of Imote2s.
The validity of the system is numerically examined in a component-by-component
manner. In the next chapter, the Imote2s are installed on a three-dimensional truss and the
algorithms are experimentally verified.

129
 PC Base Manager Cluster Leaf nodes
station sensor heads
1 1
example
1 1
1 Send a message
PC ->BS
ProcessInstruction(1)
Store Receive a message
1
-NodeIDs of leaf nodes PC->BS
-Measurement channels
of leaf nodes
NEXT Send a message
BS -> PC
10 10
NEXT Receive a message
ProcessInstruction(10) BS->PC
NEXT
1 Send long data
NEXT NEXT

1 Receive long data

2a 2a

2a 2a 10 Send short message

ProcessInstruction(2a)
Store 10 Receive short
-Measurement direction. message
Inform leaf nodes
-Measurement direction

2b 2b

ProcessInstruction(2b)
Store
-Measurement direction.

10 10

ProcessInstruction(10)
NEXT

NEXT NEXT

4 4

4 4

ProcessInstruction(4)
Store
-NExT parameters (e.g., nfft)

10 10

ProcessInstruction(10)
NEXT

NEXT NEXT

3c 3c

3c 3c

ProcessInstruction(3c)
Store
-Number of samples

10 10

ProcessInstruction(10)
NEXT

NEXT NEXT

2x 2x

2x 2x

Figure 7.26. The implementation block diagram of monitoring in a sensor community (1).

130
 PC Base Manager Cluster Leaf nodes
station sensor heads
ProcessInstruction(2x)
Store
-Parameters for DCS.

10 10

ProcessInstruction(10)
NEXT

NEXT NEXT

26 26

26 26

ProcessInstruction(26)
Store
-NodeIDs.
Call TimerDrift.start

event TimerDrift.fired() Synchronize

Broadcast Beacon signal Estimate clock drift
if repetion > 10,
ProcessInstruction (32)

ProcessInstruction(32)
Broadcast Beacon signal
Calculate t1
post ProcessInstruction(33)

ProcessInstruction(33)
Inform leaf nodes
-t1, T1, T2, etc.

34 34

ProcessInstruction(34) ProcessInstruction(34)
Store Store
-t1, T1, T2, etc. -t1, T1, T2, etc.
Adjust offset Adjust offset
Calculate t1+T2 Calculate t1+T2
Start TimerGTCheck() Start TimerGTCheck()

event TimerGTCheck.fired() event TimerGTCheck.fired()

Get global time Get global time
if global time > (t1+T1) if global time > (t1+T1)
Start Sensing Start Sensing
else
broadcast beacon signal

startSensing() startSensing()
Initialize sensing parameters Initialize sensing parameters
allocate memory for sensing allocate memory for sensing
call TimerSensing.start() call TimerSensing.start()
start sampling start sampling

WriteData.write() WriteData.write()
Copy a block of data to Copy a block of data to
global variables global variables
Get timestamp for each block Get timestamp for each block

putData() putData()
if Done if Done
stop sensing stop sensing
call TimerSensing.stop()

event TimerSensing.fired() event TimerSensing.fired()

if sensing failed if sensing failed
ProcessInstruction(32) stop sensing
else
ProcessInstruction(38)
32

Figure 7.27. The implementation block diagram of monitoring in a sensor community (2).

131
 PC Base Manager Cluster Leaf nodes
station sensor heads
ProcessInstruction(38)
Inquire of leaf nodes whether
sensing is done

39 39

ProcessInstruction(39)
Answer inquiry

3a 3a

ProcessInstruction(3a)
if failure is reported
ProcessInstruction(32)
If all nodes have done sensing
ProcessInstruction(3b)
else
ProcessInstruction(38)

32
3b 3b

ProcessInstruction(3b)
ProcessInstruction(3b)
Resampling
Resampling
Check unknown error
Check unknown error
ProcessInstruction(50)
32

ProcessInstruction(50)
Inquire of leaf nodes whether
resampling is successful

51 51

ProcessInstruction(51)
Answer inquiry

52 52

ProcessInstruction(52)
if failure is reported
ProcessInstruction(32)
If all nodes have done sensing
ProcessInstruction(53)
else
ProcessInstruction(50)
32
53 53

ProcessInstruction(53) ProcessInstruction(53)
Adjust data measurement Adjust data measurement
direction direction
Write data to Flash Write data to Flash
ProcessInstruction(3d)

ProcessInstruction(3d)
if not reported
Report to the base station

3e 3e

data data

3e 3e

ProcessInstruction(3e)
if not reported
ProcessInstruction(3d)
else if every node has reported
ProcessInstruction(54)
else
ask leaf nodes to report to
the base station

3d 3d

Figure 7.28. The implementation block diagram of monitoring in a sensor community (3).

132
 PC Base Manager Cluster Leaf nodes
station sensor heads
ProcessInstruction(3d)
3e Report to the base station

3e 3e

data data

3e 3e

ProcessInstruction(54)
Inform cluster heads that data
is ready for broadcast
If all cluster heads have
broadcast data,
-ProcessInstruction(40)

55 55

ProcessInstruction(55)
broadcast NExT parameters

5 5

ProcessInstruction(5) ProcessInstruction(5)
ProcessInstruction(6) Store
-NExT parameters
ProcessInstruction(6)
broadcast reference sensor
signal

7 7

ProcessInstruction(7) ProcessInstruction(7)
broadcast Instruction 54 Store
-reference sensor signal

54 54

40 40

Correlation function estimation Correlation function estimation

if need more averaging
ProcessInstruction(e)
else
ProcessInstruction(41)

ProcessInstruction(e)
Inquire of leaf nodes whether
correlation functions are
estimated

f f

ProcessInstruction(f)
answer inquiry

20 20

ProcessInstruction(20)
If answers from all nodes
-ready for measurement
ProcessInstruction(32)
else
-ProcessInstruction(e)
32

ProcessInstruction(41)
ProcessInstruction(48)
ProcessInstruction(48)
if not reported
Report correlation function to
the base station
else
ask leaf nodes to report to
the base station

4a 4a 49 49

Figure 7.29. The implementation block diagram of monitoring in a sensor community (4).

133
 PC Base Manager Cluster Leaf nodes
station sensor heads

data data

4c 4c

ProcessInstruction(4c)
if cluster head has reported
ProcessInstruction(41)
else
ProcessInstruction(48)

ProcessInstruction(49)

4a 4a 4a
ProcessInstruction(41)
if all clusters finished report
ready for ERA ProcessInstruction(4a)
else Store correlation functions
Inquire of the next cluster

ERATask()
apply ERA
if initialization,
compute normalization
constants, etc.
else
estimate accumulated stress

21 21

data data

2c 2c

If all the clusters have reported

if monitoring,
DCS logic
else
Inquire cluster heads

45 45

21 21

60 60

DCS logic

Figure 7.30. The implementation block diagram of monitoring in a sensor community (5).

134
Table 7.7. Instructions in DCS Code

Instruction Associated Operations Accompanying Information

The cluster heads save node IDs and Node IDs of cluster members, measurement
1
measurement channels of their leaf nodes. channels of each member node
The base station tells the PC that the base
10 station is ready to receive the next Acknowledgment
instruction.
Cluster heads save measurement directions
2a of member nodes and forward this Measurement directions of nodes
information to the leaf nodes.
The leaf nodes save measurement Measurement directions of cluster member
2b
directions. nodes
NExT parameters such as the number of
4 The cluster heads save NExT parameters.
FFT data points
Sensing parameters are saved on cluster Sensing parameters such as Number of
3c
heads. samples
The number of elements, weight of
Received parameters are saved on cluster
22-25 additional mass, results from the previous
heads.
set of measurement, etc.
The manager receives node IDs of all the
nodes and cluster heads. The manager then Node IDs of all the leaf nodes and cluster
26
starts time synchronization and estimates heads
clock drift.
The manager broadcasts beacon signal for
32
time synchronization.
The manager notifies leaf nodes of timings,
33
t1, T1, and T2.
All of the nodes save t1, T1, and T2. A timer
34 is called to periodically check timing to start Timing to start sensing, i.e. t1, T1, T2, etc.
sensing.
The receivers reply to inquiry message
39
about sensing completion.
The manager judges whether to keep
3a inquiry, restart sensing, or apply Answer to inquiry about sensing completion
resampling.
All of the sensors start resampling, check
3b for unknown error. The manager then posts
ProcessInstruction(50).
The manager asks the leaf nodes whether
50
the resampling was successful.
The leaf nodes reply to the inquiry about
51
successful resampling.
The manager judges whether to keep
Answer to inquiry about successful
52 inquiry, restart sensing, or adjust
resampling
measurement directions.

135
Table 7.7. Instructions in DCS Code (Continued)

Instruction Associated Operations Accompanying Information

All of the nodes adjust measurement data
Sensing was successfully completed at all of
53 for sensor direction. Nodes then write data
the nodes.
to the Flash memory.
If measurement data has not been reported
3d to the base station, the node reports the data
to the base station.
If the receiver has not reported its data, the
receiver reports to the base station. If all of
the leaf nodes have not reported their data,
Raw data/acknowledgment from the base
3e the manger asks them to report. If all of the
station
nodes have finished reporting, the manager
asks cluster heads to broadcast data to their
members.
Cluster heads broadcast NExT parameters
55
to their members.
The sender prepares to broadcast the
The number of FFT data points, number of
5 reference signal. The receiver stores NExT
averages, etc.
parameters.
The cluster heads broadcast reference
6
signals.
The sender reports to the manager that the
7 broadcast was successful. The receiver Reference sensor signal
stores reference sensor signal.
Report to the manager about successful
54
broadcast.
All of the nodes start correlation function
40
estimation.
The manager inquires all of the leaf nodes
e whether correlation function estimation is
complete.
Inquiry about correlation function
f The receiver replies to the inquiry.
estimation completion
If all of the nodes have replied, the manager
20 starts sensing process by posting Answer to the inquiry
ProcessInstruction(32). If not, keep inquiry.
If correlation function estimate on a cluster
head is not reported to the base station, the
cluster head report to the base station.
48
Otherwise, the cluster head asks leaf nodes
to report their correlation functions to the
cluster head and base station.
The receiver reports the correlation function
49 estimates to the cluster head and base Inquiry about correlation function estimate
station.
4a Correlation function estimates

136
Table 7.7. Instructions in DCS Code (Continued)

Instruction Associated Operations Accompanying Information

If the cluster head has collected all of the
correlation functions from its members,
4c report to the manager. Otherwise, the cluster
head keeps asking leaf nodes for their
correlation function estimates.
If all cluster heads finished collecting
correlation function estimates from their
members, the manger asks cluster heads to
41
perform ERA. Otherwise, ask the next
cluster head to collect correlation function
estimates.
21 ERA results
If all of the clusters have not reported ERA
results, the manager asks cluster heads to
report. After all of the clusters report to the
2c Acknowledgment
base station, DCS logic is initiated if the
system is monitoring structures to detect
damage.
The manager asks cluster heads to report
45
ERA results to the base station.
60 The receiver initiates DCS logic.

137
 Chapter 8

EXPERIMENTAL VERIFICATION

This chapter provides experimental verification of the SHM framework developed in

this research. The SHM framework employing the DCS for SHM was realized on
networks of Imote2s in Chapter 7. Numerical operations of the SHM system were
numerically examined in a component-by-component manner by injecting acceleration
response data to the network of Imote2s and processing the data. The system has been
shown to be capable of processing acceleration data and localizing damage.
The experimental verification of the proposed SHM framework uses acceleration
response data measured by networks of Imote2s. Acceleration measurements and the
subsequent data processing are performed before and after an element of the truss is
replaced with a thinner element having a 52.7 percent cross-section reduction. The
replaced element is detected by cluster head Imote2s. The experiment is repeated multiple
times to assess reproducibility. Also, the experiment is repeated by replacing a different
structural element to see the damage detection capability of the system in various
elements. In this way, the ability of the proposed framework to localize damage is
experimentally verified (Spencer & Nagayama, 2006).

8.1 Experimental setup

The SHM framework is experimentally verified on networks of Imote2s using the
5.6m-long, three-dimensional truss structure discussed in Chapter 7. The truss and the
structural node IDs and element IDs are shown again in Figures 8.1 and 8.2. The shaker is
attached at node 17 and excites the truss with a band-limited white noise.
The Imote2s are installed on the truss using plastic fixtures (see Figure 8.3). The
fixture has two walls, each of which has two threaded holes. A set-screw is inserted into

Figure 8.1. The Imote2 and plastic case.

138
             
6 10 14 18 22 26 30 34 38 42 46 50
3 7 11 15 19 23 31 35 39 43 47 51
27
2 5 9 13 17 21 25 29 33 37 41 45 49 53
1 4 8 12 16 20 24 28 32 36 40 44 48 52
              

Figure 8.2. Pin and roller ends.

Figure 8.3. Plastic fixture and the Imote2.

Figure 8.4. Ten Imote2s installed on nodes 2 to 11.

each of the threaded holes and tightened. The Imote2 is, thus, supported by the bottom of
the plastic fixture and the four set-screws.
Ten Imote2s are placed on the structure on the front panel of the truss creating three
overlapping local sensor communities. Because the number of available Imote2 sensors is
limited, sensors are moved on the structure, depending on which element is considered as
damaged. When damage in element 8 is considered, Imote2s are installed on nodes 2 to
11, with nodes 4, 6, and 8 becoming cluster heads that organize the local sensor
communities (see Figure 8.4). For example, node 4 organizes the local sensor community
consisting of nodes 2, 3, 4, 5, 6, and 7. When damage in element 20 is considered, Imote2s

139
 -3
x 10
8

Correlation function (g2)

4

-4

0 0.4 0. 8 1. 2 1.6
Time (sec)

Figure 8.5. Correlation function estimate from Imote2 data.

4
10
Cross spectral density (g2/Hz)

0
10

-4
10

0 20 40 60 80 100
Frequency (Hz)

Figure 8.6. Cross-spectral density estimates from Imote2 data.

are placed on nodes 6 to 15, and nodes 8, 10, and 12 become cluster heads. Local sensor
communities organized by nodes 4, 6, 8, 10, and 12 are denoted as sensor communities 1,
2, 3, 4, and 5, respectively. Each node measures acceleration in three directions. Because
this validation exercise is monitoring the vertical plane of the truss, only longitudinal and
vertical acceleration records are utilized.

8.2 NExT
The acceleration responses are measured by Imote2s, and the correlation functions are
estimated. The vertical acceleration signal at the cluster head node is used as the reference
signal. One of the estimated correlation functions and the associated frequency domain
representation, the cross-spectral densities, are shown in Figures 8.5 and 8.6. These
estimates show agreement with those estimated on a PC using the same data from
Imote2s.

140
8.3 ERA
ERA results during monitoring are summarized in Table 8.1. ERA on a PC using the
same data gives the same results numerically. Note that the structure has undergone some
structural changes since the data injected to Imote2 networks in Chapter 7 was acquired;
some of connections were tightened and some of elements were adjusted. The identified
natural frequencies in Table 8.1 are not necessarily same as those in Table 7.3. By
comparing the identified natural frequencies with the cross-spectral density plots in Figure
8.6, the identified natural frequencies are considered to well capture the dynamic
characteristics of the signals.
Table 8.1. Identification of Natural Frequencies and Damping Ratios

Natural Frequency (Hz) Damping Ratio (%)

20.7229 2.9321
32.8985 0.2127
41.2990 0.4846
64.3174 0.2422
69.1086 0.3718
94.9259 0.2810

8.4 DLV methods

The DLV method based on the mass perturbation and SDLV methods are applied to
the modal parameters identified before and after damage. The mass perturbation DLV
method, involving initialization to determine the mass normalization constants is first
examined. Subsequently, the SDLV method is investigated.
The mass normalization constants to be used in the mass perturbation DLV method
are first estimated. Imote2s are installed at nodes 7, 9, 11, 13, 15, 17, 19, 21, 23, and 25.
Measurements are taken before and after additional mass is attached to node 11. The mass
is 1.253 kg. All the nodes measure vertical acceleration. The node at the location of the
additional mass measures acceleration in the longitudinal and transverse directions as
well, because the denominator of Eq. (6.21) requires mode shapes at the degrees-of-
freedom corresponding to the mass perturbation, i.e., acceleration in all three directions is
needed. Thus, acceleration is measured in 12 channels. The mass normalization constants
are estimated and listed in Table 8.2, as well as natural frequencies of the structure with
and without the mass perturbation.
The identified mass normalization constants are utilized to reconstruct the flexibility
matrix as shown in Eqs. (6.19) and (6.32). Imote2s monitoring sensor communities 1, 2,
and 3 reconstruct the flexibility matrices before and after element 8 is replaced with an
element of reduced cross-section. The Imote2s then perform an SVD on the difference in
the flexibility matrices to estimate the DLVs. The DLVs are then multiplied by a matrix to
convert nodal force to element stresses. Elements with small stresses are identified as

141
Table 8.2. Mass Normalization Constant Estimates from Imote2 Data

Natural Frequency (Hz) Mass Normalization

Without Mass Perturbation With Mass Perturbation Constant

20.4680 20.3722 0.004776

32.7488 32.1563 0.001474
41.2410 40.9012 0.007941
64.5915 64.0984 0.009026
69.3455 69.0262 0.005865
95.5995 95.1737 0.011405

1
Normalized accumulated stress

0.8

0.6

0.4

0.2

0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID

Figure 8.7. Normalized accumulated stress in sensor community 1, 2, and 3 (the mass
perturbation DLV method).

potentially damaged elements. Figure 8.7 shows the normalized accumulated stress
calculated for the three sensor communities. Element 8 is identified as the element with a
normalized accumulated stress smaller than the threshold value of 0.3. Note that the stress
in element 18 is also small. The DLV method identifies potential damage sites as elements
with small stress. The method does not exclude undamaged elements from a set of
elements with small stress. Gao (2005) pointed out that such false-positives in damage
detection do not appear repeatedly; repetitive measurements and damage detection can
distinguish damaged elements as elements which consistently show small stress. The
results in Figure 8.7 support that the mass perturbation DLV method on Imote2s can
detect damage.
The SDLV method is then employed as the damage detection method. This method
does not require estimation of mass normalization constants illustrated in Figure 7.24. The
same set of sensor communities, 1, 2, and 3 again measures acceleration responses before
and after element 8 is replaced with an element of reduced cross section; subsequently, the
SDLV method is applied. Using the conversion matrix, the normalized accumulated
stresses are calculated from DLVs. Figure 8.8 shows the accumulated stresses estimated

142
 1

Normalized accumulated stress

0. 8

0. 6

0. 4

0. 2

0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID

Figure 8.8. Normalized accumulated stress in sensor community 1, 2, and 3 (the SDLV
method).

by the three cluster head nodes. Element 8 is detected as an element with the normalized
accumulated stress smaller than the threshold value, 0.3.
The intermediate and final results of the two DLV methods such as singular values,
DLVs, and the normalized accumulated stress are also calculated on a PC based on the
centrally collected acceleration time histories. The results calculated on the Imote2 are the
same as those on the PC with the same precision of the data type.
The damage detection capability of the proposed SHM framework implemented on a
network of Imote2s has been demonstrated in this section. Further investigation of the
capabilities of the approach will be pursued in section 8.8 by replacing other elements of
the truss.

8.5 DCS logic

Upon completing calculation of the normalized accumulated stresses, the three cluster
heads exchange their damage detection results to perform the DCS logic given in Figure
7.22. The log of the associated session recorded on the base station is presented in Figures
8.9 and 8.10. These figures correspond to the damage detection results in Figures 8.7 and
8.8, respectively. Because no inconsistency is reported, retake flags are set to zero as
indicated by lines2, 9, and 16 in Figures 8.9 and 8.10. When no inconsistency is found, the
results are reported to the base station, and the Imote2s prepare to enter a sleep mode.
Though element 18 is identified as a damaged element by the mass perturbation DLV
method, this false-positive is likely to be avoided if sensor community 4 also monitors this
element because of the redundancy. When inconsistency is detected, the DCS algorithm is
repeatedly applied. Figure 8.11 shows the report to the base station when inconsistency is
observed. Sensor communities 3, 4, and 5 monitor the truss to detect damage at element
21. Sensor communities 3 and 4 organized by cluster head nodes 73 and 135 find
inconsistency. Sensor community 3 identifies element 16 as damaged, while sensor
community 4 does not. After all three cluster heads finish reporting, only sensor
communities 3 and 4 repeat acceleration measurement and data processing. The DCS

143
 1
2 Node ID 67 RETAKE 0
3
4 ID 67 # of Damaged Element: 1
5
6 Damaged Element ID: 8
7
8
9 Node ID 81 RETAKE 0
10
11 ID 81 # of Damaged Element: 1
12
13 Damaged Element ID: 8
14
15
16 Node ID 73 RETAKE 0
17
18 ID 73 # of Damaged Element: 1
19
20 Damaged Element ID: 18
21

Figure 8.9. DCS logic log (the mass perturbation DLV method).

1
2 Node ID 67 RETAKE 0
3
4 ID 67 # of Damaged Element: 1
5
6 Damaged Element ID: 8
7
8
9 Node ID 81 RETAKE 0
10
11 ID 81 # of Damaged Element: 1
12
13 Damaged Element ID: 8
14
15
16 Node ID 73 RETAKE 0
17
18 ID 73 # of Damaged Element: 0

Figure 8.10. DCS logic log (the SDLV method).

logic is applied again, and the results are reported to the base station as shown from lines
30 to 46 in Figure 8.11. No inconsistency is reported this time and all of the Imote2s are
ready to enter the sleep mode.

144
 1
2 Node ID 73 RETAKE 1
3 INCONSISTENT
4
5 Inconsistent Element ID: 16
6
7 ID 73 # of Damaged Element: 1
8
9 Damaged Element ID: 16
10
11
12 Node ID 135 RETAKE 1
13 INCONSISTENT
14
15 Inconsistent Element ID: 16
16
17 ID 135 # of Damaged Element: 1
18
19 Damaged Element ID: 21
20
21
22 Node ID 70 RETAKE 0
23
24 ID 70 # of Damaged Element: 1
25
26 Damaged Element ID: 21
27
28 ....
29
30 Node ID 73 RETAKE 0
31
32 ID 73 # of Damaged Element: 0
33
34
35 Node ID 135 RETAKE 0
36
37 ID 135 # of Damaged Element: 1
38
39 Damaged Element ID: 21
40
41
42 Node ID 70 RETAKE 0
43
44 ID 70 # of Damaged Element: 1
45
46 Damaged Element ID: 21
47

Figure 8.11. DCS logic log with inconsistency (the SDLV method).

8.6 Calculation and communication time

The Imote2 communication and calculation speeds are limited, raising concerns over
the time necessary to perform the SHM strategy. The shorter time is preferable from two
perspectives: power consumption and monitoring interval. Although longer time required

145
for a SHM strategy does not necessarily imply larger power consumption, the time can be
a rough indicator of power consumption. If nodes do not enter a sleep mode or turn off
radio components during monitoring, their power consumption correlates well with the
time of operation. With respect to monotoring interval, its frequency is limited by the
length of execution time; an SHM strategy that takes four hours to complete, for example,
can be performed at most only once in four hours. Considering monitoring of structures on
a weekly or monthly basis, Imote2s do not need to complete the SHM strategy in a short
time, e.g. one or two hours. However, monitoring in a short time is preferable to allow
multiple measurements before environmental conditions such as wind velocity and
temperature change. When damage detection results from neighboring sensor
communities are inconsistent with each other, sensing and damage detection need to be
repeated. If the SHM strategy takes a long time, environmental conditions may change
between monitoring intervals, making comparison among measurements difficult.
Therefore, SHM strategies requiring a short execution time are preferable.
In this section, the time required for execution of the DCS for SHM on Imote2s is
determined. The implementation includes numerous reports sent to the base station for
debugging purposes. Considering that these tasks can be omitted in the future, the time
necessary for the SHM strategy is estimated.
The breakdown of the time spent during DCS for SHM is listed in Table 8.3. At the
beginning of monitoring, parameter initialization requires about 70 seconds. Parameters to
be initialized include: the conversion matrix from nodal force to stress in elements, modal
parameters identified before damage, sampling rate, and duration of sensing. A six-
second-long time synchronization activity follows the parameter initialization. Sensing
continues for approximately 50 seconds. Note that the duration of sensing can be
increased if sufficient RAM is available for temporary storage of the measured data. At
the end of sensing, the manager sensor queries each node to see if sensing was successful.
Then the resampling process is applied to the measured data; resampling takes ten
seconds. The manager node again queries each node to ensure successful resampling. The
measured data is then reported to be the base station. The communication between a leaf
node and the base station is about five seconds, while communication between the base
station and a PC takes 12 seconds. Each Imote2 node reports three axes of measurement
data. This reporting is one of the tasks requiring significant time. Then the cluster heads
multicast their own data to the leaf nodes as reference signals. NExT parameters such as
the number of averages and the number of FFT points are also multicast. This multicast of
parameters and data takes only 33 seconds for three sensor communities. Then NExT is
applied on each node. Nodes which belong to multiple sensor communities estimate
multiple sets of correlation functions. The estimated correlation functions are sent to the
cluster heads and the base station. The base station then forwards the correlation functions
to the PC. This correlation function report takes more than 400 seconds. Finally, ERA is
applied to the correlation functions, and the DLV method determines the potential
locations of damaged elements. This data processing on the cluster heads takes 150
seconds. In total, DCS for SHM takes more than 23 minutes.
Among the tasks listed in Table 8.3, some of them can be omitted when debugging is
not necessary. The raw data does not need to be sent to the base station. About 500

146
Table 8.3. Breakdown of the Time Spent During DCS for SHM

Time (sec)
Time (sec)
Without
Task With Debugging
Debugging
Purpose Tasks
Purpose Tasks
1. Parameter injection to cluster
70 0
heads and the manager node
2. Time synchronization 6 6
3. Sensing 50 50
4. Query 10 10
5. Resampling 10 10
6. Query 5 5
7. Report to the base station 5 0
8. Communication between the base
12 0
station and a PC
9. Repeat 7 and 8 (# of nodes x 3
~ 493 (=17x29) 0
axis -1) times.
10. Multicast the NExT parameters. 5 5
11. Multicast reference signals. 6 6
12. Repeat 10 and 11 (# of cluster
~ 22 (=11x2) 22
head nodes-1) times.
13. NExT 42 42
14. Repeat 13 (# of sensor
communities a node belongs to) ~84 (=42x2) 84
times.
15. Report correlation functions to
cluster heads and the base 6 6
station.
16. Communication between the base
6 0
station and a PC
17. Repeat 15 and 16 (# of cluster
heads x # of leaf nodes in a ~420 (= 35x12) 174
community x 2 - 1) times.
18. ERA/DLV 150 150
Total 1,402 570

seconds spent for this reporting are saved. Also, correlation function reports do not need to
be sent to the base station and the PC. Six seconds spent for the communication in task 16

147
are saved. The 420 seconds listed at task 17 can be reduced by a factor of two.
Furthermore, only the beginning portion of the correlation function is utilized to construct
the Hankel matrix in Eq. (6.8), while the current system reports the entire correlation
function to the base station and cluster head nodes. Therefore, the report to the cluster
heads can be even shorter, though its contribution toward shortening the SHM execution is
not counted here. Subtracting the time spent for these debugging tasks from the 1,402
seconds makes the total time for the SHM strategy approximately 676 seconds; the usage
of 400 MHz mode of the CPU may further shorten the time necessary to execute the SHM
strategy.
In summary, about one minute of sensing and 9 minutes of post processing is
necessary for the current SHM system without sending debugging information to the base
station. Note that the time needed for sensing may increase if natural frequencies of a
structure appear in a lower frequency range. As for the time for post processing, the time
does not increase as the number of sensor community increases. Though tasks 12, 14 and
17 in Table 8.3 seem to be proportional to the number of sensor communities, they are
proportional only when the number of communities is small. When one cluster is out of
the communication range of another cluster, these two clusters can perform the tasks in
12, 14, and 17 simultaneously. Monitoring damage on a large structure requiring only
about nine minutes of data processing in addition to sensing time is an appealing feature of
this approach.

8.7 Battery life

The battery life of the Imote2 while running the proposed SHM strategy
implementation is estimated in this section. When Imote2s performed sensing, the change
in the supply voltage to the Imote2s is recorded. In contrast to the previous section,
sensing was repeated four times in this experiment. That is, tasks 2 to 14 in Table 8.3 were
repeated four times. The voltage values of five Imote2s are read after parameters are
initialized, after the first sensing is performed, after reporting to the base station is
completed, after the second, third, and fourth measurement finishes, and at the end of
ERA and DLV method applications.
Figure 8.12 shows the change in the supply voltage to the Imote2s. One round of
measurement consumes about 0.03 to 0.08 V. Note that sometimes sensing fails and is
repeated until successful measurement is performed. When repetition takes place many
times, the decrease in the voltage is considered significant. If sensing becomes more stable
with less sensing failures, the number of repetitions will decrease reducing the power
consumption per round of sensing.
While Imote2s repeat sensing four times for the data shown in Figure 8.12,
monitoring of full-scale structures may not need repeated measurements, thus reducing the
power consumption per monitoring event. The advantage of repeating sensing is to reduce
the effect of observation noise by increasing the number of averages in the spectral
densities in Eq (5.1). The SHM implementation explained in sections 8.2 to 8.5, as well as
that in section 8.8, measures acceleration responses of the truss only once, yielding 21
averages. From the damage detection results shown in section 8.4, 21 averages seem to be

148
 4
Node 46
Node 43
Node 51
3.8 Node 49
Node 87

Supply voltage (V)

3.6

3.4

3.2

3
s tart report s ens ing3 E RA
s ens ing1 s ens ing2 s ens ing4

Figure 8.12. Change in the supply voltage to the Imote2.

sufficient. Although the change in supply voltage during these experiments without
repeated measurements was not recorded, the battery life is estimated by counting the
number of experiments conducted using one set of batteries. The Imote2 utilizes three
AAA batteries. A new set of batteries supplies 4.5 to 4.8 V to the Imote2. The Imote2 can
operate when the supply voltage is above 3.2 V. More than 30 sets of experiments were
performed without changing batteries. Simply dividing the change in the supply voltage
by 30 yields about a 0.05 V decrease per experiment when sensing is not repeated. If
monitoring is performed on a weekly basis, a battery life providing 30 sets of
measurements and damage detection calculations allows the Imote2 to monitor a structure
for about eight months. Note that the Imote2 consumes power even in a sleep mode,
shortening the battery life.
For monitoring of full-scale structures, sensing may continue longer, thereby
consuming more power. For example, 40 seconds of acceleration data sampled at 280 Hz
is utilized in the experiment and damage is successfully detected. However, a full-scale
structure may have lower natural frequencies than the lowest one of the truss (i.e., the
mode at about 20 Hz), necessitating longer acceleration measurements. Though a lower
sampling frequency may be allowed for such structures, the duration of sensing is likely to
be significantly longer than 40 seconds. Battery depletion may become greater than 0.05
V per monitoring event due to the longer sensing time.

8.8 Damage detection results

The damage detection capability of the SHM algorithms implemented on the Imote2
is examined in this section by monitoring the truss with its various elements replaced with
a “damaged” element, i.e., one of reduced cross section. Elements 8, 9, 10, 11, 12, 19, 20,
21, and 22 are replaced one-by-one with the damaged element. For the first five cases,
sensor communities 1, 2, and 3 participate in damage detection, while sensor communities

149
 1 1

measurement 1
measurement 1

Normalized accumulated stress

Normalized accumulated stress
0. 5 0. 5
0. 3 0. 3

0 0
1 1
measurement 2

measurement 2
0.5 0. 5
0.3 0. 3

0 0
1 1
measurement 3

measurement 3
0. 5 0.5
0. 3 0.3

0 0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19 3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID Element ID
(a) (b)
Figure 8.13. Damage detection (damaged element: Element 8): (a) mass perturbation
DLV; and (b) SDLV.

3, 4, and 5 monitor for the damaged element for the other cases. The experiment is
repeated at least three times for each case to assess the repeatability of damage detection.
The SDLV method is employed as the DLV method on Imote2s, while the damage
detection capability of the mass perturbation DLV method is also examined on a PC. As
stated in section 8.4, the DCS for SHM using known mass perturbation has been
implemented on Imote2s, and the same numerical operations on Matlab on the centrally
collected truss acceleration response data has been confirmed to produce the same
calculation results as those on Imote2s. Therefore, for convenience, the mass perturbation
DLV method is applied on Matlab to the centrally collected data to investigate the damage
detection capability.
Figures 8.13 to 8.21 show the damage detection results using the two DLV methods.
For most of the experiments, both of the DLV methods can detect the damaged element,
i.e., it has a normalized accumulated stress smaller than 0.3. In some cases, however,
false-positive and/or false-negative damage detection is observed.
False-positive damage detection can be theoretically explained. The DLV methods
identify elements with small stress as candidates for damaged elements. Undamaged
elements are not necessarily excluded from the small stress elements. Bernal (2002) and
Gao (2005) reported that such false-positives are occasional in their simulations and
experiments. In the Imote2 experiments, Figure 8.15 (a) shows false-positive damage
detection at element 11 in the second measurement. This false-positive is not observed in
the first and third measurements. In Figure 8.14 (a), however, false-positives are observed
consecutively. Element 8 has small stress in all cases. Though consecutive false-positives
do not contradict the theoretical derivation of the DLV methods, they are not reported in
Bernal (2002) or Gao (2005). Consecutive false-positives are likely to indicate that factors
such as a structure’s nonlinearity and observation noise affect the damage detection
capability in these experiments.

150
 1 1

measurement 1
measurement 1
Normalized accumulated stress

Normalized accumulated stress

0. 5 0.5

0. 3 0.3

0 0
1 1

measurement 2
measurement 2

0.5 0. 5

0.3 0. 3

0 0
1 1

measurement 3
measurement 3

0.5 0.5

0.3 0.3

0 0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19 3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID Element ID
(a) (b)
Figure 8.14. Damage detection (damaged element: Element 9): (a) mass perturbation
DLV; and (b) SDLV.

1 1

measurement 1
measurement 1

Normalized accumulated stress

Normalized accumulated stress

0. 5 0.5
0. 3 0.3

0 0
1 1
measurement 2

measurement 2

0. 5 0.5
0. 3 0.3

0 0
1 1
measurement 3

measurement 3

0. 5 0. 5
0. 3 0. 3

0 0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19 3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID Element ID
(a) (b)

Figure 8.15. Damage detection (damaged element: Element 10): (a) mass perturbation
DLV; and (b) SDLV.

From a theoretical perspective, false-negatives, on the other hand, are not expected if
conditions such as linearity of the structure and no observation noise are satisfied. No
false-negative damage detection is found in Bernal (2002) and Gao (2005). In the Imote2
experiments, false-negatives are observed. For example, Figure 8.16 (a) shows the
damage detection results when element 11 is replaced with the damaged element. Seven
false-negative damage detection cases are observed in this figure. The SHM system
cannot detect damage in this case, even though the three sensor communities commonly
monitor the replaced element. Gao (2005) stated that detecting damage in the vertical and
diagonal elements located close to the midspan of the truss structure was found to be more
difficult than for other elements. Damage detection of elements carrying small force under
self-weight is considered challenging. Small forces in vertical element 11 under self-

151
 1 1

measurement 1

measurement 1
Normalized accumulated stress
0. 5 0. 5

Normalized accumulated stress

0. 3 0. 3

0 0
measurement 2 1 1

measurement 2
0. 5 0. 5

0. 3 0. 3

0 0
1 1
measurement 3

measurement 3
0. 5 0. 5

0. 3 0. 3

0 0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19 3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID Element ID

(a) (b)

Figure 8.16. Damage detection (damaged element: Element 11): (a) mass perturbation
DLV; and (b) SDLV.

1 1
measurement 1

measurement 1
Normalized accumulated stress

Normalized accumulated stress

0.5 0.5
0.3 0.3

0 0
1 1
measurement 2

measurement 2

0. 5 0.5
0. 3 0.3

0 0
1 1
measurement 3

measurement 3

0.5 0. 5
0.3 0. 3

0 0
3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19 3 4 5 6 7 8 9 10 11 7 8 9 10 11 12 13 14 15 11 12 13 14 15 16 17 18 19
Element ID Element ID
(a) (b)
Figure 8.17. Damage detection (damaged element: Element 12): (a) mass perturbation
DLV; and (b) SDLV.

weight may explain the observed false-negatives. Nonlinearity of the structure and
observation noise are also considered possible causes of the false-negatives.
One of the two DLV methods is not always better than the other in terms of damage
detection capability. For example, damage detection on a vertical element, Element 11, in
Figure 8.16 indicates the SDLV method performs better than the mass perturbation DLV
method. On the other hand, damage on another vertical element, Element 19, in Figure
8.18 can be detected more reliably by the mass perturbation DLV method.
When the number of false-positives and false-negatives is counted, the SDLV method
seems to experience fewer false damage detections. The number of false-positives and
false-negatives is summarized in Tables 8.4. to 8.7. For example, elements 7, 8, 9, and 10

152
 1 1

measurement 1

measurement 1
Normalized accumulated stress
0. 5 0. 5

Normalized accumulated stress

0. 3 0. 3

0 0
1 1
measurement 2

measurement 2
0. 5 0.5
0. 3 0.3

0 0
1 1
measurement 3

measurement 3
0. 5 0.5
0. 3 0.3

0 0
11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27 11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27
Element ID Element ID
(a) (b)

Figure 8.18. Damage detection (damaged element: Element 19): (a) mass perturbation
DLV; and (b) SDLV.

1 1
measurement 1

measurement 1
Normalized accumulated stress
Normalized accumulated stress

0. 5 0. 5
0. 3 0. 3

0 0
1 1
measurement 2

measurement 2

0. 5 0. 5
0. 3 0. 3

0 0
1 1
measurement 3

measurement 3

0. 5 0. 5
0. 3 0. 3

0 0
11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27 11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27
Element ID Element ID
(a) (b)
Figure 8.19. Damage detection (damaged element: Element 20): (a) mass perturbation
DLV; and (b) SDLV.

are commonly monitored by sensor communities 1 and 2, resulting in two damage

detection cases per measurement. Repeating the measurement process three times results
in six damage detection cases for each of these elements. When no false damage detection
cases are observed out of six, the corresponding cells in these tables appear as “0/6.” The
numerator represents the number of false damage detection cases, while the denominator
represents the total number of damage detection cases. Elements 3, 4, 5, and 6 are
monitored by only one community; thus, the denominator of the corresponding cells is
three instead of six. Element 11 is monitored by three communities, so the denominator is
nine. When element 8 is damaged, for example, the cell corresponding to this element
represents false-negative damage detection while other cells in the same column represent

153
 1 1

measurement 1
measurement 1
Normalized accumulated stress
0. 5 0.5

Normalized accumulated stress

0. 3 0.3

0 0
1 1

measurement 2
measurement 2

0.5 0. 5

0.3 0. 3

0 0
1 1

measurement 3
measurement 3

0.5 0. 5

0.3 0. 3

0 0
11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27 11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27
Element ID Element ID

(a) (b)

Figure 8.20. Damage detection (damaged element: Element 21): (a) mass perturbation
DLV; and (b) SDLV.

1 1

measurement 1
measurement 1
Normalized accumulated stress

Normalized accumulated stress

0. 5 0. 5

0. 3 0. 3

0 0
1 1
measurement 2
measurement 2

0. 5 0. 5

0. 3 0. 3

0 0
1 1
measurement 3
measurement 3

0. 5 0. 5

0. 3 0. 3

0 0
11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27 11 12 13 14 15 16 17 18 19 15 16 17 18 19 20 21 22 23 19 20 21 22 23 24 25 26 27
Element ID Element ID
(a) (b)
Figure 8.21. Damage detection (damaged element: Element 22): (a) mass perturbation
DLV; and (b) SDLV.

false-positives. The cells corresponding to false-negatives are shaded to distinguish false-

negatives from false-positives. The number of false damage detection using the mass
perturbation DLV method is, overall, larger than the number for the SDLV method,
though the sample size is small.
Tables 8.4 to 8.7 also suggest a way for robust damage detection. The summary of
damage detection using the SDLV method in Tables 8.6 and 8.7 shows that the percentage
of false damage detection cases is always smaller than or equal to 33 percent, while the
percentage for the mass perturbation DLV method can be larger than 50 percent. From this
observation, reliable damage localization might be possible by repeating damage detection

154
with the SDLV method and by adopting the majority of damage detection results on each
element. This conjecture needs to be confirmed with a larger number of samples.

Table 8.4. Summary of False Damage Detection Using the Mass Perturbation DLV
Method (Damaged Elements 8-12)

Damaged
8 9 10 11 12
Element
False False False False False
Damage Damage Damage Damage Damage
% % % % %
Detection Detection Detection Detection Detection
/Total /Total /Total /Total /Total
3 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
4 0/3 0 1/3 33 0/3 0 0/3 0 0/3 0
5 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
6 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
7 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
8 1/6 17 3/6 50 0/6 0 2/6 33 3/6 50
9 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
10 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
11 0/9 0 2/9 22 1/9 11 7/9 78 1/9 11
12 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
13 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
14 0/6 0 2/6 33 0/6 0 0/6 0 0/6 0
15 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
16 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
17 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
18 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
19 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0

8.9 Summary
In this chapter, Imote2s were installed on a truss model, and damage to the truss
simulated by replacing an element with one of reduced cross section was identified by the
proposed SHM framework. Required time and power consumption, as well as fault
tolerance, were investigated. While some false-positive/negative damage detection cases
were observed, damaged elements were identified in most cases. The proposed SHM
framework employing DCS for SHM has, thus, been experimentally verified.

155
Table 8.5. Summary of False Damage Detection Using the Mass Perturbation DLV
Method (Damaged Elements 19-22)

Damaged
19 20 21 22
Element
False False False False
Damage Damage Damage Damage
% % % %
Detection Detection Detection Detection
/Total /Total /Total /Total
11 0/3 0 0/3 0 0/3 0 0/3 0
12 0/3 0 0/3 0 0/3 0 0/3 0
13 0/3 0 0/3 0 0/3 0 0/3 0
14 0/3 0 0/3 0 0/3 0 0/3 0
15 0/6 0 0/6 0 0/6 0 0/6 0
16 1/6 17 2/6 33 1/6 16 0/6 0
17 0/6 0 0/6 0 0/6 0 0/6 0
18 0/6 0 0/6 0 0/6 0 0/6 0
19 0/9 0 0/9 0 0/9 0 0/9 0
20 0/6 0 0/6 0 0/6 0 0/6 0
21 0/6 0 0/6 0 0/6 0 0/6 0
22 0/6 0 0/6 0 3/6 50 0/6 0
23 0/6 0 0/6 0 0/6 0 0/6 0
24 0/3 0 1/3 33 0/3 0 0/3 0
25 0/3 0 0/3 0 0/3 0 0/3 0
26 0/3 0 0/3 0 0/3 0 0/3 0
27 0/3 0 0/3 0 0/3 0 0/3 0

156
Table 8.6. Summary of False Damage Detection Using the SDLV Method (Damaged
Elements 8-12)

Damaged
8 9 10 11 12
Element
False False False False False
Damage Damage Damage Damage Damage
% % % % %
Detection Detection Detection Detection Detection
/Total /Total /Total /Total /Total
3 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
4 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
5 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
6 0/3 0 0/3 0 0/3 0 0/3 0 1/3 33
7 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
8 0/6 0 1/6 16 1/6 17 2/6 33 2/6 33
9 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
10 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
11 0/9 0 0/9 0 0/9 0 2/9 22 0/9 0
12 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
13 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
14 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
15 0/6 0 0/6 0 0/6 0 0/6 0 0/6 0
16 0/3 0 1/3 33 0/3 0 0/3 0 0/3 0
17 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
18 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0
19 0/3 0 0/3 0 0/3 0 0/3 0 0/3 0

157
Table 8.7. Summary of False Damage Detection Using the SDLV Method (Damaged
Elements 19-22)

Damaged
19 20 21 22
Element
False False False False
Damage Damage Damage Damage
% % % %
Detection Detection Detection Detection
/Total /Total /Total /Total
11 0/3 0 0/3 0 0/3 0 0/3 0
12 0/3 0 0/3 0 0/3 0 0/3 0
13 0/3 0 0/3 0 0/3 0 0/3 0
14 0/3 0 0/3 0 0/3 0 0/3 0
15 0/6 0 0/6 0 0/6 0 0/6 0
16 0/6 0 0/6 0 1/6 16 0/6 0
17 0/6 0 0/6 0 0/6 0 0/6 0
18 0/6 0 0/6 0 0/6 0 0/6 0
19 2/9 22 2/9 22 0/9 0 0/9 0
20 0/6 0 0/6 0 0/6 0 0/6 0
21 0/6 0 0/6 0 0/6 0 0/6 0
22 0/6 0 0/6 0 0/6 0 0/6 0
23 0/6 0 0/6 0 0/6 0 0/6 0
24 0/3 0 0/3 0 0/3 0 0/3 0
25 0/3 0 0/3 0 0/3 0 0/3 0
26 0/3 0 0/3 0 0/3 0 0/3 0
27 0/3 0 0/3 0 0/3 0 0/3 0

158
 Chapter 9

CONCLUSIONS AND FUTURE STUDIES

9.1 Conclusions
The research detailed in this report has established a framework for structural health
monitoring on a network of smart sensors. This framework has been experimentally
verified on networks of Imote2s, resulting in the realization of the first hierarchical smart
sensor network for structural health monitoring. The structural health monitoring (SHM)
system has essential features, such as scalability to a large number of smart sensors,
promising damage detection capability, and autonomous operation. The software
developed under this research effort is open-source and is available at: http://
shm.cs.uiuc.edu/.
Background for this research was first provided. Smart sensors with computational
and communication capabilities have been developed for various applications. These
capabilities have been considered to offer new opportunities for the monitoring of
structures. The inexpensive nature of smart sensors supports densely instrumenting
structures; dense arrays of sensors have the potential to provide structural information at a
level of detail never before available. However, smart sensor usage in SHM applications
has encountered a number of difficulties. Many of these difficulties emanate from the lack
of adequate resources on smart sensors. From a hardware perspective, smart sensors are
usually battery-powered, and have limited RAM and relatively slow communication
speed. Middleware services developed for such hardware are not necessarily suitable for
SHM applications. Smart sensors have intrinsic synchronization error, and
communication among sensors can be unreliable and/or slow. A well-developed smart
sensor platform that can be directly used for SHM applications has only recently been
reported. The Imote2 smart sensor platform will soon be released for resource demanding
applications such as SHM of civil infrastructure. The substantially richer hardware
resources on the Imote2, as compared with other smart sensors, better suits SHM
applications. However, the Imote2 still misses some middleware services for realization of
smart sensor SHM systems. Another type of difficulty regards algorithmic issues. SHM
algorithms previously deployed on smart sensors have been based on either centralized
data acquisition or independent data processing; obtaining both scalability and effective
damage detection capabilities has been difficult. The Distributed Computing Strategy
(DCS) for SHM was proposed by Gao (2005) as a promising SHM algorithm that can
benefit from a large number of sensors. Data obtained at densely instrumented smart
sensors are processed in local sensor communities in a coordinated and distributed
manner. This SHM strategy, however, was previously demonstrated only on a PC with
numerical simulation data or data from wired sensors. This research first studied and
provided basic functionality necessary to implement the DCS on the Imote2s smart sensor
platform.

159
be scalability, autonomous distributed computing, fault tolerance, etc. A system
architecture that potentially provides these desirable characteristics was proposed as a
homogeneous network of smart sensors consisting of Imote2s which run the DCS for
SHM. Realization of this system was explained in the subsequent chapters.
The availability of appropriate sensors is essential for SHM applications. Sensor
board customizability was demonstrated in Chapter 4. Strain is one of important physical
quantities utilized in SHM applications. While accelerometers are often available on smart
sensors, and their suitability for SHM applications has been studied, a smart sensor with
strain sensors is rare. The flexibility of the smart sensor platforms was shown through the
development of a strain sensor board equipped with a Wheatstone bridge circuit for the
Berkeley Mote platform. The strain sensor board is an analog circuit, needing an
antialiasing filter, which was also developed. Experimental verification of these sensor
boards demonstrated the customizability of smart sensor boards.
Middleware services frequently needed in SHM applications were studied and
realized on the Imote2. The amount of data required for SHM applications is usually so
large that centrally collecting all of the data is impractical, if not impossible. A model-
based data aggregation service was developed to estimate required correlation functions in
a distributed manner. Data transfer requirements were greatly reduced. Another important
middleware service developed was reliable communication. Lost communication packets
carrying data can degrade signals in a similar way as observation noise does. The loss of a
packet carrying a command may leave an SHM system in an unknown state. Reliable
communication is, therefore, essential. Reliable communication protocols suitable for
long data records and a short comments were each developed for both unicast and
multicast applications. Synchronized sensing is also a crucial middleware service.
Unsynchronized signals distort modal identification results, especially the phase of mode
shapes. Even when clocks on smart sensors are synchronized, synchronized sensing is not
necessarily guaranteed. When sensing commences cannot be easily controlled. To achieve
synchronized sensing, clocks on the Imote2s were first synchronized, and then signals
were acquired with time stamps; the signals were subsequently resampled based on the
time stamps. In this way, signals were synchronized with each other with high precision.
The developed middleware services allow implementation of DCS for SHM, as well as a
wide variety of SHM applications, on Imote2s.
SHM algorithms implemented on Imote2s are based on DCS for SHM. In a local
sensor community, NExT and ERA estimates the modal properties of a structure from
acceleration responses. The mass perturbation DLV method localizes damaged element
using modal properties identified before and after damage. The damage localization
results are exchanged among neighboring local sensor communities to confirm the
consistency of localization results in overlapping parts of sensor communities. If not,
measurement and damage localization are repeated. This strategy was extended by
replacing the mass perturbation DLV method with the SDLV method, which eliminates
the need for mass normalization constant estimation.
Using the middleware services and algorithms, a framework for SHM using smart
sensors was realized on the Imote2 platform. Numerical functions used in the DCS
algorithms were first ported to Imote2s and their numerical accuracy and memory size

160
requirements examined. Second, the sensing capability of the Imote2 was compared with
conventional wired accelerometers. Third, using the ported numerical functions, the DCS
algorithms were implemented. These algorithms were examined by injecting data from
wired sensors into the memory of the Imote2s and processing the data both on the Imote2s
and on a PC. Data processing on the Imote2s and the PC was shown to be numerically
identical. An autonomous capability was realized in the system by assigning tasks to be
executed to a one-byte variable and managing this variable. In this way, DCS for SHM
was realized on Imote2s, and the validity of the implementation was examined component
by component.
Finally, the smart sensor SHM system was experimentally verified. Ten Imote2s were
installed on the three-dimensional truss structure, forming three local sensor communities
with overlap. The outputs of each step of DCS for SHM were sent to the base station and
compared with the corresponding outputs processed on a PC. The Imote2 networks
performed modal analysis and damage localization as designed, and successfully localized
simulated structural damage. Calculation and communication, as well as battery life of the
SHM system, were then estimated. The results showed that the system does not need
excessive time to perform data acquisition and processing and does not consume large
amounts of power. However, the battery life needs to be improved if Imote2s are to be
used for year-round monitoring. The damage detection capability was next examined for
various damage cases. False damage detection was observed in some cases. These false-
positives/negatives need to be studied and reduced/eliminated. Nonetheless, in most
experiments, the SHM system was able to localize damage, demonstrating its damage
detection capability.
This research provided the first realization of a hierarchical SHM system for smart
sensors that is scalable to networks of densely distributed smart sensors. The system has
good damage detection capability and can be operated autonomously. Battery life of the
system is moderate. The next section describes suggested further study to extend the
developed frame to be applicable to SHM of full-scale structures using a dense array of
smart sensors.

9.2 Future studies

9.2.1 Sensing capability

As examined in section 7.2.1, acceleration signals from the Imote2s were shown to
deviate from those measured by conventional accelerometers, even after the resampling
process. For measured signals to be meaningful, the signals need to be reliable. The
difference between the Imote2 signals and the reference sensor signals requires that the
reliability of Imote2 sensors be inspected more in detail for future usage. Each of the
components involved in sensing, such as accelerometer, time stamping, resampling, etc.,
needs to be examined carefully for the Imote2 to measure acceleration reliably.
For SHM of full-scale structures, the sensing capability in the low-frequency range is
especially important, though this capability was not extensively studied in this research.

161
The lowest frequencies of interest for the truss structure used in the experiments were
around 20 Hz, and the Imote2 sensing capability was shown to be sufficient for the DCS
around these natural frequencies. However, natural frequencies of full-scale buildings and
bridges are usually below 10 Hz. Three-dimensional sensing characteristics need to be
examined in the low frequency range.
Future smart sensors used for ambient vibration measurement need to have better
resolution. The least significant bit of the digital output of the Imote2 accelerometer is
about 1 mg. Therefore, the resolution is at best 1 mg. While this resolution is sufficient for
this research, ambient vibration measurement of civil infrastructure needs better
resolution. An accelerometer chip with a low noise level and higher resolution is needed
for ambient vibration measurement; if the accelerometer does not have digital outputs, the
ADC also needs to have high resolution.
While smart sensors are often equipped with accelerometers, other types of sensors
often utilized in civil engineering applications need to be developed for the Imote2.
Though a strain sensor board for the Berkeley Mote has been developed, the Imote2 does
not have a strain sensor board. A strain sensor board for the Imote2 will be beneficial to
civil engineers.
Another problem is failure in sensing. Under the current design of the system, all of
the sensors repeat sensing if one or more sensors fail to start sensing. The system can be
modified so that sensing failure of a node makes only the sensors in the same community
retake data. Even though the expected number of repetitions can be reduced in this way,
this repetition should preferably be avoided by decreasing the chances of sensing failure.

9.2.2 Damage detection capability

The damage detection algorithms sometimes fail to detect damage. Repeated false
damage detection was observed, though previous work on the DLV method did not report
such false detection. Because the DCS for SHM on simulated numerical truss responses
does not show false-negatives or repeated false-positive damage detection, such false
damage detection in the Imote2 experiments is considered to be due primarily to causes
specific to smart sensors. One possible cause is observation noise from imperfect
measurements. As stated before, sensing hardware for the Imote2 still has room for
improvement. Inaccurate sensing results in inaccurate damage detection. Another possible
cause is a discrepancy between the physical truss and the model assumed in DCS for
SHM. NExT, ERA, and the DLV methods all assume a linear model. If linearity is not
satisfied, damage detection may become faulty. Errors in the numerical truss model used
in the stress analysis possibly introduced inaccuracy in the damage detection, too. Before
the application of the SHM system to the full-scale structures, the conditions under which
damage detection becomes faulty need to be clarified, and the reliability needs to be
improved.
The influence of other factors, such as temperature and humidity, should be
accommodated in the damage detection strategy. For example, natural frequencies of a
structure change along with temperature. The effects of temperature and humidity changes
on the DLV methods have not yet been studied.

162
 Integration of different types of information obtained from smart sensors has the
possibility of improving damage detection capabilities. For example, smart sensors
measuring acceleration, strain, wind velocity, temperature, and humidity may use four
types of sensors to detect damage more reliably than smart sensors employing only
accelerometers. Each measurand has its own characteristics. Acceleration responses
reflect global motion of a structure and have particularly rich information at higher
frequencies; on the other hand, strain is a local physical quantity, with its information
concentrated at lower frequencies. Wind velocity affects the input force, damping, and
stiffness. Temperature and humidity change at slower rate, affecting structural
characteristics. Integration of different types of information may result in more reliable
damage detection.
Furthermore, the applicability of the algorithms to nontruss structures needs to be
examined and experimentally verified. The applicability has been studied mainly for truss
structures. For DCS for SHM to be more widely used in monitoring civil infrastructure,
the applicability to a wide range of structure types should be demonstrated.

9.2.3 Power harvesting

The three AAA batteries powering the Imote2 have a limited life. For some
applications needing only a few samples taken infrequently, the batteries may last for
months or even years. However, data-intensive applications such as SHM consume
significant power, shortening the battery life. A battery life of years using only a few AAA
batteries is not likely to be achieved in the near future. Using larger batteries is one
practical solution when a small form factor is not required. The life of a smart sensor can
be lengthened by increasing the battery capacity. Alternatively, power harvesting at the
smart sensor nodes is a promising approach to achieve semipermanent monitoring of
structures using nonplugged-in smart sensors.
Several energy sources can be identified for development of SHM power harvesting
strategies. Wind energy is a potential energy source. Bridges are most likely constructed
over a river, street, or railroad tracks where obstacles to block the wind are limited. Wind
velocities at bridge cites are expected to be relatively high, making power harvesting with
wind energy promising. Solar power is another candidate, though available energy
depends on the climate at the site and sunlight availability of sensor locations. As opposed
to wind power, solar power harvesting does not have any moving parts; therefore,
vibration originating from solar power harvesting will not contaminate measurement of
structural vibration signals. Another energy source, structural vibration energy, can
conceptually be converted to electrical energy. Considering that vibration of civil
infrastructure is normally in the low frequency range, power harvesting from such
vibration might be difficult. Devices to capture small sources of energy and condition
stored energy for smart sensor use have also been reported (Advanced Linear Devices,
Inc., 2007). Further study is required to use these energy sources reliably for smart
sensors.
Even when power harvesting is realized, power consumption on the smart sensor
should be kept moderate. The amount of energy available from wind, solar power, or

163
structural vibration is considered relatively small. Providing ample power to the smart
sensors from these energy sources is not likely to happen soon. Power consumption at
smart sensors, therefore, needs to be well-managed.

9.2.4 Power management and scheduling

Although the realized SHM system manages power consumption, the management
still has room for improvement. The system does change the CPU speed depending on the
tasks to be performed, and the system does prepare to enter a sleep mode when all of the
tasks are performed. However, the change in CPU speed is only between two of four
available speeds, and the Imote2 does not actually enter a sleep mode. By addressing these
problems, power consumption can be further reduced.
In the future, DCS for SHM can be performed periodically. For example, the
measurement and data processing can take place once every few days. While not active,
the Imote2 can be in a sleep mode, only checking its clock. When the monitoring time
approaches, the Imote2 can wake up and perform the necessary monitoring.

9.2.5 Monitoring of occasional events such as earthquakes

For monitoring of occasional events such as earthquakes, a scheme to wake up

networks of smart sensors in a sufficiently short time is needed. Assigning one sensor
node to monitor possible earthquakes continuously and to disseminate wake-up signals to
all of the other nodes is one approach. However, the radio is turned off to save power in
the sleep mode. Therefore, waking up all of the smart sensors by a wake-up command sent
through the radio is not straightforward. Periodically turning on the radio to listen to a
wake-up preamble is one possible solution, though such periodic usage of the radio
consumes power. The interval of the periodic radio listening can be shortened to improve
the response time in system wake-up at the expense of power consumption. If the start of
an earthquake can be predicted, for example by measuring primary waves of an
earthquake near the epicenter, the response time need not be so short, allowing relatively
long intervals between the periodic radio listening. Another approach is to equip all of the
sensor nodes with a mechanism to detect large motion and trigger sensing. A mechanism
to detect large motion mechanically without power is preferable. Thus, system wake-up at
the beginning of a rare event is a key issue.

9.2.6 Multihop

SHM systems to monitor full-scale structures may need multihop communication,

while the SHM system realized in this research is based on single-hop communication. A
sensor community for full-scale structures can be physically larger than the
communication range for a smart sensor. Multihop communication is necessary in such
cases. The unicast and multicast protocols need to be extended for multihop
communication.

164
9.2.7 Communication range adjustment

When the communication range of a smart sensor includes neighboring sensor

communities, the communication range should be shortened to reduce RF interference.
When two sensor communities are out of communication range of each other, these
communities can perform communication in parallel without interfering with each other.
Communication range adjustment results in faster execution of DCS and smaller power
consumption.

9.2.8 Reliability of the system

For the system to be used for years, the reliability of smart sensor nodes needs to be
improved. In the current system, the Imote2s sometimes hangs up. The possibility of
hang-ups should be reduced.
Additionally, node failure tolerance should preferably be realized, because the
possibility of node failure cannot be completely eliminated; for example, even a PC with
better resources sometimes hangs up. Failure of a single node should not provoke failure
of the entire system.
Even when all of the nodes seem to be working, sensor signals may be erroneous. For
example, sensors may not be firmly attached to a structure. Such Byzantine-type error can
possibly be detected by checking the linearity between measured signals. When the
linearity between two signals is much lower than that of other pairs, one of the two signals
may come from a sensor which is not adequately installed. Implementation of such a
Byzantine error detection algorithm can improve the reliability of the damage detection
system.

9.2.9 Environmental hardening

The Imote2 needs to be packaged before the node can be installed on full-scale
structures, especially when the Imote2 is used outside. The package needs to protect the
Imote2 from weather, birds, insects, etc. Attention needs to be paid to not blocking radio
reception/transmission. Also, the packaging should not absorb or amplify vibration.

9.2.10 Multiple purpose usage of smart sensors

Smart sensor networks can potentially serve multiple purposes, while the developed
SHM framework currently serves only for SHM purposes. Smart sensors can be used for
traffic monitoring, local weather monitoring, fire detection, etc. Integration of these
services enhances the value of smart sensors and makes introduction of smart sensor
systems more attractive from a cost-benefit perspective.

165
 REFERENCES
Actis, R. L. & Dimarogonas, A. D. (1989). “Non-linear effects due to closing cracks in vi-
brating beams.” ASME Design Engineering Division Publication DE-Structural Vi-
bration and Acoustics, 18(3), 99-104.
Adler, R., Flanigan, M., Huang, J., Kling, R., Kushalnagar, N., Nachman, L., Wan, C-Y.,
& Yarvis, M. (2005). “Intel mote 2: an advanced platform for demanding sensor net-
work applications.” Proceedings of 3rd Int. Conference on Embedded Networked Sen-
sor Systems, 298-298.
Advanced Linear Devices, Inc. <http://www.aldinc.com> (Mar. 2, 2007).
Agardh, L. (1991). “Modal analysis of two concrete bridges.” Structural Engineering In-
ternational, 1(4), 35-39.
Agre, J. R., Clare, L. P., Pottie, G. J., & Romanov, N. P. (1999). “Development platform
for self-organizing wireless sensor networks.” Proceedings of SPIE - Conference on
Unattended Ground Sensor Technologies and Applications, Vol. 3713, 257-268.
Aktan, A. E., Lee, K. L., Chuntavan, C., & Aksel, T. (1994). “Modal testing for structural
identification and condition assessment of constructed facilities.” Proceedings of In-
ternational Modal Analysis Conference, Honolulu, HI, (1), 464-468.
Alampalli, S., Fu, G., & Aziz, I. A. (1992). “Modal analysis as a bridge inspection tool.”
Proceedings of 10th International Modal Analysis Conference, San Diego, CA, 2,
1359-1366.
Al-Najjar, B. (2000). “Accuracy, effectiveness and improvement of vibration-based main-
tenance in paper mills; case studies.” Journal of Sound and Vibration, 229(2), 389-
410.
Analog Devices, Inc. (2007). “AD623 - single supply, rail-rail, low cost instrumentation
amplifiers.” Datasheet, <http://www.analog.com/UploadedFiles/Data_Sheets/
516895375AD623_c.pdf> (Mar. 2, 2007).
Anderson, E., Bai, Z., Bischof, C., Blackford, S., Demmel, J., Dongarra, J., Du Croz, J.,
Greenbaum, A., Hammarling, S., McKenney, A., & Sorensen, D. (1999). LAPACK
User’s Guide, Philadelphia: Society for Industrial and Applied Mathematics.
Aoki, S., Fujino, Y., & Abe, M. (2003). “Intelligent bridge maintenance system using
MEMS and network technology.” Proceedings of SPIE - Smart Systems and NDE for
Civil Infrastructures, 5057, 37-42.
Arici, Y. & Mosalam, K. M. (2003). “Modal analysis of a densely instrumented building
using strong motion data.” Proceedings of Int. Conference on Applications of Statis-
tics & Probability in Civil Engineering, San Francisco, CA., 419-426.
Arms, S. P., Townsend, C. P., Churchhill, D. L., Hamel, M. J., Galbreath, J. H., & Mun-
dell, S. W. (2004). “Frequency agile wireless sensor networks.” Proceedings of SPIE -
Smart Electronics, MEMS, BioMEMS, and Nanotechnology, 5389, 468-475.

166
Basheer, M. R., Rao, V., & Derriso, M. (2003). “Self organizing wireless sensor networks
for structural health monitoring.” Proceedings of the 4th Int. Workshop on Structural
Health Monitoring, Stanford, CA, 1193-1206.
Becker, K. C., Byington, C. S., Forbes, N. A., & Nickerson G. W. (1998). “Predicting and
preventing machine failures, The Industrial Physicist, American Institute of Physics,
20-23.
Begg, R. D., Mackenzie, A. C., Dodds, C. J., & Loland, O. (1976). “Structural integrity
monitoring using digital processing of vibration signals.” Proceedings of 8th Annual
Offshore Technology Conference, Houston, TX, 305-311.
Bendat, J. S., & Piersol, A. G. (2000). Random data: analysis and measurement proce-
dures, New York: Wiley.
Bernal, D. (2002). “Load vectors for damage localization.” Journal of Engineering Me-
chanics, 128(1), 7-14.
Bernal, D. (2004). “Modal scaling from known added masses.” Journal of Engineering
Mechanics, 130(9), 1083-1088.
Bernal, D. (2006). “Flexibility-based damage localization from stochastic realization re-
sults.” Journal of Engineering Mechanics, 132(6), 651-658.
Bernal, D. & Gunes, B. (2004). “Flexibility based approach for damage characterization:
benchmark application.” Journal of Engineering. Mechanics, 130(1), 61-70.
Bhatti, S. Carlson, J., Dai, H., Deng, J., Rose, J., Sheth, A., Shucker, B., Gruenwald, C.,
Torgerson, A., & Han, R. (2005). “MANTIS OS: an embedded multithreaded operat-
ing system for wireless micro sensor platforms.” Mobile Networks and Applications,
10(4), 563-579.
Bishop, C. M. (1994). “Neural networks and their applications.” Review of Scientific In-
struments, 65, 1803-1832.
Biswas, M., Pandey, A. K. & Samman, M. M. (1990). “Diagnostic Experimental Spectral/
Modal Analysis of a Highway Bridge.” Int. Journal of Analytical and Experimental
Modal Analysis, 5, 33-42.
Brincker, R., Zhang, L., & Andersen, P. (2001). “Modal identification of output-only sys-
tems using frequency domain decomposition.” Smart Materials and Structures, 10,
441-445.
Bult, K., Burstein, A., Chang, D., Dong, M., Fielding, M., Kruglick, E., Ho, J., Lin, F. Lin,
T. H., Kaiser, W. J., Mukai, R., Nelson, P., Newburg, F. L., Pister, K. S. J., Pottie, G.,
Sanchez, H., Stafsudd, O. M., Tan, K. B., Ward, C. M., Yung, G., Xue, S., Marcy, H.
& Yao, J. (1996). “Wireless integrated microsensors.” Technical Digest of the 1996
Solid State Sensor and Actuator Workshop, Transducers Research Foundation, Hilton
Head Island, SC, 205-210.

167
Casciati, F., Faravelli, L, Borghetti, F., & Fornasari, A. (2003). “Future trends in infra-
structure monitoring.” Structural Health Monitoring and Intelligent Infrastructure, 2,
997-1002.
Casciati, F., Casciati, S., Faravelli, L., & Rossi, R. (2004). “Hybrid wireless sensor net-
work.” Proceedings of SPIE - Smart Structures and Materials: Sensors and Smart
Structures Technologies for Civil, Mechanical, and Aerospace Systems, Berkeley, CA,
5391, 308-313.
Celebi, M. (2002). “Seismic instrumentation of buildings (with emphasis on federal build-
ings)” Special GSA/USGS Project, an administrative report, United States Geological
Survey, Menlo Park, CA.
Chance, J., Tomlinson, G. R., & Worden, K. (1994). “A simplified approach to the numer-
ical and experimental modeling of the dynamics of a cracked beam.” Proceedings of
12th International Modal Analysis Conference, 778-785.
Chen, M., Glaser, S. D., & Oberheim, T. (2006). “Terra-Scope - A MEMS-based Vertical
Seismic Array.” Smart Structures & Systems, 2(2), 115-126.
Chintalapudi, K., Paek, J., Gnawali, O., Fu, T., Dantu, K., Caffrey, J., Govindan, R., &
Johnson, E. (2006). “Structural damage detection and localization using NetSHM.”
Proceedings of 5th Int. Conference on Information Processing in Sensor Networks.
Nashville, TN.
Chipcon, Inc. <http://www.chipcon.com> (Mar. 2, 2007).
Chowdhury, M. R. (1990). “Experimental modal testing and analysis of fracture damage
in structures by the modal frequency method.” Proceedings of 8th International Modal
Analysis Conference. Florida, (1), 109-114.
Chung, H.-C., Enomoto, T., Shinozuka, M., Chou, P., Park, C., Yokoi, I., & Morishita, S.
(2004). “Real time visualization of structural response with wireless MEMS sensors.”
Proceedings of 13th World Conference on Earthquake Engineering, No. 121, 1-10.
Clayton, E. H., & Spencer, B. F., Jr. (2001). “Development of an experimental model for
the study of infrastructure preservation.” Proceedings of the National Conf. on Under-
graduate Research (NCUR), Whitewater, WI.
Collacott, R. A. (1977) Mechanical fault diagnosis and condition monitoring, London:
Chapman and Hall.
Cooley, J. M. & Tukey, J. W. (1965). “An algorithm for the machine calculation of com-
plex Fourier series.” Mathematics of Computation, 19:297-301.
Crossbow Technology, Inc. <http://www.xbow.com> (Mar. 2, 2007).
Danielson, G. C. & Lanczos, C. (1942). “Some improvements in practical Fourier analysis
and their application to X-ray scattering from liquids.” Journal of the Franklin Insti-
tute, 233, 365-380 and 435-452.
DataGrid Project. <http://eu-datagrid.web.cern.ch/eu-datagrid/default.htm> (Mar. 2,
2007).

168
Doebling, S. W. (1995). Measurement of structural flexibility matrices for experiments
with incomplete reciprocity, (Doctoral dissertation, University of Colorado, 1995).
Doebling, S. W. & Farrar, C. R. (1999). The state of the art in structural identification of
constructed facilities, A report by American Society of Civil Engineers Committee on
Structural Identification of Constructed Facilities.
Doebling, S. W., Farrar, C. R., & Prime, M. B. (1998). "A summary review of vibration-
based damage identification methods." Shock Vib. Dig., 30(2), 91-105.
Doebling, S. W., Farrar, C. R., Prime, M. B., & Shevitz, D. W. (1996). “Damage identifi-
cation and health monitoring of structural and mechanical systems from changes in
their vibration characteristics: a literature review.” Los Alamos National Laboratory
Report, LA-13070-MS.
Doherty, L., Pister, K. S. J., & Ghaoui. L. E. (2001). “Convex position estimation in wire-
less sensor networks.” Proceedings of the IEEE Infocom, AK., 1655-1663.
Dubin, E. E. & Yanev, B. S. (2001). “Managing the East river bridges in New York City.”
Proceedings of SPIE - 7th International Symposium on Smart Structures and Materi-
als, Newport, CA, 60-74.
Dust Networks. <http://www.dust-inc.com> (Mar. 2, 2007).
Electronics Services Shop. <http://www.ece.uiuc.edu/eshop> (Mar. 2, 2007).
Elson, J., Girod, L., & Estrin, D. (2002). “Fine-grained network time synchronization us-
ing reference broadcasts.” Proceedings of 5th Symposium on Operating Systems De-
sign and Implementation, Boston, MA.
Ember Corporation. <http://www.ember.com> (Mar. 2, 2007).
EYES. <http://www.eyes.eu.org> (Mar. 2, 2007).
Farrar, C. R. (2001). “Historical overview of structural health monitoring.” Lecture Notes
on Structural Health Monitoring using Statistical Pattern Recognition. Los Alamos
Dynamics, NM.
Farrar, C. R., Allen, D. W., Ball, S., Masquelier, M. P., & Park, G. (2005). “Coupling
sensing hardware with data interrogation software for structural health monitoring.”
Proceedings of 6th International Symposium on Dynamic Problems of Mechanics,
Ouro Preto, Brazil.
Farrar, C. R., Baker, W. E., Bell, T. M., Cone, E. M., Darling, T. W., Duffey, T. A., Ek-
lund, A., & Migliori, A. (1994). “Dynamic characterization and damage detection in
the I-40 bridge over the Rio Grande.” Los Alamos National Laboratory Report LA-
12767-MS.
Farrar, C. R., Sohn, H., Hemez, F. M., Anderson, M. C., Bement, M. T., Cornwell, P. J.,
Doebling, S. W., Lieven, N., Robertson, A. N., & Schultze, J. F. (2003). “Damage
prognosis: current status and future needs.” Los Alamos National Laboratory Report
LA-14051-MS.

169
Feldman, M. & Braun, S. (1995). “Identification of nonlinear system parameters via the
instantaneous frequency: application of the Hilbert transform and Wigner-Wille tech-
niques.” Proceedings of 13th International Modal Analysis Conference, 637-642.
Fox, C. H. J. (1992). “The location of defects in structures: a comparison of the use of nat-
ural frequency and mode shape data.” Proceedings of 10th International Modal Analy-
sis Conference, 522-528.
Galbreath, J. H., Townsend, C. P., Mundell, S. W., Hamel, M. J., Esser, B., Huston, D., &
Arms, S. W. (2003). “Civil structure strain monitoring with power-efficient, high-
speed wireless sensor networks.” Proceedings of 4th Int. Workshop on Structural
Health Monitoring, 1215-1222.
Ganeriwal, S., Kumar, R., & Srivastava, M. B. (2003). “Timing-sync protocol for sensor
networks.” Proceedings of 1st International Conference On Embedded Networked
Sensor Systems, Los Angeles, CA., 138-149.
Gao, Y. (2005). Structural health monitoring strategies for smart sensor networks (Doctor-
al dissertation, University of Illinois at Urbana-Champaign, 2005).
Gao, Y., Spencer, B. F., Jr. & Ruiz-Sandoval, M. (2005). "Distributed computing algo-
rithm for structural health monitoring." Structural Control and Health Monitoring, in
review.
Gardner-Morse, M. G. & Huston, D. R. (1993). “Modal identification of cable stayed pe-
destrian bridge.” Journal of Structural Engineering, ASCE, 119(11), 3384-3404.
Gay, D., Levis, P. Behren, R., Welsh, M., Brewer, E. & Culler, D. (2003). “The nesC lan-
guage: a holistic approach to networked embedded systems.” Proceedings of Pro-
gramming Language Design and Implementation.
Girod, L., Elson, J., Cerpa, A., Stathopoulos, T., Ramanathan, N., & Estrin, D. (2004).
“EmStar: a software environment for developing and deploying wireless sensor net-
works.” Proceedings of USENIX ‘04, Boston, MA, 283-296.
Golden Gate Bridge, Highway and Transportation District. <http://goldengatebridge.org>
(Mar. 2, 2007).
Gros, X. E. (1997). NDT data fusion, London: Arnold.
He, J. & Ewins, D. J. (1986). “Analytical stiffness matrix correction using measured vi-
bration modes.” Int. Journal of Analytical and Experimental Modal Analysis, 1(3), 9-
14.
Hearn, G. & Testa, R. B. (1991). “Modal analysis for damage detection in structures.”
Journal of Structural Engineering, 117(10), 3042-3063.
Hermans, L. & Auweraer, H. V. (1999). “Modal testing and analysis of structures under
operational conditions: industrial applications.” Mechanical Systems and Signal Pro-
cessing, 13(2), 193-216.
Hill, J., & Culler, D. (2002). “Mica: A wireless platform for deeply embedded networks.”
IEEE Micro., 22(6), 12-24.

170
Hill, J., Szewczyk, R., Woo, A., Hollar, S., Culler, D., & Pister, K. (2000). “System archi-
tecture directions for networked sensors.” Proceedings of 9th International Confer-
ence Architectural Support for Programming Languages and Operating Systems,
Cambridge, MA, 93-104.
Hollar, S. (2000). "COTS Dust." Master's Thesis, University of California, Berkeley, CA.
Hou, T.-C., Lynch, J. P., & Parra-Montesinos, G. (2005). “In-situ wireless monitoring of
fiber reinforced cementitious composite bridge piers.” Proceedings of Int. Modal
Analysis Conference, Orlando, FL.
Ibrahim, S. R. & Mikulcik, E. C. (1977). "A method for the direct determination of vibra-
tion parameters from the free responses." Shock and Vibration Bulletin, 47(4), 183-
198.
Intel Corporation. <http://www.intel.com> (Mar. 2, 2007).
James, G. H., Carne, T. G., & Lauffer, J. P. (1993). "The natural excitation technique for
modal parameter extraction from operating wind turbine,"Report No. SAND92-1666,
UC-261, Sandia National Laboratories, NM.
James, G. H., Carne, T. G., Lauffer, J. P. & Nord, A. R. (1992). "Modal testing using nat-
ural excitation," Proceedings of 10th Int. Modal Analysis Conference, San Diego, CA.
Juang, J.-N. and Pappa, R. S. (1985). “An eigensystem realization algorithm for modal pa-
rameter identification and model reduction.” Journal of Guidance, Control, and Dy-
namics, 8(5):620-627.
Kaouk, M. & Zimmerman. D. C. (1994). ”Structural damage assessment using a general-
ized minimum rank perturbation theory.” AIAA Journal, 32(4), 836-842.
Kaouk, M. & Zimmerman, D. C. (1995). “Reducing the required number of modes for
structural damage assessment.” Proceedings of 36th AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics and Materials Conference, 2802-2812, AIAA-95-
1094-CP.
Karbhari, V. M., Guan, H. & Sikorsky, C. (2003). “Web-based structural health monitor-
ing of a FRP composite bridge.” Structural Health Monitoring and Intelligent Infra-
structure, 1, 217-225.
Kawahara, Y., Minami, M., Morikawa, H., & Aoyama, T. (2003). "Design and implemen-
tation of a sensor network node for ubiquitous computing environment." Proceedings
of IEEE Semiannual Vehicular Technology Conference, Orlando, FL., 5, 3005-3009.
Kiremidjian, A. S., Straser, E. G., Meng, T. H., Law, K. & Soon, H. (1997). "Structural
damage monitoring for civil structures." Proceedings of Int. Workshop on Structural
Health Monitoring, Stanford, CA, 371-382.
Kirianaki, N. V., Yurish, S. Y., & Shpak, N. O. (2000). “Smart sensors with frequency
output: state-of-the-art and future development.” Proceedings of IFAC Workshop on
Progrmmable Devices and Systems, Ostrava, Czech Republic, 37-42.

171
Kling, R. (2003). “Intel Mote: an enhanced sensor network node.” Proceedings of Int.
Workshop on Advanced Sensors, Structural Health Monitoring, and Smart Structures,
Tokyo, Japan.
Kling, R., Adler, R., Huang, J., Hummel, V., & Nachman, L. (2005). “Intel Mote-based
sensor networks.” Structural Control and Health Monitoring, 12, 469-479.
Ko, J. M., & Ni, Y. Q. (2005). “Technology developments in structural health monitoring
of large-scale bridges.” Engineering Structures, 27(12), 1715-1725.
Kottapalli, V. A., Kiremidjian, A. S., Lynch, J. P., Carryer, E., Kenny, T. W., Law, K. H.,
& Lei, Y. (2003). “Two-tiered wireless sensor network architecture for structural
health monitoring.” Proceedings of SPIE - The International Society for Optical Engi-
neering - Smart Structures and Materials, 5057, 8-19.
Krawczuk, M. & Ostachowicz, W. M. (1992). “Parametric vibrations of a beam with
crack.” Archive of Applied Mechanics, 62, 463-473.
Kurata, N., Spencer, B. F., Jr., & Ruiz-Sandoval, M. (2004). “Building risk monitoring us-
ing wireless sensor network.” Proceedings of 13th World Conference on Earthquake
Engineering, Vancouver, BC., Canada, No. 1406.
Kwon, Y., Mechitov, K. A., Sundresh, S., Kim, W., & Agha, G. A. (2005a). “Resilient lo-
calization for sensor networks in outdoor environments.” Proceedings of 25th IEEE
International Conference on Distributed Computing Systems, Colombus, OH, 643-
652.
Kwon, Y., Sundresh, S., Mechitov, K. A., & Agha, G. A. (2005b). "ActorNet: an actor
platform for wireless sensor networks." University of Illinois at Urbana-Champaign
Technical Report.
Law, Y., Dulman, S., Etalle, S., & Havinga, P. (2002). “Assessing security-critical ener-
gy-efficient sensor networks.” Technical Report TR-CTIT-02-18, Centre for Telemat-
ics and Information Technology, University of Twente, Netherlands.
Lieven, N. A. J. & Waters, T. P. (1994). “Error location using normalised orthogonality.”
Proceedings of 12th International Modal Analysis Conference, Honolulu, HI, 1, 761-
764.
Lin, C. S. (1990). “Location of modeling errors using modal test data.” AIAA Journal, 28,
1650-1654.
Lin, R. M. & Ewins, D. J. (1990). “On the location of structural nonlinearity from modal
testing - a feasibility study.” Proceedings of 8th International Modal Analysis Confer-
ence, 358-364.
Linear Technology. (2007). “LTC1682 - doubler charge pumps with low noise linear reg-
ulator.” Datasheet, <http://www.linear.com/pc/downloadDocument.do?nav-
Id=H0,C1,C1003,C1039,C1133,P1749,D1952> (Mar. 2, 2007).
Lynch, J. P. (2006). “Design of a wireless active sensing unit for localized structural
health monitoring.” Structural Control and Health Monitoring, in press.

172
Lynch, J. P., Law, K. H., Kiremidjian, A. S., Carryer, E., Kenny, T. W., Patridge, A., &
Sundararajan, A. (2002). “Validation of a wireless modular monitoring system for
structures.” Proceedings of SPIE - Smart Structures and Materials: Smart Systems for
Bridges, Structures, and Highways, San Diego, CA, 4696(2), 17-21.
Lynch, J. P. & Loh, K. (2006). “A summary review of wireless sensors and sensor net-
works for structural health monitoring.” Shock and Vibration Digest, in press.
Lynch, J. P., Sundararajan, A., Law, K. H., Kiremidjian, A. S., Carryer, E., Sohn, H., &
Farrar, C. R. (2003). “Field validation of a wireless structural health monitoring sys-
tem on the alamosa canyon bridge.” Proceedings of SPIE - Smart Structures and Ma-
terials: Smart Systems and Nondestructive Evaluation for Civil Infrastructures, 5057,
267-278.
Lynch, J. P., Wang, Y., Law, K. H., Yi, J.-H., Lee, C. G., & Yun, C. B. (2005). “Valida-
tion of a large-scale wireless structural monitoring system on the Geumdang bridge.”
Proceedings of the Int. Conference on Safety and Structural Reliability, Rome, Italy.
Maia, N. M. M., & Silva, J. M. M. (1997). Theoretical and experimental modal analysis.
Taunton Somerset, England: Research Studies Press.
Maroti, M. Kusy, B., Simon, G., & Ledeczi, A. (2004). “The flooding time synchroniza-
tion protocol.” Proceedings of 2nd International Conference On Embedded Net-
worked Sensor Systems, Baltimore, MD, 39-49.
Mason, A., Yazdi, N., Najafi, K., & Wise, K. (1995). “A low-power wireless microinstru-
mentation system for environmental monitoring.” Dig. 8th Int. Conference on Solid-
State Sensors and Actuators, Stockholm, Sweden, 107-110.
Maxim Integrated Products, Inc. (2007). “Single/dual/quad, wide-bandwidth, low-power,
single-supply, rail-to-rail i/o op amps.” Datasheet, <http://pdfserv.maxim-ic.com/en/
ds/MAX4130-MAX4134.pdf> (Mar. 2, 2007).
Mechitov, K. A., Kim, W., Agha, G. A., & Nagayama, T. (2004). “High-frequency dis-
tributed sensing for structure monitoring.” Proceedings of 1st Int. Workshop on Net-
worked Sensing Systems, Tokyo, Japan, 101–105.
MicroStrain, Inc. <http://www.microstrain.com> (Mar. 2, 2007).
Millennial Net. <http://www.millennial.net> (Mar. 2, 2007).
Miller, B. (2001). “Bluetooth and other wireless technologies.” Prentice Hall. <http://
www.phptr.com/articles/printerfriendly.asp?p=24265> (Mar. 2, 2007).
Moore, M., Phares, B., Graybeal, B., Rolander, D., & Washer, G. (2001). “Reliability of
visual inspection of highway bridges.” FHWA Report No. FHWA-RD-01-020, FHWA,
U.S. Dept. of Transportation, Washington, DC.
Morgan, B. J. & Osterle R. G. (1985). “On-site modal analysis - a new powerful inspec-
tion technique.” Proceedings of 2nd International Bridge Conference, Pittsburg, PA,
108-114.

173
Mufti, A. A. (2003). “Integration of sensing in civil structures: development of the new
discipline of civionics.” Structural Health Monitoring and Intelligent Infrastructure,
1, 119-129.
Nagayama, T., Abe, M., Fujino, Y., & Ikeda, K. (2005). “Structural identfication of a non-
proportionally damped system and its application to a full-scale suspension bridge.”
Journal of Structural Engineering, 131(10), 1536-1545.
Nagayama, T., Rice, J. A., & Spencer, B. F., Jr. (2006a). “Efficacy of Intel's Imote2 wire-
less sensor platform for structural health monitoring applications.” Proceedings of
Asia-Pacific Workshop on Structural Health Monitoring, Yokohama, Japan.
Nagayama, T., Ruiz-Sandoval, M., Spencer, B. F., Mechitov K. A., & Agha, G. A. (2004).
“Wireless strain sensor development for civil infrastructure.” Proceedings of 1st Int.
Workshop on Networked Sensing Systems, Tokyo, Japan, 97-100.
Nagayama, T. Sim, S-H., Miyamori, Y., & Spencer, B. F., Jr. (2007). “Issues in structural
health monitoring employing smart sensors.” Smart Structures and Systems (accept-
ed).
Nagayama, T., Spencer, B. F., Jr., Agha, G. A., & Mechitov, K. A. (2006b), “Model-based
data aggregation for structural monitoring employing smart sensors”, Proceedings of
3rd Int. Conference on Networked Sensing Systems (INSS 2006), Chicago, IL, 203-
210.
Nakamura, Y. (2004). “UrEDAS, urgent earthquake detection and alarm system, now and
future.” Proceedings of 13th World Conference on Earthquake Engineering, Vancou-
ver, B.C., Canada, Paper No. 908.
Nakamura, M., Masri, S. F., Chassiakos, A. G., & Caughey, T. K. (1998). “A method for
non-parametric damage detection through the use of neural networks.” Earthquake
Engineering and Structural Dynamics, 27, 997-1010.
Nashif, A. D., Jones, D. I., & Henderson, J. P. (1985). Vibration damping. New York:
Willey.
National Semiconductor. (1991). “A basic introduction to filtersdactive, passive, and
switched-capacitor.” National Semiconductor Application Note 779, <http://www.na-
tional.com/an/AN/AN-779.pdf> (Mar. 2, 2007).
Nitta, Y., Nagayam, T., Spencer, B. F. Jr., & Nishitani, A. (2005). “Rapid damage assess-
ment for the structures utilizing smart sensor MICA2 MOTE.” Proceedings of 5th Int.
Workshop on Structural Health Monitoring, Stanford, CA, 283-290.
Nwosu, D. I., Swamidas, A. S. J., Guigne, J. Y., & Olowokere, D. O. (1995). “Studies on
influence of cracks on the dynamic response of tubular T-joints for nondestructive
evaluation.” Proceedings of 13th International Modal Analysis Conference, 1122-
1128.
Oppenheim, A. V., Schafer, R. W., & Buck, J. R. (1999). Discrete-time signal processing.
Upper Saddle River, NJ: Prentice Hall.

174
Oshima, T., Rahman, M. S., Mikami, S., Yamazaki, T., Takada, N., Lesko, J. J., & Kriz,
R. D. (2000). “Application of smart materials and systems to long-term bridge health
monitoring.” Proceedings of SPIE - Nondestructive Evaluation of Highways, Utilities,
and Pipelines, 3995, 253-263.
Ou, J. P., Li, H. W., Xiao, Y. Q., & Li, Q. S. (2005). “Health dynamic measurement of tall
building using wireless sensor network.” Proceedings of SPIE - Smart Structures and
Materials, 5765(1), 205-215.
Pandey, A. K., Biswas, M. & Samman, M. M. (1991). “Damage detection from changes in
curvature mode shapes.” Journal of Sound and Vibration, 145(2), 321-332.
PCBExpress. <http://www.pcbexpress.com> (Mar. 2, 2007).
Pines, D. & Aktan, A. E. (2002). “Status of structural health monitoring of long-span
bridges in the United States.” Progress in Structural Engineering and Materials, 4(4),
372-380.
Polastre, J., Szewczyk, R. & Culler, D. (2005). “Telos: enabling ultra-low power wireless
research.” Proceedings of 4th International Conference on Information Processing in
Sensor Networks, Los Angeles, CA.
Press W. H., Teukolsky, S. A., Vetterling, W. T., & Flannery, B. P. (1992). Numerical
recipes in C: the art of scientific computing. New York: Cambridge University Press.
Rahimi, M., Hardik, S., Sukhatme, G. S., Heideman, J., & Deborah, E. (2003). "Studying
the feasibility of energy harvesting in a mobile sensor networks." Proceedings of IEEE
Int. Conference on Robotics and Automation, Taipai, Taiwan, 1, 19-24.
Romer, K., Kasten, O., & Mattern, F. (2002). “Middleware Challenges for Wireless Sen-
sor Networks.” Mobile Computing and Communications Review, 6(4), 59-61.
Ruiz-Sandoval, M. (2004). “Smart sensors” for civil infrastructure systems (Doctoral dis-
sertation, University of Notre Dame, 2004).
Ruiz-Sandoval, M., Nagayama, T., & Spencer, B. F., Jr. (2006). “Sensor development us-
ing Berkeley Mote platform.” Journal of Earthquake Engineering, 10(2), 289-309 .
Salawu, O. S. (1997). “Detection of structural damage through changes in frequency: a re-
view.” Engineering Structures, 19(9), 718-723.
Salawu, O. S. & Williams, C. (1994). “Damage location using vibration modeshapes.”
Proceedings of 12th International Modal Analysis Conference, 933-939.
Salawu, O. S. & Williams, C. (1995). “Review of full-scale dynamic testing of bridge
structures.” Engineering Structures 17(2), 113-121.
Sallen, R. P. & Key, E. L. (1955). "A practical method of designing active filters," IRE
Transactions on Circuit Theory, CT-2, 74-85.
Satpathi, D., Victor, J. P., Wang, M. L., & Yang, H. Y. (1999). “Development of a PVDF
film sensor for infrastructure monitoring.” Proceedings of SPIE - 6th Annual Int. Sym-
posium on Smart Structures and Materials, Newport Beach, CA, 3671, 90-99.

175
Sazonov, E., Janoyan, K., & Jha, R. (2004). “Wireless intelligent sensor network for au-
tonomous structural health monitoring.” Proceedings of SPIE - Smart Structures and
Materials: Smart Sensor Technology and Measurement Systems, 5384, 305-314.
Screaming Circuit. <http://www.screamingcircuits.com> (Mar. 2, 2007).
Sensametrics. <http://www.sensametrics.com> (Mar. 2, 2007).
Sensicast Systems. <http://www.sensicast.com> (Mar. 2, 2007).
Shinozuka, M. (2003). “Homeland security and safety.” Proceedings of Structural Health
Monitoring and Intelligent Infrastructure, Tokyo, Japan, 1139-1145.
Silicon Designs, Inc. (2007). “Model 1221 - Low noise analog accelerometer.” Datasheet,
<http://silicondesigns.com/Pdf/1221.pdf> (Mar. 2, 2007).
Simon G., Maroti, M., & Ledeczi, A. (2004). “Sensor network-based countersniper sys-
tem.” Proceedings of SenSys ‘04, Baltimore, MD.
Skolnick, D. & Levine, N. (1997). “Why Use DSP? Digital signal processing 101- an in-
troductory course in DSP system design:” Analog Dialogue, 31(1), <http://www.ana-
log.com/library/analogDialogue/archives/31-1/DSP.html> (Mar. 2, 2007).
Sohn, H., Farrar, C. R., Hemez, F. M., Shunk, D. D., Stinemates, D. W., & Nadler B. R.
(2003). “A review of structural health monitoring literature: 1996-2001” Los Alamos
National Laboratory Report, LA-13976-MS.
Sohn, H., Park, H. W., Law, K., & Farrar, C. R. (2007). “Combination of a time reversal
process and a consecutive outlier analysis for baseline-free damage diagnosis.” Jour-
nal of Intelligent Material Systems and Structures, 18(4), 335-346.
Sohn, H., Worden, K., & Farrar, C. R. (2002). “Statistical damage classification under
changing environmental and operational conditions.” Journal of Intelligent Material
Systems and Structures, 13, 561-574.
Spectral Dynamics, Inc. <http://www.spectraldynamics.com> (Mar. 2, 2007).
Spencer, B. F., Jr. & Nagayama, T. (2006). “Smart sensor technology: A new paradigm
for structural health monitoring.” Proceedings of Asia-Pacific Workshop on Structural
health Monitoring, Yokohama, Japan.
Srinivasan, M. G. & Kot, C. A. (1992). “Effects of damage on the modal parameters of a
cylindrical shell.” Proceedings of 10th International Modal Analysis Conference, 529-
535.
Staszewski, W. J. (1998). “Structural and mechanical damage detection using wavelets.”
The Shock and Vibration Digest, 30(6), 457-472.
Stewart, R. W. & Pfann, E. (1998). “Oversampling and sigma-delta strategies for data
conversion.” Electronics & Communication Engineering Journal, 10(1), 37-47.
STMicroelectronics. <http://www.st.com/stonline> (Mar. 2, 2007).

176
Straser, E. G. & Kiremidjian A. S. (1996). "A modular visual approach to damage moni-
toring for civil structures." Proceedings of SPIE - Smart Structures and Materials, San
Diego, CA, 2719, 112-122.
Straser, E. G. & Kiremidjian A. S. (1998). "A modular, wireless damage monitoring sys-
tem for structures." The John A. Blume Earthquake Engineering Center Technical Re-
port, 128.
Studer, M. & Peters, K. (2004). “Multi-scale sensing for damage identification”, Smart
Materials and Structures, 13, 283-294.
Sun, C. T. and Lu, Y. P. (1995). Vibration damping of structural elements. Englewood
Cliffs, NJ: Prentice Hall.
Tang, J. P. and Leu, K. M. (1989). “Vibration tests and damage detection of P/C bridges”
ICOSSAR 1989, Proceedings of 5th International Conference on Structural Safety
and Reliability, ASCE, San Francisco, CA, 2263-2266.
Tanner, N. A., Farrar, C. R., & Sohn, H. (2002). “Structural health monitoring using wire-
less sensing systems with embedded processing.” Proceedings of SPIE - NDE and
Health Monitoring of Aerospace Materials and Civil Infrastructure, 4704, 215-224.
Tanner, N. A., Wait, J. R., Farrar, C. R., & Sohn, H. (2003). “Structural health monitoring
using modular wireless sensors.” Journal of Intelligent Material Systems and Struc-
tures, 14(1), 43-56.
Texas Instruments. (2007) “Is DSP Right for You?” <http://focus.ti.com/dsp/docs/dspsup-
porto.tsp?sectionId=4&tabId=1443#what_why> (Mar. 2, 2007).
The MathWorks, Inc. <http://www.mathworks.com> (Mar. 2, 2007).
TinyOS. <http://www.tinyos.net> (Mar. 2, 2007).
Toksoy, T. & Aktan, A. E. (1994). “Bridge-condition assessment by modal flexibility.”
Experimental Mechanics, 34, 271-278.
Vapnik, V. (1998). Statistical Learning Theory, New York: Wiley.
Wang, M. L., Gu, H., Lloyd, G. M., & Zhao, Y. (2003). “A multi-channel wireless PVDF
displacement sensor for structural monitoring.” Proceedings of Int. Workshop on Ad-
vanced Sensors, Structural Health Monitoring and Smart Structures, Tokyo, Japan, 1-
6.
West, W. M. (1984). “Illustration of the use of modal assurance criterion to detectstructur-
al changes in an orbiter test specimen.” Proceedings of Air Force Conference on Air-
craft Structural Integrity, 1-6.
Williams, C. & Salawu, O. S. (1997). “Damping as a damage indication parameter.” Pro-
ceedings of 15th International Modal Analysis Conference, 1531-1536.
Wong, K.-Y. (2004). “Instrumentation and health monitoring of cable-supported bridges.”
Structural Control and Health Monitoring, 11(2), 91-124.

177
Yi, J.-H. & Yun, C.-B. (2004). “Comparative study on modal identification methods using
output-only information.” Structural Engineering and Mechanics, 17(3-4), 445-466.
Yurish, S. Y. (2005) “Intelligent opto sensors’ interfacing based on universal frequency-
to-digital converter.” Sensors and Transducers Magazine, 56(6), 326-334.
Zhang, Z. & Atkan, A. E. (1995). “The damage indices for constructed facilities.” Pro-
ceedings of 13th International Modal Analysis Conference, 1520-1529.
Zhao, F. & Guibas, L. J. (2004). Wireless sensor networks: an information processing ap-
proach. San Francisco, CA: Morgan Kaufmann.
Zhao, J., Ivan, J. N., & DeWolf, J. T. (1998). “Structural damage detection using artificial
neural networks.” Journal of Infrastructure Systems, 4(3), 93-101.
Zimmerman, D. C. & Kaouk, M. (1994). “Structural damage detection using a minimum
rank update theory.” Journal of Vibration and Acoustics, 116, 222-230.
Zimmerman, D. C. & Simmermacher, T. (1995). “Model correlation using multiple static
load and vibration tests.” AIAA Journal, 33(11), 2182-2188.
Zimmerman, D. C., Kaouk, M. & Simmermacher, T. (1995). “Structural damage detection
using frequency response functions.” Proceedings of 13th International Modal Analy-
sis Conference, 179-184.
Zimmermann, H. (1980). “OSI reference model - the ISO model of architecture for open
systems interconnection.” IEEE Trans. Communications, 28(4), 425-432.

178
 List of Recent NSEL Reports

No. Authors Title Date

001 Nagayama, T. and Structural Health Monitoring Using Smart Sensors Nov. 2007
Spencer, B.F.
002 Sun, S. and Shear Behavior and Capacity of Large-Scale Prestressed Nov. 2007
Kuchma, D.A. High-Strength Concrete Bulb-Tee Girders
003 Nagle, T.J. and Nontraditional Limitations on the Shear Capacity of Prestressed Dec. 2007
Kuchma, D.A. Concrete Girders
004 Kwon, O-S. and Probabilistic Seismic Assessment of Structure, Foundation, Dec. 2007
Elnashai, A.S. and Soil Interacting Systems
005 Nakata, N., Multi-dimensional Mixed-mode Hybrid Simulation: Dec. 2007
Spencer, B.F., and Control and Applications
Elnashai, A.S.
006 Carrion, J. and Model-based Strategies for Real-time Hybrid Testing Dec. 2007
Spencer, B.F.