Design and Implementation of Scalable Wireless Sensor
Network for Structural Monitoring
Shamim N. Pakzad1; Gregory L. Fenves2; Sukun Kim3; and David E. Culler4
Abstract: An integrated hardware and software system for a scalable wireless sensor network 共WSN兲is designed and developed for
structural health monitoring. An accelerometer sensor node is designed, developed, and calibrated to meet the requirements for structural
vibration monitoring and modal identiﬁcation. The nodes have four channels of accelerometers in two directions and a microcontroller for
processing and wireless communication in a multihop network. Software components have been implemented within the TinyOS oper-
ating system to provide a ﬂexible software platform and scalable performance for structural health monitoring applications. These
components include a protocol for reliable command dissemination through the network and data collection, and improvements to
software components for data pipelining, jitter control, and high-frequency sampling. The prototype WSN was deployed on a long-span
bridge with 64 nodes. The data acquired from the testbed were used to examine the scalability of the network and the data quality. Robust
and scalable performance was demonstrated even with a large number of hops required for communication. The results showed that the
WSN provides spatially dense and accurate ambient vibration data for identifying vibration modes of a bridge.
CE Database subject headings: Sensors; Design; Implementation; Networks; Monitoring.
Structural health monitoring 共SHM兲is a rapidly developing ﬁeld
encompassing the technology and algorithms for sensing the state
of a structural system, diagnosing the structure’s current condi-
tion, performing a prognosis of expected future performance, and
providing information for decisions about maintenance, safety,
and emergency actions 共Farrar 2001; Doebling et al. 1996; Lynch
and Loh 2006兲. Advances in micro-electro-mechanical-systems
共MEMS兲technology in the past decade provide opportunities for
sensing, wireless communication, and distributed data processing
for a variety of new SHM applications. There have been several
prototypes of sensor networks, emphasizing the sensing devices,
such as Wang and Pran 共2000兲, Westermo and Thompson 共1997兲,
and Zimmermann 共1999兲; wireless communication, such as Ihler
et al. 共2000兲, Paek et al. 共2005兲and Pei et al. 共2005兲; and data
processing for SHM, such as Williams and Messina 共1999兲,
Strubbs et al. 共1999兲, and Sohn and Farrar 共2000兲. The rapid
reduction in physical size and cost of MEMS-based wireless sen-
sors has driven increased interest in the scalability of wireless
sensor networks 共WSN兲to hundreds or even thousands of nodes.
While the current research in SHM has made substantial
progress, the scalability of the WSN for structural monitoring
applications has not been thoroughly investigated or demon-
strated. Wireless communication is essential for scalability be-
cause installation and maintenance of a monitoring system with a
wired or tethered communication network would be too expen-
sive and complex for hundreds to thousands of nodes. Scalability
of a wireless sensor network involves the sensors, data acquisition
and processing, and wireless communication. A scalable network
is one that can be expanded in terms of the number of sensors,
complexity of the network topology, data quality 共e.g., sampling
rate, sensor sensitivity兲, and amount of data while the cost of the
expansion 共installation and operational cost, communication time,
processing time, power, and reliability兲is no worse than a linear,
or nearly linear, function of the number of sensors. WSN scalabil-
ity needs to consider an integrated view of the hardware and
software. For hardware, scalability involves sensitivity and range
of MEMS sensors, communication bandwidth of the radio, and
power usage. The software issues include reliability of command
dissemination and data transfer, management of large volume of
data, and scalable algorithms for analyzing the data. The com-
bined hardware-software issues include high-frequency sampling,
which is necessary for structural health monitoring, and the
tradeoffs between on-board computations compared with wireless
communication between nodes. Addressing these problems is es-
sential for the application of WSN beyond laboratory prototypes
to the scale needed for structural monitoring applications.
Wireless sensor networks have been developed for a variety of
purposes that range from low duty-cycle, low-power environmen-
tal monitoring applications, such as described by Mainwaring et
al. 共2002兲and Tolle et al. 共2005兲, to high-ﬁdelity applications
共accurate measurements, high sampling rate, lossless communica-
tion兲for monitoring of mechanical and structural systems.
Mastroleon et al. 共2004兲examined the architecture of a wireless
1Doctoral Student, Dept. of Civil and Environmental Engineering,
Univ. of California, Berkeley, CA 94720 共corresponding author兲. E-mail:
2Professor, Dept. of Civil and Environmental Engineering, Univ. of
California, Berkeley, CA 94720.
3Doctoral Student, Dept. of Electrical Engineering and Computer
Sciences, Univ. of California, Berkeley, CA 94720.
4Professor, Dept. of Electrical Engineering and Computer Sciences,
Univ. of California, Berkeley, CA 94720.
Note. Discussion open until August 1, 2008. Separate discussions
must be submitted for individual papers. To extend the closing date by
one month, a written request must be ﬁled with the ASCE Managing
Editor. The manuscript for this paper was submitted for review and pos-
sible publication on April 9, 2007; approved on June 8, 2007. This paper
is part of the Journal of Infrastructure Systems, Vol. 14, No. 1, March 1,
2008. ©ASCE, ISSN 1076-0342/2008/1-89–101/$25.00.
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 89
system for SHM with an emphasis on low-power multisensor
modular units. Ruiz-Sandoval et al. 共2006兲developed a MEMS-
based accelerometer and strain sensor board for a WSN for struc-
tural health monitoring using Mica motes for communication and
control. Maser et al. 共1996兲described the ﬁeld deployment of a
wireless sensor network on a highway bridge. The wireless trans-
ceiver used one-hop communication with a base station, and the
base stations were interconnected using cellular telephony. Lynch
et al. 共2003兲examined the quality of data from a seven-node
wireless network with MEMS accelerometers by comparison with
conventional wired accelerometers. Lynch et al. 共2005兲presented
the deployment of 14 wireless sensors to monitor forced accelera-
tion response of Geumdang Bridge in Korea. These studies have
been advances in wireless sensor networks for structural health
monitoring, but since the systems rely on one-hop wireless com-
munication between a sensor and a base station, the studies do not
address the question of scalability.
The objective of this paper is to present the design, develop-
ment, and large-scale deployment and testing of a scalable wire-
less sensor network for structural health monitoring. The key
technology innovations are: 共1兲a new approach for maximizing
the effective network bandwidth with a large number of commu-
nication hops; and 共2兲reliable command dissemination and data
transfer for high-frequency sampling, all using a low-power mi-
crocontroller and radio. Both aspects use multihop communica-
tion between nodes, which is essential for scalable networking
because the radio power required for single-hop communication
in a large network is impractical. The ﬁrst section of the paper
presents the requirements and design for the WSN architecture,
describes the hardware and software components, and summa-
rizes the calibration of the system. The second section describes
the network topology and deployment of the scalable WSN for a
long-span bridge, analyzes the network performance, and presents
ambient vibration data of the bridge. The third section describes
the algorithms used to analyze the data and presents information
about the spatial sampling of the vibration properties of the
bridge. The paper concludes with recommendations for achieving
scalability in wireless sensor networks.
Wireless Sensor Network Architecture
The ﬁrst step in designing a sensor network is deciding on the
physical quantities to measure. In the case of a structure as a
dynamic system, the measurement of acceleration is the most
straightforward, but is recognized that acceleration is an indirect
function of damage and structural condition. Although sensing
displacement is possible 关using global positioning system 共GPS兲
technology, for example兴, the reliability, accuracy, and sampling
rate are not yet sufﬁcient for many applications, particularly those
needing high-frequency sampling of small displacements. Strain
measurements could provide a direct measure of damage, but the
installation, operation, and interpretation of reliable strain sensors
on a large structure is difﬁcult. Consequently, accelerations pro-
vide useful information about structural vibration characteristics,
so they are adequate for the primary goal of this investigation,
which is to examine the scalability of wireless sensor networks.
For bridge applications, the accelerometer range should be large
enough to capture 1–2 gduring an earthquake, yet they should be
sensitive enough to measure ambient vibrations due to wind and
trafﬁc on the order of tens to hundreds of g.
The network must be designed for fast sampling rates for tem-
poral scalability, and reliable command dissemination and data
collection over many nodes to provide spatial scalability. Consid-
ering the sampling rate for accelerations or other structural re-
sponse quantities, the lower vibration frequencies of a structure
are generally on the order of 10−1–10+1 Hz, but higher sampling
rates are desirable for two reasons. Local features of response are
characteristic of much higher vibration frequencies and, second,
high-frequency sampling can be used to reduce noise and increase
the signal-to-noise ratio 共Oppenheim and Schafer 1999兲. High-
frequency sampling, however, complicates time synchronization
of nodes over the network and may generate large volumes
of data that need to be managed, processed, and possibly
Another design requirement is that the network must have high
communication reliability to transmit data and disseminate com-
mands without loss of information 共packet loss兲. Particularly for
rare events, such as an earthquake, data loss is not acceptable. For
ambient vibration applications, data loss would have the affect of
increasing the noise, which would make modal identiﬁcation
more difﬁcult. In this work the requirement of no data loss is
High-frequency sampling, multihop communication, and reli-
able data transmission are stringent requirements. The following
subsections provide details about how the hardware and software
components of the scalable WSN are designed to address these
Sensor Node Hardware Design
The network consists of a set of sensor nodes, and each node has
three main hardware components, sensors, ﬁlters and microcon-
troller, and radio for wireless communication. Fig. 1 is a sche-
matic of the major components of a node. For the measurement of
low-level and high-level accelerations, two commercially avail-
able MEMS accelerometer sensors are used each in two directions
共one horizontal and one vertical兲. The use of two sensor types is a
cost-effective solution and allows examination of performance-
price tradeoffs. The high-level sensor is Analog Device’s
ADXL202, a widely used device that provides a ±2 grange with
a sensitivity of 1 mg at 25 Hz 共Analog Devices 1999兲. For low-
level ambient vibrations, a Silicon Design 1221L provides accept-
able sensitivity for ambient structural vibrations at a relatively
low cost 共Silicon Designs 2007兲. Tests show that the Silicon De-
sign accelerometers have a hardware noise ceiling of 10 g
Fig. 1. Schematic diagram of sensor node
90 / JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008
共Pakzad and Fenves 2004兲.共Pakzad et al. 2005兲, which is small
enough to resolve signals with amplitude of a few hundreds of
Each channel from the MEMS accelerometers provides an
analog voltage that is fed to a single-pole anti-aliasing low-pass
ﬁlter with a cutoff frequency of 25 Hz. The ﬁlter was set for a
long-span bridge application because even very high vibration
modes have frequencies well below the anti-aliasing ﬁlter. The
ﬁltered analog signal is fed to a 16-bit analog-to-digital converter
共ADC兲for each of the four channels. Sampling is done at a high
frequency 共1 kHz兲, but the digitized signal is downsampled by
averaging, which acts as a digital ﬁlter and reduces the Gaussian
noise level by a factor of 冑n, when every nsample is averaged.
This analog anti-aliasing ﬁlter at 25 Hz, high-frequency sampling
at 1,000 Hz, and downsampling to 50 Hz provide a simple
and power-efﬁcient approach for high-resolution acceleration
For each node, a mote with a microcontroller provides local
processing and storage capability and a low-power radio commu-
nication. The MicaZ mote was selected because it has a good
tradeoff between processing and communication capability, and
power requirements 共Crossbow 2007a兲. The MicaZ has 512 kB
ﬂash memory, which can store up to 250,000 2-byte data samples,
and a 2.4 GHz radio-frequency 共RF兲Chipcon CC2420 transceiver
with a hardware interface that can support commercially available
bi-directional antennas. The ability of the mote to connect with a
bi-directional antenna was an important factor because the long-
span bridge application required a linear topology and a standard
omni-directional antenna would have wasted a signiﬁcant amount
of radio power.
Power consumption is a critical factor in scalability of wireless
sensor networks. To analyze power for the sensor node with the
MicaZ mote, Fig. 2 shows the power draw for the major compo-
nents of the node in various operational modes based on a 9 V
power source. The power consumption of the sensor board
共MEMS sensors, anti-aliasing ﬁlters, and ADC兲is more than
twice that of the MicaZ mote. This is a result of a design decision
made to use a single power regulator for the node. An improved,
but more complex, hardware design would have separate power
regulators for the sensors and mote, thus allowing the mote to
operate while the sensors are in a sleep mode. Fig. 2 shows that
the mote draws a signiﬁcant portion of power while idle because
the radio is in a listening mode. The broadcast mode only in-
creases the power consumption by about 10%. The test data were
consistent with the speciﬁcation sheet that estimates that the radio
uses 62 mW in the listening mode versus 57 mW in broadcast
mode 共Chipcon Products 2007兲. Based on the power usage testing
of the node, it was decided to use four 6 V lantern batteries to
provide 12 V and 15 A-h in the deployment.
An alternative to the MicaZ mote would have been iMote2,
which has similar functionality as MicaZ but consumes more
power. Crossbow 共2007a,b兲estimates power consumption of
24 mW for MicaZ in active mode with an 8-bit bus size and
8 MHz clock speed versus 139.5 mW for iMote2 with 32-bit bus
size and 12 MHz clock speed. The greater computational power
of iMote2, however, does not make a critical difference in the
performance of the node for the bridge testbed application. The
power consumption was the principal factor in the selection of the
The system software for a scalable wireless sensor network
is based on the TinyOS operating system 共TinyOS 2007兲an open-
source, framework for programming Mica motes 共Hill et al.
2000兲. TinyOS is multilevel component-oriented software that
supports a wide variety of applications for wireless sensor net-
works. Low-level components perform basic tasks, and higher-
level components use sequences of low-level components to
achieve more complex functionality while maintaining efﬁciency
and simplicity of coding. The components range from providing
simple diagnostic operations such as turning an indicator light
emitting diode 共LED兲on/off, to sophisticated components for
routing of data packets in a self-conﬁguring wireless communica-
tion network. With this background, Fig. 3 shows the system
software architecture and the main components of the TinyOS
operating system and application layer that were developed or
adapted for scalable structural health monitoring. The important
components are discussed in the following subsections.
A focus of the recent work on WSN has been on multihop com-
munication. Multihop communication is the transfer of data and
commands between two nodes that are not in the direct radio
range, using intermediary nodes. Multihop communication is es-
sential for scalability of low-power wireless sensor networks be-
Fig. 2. Power consumption of sensor node in operational modes
Fig. 3. Software architecture of TinyOS components for reliable and
high-frequency sampling wireless network
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 91
cause single-hop networks are spatially limited by the radio range
and cannot span long distances without a large power supply.
However, multihop presents major challenges to several aspects
of a WSN. The routing of data packets in a single-hop network
only needs a queue for all of the nodes to transmit their packets
directly to a base station. In a multihop network, routing is more
complex because each node has to determine how to ﬁnd the most
efﬁcient way to forward packets to the base station and coordinate
transmission of packets received from other nodes. The routing
needs to reconﬁgure dynamically for robustness if a node fails
and is no longer able to serve as an intermediary 共such as because
of radio interference兲for multihop communication.
One of the most important aspects of a multihop network is
establishing and updating the routing information 共typically re-
ferred to as routing tables or charts兲for each node to communi-
cate with the base station. The TinyOS component MintRoute
共Woo et al. 2003兲provides multihop connectivity in the wireless
network. MintRoute establishes the routing of packets by mini-
mizing the power cost of the multihop travel of a packet from a
generating node to the base station, subject to a constraint on
minimally acceptable transmission quality for each one-hop link.
If a link falls below a threshold quality level, MintRoute seeks an
alternative route that bypasses the weak link with the minimum
power requirement. The TinyOS component GenericComm pro-
vides low-level communication between a node and one nearby
node that requires the least amount of power 共Hill et al. 2003兲.
The component Broadcast 共Buonadonna 2003兲builds upon Ge-
nericComm to provide radio connectivity between the two nodes.
Each node in a multihop network has two functions. Its ﬁrst func-
tion is to generate data by sampling and sending data packets to
the base station through the network. A node also acts as an
intermediary relay by receiving data packets generated by other
nodes and passing them towards the base station. Spatial reuse of
network bandwidth is essential for a scalable multihop network.
Bandwidth reuse by pipelining means that several nodes in addi-
tion to the one that generated the data, transmit packets at the
same time within a network. Although this increases the amount
of data communicated within the network, pipelining must be
designed so that the nodes transmitting simultaneously do not
interfere with each other. Single-hop networks cannot reuse band-
width because only one node broadcasts at any time to the base
station. Pipelining signiﬁcantly increases the effective bandwidth
of large networks by maintaining a higher throughput for the net-
work regardless of the number of hops.
Fig. 4 illustrates how pipelining allows packets from a node to
travel through different parts of the network at the same time, thus
increasing the effective bandwidth of the network. The parameter
K⫽number of hops between transmitting nodes. If K=n, where
n⫽number of hops in the route, there is no pipelining and the
sender waits until the ﬁrst packet is received by the base station
before sending another packet. As Kdecreases, pipelining reuses
more of the bandwidth, but with an increased potential for radio
interference. The interference causes a higher rate of packet loss
and data retransmission, thus reducing bandwidth. In the limit, K
has a lower bound of 3 because radio interference would jam the
network if K=2, and K=1 is not possible.
For a multihop network without pipelining 共K=n兲the total
transfer time for mpackets from a sender that is nhops away
from the base station is mnT, where T⫽transfer time of one
packet for one hop. Pipelining with a length Kreduces this time
to 共n+mK−K兲T. The effective bandwidth of a network with a
pipeline length of K⬍nto a network without pipelining 共K=n兲is
Fig. 4. Transfer time for n-hop network using pipelining
92 / JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008
bandwith with pipelining
bandwith without pipelining =
The increase in bandwidth is signiﬁcant for a network with large
number of hops, n, and using a small pipeline length K.
To provide spatial reuse of bandwidth in a multihop wireless
sensor network, Kim et al. 共2006兲developed a new TinyOS com-
ponent for data pipelining. The length Kis selected based on the
link quality and radio interference. In this application, Kis set
when the routing table is established and held constant. Addi-
tional work by Kim et al. 共2006兲extends pipelining to set the
optimal value of Kin different parts of a network dynamically.
Reliable Data Communication
Even with multihop routing and data pipelining in a wireless sen-
sor network, data packets will be lost because of electromagnetic
interference and collision of packets that arrive at a node at the
same time. In many applications it is not critically important if a
small percentage of packets are lost, but for SHM it is vital to
have a protocol that guarantees reliable transmission of data be-
cause features of the response characteristics may be affected by
data loss. Furthermore, for monitoring critical events such as an
earthquake, data loss is also unacceptable because of the value of
the information. The challenge is to provide reliability with mini-
mal network resources, in terms of computation and memory, and
hence power. To satisfy the requirement of reliable data commu-
nication, a new protocol, scalable thin and rapid amassment with-
out loss 共Straw兲was developed, tested, and deployed. It is a
selective negative-acknowledgement 共NACK兲collection protocol,
in which the data transfer is always initiated by the receiver. The
sender transmits the data when it is requested by the receiver, and
the receiver then identiﬁes and returns a list of missing packets
back to the sender. The sender retransmits those packets again
until all packets are received. Straw provides reliable data transfer
when the sender and receiver are separated by an arbitrary num-
ber of communication hops 共Kim et al. 2007兲.
Time Synchronization and Jitter Control
Multihop networks have to perform time synchronization. In a
single-hop network this is a trivial task since the base station
sends a start command and all of the nodes receive it virtually at
the same time. The roundtrip communication time from a node to
the base station over multiple hops needs to be accurately esti-
mated for efﬁcient transmission, recognizing that the time is non-
deterministic in a multihop network. This problem is addressed by
a global synchronized clock mechanism with time stamping and
regression algorithms 共Maróti et al. 2004兲.
Jitter is the distortion of a signal caused by variance in the
time-sampling interval as a result of poor synchronization. Ying
et al. 共2005兲explains how accurate jitter control is essential for
high-frequency sampling, since asynchronous data can produce
errors in identifying the mode shapes of a structural system. As
illustrated in the schematic shown in Fig. 5, there are two sources
of jitter: temporal and spatial. Temporal jitter occurs at a node
when the sampling intervals are not uniform due to uneven clock
ticks, event handling, or other hardware bias. Spatial jitter is the
time-synchronization error between different nodes, which occurs
because of uncertainty in estimating latency in propagating a glo-
bal time across the network and the drift of the clock at a node.
To provide time synchronization in a network by controlling
jitter, a TinyOS component named ﬂooding time-synchronization
protocol 共FTSP兲is used. FTSP propagates the global time gener-
ated by the base station through the network by a series of hand-
shakes between adjacent nodes, until the entire network is in the
same time zone. Experiments by Maróti et al. 共2004兲show that
the protocol limits the spatial jitter to 67 s over a network of 59
nodes and 11 hops. High-frequency sampling and logging can
increase this jitter, so Kim et al. 共2007兲performed a jitter analysis
and the tests showed that the temporal jitter using the Timer com-
ponent of TinyOS is limited to 10 s for a sampling rate of up to
6.67 kHz. For a harmonic signal of 25 Hz, the highest frequency
of interest in the current application, this time synchronization
error causes a maximum 0.16% error in the measured value of the
signal. For an ambient acceleration signal of 10 mgamplitude,
the jitter is equivalent to 16 gnoise level, which falls below the
sensitivity of the MEMS accelerometers.
The jitter analysis of the nodes with FTSP provided insight
into the limitations of the motes for a WSN in structural health
monitoring. While the MicaZ microcontrollers are faster than the
ﬂash memory, other microcontroller tasks are delayed because of
sampling, which is a time-consuming operation and thus blocking
computation and communication. Using multiplexed ADC with
its own clock would marginally limit temporal jitter, but at the
cost of consuming additional power for the separate clock. A
faster microcontroller would have smaller jitter but still have the
same fundamental problem that would need to be addressed.
For structural health monitoring, Kim et al. 共2007兲developed an
application named Sentri, based on TinyOS, for high-level control
of a wireless network from a base station. The control program
consists of two components: one for the individual nodes and the
second one for the base station. The node software allows the
mote to listen to the network, join the network, control the sensor
board 共sampling, ﬁltering, logging兲, and be a sender/receiver for
multihop communication. It is designed for a very small memory
footprint because of the limited resources for the motes. The base
station control software has more functionality for sending inquir-
ies to all nodes in the network, evaluating connectivity and com-
munications, and executing commands on parts or the entire
Testing and Calibration
The sensor nodes, network, and the system software were tested
by several laboratory and ﬁeld experiments to determine sensitiv-
ity of the sensors, communication reliability and bandwidth, and
the robustness of the system components. Pakzad et al. 共2005兲
describe these tests, which included quiet-environment tests to
determine the noise ﬂoor of the MEMS accelerometers, shaking
table tests to asses the accuracy of the sensors over a wide fre-
quency range, and a variety of ﬁeld tests to study the multihop
networking and reliable data collection components. Each sensor
node was individually calibrated by a rotary tilt table and shift
and scale factors for converting the ADC output to acceleration
were determined for all channels.
Full-Scale Deployment on Long-Span Bridge
Long-span, suspension bridges have been the subject of study for
structural health monitoring because they are important physical
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 93
infrastructure and can have unique vibration properties and re-
sponse to earthquake ground motion. As two examples of studies
using data from wired sensors, Smyth et al. 共2003兲examined the
Vincent Thomas Bridge in the 1987 Whittier and 1994 Northridge
earthquakes with linear and nonlinear system identiﬁcation tech-
niques to develop a multiinput, multioutput dynamic model of the
bridge using data from 26 accelerometer sensors on the super-
structure and the footings. Abdel-Ghaffar and Scanlan 共1985a,b兲
used spectral densities and ambient vibration data caused by wind
and trafﬁc and collected at 28 locations on the span and a tower of
the Golden Gate Bridge to estimate vibration frequencies and
mode shapes of the bridge.
Building upon the Abdel-Ghaffar and Scanlan study, the
Golden Gate Bridge was selected for the full-scale deployment of
the scalable wireless sensor network described in the previous
sections. The bridge has a main-span 共1,280 m兲, two side-spans
共342.9 m兲, and two towers 共210.3 m above the water level兲. The
objective of the WSN deployment was to identify the vibration
characteristics of the main span and the south tower. The deploy-
ment on the Golden Gate Bridge provided the opportunity to test
the WSN in a difﬁcult environment and with a linear topology
that required a large number of hops for communication.
Fig. 6 shows the instrumentation plan for the bridge with a
total of 64 nodes, 56 on the main span 共measuring transverse and
vertical acceleration兲and eight on the South Tower 共measuring
transverse and longitudinal acceleration兲. On the main span 53
nodes were deployed on the west side, and three nodes on the east
side. Each main span node was attached to the top ﬂange of the
ﬂoor girder directly inside of the cable. Fig. 7 shows a node with
the bidirectional antenna, along with the clamps and guy wires for
temporary installation of the testbed. The node spacing on the
west side was selected based on the range of the radio with a
majority of nodes placed 30.5 m apart, but at places where an
obstacle obstructs a clear line of sight this distance was reduced to
15.25 m. The three nodes on the east side, added in the second
phase after changing the batteries, were located at the two quarter
spans and the midspan of the bridge. The east side nodes have
radio communication with the west side nodes under the roadway
deck. For the South Tower, there is a node on each side of each
strut. The tower nodes have a clear line of sight between them and
hence have greater radio range than the main-span nodes. The
node on the west side of the strut above the superstructure collects
data from all the nodes on the tower and transmits them to the
network on the main-span.
Installation of the network began on July 14, 2006 and the last
set of data was collected on September 22, 2006. The 512 kB
ﬂash memory of each node can buffer about 250,000 samples of
data, which can be allocated to any combination of the ﬁve sensor
channels on the node 共four accelerometers and a temperature sen-
sor兲. Each run starts with a pause to synchronize the network and
disseminate a command to start sampling at a future time. After
the scheduled sampling takes place there is a pause to establish
the network routing. The recorded data are then transferred from
Fig. 5. Spatial and temporal sources of jitter
Fig. 6. Instrumentation plan for 56 nodes on main span of the Golden Gate Bridge
94 / JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008
each node to the base station using the reliable data communica-
tion and pipelining. A complete cycle of sampling and data col-
lection for the full network produces 20 MB of data and takes
about 9 hours. There were a total of 174 such runs during the
deployment. This total includes runs where the network was being
installed and tested so all of the collected data sets do not contain
data from all of the nodes.
The network was fully installed on the west side of the main
span and the south tower on August 1, 2006 and 13 sets of data
were collected with the ﬁrst set of batteries. At the time of chang-
ing batteries on September 22, 2006, the nodes on the east side of
the main span were installed and three more sets of data collected.
The runs include a variety of combinations of sensor channels,
which always included the two low-level Silicon Design 1221
accelerometers, but the other two high-level ADXL202 acceler-
ometers and the temperature sensor were turned off in some of the
runs to reduce the volume of the data or increase the sampling
Ambient Vibration Data
As an example of the ambient vibration data, the vertical accel-
erations from the low-level accelerometers in a typical run 共174兲
are shown in Figs. 8–10 for the three quarter points on the main-
span. The sampling frequency was 50 Hz over 1,600 s, resulting
in 80,000 samples per channel. Each ﬁgure includes plots of the
signal and the power spectral density 共PSD兲using the Welch
method 共Welch 1967兲.
The amplitudes of the ambient accelerations are about
±10 mg, but spikes of up to 50 mgare apparent, presumably
caused by heavy vehicles traveling on the roadway. The PSD
plots show clear and consistent peaks at frequencies at several
nodes. These spectral peaks are distinct in lower frequencies.
Twenty peaks are visible in the frequency plots, which correspond
to vertical and torsional vibration mode shapes of the bridge, as
will be shown later. The PSD plots indicate that the low-
frequency noise level is very small compared with the peaks of
the spectra 共the power of the noise is about two orders of magni-
tude smaller than the power at peak frequencies兲.
Wireless Network Performance
Three aspects of the wireless communication performance for the
network were examined empirically using data from the bridge:
effective bandwidth, loss rate, and average network bandwidth.
These three metrics are important indications of network quality
and are critical to scalability of the network. The effective band-
width is deﬁned as the amount of data per unit time that is sent to
the base station from a node n-hops away. Fig. 11 shows the
effective bandwidth based on empirical measurements of network
performance for four different runs. The one-hop bandwidth of
about 1,200 bytes/s is reduced by each additional hop because
each node has to receive, buffer, and transmit the packets in the
communication stream. A pipeline length of K= 5 was established
for the network, so the effective bandwidth remains relatively
constant for nodes that are beyond the ﬁfth node up to the 45th
hop, which was the deepest hop of the network. These results
show that the data pipelining is very effective for producing a
constant effective bandwidth in a large multihop network.
Considering communication packet loss and retransmission,
Fig. 12 plots the loss rate versus hop count for four runs. Al-
though the loss rate increases as hop count goes up, it is less than
2.5% for 45 hops, which means that the effective volume of trans-
mitted data from the deepest node in the network is only in-
creased slightly due to packet loss. The ﬁgure also shows the loss
rate for the four runs compared with the estimated loss rates for
different values of pipeline length. Using a smaller pipeline length
increases the effective bandwidth but the volume of transmitted
data increases because interference causes higher losses and more
retransmission. In the current deployment K=5 provides a good
balance between effective bandwidth and loss rate.
Another measure of performance is the average network band-
width, deﬁned as the average amount of data collected during a
run per unit time. Fig. 13 shows this metric during the 2-month
deployment. In the installation phase only a few nodes were op-
Fig. 7. Node with its battery pack and bi-directional antenna on main
span of the Golden Gate Bridge
Fig. 8. Time-history and power spectral density for vertical sensor at
west-side south 1/4-span 共run 174, node 64兲
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 95
erational, and the average network bandwidth was dominated by
the effective bandwidth of nodes that are close to the base station.
When additional nodes were being installed and the network de-
bugged, the average bandwidth ﬂuctuated based on which part of
the network was operational and how far the nodes were from the
base station. When the entire network was operational, the aver-
age bandwidth stabilized at 550 bytes/s, taking full advantage of
pipelining. In summary, the empirical data on bandwidth and
packet loss in a large-scale deployment show that the pipelining
of data is very effective in providing a constant and reliable ef-
fective bandwidth for a large number of hops.
Analysis of Vibration Modes
Although the vibration modes of a structure, particularly the
lower modes, are not very sensitive indicators of the health of a
structure 共Doebling et al. 1998兲, they are useful measures to study
the quality of data acquired in the WSN. The ambient vibration
data from the Golden Gate Bridge is used to estimate modal prop-
erties using off-line modal realization methods. The analysis ex-
amines the repeatability of the data and the effect of a spatially
dense sensor network on estimated modal properties.
Modal Identiﬁcation Methods
The continuous-time dynamic system can be modeled by a
multiinput-multioutput 共MIMO兲system. The system reacts to the
input signals x1,...,xp, and produces response signals y1, ...,yq.
A multi-degree-of-freedom mass-damper-spring system is an ex-
ample of such a system and can be mathematically modeled by
where M=mass matrix; C=damping matrix; K=stiffness matrix;
U共t兲=displacement vector; and F共t兲=external forces.
The objective of the system identiﬁcation process is to esti-
mate properties of the system transfer function H共s兲, using the
observed input and output response.
A system identiﬁcation method using multivariate autoregres-
sive models 共ARX兲is used in this study to estimate modal prop-
erties of the bridge 共Ljung 1999兲. The multivariate ARX共M,N兲
consists of Mautoregressive matrices and Nexogenous matrices
Fig. 9. Time-history and power spectral density for vertical sensor at
west-side midspan 共run 174, node 78兲
Fig. 10. Time-history and power spectral density for vertical sensor
at west-side north 1/4–span 共run 174, node 45兲
Fig. 11. Effective bandwidth of wireless sensor network with
pipelining for four runs
96 / JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008
where A共q兲and B共q兲=autoregressive- and exogenous operators in
discrete time domain, respectively; and H共q兲=B共q兲/A共q兲
=discrete system transfer function. The discrete functions y
共n兲, and e
共n兲=output, input, and noise vectors, respectively,
共nT0兲, and T0=sampling period. This
equation can be rewritten in time domain as
The discrete-time input and output vectors are p- and
q-dimensional vectors x
=关y1共n兲y2共n兲¯yq共n兲兴T.Aiare q⫻qmatrices of AR coefﬁcients
while Biare q⫻pcoefﬁcient matrices of exogenous terms. The
noise vector e
共n兲is assumed to be independent and identically
distributed 共IID兲. For a MIMO system with measured input and
output signals, i.e., known as x
number of computational algorithms to estimate Aiand Biparam-
eters are described in Ljung 共1999兲. Stochastic 共output-only兲sys-
tems are special cases of this general model, and when the input
can be reasonably assumed to have characteristics of white noise,
an equivalent autoregressive 共AR兲or autoregressive with moving
average 共ARMA兲model can be used 共Peeters and Roeck 2001兲.
The computational steps for determining the parameters in both
cases are similar to that of ARX methods 共Juang and Phan 2001兲.
After the modal vibration properties are identiﬁed, modal
phase collinearity 共MPC兲is used to distinguish actual modes from
spurious ones that are an artifact of the computation. Vibrations of
different parts of a structure in a classical normal mode are mono-
phase, i.e., the difference between their phases is either 0 or .
MPC is a measure to quantify this monophase behavior 共Pappa
et al. 1993兲. The MPC value for a mode is close to unity for a
noise-free set of data. In this study, a cutoff MPC of 0.90 is
chosen; however, most of the selected modes have MPC values of
Modal Results for Main-Span
With a 50 Hz sampling rate, the signals have a Nyquist frequency
of 25 Hz, but analyzing a large system in this frequency range
requires a very high-order model. To reduce model order and
concentrate on important vibration modes of the main-span below
5 Hz, the signals were low-pass ﬁltered off-line with a Chebychev
Type-II ﬁlter with a 5 Hz cutoff frequency and then downsampled
accordingly. Figs. 14–17 show several estimated vertical and tor-
Fig. 12. Network loss rate versus hop count using pipelining for four
runs compared with theoretical rates with K-nodes pipelining
Fig. 13. Average network bandwidth for individual runs of wireless
Fig. 14. Estimated vertical mode shapes of main span for four data
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 97
sional modes with frequencies in the 0–5 Hz frequency band.
The various vibration mode estimates shown in the ﬁgures are
used to evaluate the repeatability of the information in the ambi-
ent vibration data.
With a high spatial resolution network on the main-span of the
bridge, it is possible to examine how the spatial sampling affects
the repeatability of identiﬁed vibration properties. Two aspects of
node conﬁguration are examined studying relation to the repeat-
ability of modal information: spatial density of the network and
shift in node locations. To study spatial repeatability, the modal
properties using different data sets with different spatial densities
are compared. For shift repeatability, the modal properties from
two different data sets with the same number of nodes but at
different locations are compared with reference modal properties
estimated using the full network data.
Three modes in vertical and torsional directions are used to
examine the effect of spatial resolution in estimated vibration
properties: Modes 1, 11, and 23 in the vertical direction, and
torsional modes 1, 8, and 18. These modes represent the lower,
middle, and higher range of identiﬁed frequencies in each direc-
tion. Figs. 14 and 15 show the mode shapes, frequencies, and
damping ratios of four sets of data. Each set of data consists of a
series of nodes: six nodes in the smallest set, ten nodes in the
second set, 20 nodes in the third set, and 49 nodes in the largest
set. Each set includes three nodes on the east side of the main-
span, and the remaining nodes are on the west side. The east-side
nodes are included to distinguish between vertical and torsional
modes. The data sets used in spatial repeatability analysis are
increasing in size with the doubling of nodes for each subsequent
set. The solid lines in the ﬁgures are ﬁtted splines to the 49-node
data set, which are considered the reference mode shapes. Figs.
14 and 15 show that the mode shapes from all data sets are con-
sistent with the reference shapes. Furthermore, the estimated
frequencies are comparable. On the other hand, the estimated
damping ratios decrease as the number of nodes used in the sys-
tem identiﬁcation increase. This is because the damping estimate
is sensitive to measurement noise. As the number of data points
increases, the model certainty improves, the effect of noise de-
creases, and hence the damping ratio estimate decreases.
Repeatability with respect to the location of the nodes gives a
different measure of consistency of the collected data, because
mutually exclusive data sets with similar sizes can be selected to
compare their vibration contents. Using data sets with similar
sizes makes the effect of measurement noise in estimated param-
Fig. 15. Estimated torsional mode shapes of main span for four data
sets Fig. 16. Estimated vertical mode shapes of main span for shifted
98 / JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008
eters equal. Additionally, using mutually exclusive data sets guar-
antees that the compatibility of the estimated modal properties by
each set is because they independently reﬂect the vibration prop-
erties of the bridge. Two sets of 20 nodes are selected, where the
two sets only share one node at the midspan and all other nodes
are between 15.25 and 30.5 m apart from each other such that one
set is a shifted version of the other set. The same three modes in
vertical and torsional directions are considered in Figs. 16 and 17.
The plots show shift repeatability in the data, and the estimated
frequencies and damping ratios of the two sets match each other
and the mode shapes are consistent with the reference mode shape
estimated using the full record.
The design, implementation, and deployment of a wireless sensor
network for structural health monitoring applications has pro-
vided new information about the scalability and performance of
the network. Hardware components were developed to provide
the range and sensitivity for sensing acceleration from strong mo-
tion as well as ambient structural vibration. Software components
were developed for the TinyOS operating system for reliable
command dissemination and data collection and spatial reuse of
network bandwidth is implemented through pipelining. Time syn-
chronization for high-frequency sampling is provided by modify-
ing TinyOS components to limit total network jitter.
The performance of the network is analyzed during the testbed
deployment on a long-span bridge. The data on network perfor-
mance conﬁrm the effectiveness of spatial reuse of radio band-
width through pipelining in maintaining a scalable network. The
quality of ambient structural vibration data is investigated by
studying the repeatability of the estimated modal properties of the
structure with variable numbers and locations of the nodes. The
estimated structural modal properties are consistent as the number
of nodes increases. The scalable WSN enables a high spatial den-
sity network, which makes it possible to identify higher modes of
vibration with greater accuracy. The identiﬁed modes are also
consistent for data sets with different conﬁguration of nodes that
independently reﬂect the vibration properties of the bridge.
Tests and analysis of the hardware and software provided valu-
able insight in designing real-time WSN for SHM and a roadmap
to future work. In order to guarantee a real-time system that can
instantly react to commands and triggers, two main principles
should be observed in the architecture of software and hardware.
First, the operating system needs to be capable of supporting mul-
tiple threads to avoid delays of computation and communication
tasks while sampling data from the sensors. Second, the hardware
needs to be equipped with a separate microcontroller that is dedi-
cated to sampling job, to let a sampling-only process be enforced.
This is clearly an extra component in the hardware architecture,
which results in more power usage by the node, but is a necessary
feature for a real-time WSN. The addition of a second microcon-
troller also provides more computational power to the node,
which can be used for in-network processing to reduce the vol-
ume of transmitted data and transform the WSN from a sensing/
routing entity to a parallel computing unit as well.
This paper reﬂects the advice and guidance of Professor James
Demmel and Professor Steven Glaser, who participated and sup-
ported the research. The writers provide special thanks to the staff
and management of Golden Gate Bridge District, in particular
Dennis Mulligan and Jerry Kao, for their close cooperation in
every step of the project. Jorge Lee provided extraordinary help in
the deployment, which made this project possible. Thanks to Tom
Oberheim who helped design and develop the sensor board. This
research is supported by the National Science Foundation under
Grant No. EIA-0122599 and by the Center for Information Tech-
nology Research in the Interest of Society 共CITRIS兲at the Uni-
versity of California, Berkeley.
Abdel-Ghaffar, A. M., and Scanlan, R. H. 共1985a兲. “Ambient vibration
studies of Golden Gate Bridge: I. Suspended structure.” J. Eng.
Mech., 111共4兲, 463–482.
Fig. 17. Estimated torsional mode shapes of main span for shifted
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 99
Abdel-Ghaffar, A. M., and Scanlan, R. H. 共1985b兲. “Ambient vibration
studies of Golden Gate Bridge: II. Pier-tower structure.” J. Eng.
Mech., 111共4兲, 483–499.
Analog Devices. 共1999兲. “ADXL202 low cost ±2gdual axis iMEMS®
accelerometer with duty cycle output.” 具http://www.analog.com/
Buonadonna, P. 共2003兲. “Index of/tinyos-1.x/tos/lib/Broadcast.” 具http://
Chipcon Products. 共2007兲. “2.4 GHz IEEE 802.15.4/ZigBee-Ready RF
Tranceiver 共Rev. B兲.” 具http://focus.ti.com/lit/dx/symlink/cc2420.pdf典.
Crossbow. 共2007a兲. “MICAz, wireless measurement system.”
Crossbow. 共2007b兲. “iMote2, high-performance wireless sensor network
Doebling, S. W., Farrar, C. R., and Prime, M. B. 共1998兲. “A summary
review of vibration-based damage identiﬁcation methods.” Shock Vib.
Doebling, S. W., Farrar, C. R., Prime, M. B., and Shevitz, D. W. 共1996兲.
“Damage identiﬁcation and health monitoring of structural and me-
chanical systems from changes in their vibration characteristics: A
literature review.” Rep. No.LA-13070-MS, Los Alamos National
Laboratory, Los Alamos, N.M.
Farrar, C. R. 共2001兲. “Historical overview of structural health monitor-
ing.” Lecture notes on structural health monitoring using statistical
pattern recognition, Los Alamos Dynamics, Los Alamos, N.M.
Hill, J., Gay, D., and Levis, P. 共2003兲. “index of tinyos-1.x/tos/system.”
Hill, J., Szewczyk, R., Woo, A., Hollar, S., Culler, D. E., and Pister, K. S.
J. 共2000兲. “System architecture directions for networked sensors.”
Proc., 9th Int. Conf. on Architectural Support for Programming Lan-
guages and Operating Systems (ASPLOS 2000), Cambridge, Mass.,
Ihler, E., Zaglauer, W., Herold-Schmidt, U., Dittrich, K. W., and Wies-
beck, W. 共2000兲. “Integrated wireless piezoelectric sensors.” Proc.
Juang, J. N., and Phan, M. Q. 共2001兲.Identiﬁcation and control of me-
chanical systems, Cambridge University Press, Cambridge, U.K.
Kim, S., Fonseca, R., Kumar Dutta, P., Tavakoli, A., Culler, D. E., Levis,
P., Shenker, S., and Stoica, I. 共2006兲. “Flush: A reliable bulk transport
protocol for multihop wireless network.” Technical Rep. No.UCB/
EECS-2006-169, University of California, Berkeley, Calif.
Kim, S., Pakzad, S. N., Culler, D., Demmel, J., Fenves, G. L., Glaser, S.,
and Turon, M. 共2007兲. “Health monitoring of civil infrastructures
using wireless sensor networks.” Proc., 6th Int. Conf. on Information
Processing in Sensor Networks (IPSN 2007), Cambridge, Mass.
Ljung, L. 共1999兲.System identiﬁcation—Theory for the user, 2nd Ed.,
Prentice-Hall, Upper Saddle River, N.J.
Lynch, J. P., and Loh, K. J. 共2006兲. “A summary review of wireless
sensors and sensor networks for structural health monitoring.” Shock
Vib. Dig.,38共2兲, 91–128.
Lynch, J. P., Sundararajan, A., Law, K. H., Kiremidjian, A. S., Carryer,
E., Sohn, H., and Farrar, C. R. 共2003兲. “Field validation of a wireless
structural health monitoring system on the Alamosa Canyon Bridge.”
Proc. SPIE, 5057, 267–278.
Lynch, J. P., Wang, Y., Law, K. H., Yi, J.-H., Lee, C. G., and Yun, C. B.
共2005兲. “Validation of a large-scale wireless structural monitoring sys-
tem on the Geumdang Bridge.” Proc., Int. Conf. on Safety and Struc-
tural Reliability (ICOSSAR), Rome, Italy.
Mainwaring, A., Polastre, J., Szewczyk, R., Culler, D., and Anderson, J.
共2002兲. “Wireless sensor networks for habitat monitoring.” Proc.,
2002 ACM Int. Workshop on Wireless Sensor Networks and Applica-
tions, WSNA’02, Atlanta.
Maróti, M., Kusy, B., Simon, G., and Lédeczi, Á. 共2004兲. “The ﬂooding
time synchronization protocol.” Proc. ACM 2nd Int. Conf. on Embed-
ded Networked Sensor Systems (SenSys 2004), Baltimore, 39–49.
Maser, K., Egri, R., Lichtenstein, A., and Chase, S. 共1996兲. “Field evalu-
ation of a wireless global bridge evaluation and monitoring system.”
Proc., 11th Conf. on Engineering Mechanics, Vol. 2, Fort Lauderdale,
Mastroleon, L., Kiremidjiana, A. S., Carryerb, E., and Law, K. H. 共2004兲.
“Design of a new power-efﬁcient wireless sensor system for structural
health monitoring.” Proc. SPIE, 5395, 51–60.
Oppenheim, A. V., and Schafer, R. W. 共1999兲.Discrete-time signal pro-
cessing, 2nd Ed., Prentice-Hall, Upper Saddle River, N.J.
Paek, J., Chintalapudi, K., Cafferey, J., Govindan, R., and Masri, S.
共2005兲. “A wireless sensor network for structural health monitoring:
Performance and experience.” Proc., 2nd IEEE Workshop on Embed-
ded Network Sensors (EmNetS-II), Sydney, Australia.
Pakzad, S. N., and Fenves, G. L. 共2004兲. “Structural health monitoring
applications using MEMS sensor networks.” Proc., 4th Int. Workshop
on Structural Control, Columbia Univ., New York, 47–56.
Pakzad, S. N., Kim, S., Fenves, G. L., Glaser, S. D., Culler, D. E., and
Demmel, J. W. 共2005兲. “Multipurpose wireless accelerometers for
civil infrastructure monitoring.” Proc., 5th Int. Workshop on Struc-
tural Health Monitoring (IWSHM 2005), Stanford, Calif.
Pappa, R. S., Elliott, K. B., and Schenck, A. 共1993兲. “Consistent mode
indicator for eigen system realization algorithm.” J. Guid. Control
Peeters, B., and Roeck, G. D. 共2001兲. “Stochastic system identiﬁcation
for operational modal analysis: A review.” J. Dyn. Syst., Meas., Con-
trol, 123, 659–667.
Pei, J.-S., Kapoor, C., Graves-Abe, T. L., Sugeng, Y., and Lynch, J. P.
共2005兲. “Critical design parameters and operating conditions of wire-
less sensor units for structural health monitoring.” Proc., 23rd Int.
Modal Analysis Conf. (IMAC XXIII), Orlando, Fla.
Ruiz-Sandoval, M., Nagayama, T., and Spencer, B. F. 共2006兲. “Sensor
development using Berkeley Mote platform.” J. Earthquake Eng.,
Silicon Design 共2007兲. “Model 1221, low noise analog accelerometer.”
Smyth, A. W., Pei, J., and Masri, S. F. 共2003兲. “System identiﬁcation of
the Vincent Thomas suspension bridge using earthquake records.”
Earthquake Eng. Struct. Dyn., 32, 339–367.
Sohn, H., and Farrar, C. R. 共2000兲. “Statistical process control and pro-
jection techniques for damage detection.” Proc., European COST F3
Conf. on System Identiﬁcation and Structural Health Monitoring,
Madrid, Spain, 105–114.
Strubbs, N., Sikorsky, C., Park, S. C., and Bolton, R. 共1999兲.Veriﬁcation
of a methodology to nondestructively evaluate the structural proper-
ties of bridges, Structural Health Monitoring Workshop, Stanford
Univ., Palo Alto, Calif., 440–449.
TinyOS. 共2007兲. “TinyOS community forum, and open-source for the
networked sensor regime.” 具http://www.tinyos.net/典.
Tolle, G., et al. 共2005兲. “A macroscope in the redwoods.” Proc., 3rd ACM
Conf. on Embedded Networked Sensor Systems (Sensys 05), San
Wang, G., and Pran, K. 共2000兲. “Ship hull structure monitoring using
ﬁber optic sensors.” Proc., European COST F3 Conf. on System Iden-
tiﬁcation and Structural Health Monitoring, Madrid, Spain, 15–27.
Welch, P. D. 共1967兲. “The use of fast Fourier transform for the estimation
of power spectra: A method based on time averaging over short, modi-
ﬁed periodograms.” IEEE Trans. Audio Electroacoust.,15共2兲, 70–73.
Westermo, B., and Thompson, L. D. 共1997兲. “A peak strain sensor for
damage assessment and health monitoring.” Structural health moni-
toring, Current status and perspectives, Stanford University Press,
Palo Alto, Calif., 515–526.
Williams, E. J., and Messina, A. 共1999兲. “Applications of the multiple
100 / JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008
damage location assurance criterion.” Proc., Int. Conf. on Damage
Assessment of Structures, Dublin, Ireland, 256–264.
Woo, A., Tong, T., and Culler, D. 共2003兲. “Taming the underlying chal-
lenges of reliable multihop routing in sensor networks.” Proc., SenSys
2003, Los Angeles.
Ying, L., Kiremidjian, A. S., Nair, K. K., Lynch, J. P., and Law, K. H.
共2005兲. “Algorithms for time synchronization of wireless structural
monitoring sensors.” Earthquake Eng. Struct. Dyn., 34, 555–573.
Zimmermann, D. C. 共1999兲. “Looking into the crystal ball: The continued
need for multiple viewpoints in damage detection.” Damage Assess-
ment of Structures, Proc., Int. Conf. on Damage Assessment of Struc-
tures, Dublin, Ireland, 76–90.
JOURNAL OF INFRASTRUCTURE SYSTEMS © ASCE / MARCH 2008 / 101