NSEL Report Series

Report No. NSEL-011

May 2008

Structural Health Monitoring Strategies

for Smart Sensor Networks

Yong Gao
and
Billie F. Spencer, Jr.

NEWMARK STRUCTURAL ENGINEERING LABORATORY

Department of Civil and Environmental Engineering

University of Illinois at Urbana-Champaign
 UILU-ENG-2008-1805

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 July 2005. 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 grant CMS 03-01140 (Dr. S. C. Liu, Program Manager). 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

Structural health monitoring (SHM) is an emerging field in civil engineering,

offering the potential for continuous and periodic assessment of the safety and integrity of
civil infrastructure. Based on knowledge of the condition of the structure, certain
preventative measures can be carried out to prolong the service life of the structure and
prevent catastrophic failure. However, challenges remain to apply SHM to civil
engineering structures.
The research detailed in this report has three complimentary efforts that seek to
address some of those challenges. The first component is to experimentally verify an
existing damage detection method utilizing a three-dimensional 14-bay truss structure at
the Smart Structures Technology Laboratory (SSTL) of University of Illinois at Urbana-
Champaign (UIUC). This flexibility-matrix-based method has drawn considerable
attention recently; however, only numerical validation had been previously provided.
Experimental verification allows assessment of the efficacy of the method in practice.
The second part of the work is directed toward extending the flexibility-matrix-
based approach to continuous online SHM employing ambient vibration (i.e., unmeasured
input excitations). Continuous online SHM of civil infrastructure is highly desired,
because it allows early detection of the damage in a structure and therefore offers the
possibility to extend the service life of the structure.
Finally, a new distributed computing SHM strategy, which is suitable for
implementation on arrays of densely distributed smart sensors, is proposed for monitoring
of civil infrastructure. Recent development of smart sensor technology has the potential to
fundamentally change how civil infrastructure will be monitored. Damage detection
algorithms which can take advantage of smart sensor technology are highly desired, but
currently very limited. The new approach proposed in this research is different from the
traditional ones which have relied on central data acquisition and processing, and
therefore meshes well with the distributed computing environment offered by smart sensor
technology. A strong basis for application of SHM to civil engineering structures using
smart sensors has been provided.
 CONTENTS

CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Importance of Damage Detection of Civil Infrastructure . . . . . . . . . . . . . . . . 1
1.2 Damage in Civil Engineering Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Challenges of Applying Existing SHM Methods to Civil Infrastructure . . . . 2
1.4 Overview of Report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
CHAPTER 2 LITERATURE REVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1 Vibration Based Damage Detection Techniques . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Smart Sensor Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
CHAPTER 3 EXPERIMENTAL VERIFICATION OF THE DAMAGE
LOCATING VECTOR METHOD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1 Motivation for Flexibility-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2 The DLV Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.1 General concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Calculation of DLVs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.3 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Construction of Flexibility Matrix Using Limited Sensor Information . . . . . 20
3.4 Experimental Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4.1 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4.2 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
CHAPTER 4 CONTINUOUS STRUCTURAL HEALTH MONITORING
EMPLOYING AMBIENT VIBRATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1 Construction of the Flexibility Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.1.1 Formulation of the flexibility matrix based on forced vibration . . 32
4.1.2 Formulation of the flexibility matrix based on ambient vibration . 34
4.2 Extension of the DLV Method for Online Damage Diagnosis . . . . . . . . . . . 36
4.2.1 Evaluation of modal normalization constant change due to damage36
4.2.2 Algorithm initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2.3 Algorithm operation – detecting damage . . . . . . . . . . . . . . . . . . . . 39
4.3 Numerical Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.1 Algorithm initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.2 Damage diagnosis results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
CHAPTER 5 DISTRIBUTED COMPUTING SHM STRATEGY
USING SMART SENSORS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2 Locating Damage Using Local Sensor Information . . . . . . . . . . . . . . . . . . . 49
5.3 Distributed Computing Strategy (DCS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3.1 Hierarchical organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
 5.3.2 Strategy implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.4 Numerical Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.4.1 Constructing undamaged flexibility matrix in communities . . . . . 57
5.4.2 Damage detection results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
CHAPTER 6 EXPERIMENTAL VALIDATION OF THE DISTRIBUTED
COMPUTING SHM STRATEGY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.2 Algorithm Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
6.3 Damage Detection and Decision Making . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.3.1 Excitation condition 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
6.3.2 Excitation condition 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.3.3 Excitation condition 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
CHAPTER 7 IMPLEMENTATION OF THE DCS APPROACH ON A
SIMULATED SMART SENSOR NETWORK . . . . . . . . . . . . . . . . . . . . . . . 91
7.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
7.2 Overview of the Simulink and Stateflow Model . . . . . . . . . . . . . . . . . . . . . . 92
7.3 Numerical and Experimental Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
CHAPTER 8 CONCLUSIONS AND FUTURE STUDIES . . . . . . . . . . . . . . . . . . 107
8.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
8.2 Future Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.2.1 Effect of excitation conditions and damage scenarios . . . . . . . . . 109
8.2.2 Probability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
8.2.3 Implementation of the DCS approach on smart sensor networks . 110
8.2.4 Extension of the DCS approach to more complicated structures . 110
8.2.5 Optimal sensor topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
8.2.6 SHM strategies employing multi-scale information . . . . . . . . . . . 111
REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
 CHAPTER 1

INTRODUCTION

1.1 Importance of Damage Detection of Civil Infrastructure

Our daily lives are becoming more and more dependent on civil infrastructure,
including bridges, buildings, pipelines, offshore structures, etc. Much of the existing
infrastructure in the United States has been in service for many years. These structures
continue to be used, despite aging and the associated accumulation of damage. For
example, the George Washington Bridge which crosses the Hudson River between upper
Manhattan and Fort Lee, New Jersey has been in service for more than 70 years. This
bridge is still in use and deemed as one of the busiest bridges in the world. Any
malfunction of this bridge could cause tremendous economic loss. Aktan et al. (2001)
pointed out that most of the long-span bridges in the United States, which have 100 meters
or longer spans, are over 50 years old and several are even over 100 years old; more than
800 of these kinds of bridges are classified as fracture-critical in the National Bridge
Inventory. Monitoring the condition of these structures to provide the necessary
maintenance has become critically important to our society. Moreover, evaluation of the
condition of critical facilities and civil infrastructure is extremely important after natural
hazards such as earthquake, or man-made disasters such as terrorist attack. These
emergency facilities have to be evaluated and repaired immediately to minimize the
impact of the disaster and to facilitate the recovery of our society.
Structural health monitoring (SHM) is an emerging field in civil engineering, offering
the potential for continuous and periodic assessment of the safety and integrity of the civil
infrastructure. Based on knowledge of the condition of the structure, certain preventative
measures can be carried out to prolong the service life of the structure and prevent
catastrophic failure. Damage detection strategies can ultimately reduce life cycle cost.

1.2 Damage in Civil Engineering Structures

In the most general terms, damage can be defined as changes introduced into a system
that adversely affects its performance (Farrar et al. 1999). As for civil engineering
structures, changes in materials, connections, boundary conditions, etc., which cause
deteriorated performance of the structure, can be defined as damage. For example,
material aging usually reduces the load capacity of structural elements which leads to
stress redistribution in the structure. This stress redistribution can result in loads that are
substantially different from those expected based on the original structural design,
potentially undermining the safety of the structure and even leading to its failure.
Structural damage can be caused in various ways. Normal activities can introduce
damage to the structure. Buildings can be damaged due to corrosion, aging, and daily
activities. Traffic and wind loads cause damage on bridges, while offshore structures
suffer from wave loading and corrosion due to the seawater. On the other hand, excessive

1
loads produced by tornados, hurricanes, and earthquakes also can potentially cause
damage in structures.
The effect of damage on structures can be classified as linear and nonlinear (Doebling
et al. 1996). Linear damage can be defined as the case when structures still behave linear-
elastically after damage is introduced, while nonlinear damage causes structures to show
nonlinear behavior after damage has occurred. In civil engineering structures, metal
corrosion and concrete spalling/scour are typical damage events that may be defined as
linear damage. Both corrosion and spalling/scour can significantly reduce the cross
section of structural members, and therefore degrade the load capacity of the structure. For
example, severe spalling resulted from rapid current often happens at the waterline of
bridge piles with a considerable loss of cross-section. Examples of nonlinear damage in
civil engineering include cracks formed in concrete or metal members, loose connections
of steel members, etc. For example, structural members with fatigue cracks often show
nonlinear behaviors due to the open and close of the cracks even under the normal
operating vibration environment.
The research detailed in this report only addresses the case of linear damage in
structures. In particular, structural damage is simulated by decreasing the elastic modulus
in numerical studies and by reducing the cross section of structural members in
experimental studies. The methods developed here are expected to be able to
accommodate various cases of linear damages that result in a loss of structural stiffness.

1.3 Challenges of Applying Existing SHM Methods to Civil

Infrastructure

Numerous SHM methods have been proposed in the past few decades; however,
challenges remain before they can be applied to civil engineering structures. Most existing
SHM methodologies require direct measurement of the input excitation for
implementation. However, in many cases, there is no easy way to measure these inputs –
or alternatively, to externally excite the structure. This difficulty has limited the
application of existing SHM methods which require the measurements of input
excitations. Methods based on ambient vibration have become more important in the field
of SHM and damage detection. More research efforts should be directed toward the
development of SHM methodologies which minimize the needs to measure the input
excitation and can handle the ambient vibration case.
Another challenge results from the fact that damage in structures is an intrinsically
local phenomenon. Sensors close to the damaged site are expected to be more heavily
influenced than those remote to the damage. Therefore, to effectively detect damage at an
arbitrary location in a structure, sensors must be densely distributed throughout the
structure. Most existing SHM methods assume the measured data is to be centrally
acquired. Using traditionally wired sensors to implement such a SHM system with a dense
array of sensors is quite challenging because of the cost and difficulties in deploying and
maintaining associated wiring plant. A traditional SHM system employing wired sensors
is shown in Fig. 1.1. The wiring system for a large civil infrastructure is obviously much
more complicated and therefore more difficult to manage. In addition, a tremendous

2
 Sensors Sensors Sensors

Centralized
Data Acquisition

Figure 1.1: Traditional SHM system using centralized

data acquisition (Spencer et al. 2004).

amount of data is expected to be generated that would need to be sent to the central station.
Managing such a large amount of data is also challenging and is not cost-effective.
Therefore, damage detection of large civil infrastructure employing traditionally wired
sensors is intractable.
Recent development of smart sensors has made SHM using a dense array of sensors
feasible (Spencer et al. 2002, 2004). The essential feature of a smart sensor is the on-board
microprocessor, which grants sensors the “smart” characteristics. Programming can be
embedded in the sensor’s microprocessor, which allows smart sensors to save data locally,
perform desired computation, make “if-then” decisions, scan valuable information, send
results quickly, etc. Therefore, a portion of the computation can be done at the local sensor
level for damage detection. Extraneous information can be discarded, reducing the
information that needs to be transferred back to the central station. Note that all smart
sensors to date are wireless as well, with data transmissions based on radio frequency (RF)
communications. A typical wireless SHM system using smart sensors is shown in Fig. 1.2.
Damage detection algorithms which can take advantage of the distributed computing
environment offered by smart sensor technology are highly desired but currently very
limited.

1.4 Overview of Report

The research detailed in this report seeks to address some of these challenges, with
the final goal of developing a new SHM strategy that is suitable for implementation on a
dense array of smart sensors.
Chapter 2 reviews previous research on vibration-based damage detection methods,
as well as the smart sensor technology.

3
 Smart Sensors Smart Sensors Smart Sensors

Base-Station

Figure 1.2: Wireless SHM system using smart sensors

(Spencer et al. 2004).

Chapter 3 presents the experimental verification of a flexibility matrix based damage

detection method, the damage locating vector (DLV) method. The DLV method has drawn
considerable attention recently; however, only numerical examples have been provided to
date. Experimental verification needs to be conducted to assess the efficacy of the method
in practice. Following presentation of the motivation for the flexibility-based methods, the
basic concept of the DLV method is introduced. Construction of the flexibility matrix for
proportionally damped structure employing forced vibration is then reviewed. Finally,
details of the experimental setup and experimental verification results are presented in this
chapter.
Extension of the DLV method to continuous online SHM employing ambient
vibration is described in Chapter 4. Continuous online monitoring of civil infrastructure is
highly desirable, because it allows the damage in the structure to be detected at an early
stage so that necessary measures can be carried out in time to prevent further damage to
the structure. In this chapter, construction of the flexibility matrix employing forced
vibration for structures with general viscous damping and employing ambient vibration
for proportionally damped structures is described. Changes of the modal normalization
constants due to structural damage are then investigated, which leads to the extension of
the DLV method for continuous online SHM. Finally, numerical validation of the
proposed approach is presented.
Chapter 5 presents the development of a new distributed computing strategy (DCS)
suitable for implementation of SHM on a dense array of smart sensors. The proposed
algorithm employs only locally measured sensor information to monitor the portion of the
structure which is in the vicinity of the sensors. Measured data are aggregated locally with
extraneous information being discarded before sending to a central station. This new
approach is therefore different from the traditional algorithms which are reliant on central

4
data acquisition and processing. The concept of damage localization using local sensor
information is first presented. The hierarchical distributed computing SHM strategy is
then proposed. Numerical validation of the proposed approach employing a 14-bay planar
truss is provided.
In Chapter 6, experimental validation of the proposed DCS strategy is presented using
a 6.5 m long truss structure. This three-dimensional 14-bay truss structure is tested at the
Smart Structures Technology Laboratory (SSTL) of University of Illinois at Urbana-
Champaign (UIUC). Following the discussion of the experimental setup, experimental
results for the algorithm initialization are first presented. Damage detection results of
different damage scenarios under various excitation conditions are then provided. Finally,
preliminary reliability study using experimental data is conducted.
Reference implementation of this distributed computing SHM strategy in a Simulink
and Stateflow model is presented in Chapter 7. Numerical and experimental studies of the
efficacy of the proposed DCS approach are carried out in Chapter 5 and 6. However, some
important issues of implementing the proposed approach on real smart sensor networks
have not been studied. A Simulink and Stateflow model which simulates a smart sensor
network consisting of eleven sensor communities has been developed in this chapter to
better understand the application of the proposed DCS approach in practice. Background
knowledge of Simulink and Stateflow is first provided. Detailed description of the
proposed Simulink and Stateflow model is then presented, followed by implementation of
the proposed DCS approach in the proposed model employing both simulation and
experimental data.
Finally, Chapter 8 summarizes the research detailed in this report and presents
possible directions for future research on SHM employing smart sensors.

5
 CHAPTER 2

LITERATURE REVIEW

This chapter presents a brief review of some of the existing vibration-based damage
detection methods, as well as recent developments in smart sensor technology.

2.1 Vibration Based Damage Detection Techniques

Numerous existing health monitoring methods were carefully reviewed by Doebling et

al. (1996) and Sohn et al. (2003). Vibration-based damage detection techniques, which
usually do not require a prior knowledge of the damage location and by which damage
covered by non-structural elements can still be detected, have been an important research
area for SHM. Vibration-based damage detection methods are often classified based on
the type of the measured data used and/or the technique used to identify the damage.
One category of these methods uses frequency changes. Vandiver (1975) detected the
damage in an offshore light station tower by examining the frequency changes in the first
two bending modes and the first torsional mode. By developing an analytical model and
systematically removing members from the model to simulate structural damage, the
author demonstrated that failure of most members produces a frequency change greater
than 1%, and thus, damage in most of the elements is detectable. Adams et al. (1978) and
Cawley and Adams (1979) proposed a sensitivity-based damage localization method
using only frequency measurements. This method is based on the assumption that the ratio
of the frequency change between two modes is only a function of the damage locations.
By calculating the theoretical frequency ratios due to the damage in selected locations and
comparing them with the measured ones, the damage locations were determined.
However, some erroneous results were obtained when implementing this method.
Additionally, the accuracy of this method is highly dependent on the accuracy of the
analytical model. Stubbs (1985) and Stubbs and Osegueda (1987) presented a method
which relates the change of the element stiffness to the change of the modal frequency.
This method assumes that the structural stiffness matrix can be expanded in terms of a
diagonal element stiffness matrix. This assumption is only valid for simple structures, for
example, truss structures. By removing this assumption, Stubbs and Osegueda (1990a, b)
gave a more general expression relating the element stiffness matrix and the modal
frequencies. For multiple damage scenarios, false damage locations were predicted by this
method. Spyrakos et al. (1990), Chen et al. (1995), and Brincker et al. (1995) reported
experimental studies on frequency changes due to damage in the structure. As expected,
the frequency decreases when the damage extent increases. Messina et al. (1996) and
Williams et al. (1997) introduced the damage location assurance criterion (DLAC), which
is a correlation criteria between the measured frequency change and the frequency change
due to damage in assumed locations.
Although numerous researchers have been working on the techniques only employing
frequencies, several limitations exist. Salawu (1997) pointed out that a 5% frequency shift

6
might be required to detect structural damage with confidence when using frequencies
only. In addition, frequency shifts alone might not necessarily indicate that damage has
occurred in the structure. As Aktan et al. (1994) reported, significant frequency shifts
(exceeding 5%) caused by changes in ambient conditions have been measured for bridges
in a single day. Additionally, different damage locations can produce the same degree of
the frequency shift; therefore, using only frequency changes might not be sufficient to
uniquely determine the damage location. Moreover, damage in low stress regions might be
particularly difficult to detect employing frequency data alone (Salawu 1997).
Another category of these methods uses the change in mode shapes. West (1984) was
perhaps the first to implement systematic use of mode shape information for damage
detection. His technique uses the modal assurance criteria (MAC) to localize structural
damage. To implement this method with confidence, numerous sensors appear to be
necessary. After numerical and experimental study of a beam, Fox (1992) showed that the
MAC value is not so sensitive to damage and suggested that graphical comparison of the
mode shapes might be a good way to locate damage. Kim et al. (1992) successfully
isolated damage in a structure by employing the Coordinate MAC (COMAC) in
conjunction with the Partial MAC (PMAC). Ko et al. (1994), Lam et al. (1995), and
Salawu (1995) also did extensive research in this area in the past decade.
Modal strain energy has also been employed to detect structural damage. Carrasco et
al. (1997a, b) located and quantified the damage in a truss structure using changes in the
modal strain energy before and after damage. However, damage simulated by a cut
through half the depth of the element could not be located. Shi et al. (1998, 2000) also
applied the concept of modal strain energy change to detect structural damage. Elements
with a high modal strain energy change ratio were identified as possibly damaged
elements. Then, a damage quantification algorithm was developed by expressing the
modal strain energy change in terms of the contributions from the possible damage
elements. In this algorithm, iteration is required to obtain a converged damage measure,
and results showed that many modes are needed to obtain an accurate estimate of the
damage extent. In practice, measuring higher modes is challenging. Shi et al. (2002)
improved their quantification algorithm by using only the first few modes. The numerical
example showed good results. For experimental verification, the authors assumed that the
type of the damage was known a priori, which is usually not the case.
Ricles and Kosmatka (1992) used residue force as an indicator of the damage in the
structure. Damage locations are first identified at the degrees of freedom (DOFs) with
non-zero residue forces. The damage extent is then determined by expressing the damaged
modal parameters as a Taylor series in terms of the structural parameters in the damaged
region and the corresponding undamaged modal parameters. The results indicated that this
method is quite model dependent. Baruh and Ratna (1993) and Sheinman (1996) also used
the residue force to detect the damage with a different quantification algorithms.
Parameter estimation methods have also shown promising. By estimating structural
parameters, not only the location but the extent of the damage can be obtained. Yao et al.
(1992) successfully performed parameter estimation on a steel frame structure for damage
detection using experimental data. Shin and Hjelmstad (1994) proposed a parameter
estimation method employing an adaptive element-grouping scheme. This method has

7
shown to be successful using noisy and limited measurement information. However,
errors arise in some of the cases when multiple damage locations exist. Dos Santos and
Zimmerman (1996) propose a method employing minimum rank perturbation theory in
conjunction with ordinary least-squares estimation for damage detection. Hjelmstad
(1996) observed that there will not be a unique solution for structural parameter estimation
in some of the cases when measurements are too limited and sparsely distributed. Pothisiri
and Hjelmstad (2003) improved the method proposed by Shin and Hjelmstad (1994) by
successfully tackling the multiple damage scenario using a new element-group updating
scheme. The proposed approach was also shown to be able to handle the nonuniqueness
issue in parameter estimation. In this method, errors in the estimated parameters are
reduced by selecting a near-optimal measurement set.
Another group of these methods takes advantage of the changes in the flexibility
matrix. Because there is an inverse relationship between the flexibility and stiffness
matrix, damage in the structure, which leads to a decrease in the stiffness, will increase the
flexibility. Pandey and Biswas (1994, 1995b) employed the flexibility matrix to locate
damage. First, the flexibility matrices before and after damage are constructed from the
measured data. The maximum value in each column (each column corresponding to one
DOF) of the flexibility matrix change was then selected as indicator of the damage
location. The columns which have larger values are the possible damage locations.
However, directly using the flexibility matrix change to do damage localization might be
difficult for multiple damage locations. Pandey and Biswas (1995a) further improved the
method based on the flexibility matrix to locate and quantify the damage in the structure.
A formulation describing the relationship between the change in the flexibility matrix as a
function of the change in the elemental stiffness matrix was developed. A pseudo inverse
technique was then applied to obtain the change in the elemental stiffness. Toksoy and
Aktan (1994) examined the damage on a real bridge based on the measured modal
flexibility matrix. The modal flexibility matrices before and after damage were developed
from the measurements. Then, these matrices were loaded with different load
configurations and the deflection profiles obtained. Comparison of the deflection profiles
gave the damage location information in the bridge. More recently, Bernal (2002)
proposed a flexibility-matrix-based damage localization method — the damage locating
vector (DLV) method. A set of load vectors, designated as damage locating vectors
(DLVs), were computed from the change in the flexibility matrix. 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.
Numerous methods, including mode shape curvatures based methods (Kim et al. 1997;
Garcia et al. 1998; Maeck and De Roeck 1999; Ho and Ewins 1999, 2000), neural network
techniques (Wu et al. 1992; Chen and Kim 1994; Masri et al. 1996; Luo and Hanagud
1997; Ni et al. 1999), genetic algorithms (Friswell et al. 1995; Mares and Surace 1996;
Carlin and Garcia 1996; Zimmerman et al. 1997), etc., have also been employed by
researchers in the field of damage detection.

8
2.2 Smart Sensor Technology

Another important research area pertaining to SHM that has attracted significant
attentions in the past decade is the smart sensor technology. Indeed, a National Research
Council report recently noted that the use of networked systems of embedded computers
and sensors throughout society could well dwarf all previous milestones in the
information revolution (National Research Council Computer Science and
Telecommunications Board 2002). Smart sensors to date have four important
characteristics: (1) an on-board microprocessor, (2) wireless communication, (3) a small
size, and (4) a low cost. Smart sensing technology may be the only way to fulfill the vision
of SHM of civil engineering structures using a densely distributed sensor network. Some
of the recently developed smart senors will be reviewed in this section.
Straser and Kiremidjian (1996, 1998) proposed a wireless modular monitoring system
(WiMMS) for SHM of civil engineering structures using smart sensor technology. The
proposed network provides ease of installation, low cost, portability, and broad
functionality. The sensor unit consists of a microprocessor, radio modem, data storage,
and batteries. To reduce the battery consumption, the smart sensor can be either in the
waiting mode or the operation mode. Agre et al. (1999) presented a prototype smart sensor
called “AWAIRS I”. This sensor supports bidirectional, peer-to-peer communications with
a small number of neighbors. Brooks (1999) emphasized the importance of the sensor’s
computational capacity and defined the fourth-generation sensors as having a number of
attributes: bi-directional command and data communication, all digital transmission, local
digital processing, preprogramming decision algorithms, user-defined algorithms, internal
self-verification and diagnosis, compensation algorithms, on-board storage and extensible
sensor object models. Lynch et al. (2001) demonstrated a proof-of-concept smart sensor
that uses a standard integrated circuit component. Key features of the unit include wireless
communications, high-resolution 16-bit digital conversion of interfaced sensors, and a
powerful computational core that can perform various data interrogation techniques in
near real-time. The sensor unit was validated through various controlled experiments in
the laboratory. Numerous researchers, including Maser et al. (1997), Mitchell et al. (1999,
2001), and Liu et al. (2001), also contributed to this research area.
While substantial research has been conducted, the above-mentioned smart sensor
systems are of a proprietary nature. To advance the technology more efficiently, an open
hardware/software platform is needed.
The Smart Dust project supported by the US Defense Advanced Research Projects
Agency (DARPA) has the ultimate objective of creating massively distributed sensor
networks, which consist of hundreds or thousands sensor nodes; these nodes have been
termed as Smart Dust or Motes. The goal is to have small dust that is fully autonomous
and is cubic millimeter in size (http://www.darpa.mil/). The first Smart Dust, termed
COTS Dust (Hollar 2000), incorporated communications, processing, sensors, and
batteries into a package about a cubic inch in size.
As part of the Smart Dust project, researchers in University of California at Berkeley
have developed an open hardware/software platform for smart sensing research. The
generation of Motes following COTS Dust was the Rene, which was developed in the

9
summer of 2000. This version of the Berkeley Mote had 8KB of programming memory
and could be reprogrammed over its radio link. Measured data could be communicated
over a range of about 15 feet. The Mica node, which increases the radio transmission rate
to 40KB per second, was subsequently developed (Hill and Culler 2002). It contains the
same expansion bus as the Rene node allowing it to utilize all existing sensor boards. The
programming memory and the data storage memory were also improved to 128KB and
4KB, respectively. The Mica has a transmission range over 100 feet, and has been
proposed for applications ranging from military vehicle tracking to remote environmental
monitoring. a series of development has been carried out to improve Mica platform that
results in Mica2, Mica2Dot, and MicaZ (Crossbow Technology Inc., 2004).
In the fall of 2002, the Spec node that is resulted from analyzing the Mica platform
was designed. The size of the Spec node was only 2.5x2.5 mm (see Fig. 2.1). The CPU,
memory, and RF transceiver are all integrated into a single piece of silicon. It includes an
ultra-low power transmitter that drastically reduces overall power consumption. The Spec
node also supports the multi-hop mesh networking protocol.
The latest generation of the of Berkeley Mote is Telos platform, that seeks to achieve
three major goals: ultra-lower power operation compared with previous Mote generations,
easy to use, and robust hardware and software implementation (Polastre et al. 2005).
While Spec node provides significant advantages in power consumption, it offers limited
interface flexibility. Telos not only achieved the goal of ultra-lower power operation, its
integrated design also allows researchers to develop more robust systems.
The hardware design for the Mote can be downloaded from the Berkeley web site
(http://www.tinyos.net/scoop/special/hardware/) and the components needed for this
design can be purchased from Crossbow Inc. (http://www.xbow.com/).

Figure 2.1: Spec node.

10
 Tiny OS is a component based operating environment designed for the Berkeley Mote.
More specifically, it is designed to support the concurrency intensive operations required
by networked sensors with minimal hardware requirements (http://today.cs.berkeley.edu/
tos/). The autonomous characteristic of the smart sensor can be realized by developing
programs using Tiny OS and then running these program on the on-board microprocessor.
This software is open source.
Intel Inc. announced the development of the Intel-Mote platform (Kling 2003). The
objective of the Intel-Mote is to create a new platform for smart sensing that provides a
high level of integration as well as low-power operation and small physical size. This new
platform will fully support Tiny OS. The Berkeley-Mote and Intel-Mote platforms are
expected to be an important impetus for smart sensing software and hardware
developments.

2.3 Summary

This chapter reviewed some of the vibration-based damage detection methods, as well
as recent developments in smart sensor technology. Vibration-based damage detection
methods are important because there is no requirement of a prior knowledge of damage
location. In addition, damage covered by the non-structural elements can also be detected.
Another important research area is the fast-growing smart sensor technology, which is
expected to have a significant impact on SHM for civil infrastructure. Development of the
vibration-based SHM methods that mesh well with the smart sensor technology is highly
desired.

11
 CHAPTER 3

EXPERIMENTAL VERIFICATION OF THE

DAMAGE LOCATING VECTOR METHOD

Health monitoring methods based on the flexibility matrix have recently been shown
promising. In particular, Bernal (2002) proposed a flexibility-based damage localization
method, the damage locating vector (DLV) method. This method has drawn considerable
attention recently; however, only numerical validation has been previously provided.
Experimental verification needs to be conducted to assess the efficacy of this method in
practice.
Following the motivation for flexibility-matrix-based methods, the concept of the
DLV method is introduced. Construction of the flexibility matrix for proportionally
damped structures employing forced vibration is then provided. Finally, details of the
experimental setup and experimental verification results are presented (Gao et al. 2004,
2007).

3.1 Motivation for Flexibility-Based Approach

Because an inverse relationship exists between the flexibility matrix and the square of
the modal frequencies, the flexibility matrix is frequently insensitive to high frequency
modes, which are typically quite difficult to determine experimentally. This unique
characteristic allows the use of a small number of truncated modes to construct a
reasonably accurate representation of the flexibility matrix (Gao and Spencer 2002).
To better understand this point, consider the response of a structure described by the
following linear equations of motion
M x·· + C d x· + K x = f (3.1)

where M = mass matrix; C d = damping matrix; K = stiffness matrix; x = displacement

vector; and f = force vector. Assuming proportional damping, the orthogonality property
of the mode shapes with respect to the mass and stiffness matrix leads to
T T
M = ψ Mψ and K = ψ Kψ (3.2)

in which M = diagonal modal mass matrix; K = diagonal modal stiffness matrix; and ψ
= undamped arbitrarily normalized mode shapes. The square of the modal frequencies can
be expressed in a matrix sense as
–1
Λ = M K (3.3)
2
in which the square of the modal frequencies ω j are on the main diagonal of matrix Λ .
Combining Eqs. (3.2) and (3.3) gives

12
 T T
ψ Kψ – ψ MψΛ = 0 (3.4)

Let
T 1⁄2 1⁄2
v = ( ψ Mψ ) = M (3.5)

where v = diagonal matrix with the mass normalized indices on the main diagonal.
Substituting Eq. (3.5) into Eq. (3.4) and keeping in mind that v is a diagonal matrix, one
has
T 2 T T
ψ Kψ – v Λ = 0 or ψ Kψ = v Λv (3.6)

Therefore, the stiffness matrix can be obtained from Eq. (3.6) as

T –1 T –1
K = ( ψ ) v Λvψ (3.7)

From Eq. (3.5), one can get

–2 T T –2
( v ψ M )ψ = I and ψ ( Mψv ) = I (3.8)

Therefore
–1 –2 T T –1 –2
ψ = v ψ M and (ψ ) = Mψv (3.9)

Eq. (3.7) can be rewritten as

–1 –1 T
K = Mψv Λv ψ M (3.10)
–1
From the relationship between the stiffness and flexibility matrix F = K , the flexibility
matrix is derived from Eq. (3.7) as
–1 –1 –1 T
F = ( ψv )Λ ( ψv ) (3.11)

Equations (3.10) and (3.11) indicate the different influences of the various frequency
modes on the stiffness and flexibility matrices, respectively. The influence of the jth mode
2
on the stiffness matrix K increases with the square of the modal frequency ω j , whereas
–2
for the flexibility matrix F , the influence decreases with ω j .
To quantitatively see this effect, consider the 53-bar planar truss with 53 DOFs given
in Fig. 3.1. The truncated stiffness matrix K n at all DOFs, which contains the contribution
of the first n modes, can be derived from Eq. (3.10) and is written as
n

∑ Mψj vj
–2 2 T
Kn = ω j ψj M , n≤ N (3.12)
j=1

13
 14 x 0.4 m = 5.6 m

2 4 6 8 10 12 14 16 18 20 22 24 26
y 0.4 m
1 28
x 3 5 7 9 11 13 15 17 19 21 23 25 27

Figure 3.1: 53-bar planar truss.

in which N = number of DOFs of the structure. Two error norms are defined here to
measure the difference between the exact and the truncated stiffness matrices at all DOFs.
The first is the 2-norm given by the maximum singular value, designated by s , of the
difference matrix
K – Kn 2
= s { K – Kn } (3.13)

The second norm calculated is the Frobenius norm

N N 1⁄2

∑ ∑
2
K – Kn F
= ( K – K n ) ij (3.14)
i=1 j=1

Figure 3.2 shows these two error norms, which are normalized to have a value of 1.0
for n = 0 . As can be seen from the graph, nearly all of the modes are required to obtain a
reasonably accurate representation of the stiffness matrix. Because experimentally
obtaining the higher modes of a structure is often challenging, stiffness-matrix-based
damage detection strategies may be difficult to implement in practice.
Similarly, the truncated flexibility matrix F n can be derived from Eq. (3.11) as

2-Norm Frobenius Norm

1 1

0.8 0.8
K-K n F
2

K F
K 2
K-K n

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 10 20 30 40 50 0 10 20 30 40 50

Number of Modes Number of Modes

Figure 3.2: Normalized error in truncated stiffness

matrix at all DOFs.
14
 n

∑ ψj vj
–2 –2 T
Fn = ω j ψj , n≤ N (3.15)
j=1

The counterparts to the error norms defined in Eqs. (3.13) and (3.14) are then
F – Fn 2
= s { F – Fn } (3.16)

N N 1⁄2

∑ ∑ (F –
2
F – Fn F
= F n ) ij (3.17)
i = 1j = 1

These two norms are shown in Fig. 3.3, again normalized to have a value of 1.0 for
n = 0 . As can be seen, only a few modes are required to achieve good accuracy in the
flexibility matrix. These results indicate that the truncated flexibility matrix at measured
DOFs also can be accurately estimated using the first lower frequency modes.
Because of the high accuracy of the results for the truncated flexibility matrix and its
inverse relationship to the stiffness matrix, one might speculate that the truncated stiffness
matrix can be recovered by taking the inverse of the truncated flexibility matrix. To
investigate this possibility, consider again the planar truss in Fig. 3.1. For this truss, the
mode shapes at the DOFs in the y-direction at nodes [3, 5, 7, 9, 11] are computed from the
analytical model and used to construct the associated stiffness and flexibility matrices.
The truncated flexibility matrix for these DOFs can be obtained using Eq. (3.15) with
ψ j representing the mode shape at the measured DOFs. Denoting the exact flexibility
matrix at the measured DOFs as F m and the truncated one as F m, n , similar error norms
can be computed using Eqs. (3.16) and (3.17) and are shown in Fig. 3.4. These results, of
course, are essentially similar to those previously presented in Fig. 3.3 as the number of
DOFs does not change the accuracy of the computed flexibility coefficients.

1 1

2-Norm Frobenius Norm

0.8 0.8
F-Fn F
F-Fn 2

0.6 0.6
F F
F 2

0.4 0.4

0.2 0.2

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Number of Modes Number of Modes

Figure 3.3: Normalized error in truncated flexibility

matrix at all DOFs.

15
 1 1

2-Norm Frobenius Norm

0.8 0.8

F
2
Fm -Fm , n

Fm -Fm , n
0.6 0.6

F
Fm 2

Fm
0.4 0.4

0.2 0.2

0 0
0 10 20 30 40 50 0 10 20 30 40 50
Number of Modes Number of Modes

Figure 3.4: Normalized error in truncated

flexibility matrix at partial DOFs.

3 2
2-Norm 1.8 Frobenius Norm
2.5 1.6

F
2

2 1.4
Km -K m , n

Km -K m , n

F
2

1.2
Km

Km
V

1.5
V

V
1
0.8
V

1
0.6
0.5 0.4
0.2
0 0
15 20 30 40 50 15 20 30 40 50

Figure 3.5: Normalized error in condensed

truncated stiffness matrix.

The exact condensed stiffness matrices K ˆ and the truncated condensed stiffness
m
matrix Kˆ associated with the measured DOFs can be derived from the flexibility
m, n
matrices, i.e.,
ˆ = F–1
K and ˆ
K
–1
m m m, n = F m, n (3.18)

Similar norms for truncated condensed stiffness matrix can be computed from Eqs. (3.13)
and (3.14) and are shown in Fig. 3.5. As compared to the flexibility matrix, a larger
number of high frequency modes are needed to obtain an accurate estimation of the
condensed stiffness matrix.
While these results are problem dependent, they indicate significant potential for
practical application of flexibility-based damage detection approaches.

16
3.2 The DLV Method

3.2.1 General concepts

Bernal (2002) proposed a flexibility-based damage localization method, the DLV

method. This technique is based on determination of a special set of load vectors, the so-
called damage locating vectors (DLVs). These DLVs have the property that when they are
applied to the undamaged structure as static forces at the sensor locations, no stress is
produced in the damaged elements. This unique characteristic can be employed to localize
structural damage.
Assuming nominally linear structural behavior, the flexibility matrices at sensor
locations are constructed from measured data before and after damage and denoted as F u
and F d , respectively. Then, all of the linear-independent load vectors L are collected
which satisfy the following relationship
Fd L = Fu L or F Δ L = ( F d – F u )L = 0 (3.19)

This equation implies that the load vectors L produce the same displacements at the
sensor locations before and after damage. From the definition, the DLVs are seen to satisfy
Eq. (3.19); that is, because the DLVs induce no stress in the damaged structural elements,
the damage of those elements does not affect the displacements at the sensor locations.
Therefore, the DLVs are indeed the vectors in L .
To better understand the general concept behind the DLV method, consider the truss
shown in Fig. 3.6. Note that the element number has a circle around it. In this structure,
assume that three sensors are installed in the y-direction at nodes [2, 3, 4]. Therefore, the
dimension of the flexibility matrix at the sensor locations is 3 × 3 and the DLVs are 3 × 1
vectors. For the case when element 8 is damaged, the DLVs are shown in Fig. 3.7. As can
be seen, both DLVs have a zero force in the y-direction at node 3. As a result, there will be
no stress in element 8 under either of the DLVs, which indicates that element 8 is the
damage location.
However, damage in elements for which no stresses are induced by any combination
of loads applied at the sensor locations can not be determined by the DLV method. To
understand this situation more clearly, refer to Fig. 3.6 again. If the sensor at node 3 is
moved to node 7, then the damage in element 8 cannot be located by the DLV method.

4 x 4.5 m = 18 m
6 5 7 6 8
10 11 12 13
7 8 9 1.5 m
y 1 1 2 3 4 5
2 3 4

x
Accelerometers in y-direction

Figure 3.6: 13-bar planar truss.

17
 6 7 8
8
y 1 5
2 3 4

x
DLV1
f1 f2 = 0 f3

DLV2
f1 f2 = 0 f3

Figure 3.7: Illustration of the DLVs.

This result is due to the fact that no matter what combination of the loads at nodes [2, 4,
7], there is always zero stress in element 8. Therefore, it is important to point out that, for
a specific sensor configuration and damage extent, there can be elements that are not
damaged but also have small stresses under DLVs. The DLV method identifies a small
group of damaged elements that contains the actually damaged elements (Bernal 2002).

3.2.2 Calculation of DLVs

To calculate the load vectors L , the singular value decomposition (SVD) is

employed. The SVD of the flexibility difference matrix F Δ leads to

T S1 0 T
F Δ = USV = U 1 U 0 V1 V0 (3.20)
0 0

Recall from the properties of the SVD

T
V1 V0 V1 V0 = I (3.21)

Equation (3.20) can be rewritten as

FΔ V1 FΔ V0 = U1 S1 0 (3.22)

From Eq. (3.22), one obtains

FΔ V0 = 0 (3.23)

Eqs. (3.19) and (3.23) indicate that L = V 0 , i.e., DLVs can be obtained from the SVD of
the difference matrix F Δ .
In Eq. (3.20), because of noise and computational errors, the singular values
corresponding to V 0 are generally not exactly zero. To select the DLVs from the SVD of
the matrix F Δ , an index svn was proposed by Bernal (2002) and defined as

18
 2
si ci
svn i = -------------------------
- (3.24)
2
max ( s k c k )
k

in which s i = ith singular value of the matrix F Δ ; c i = constant that is used to normalize
the maximum stress in the structural element, which is induced by the static load c i V i , to
have a value of one; and V i = right singular vector of the matrix F Δ . Bernal (2002)
suggested that a value of 0.2 might be a good cutoff for the index svn; that is to say, the
singular vectors corresponding to the singular values having the index svn smaller than 0.2
can be selected as the DLVs.
Each of the DLVs is then applied to an undamaged analytical model of the structure.
The stress in each structural element is calculated and a normalized cumulative stress is
obtained. If an element has zero normalized cumulative stress, then this element is a
possible candidate of damage. The normalized cumulative stress σ j for the jth element is
defined as
m
σj σ ij
σ j = ----------------------
max ( σ k )
where σj = ∑ abs ⎛⎝ -----------------------
max ( σ ik )⎠
-⎞ (3.25)
k i=1 k

In Eq. (3.25), σ j = cumulative stress in the jth element; σ ij = stress in the jth element
induced by the ith DLV; and m = number of DLVs. In practice, the normalized cumulative
stresses induced by the DLVs in the damaged elements may not be exactly zero due to
noise and uncertainties. Reasonable thresholds should be chosen to select the damaged
elements (Bernal 2002).

3.2.3 Numerical example

To quantitatively illustrate the idea of the DLV method, a simple numerical example
is given. The 13-bar planar truss, shown in Fig. 3.6 is selected as the test structure. In this
structure, each bar has a pipe cross section with an outer diameter of 1.0 cm, and a wall
11 2
thickness of 0.2 cm. The elastic modulus of the material is 2 × 10 N ⁄ m and the mass
3 3
density is 7.86 × 10 kg ⁄ m . The bars are connected at pinned joints. The total length of
this truss is 18 m, with 4.5 m in each bay, and the height of the truss is 1.5 m. There are
two supports in this truss structure — a pin support at the left end and a roller support at
the right end of the lower chord. The roller support at the right end is constrained in the y-
direction (vertical direction).
A finite element model consisting of 13 bars, 8 nodes, and 13 DOFs is developed
using Matlab. In this numerical example, sensor locations are in the y-direction of nodes
[2, 3, 4]. Modal parameters of the lowest three modes before and after damage are
obtained from the finite element model. The case when element 12 (i.e., node 4–7) has a
20% stiffness reduction is studied to illustrate the idea of the DLV method.

19
 After the modal parameters are obtained before and after damage, the corresponding
flexibility matrices at the sensor locations with a dimension of 3 × 3 are constructed using
Eq. (3.15). The SVD is then applied to obtain the DLVs. To select the DLVs from the
results of the SVD, index svn is calculated according to Eq. (3.24). Results are shown in
Table 3.1. Because the 2nd and 3rd singular values in Table 3.1 have the index svn smaller
than 0.2, the corresponding singular vectors are selected as the DLVs. These two DLVs are
then applied as static forces to the undamaged structure at the sensor locations, i.e., in the
y-direction of nodes [2, 3, 4]. The normalized cumulative stress for each element is
calculated using Eq. (3.25). The results are shown in Fig. 3.8. As can be seen, element 12
has a small normalized cumulative stress, which indicates that element 12 is damaged.

Table 3.1: Results for index svn.

1st singular 2nd singular 3rd singular

value value value

svn 1.0 0.0341 0.0081

3.3 Construction of Flexibility Matrix Using Limited Sensor

Information

As shown in the previous section, the flexibility matrix needs to be constructed from
the measurement data to implement the DLV method. When the input excitation is
measured, there is a least one co-located sensor and actuator pair, and the measured
responses are either displacements or velocities, the experimental data can be used to
obtain the mass normalized mode shapes (Alvin and Park 1994). Then the flexibility
matrix can be computed. Bernal (2000) presented an approach to construct the flexibility
matrix at the sensor locations that can employ displacement, velocity, and acceleration
responses. This approach will be reviewed in this section.
Normalized Cumulative Stress σ j

0.8

0.6

0.4

0.2

0
1 2 3 4 5 6 7 8 9 10 11 12 13
Element Number

Figure 3.8: Normalized cumulative stress when

element 12 is damaged.

20
 First, a state space representation of the structure is obtained using a system
identification algorithm, such as the Eigensystem Realization Algorithm (ERA) (Juang
and Pappa 1985; Juang 1994). By converting the discrete time state space representation
identified from the measured data to continuous time, one obtains
z· = Az + Bu
(3.26)
y = Cz + Du

where A = system matrix; B = input influence matrix; C = output influence matrix; D =

direct transmission matrix; z = state vector; u = input excitation vector; and y = output
vector. Taking the Fourier Transform of Eq. (3.26) yields the following relationship
between the inputs and outputs
–1
y ( ω ) = { C [ I ⋅ iω – A ] B + D }u ( ω ) (3.27)

Based on the type of the sensors used in the experiment, the displacement vector y D ( ω )
can then be expressed as
1 –1
y D ( ω ) = ------------p- { C [ I ⋅ iω – A ] B + D }u ( ω ) (3.28)
( iω )

where p = 0, 1, and 2 when measured outputs are displacement, velocity, and acceleration,
respectively. If we note that the flexibility matrix relates the inputs to the outputs at
ω = 0 , the flexibility matrix can be obtained as
1 –1 –( p + 1 )
F f = lim ------------p- { C [ I ⋅ iω – A ] B + D } = – C A B (3.29)
ω → 0 ( iω )

Therefore, the ith column of F f represents the displacements of the structure at the sensor
locations due to a unit force being applied at the ith excitation location, u i .
Let’s denote the flexibility matrix at the sensor locations as F m , then the jth column
of F m represents the displacements of the structure at the sensor locations due to a unit
force being applied at the jth sensor location. If any of the inputs are co-located with
sensors, then the corresponding columns in the matrices F f and F m will be equal.
Defining two Boolean matrices q f and q m which pick out these columns from F f and
F m , respectively, we have
–( p + 1 )
Fm qm = Ff qf and Fm qm = –C A Bq f (3.30)

Expressing the flexibility F m and the system matrix A in terms of their eigenvalues and
eigenvectors gives
–1 –2 –1 T –( p + 1 ) –1
( ψ m v )ω ( ψ m v ) q m = – C φλ φ Bq f (3.31)

21
where ψ m = undamped mode shapes at the measured DOFs; v = matrix of mass
normalized indices; λ and φ = eigenvalue and eigenvector matrices of the system matrix
A , respectively. For classical damping, a mode-by-mode equality holds and Eq. (3.31) can
be written as
–2 T –( p + 1 ) –1
[ ψ m ] j ( ω j v j ) [ ψ m ] j q m = – 2 ⋅ real ( Cφ j λ j φ j Bq f ) (3.32)

In Eq. (3.32), only v j , the diagonal terms of v , are unknown. After v j is calculated (i.e.,
the diagonal matrix v is obtained), the flexibility matrix at the sensor locations can be
constructed from the relationship
–1 –2 –1 T
F m = ( ψ m v )ω ( ψ m v ) (3.33)

The DLV method can then be applied to localize damage in the structure. A flow chart
for implementing the DLV method for the case when inputs are measured is shown in Fig.
3.9.

3.4 Experimental Verification

The DLV method is experimentally verified using a three-dimensional truss structure

(Gao et al. 2004, 2007). In this section, the experimental setup is first described and the
experimental results are then presented.

3.4.1 Experimental setup

This 5.6 m long, three-dimensional truss structure (see Fig. 3.10) is tested at the Smart
Structures Technology Laboratory (SSTL) of the University of Illinois at Urbana-
Champaign (http://cee.uiuc.edu/sstl/). It is originally designed at the Structural Dynamics
and Control/Earthquake Engineering Lab in the University of Notre Dame (Clayton and
Spencer 2001). The length of each bay of the truss is 0.4 m on each side. The truss sits on
two rigid supports. One end of the truss is pinned to the support, and the other is roller-
supported (see Fig. 3.11). 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 1.09 cm and an outer
diameter of 1.71 cm. The joints of the elements are specially designed so that the truss
member can be easily removed or replaced to simulate damage without dissembling the
entire structure. A detailed picture of the joint is shown in Fig. 3.12. As can be seen, the
truss member can be removed by unscrewing the collars at the both ends of the member
towards the joint. On the other hand, if the collar is screwed away from the joint, this
member can be easily installed.
The truss is excited vertically by a Ling Dynamic Systems permanent magnetic V408
shaker (see Fig. 3.13) that can generate a maximum force of 20 lbs with a dynamic
performance ranging from 5 Hz to 9000 Hz. A band-limited white noise is sent from the

22
 measured data (inputs and outputs)

transfer functions

impulse responses
extraction of
flexibility from system realization algorithm (e.g. ERA)
measurements
system matrices (A, B, C) and
modal parameters (ω, φ)

flexibility matrices (Fu and Fd )

FΔ = Fd - Fu

singular value decomposition of FΔ

the DLV method
damage locating vectors (DLVs)

obtain normalized cumulative

stress under DLVs

damage

Figure 3.9: Implementation of the DLV method.

Outer Vertical Panel

Vertical
Vertical Element 112 Element 112

Longitudinal Element 82
Longitudinal Element 82

Figure 3.10: Three-dimensional 5.6 m long truss structure.

23
 Figure 3.11: Pin and roller ends.

Figure 3.12: Details of the joint.

computer to the shaker to excite the truss structure up to 200 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 monitor
and measure the input to the structure. This load cell has a sensitivity of 50 mV/lb, a
frequency range of 0.001 to 36000 Hz, and a measurement range of 100 lbs in both
compression and tension. A picture depicting the load cell and the connection between the
steel rod to the joint node is shown in Fig. 3.14.
PCB high sensitivity piezotronics accelerometers (model 353B33, see top left of Fig.
3.15) are used to measure the response of the structure. These accelerometers have a

24
 Figure 3.13: Magnetic shaker.

Figure 3.14: Load cell.

Figure 3.15: Accelerometer and magnetic base.

25
sensitivity of 100 mV/g, a frequency range of 1 to 4000 Hz, and a measurement range of
± 50g . These accelerometers are mounted on the structure through a magnetic base (see
bottom left of Fig. 3.15). These magnetic bases significantly facilitate the relocation of the
accelerometers during testing.
Using a limited number of sensors to monitor all members in a complex structure
might be difficult. In this experimental verification, the 53 elements in the outer panel (see
Fig. 3.16), which are elements 14 through 25 and 79 through 119, are monitored using 13
accelerometers, which are installed vertically at the joints of the lower chord with
magnetic bases. A picture of the accelerometer mounted on the structure is shown at the
right side of Fig. 3.15.
Two four-channel 20-42 Siglab spectrum analyzers (see Fig. 3.17) are used to drive
the shaker and measure the inputs and responses. The spectrum analyzers are
synchronized to measure eight channels of data simultaneously, one for the load cell and
the other seven channels for the accelerometers. Due to the fact that only eight channels

2 14 4 15 6 16 8 17 10 18 12 19 14 20 16 21 18 22 20 23 22 24 24 25 26

93 95 97 99 101 103 105 107 109 111 113 115 117 119
94 96 98 100 102 104 106 108 110 112 114 116 118

1 79 3 80 5 81 7 82 9 83 11 84 13 85 15 86 17 87 19 88 21 89 23 90 25 91 27 92 28

y
shaker location
x

13 accelerometers in the vertical direction

Figure 3.16: Sketch of the outer vertical panel

(elements 14 through 25 and 79 through 119).

Figure 3.17: 20-42 Siglab spectrum analyzer.

26
are available, these thirteen accelerations are measured in two sequential experiments. In
this way, thirteen transfer functions between the shaker and different accelerations can still
be measured for system identification, by which mode shapes at these thirteen DOFs can
be obtained. The shaker is connected to one of the joints at the bottom, so there is one co-
located sensor and actuator pair.
Eight-pole elliptical anti-aliasing filters are employed for both the input and output
measurements with a cutoff frequency of 200 Hz. The sampling rate for all the
measurements is 512 Hz. Measured transfer functions, ranging from 0 to 200 Hz, have a
length of 8192 spectra lines with the frequency resolution as 0.0625 Hz. The Hanning
window with a 50% overlap is used to compute the transfer functions in the Siglab
spectral analyzer. The final transfer function is the average of 30 realizations.
Other equipment used during the testing consist of:
• Amplifier: a Sony STR-D315 amplifier (see Fig. 3.18) is used to magnify the input
voltage from the Siglab spectrum analyzer to drive the magnetic shaker.
• Signal conditioner: PCB 4-channel signal conditioners (model 441B104) with AC
power supply (model 441A101), as well as single-channel battery powered signal
conditioners (model 480E09) have been used in this experiment.

3.4.2 Experimental results

Different damage scenarios have been studied. Damage is simulated by replacing the
original member with one having a 52.7% cross section reduction. Results from two of
these damage cases (see Table 3.2) in this experimental verification will be presented in
the remainder of the chapter.
The use of a single cut-off value of svn index (Bernal and Gunes 2004) to select the
DLVs has difficulties in the detection of some of the damage cases. For this truss structure
and current sensor configuration, a total of five DLVs corresponding to the smallest
singular values have been found working well and therefore are used to identify damage
for different damage scenarios.

Figure 3.18: Amplifier.

27
 Table 3.2: Two damage cases.

Cross section Equivalent axial

Cases Damaged element
reduction stiffness reduction

Case 1 longitudinal element 82 52.7% 44%

Case 2 vertical element 112 52.7% 43.5%

Case 1

In this case, longitudinal element 82 (see Figs. 3.10 and 3.16) in the lower chord is
replaced by a tube with a 52.7% cross section reduction.
The transfer functions are measured first. Four typical experimentally-measured
transfer functions, which are associate the input excitation with measurements from four
accelerometers at nodes [2, 5, 8, 11], are shown in Fig. 3.19. The modal parameters are
then obtained from the measured transfer functions before and after damage using the
ERA method. Here, the first five dominant natural frequencies, which are numbered in
Fig. 3.19, are identified. The corresponding mode shapes are also extracted as shown in
Fig. 3.20. As noted, there is very little change in the frequencies and mode shapes. Direct
comparison between the frequencies and mode shapes to detect damage is difficult if not
impossible in this case. Herein, the DLV method is employed to detect the damage.
Once the modal parameters before and after damaged are obtained, the flexibility
matrix at the sensor locations can be constructed following the procedure described above.
The DLVs are then computed from the difference matrix F Δ using Eq. (3.20), and applied
as vertical loads at the sensor locations to the undamaged analytical model for static
computation. The normalized cumulative stress can then be obtained and employed to
locate the damage in the structure.

40 2 5
20
1 3 4
Amplitude (dB)

- 20

- 40

- 60 node 2 node 5
node 8 node 11
- 80
0 20 40 60 80 100 120
Frequency (Hz)

Figure 3.19: Experimentally measured transfer functions.

28
 21.3626 Hz 21.3403 Hz

41.5719 Hz 41.5034 Hz

69.3878 Hz 69.3932 Hz

94.6607 Hz 94.5922 Hz

120.1610 Hz 119.8784 Hz

Undamaged Modes Damaged Modes

Figure 3.20: Modal parameters identified from

experimental data.
j
Normalized Cumulative Stress σ
_

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
15 20 25 80 85 90 95 100 105 110 115 120

Element Number

Figure 3.21: Normalized cumulative stress when

element 82 is damaged.

The results of the normalized cumulative stress are shown in Fig 3.21. As can be seen,
the normalized cumulative stress for element 82 is considerably smaller than others
elements, so element 82 is correctly identified as the damage candidate. However,
elements 17 and 108, which are not a damaged element, also have a small value of
normalized cumulative stress. Therefore, the DLV method identifies a small group of
potentially damaged candidates that contain the damaged elements. However, as will be
shown later on, when number of sensors increases, the likelihood of false identification of
damaged elements is significantly reduced.

29
 More specifically, the reason why element 17 has a small cumulative stress is
apparent by looking at the force balance of the truss structure at the horizontal direction if
we cut the truss through 82, 100, and 17, and consider the right portion of the truss. The
current sensor configuration indicates that the DLVs are only applied in the vertical
directions at the lower chord joints in the outer vertical panel; the structural member force
which is not in the outer vertical panel (see Fig. 3.10) are indeed very small. Therefore, if
we consider the horizontal force balance of the right portion of the truss structure, the axial
force in elements 82 and 17 have to be equal and opposite (note that the normalized
cumulative stress defined in Eq. (3.25) is always positive). Therefore, if either of these
two elements has a small stress, so does the other. A detailed discussion on the theoretical
limitations of the spatial resolution of damage that can be obtained from examination of
changes in flexibility for a given sensor set is presented in Bernal (2002).

Case 2

In this case, the vertical element, element 112, is replaced by one with a 52.7% cross
section reduction instead of a longitudinal element as presented in Case 1. The results are
shown in Fig. 3.22, in which elements 107, 112, and 113 have a small normalized
cumulative stress compared with others. Obviously, with the current sensor configuration
and resulting DLVs, if either one of the two elements (112 and 113) has a small axial
force, so does the other. This can be seen from the force balance in the vertical direction of
the joint 20 (see Fig. 3.16). Therefore, the DLV method successfully detects element 112
as a possibly damaged location.

3.5 Summary

In this chapter, the flexibility-based damage localization method, the DLV method,
was experimentally verified employing a 5.6 m long truss structure. Motivation of the
flexibility-based damage detection method was first discussed. The numerical example
shows that a reasonably accurate flexibility matrix can be constructed using only a limited
j
Normalized Cumulative Stress σ
_

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
15 20 25 80 85 90 95 100 105 110 115 120

Element Number

Figure 3.22: Normalized cumulative stress when

element 112 is damaged.

30
number of lower frequency modes, which is a desirable characteristic for vibration-based
damage detection methods. The concepts of the DLV method were then introduced and
numerical examples were provided to better understand this method.
To assess the efficacy of the DLV method in practice, this method was successfully
verified using experimental data. Experimental results show that the change of modal
properties subjected to a 52.7% cross section reduction (equivalent to around a 44% axial
stiffness reduction) of a single member is very small. Direct comparison of the modal
properties to detect damage is difficult if not impossible for this truss structure. By using
the flexibility-based DLV method, damage in this truss can be correctly located using only
a limited number of sensors and truncated modes.

31
 CHAPTER 4

CONTINUOUS STRUCTURAL HEALTH MONITORING

EMPLOYING AMBIENT VIBRATION

Numerous vibration-based damage detection methods have been proposed recently.

Although these methods can be effective, they are not yet easily employed for continuous,
online SHM. Most of these methods require measurement of the input excitation to be
implemented. However, in numerous cases, externally exciting the structure (e.g., with an
impactor or a rotating imbalance vibrator) is difficult, especially for large complicated
civil infrastructure systems. In addition, Peeters et al. (2001) pointed out that the price of a
shaker and the additional manpower to install it on a structure makes it cost-prohibitive
and suggested that only ambient vibration should be employed for long-term, continuous
monitoring. SHM strategies employing ambient vibration that are suitable for continuous
online diagnosis are currently limited.
In this chapter, an approach is proposed for continuous online SHM that is an
extension of the DLV method (Gao and Spencer 2005a, 2006). The essence of the
proposed approach is to construct an approximate flexibility matrix for the damaged
structure utilizing the modal normalization constants from the undamaged structure. This
extended DLV method can then be applied for online damage diagnosis employing
ambient vibration. Following an overview of the construction of the flexibility matrix
based on both forced and ambient vibration, extension of the DLV method for the
continuous online SHM is proposed and then numerically validated using a 14-bay planar
truss structure.

4.1 Construction of the Flexibility Matrix

As shown in Chapter 3, the flexibility matrices before and after damage need to be
constructed to implement the DLV method. Depending on whether or not the input
excitations are measured, different formulations are needed to construct the flexibility
matrix from the measured data. For the case when the inputs are measured and there is at
least one co-located sensor and actuator pair, the flexibility matrix at the sensor locations
can be constructed for structures with general viscous damping. For the ambient vibration
case (i.e. the input excitation is not measured), a formulation is utilized that requires the
structure to have classical damping. Both approaches are reviewed in this section.

4.1.1 Formulation of the flexibility matrix based on forced vibration

Consider the homogeneous equations of motion for a linear structure

Mx·· + C d x· + Kx = 0 (4.1)

where M, C d , and K = mass, damping, and stiffness matrix, respectively; and x =

displacement vector. A state space representation of the structure can be expressed as

32
 ·
Eη = Gη (4.2)

Cd M
E = G = –K 0 η = x (4.3)
M 0 0 M x·

Solving the standard eigenproblem of Eq. (4.2) yields the eigenvalues Λ E and
eigenvectors Φ E

ΛE =
Λ 0
ΦE = ϕ ϕ∗ (4.4)
0 Λ∗ ϕΛ ϕ∗ Λ∗

where ∗ represents complex conjugate. Orthogonality of the eigenvectors yields

–1 T –1 T
D g = Φ E GΦ E G = ΦE Dg ΦE (4.5)

From Eq. (4.5), one obtains

–1 T
K = – ΨD g Ψ (4.6)

T
F = – ΨD g Ψ (4.7)

where Ψ = [ ϕ ϕ∗ ] , arbitrarily normalized complex mode shapes;

D g = diag ( [ d 1, d 2, …, d j, … ] ) ; and d j = modal normalization constant.

Determination of modal normalization constants

When the input is measured and there is at least one co-located sensor and actuator
pair, the experimental data can be used to obtain the modal normalization constant d j for
structures with general viscous damping (Bernal and Gunes 2004).
First, the flexibility matrix related the input locations to the output locations is
constructed from measured data following the same procedure presented in section 3.3
and expressed as
–( p + 1 )
Ff = –C A B (4.8)

where all the variables have been defined in section 3.3. Again, expressing the system
matrix A in terms of its eigenvalues and eigenvectors yields
–( p + 1 ) –1 –( p + 1 ) –1
F f = – C φλ φ B = –Ψm λ φ B (4.9)

where Ψ m = Cφ and Ψ m = complex mode shapes at sensor locations.

33
 Similar as the derivation in section 3.3, we denote the flexibility matrix at the sensor
locations as F m . If any of the inputs are co-located with sensors, then the corresponding
columns in the matrices F f and F m will be equal. Defining two Boolean matrices q f and
q m which pick out these columns from F f and F m , respectively, we have
T –( p + 1 ) –1
–Ψm Dg Ψm qm = –Ψm λ φ Bq f (4.10)

The modal normalization constant d j can then be solved from Eq. (4.10) as
–( p + 1 ) T T –1
dj = λj φ j Bq f [ diag ( Ψ m, j q m ) ] (4.11)

T –1 T T
where φ j = jth row of matrix φ ; and Ψ m, j = jth row of matrix Ψ m . When there is more
than one co-located sensor and actuator pair, multiple estimations of d j will be obtained.
Bernal and Gunes (2004) suggested that the d j corresponding to the component in vector
T
Ψ m, j q m with the largest magnitude might be used.

4.1.2 Formulation of the flexibility matrix based on ambient vibration

Consider again Eq. (4.1) for a linear structure. Recall that for classical damping, the
stiffness matrix can be expressed as
T T –1 –1
Φ KΦ = Λ and K = ( Φ ) ΛΦ (4.12)

Therefore the flexibility matrix takes the form

–1 T –1 T
F = ΦΛ Φ = ( ψα )Λ ( ψα ) (4.13)

where Φ = mass normalized mode shapes and Φ = ψα ; ψ = undamped arbitrarily

normalized mode shapes; Λ and ψ = solutions of the eigenproblem Kψ = MψΛ ;
α = diag ( [ α 1, α 2, …, α j, … ] ) ; and α j = modal normalization constant, which can be
obtained from the output measurements for structures with classical damping. When input
excitations are not measured, Eq. (4.13) can be used to construct the flexibility matrix.

Determination of modal normalization constants

For the ambient vibration case, the approach presented in the previous section can’t
be employed to obtain the modal normalization constant α j because of the lack of a co-
located sensor and actuator pair. To overcome this difficulty, researchers (Parloo etc. 2001,
2002; Brinker and Pade 2003; Bernal 2004) have proposed methods based on testing of
the structure with known modification masses. The mass perturbation method provided by
Bernal (2004) for a linear structure with proportional damping will be presented in the
remainder of this section.
The mass matrix of the modified structure can be expressed as
M 1 = M 0 + ΔM (4.14)

34
in which M 0 = mass matrix of the original structure; M 1 = mass matrix of the modified
structure; ΔM = matrix describing the mass perturbation. The eigenvalue problem for the
modified structure is
Kψ 1, j = λ 1, j ( M 0 + ΔM )ψ 1, j (4.15)

where λ 1, j = jth eigenvalue of the modified structure; and ψ 1, j = jth eigenvector of the
modified structure. The corresponding eigenvalues and eigenvectors for the original
structure are λ 0, j and ψ 0, j , respectively. The mode shape ψ 1, j can be written as
ψ 1, j = ψ 0 q j + N ( ψ 0 )g j (4.16)

where ψ 0 = [ ψ 0, 1, ψ 0, 2, …, ψ 0, j, … ] ; N ( ψ 0 ) = column null space of ψ 0 ; q j and g j =

coefficient vectors.
T
Substituting Eq. (4.16) into Eq. (4.15) and multiplying both sides by ψ 0 yields
2
α ( Rj + εj ) = Ej (4.17)

where
T
R j = λ 1, j ψ 0 ΔMψ 1, j
T T
ε j = λ 1, j ψ 0 M 0 N ( ψ 0 )g j – ψ 0 KN ( ψ 0 )g j (4.18)

E j = Λ 0 q j – λ 1, j q j

T –1 T
with Λ 0 = diag [ λ 0, 1, λ 0, 2, …, λ 0, j, … ] and q j = ( ψ 0 ψ 0 ) ψ 0 ψ 1, j . By neglecting the
error term ε j , Eq. (4.17) can be rewritten as
2 T
α λ 1, j ψ 0 ΔMψ 1, j = Λ 0 q j – λ 1, j q j (4.19)

which can be de-coupled to solve for the unknown α i as

2 λ 0, i – λ 1, j q ij
α i = ------------------------ -----------------------------
- (4.20)
λ 1, j ψ T ΔMψ
0, i 1, j

Equation (4.20) indicates that there is one set of [ α 1, α 2, …, α i, … ] for each ψ 1, j

(i.e., there are various estimations for different ψ 1, j ). Bernal (2004) pointed out that the
most accurate estimation for the normalization index α i is when i = j in Eq. (4.20).
The modal parameters for both the original and modified structure can be obtained
employing the Natural Excitation Technique (NExT) (James et al. 1993) in conjunction
with the ERA method. To apply the NExT, the first step is to measure responses. Then
auto- and cross-spectral density functions of these responses can be calculated.
Afterwards, the inverse Fourier transform is applied to obtain the auto- and cross-
correlation functions, which are in turn used by the ERA to extract the modal parameters.

35
To obtain expected modal parameters, the reference output selected in calculating
correlation functions should include the information on all modes of interest (Dyke et al.
2000).

4.2 Extension of the DLV Method for Online Damage Diagnosis

The DLV method is not easily employed for continuous online damage diagnosis.
Eqs. (4.7) and (4.13) show that the modal normalization constants d j or α j need to be
obtained to construct the flexibility matrix. As suggested in the derivation of Eqs. (4.11)
and (4.20), a certain degree of interruption of structural operation will be needed to obtain
these normalization constants from measured data. The method based on forced vibration
requires employing an external exciter, e.g. an impactor or a rotating imbalance vibrator,
to shake the structure. The method based on ambient vibration requires adding mass to the
structure to conduct dynamic testing. These methods can be used to constructed the
undamaged flexibility matrix. However, computing the damaged flexibility matrix by one
of these methods each time the health of the structure is assessed is intractable. An
approach which can take the advantage of the DLV method and is more suitable for
continuous online monitoring is highly desired.
Gao and Spencer (2005a, 2006) proposed an extension of the DLV method to handle
the continuous online damage diagnosis employing ambient vibration and will be
presented in the remainder of this chapter.

4.2.1 Evaluation of modal normalization constant change due to damage

The main difficulty in applying the DLV method for continuous online damage
diagnosis is the computation of the modal normalization constants for the damaged
structure. For complex civil infrastructure, the local damage is not expected to
significantly change the global structural characteristic, including the modal normalization
constants, which leads to the conjecture that the undamaged normalization constants
possibly can be employed for constructing an approximate damaged flexibility matrix for
the purpose of damage detection.
To understand the situation better, the 14-bay planar truss shown in Fig. 4.1 is
considered. A Matlab finite element model consisting of 53 bars and 28 nodes has been

2 6 4 10 6 14 8 18 10 22 12 26 14 30 16 34 18 38 20 42 22 46 24 50 26

2 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53
1 28
1 3 4 5 8 7 12 9 16 11 20 13 24 15 28 17 32 19 36 21 40 23 44 25 48 27 52
y
x
shaker 1 shaker 2

13 accelerometers in the vertical direction

Figure 4.1: 14-bay planar truss.

36
developed. Structural nodes are numbered starting from left to right, with all the odd
numbers at the lower chord, and all the even numbers at the upper chord, except node 28
which is at the right support. Each structural node number has a circle around it. Mode
shapes in the y-direction of all the lower chord nodes except the supports, and in the x-
direction for node 7, are used to construct the flexibility matrix in this example.
Now, let’s evaluate the change in the modal normalization constants due to damage on
this truss structure. Because the mode shapes obtained from system identification based on
measured data are arbitrarily normalized, the mode shapes before and after damage have
to be properly scaled to make a meaningful comparison of the modal normalization
constants. The idea is to have the mode shape at the measured DOFs scaled in such a way
that they have a unit magnitude, i.e.
T T
Ψ j Ψ j = 1.0 or ψ j ψ j = 1.0 (4.21)

The ratios between the damaged and undamaged normalization constant are evaluated
here for various damage scenarios.
The sample results when element 13 is damaged are shown in Fig. 4.2. Another
example when element 16 and 33 are damaged are displayed in Fig. 4.3. Both of these

Modal normalization constant d j

1.5

1.4

1.3
3rd index
Ratio

1.2 2nd index 1st index

4th index
1.1

1
5th index
0.9
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Damage extent

Modal normalization constant α j

1.16

1.12
Ratio

1.08
3rd index
1.04 2nd index
1st index

0.96 4th index 5th index

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Damage extent

Figure 4.2: Ratio between the damaged and undamaged

normalization constants when element 13 is damaged.

37
 Modal normalization constant d j

1.4

1.3 3rd index

Ratio
1.2 1st index

5th index
1.1
4th index
2nd index
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Damage extent

Modal normalization constant α j

1.12

1.08
3rd index
Ratio

1.04
4th index 2nd index

1st index 5th index

0.96
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Damage extent

Figure 4.3: Ratio between the damaged and undamaged

normalization constants when elements 16 and 33 are damaged.

results indicate a small change of modal normalization constants for the first five modes,
especially for small damage. This result implies that the undamaged normalization
constants may provide a reasonable approximation to the damaged ones when damage is
small. Note that, if the damping in the forced vibration case is proportional, the
normalization constant d j can be easily associated with α j for each mode by comparing
Eqs. (4.7) and (4.13).

4.2.2 Algorithm initialization

The first step of the proposed algorithm is to compute the undamaged normalization
constants and then the undamaged flexibility matrix from the measured data. Herein, this
step is termed as algorithm initialization. Based on how the undamaged normalization
constants are obtained, there are two different initialization approaches (see Fig. 4.4).

Initialization based on forced vibration

To obtain the undamaged normalization constant d j from forced vibration, the modal
parameters of the undamaged structure are first identified using the ERA method; the

38
 measured undamaged data measured undamaged data
(inputs and outputs) (outputs only)

impulse responses auto and cross-correlation auto and cross-correlation functions

(inputs and outputs) functions (outputs only)
NExT & ERA
ERA method
NExT & ERA undamaged modal
system matrices (A, B, C) and parameters (λ u, ψ u)
complex mode shapes Ψm undamaged complex
mode shapes Ψu ψu normalized to have
a unit magnitude
Ψm normalized to have
a unit magnitude
Ψu normalized to have mass purterbation method
a unit magnitude
equation (4.11) equation (4.20)

undamaged normalization T undamaged normalization

constants Dg Fu = Ψu Dg Ψu
constants α

undamaged flexibility undamaged flexibility

matrix F u matrix F u

Based on forced vibration Based on ambient vibration

Figure 4.4: Algorithm initialization.

normalization constant d j is then computed from Eq. (4.11) with the mode shapes
normalized to have a unit magnitude.
Because the mode shapes for the damaged structure will be identified from output
measurements employing the NExT algorithm in conjunction with the ERA method,
system identification using the same approach is applied on the output measurements to
obtain the undamaged mode shapes. These mode shapes are then normalized to have a unit
magnitude to compute the undamaged flexibility matrix employing Eq. (4.7).

Initialization based on ambient vibration

To obtain the undamaged normalization constant α j from ambient vibration, the

NExT algorithm in conjunction with the ERA method is employed for system
identification. Again, the identified mode shapes are normalized to have a unit magnitude.
α j can then be computed employing the mass perturbation method outlined above; and
the undamaged flexibility matrix is then constructed by Eq. (4.13).

4.2.3 Algorithm operation – detecting damage

Once the initialization is completed, the second step of the proposed approach is to
construct an approximate flexibility matrix for the potentially damaged structure

39
 measured data (outputs only)

auto and cross-correlation functions

NExT & ERA

modal parameters (Ψd or λd , ψd )

Ψd or ψd normalized to
have a unit magnitude

undamaged normalization
constants Dg or α

flexibility matrix Fd
(Eq. 4.7 or 4.13)

FΔ = Fd - Fu

singular value decomposition of FΔ

damage locating vectors (DLVs)

obtain normalized cummulative

stress under DLVs

damage?

Figure 4.5: Algorithm operation.

employing ambient vibration and then to apply the DLV method to detect damage in the
structure. Herein, this step is referred as algorithm operation (see Fig. 4.5).
During the operation of the SHM system, only outputs are measured and used to
identify the modal parameters. The NExT algorithm in conjunction with ERA method is
applied for system identification. The identified mode shapes are then normalized to have
a unit magnitude. The flexibility matrix is computed employing these identified modal
parameters and the undamaged normalization constants D g or α obtained during the
algorithm initialization. After the flexibility matrices for the undamaged and potentially
damaged structures are constructed, the DLV method can be applied to detect damage.
The DLVs are first calculated from the difference of the flexibility matrices, and then
statically applied to undamaged structure at the sensor locations. A small value of the
normalized cumulative stress in an element indicates that this element is a damage
candidate.

4.3 Numerical Validation

The proposed damage diagnosis approach is verified using the 14-bay planar truss
structure shown in Fig. 4.1. This planar truss consists of 53 steel circular bars, which have

40
 –4 2 –9 4
a cross section of 1.122 × 10 m and an area moment of inertia of 2.111 × 10 m . The
11 2
elastic modulus of the material is 2 × 10 N ⁄ m , and the mass density is
3 3
7.83 × 10 kg ⁄ m . The total length of this truss is 5.6 m, with each bay being
0.4m × 0.4m .
A total of 13 accelerometers are installed in the y-direction at all the lower chord
nodes except at the supports; and there is one accelerometer installed in the x-direction of
node 7. A band-limited white noise with RMS amplitude about 5% of the measured signal
is added to each of the measurements to simulate measurement noises. The truss is excited
using two independent band-limited white noises in the y-direction at nodes 13 and 21. To
obtain the undamaged normalization constants from forced vibration, excitation at node
13 is measured as the input. For the case of computing undamaged normalization
constants from ambient vibration, these two inputs are not measured.

4.3.1 Algorithm initialization

Algorithm initialization based on forced vibration

To obtain the modal normalization constant d j from forced vibration, the ERA
method is employed for system identification. The sample identification results are shown
in Fig. 4.6. As can be seen, system identification using the impulse responses shows good
agreement with the exact results. The undamaged normalization constants D g are then
computed using Eq. (4.11).
After the undamaged normalization constants D g are obtained, system identification
employing the NExT algorithm in conjunction with the ERA method is applied on the
output measurements only to obtain the undamaged mode shapes. First, acceleration
measurements are collected. The auto- and cross-spectral density functions are then
computed by selecting a reference output. The sample spectral density function between y-
direction acceleration at node 3 and node 9 (reference output location) is shown in Fig.
4.7; and the corresponding correlation function is shown in Fig. 4.8. These correlation
functions can then be used in the ERA method to extract the modal parameters. Sample
results are shown in Fig. 4.9. As can be seen, system identification using correlation
functions of outputs shows a good agreement with the exact results. This figure also
indicates that the difference between the natural frequencies before and after damage is as
small as 0.1% when element 13 has a 20% stiffness reduction for this specific truss.
The undamaged flexibility matrix is then constructed using Eq. (4.7) with Ψ being
the identified complex mode shapes at the sensor locations. Comparison of portion of the
identified and exact flexibility matrices constructed using the first four modes is shown in
Table 4.1. As can be seen, accurate results have been obtained for the flexibility matrix
associated with the first four sensor locations, i.e., y-direction of nodes [3 5], x- and y-
direction of node 7.

41
 Estimates: 26.1855 Hz Exact: 25.9587 Hz
2

-2
Estimated: 88.2100 Hz Exact: 88.0730 Hz
2

-2
Estimated: 126.5754 Hz Exact: 126.3466 Hz
2

-2
Estimated: 176.0944 Hz Exact: 176.0317 Hz
5

-5
2 4 6 8 10 12 14

Figure 4.6: Undamaged mode parameters

from ERA method ( ∗ : identified mode
shapes; : exact mode shapes).

Algorithm initialization based on ambient vibration

For the case of ambient vibration, system identification is conducted by employing

the NExT algorithm in conjunction with the ERA method. The undamaged normalization
constants α are obtained utilizing the mass perturbation method. These normalization
constants are then used to construct the undamaged flexibility matrix from Eq. (4.13) with
ψ being the identified undamped mode shapes at the sensor locations. A comparison of
the identified and exact flexibility matrices is shown in Table 4.1. Again, reasonably
accurate results have been obtained.
–5
Table 4.1: Comparison of undamaged flexibility matrices using the first four modes ( x10
m/N)

Estimated flexibility matrix

Exact flexibility matrix
Forced vibration Ambient vibration
0.0160 0.0299 -0.0042 0.041 0.0159 0.0298 -0.0042 0.0403 0.0167 0.0312 -0.0042 0.0420
0.0299 0.0560 -0.0069 0.0760 0.0298 0.0562 -0.0071 0.0765 0.0312 0.0587 -0.0070 0.0799
-0.0042 -0.0069 0.0031 -0.0082 -0.0042 -0.0071 0.0029 -0.0086 -0.0042 -0.0070 0.0030 -0.0084
0.0401 0.0760 -0.0082 0.1046 0.0403 0.0765 -0.0086 0.1052 0.0420 0.0799 -0.0084 0.1101

42
 60

40

Magnitude (dB)
20

- 20
- 40

- 60

- 80
0 20 40 60 80 100 120 140 160 180 200
Frequency (Hz)

Figure 4.7: Cross-spectral density function.

2500

2000

1500
Magnitude

1000

500

0
- 500

- 1000

- 1500
0 0. 2 0. 4 0. 6 0. 8 1 1. 2 1. 4 1. 6
Time (sec)

Figure 4.8: Correlation function.

4.3.2 Damage diagnosis results

Once the algorithm initialization is completed, the sensors on the structure can start
collecting data to monitor the structural condition. The approximate flexibility matrix is
constructed from the output measurements using the undamaged normalization constants.
The DLV method is then applied to detect damage in the structure. A cut-off value of 0.15
for normalized cumulative stress while using a total of six DLVs associated with the
smallest singular values yields good results for a wide range of damage cases; and
therefore will be used here to select the damaged candidates.
The damage cases considered in this numerical example are shown in Table 4.2. Case
1-1 demonstrates performance of the extended DLV method under the situation when only
a single element is damaged. Cases 2-1 and 2-2 illustrate the multiple damage scenarios.
These three cases are selected to illustrate the effectiveness of the method for damage in
different types of elements, including longitudinal, diagonal, and vertical elements.

43
 Estimates: 25.9521 Hz Exact: 25.9587 Hz Estimates: 25.9355 Hz Exact: 25.9452 Hz
2 2

0 0

-2 -2
Estimated: 88.0642 Hz Exact: 88.0730 Hz Estimated: 88.0549 Hz Exact: 88.0633 Hz
2 2

0 0

-2 -2
Estimated: 126.3201 Hz Exact: 126.3466 Hz Estimated: 126.1842 Hz Exact: 126.2030 Hz
2 2

0 0

-2 -2
Estimated: 176.0823 Hz Exact: 176.0317 Hz Estimated: 175.7030 Hz Exact: 175.6099 Hz
5 5

0 0

-5 -5
2 4 6 8 10 12 14 2 4 6 8 10 12 14
Undamaged Mode Shapes Damaged Mode Shapes

Figure 4.9: Identified modal parameters from NExT & ERA

method ( ∗ : identified mode shapes; : exact mode shapes).

Table 4.2: Two damage scenarios: three damage cases.

Damage extent
Cases Damaged elements
(stiffness reduction)
Single damage Case 1-1 20% diagonal element 13
longitudinal element 12 and
Case 2-1 20%
diagonal element 33
Multiple damage
longitudinal element 20 and
Case 2-2 20%
vertical element 39

Case 1-1

The results of the normalized cumulative stress induced by the DLVs, when element
13 is damaged, are shown in Fig. 4.10. As can be seen, for both cases of initialization from
forced and ambient vibration, elements 13 and 15 have a considerably smaller normalized
cumulative stress compared with other elements and smaller than the threshold. The
reason why element 15 has a small cumulative stress is apparent by looking at the force
balance at node 8 in Fig. 4.1. Under the current load configuration where the vertical
forces are only applied at the lower chord nodes, if either of these two elements has small
stress, so does the other. Therefore, element 13 is successfully identified as a possibly
damaged element in this truss structure employing the extended DLV method.

44
 Normalized Cumulative Stress σ j
1

_
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

Normalized Cumulative Stress σ j

1
_

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 4.10: Normalized cumulative stress when

element 13 is damaged.

Case 2-1

For the case in which elements 12 and 33 are damaged, the results are shown in Fig.
4.11. As can be seen, the damaged elements have a comparatively smaller stress which is
smaller than the threshold. Obviously, under the current load configuration, if element 12
has a small stress, element 18 should have a small stress as well. Similarly for elements 31
and 33, both of them will have a small stress, if either of them does. Therefore, elements
12 and 33 are determined as the candidates of the damage locations in both cases.

Case 2-2

The results when elements 20 and 39 are damaged are shown in Fig. 4.12, in which
the damaged elements have a comparatively smaller stress. Similarly, if element 20 or 39

45
 Normalized Cumulative Stress σ j
1

_
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

Normalized Cumulative Stress σ j

1
_

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 4.11: Normalized cumulative stress when

elements 12 and 33 are damaged.

is damaged, so does the element 26 or 41. Therefore, elements 20 and 39 are determined
as potentially damaged elements for both cases.

4.4 Summary

In this chapter, an approach extending the DLV method to continuous online damage
diagnosis was proposed. Different formulations of the flexibility matrix were reviewed.
The flexibility matrix for a structure with general viscous damping can be constructed
when there is at least one co-located sensor and actuator pair; while for the case when only
output measurements are available, flexibility matrix can be computed by employing a
mass perturbation method for structures with proportional damping. Changes in the modal
normalization constants due to structural damage were investigated. Results for a planar
truss structure indicate that modal normalization constants of lower frequency modes may
not change significantly when damage is small. This results lead to the extension of the
DLV method for continuous online monitoring.

46
 Normalized Cumulative Stress σ j
1

_
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

Normalized Cumulative Stress σ j

1
_

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 4.12: Normalized cumulative stress when

elements 20 and 39 are damaged.

The essence of the extended DLV method is to construct an approximate flexibility

matrix for the potentially damaged structure utilizing the modal normalization constants
from the undamaged structure. The DLV method can then be applied for online damage
diagnosis employing ambient vibration. The proposed approach was numerically
validated using a 14-bay planar truss structure, with results showing it working well for
both single and multiple damage scenarios.

47
 CHAPTER 5

DISTRIBUTED COMPUTING SHM STRATEGY

USING SMART SENSORS

As shown in Chapter 4, good results have been obtained using a limited number of
sensors distributed throughout the structure; however, questions arise about whether the
sensor information away from the damaged region is needed. Could the information from
the sensors in the vicinity of the damaged region alone be used to locate the damage? This
question will be addressed in this chapter.
Following the introduction of the necessary background knowledge, the concept of
damage detection using localized sensors is demonstrated. Development of a new
distributed computing strategy (DCS) for SHM is then proposed that is suitable for
implementation on a network of densely distributed smart sensors. Finally, numerical
validation of the DCS approach is provided with results showing the proposed approach
promising (Spencer and Gao 2005; Gao et al. 2006).

5.1 Background

Monitoring of complex structures to provide real-time safety and reliability

information of the structure poses significant technical challenges. To detect damage in
large civil infrastructure systems, densely distributed sensors are expected to be required.
Use of traditional wired sensors is challenging for such applications because of the cost
and difficulty in deploying and maintaining a large wiring plant. Using a wireless sensor
network is also difficult because large amounts of measured data need to be transferred to
a central station. The bandwidth and power requirement to transfer these data may easily
exceed the limit of the wireless sensor. In both cases, a tremendous amount of data is
expected to be generated that would need to be sent to a central station. Managing this
large amount of data is challenging. Unnecessary information needs to be eliminated to
efficiently utilize the network.
Recently rapid advances in smart sensor technology have made damage detection
using a dense array of sensors feasible (Spencer et al. 2002, 2004). The essential feature of
a smart sensor is the on-board microprocessor, which allows smart sensors to make
decisions, perform computation, save data locally, etc. By conducting a portion of the
computation at the sensor level, only limited information needs to be transferred back to a
central station. Damage detection algorithms which can take advantage of the distributed
computing environment offered by smart sensors are highly desired but currently very
limited.

48
5.2 Locating Damage Using Local Sensor Information

In this section, damage detection using localized sensors to monitor the local
elements is investigated, which directly leads to the proposed DCS approach in the next
section.
This concept is demonstrated through a numerical example employing the DLV
method. The flexibility matrices before and after damage are constructed using the first
four modes from the analytical model. The planar truss shown in Fig. 4.1 is utilized in this
numerical example. For convenience, a picture of this truss structure is re-displayed in
Fig. 5.1.
For this specific structure, assume there are two sensors at every structural node
except the supports, one in the x-direction and the other in the y-direction. Sensors on
every three consecutive lower chord nodes and the corresponding upper chord nodes are
grouped together to monitor those elements connected to these nodes. A total of 11
different localized sensor groups are formed from left to right in this truss structure. To
allow some computational redundancy, every two adjacent groups have some overlap. For
example, group 1 includes nodes [2, 3, 4, 5, 6, 7], and group 2 includes structural nodes [4,
5, 6, 7, 8, 9], etc. Therefore, for each group of localized sensors there are a total of 12
sensors that are installed in the x- and y-direction at six structural nodes.
In this initial example, system identification is not employed; rather, the exact modal
parameters (i.e., mode shapes and frequencies) are used to construct the truncated
flexibility matrix which is then employed in the DLV method. The flexibility matrix at the
sensor locations is constructed using the first four modes. For the case when element 5 has
a 20% stiffness reduction, the normalized cumulative stress when using a total of seven
DLVs associated with the smallest singular values is computed for each localized sensor
group and shown in Fig. 5.2. Results from group 1 to 11 are shown in the order from top to
the bottom. In each plot, a circle is used to mark elements in which both nodes are
members of the group, and a cross signifies the elements in which only one of its nodes is
a member of a group. Only the condition of elements in which both nodes are in a given
group are evaluated here for damage by the localized sensor group.
In Fig. 5.2, results in group 1 shows that element 5 has a very small normalized
cumulative stress compared with other elements in the group, which indicates that element
5 is a possibly damaged element. The results from all other groups indicate that there is no
other damaged element in this truss. Similar results when elements 11 and 33 are damaged

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

Figure 5.1: 14-bay planar truss.

49
 5
1
: members being monitored

1
0.5
0 : members with one end
1
connected to community

2
0.5
0
1

3
0.5
0
1

4
0.5
0
1
5
0.5
Sensor Group

0
1
6

0.5
0
1
7

0.5
0
1
8

0.5
0
1
9

0.5
0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Figure 5.2: Normalized cumulative stress when element 5 has

a 20% stiffness reduction.

are shown in Fig. 5.3, in which element 11 is identified as having damage by local groups
1, 2, and 3; and element 33 is determined as a damage candidate by local groups 7 and 8.
It has been demonstrated that the damaged elements can be determined based on
information of a small group of sensors localized in the damaged region. This implies that
the information away from the damage location might not be necessary for damage
detection. This concept can be employed for damage localization for structures with a
dense array of sensors.

5.3 Distributed Computing Strategy (DCS)

The previous section provides the basis for development of the scalable SHM strategy
presented in this section (Spencer and Gao 2005; Gao et al. 2006).

5.3.1 Hierarchical organization

The conceptual hierarchical organization of the proposed DCS approach is shown in

Fig. 5.4. Different hierarchical organizations have been proposed by researchers for
different applications in recent years (Frampton 2001; Akyildiz et al. 2002; Lynch 2002).
Note that Fig. 5.4 is just a simple example to demonstrate the concept of the DCS
approach. In contrast to traditional SHM algorithms which require all the measured
information to be transferred to a central station, the measured information is aggregated
locally by a selected sensor within the sensor group, termed the manager sensor, and only

50
 11 33
1
: members being monitored

1
0.5
0 : members with one end
1 connected to community

2
0.5
0
1

3
0.5
0
1

4
0.5
0
5 1
0.5
Sensor Group

0
1
6

0.5
0
1
7

0.5
0
1
8

0.5
0
1
9

0.5
0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Figure 5.3: Normalized cumulative stress when elements

11 and 33 have a 20% stiffness reduction

community 1 2 3 4
smart sensors

manager sensors

central station

Figure 5.4: Sketch of hierarchical organization.

limited information is sent back to the central station to provide the condition of the
structure. Small numbers of smart sensors are grouped to form different communities. For
clarity, this figure shows each sensor as being included in only one community; however,
in the proposed approach, each sensor can participate in multiple communities. For each
community, the manager sensor collects measured responses and implements the damage
detection algorithm for this community. Adjacent manager sensors need to interact with
each other to exchange information. Referring again to Fig. 5.4, manager sensors in
communities 1, 2, and 3 interact with each other while community 4 only interacts with
community 3.

51
 After the measured information is aggregated, the manager sensor determines what
information needs to be sent back to the central station. In the proposed approach, each of
the communities, in which damage has not occurred only transmits an “ok” signal to the
central station, which is reflected by the dotted line connection in Fig. 5.4. The
communities in which damage has occurred need to send information about the damaged
elements, which is reflected by the solid line connection in Fig. 5.4. In this way, only
limited information needs to be transferred between sensors throughout the entire
structure. This approach will significantly reduce the communication traffic in the sensor
network.

5.3.2 Strategy implementation

The planar truss structure shown in Fig. 5.1 is employed to illustrate details of the
implementation of the proposed SHM strategy.

Community development

First, different communities are formed. A single community includes a set of

adjacent structural nodes, sensors on these nodes, and members. These structural members
have both ends connected to the structural nodes in the same community. Fig. 5.5 shows
an example how communities can be formed. Different communities are developed from
left to right in the truss structure. To facilitate communication among the smart sensors,
the structural nodes within the same community should be close to each other. As an
example, community 6 in Fig. 5.5 includes nodes [12, 13, 14, 15, 16, 17] and elements
[23, 24, 25, 26, 27, 28, 29, 30, 31].
Only the elements which have both ends connected to community nodes are
monitored by the manager sensor for this community. There is an exception for those
communities close to the supports. If a structural member has one end connected to the
support and the other end connected to a community node, then this structural member can
also be monitored by this community. For example, in Fig. 5.5, elements 1 and 2 only have
one end connected to the community nodes, but they both have the other end connected to
a support; therefore these two elements will be investigated by the smart sensors in
community 1.

community 1 community 6
community 2 community 7

2 6 4 10 6 14 8 18 22 12 26 14 30 16 34 18 38 42 46 50

2 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53
1 28
1 3 4 5 8 7 12 9 16 20 13 24 15 28 17 32 19 36 40 44 48 52

Figure 5.5: Forming communities for damage detection.

52
 To allow some computational redundancy, adjacent communities are recommended to
have some overlaps so that each structural member is monitored by more than a single
community.

Computing undamaged normalization constants for communities

Once the community is formed, the modal normalization constants for the undamaged
structure have to be computed based on either forced or ambient vibration. For this
purpose, one sensor from each community, named as the reference sensor, will be required
to send the recording data to the central station. These reference sensors are selected so
that their mode shape magnitudes will not be zero for the modes interested. The modal
parameters associated with these reference sensors, denoted as Ψˆ j or λ̂ j and ψ̂ j
respectively, are identified using the ERA method or the NExT in conjunction with the
ERA method. The undamaged normalization constants can then be computed based on the
methods presented in Chapter 4, and denoted as dˆ j or α̂ j .
However, the normalization constants dˆ j or α̂ j cannot be employed directly by each
community for damage detection, because the scalar for the j mode shape Ψˆ or ψˆ and j j
i i
the mode shape in ith community Ψ j or ψ j can be different. The jth undamaged
i i
normalization constant for the ith community d j or α j can be obtained as
ˆ 2
i ˆ ⎛ Ψj ( i ) ⎞ i ψ̂ j ( i )
dj = d j ⋅ ⎜ -------------
-⎟ and α j = α̂ j ⋅ ------------
- (5.1)
⎝ Ψ ij ( k )⎠ i
ψj ( k )

where j = jth mode; i = ith community; and k = reference sensor location in the ith
community. The undamaged normalization constants for the ith community are
i i i i i i i i
D g = [ d 1, d 2, …d j ] or α = [ α 1, α 2, …α j ] .

Data aggregation

To minimize the communication traffic in the sensor network, measured data need to
be transferred to the manager sensor for data processing.
Clocks of smart sensors in the same community are first synchronized with each other
and these sensors can then start measuring data. The measurements are transferred to the
manager sensor for computation. To facilitate communication, the manager sensor should
have an overall shorter distance to other sensors in the community. For example, in Fig.
5.6, the sensor on structural node 15 is selected as the manager sensor for community 6.
As can be seen from the figure, there are smart sensors which may need to transfer
information to more than one manager sensors. This happens when the smart sensors
participate in different communities. For example, if communities are developed as shown
in Fig. 5.5, sensors at node 6 need to send measured data to the manager sensor on node 5
for community 1, node 7 for community 2, as well as node 9 for community 3.
After the data has been transferred to the manager sensor, computation can be
conducted using the on-board microprocessor to locate the damage within the community.

53
 community 1 community 6
2 6 12 14 16

1 28

5 7 9 11 15 19

Figure 5.6: Data collection.

The extended DLV method presented in Chapter 4 is incorporated in this DCS approach to
localize the damage in each community. The flow chart for data aggregation in a
community is shown in Fig. 5.7. Note that the undamaged flexibility matrix Fu in Fig. 5.7
i
is constructed using Eq. (4.7) or (4.13) once the undamaged normalization constants d j or
i
α j for community i are computed from experimental data.

Decision making

Actions need to be taken after the data aggregation is done for a community. If there
is no damage detected in a community, the manager sensor does not initiate the interaction
with other manager sensors and just sends an “ok” signal back to the central station. If
there is damage identified in a community, the manager sensor needs to send queries to its
counterpart in adjacent communities. There are three possibilities after sending the
queries:
• The damage candidate in community i does not participate in adjacent communities.
The manager sensor in community i sends the damage information back to the
central station.
•The damage candidate in community i participates in other communities and is
identified as the potentially damaged location in all of these communities. This
damage candidate is confirmed and reported to the central station by these
communities.
• The damage candidate in community i participates in other communities, but not all
of the communities identify it as the potentially damaged element. These
communities then need to retake data, and re-conduct data aggregation and decision
making.
A flow chart for the decision making is shown in Fig. 5.7.
To better illustrate how decision making works in a smart sensor network, a simple
example is shown in Fig. 5.8. The manager sensors in each community first conduct
damage detection locally. Those communities having no damage send “ok” signal back to
the central station. Because damage is detected in communities 3, 4, and 8, these
communities interact with the surrounding communities. As a result, communities 3 and 4
confirm that element 16 is the damage location and send this damage information back to
the central station. Communities 7 and 8 are required to retake data and re-conduct
damage detection as inconsistent information has been obtained regarding element 33.

54
 measured data (outputs only)

data collection

auto and cross-correlation functions

NExT & ERA

modal parameters
i i i
(Ψd or λd , ψd )

i
Ψd ior ψ d normalized to
have a unit magnitude

data aggregration undamaged normalization

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

potentially damaged
flexibility matrix Fd

FΔ = Fd - Fu

singular value decomposition of FΔ

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 these communities?
communities
no
retake data, and re-conduct data
aggregation and decision making

Figure 5.7: Data aggregation and decision making.

55
 action action action
damage
send receive send receive send receive

G1 No S "ok" B - SLEEP

G2 No S "ok" B R ?16 G3 S "0" G3 - SLEEP

S ?16 G2 - - R "0" G2
G3 Yes S "D16" B - SLEEP
S ?16 G4 R ?16 G4 S "1" G4 R "1" G4

S ?16 G3 R ?16 G3 S "1" G3 R "1" G3

G4 Yes S "D16" B - SLEEP
S ?16 G5 - - R "0" G5

G5 No S "ok" B R ?16 G4 S "0" G4 - SLEEP

G6 No S "ok" B - SLEEP

G7 No S "ok" B R ?33 G8 S "-1" G8 - - R "C" G8 P "C"

S ?33 G7 - - R "-1" G7 S "C" G7

G8 Yes - P "C"
S ?33 G9 - - R "0" G9 S "I33 7-8" B

G9 No S "ok" B R ?33 G8 S "0" G8 - SLEEP

G10 No S "ok" B - SLEEP

G11 No S "ok" B - SLEEP

Notes: Examples
G1 Manager sensor in community 1. S ?16 G2 Send query about element 16 TO group 2.
S Send. S "1" G2 Send signal "1" TO group 2.
R Receive. S "ok" B Send ok signal TO base station.
B Base station. S "D16" B Send damage information about element 16 TO
P Perform. base station.
SLEEP Sleep mode. S "I33 7-8" B Send signal to base station indicating that
"-" Idle or listening mode. inconsistent information regarding element
? Query. 33 has been obtained by groups 7 and 8.
"0" No. It is not a member of our group. S "C" G3 Send "Cycle" signal TO group 3 advising it to
"1" Yes. It is a member of our group and is damaged. retake data and re-conduct damage evaluation.
"-1" Yes. It is a member of our group but is NOT damaged. R ?16 G2 Receive query about element 16 FROM group2.
"D" Damage information. R "1" G2 Receive signal "1" FROM group 2.
"ok" No damage found in this group. R "C" G4 Receive "Cycle" signal FROM group 4.
"C" Cycle: cycle of retaking data and re-conducting P "C" Perform another cycle of retaking data and
damage evaluation. re-conducting damage evaluation.
"I" Inconsistent.

Figure 5.8: Decision making example (element 16 is consistently identified as

having damage in communities 3 & 4; and inconsistent information has been
obtained by communities 7 & 8 regarding element 33).

56
Damage detection results regarding element 33 need to be reported to the central station
after retaking data.

5.4 Numerical Validation

The proposed DCS approach is verified using the planar truss structure shown in Fig.
5.1. Properties of this truss structure have been presented in section 4.4 of the previous
chapter.
In this numerical example, assume that accelerometers are installed in the x- and y-
direction at all the nodes except the supports. Communities are developed as described in
section 5.3.2. A band-limited white noise with an RMS amplitude around 5% of the
measured signal is added to each measurement to simulate measurement noises. Only the
first three modes identified from the measured data are utilized for damage detection.
This truss structure is excited in the y-direction using two independent band-limited
white noises. These excitations are not measured except for the case of algorithm
initialization employing forced vibration. To better assess efficacy of the proposed
approach in practice, both excitation location and magnitude have been changed before
and after damage. These excitation conditions are shown in Table 5.1.

Table 5.1: Excitation conditions before and after damage.

Magnitude (height of Power

Location
Spectral Density function)
Before damage nodes 7 and 11 0.12 and 0.18
After damage nodes 9 and 23 0.18 and 0.09

5.4.1 Constructing undamaged flexibility matrix in communities

The undamaged normalization constants dˆj or αˆ j in Eq. (5.1) are first obtained from
forced or ambient vibration. Then the SHM system can start measuring data to construct
the undamaged flexibility matrix employing ambient vibration.
First, the acceleration measurements are collected and the auto- and cross-spectral
density functions are computed by selecting a reference output, which are used to compute
the corresponding auto- and cross-correlation functions. The sample correlation function,
between the acceleration in the y-direction of node 5 and node 7 (reference output in
community 1), is shown in Fig. 5.9. These correlation functions can then be used by the
ERA method to extract the modal parameters and sample results are shown in Fig. 5.10. In
this figure, the star represents the identified mode shape, and the solid line represents
exact mode shape from the undamaged analytical model. As can be seen, system
identification using only outputs has shown good agreement with the analytical model.
Once the undamaged modal parameters in each community are identified, the
associated normalization constants can be obtained using Eq. (5.1), and the undamaged
flexibility matrix can then be constructed employing Eq. (4.7) or (4.13). Comparison of

57
 1500

1000

Magnitude
500

- 500

-1000
0 0.5 1 1.5 2 2.5
Time (sec)

Figure 5.9: Correlation function.

Exact: 25.9587 Hz Estimates: 25.9657 Hz

Exact: 88.0730 Hz Estimated: 88.0758 Hz

Exact: 126.3466 Hz Estimated: 126.4384 Hz

Figure 5.10: Undamaged modal parameters ( ∗ :

identified mode shapes; : exact mode shapes).

the identified flexibility matrices with the exact one at the measured DOFs in the x- and y-
direction of nodes 3 and 5 is shown in Table 5.2. As can be seen, accurate results have
been obtained using the modal normalization constants identified from either forced or
ambient vibration.

5.4.2 Damage detection results

For this specific truss structure and sensor configuration, a cut-off value of 0.3 for
normalized cumulative stress has been found to work well for a wide range of damage

58
 Table 5.2: Comparison of undamaged flexibility matrices using the first three modes
–6
( x10 m/N).

Estimated flexibility matrix

Exact flexibility matrix Initialization based on forced Initialization based on ambient
vibration vibration
0.0038 -0.0083 0.0072 -0.0128 0.0032 -0.0089 0.0063 -0.0145 0.0034 -0.0082 0.0065 -0.0130
-0.0083 0.1428 -0.0201 0.2710 -0.0089 0.1460 -0.0215 0.2766 -0.0082 0.1488 -0.0200 0.2832
0.0072 -0.0201 0.0139 -0.0328 0.0063 -0.0215 0.0123 -0.0360 0.0065 -0.0200 0.0126 -0.0332
-0.0128 0.2710 -0.0328 0.5169 -0.0145 0.2766 -0.0360 0.5264 -0.0130 0.2832 -0.0332 0.5417

scenarios while using a total of seven DLVs associated with the smallest singular values.
This cut-off values is therefore used herein to select the damaged elements.
By constructing an approximate flexibility matrix for each community employing the
locally measured information, the DLV method can be applied for continuous monitoring
of the structure. Both single and multiple damage scenarios are investigated and the
representative damage cases are shown in Table 5.3. The single damage scenario is studied
in Cases 1-1 and 1-2, in which different damage extents are considered. Additionally,
three multiple damage scenarios are illustrated. Cases 2-1 and 2-2 have a same damage
extent for the damaged elements, while Case 2-3 has a different damage extent for the
damaged element. Again, damage in different types of elements, including longitudinal,
diagonal, and vertical elements, are studied through these selected damaged cases.

Table 5.3: Two damage scenarios: five damage cases.

Damage extent
Cases Damaged elements
(stiffness reduction)
Case 1-1 20% diagonal element 13
Single damage
Case 1-2 40% longitudinal element 18
diagonal element 17 and
Case 2-1 20%
longitudinal element 36
vertical element 11 and
Multiple damage Case 2-2 20%
longitudinal element 40
longitudinal element 20 and
Case 2-3 40% and 30%
diagonal element 45

Case 1-1

The results of the normalized cumulative stress for each community when diagonal
element 13 has a 20% stiffness reduction is displayed in Fig. 5.11. For both cases of
initialization from forced and ambient vibration, results from communities 2 and 3 show
that element 13 has a normalized cumulative stress smaller than the threshold value of 0.3.
Therefore, this element is confirmed as a damage location in these communities. Results
from other communities show no elements having a small normalized cumulative stress,

59
 13
1
0.5 : members being monitored

1
0 : members with one end
1
connected to community
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

13
1
0.5 : members being monitored
1

0 : members with one end

1
0.5 connected to community
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 5.11: Normalized cumulative stress when

element 13 has a 20% stiffness reduction.

60
 Note that for a densely distributed sensor network, the possibility of identifying
undamaged elements as potentially damaged locations has been significantly reduced. For
this specific case, the damaged element is correctly identified as the damage candidate by
communities 1 and 2, and no more damage is found by the other communities.
Therefore, the manager sensors in communities 2 and 3 send the damage information
to the central station; and other communities which have no damage only send an “ok”
signal back to the central station.

Case 1-2

Figure 5.12. displays the results of the computed normalized cumulative stress when
longitudinal element 18 has a 40% stiffness reduction. Results from communities 3 and 4
show that the normalized cumulative stress for element 18 is smaller than the threshold.
Therefore, this element is confirmed as a damage location by these two communities.
For the case of initialization based on forced vibration, results from other
communities show no elements having a small normalized cumulative stress. The
manager sensors in communities 3 and 4 send the damage information to the central
station; and other communities only send an “ok” signal back to the central station.
For the case of initialization based on ambient vibration, community 11 identifies
element 43 as a damage candidate because it has a small normalized cumulative stress
which is smaller than the threshold. However, communities 9 and 10 report this element as
undamaged because it has a larger normalized cumulative stress. Inconsistent information
has been obtained by communities 9, 10, and 11 regarding element 43. These communities
need to retake data and re-conduct damage detection. The results are shown in Fig. 5.13.
Because element 43 has a larger stress than the threshold in all these communities, so it is
identified as an undamaged element. The manager sensors in communities 3 and 4 send
the damage information regarding element 18 to the central station; and other
communities only send “ok” signal back to the central station.
These results show that the proposed DCS approach works well when a single
element is damaged. Multiple damage scenarios are also studied and results from three
different cases will be illustrated in the following paragraphs.

Case 2-1

Figure 5.14 shows the results when diagonal element 17 and longitudinal element 36
have a 20% stiffness reduction. Communities 3 and 4 determine element 17 as a
potentially damaged element; and communities 8 and 9 identify element 36 as the damage
candidate.
For the case of initialization from forced vibration, community 6 reports element 23
as a damage candidate because it has a normalized cumulative stress smaller than the
threshold. However, this element is determined as undamaged by communities 4 and 5.
Inconsistent information is obtained by communities 4, 5, and 6 regarding element 23.
These communities need to retake data and re-conduct damage detection. The results are

61
 18
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5

4
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

18 43
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 5.12: Normalized cumulative stress when element

18 has a 40% stiffness reduction.

62
 18 43
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
4 0.5
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 5.13: Normalized cumulative stress when element

18 has a 40% stiffness reduction (retaking data regarding
element 43).

shown in Fig. 5.15 which determines element 23 as an undamaged location as it has a

large stress in all these communities.
Therefore, for both the cases of initialization based on forced and ambient vibration,
communities 3 and 4 report damage information regarding element 17, and communities 8
and 9 send damage information regarding element 36 back to the central station. Other
communities having no damage only send back an “ok” signal.

Case 2-2

Figure 5.16 shows the results when vertical element 11 and longitudinal element 40
have a 20% stiffness reduction. Similarly, communities 1, 2, and 3 determine that element
11 is the potentially damaged element; and communities 9 and 10 show element 40 as a
damaged candidate. For both cases, these communities report damage information and
other communities only send an “ok” signal back to the central station.

63
 17 23 36
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

17 36
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 5.14: Normalized cumulative stress when elements 17 and 36

have a 20% stiffness reduction.

64
 17 23 36
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

Figure 5.15: Normalized cumulative stress when elements 17 and 36

have a 20% stiffness reduction (retaking data regarding element 23).

Case 2-3

The results when longitudinal element 20 has a 40% stiffness reduction and diagonal
element 45 has a 30% stiffness reduction are displayed in Fig. 5.17. Again, element 20 has
been correctly identified as a damage candidate by communities 3 and 4; and element 45
is determined as a potentially damaged element by communities 10 and 11. These
communities send the damage information to the central station. Communities with no
damage detected only send an “ok” signal back to central station.
These simulation results demonstrate that the proposed DCS approach is promising
not only for the single damage scenario but also for the multiple damage scenario. In these
examples, damaged elements are consistently identified by using localized sensor groups,
and retaking data and re-conducting damage detection was also studied.

5.5 Summary

In this chapter, a new distributed computing strategy (DCS) for SHM was proposed
that is suitable for implementation on a network of densely distributed smart sensors. The
proposed DCS approach differs from the traditional damage detection algorithms because

65
 11 40
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

11 40
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 5.16: Normalized cumulative stress when elements

11 and 40 have a 20% stiffness reduction.

66
 20 45
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

20 45
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 5.17: Normalized cumulative stress when element 20 has a 40%

stiffness reduction and element 45 has a 30% stiffness reduction.

67
it does not rely on central data acquisition and processing. In this approach, a hierarchical
strategy is proposed in which adjacent smart sensors are grouped together to form sensor
communities. The extended DLV method developed in Chapter 4 is employed to evaluate
the condition of the local elements within these communities by utilizing only locally
measured information. The damage detection results in these communities are then
communicated with the surrounding communities and sent back to a central station. In this
way, only limited information needs to be transferred between sensors throughout the
entire structure.
Numerical simulations with noise included in measurements were conducted by
employing a 14-bay truss structure. To better assess the performance of the approach,
excitations were changed before and after damage. Both single and multiple damage
scenarios were studied. The possibility of falsely identifying undamaged elements as the
site of potential damage has been significantly reduced through the use of a dense array of
sensors. Numerical results have shown the proposed DCS approach promising for
continuous SHM with a densely distributed sensor network.

68
 CHAPTER 6

EXPERIMENTAL VALIDATION OF THE DISTRIBUTED

COMPUTING SHM STRATEGY

To better assess the performance of the distributed computing SHM strategy proposed
in Chapter 5, experimental validation of the proposed approach has been conducted at the
Smart Structures Technology Laboratory of the University of Illinois at Urbana-
Champaign (Gao and Spencer 2005b, 2007). Following a detailed description of the
experimental setup, experimental results employing the distributed computing SHM
strategy are provided which show the proposed approach to be very promising.

6.1 Experimental Setup

The distributed computing SHM strategy is experimentally verified using the 5.6 m
long three-dimensional truss structure discussed in Chapter 3 and shown in Fig. 3.10. Note
that the roller support has been changed in this experiment to prevent the movement in the
vertical direction. This new roller support is depicted in Fig. 6.1.
The dynamic testing system shown in Fig. 6.2 has been developed to support this
experiment. The magnetic shaker and the truss structure are not in the picture. Some of the
components have been described in details in Chapter 3. Herein, only the new
experimental components will be discussed.
A new data acquisition system from National Instruments was purchased and set up in
the SSTL. The key components for this data acquisition system are the data acquisition
(DAQ) board NI PCI-6052E, which has a 16-bit input and output resolution and a

Figure 6.1: Roller support.

69
 spectral analyzer

signal conditioner

multifunction
multifunction DAQ
DAQ

lowpass filter cascaded

with simultaneous
sampling amplifier

labview
labview program
program

Figure 6.2: Experimental setup.

maximum sampling rate of 333 kS/s, the lowpass elliptical filter NI SCXI-1141, and the
simultaneous-sampling differential amplifier NI SCXI-1140.
The National Instruments SCXI-1141 is an 8-channel, programmable 8th-order
lowpass elliptical filter, which allows a sharp roll off of the response. By passing the
measured data through this lowpass filter, the component of the signal at frequencies
higher than the cut-off frequency will be significantly reduced. The cut-off frequency of
this filter is user selectable in the range from 10Hz to 25KHz. The SCXI-1141 has a
maximum sampling rate of 333KS/s and a programmable gain for each channel with a
input range from ± 50 mV to ± 5 V.
The National Instruments SCXI-1140 is an 8-channel simultaneous sample and hold
differential amplifier module. Simultaneous sampling is very important for dynamic
testing with multiple channel measurements, because any differences in the sample times
will result in artificial phase lags in the data. With the SCXI-1140, users can sample
multiple signals with negligible phase delay between channels. The SCXI-1140 can be
cascaded with the SCXI-1141 for applications requiring filtering and simultaneous
sampling to provide high quality dynamic measurements.

70
 A Labview program has been developed to acquire and save the data from this
National Instruments data acquisition system. This program allows users to set the
programmable cut-off frequencies for the lowpass filter, change the gain on each
measurement channel, set sampling rate and measuring time, etc.
Ideally, all the structural members should be monitored for damage. Due to the
limited experimental equipment, only elements in the outer vertical panel of the truss will
be monitored employing the proposed DCS approach. A sketch of this panel is depicted in
Fig. 6.3 with each node number having a circle around it.
A Ling Dynamic Systems permanent magnetic V408 shaker is attached to the bottom
nodes of this panel to excite the structure vertically. Connection among the shaker, a PCB
load cell (model 208B02), and a structural node is described in section 3.4.1 and shown in
Figs. 3.13 and 3.14.
A steel bracket has been installed at each of the structural joints except the supports in
this panel to facilitate installation of accelerometers in the x-direction (see Fig. 6.4). The
accelerometers are attached to the bracket and structural node using magnetic bases.

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

Figure 6.3: Outer vertical panel of the three-dimensional

truss structure.

Figure 6.4: Installation of accelerometers.

71
 Sensor communities are formed in the way same as Fig. 5.5. There are a total of 12
accelerometers in a sensor community with one accelerometer in the x-direction and the
other in the y-direction at each structural node. These localized sensors are used to monitor
the condition of nine local structural members which have both ends connected to the
structural nodes in the same community. Again, there are a total of 11 sensor communities
in this experiment.
One of the communities with a single member replaced by the damaged one is
displayed in Fig. 6.5. Due to the limited experimental equipment, these 12 accelerometers
shown in Fig. 6.5 have to be moved along the truss structure to measure the response for
each community.

two accelerometers at each

structural node (see Fig. 6.4).

damaged element

Figure 6.5: Example of a single sensor community for the

outer vertical panel of the truss.

6.2 Algorithm Initialization

To detect the damage in this truss structure employing the proposed DCS approach,
the undamaged flexibility matrix at the sensor communities needs to be constructed.
Construction of the flexibility matrix based on both forced and ambient vibration
employing experimental data will be investigated. Details of these two approaches have
been reviewed in sections 4.1.
For the case of initialization based on forced vibration, dynamic testing is first
conducted when seven accelerometers are installed at nodes [7, 11, 13, 17, 19, 21, 23] to

72
measure the acceleration in the y-direction and the shaker is attached to the bottom of node
17. Therefore, there is one co-located sensor and actuator pair which is at node 17. Modal
normalization constants associated with these mode shapes can then be obtained from
experimental data using Eq. (4.11).
For the case of algorithm initialization based on ambient vibration, a total of 11
accelerometers were installed in the y-direction of nodes [5, 7, 9, 11, 13, 15, 17, 19, 21, 23,
25], and two accelerometers in the x- and z-direction (perpendicular to x-y plane) of node
11. A lumped mass of 1.124 kg, which is about 48% of the original nodal mass (including
the structural joint and half weight of each member connected to it), was attached to the
bottom of node 11 to simulate mass perturbation. The modal normalization constants can
be obtained based on the mass perturbation method using Eq. (4.20).
To verify the results from both of these approaches, undamaged flexibility matrices at
the y-direction of nodes [7, 11, 13, 17, 19, 21, 23] have been computed using Eqs. (4.7)
and (4.13) using the first six dominant modes (see Fig. 6.6, in which acceleration at node
21 is selected as the reference). Good agreement has been shown in Tables 6.1. and 6.2
Table 6.1: Undamaged flexibility matrices based on forced vibration using the first six
–5
dominant modes ( x10 m/N).

Normalization constants based on forced vibration

0.0533 0.0704 0.0703 0.0604 0.0533 0.0414 0.0321
0.0704 0.0958 0.0977 0.0885 0.0809 0.0651 0.0526
0.0703 0.0977 0.1012 0.0949 0.0887 0.0732 0.0604
0.0604 0.0885 0.0949 0.0960 0.0936 0.0804 0.0689
0.0533 0.0809 0.0887 0.0936 0.0935 0.0819 0.0714
0.0414 0.0651 0.0732 0.0804 0.0819 0.0730 0.0646
0.0321 0.0526 0.0604 0.0689 0.0714 0.0646 0.0579

After the modal normalization constants associated with the mode shapes at reference
sensor locations were determined, a series of dynamic testing was conducted with 12

20 2
0 1 3 6
- 20
4 5
Magnitude (dB)

- 40
- 60
- 80
-100
-120
-140
-160
-180
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)

Figure 6.6: Auto- and cross-spectral density

functions at nodes [7, 11, 13, 17, 19, 21, 23].

73
 Table 6.2: Undamaged flexibility matrices based on ambient vibration using the first six
–5
dominant modes ( x10 m/N).

Normalization constants based on ambient vibration

0.0539 0.0695 0.0690 0.0568 0.0489 0.0375 0.0286
0.0695 0.0930 0.0951 0.0838 0.0756 0.0609 0.0488
0.0690 0.0951 0.0993 0.0917 0.0850 0.0705 0.0579
0.0568 0.0838 0.0917 0.0930 0.0906 0.0788 0.0674
0.0489 0.0756 0.0850 0.0906 0.0906 0.0805 0.0700
0.0375 0.0609 0.0705 0.0788 0.0805 0.0728 0.0642
0.0286 0.0488 0.0579 0.0674 0.0700 0.0642 0.0572

sensors installed in each community to obtain the mode shapes in these communities
under ambient vibration. The associated modal normalization constants in each of these
communities can then be obtained from Eq. (5.1) and the undamaged flexibility matrix at
each community can be constructed using Eq. (4.7) or (4.13).
Based on contribution of the frequency modes to the measurements in each
community, different communities might need to utilize different frequency modes to
construct the flexibility matrix. Fig. 6.7 shows a total of 12 cross spectral density functions
for communities 4 and 8, respectively. These cross spectral density functions are

20 2
5
0 1 3 4
- 20
Magnitude (dB)

- 40
- 60
- 80
-100
-120
-140
-160
-180
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)

Community 4

20 2
1 5
0 3 4
- 20
Magnitude (dB)

- 40
- 60
- 80
-100
-120
-140
-160
-180
0 10 20 30 40 50 60 70 80 90 100
Frequency (Hz)

Community 8

Figure 6.7: Auto- and cross-spectral density functions.

74
computed from the measurements of the 12 accelerometers in each community, with the
vertical acceleration at node 9 being selected as the reference output for community 4 and
the vertical acceleration at node 21 being selected as the reference output for community
8. The first five dominant modes are identified for both communities 4 and 8. As shown in
Fig. 6.7, the fourth dominant mode is different for these two communities.
In this experimental validation, flexibility matrices in communities 1 through 5 are
constructed using the same dominant modes as community 4, and communities 6 through
11 use the same dominant modes as community 8. Therefore, the flexibility matrix in each
community is constructed using the first five dominant modes identified from the
measurements.

6.3 Damage Detection and Decision Making

Once the algorithm initialization is finished, the sensors can start collecting data from
the structure to monitor its condition.
A finite element model of the truss structure has been developed using Matlab with a
total of 160 elements and 56 nodes (see Fig. 6.8). Pin supports are defined at nodes 1 and
29 and roller supports are defined at nodes 28 and 56. For the roller supports, translation is
only allowed in the x-direction, and rotation around x-axis is also permitted. Matching the
dynamic properties between the finite element model and this experimental structure is
difficult. However, the proposed approach is not sensitive to the dynamic properties of this
structure as the analytical model is utilized for a static computation after the DLVs are
obtained.
A series of damage detection tests have been conducted using the three-dimensional
14-bay truss structure located at SSTL of UIUC. Structural damage is simulated by
replacing the original elements with ones having a 52.7% cross section reduction, which
results in a slightly different axial stiffness reduction for different members (see Table 6.3)
Three excitation conditions (see Table 6.4) have been investigated. As shown in the
table, excitation condition 1 has a same excitation location, magnitude, and bandwidth
before and after damage. To better simulate the real ambient vibration condition in which
the excitation contents can be changed, performance of the proposed approach under
excitation condition 2 and 3 is also studied. Condition 2 has a different magnitude and

54
52
50
48 26
46 24
44 22 56
42 20 55
40 18 53
38 16 51 28
36 14 49 27
34 12 47 25
32 10 45 23
30 8 43 21
6 41 19
4 39 17
y 2 37 15
35 13
29 33 11
31 9
x 7
5
3
1
z

Figure 6.8: Three-dimensional finite element model.

75
 Table 6.3: Structural damage.

Longitudinal Vertical Diagonal

element element element
Cross section reduction 52.7% 52.7% 52.7%
Equivalent axial stiffness
44% 43.5% 46.9%
reduction

Table 6.4: Various excitation conditions before and after damage.

Location Magnitude Bandwidth

Condition 1 same same same
Condition 2 same different different
Condition 3 different different different

bandwidth, and condition 3 has a different location, magnitude, and bandwidth before and
after damage.
A cut-off value of 0.3 for the normalized cumulative stress is used and provides good
performance for a wide range of damage cases by using a total of five DLVs associated
with the smallest singular values. This cut-off value will be used to select the damaged
elements from the detection results.

6.3.1 Excitation condition 1

Under this excitation condition, the Ling Dynamic Systems permanent magnetic
V408 shaker is attached to the bottom of node 17 to excite the structure vertically. A band-
limited white noise with a 0.1 RMS value of magnitude and a bandwidth of 100 Hz is sent
from the Siglab spectral analyzer to drive the shaker.
Single and multiple damage scenarios, as well as damage in different types of
elements, are studied. The process of retaking data when there is inconsistent damage
information is not investigated in this stage. Instead it will be studied under excitation
conditions 2 and 3. Results of three damage detection cases (see Table 6.5) are illustrated
and discussed herein to demonstrate the performance of the proposed DCS approach.

Table 6.5: Excitation condition 1: three damage cases.

Cases Damaged elements

Case 1-1 diagonal element 17
Case 1-2 longitudinal element 12 and diagonal element 41
Case 1-3 vertical element 7 and longitudinal element 36

Case 1-1

Figure 6.9 shows the results when element 17 is damage. For both the cases of

76
 17 19
1
: members being monitored
0.5

1
0 : members with one end
1 connected to community
0.5

2
0
1
0.5

3
0
1
0.5
4
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

17
1
0.5 : members being monitored
1

0 : members with one end

1
connected to community
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.9: Normalized cumulative stress when

element 17 is damaged.

77
cumulative stress smaller than the threshold of 0.3 and therefore is determined as a
damage candidate.
For the case of initialization from forced vibration, community 3 identifies element
19 as possibly having damage because its normalized cumulative stress is smaller than the
threshold. However, communities 4 and 5 report this element as undamaged because its
stress is larger than the threshold in both these communities. Inconsistent information
regarding the condition of element 19 is obtained in communities 3, 4, and 5, therefore
these communities need to retake data and re-conduct damage detection. As mentioned
above, the process of re-conducting damage detection is not investigated under excitation
condition 1, but it will be studied under excitation conditions 2 and 3.
Figure 6.9 shows that elements 52 and 53 also have a small normalized cumulative
stress for the case of initialization based on forced vibration. These elements are not part
of community 11. Instead, they have one end connected to community 11 and the other
end connected to the support. Numerical simulation in Chapter 5 shows that these type of
elements can still be monitored by the adjacent community (i.e., elements 1 and 2 are
monitored by community 1 and elements 52 and 53 are monitored by community 11).
However, the experimental results do not confirm that assumption. Part of the reason may
be due to the difficulty to model the real support condition. In this experiment, only the
elements which have both ends connected to structural nodes in the same community will
be monitored.

Case 1-2

Multiple damage scenarios are also investigated. Figure 6.10 shows the damage
detection results when longitudinal element 12 and diagonal element 41 are damaged. As
can be seen from the figure, element 12 is determined as a potentially damaged element by
communities 2 and 3, and element 41 is identified as a damage candidate by communities
9 and 10.
For the case of initialization from forced vibration, communities 8, 9, and 10 need to
retake data and re-conduct damage detection as inconsistent information regarding the
condition of element 39 is obtained. Similarly, for the case of initialization based on
ambient vibration, communities 6 and 7 need to re-conduct damage detection as
inconsistent information regarding element 28 is obtained.

Case 1-3

The detection results regarding damage in vertical element 7 and longitudinal element
36 are shown in Fig. 6.11. The damaged elements have been correctly identified for both
cases of initialization based on forced and ambient vibration. For the case of initialization
based on ambient vibration, communities 1 and 2 need to re-conduct damage detection
regarding element 9, and communities 9 and 10 need to retake data and re-conduct
damage detection regarding element 40.
These experimental results show that the proposed DCS approach works well when
excitation remains the same before and after damage. To better evaluate the performance

78
 12 39 41
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

12 28 41
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.10: Normalized cumulative stress when

elements 12 and 41 are damaged.

79
 7 36
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

7 9 36 40
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
: members being monitored
0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.11: Normalized cumulative stress when

elements 7 and 36 are damaged.

80
of the proposed approach, experiments with excitation contents being changed before and
after damage are also studied.

6.3.2 Excitation condition 2

For excitation condition 2, the magnitude of the band-limited white noise excitation is
changed from a RMS value of 0.1 to 0.08, and the bandwidth is changed from 100 Hz to
200 Hz before and after damage. Two damage scenarios (see Table 6.6) are studied under
this excitation condition.
Table 6.6: Excitation condition 2: two damage cases.

Cases Damaged element

Case 2-1 no damage
Case 2-2 longitudinal element 8

Case 2-1

Experiments have been conducted employing the DCS approach when there is
damage in the structure and good damage detection results are obtained. Questions arise
regarding its performance when there is no damage in the structure.
Figure 6.12 shows the computed normalized cumulative stress when there is no
damage in the structure. As shown in the figure, the normalized cumulative stresses in
each community are larger than the threshold, so no damage is detected. All sensor
communities send “ok” signal to the central station.
This example demonstrates that the proposed approach also works well when there is
no damage in the structure.

Case 2-2

This case demonstrates the process of retaking data and re-conducting damage
detection when inconsistent information has been obtained. Damage detection results
when longitudinal element 8 is damaged are shown in Fig. 6.13. Element 8 is successfully
identified as the damage candidate in both cases as its computed normalized cumulative
stress is smaller than the threshold.
For the case of initialization based on forced vibration, community 6 reports that
element 24 is damaged as it has a normalized cumulative stress smaller than the threshold
0.3. However, community 5 identify element 24 as having no damage as its stress exceeds
the threshold. Inconsistent information is obtained by communities 5 and 6 regarding the
condition of element 24, therefore these two communities need to retake data and re-
conduct damage detection to determine its condition. Fig. 6.14 shows the results after re-
conducting damage detection in these two communities, element 24 is determined as
undamaged because both communities report that it has a normalized cumulative stress
larger than the threshold.

81
 1
0.5 : members being monitored

1
0 : members with one end
1
0.5
connected to community

2
0
1
0.5

3
0
1
0.5

4
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

1
0.5 : members being monitored
1

0 : members with one end

1
connected to community
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.12: Normalized cumulative stress

when there is no damage in the structure.

82
 8 24
1
0.5 : members being monitored

1
0 : members with one end
1 connected to community
0.5

2
0
1
0.5

3
0
1
4 0.5
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

8
1
0.5 : members being monitored
1

0 : members with one end

1
connected to community
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.13: Normalized cumulative stress when

element 8 is damaged.

83
 8 24
1
0.5 : members being monitored

1
0 : members with one end
1 connected to community
0.5

2
0
1
0.5

3
0
1
0.5

4
0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

Figure 6.14: Normalized cumulative stress when element 8

is damaged (retaking data regarding element 24).

Therefore, only communities 1 and 2 report damage information to the central station,
while other communities send “ok” signal back to the central station.

6.3.3 Excitation condition 3

For excitation condition 3, not only are the excitation magnitude and level changed
before and after damage, but the shaker location is also moved from node 17 to node 23 to
simulate the change of the excitation location. Two damage scenarios (see Table 6.7) are
considered under this excitation condition, with case 3-2 illustrating the process of
retaking data and re-conducting damage detection.

Table 6.7: Excitation condition 3: two damage cases.

Cases Damaged element

Case 3-1 diagonal element 9
Case 3-2 longitudinal element 16

84
Case 3-1

Figure 6.15 shows the damage detection results when element 9 is damaged. In both
cases, element 9 is consistently identified as the damage candidate and there is no other
damage. Therefore, only communities 1 and 2 report the damage information and other
communities only need to send a “ok” signal back to the central station.

Case 3-2

Detection results when element 16 is damaged is shown in Fig. 6.16, where element
16 is correctly identified as the damage candidate by communities 3 and 4 for both cases.
For the case of initialization based on forced vibration, communities 9, 10, and 11
report inconsistent information regarding the condition of element 43. These communities
need to retake data and re-conduct damage detection regarding element 43. Fig. 6.17
illustrates the results after communities 9, 10, and 11 re-conduct damage detection. These
results confirmed that element 43 is not damaged.
For the case of initialization based on ambient vibration, inconsistent information
regarding elements 20 and 34 is obtained. Fig. 6.18 shows results when communities 4
and 5 retake data regarding element 20, and communities 7 and 8 retake data regarding
element 34. Again, these elements are identified as having no damage as their normalized
cumulative stresses are larger than the threshold.
These experimental results demonstrate that the proposed DCS approach not only
works robustly when excitation remains the same before and after damage, it also
performs well when excitation contents have been changed. Therefore, the proposed
approach is a promising algorithm which has the potential application to practice with a
dense array of smart sensors.

6.4 Discussions

In this experiment, inconsistent damage information is obtained in various damage

cases. Questions arise how frequently this situation happens, as well as how reliably the
proposed algorithms can detect structural damage. This section provides some insights on
these questions.
To investigate the situation of inconsistent damage information, let’s consider Case 2-
2 in which element 8 is damaged. As shown in Fig. 6.12, for the case of initialization from
forced vibration, inconsistent information regarding element 24 is obtained by community
6. For this specific community, a total of ten sequential tests of retaking data and re-
conducting damage detection for the damaged structure and have been conducted under
various excitation conditions (see Table 6.8). The damage detection results for community
6 that includes elements [23, 24, 25, 26, 27, 28, 29, 30, 31] are shown in Fig. 6.19. As can
be seen, among these ten tests, only tests 1 and 7 reports possible damage information for
a different location. Although these results are problem dependent, they indicate that the

85
 9
1
0.5 : members being monitored

1
0 : members with one end
1
connected to community
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

9
1
0.5 : members being monitored
1

0 : members with one end

1
connected to community
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5
0
1
11

0.5
0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.15: Normalized cumulative stress when

element 9 is damaged.

86
 16 43
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

16 20 34
1
0.5
1

0
1
0.5
2

0
1
0.5
3

0
1
0.5
4

0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.16: Normalized cumulative stress

when element 16 is damaged.

87
 16 43
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on forced vibration

Figure 6.17: Normalized cumulative stress when element 16

is damaged (retaking data regarding element 43).

situation of inconsistent damage information is rather random and does not occur in the
majority of the tests.
To provide some insights on how reliably the proposed DCS approach detects
structural damage, Case 2-2 is considered again. Ten sequential tests are conducted for
community 2 that includes elements [7, 8, 9, 10, 11, 12, 13, 14, 15]. The excitation
conditions are the same as Table 6.8. The damage detection results are displayed in Fig.
6.20, which shows element 8 is identified as the damage candidate. These results indicate
the proposed DCS approach can consistently identify the damage location, in this case,
element 8.

6.5 Summary

In this chapter, the distributed computing strategy for SHM was experimentally
verified using a three-dimensional 14-bay truss structure. Structural damage was
simulated by replacing various truss elements with ones having a 52.7% cross section
reduction. Both single and multiple damage scenarios were studied. Various excitation
conditions, including changing excitation level, bandwidth and location before and after
damage, were also investigated. Retaking data and re-conducting damage detection when
there is inconsistent information were studied using experimental data.

88
 16 20 34
1
0.5

1
0
1
0.5

2
0
1
0.5

3
0
1
0.5
4 0
1
Community

0.5
5

0
1
0.5
6

0
1
0.5
7

0
1
0.5
8

0
1
0.5
9

0
1
10

0.5 : members being monitored

0
1 : members with one end
11

0.5 connected to community

0
0 5 10 15 20 25 30 35 40 45 50
Element Number

Initialization based on ambient vibration

Figure 6.18: Normalized cumulative stress when element 16

is damaged (retaking data regarding elements 20 and 34).

Table 6.8: Excitation conditions for ten sequential experiments: element 8 is damaged.

Excitation Excitation
Tests bandwidth magnitude
(Hz) (RMS value)
Undamaged 100 0.10
Test 1 200 0.06
Test 2 200 0.08
Test 3 200 0.10
Test 4 200 0.12
Test 5 200 0.14
Damaged
Test 6 100 0.06
Test 7 100 0.08
Test 8 100 0.10
Test 9 100 0.12
Test 10 100 0.14

89
 1.0 1.0

Test 1
Test 1
0.5 0.5
0 0
1.0 1.0

Test 2
Test 2
0.5 0.5
0 0

1.0 1.0

Test 3
Test 3
0.5 0.5
0 0
1.0 1.0

Test 4
Test 4

0.5 0.5
0 0

1.0 1.0

Test 5
Test 5

0.5 0.5
0 0
1.0 1.0

Test 6
Test 6

0.5 0.5
0 0
1.0 1.0

Test 7
Test 7

0.5 0.5
0 0
1.0

Test 8
1.0
Test 8

0.5 0.5
0 0
1.0
Test 9
1.0
Test 9

0.5 0.5
0 0
Test 10

1.0
Test 10

1.0
0.5 0.5
0 0
23 25 27 29 31 7 9 11 13 15
Element Number Element Number

Figure 6.19: Normalized Figure 6.20: Normalized

cumulative stress when element cumulative stress when element
8 is damaged (community 6) 8 is damaged (community 2)

The experimental results have shown that the proposed DCS approach can
successfully monitor local community members only utilizing locally measured
information for various damage scenarios under different excitation conditions. The
proposed DCS approach was shown promising for application of SHM on a densely
distributed sensor network.

90
 CHAPTER 7

IMPLEMENTATION OF THE DCS APPROACH

ON A SIMULATED SMART SENSOR NETWORK

To better illustrate the implementation of the proposed DCS approach, a Simulink

model with the support of Stateflow (MathWorks, Inc., 2004) has been developed that
simulates a smart sensor network. In Chapters 5 and 6, the proposed DCS approach was
shown to be promising through numerical and experimental examples. However, the
flowchart shown in Fig. 5.7 needs to be implemented in an autonomous manner so that
important implementation issues for the DCS approach can be studied. These issues
include: how sensor communities communicate with each other to exchange damage
information, and when and how they automatically initiate retaking data and re-conduct
damage detection, etc. The Simulink and Stateflow model, which has been developed and
will be presented in this chapter, is used to address these issues.
In this chapter, following brief introduction of Simulink and Stateflow, an overview
of the proposed Simulink and Stateflow model is provided. Implementation of the
proposed DCS approach in this Simulink model employing both numerical and
experimental measurements is then presented.

7.1 Background

Simulink is an extension to Matlab which provides a graphical interface for the

construction of a block diagram representation of dynamic systems. It can represent either
linear or nonlinear systems whether they are in continuous or discrete time domain. It has
a wide range of pre-programmed block libraries which allow users to accurately design,
simulate, and analyze different dynamic systems.
Stateflow is a graphical design and development tool for control and supervisory
logic used with Simulink (The MathWorks, Inc. 2004). Stateflow is a finite state machine,
which means it is a representation of an event-driven system. This system reacts to events
by making a transition from one state to another state if the condition of this transition is
true (see Fig. 7.1). Programming for smart sensors is based on event-driven language,
therefore Stateflow in conjunction with Simulink is a suitable tool to simulate a smart
sensor network.
Some practical issues pertaining to the smart sensor technology will not be studied
herein due to the technical limitation of Simulink and Stateflow. The wireless transmission
between smart sensors is replaced by the direct link between Simulink or Stateflow
blocks. Therefore, communication of sensor communities through radio frequency
technology is not studied. Data loss or confliction which may happen in practice can not
be modeled. In addition, synchronization of the measurements is not simulated because
Simulink cannot run under different clocks.

91
 event[condition]
State 1 State 2

Figure 7.1: Stateflow example.

7.2 Overview of the Simulink and Stateflow Model

A Simulink and Stateflow model that simulates a smart sensor network consisting of
eleven sensor communities has been developed (see Fig. 7.2). Sensor communities are
formed in the manner same as Fig. 5.5. This sensor network can be initiated to detect
damage upon the user’s request, and provide damage information and the inconsistent
information regarding condition of structural members. It has the capability to allow
adjacent sensor communities to communicate with each other in order to share damage
information, and automatically initiate those communities with inconsistent damage
information to retake data and re-conduct damage detection. This sensor network is
capable of detecting both single and multiple damage scenarios, but only a single damage
is allowed in one community. Herein, each sensor community communicates with the
communities immediately before and after it. Major components of this Simulink and
Stateflow model will be introduced in the remainder of this section.
Figure 7.3 shows the user switch and output scopes of the Simulink and Stateflow
model. The user switch, which is at the lower left corner of the graph, can simulate
sending a request from a central station to the sensor network to initiate a process of
damage detection. This is done by broadcasting an event throughout the sensor network
whenever the user switch changes its output value. There are a total of 11 columns LEDs/
scopes. Each column represents outputs from one sensor community, e.g., LED1 is
associated with community 1, and LED2 is associated with community 2, and so on.
The green LED at the first row turns to red if there is damage confirmed in the
associated sensor community, and turns to half green and half red if there is inconsistent
information. For example, LED7 changes to half green and half red because inconsistent
damage information is obtained in community 7.
The second row from the top in Fig. 7.3 shows the final damage detection results by
displaying a number bearing a specific meaning (see Table 7.1). For example, scope
F_dele3 displays a value of 17, which means element 17 has been determined as a damage
candidate in community 3.
The third row from the top shows the same value as the one in the second row if there
is no inconsistent information. Otherwise, it displays the number of the element which has
inconsistent damage information. For example, scope dele10 displays a number of 41.
This result means inconsistent information has been obtained in community 10 regarding

92
 LED1 LED2 LED3 LED4 LED5 LED6 LED7 LED8 LED9 LED10 LED11

u j3 y u j3 y u j3 y u j3 y u j3 y u j3 y u j3 y u j3 y u j3 y u j3 y u j3 y

J3_1 J3_2 J3_3 J3_4 J3_5 J3_6 J3_7 J3_8 J3_9 J3_10 J3_11

0 0 0 0 0 0 0 0 0 0 0

F_dele1 F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10 F_dele11

0 0 0 0 0 0 0 0 0 0 0

dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 dele11

0 0 0 0 0 0 0 0 0
0 0 0 0
0 0 0 0 0 0 0
PC1
PC2 PC3 PC5 PC6 PC7 PC9 PC10 PC11
PC4 PC8
In1 In1
In1 In1 In1 In1 In1 In1 In1
Out1 Out1
Out1 Out1 Out1 Out1 Out1 Out1 Out1
In2 In2
In2 In2 In2 In2 In2 In2 In2
J2_2 J2_3 J2_4 J2_5 J2_6 J2_7 J2_8 J2_9 J2_10

User User User User User User User User User

dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10
User dele1 User dele11
1 isDam1 isDam2 isDam3 isDam4 isDam5 isDam6 isDam7 isDam8 isDam9

93
Constant F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10
isDam2 F_dele1 isDam10 F_dele11
isDam3 isDam4 isDam5 isDam6 isDam7 isDam8 isDam9 isDam10 isDam11
-1 User Switch
isDam2_1 isDam3_2 isDam4_3 isDam5_4 isDam6_5 isDam7_6 isDam8_7 isDam9_8 isDam10_9
Constant1 dele2 isDam1_2 dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 isDam11_10

Community 1 Community 11
dele3 isDam2_3 dele4 isDam3_4 dele5 isDam4_5 dele6 isDam5_6 dele7 isDam6_7 dele8 isDam7_8 dele9 isDam8_9 dele10 isDam9_10 dele11 isDam10_11

Community 2 Community 3 Community 4 Community 5 Community 6 Community 7 Community 8 Community 9 Community 10

Figure 7.2: Simulink and Stateflow model.

 LED1 LED2 LED3 LED4 LED5 LED6 LED7 LED8 LED9 LED10 LED11

0 0 17 17 0 0 -1 0 0 -1 0

F_dele1 F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10 F_dele11

0 0 17 17 0 0 33 0 0 41 0

dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 dele11

-2 0 1 -2 -2 0 -2 -2 -1
-2 -2
-2 1 0 -2 -2 -1 -2 -2 0
PC1
PC2 PC3 PC5 PC6 PC7 PC9 PC10 PC11
PC4 PC8

94
1

Constant

-1 User Switch

Constant1

Figure 7.3: User switch and output scopes.

element 41, because LED10 turns to half green and half red and F_dele10 displays a value
of -1.

Table 7.1: Final damage detection result in a community.

Number Meaning

0 no damage in this community

-1 inconsistent information obtained in this community

other damaged element number

The last row in Fig. 7.3 shows the query results from adjacent communities. The top
scope shows the query result from the left community and the bottom one displays the
result from the right community with a different number bearing a different meaning (see
Table 7.2). For example, the top scope in PC7 shows a value of 0 representing that the
damage candidate 33 is not a member of community 6. The bottom scope displays a value
of -1, which means element is a member of community 7 but it is not identified as a
damage location.

Table 7.2: Query result from an adjacent community.

Number Meaning

0 damage candidate is not a member of the adjacent community

damage candidate is a member of the adjacent community and is

1
confirmed as having damage

damage candidate is a member of the adjacent community but is

-1
not identified as having damage

-2 no query sent to this adjacent community

A total of 11 sensor communities have been formed (see Fig. 7.4) and the content of a
typical sensor community, e.g. community 4, is shown in Fig. 7.5. There are four different
Stateflow blocks to simulate different functions of the manager sensor, including detecting
damage, determining the condition of local elements based on detection results in this
community and query results from adjacent communities, and processing queries from
adjacent communities. There is an embedded Matlab function that advises community 4 to
retake data and re-conduct damage detection when inconsistent damage information is
obtained.
The Stateflow block “Detect 4” (see Fig. 7.6) simulates damage detection in the
sensor community. There are two states in this block: Off and Detect4 (see Table 7.3).
State Off is the default state for a manger sensor. There are two possibilities when the
manager sensor transits from state Off to Detect4 (see Table 7.4). In this table, y = -1
means there is inconsistent information; and n <= 2 means that the manager sensor will at
most re-conduct damage detection for two times. In state Detect4, a Matlab program is

95
 User User User User User User User User User
dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10
User dele1 User dele11

isDam1 isDam2 isDam3 isDam4 isDam5 isDam6 isDam7 isDam8 isDam9

F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10
isDam2 F_dele1 isDam10 F_dele11
isDam3 isDam4 isDam5 isDam6 isDam7 isDam8 isDam9 isDam10 isDam11

isDam2_1 isDam3_2 isDam4_3 isDam5_4 isDam6_5 isDam7_6 isDam8_7 isDam9_8 isDam10_9

dele2 isDam1_2 dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 isDam11_10

Community 1 Community 11
dele3 isDam2_3 dele4 isDam3_4 dele5 isDam4_5 dele6 isDam5_6 dele7 isDam6_7 dele8 isDam7_8 dele9 isDam8_9 dele10 isDam9_10 dele11 isDam10_11

96
Community 2 Community 3 Community 4 Community 5 Community 6 Community 7 Community 8 Community 9 Community 10

Figure 7.4: Sensor communities.

 y dele4 1
dele4

Detect 4

dele4

2 isDam3
isDam3
F_dele4 2
F_dele4
3 isDam5
isDam5

y
1
User R_results 4

Pulse 4 dele3
Generator dele3

dele4 isDam 3
isDam4_3

R_query 4_3
retakeData

Embedded
MATLAB Function
dele4

5 dele5 isDam 4
dele5 isDam4_5

R_query 4_5

Memory
0

Constant

Figure 7.5: A typical sensor community:

community 4.

97
 Detect4
en: gn = 4;
ele = ml('[15 16 17 18 19 20 21 22 23]');
dele4 = ml.detect3(gn, y);

{n =n+1;}

on_switch{n = 0;}

Off
[y == -1 & n <= 2]

Figure 7.6: Typical Stateflow block in

Simulink model: Detect 4.

invoked to load the measured data and implement the damage detection algorithm for the
community. The manager sensor automatically returns from state Detect4 to Off after a
damage detection process is finished.

Table 7.3: Two states in Stateflow block “Detect 4”.

States Meaning

Off manager sensor is not detecting damage

Detect4 manager sensor is detecting damage

Table 7.4: Two possibilities of transition from state Off to Detect4.

Possibility Condition Meaning

an on_switch event is generated from the central

First on_switch{n = 0} station to advise the manager sensor to conduct
damage detection

inconsistent damage information has been

Second [y = -1 & n <= 2] obtained, therefore this community is required to
re-conduct damage detection

The Stateflow block “R_results 4” (see Fig. 7.7) receives the query results from
adjacent communities and processes these results to provide condition of local elements in
community 4. There are three states in this Stateflow block: Off, R_results4, and
No_results (see Table 7.5); and there are three possibilities that the manager sensor needs

98
 {n=n+1;} [isDam3 == -2 & isDam5 == -2]

R_results4
en: det(); No_results
en: F_dele4 = 0;

on_switch{n = 0;}

[(y == -1 | z == -1)...
& n <=2]
{n = n+1;}

Off
en: pre_isDam3 = isDam3; ...
pre_isDam5 = isDam5;
du: det_z2(pre_isDam3, isDam3, ...
pre_isDam5, isDam5);

eM
eM det_z2(px1, x1, px2, x2)
det

Figure 7.7: Typical Stateflow block in

Simulink model: R_results 4

to re-evaluate the condition of local elements, i.e., transit from state Off to either state
R_results4 or No_results (see Table 7.6). Under all these three conditions, the manager
sensor exits the Off state. If there is damage in community 4, it transits to state R_results4
which makes final judgements on which element is damaged. Otherwise, it moves to state
No_results which assigns a value of 0 to variable F_dele4.

Table 7.5: Three states in Stateflow block “R_results 4”.

State Meaning

Off manager sensor is not processing damage detection and query results

manager sensor is processing the damage detection and query results,

R_results4
and damage is found in this community

manager sensor is processing the damage detection and query results

No_results
and no damage is found in this community

99
 Table 7.6: Three possibilities to exit state Off.

Possibility Condition Meaning

new damage detection has been conducted in this

First on_switch{n = 0}
community reacting to event on_switch

new damage detection has been conducted in this

Second [y = -1 & n <= 2] community because inconsistent information is
obtained

new damage detection has been conducted in

Third [z = -1 & n <= 2] adjacent communities therefore query results from
adjacent communities may be changed

The Stateflow block “R_query 4_3” (see Fig. 7.8) receives query from adjacent
community 3, processes the query, and sends the results back to community 3. Similarly,
there are three states in this block: Off, R_query4_3, and No_query (see Table 7.7); and
there are three possibilities when the manager sensor exits state Off (see Table 7.8). After
the manager sensor exits the Off state, it enters state R_query4_3 if a query is detected,
otherwise it enters state No_query. The query results from this Stateflow block are then
sent back to community 3.

Table 7.7: Three states in Stateflow block “R_query 4_3”.

State Meaning

Off manager sensor is not detecting and processing query

R_query4_3 manage sensor detects there is query and is processing it

manager sensor detects no query and is assigning a value -2 to

No_query
variable isDam which means no query is received from community 3

Table 7.8: Three possibilities to exit state Off.

Possibility Condition Meaning

new damage detection has been conducted in this

community due to event on_switch, so the damage
First on_switch{n = 0}
information in community 4 may be changed and the
query from community 3 needs to be re-valuated

new damage detection has been conducted in this

Second [y = -1 & n <= 2]
community due to inconsistent damage information

new damage detection has been conducted in the

Third [z = -1 & n <= 2] adjacent communities whose damage information may
have been changed

100
 [dele3 ~= 0]{n=n+1;}
R_query4_3 [dele3 == 0]
en: comp();

No_query
on_switch{n = 0;} en: isDam = -2;

[(y == -1 | z == -1)...
& n <=2]

{n = n+1;}

Off
en: pre_dele = dele3;
du: det_z(pre_dele, dele3)

eM
eM
comp det_z(px, x)

Figure 7.8: Typical Stateflow block in

Simulink model: R_query 4_3.

Similarly, the Stateflow block “R_query 4_5” receives query from adjacent
community 5, processes it, and sends the results back to community 5.
The embedded Matlab function “retakeData” collects outputs from Stateflow blocks
“R_results 4”, and “R_query 4_3”, and “R_query 4_5” to determine whether re-
conducting damage is needed. If any of the results from these three Stateflow blocks is -1,
then “retakeData” generates a value of -1 to advise the community to re-conduct damage
detection. A number of -1 from these three Stateflow blocks means that inconsistent
information has been obtained either in community 4 or the adjacent community (see
Table 7.9), therefore re-conducting damage detection is needed in community 4.
Table 7.9: Meaning of a number of -1 from different Stateflow blocks.

-1 Meaning

From “R_results 4” inconsistent information has been obtained in community 4

inconsistent information is obtained in community 3 whose

From “R_query 4_3”
damage candidate is a member of community 4.

inconsistent information is obtained in community 5 whose

From “R_query 4_5”
damage candidate is a member of community 4

This section illustrated major components of the proposed Simulink and Stateflow
model. Implementation of the DCS approach employing both simulation data and
experimental measurements utilizing the proposed Simulink and Stateflow model will be
presented in the next section.

101
7.3 Numerical and Experimental Validation

Damage detection results of two damage cases (see Table 7.10) employing the
proposed Simulink and Stateflow model are illustrated herein. In both cases, algorithm
initialization in the proposed DCS approach is based on forced vibration.

Table 7.10: Numerical and experimental validation.

Cases Case type Structure Damaged elements

longitudinal element 20 has a 40%

14-bay planar truss
Case 1 Simulation stiffness reduction, and diagonal element
(see Fig. 5.1)
45 has a 30% stiffness reduction

outer vertical panel of a

three-dimensional 14-bay longitudinal element 8 has a 52.7% cross
Case 2 Experiment
truss structure section reduction
(see Figs. 3.10 and 6.3)

Case 1

In this case, structural damage in elements 20 and 45 is investigated using simulation

data from a 14-bay planar truss (see Fig. 5.1). Details regarding this structural and the
computer simulation have been presented in Chapter 5.
The damage detection results without employing the Simulink and Stateflow model
are shown in Fig. 5.17, where communities 4 and 5 identify element 20 as a potentially
damaged element and communities 10 and 11 determine element 45 as a damage
candidate.
The damage detection results employing the Simulink and Stateflow model are shown
in Fig. 7.9. As can be seen, the LEDs for community 4, 5, 10, and 11 turn to red which
means damage has been detected in these sensor communities. The second row from the
top shows the final damage detection results, where communities 4 and 5 show that
element 22 is the damaged element and communities 10 and 11 show that element 45 is
the damaged element. The third row from the top displays the same element numbers as
the second row as there is no inconsistent damage information. In the fourth row, the top
scope of PC4 shows a number of 0, which means damaged element 20 is not a member of
community 3. The bottom scope of PC4 displays a number 1 indicating element 20 is a
member of community 5 and is confirmed as a damage candidate. Scopes in PC5, PC10
and PC11 display the similar information. Therefore, the damaged elements have been
consistently identified using the proposed Simulink and Stateflow model.

Case 2

In this case, structural damage in element 8 is studied using experimentally measured

data from the outer vertical panel of a three-dimensional 14-bay truss structure (see Figs.

102
 LED1 LED2 LED3 LED4 LED5 LED6 LED7 LED8 LED9 LED10 LED11

0 0 0 0 20 20 0 0 0 45 45

F_dele1 F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10 F_dele11

0 0 0 0 20 20 0 0 0 45 45

dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 dele11

103
-2 -2 -2 0 1 -2 -2 -2 0
-2 1
-2 -2 -2 1 0 -2 -2 -2 1
PC1
PC2 PC3 PC5 PC6 PC7 PC9 PC10 PC11
PC4 PC8

Figure 7.9: Outputs from Simulink and Stateflow model

when elements 20 and 45 are damaged.
3.10 and 6.3). Details regarding the experimental set up have been presented in Chapters 3
and 6.
The damage detection results without employing the Simulink and Stateflow model
are shown in Figs. 6.13 and 6.14. Damage detection results in Fig. 6.13 show that
communities 1 and 2 identify element 8 as a potentially damaged element and
communities 5 and 6 obtain inconsistent damage information regarding element 24. The
results after retaking data in these two communities are displayed in Fig. 6.14 which
identifies element 24 as undamaged.
The damage detection results employing the proposed Simulink and Stateflow model
are illustrated in Figs. 7.10 and 7.11. The LEDs in communities 1 and 2 turn to red in Fig.
7.10 and the associated scopes in the second and third rows display element 8 is identified
as a damage candidate. PC1 and PC2 in the fourth row indicate element 8 is only a
member of communities 1 and 2. As can be seen from Fig. 7.10, the LED in community 6
turns to half red and half green and the associated scope in the second row shows a value
of -1, which indicate inconsistent damage information has been obtained in this
community. Output scope in the third row shows that element 24 is identified as damaged
in community 6. Top scope of PC6 shows a value of -1 indicating element 24 is a member
of community 5 but is not identified as damaged; and bottom scope of PC6 displays value
of 0 which means element 24 is not a member of community 7. Therefore, communities 5
and 6 are required to re-conduct damage detection.
Figure 7.11 shows the results after retaking data and re-conducting damage in
communities 5 and 6. Both LEDs in communities 5 and 6 turn to green which means
element 24 is determined by these communities as undamaged after re-conducting damage
detection.

7.4 Summary

A Simulink and Stateflow model, which simulates a smart sensor network consisting
of eleven sensor communities, has been developed. The proposed DCS approach was
successfully implemented on this sensor network employing both numerical and
experimental data. An detailed overview of major components of the proposed model was
provided. This Simulink and Stateflow model has been shown capable of demonstrating
important features, such as damage information exchange between sensor communities,
providing inconsistent damage information, automatically initiating the process of
retaking data, etc. Numerical and experimental results indicate that the proposed DCS
approach is promising for SHM application on densely distributed sensor networks.

104
 LED1 LED2 LED3 LED4 LED5 LED6 LED7 LED8 LED9 LED10 LED11

8 8 0 0 0 -1 0 0 0 0 0

F_dele1 F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10 F_dele11

8 8 0 0 0 24 0 0 0 0 0
dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 dele11

105
1 -2 -2 -2 -1 -2 -2 -2 -2
1 -2
0 -2 -2 -2 0 -2 -2 -2 -2
PC1
PC2 PC3 PC5 PC6 PC7 PC9 PC10 PC11
PC4 PC8

Figure 7.10: Outputs from Simulink and Stateflow model

when 8 is damaged.
 LED1 LED2 LED3 LED4 LED5 LED6 LED7 LED8 LED9 LED10 LED11

8 8 0 0 0 0 0 0 0 0 0

F_dele1 F_dele2 F_dele3 F_dele4 F_dele5 F_dele6 F_dele7 F_dele8 F_dele9 F_dele10 F_dele11

8 8 0 0 0 0 0 0 0 0 0
dele1 dele2 dele3 dele4 dele5 dele6 dele7 dele8 dele9 dele10 dele11

106
1 -2 -2 -2 -2 -2 -2 -2 -2
1 -2
0 -2 -2 -2 -2 -2 -2 -2 -2
PC1
PC2 PC3 PC5 PC6 PC7 PC9 PC10 PC11
PC4 PC8

Figure 7.11: Outputs from Simulink and Stateflow model when 8 is damaged
(retaking data in communities 5 and 6 regarding element 24).
 CHAPTER 8

CONCLUSIONS AND FUTURE STUDIES

8.1 Conclusions

The research detailed in this report has addressed some of the challenges of applying
SHM to civil infrastructure. Focus has been placed on development of SHM algorithms
suitable for application on densely distributed sensor networks.
An extensive literature review of vibration-based damage detection methods, as well
as recent developments in smart sensor technology, have been provided. To effectively
detect damage at an arbitrary location in a structure, sensors must be densely distributed
throughout the structure. However, most of existing SHM methodologies assume the
measured data to be centrally acquired and processed. Implementation of SHM systems
with central data processing on densely distributed sensor networks is challenging either
using traditionally wired or wireless sensors. By offering distributed computing capacities
and environment, recent fast-growing smart sensor technology has made structural health
monitoring using a dense array of sensors feasible, and therefore has the potential to
change fundamentally the way how civil infrastructure is monitored. SHM algorithms
which can take advantage of the smart sensor technology are desired but currently very
limited. Therefore, development of new SHM algorithms that mesh well with the smart
sensor technology is very important.
A flexibility-based damage localization method, the DLV method that has been
shown to be quite promising, was investigated. Motivation of flexibility-matrix-based
methods was first presented. Because an inverse relationship exists between the flexibility
matrix and the square of the modal frequencies, the flexibility matrix is frequently
insensitive to high frequency modes, which are typically quite difficult to determine
experimentally. Numerical examples were employed to demonstrate this unique
characteristic with results indicating significant potential for practical application of
flexibility-based damage detection approaches. Concepts of the DLV method were then
introduced. The essence of the DLV method is to compute a set of load vectors which
produce zero stress at the damaged region. Numerical examples have been provided to
better understand this method. As the flexibility matrix needs to be constructed for
implementation of the DLV method, formulation of the flexibility matrix for
proportionally damped structures was reviewed. This formulation is based on the
assumption that input excitations are measured and there is at least one co-located sensor
and actuator pair. To assess the efficacy of the DLV method in practice, this method was
experimentally verified using a three-dimensional 14-bay truss structure. Damage was
simulated by replacing the original element with one having a 52.7% cross section
reduction, which results in around a 45% axial stiffness reduction. Structural damage was
successfully identified using only a limited number of frequency modes.
The DLV method was then extended for continuous online damage diagnosis
employing ambient vibration. The essence of the extended DLV method is to construct an

107
approximate flexibility matrix for the potentially damaged structure utilizing the modal
normalization constants from the undamaged structure. The DLV method can be then
applied for continuos online damage diagnosis employing ambient vibration. Different
formulations of the flexibility matrix were reviewed. For the case when input excitations
are measured, the flexibility matrix can be constructed for structures with general viscous
damping; while a mass perturbation method was introduced to compute the flexibility
matrix for proportionally damped structures when input excitations are not available.
Changes of the modal normalization constants due to structural damage were then studied
with results implying that the undamaged normalization constants may provide a
reasonable approximation to the damaged ones when damage is small. Based on these
results, the extension of the DLV method was proposed and illustrated with flow charts.
Finally, the proposed approach was numerically validated using a 14-bay planar truss
structure.
A new distributed computing strategy for SHM based on the extended DLV methods
was then proposed which is suitable for implementation on a network of densely
distributed smart sensors. In this approach, a hierarchical strategy is proposed in which
adjacent smart sensors are grouped together to form sensor communities. The extended
DLV method is employed to evaluate the condition of local elements within these
communities by utilizing only locally measured information. Therefore, only limited
information needs to be transferred to a central station. The concept of locating damage
using local information was first presented, followed by the development of the DCS
approach. Numerical simulations with noise included in measurements were then
presented. Both the excitation magnitude and location were changed before and after
damage to better simulate ambient vibration. Different damage extents were investigated.
Both single and multiple damage scenarios have been studied with results showing the
proposed DCS approach promising for online SHM with a densely distributed sensor
network.
Experimental verification of the proposed DCS approach for SHM was carried out
using a three-dimensional 14-bay truss structure. Damage was simulated by replacing
structure members with the ones having a 52.7% cross section reduction that results in
around a 45% axial stiffness reduction. Both single and multiple damage scenarios were
studied in this experiment. Various excitation conditions, including changing excitation
level, bandwidth and location before and after damage, were also investigated. The
process of retaking data and re-conducting damage detection when there is inconsistent
damage information was studied using experimental data. The experimental results show
that the proposed DCS approach can successfully monitor local community members only
utilizing locally measured information for various damage scenarios under different
excitation conditions.
A Simulink and Stateflow model which simulates a smart sensor network consisting
of eleven sensor communities was then developed to better understand application of the
proposed DCS approach in practice. Numerical and experimental verification of the
proposed DCS approach were successfully conducted in Chapter 5 and 6. However, some
important issues of implementation of the proposed approach in smart sensor networks
were not studied in those two chapters. The proposed Simulink and Stateflow model

108
simulated a smart sensor network, which can be initiated to detect damage upon the user’s
request, and provide damage information and the inconsistent damage information. It has
the capability to allow adjacent sensor communities to communicate with each other, and
automatically initiate those communities with inconsistent damage information to re-
conduct damage detection. This Simulink and Stateflow model is also capable of
illustrating damage detection of both single and multiple damage scenarios. Both
simulation data from a 14-bay planar truss and experimental measurements from a three-
dimensional 14-bay truss structure have been utilized to successfully demonstrate
important features for implementation of the DCS approach on densely distributed sensor
networks.

8.2 Future Studies

Although this research has successfully addressed some of the challenges for
application of SHM algorithms to civil infrastructure, several questions remain. This
section presents a few directions for further research in the future.

8.2.1 Effect of excitation conditions and damage scenarios

The effect of various excitation conditions on the proposed DCS approach should be
further studied. In this research, numerous simulations and experiments have been
successfully carried out to evaluate the performance of the DCS approach. Band-limited
white noise was utilized as excitations for these simulations and experiments. The
proposed DCS approach performed well under these excitations. Questions arise about
how well this method works under different excitations from band-limited white noise,
e.g. non-stationary excitations. More efforts should be directed to the study using various
excitation conditions.
In addition, different damage scenarios should be studied to gain more insight on the
performance of the proposed approach. In both the simulations and experiments which
were conducted in this research, structural damage was simulated by uniformly reducing
the member stiffness along its length. Elastic modulus of structural members was changed
to simulate the damage during simulations; while this goal was achieved in experiments
by replacing original elements with ones having a uniformly smaller cross section. Other
types of damage scenarios, e.g. cracks on structural members, loose-joint at the structural
node, etc. were not studied. Performance of the proposed approach should be studied for
these type of damage scenarios to obtain a better understanding.

8.2.2 Probability analysis

Probability analysis should also be performed to better assess the efficacy of the
proposed DCS approach in practice. Many uncertainties exist when detecting damage
under real conditions. Excitations of real structures can be quite different from the
simulated ones in numerical examples and experiments. Environmental conditions (e.g.
temperature, humidity contents, etc.) can change from time to time. Development of
analytical model for real structures is certainly more difficult than modelling relatively

109
simpler structures in laboratory. Therefore, modelling errors are inevitable. All these
factors create uncertainties which affect the performance the proposed approach.
Modeling these uncertainties and evaluating their effects on the performance of the
proposed DCS approach using numerical methods, e.g. Monte Carlo simulation, is
certainly desired and very useful for better understanding this approach.

8.2.3 Implementation of the DCS approach on smart sensor networks

Implementation of the DCS approach on real smart sensor networks is important to

gain insight on application of the proposed approach in practice. The research detailed in
this report has established the feasibility of the proposed DCS approach through
simulations and experiments. A Simulink model with the support of the Stateflow was
also developed to demonstrate important features of implementing this approach with a
dense array of smart sensors. However, due to the technical limitations of Simulink and
Stateflow, practical issues, such as sensor communications through radio frequency
technology, data synchronization among community sensors, data loss during
communication and so on, could not have been studied. Effects of these practical issues
need to be investigated by implementing the proposed DCS approach on real smart sensor
networks.

8.2.4 Extension of the DCS approach to more complicated structures

Extending the proposed DCS approach to more complicated structures is important.

The proposed DCS approach has been shown to work well for truss structures. However,
real structures consist of different types of members, such as truss members, frame
members, plate members, and so on. It is desirable to extend the proposed approach to
handle more complicated structures.
The essence of the extension is to define appropriate member stresses which are
produced by the DLVs for different type of elements. For truss members, the axial force
can be defined as the generalized member stress when applying the DLV method. For
other type members, different generalized member stress is needed. For example, Bernal
(2002) suggested to 1⁄2
define the member stress for a planar beam element as
2 2
( M i + M j + 2M i M j ) where M i and M j are the two end moments. After appropriate
generalized stresses are defined, the proposed DCS approach can still be applied to detect
damage for different types of structures.

8.2.5 Optimal sensor topology

Optimizing sensor topology in a single community is also desired. For the relatively
simple truss structures studied herein, each community was developed employing a total
of 12 sensors on six structural nodes to monitor nine local structural members, and good
results have been reported. For more complicated structures, each community may need to
cover a larger portion of the structure, and therefore includes more localized sensors. It is
important to study the effect of different sensor topology on the proposed DCS approach.
With an optimal sensor topology, good results can potentially be achieved with less smart

110
sensors. Therefore, communication and computation in a single community can be
reduced, and the efficiency of the sensor network can be improved.

8.2.6 SHM strategies employing multi-scale information

Detecting damage in some structural members is more difficult than others due to
their different contributions to the dynamic behavior of the structure. During the
experimental verification using the DCS approach, detecting damage in the vertical and
diagonal elements located close to the mid-span of the truss structure was found to be
more difficult than other elements. A rational explanation of this phenomenon can be
sought by analyzing the structure under its self-weight. For those elements that carry more
forces under self-weight, their damage is expected to have larger impact on structural
dynamic behavior, and therefore is easier to be detected. Fig. 8.1 shows actual axial forces
in vertical panel of the truss structure (see Fig. 3.10). In this figure, for illustration
purpose, the member width represents magnitude of the axial force. As can be seen, the
vertical and diagonal elements close to the mid-span of the truss carry much smaller forces
compared with others. Therefore, detecting damage in these elements is more challenging.
A detailed sensitivity analysis should be conducted to identify those elements in which
damage is more difficult to detect, so that different SHM strategies could be possibly
developed to monitor these type of structural members.

Figure 8.1: Axial forces under self-weight.

Development of new SHM strategies which do not rely only on dynamic property
changes is therefore desired. As discussed above, damage in some structural elements can
be difficult to detect by vibration-based damage detection methods. These structural
members do not contribute significantly to structural stiffness, but they may still be critical
to the structure (e.g. to maintain the structural stability). Monitoring the condition of these
structural members is also an important issue. Using different types of measurements
other than accelerations (e.g. strain gage measurement) might be more effective in
monitoring these elements. To optimize performance of a SHM system, more effort should
be directed to the development of new SHM strategies which take advantage of multi-
scale information from different types of measurements.

111
 REFERENCES

Adams, R.D., Cawley, P., Pye, C.J., and Stone, B.J. (1978). “A vibration technique for
nondestructive assessing the integrity of structures.” J. of Mech. Engrg. Science,
20:93–100.
Agre, J.R., Clare, L.P., Pottie, G.J., and Romanov, N.P. (1999). “Development platform for
self-organizing wireless sensor networks.” Proc. of SPIE - The International Society
for Optical Engineering, Orlando, Florida, 3713:257–267.
Aktan, A.E., Lee, K.L., Chuntavan, C., and Aksel, T. (1994). “Modal testing for structural
identification and condition assessment of constructed facilities.” Proc. of the 12th
International Modal Analysis Conf., Honolulu, Hawaii, Vol.I, 462–468.
Aktan, A.E., Chase, S., Inman D.J., and Pines, D.J. (2001). “The paradigm of monitoring
and managing the health of infrastructure systems.” A Workshop on Health Monitoring
of Long Span Bridges, Irvine, CA, 1–14.
Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., and Cayirci, E. (2002). “Wireless sensor
networks: a survey.” Computer Networks, 38:393–422.
Alvin, K.F., and Park, K.C. (1994). “Second-order structural identification procedure via
state-space-based system identification.” AIAA J., 32(2):397–406.
Baruh, H., and Ratan, S. (1993). “Damage detection in flexible structures.” J. of Sound
and Vibration, 166:21–30.
Bernal, D. (2000). “Extracting flexibility matrices from state-space realizations.” COST
F3 Conf., Madrid, Spain, 127–135.
Bernal, D. (2002). “Load vectors for damage localization.” J. of Engrg. Mech., 128(1):7–
14.
Bernal, D., and Gunes, B. (2004). “Flexibility based approach for damage
characterization: benchmark application”, J. of Engrg. Mech., 130(1), 61–70.
Bernal, D. (2004). “Modal Scaling from Known Added Masses,” J. of Engrg. Mech.,
130(9), 1083–1088.
Brincker, R., Andersen, P., Kirkegaard, P.H., and Ulfkjær, J.P. (1995). “Damage detection
in laboratory concrete beams.” Proc. of the 13th International Modal Analysis Conf.,
Nashville, Tennessee, Vol.I, 668–674.
Brinker, R., and Andersen, P. (2002). “A way of getting scaled mode shapes in output-only
modal testing”, IMAC-XXI, Orlando, Florida, paper #141.
Brooks, T. (1999). “Using smart accelerometers and wireless interfaces for condition
monitoring.” Machine Plant and System Monitoring, May/June issue.
Carlin, R.A., and Garcia, E. (1996). “Parameter optimization of a genetic algorithm for
structural damage detection.” Proc. of the 14th International Modal Analysis Conf.,
Dearborn, Michigan, Vol.2, 1292–1298.

112
Carrasco, C.J., Osegueda, R.A., Ferregut, C.M., and Grygier, M. (1997a). “Damage
localization in a space truss model using modal strain energy.” Proc. of the 15th
International Modal Analysis Conf., Orlando, Florida, Vol.II, 1786–1792.
Carrasco, C.J., Osegueda, R.A., Ferregut, C.M., and Grygier, M. (1997b). “Localization
and quantification of damage in a space truss model using modal strain energy.” Smart
Structures and Materials 1997: Smart Systems for Bridges, Structures, and Highways,
SPIE, Vol.3043, 181–192.
Cawley, P., and Adams, R.D. (1979). “The location of defects in structures from
measurements of natural frequencies.” J. of Strain Analysis, 14:49-57.
Chen, H.L., Spyrakos, C.C., and Venkatesh, G. (1995). “Evaluating structural deterioration
by dynamic response.” J. of Struct. Engrg., 121:1197–1204.
Chen, S.S., and Kim, S. (1994). “Neural network based signal monitoring in a smart
structural system.” Smart Structures and Materials 1994: Smart Sensing, Processing,
and Instrumentation, SPIE, Vol.2191, 176–186.
Clayton, E.H., and Spencer, B.F., Jr. (2001). “Development of an experimental model for
the study of infrastructure preservation.” Proc. of the National Conf. on
Undergraduate Research (NCUR), Whitewater, WI.
Crossbow Technology Inc. “MICAz wireless measurement system.” http://
www.xbow.com, June, 2004.
Doebling, S.W., Farrar, C.R., Prime, M.B., and Shevitz, D.W. (1996). “Damage
identification 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.
Dos Santos, J.M.C., and Zimmerman, D.C. (1996). “Structural Damage Detection Using
Minimum Rank Update Theory and Parameter Estimation,” Proc. of the AIAA/ASME/
AHS Adaptive Structures Forum, 168–175, AIAA-96-1282.
Dyke, S.J., Caicedo, J.M., and Johnson, E.A. (2000). “Monitoring a benchmark structure
for damage identification.” Proc. of the Engineering Mechanics Specialty Conference,
Austin, Texas.
Farrar, C.R., Doebling, S.W., and Duffey, T.A. (1999). “Vibration-based damage
detection.” Proc. of the Structural Dynamics Forum SD2000, Los Alamos, NM, 11–
17.
Friswell, M.I., Penny, J.E.T., and Lindfield, G. (1995). “The location of damage from
vibration data using genetic algorithms.” Proc. of the 13th International Modal
Analysis Conf., Nashville, Tennessee, Vol.II, 1640–1646.
Fox, C.H.J. (1992). “The location of defects in structures: a comparison of the use of
natural frequency and mode shape data.” Proc. of the 10th International Modal
Analysis Conf., Las Vegas, Nevada, Vol.I, 522–528.

113
Frampton, K.D. (2001). “Decentralized control of structural acoustic radiation.” Proc. of
International Mechanical Engineering Congress and Exhibition, New York, NY,
Vol.28, 232–238.
Gao, Y., and Spencer, B.F., Jr. (2002). “Damage localization under ambient vibration
using changes in flexibility.” J. of Earthquake Engineering and Earthquake Vibration,
1(1):136–144.
Gao, Y., and Spencer, B.F., Jr. (2005a). “Flexibility-based approach for continuous
monitoring of civil infrastructure.” Proc. of the 9th International Conference on
Structural Safety and Reliability, Rome, Italy, June 19-23.
Gao, Y., and Spencer, B.F., Jr. (2005b). “Damage diagnosis of a truss structure employing
a distributed computing algorithm.” Proc. of the 2nd International Conference on
Structural Health Monitoring of Intelligent Infrastructure, Shenzhen, China,
November 16-18.
Gao, Y. and Spencer, B.F., Jr. (2006). “Online damage diagnosis of civil infrastructure
employing a flexibility-based approach.” J. of Smart Materials and Structures, 15:9–
19.
Gao, Y., and Spencer, B.F. Jr. (2007). “Experimental verification of a distributed
computing algorithm for structural health monitoring.” J. of Smart Structures and
Systems, Vol. 3, No. 4.
Gao, Y., Spencer, B.F., Jr., and Bernal, D. (2004). “Experimental verification of the
damage locating vector method.” Proc. of the 1st International Workshop on Advanced
Smart Materials and Smart Structures Technology, Honolulu, Hawaii.
Gao, Y., Spencer, B.F., Jr., and Bernal, D. (2007). “Experimental verification of the
damage locating vector method.” J. of Eng. Mech., 133(10):1043–1049.
Gao, Y., Spencer, B.F. Jr., and Ruiz-Sandoval, M. (2006). “Distributed computing
algorithm for structural health monitoring.” J. of Structural Control and Health
Monitoring, 13(1): 488–507.
Garcia, G., Osegueda, R., and Meza, D. (1998). “Comparison of the damage detection
results utilizing an ARMA model and a FRF model to extract the modal parameters.”
Proc. of SPIE: Smart Systems for Bridges, Structures, and Highways, San Diego, CA,
Vol.3325, 244–252.
Hill, J. and Culler, D. (2002). “Mica: a wireless platform for deeply embedded networks.”
IEEE Micro, Vol. 22, No. 6, November/December, 12–24.
Hjelmstad, K.D. (1996). “On the uniqueness of modal parameter estimation.” J. Sound
Vib., 192(2):581–598.
Ho, Y.K., and Ewins, D.J. (1999). “Numerical evaluation of the damage index.” Proc. of
2nd Int. Workshop on Structural Monitoring, Stanford, CA, 995–1011.

114
Ho, Y.K., and Ewins, D.J. (2000). “On the structural damage identification with mode
shapes.” European COST F3 Conf. on System Identification and Structural Health
Monitoring, Madrid, Spain, 677–686.
Hollar, S. (2000). “COTS Dust.” Master of Science Thesis, University of California at
Berkeley.
James, G.H. III, Carne, T.G., and Lauffer, J.P. (1993). “The natural excitation technique
(NExT) for modal parameter extraction from operating wind turbines.” SAND92-
1666, UC-261, Sandia National Laboratories.
Juang, J.N., and Pappa, R.S. (1985). “An eigensystem realization algorithm for modal
parameter identification and model reduction.” J. of Guidance Control and Dyn.,
8:620–627.
Juang, J.N. (1994). Applied System Identification, New Jersey:Prentice Hall PTR.
Kim, K., Ryu, J., Lee, S., and Choi, L. (1997). “In-situ monitoring of Sungsan Bridge in
Han River with a optical fiber sensor system.” Proc. of SPIE: Smart Systems for
Bridges, Structures, and Highways, San Diego, CA, Vol.3043, 72–76.
Kim, J.-H., Jeon, H.-S., and Lee, C.-W. (1992). “Application of the modal assurance
criteria for detecting structural faults.” Proc. of the 10th International Modal Analysis
Conf., Las Vegas, Nevada, Vol.I, 536–540.
Kling, R. (2003). Intel Research Mote. Intel Corporation Research, Santa Clara,
California.
Ko, J.M., Wong, C.W., and Lam, H.F. (1994). “Damage detection in steel framed
structures by vibration measurement approach.” Proc. of 12th International Modal
Analysis Conf., Honolulu, Hawaii, 280–286.
Lam, H.F., Ko, J.M., and Wong, C.W. (1995). “Detection of damage location based on
sensitivity analysis,” Proc. of the 13th International Modal Analysis Conf., Nashville,
Tennessee, 1499–1505.
Liu, R.C., Zhou, L., Cheng X., and Mau. S.T. (2001). “Wireless sensors for structural
monitoring.” Strong Motion Instrumentation for Civil Engineering Structures, 253–
266.
Luo, H., and Hanagud, S. (1997). “Dynamic learning rate neural network training and
composite structural damage detection.” AIAA J., 35:1522–1527.
Lynch, J.P. (2002). Decentralization of wireless monitoring and control technologies for
smart civil structures. PhD Thesis, Stanford University, CA.
Lynch, J.P., Law, K.H., Kiremidjian, A.S., Kenny, T.W., Carryer, E., and Patridge, A.
(2001). “The design of a wireless sensing unit for structural health monitoring.” 3rd
International Workshop on Structural Health Monitoring, Stanford, California, 1041–
1050.

115
Maeck, J., and De Roeck, G. (1999). “Damage detection on a prestressed concrete bridge
and RC beams using dynamic system identification.” Proc. of the International Conf.
on Damage Assessment of Structures (DAMAS 99), Dublin, Ireland, 320–327.
Mares, C., and Surace, C. (1996). “An application of genetic algorithms to identify
damage in elastic structures.” J. of Sound and Vibration, 195:195–215.
Maser, K., Egri, R., Lichtenstein, A., and Chase, S. (1997). “Development of wireless
global bridge evaluation and monitoring system (WGBEMS).” Proc. of the Specialty
Conf. on Infrastructure Condition Assessment: Art, Science, Practice, 91–100.
Masri, S.F., Nakamura, M., Chassiakos, A.G., and Caughey, T.K. (1996). “Neural network
approach to detection of changes in structural parameters.” J. of Engrg. Mech.,
122:350–360.
Messina, A., Jones, I.A., and Williams, E.J. (1996). “Damage detection and localization
using natural frequency changes.” Identification in Engrg. Systems: Proc. of the
International Conf., Swansea, U.K., 67–76.
Mitchell, K., Sana, S., Balakrishnan, V.S., Rao, V., and Pottinger, H.J. (1999). “Micro
sensors for health monitoring of smart structures.” SPIE Conf. on Smart Electronics
and MEMS, 3673:351–358.
Mitchell, K., Dang, N., Liu, P., Rao, V., and Pottinger H.J. (2001). “Web-controlled
wireless network sensors for structural health monitoring.” Proc. of SPIE the
International Society for Optical Engrg., 4334:234–243.
National Bridge Inventory Data (NBID) (1998). Rep. No.FHWA-PD-96-001, Washington
D.C.
National Research Council Computer Science and Telecommunications Board (2002).
Embedded everywhere, a research agenda for networked systems of embedded
computers. National Research Council, National Academy Press; Washington, D.C.
236.
Ni, Y.Q., Wang, B.S., and Ko, J.M. (1999). “Selection of input vectors to neural networks
for structural damage identification.” Smart Structures and Materials 1999: Smart
Systems for Bridges, Structures, and Highways, SPIE, 3671:270–280.
Pandey, A.K., and Biswas, M. (1994). “Damage detection in structures using changes in
flexibility.” J. of Sound and Vibration, 169(1):3–17.
Pandey, A.K., and Biswas, M. (1995a). “Damage diagnosis of truss structures by
estimation of flexibility change.” Modal Analysis, 10:104–117.
Pandey, A.K., and Biswas, M. (1995b). “Experimental verification of flexibility difference
method for locating damage in structures.” J. of Sound and Vibration, 184:311–328.
Parloo, E., Verboven, P. Cuillame, P., and Overmeire, M.V. (2001). “Sensitivity-based
mass normalization of mode shape estimates from output-only data.” Proc. of the
International Conf. on Structural System Identification, Kassel, Germany, 627–636.

116
Parloo, E., Verboven, P. Cuillame, P., and Overmeire, M.V. (2002). “Iterative calculation
of nonlinear changes by first order approximations.” Mechanical Systems and Signal
Proc., 16(5):757–767.
Peeters, B., Maeck, J., and Roeck, G.D. (2001). “Vibration-based damage detection in civil
engineering: excitation sources and temperature effects,” Smart Materials and
Structures, 10(3):518–527
Polastre, J., Szewczyk, R., and Culler, R. (2005). “Telos: enabling ultra-low power
wireless research.” Proc. of the Fourth International Conf. on Information Processing
in Sensor Networks: Special track on Platform Tools and Design Methods for Network
Embedded Sensors (IPSN/SPOTS), April 25-27.
Pothisiri, T., and Hjelmstad, K.D. (2003). “Structural damage detection and assessment
from modal response.” J. of Engrg. Mech., 129(2):135–145.
Ricles, J.M., and Kosmatka, J.B. (1992). “Damage detection in elastic structures using
vibratory residual forces and weighted sensitivity.” AIAA Journal, 30:2310–2316.
Salawu, O.S., (1995). “Nondestructive assessment of structures using the integrity index
method applied to a concrete highway bridge.” Insight, 37(11):875–878.
Salawu, O.S. (1997). “Detection of structural damage through changes in frequency: a
review.” Engrg. Struct., 19:718–723.
Sheinman, I. (1996). “Damage detection and updating of stiffness and mass matrices using
mode data.” Computers & Structures, 59:149–156.
Shi, Z.Y., and Law, S.S. (1998). “Structural Damage Localization from Modal Strain
Energy Change.” J. of Sound and Vibration, 218(5):825–844.
Shi, Z.Y., Law, S.S., and Zhang, L.M. (2000). “Structural Damage Detection from Modal
Strain Energy Change.” J. of Engrg. Mech., 126(12):1216–1223.
Shi, Z.Y., Law, S.S., and Zhang, L.M. (2002). “Improved Damage Quantification from
Elemental Modal Strain Energy Change.” J. of Engrg. Mech., 128(5):521–529.
Shin, S., and Hjelmstad, K.D. (1994). “Damage detection and assessment of structural
systems from measured response.” Civil Engrg. Studies, SRS 593, UILU-ENG-94-
2013, Univ. of Illinois at Urbana-Champaign, IL.
Sohn, H., Farrar, C.R., Hemez, F.M., Shunk, D.D., Stinemates, D.W., and Nadler, B.R.
(2003). “A review of structural health monitoring literature: 1996-2001.” Los Alamos
National Laboratory Report LA-13976-MS.
Spencer, B.F., Jr., Ruiz-Sandoval, M., and Gao, Y. (2002). “Frontiers in structural health
monitoring.” Proc. of China-Japan Workshop on Vibration Control and Health
Monitoring of Structures and 3rd Chinese Symposium on Structural Vibration Control,
Shanghai, China.
Spencer, B.F., Jr., Ruiz-Sandoval, M.E. and Krata, N. (2004). “Smart sensing technology:
opportunities and challenges.” J. of Structural Control and Health Monitoring,
11:349–368.

117
Spencer, B.F., Jr., and Gao, Y. (2005). “New paradigms for structural health monitoring
employing smart sensor technology.” Proc. of the 2nd International Workshop on
Advanced Smart Materials and Smart Structures Technology, Gyeong-ju, Korea, July
21-24.
Spyrakos, C.C., Chen, H.L., Stephens, J., and Govindaraj, V. (1990). “Evaluating
structural deterioration using dynamic response characterization.” Intelligent Struct.,
edited by K.P. Chong, S.C. Liu and J.C. Li, Elsevier Applied Science, London, 137–
154.
Straser, E.G., and Kiremidjian A.S. (1996). “A modular visual approach to damage
monitoring for civil structures.” Proc. of SPIE v2719, Smart Structures and Materials,
112–122.
Straser, E.G., and Kiremidjian A.S. (1998). “A modular wireless damage monitoring
system for structures.” The John A. Blume Earthquake Engineering Center, Report
No. 128.
Stubbs, N. (1985). “A general theory of non-destructive damage detection in structures.”
Proc. of the 2nd International Symposium on Struct. Control, Ontario, Canada, 694–
713.
Stubbs, N., and Osegueda, R. (1987). “Global nondestructive damage evaluation of
offshore platforms using modal analysis.” Proc. of the 6th International Offshore
Mech. and Arctic Engrg. Symposium, Houston, Texas, Vol.II, 517–524.
Stubbs, N., and Osegueda R. (1990a). “Global non-destructive damage evaluation in
solids.” Modal Analysis: The International J. of Analytical and Experimental Modal
Analysis, 5(2):67–79.
Stubbs, N., and Osegueda R. (1990b). “Global damage detection in solids–experimental
verification.” Modal Analysis: The International J. of Analytical and Experimental
Modal Analysis, 5(2):81–97.
Toksoy, T., and Aktan, A.E. (1994). “Bridge-condition assessment by modal flexibility.”
Experimental Mech., 34:271–278.
Vandiver, J.K. (1975). “Detection of Structural Failure on Fixed Platforms by
Measurement of Dynamic Response,” Proc. of the 7th Annual Offshore Technology
Conf., 243–252.
West, W.M. (1984). “Illustration of the use of modal assurance criterion to detect
structural changes in an orbiter test specimen.” Proc. Air Force Conf. on Aircraft
Struct. Integrity, 1–6.
Williams, E.J., Messina, A., and Payne, B.S. (1997). “A frequency-change correlation
approach to damage detection.” Proc. of the 15th International Modal Analysis Conf.,
Orlando, Florida, Vol.I, 652–657.
Wu, X., Ghaboussi, J., and Garrett, J.H. (1992). “Use of neural networks in detection of
structural damage.” Computers & Structures, 42:649–659.

118
Yao, G.C., Chang, K.C., and Lee, G.C. (1992). “Damage diagnosis of steel frames using
vibrational signature analysis.” J. of Engrg. of Mech., 118(9):1946–1961.
Zimmerman, D.C., Yap, K.C., and Hasselman, T. (1997). “Evolutionary approach for
model refinement.” Proc. of the 15th International Modal Analysis Conf., Orlando,
Florida, Vol.I, 551–557.

119
 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.
007 Kim, Y.S., Seismic Loss Assessment and Mitigation for Critical Urban Jan. 2008
Spencer, B.F., and Infrastructure Systems
Elnashai, A.S.

008 Gourley, B.C., Tort, C., A Synopsis of Studies of the Monotonic and Cyclic Behavior of April 2008
Denavit, M.D., Concrete-Filled Steel Tube Members, Connections, and Frames
Schiller, P.H., and
Hajjar, J.F.

009 Xu, D. and A New Node-to-node Approach to Contact/Impact Problems for May 2008
Hjelmstad, K.D. Two Dimensional Elastic Solids Subject to Finite Deformation

010 Zhu, J. and Non-contact NDT of Concrete Structures Using Air Coupled Sensors May 2008
Popovics, J.S.

011 Gao, Y. and Structural Health Monitoring Strategies for Smart Sensor Networks May 2008
Spencer, B.F.