DECENTRALIZED DAMAGE DETECTION IN CIVIL ......Develop a wireless sensor network philosophy that...
Transcript of DECENTRALIZED DAMAGE DETECTION IN CIVIL ......Develop a wireless sensor network philosophy that...
DECENTRALIZED DAMAGE DETECTION IN CIVIL
INFRASTRUCTURE USING MULTI-SCALE WIRELESS SENSOR NETWORKS
A Dissertation
Submitted to the Graduate School
of the University of Notre Dame
in Partial Fulfillment of the Requirements
for the Degree of
Doctor of Philosophy
by
Su Su, B.S.C.E., M.S.C.E.
Tracy Kijewski-Correa, Director
Department of Civil Engineering and Geological Sciences
Notre Dame, Indiana
July 2011
© Copyright 2011
Su Su
DECENTRALIZED DAMAGE DETECTION IN CIVIL
INFRASTRUCTURE USING MULTI-SCALE WIRELESS SENSOR NETWORKS
Abstract
by
Su Su
The interest in the ability to monitor a structure and detect, at the earliest
possible stage, any damage to it has been pervasive through the Civil Engineering
community, even before the catastrophic collapse of the I-35W Bridge over the
Mississippi in the summer of 2007. This was driven largely by the fact that the current
manual inspection and maintenance philosophy charged with preventing such failures
cannot detect damage in its early stages, and the labor burdens associated with it are
extremely heavy. In response, this dissertation proposed a two stage wireless structural
health monitoring process, including damage detection and localization, to replace the
manual and subjective paradigm. To enhance performance, this dissertation offers a
network architecture that is organized into a multi-scale format, with data fusion of
decentralized real-time damage decisions based on spatially distributed heterogeneous
sensors, operating under a restricted activation scheme and within the computational
constraints of the wireless platform with the objective of minimizing intrusion,
enhancing the reliability of automated detection, maximizing network lifetime and
eliminating the need for strict synchronization and transmission of large amounts of
data.
Thus the primary research tasks in this dissertation can be summarized as:
Develop a wireless sensor network philosophy that provides reliable data for detection and localization of damage in complex Civil Infrastructure, while maximizing the performance and lifetime of the hardware
Develop an assessment framework suitable for damage detection and localization using data measured from a distributed wireless sensors and suitable for operation within said network, i.e., recognizing the computational resources, communications constraints, and power available to the network
Analytically and experimentally verify, at various scales and levels of complexity, the proposed network philosophy and assessment framework.
The result of this effort is number of novel contributions achieved through the
integrated development of the wireless sensing philosophy, network activation scheme
and condition assessment framework to offset inherent limitations of the hardware and
optimize performance for the challenging problem of output only, ambient vibration
monitoring. These contributions include (1) a Bivariate Regressive Adaptive INdex
(BRAIN) for damage detection that proves to be more robust and accurate than previous
formats, (2) a Restricted Input Network Activation Scheme (RINAS) with a new image-
based vehicle classification algorithm that not only reduces the size of reference
databases and enhances detection reliability, but also relieves computational burdens
and extends network lifetime and (3) an offline damage localization technique
employing Dempster-Shafer Evidence Theory that is capable of effectively isolating
damage positions even for minor loss levels. In total, this dissertation offers a definitive
step in translating research to practice to advance the notion of ubiquitous sensing to
address the 21st Centrury Infrastructure Challenges facing society.
ii
This is for My Family
iii
CONTENTS
Figures ................................................................................................................................. vi
Tables .................................................................................................................................. xi
Acknowledgments............................................................................................................. xiii
Chapter 1: Motivation and background ............................................................................. 1
1.1 Background ....................................................................................................... 1
1.1.1 Structural Failures within the Current Inspection Paradigm ............. 2
1.1.2 Current State-of-the-Art in Inspection............................................... 8
1.2 The Need for a Paradigm Shift ........................................................................ 10
1.2.1 Structural Health Monitoring ........................................................... 10
1.2.2 The Role of Wireless Networks ........................................................ 13
1.2.3 Current State-of-the-Art in Wireless Sensor Networks ................... 15
1.3 Objectives of Proposed Research ................................................................... 17
Chapter 2: Test Beds and Hardware ................................................................................. 20
2.1 Bench scale Experimental Datasets ................................................................ 20
2.1.1 LANL Vibrating Disc System ............................................................. 20
2.1.2 LANL Bookshelf Structure ................................................................ 22
2.1.3 Thin Cantilever Beam ....................................................................... 24
2.1.4 Bridge Model .................................................................................... 29
2.2 Simulated Responses ...................................................................................... 39
2.2.1 Benchmark problem by the ASCE Task Group on Health Monitoring39
2.2.2 Thin Cantilever Beam Model ........................................................... 42
2.3 Summary ......................................................................................................... 45
Chapter 3: Overview of WIRELESS SENSOR NETWORK Concept ...................................... 47
3.1 Challenges to WSNs ........................................................................................ 47
3.2 Proposed WSN Monitoring Processes ............................................................ 48
3.3 Restricted Input Network Activation Scheme (RINAS) ................................... 50
3.4 Multi-Scale Network Architecture .................................................................. 53
3.5 Data Fusion within the Network ..................................................................... 57
3.6 Summary ......................................................................................................... 59
Chapter 4: Online Damage Detection ............................................................................... 60
iv
4.1 Time Series Models ......................................................................................... 67
4.1.1 Homogeneous Representations ...................................................... 68
4.1.2 Heterogeneous Representations ..................................................... 69
4.1.3 Performance Assessment ................................................................ 70
4.1.4 Computational Burden..................................................................... 70
4.1.5 Computational Demands ................................................................. 72
4.2 Online Damage Detection ............................................................................... 74
4.2.1 Damage Sensitive Features for Homogeneous Representations .... 74
4.2.1.1 Validation Using Simulated Thin Beam Model ................. 78
4.2.1.2 Validation Using Vibrating Disk Assembly ........................ 82
4.2.1.3 Validation Using LANL Bookshelf Structure ...................... 85
4.2.1.4 Validation Using Steel Truss Bridge Model ....................... 88
4.2.1.5 Validation Using Phase I IASC-ASCE Benchmark Problem 90
4.2.2 Damage Sensitive Features for Heterogeneous Representations ... 94
4.2.2.1 Validation Using Simulated Thin Beam Model ................. 95
4.2.2.2 Validation Using Experimental Thin Beam ..................... 112
4.2.2.3 Validation Using Steel Truss Bridge Model ..................... 114
4.3 Data fusion at the Meso-net ......................................................................... 117
4.4 Summary ....................................................................................................... 119
Chapter 5: RESTRICTED Input Activation Strategies ....................................................... 121
5.1 Camera-Based Traffic Classification with Illustrative Example ..................... 122
5.1.1 Video Conversion ........................................................................... 124
5.1.2 Lane Masking ................................................................................. 124
5.1.3 Background Removal ..................................................................... 125
5.1.4 Noise Filtration ............................................................................... 126
5.1.5 Contour Extraction ......................................................................... 128
5.2 RINAS Concept Verification ........................................................................... 130
5.2.1 Validation Using Vibrating Disk Assembly ..................................... 131
5.2.2 Validation Using Steel Truss Bridge Model .................................... 134
5.3 Summary ....................................................................................................... 138
Chapter 6: Offline damage localization .......................................................................... 140
6.1 Revisiting Damage Localization using AR Model Coefficients ...................... 142
6.2 Introduction to Evidence Theory .................................................................. 145
6.3 Application of Evidence Theory and Proof-of-Concept ................................ 147
6.3.1 Proof-of-Concept for Single Damage Site ...................................... 148
6.3.2 Proof-of-Concept for Multiple Damage Sites ................................ 156
6.4 Weighted Balance Evidence Theory ............................................................. 160
6.5 Summary ....................................................................................................... 164
Chapter 7: CONCLUSIONS AND FUTURE DIRECTIONS .................................................... 166
7.1 Contributions of This Work ........................................................................... 166
7.1.1 Multi-scale Wireless Sensor Network ............................................ 167
v
7.1.2 Restricted Input Network Activation Scheme (RINAS) .................. 167
7.1.3 Data Reduction Using Time Series Models .................................... 168
7.1.4 Data-driven Bivariate Regressive Adaptive Index (BRAIN) ............ 169
7.1.5 Novel Damage Localization Index and Evidence Theory ............... 170
7.2 Future Directions .......................................................................................... 171
7.2.1 Prototype Hardware ...................................................................... 171
7.2.2 Full-scale Validation ....................................................................... 184
7.2.3 Genetic Algorithm Methods for Damage Localization .................. 187
Appendix A: PUBLICATIONS RELATED TO THIS RESEARCH ............................................. 188
Bibliography .................................................................................................................... 190
vi
FIGURES
Figure 1.1: Notable structural failures in recent years: (Left) I-35W Mississippi River bridge (Source: wikipedia.org); (Middle) Lake View Drive Bridge (Source: CBS Broadcasting); (Right) cracked column holding up an I-95 overpass (Source: The Philadelphia Inquirer) ............................................................................................. 7
Figure 1.2: Overview of key features of proposed wireless sensor network for structural health monitoring ................................................................................................. 19
Figure 2.1: LANL eight degree-of-freedom system shaker with accelerometers mounted on each mass (Farrar 1999). ................................................................................. 21
Figure 2.2: Elevation (left) and plan (right) view of the LANL three story frame test structure (Fasel, et al. 2003). ................................................................................ 23
Figure 2.3: Locations of response measurement and load application for cantilever beam specimen: (a) plan view schematic with dimensions; (b) topside with accelerometers; (c) underside with strain gages. ................................................. 26
Figure 2.4: Archer’s trigger used to impart initial displacements to beam: deformed position (left) and released position (right). ......................................................... 28
Figure 2.5: Rendering of thin beam with damage. ........................................................... 28
Figure 2.6: Vertical cantilever beam test with inset photo of base mount. ..................... 29
Figure 2.7: Test Assembly Bridge: (a) 3-D view; (b) Side View; (c) Top view; (d) Bottom View; Note: Units shown are feet [circles denote locations where impacts were imparted]. ............................................................................................................. 31
Figure 2.8: Detailed view of replaceable structural member (right side is before connecting; left side is after connecting). ............................................................ 32
Figure 2.9: Node connections and changeable parts (Dimensions in inches). ................. 32
Figure 2.10: Mode shapes of the bridge model as predicted by SAP 2000. ..................... 33
vii
Figure 2.11: Distribution of sensors on bridge model (=accelerometers, = strain gages, number 1~5 are names of the nodes). ...................................................... 34
Figure 2.12: Bridge model instrumented with accelerometer and strain gages. ............. 34
Figure 2.13: Data acquisition equipment for bridge testing: (a) NI cRIO-9074; (b) NI 9234; (c) NI 9236. ............................................................................................................ 35
Figure 2.14: Excitation equipment for bridge testing: (left) 086C03 impact hammer; (right) K2004E01 electrodynamic vibration shaker. ............................................. 36
Figure 2.15: Examples of acceleration and strain signals with power spectral densities for (a) hammer and (b) shaker tests. .......................................................................... 37
Figure 2.16: Four damage scenarios independently simulated for truss bridge model, shaded circles indicate “damaged” members. ..................................................... 38
Figure 2.17: Schematic of ASCE benchmark building finite element model (Johnson, et al. 2000). .................................................................................................................... 40
Figure 2.18: Six damage patterns in ASCE Benchmark Building (Johnson, et al. 2000). .. 42
Figure 2.19: Damage cases simulated on thin beam model for offline localization proof-of-concept: (a) undamaged beam with element numbering convention, (b)-(g) damage patterns 1-6. ............................................................................................ 45
Figure 3.1: One cycle of structural health monitoring and condition assessment........... 49
Figure 3.2: Two stage process of input selection and restricted activation for assessment. = gateway sensor, = meteorological station, = wireless nodes, = response sensors. ........................................................................................... 52
Figure 3.3: Structural health monitoring and condition assessment period with RINAS (event triggered). .................................................................................................. 53
Figure 3.4: (a) Traditional wired hub and spoke architecture, (b) wireless hub and spoke architecture, (c) proposed multi-scale wireless network. .................................... 54
Figure 3.5: Diagram of two-tiered wireless sensor networks........................................... 57
Figure 3.6: Overview of key features of proposed wireless sensor network for structural health monitoring with addition of new benefits introduced in Chapter 3. ........ 59
Figure 4.1: Relationship between time series coefficients and dynamic properties. ...... 61
Figure 4.2: The relationship among AR coefficients, z-transform poles, dynamic properties, and structural damage. ...................................................................... 64
viii
Figure 4.3: Representation of (a) acceleration signal by (b) AR-ARX, (c) AR, (d) ARMA, and (e) BAR, with residual errors by (f) AR-ARX, (g) AR, (h) ARMA, and (i) BAR. . 72
Figure 4.4: Normal distribution test on homogeneous dynamic DSF for reference pool of undamaged acceleration data. ............................................................................. 77
Figure 4.5: Damage detection rate comparison between static DSF (Grey Bars) and dynamic DSF (Black Bars) based on only acceleration responses of model bridge................................................................................................................................ 90
Figure 4.6: Damage detection results for IASC-ASCE benchmark building. ..................... 92
Figure 4.7: First five normalized mode shapes of simulated thin beam with measurement points superimposed. ......................................................................................... 101
Figure 4.8: Average stiffness lost in the first five modes of simulated thin beam as a function of cross sectional area removed and location of damage (damage pattern). .............................................................................................................. 103
Figure 4.9: Matrix of damage detection rates on simulated thin beam under random excitation results (columns are damage locations, rows are measurement locations). ............................................................................................................ 105
Figure 4.10: Matrix of standard deviation of residual error in underlying regressive model fit to simulated thin beam under random excitation (columns are damage locations, rows are measurement locations). .................................................... 108
Figure 4.11: Definition of damage lengths on thin cantilever beam (plan view). .......... 112
Figure 4.12: Damage detection rate comparison between homogenous dynamic DSF (Grey Bars) and heterogeneous dynamic DSF (Black Bars) on experimental thin beam. .................................................................................................................. 115
Figure 4.13: Damage detection rate comparison between 8th order homogenous dynamic DSF (Grey Bars) and 11th order (na=8, nb=3) heterogeneous dynamic DSF (Black Bars) for experimental thin beam. .................................................... 117
Figure 4.14: Overview of key features of proposed wireless sensor network for structural health monitoring, with addition of new benefits introduced in Chapter 4. ............................................................................................................ 120
Figure 5.1: Example traffic scene (a) before masking and (b) after lane masking. ........ 125
Figure 5.2: Simulation scene (a) before background elimination and (b) after background elimination. ......................................................................................................... 126
ix
Figure 5.3: Simulation scene (a) before noise filtration and (b) after noise filtration ... 127
Figure 5.4: Designation of adjacent pixels defining the bounding circle for contour area extraction. ........................................................................................................... 129
Figure 5.5: Logic tree for contour area calculation......................................................... 130
Figure 5.6: Overview of key features of proposed wireless sensor network for structural health monitoring with addition of new benefits introduced in Chapter 5. ...... 139
Figure 6.1: Examples of over fitting, optimal fitting and under fitting (left to right). .... 144
Figure 6.2: The tree structure of Dempster-Shafer evidence theory data fusion. ......... 147
Figure 6.3: Schematic representation of Evidence Theory applied to single site damage detection in thin beam model. ........................................................................... 150
Figure 6.4: Evidence Theory localization results for damage case 1 (actual damage location at element 4). ........................................................................................ 152
Figure 6.5: Evidence Theory localization results for damage case 2 (actual damage location at element 8). ........................................................................................ 153
Figure 6.6: Evidence Theory localization results for damage case 3 (actual damage location at element 12). ...................................................................................... 154
Figure 6.7: Evidence Theory Localization results for damage case 4 (actual damage location at element 16). ...................................................................................... 155
Figure 6.8: Evidence Theory localization results for damage case 5 (actual damage locations at elements 4 and 13). ......................................................................... 157
Figure 6.9: Evidence theory localization results for damage case 6 (actual damage locations at elements 8 and 13). ......................................................................... 158
Figure 6.10: The weighted tree structure of Dempster-Shafer evidence theory data fusion. ................................................................................................................. 161
Figure 6.11: Damage localization results using unweighted/Dempster (Column 1) and weighted (Column 2) evidence theory. First row is results for damage case 1 and second row is results for damage case 4. ........................................................... 163
Figure 6.12: Overview of key features of proposed wireless sensor network for structural health monitoring, with addition of new benefits introduced in Chapter 6. ..... 165
Figure 7.1: Prototype wireless unit to support -net for structural health monitoring (left) and deployed gateway node or M-node (right) (Source: EmNet LLC). ...... 173
x
Figure 7.2: DT-3716 strain gauge: photo and schematic side and plan views (Source: Columbia Research Labs). ................................................................................... 174
Figure 7.3: Vaisala Weather Transmitter WXT510, interfaced with EmNet gateway in field deployment in Chicago (left) with elevation and plan view schematics (right) (Source: Vaisala Inc.). .......................................................................................... 176
Figure 7.4: Photo of SKY5303V CCTV Camera (Source: Skyway Security). ..................... 178
Figure 7.5: Diagram of Passive RFID tag components. ................................................... 179
Figure 7.6: Illustration of steps in RFID RINAS concept. ................................................. 182
Figure 7.7: Common configuration of different WIM systems. ...................................... 184
xi
TABLES
Table 1.1 ASCE Rating for National Bridges ....................................................................... 3
Table 2.1 Nominal values of LANL eight disc system ........................................................ 22
Table 2.2 Summary of damage cases for LANL three story frame structure ................... 24
Table 4.1 estimated AR COEFFICIENTS DEMONSTRATING THE PERFORMANCE OF EMBEDDED ALGORITHM ON WIRELESS PLATFORM WITH 4 KB OF RAM ............ 74
Table 4.2 Damage detection results for static (Eq. 4.18) and dynamic (Eq. 4.17) DSF for simulated thin beam ............................................................................................. 80
Table 4.3 Damage detection results for static (Eq. 4.18) and Dynamic (Eq. 4.17) DSF for 8DOF system under 3V input voltage level........................................................... 84
Table 4.4 Damage detection results for static (Eq. 4.18) and Dynamic (Eq. 4.17) DSF for 8DOF system under 5V input voltage level........................................................... 85
Table 4.5 Damage detection results for static (Eq. 4.18) and Dynamic (Eq. 4.17) DSF for BOOKSHELF STRUCTURE ....................................................................................... 87
Table 4.6 Damage Patterns of Phase I IASC-ASCE Benchmark Problem and average damage detection rates ........................................................................................ 91
Table 4.7 Damage LOCALIZATION index results for first two damage patterns of IASC-ASCE Benchmark Problem .................................................................................... 94
Table 4.8 Detection results of homogeneous and heterogeneous dynamic DSF for SIMULATED thin beam .......................................................................................... 97
Table 4.9 Summary of detection results for SIMULATED thin cantilever beam: comparison of static homogeneous and homogeneous/heterogeneous dynamic damage sensitive features .................................................................................... 99
Table 4.10 Percent stiffness lost and modal participation factor change (absolute) for each damage pattern for first five modes of SIMULATED thin beam ................ 104
xii
Table 4.11 Comparison of dynamic DSF in homogeneous and heterogeneous formats using experimental thin cantilever beam under white noise excitation ............ 113
Table 4.12 damage detection rate before and after local voting process ..................... 119
Table 5.1 Damage detection results for TRADITIONAL AND RINAS IMPLEMENTATIONS ONVIBRATING DISK ASSEMBLY ........................................................................... 133
Table 5.2 Damage detection results (Damage scenario I) for TRADITIONAL AND RINAS IMPLEMENTATIONS ON Steel Truss Bridge Model ............................................. 136
Table 5.3 Damage detection results (Damage scenario III) for TRADITIONAL AND RINAS IMPLEMENTATIONS ON Steel Truss Bridge Model ............................................. 137
Table 7.1 Accelerometer comparison for different monitoring projects ....................... 172
Table 7.2 DT-3716 strain gauge specifications ............................................................... 174
Table 7.3 Gnode specifications for sewer overflow monitoring (Source: EmNet LLC)... 177
Table 7.4 SKY5303V CCTV Camera specifications. .......................................................... 178
xiii
ACKNOWLEDGMENTS
The author would like to express his gratitude to Dr. Tracy Kijewski-Correa for
supervising this research. Her knowledge, enthusiasm and imagination have been a
constant source of encouragement for me. I gratefully acknowledge the help and
support provided by the other members of the committee, Dr. Ahsan Kareem, Dr.
Panos Antsaklis, and Dr. Yahya Kurama.
All the author’s gratitude and respect is devoted to his beloved wife, Yanxin, for
her unreserved support and patience. Special thanks to my lovely kids, Henry and Kelly,
for the happy time sharing with them.
1
CHAPTER 1:
MOTIVATION AND BACKGROUND
1.1 Background
Even before the catastrophic collapse of the I-35W Bridge over the Mississippi in
the summer of 2007, many within the Civil Engineering community were aware of the
growing deterioration of the nation’s infrastructure. Damage within this infrastructure
can be defined as changes from the original condition that adversely affect current or
future performance by resulting in undesirable responses (Doebling, et al. 1998). While
most infrastructures are damaged at some phase of its operational life, due to a variety
of common progressive causes such as creep, fatigue, weathering or overloading, these
damages are not always disruptive to performance. Damage may also have a sudden
onset result due to natural disasters or man-made actions, e.g., blast or impact.
Understandably, with ample severity, any of these damages may cause loss of the load-
carrying capacity and potential severe consequences, i.e., partial or complete collapse.
The goal of this research is to develop cost-effective, reliable means to autonomously
perform structural damage identification, localization and quantification so as to
accurately assess the current condition of civil infrastructure, enabling the prioritization
of rehabilitation and maintenance efforts.
2
1.1.1 Structural Failures within the Current Inspection Paradigm
The inadequacy of the nation’s current inspection and maintenance programs
has been documented by a variety of indirect and sadly direct measures. These indirect
measures have been offered by agencies such as the American Society of Civil Engineers
(ASCE) who recently awarded national bridges a C “grade,” while the whole
infrastructure system received a D “grade point average” (ASCE 2009), necessitating an
investment of $850 billion for repairs. Table 1.1 summarizes these two most recent
assessments by ASCE underscoring the fact that the response to this issue over the last
few years has been incapable of keeping up with the speed of bridge deterioration. In
particular note that the projected cost to eliminate all bridge deficiencies in the next 20
years has nearly doubled since the 2005 report card (ASCE 2005), underscoring the
mounting consequences of inaction to address the infrastructure crisis.
3
TABLE 1.1
ASCE RATING FOR NATIONAL BRIDGES
Year Subject Grade Comments
2005 Bridges C
It will cost $9.4 billion a year for 20 years to eliminate all bridge deficiencies. Long-term underinvestment is compounded by the lack of a Federal transportation program.
2009 Bridges C
The cost of eliminating all existing bridge deficiencies as they arise over the next 50 years is estimated at $850 billion, equating to an average annual investment of $17 billion.
Notes: A = Exceptional, B = Good, C = Mediocre, D = Poor, F = Failing, I = Incomplete; Source: 2005 Report Card for American Infrastructure; 2009 Report Card for American Infrastructure
These sentiments have been echoed by the Federal Highway Administration
(FHWA), which has cataloged bridges in varying degrees of degradation and disrepair,
termed structurally deficient, as well as many older structures who simply do not meet
current minimum provisions specified in modern design codes, termed functionally
obsolete. According to a FHWA (USDOT 2007) report, as summarized in Table 1.2, there
are more than 25% bridges under the National Bridge Inspection Program that fall into
one or both of these categories. This equates to approximately 150,000 bridges. Rural
bridges tend to have a higher percentage of structural deficiencies, while urban bridges
have a higher incidence of functional issues due to rising traffic volumes. For long-span
bridges, the situation is even worse. More than 40% are either structurally deficient or
functionally obsolete and more than 800 long-span bridges in the national bridge
inventory are classified as fracture-critical (Pines 2002).
4
Based on projections of the FHWA in 2006, removing or repairing all deficient
bridges national wide could take 57 years, and the average annual cost to maintain
highways and bridges is projected to be $78.8 billion from all sources for 2005 to 2024
(Holt, et al. 2008). Sadly, according to the U.S. Department of Transportation Fiscal Year
2009 Budget, only $4.5 billion has been appropriated for the bridge program that
enables states to improve the condition of their bridges through replacement,
rehabilitation, and systematic preventive maintenance (USDOT 2008).
TABLE 1.2
PERCENTAGES OF RURAL AND URBAN BRIDGE DEFICIENCIES, BY NUMBER OF BRIDGES
2002 2004
Rural Bridges
Structurally Deficient 15.1% 14.4%
Functionally Obsolete 11.4% 11.0%
Total Deficiencies 26.5% 25.4%
Urban Bridges
Structurally Deficient 9.2% 8.8%
Functionally Obsolete 21.9% 21.6%
Total Deficiencies 31.2% 30.4%
Total Bridges
Structurally Deficient 13.7% 13.1%
Functionally Obsolete 13.8% 13.6%
Total Deficiencies 27.5% 26.7%
Source: USDOT (2006) Status of the Nation’s Highways, Bridges and Transit
Direct evidence of the inadequacy of our current paradigm is evidenced by the
more than 1,500 bridges that have collapsed in the United States alone between 1966
and 2005 (Reid 2008). Table 1.3 summarizes the most severe bridge collapses in the
United States in last 20 years, which have been documented in both steel and
reinforced concrete bridges. For example, the infamous I-35 bridge in Minnesota, a steel
5
bridge built in 1967 and carrying 140,000 vehicles daily, collapsed suddenly on August 1,
2007 while undergoing repairs (Fig. 1.1), taking the lives of eight citizens and levying an
economic impact of more than $60 million. The bridge was inspected annually and
maintained by the Minnesota Department of Transportation (Mn/DOT). The inspection
carried out June 15, 2006 (no inspection report was completed in 2007 due to the
construction work) found evidence of cracking and fatigue. The bridge was rated as
"structurally deficient" and in possible need of replacement (USDOT 2006). The Federal
National Bridge Inventory database of inspection records showed that the I-35W bridge
was rated at 50/100, which placed it near the bottom of federal inspection ratings
nationwide1.
1 A score below 80 indicates that some rehabilitation may be needed, while a score of 50 or less
shows that replacement may be in order.
6
TABLE 1.3
SEVERE BRIDGE COLLAPSES IN UNITED STATES IN LAST 20 YEARS
Bridge Date Death/injuries
Hatchie River Bridge, Memphis, Tennessee Apr 1st ,1989 8 dead
Cypress Street Viaduct, Oakland, California Oct 17th, 1989 42 dead
CSXT Big Bayou Canot rail bridge, Mobile, Alabama Sep 22nd, 1993 47/103
Hoan Bridge, Milwaukee, Wisconsin Dec 13th, 2000 0/0
Queen Isabella Causeway, Texas Sep 15th, 2001 8 dead
I-40 bridge, Webbers Falls, Oklahoma May 26th, 2002 14 dead
Howard Avenue Overpass, Bridgeport, Connecticut Mar 26th, 2003 0/0
Kinzua Bridge, Kinzua Bridge State Park, Pennsylvania Jul 21st, 2003 0/0
Minneapolis I-35W bridge, Minneapolis, Minnesota Aug 1st, 2007 13/100
Harp Road bridge, Oakville, Washington Aug 15th, 2007 0/0
The Cedar Rapids and Iowa City Railway (CRANDIC) bridge, Cedar Rapids, Iowa Jun 12th, 2008 0/0
9 Mile Road Bridge at I-75Hazel, Park, Michigan Jul 15th, 2009 0/1
San Francisco – Oakland Bay Bridge, California Oct 27th, 2009 0/1
Salem Bridge, Naugatuck, Connecticut Jun 15th, 2010 0/1
Source: Wikipedia.org
Preceding this, on December 27, 2005, a 60 ton concrete fascia beam that was
part of the Lake View Drive Bridge collapsed onto Interstate 70 southwest of Pittsburgh,
injuring two people (Fig. 1.1). The bridge was last inspected as part of national
inspection requirements in March 2004 and found to be structurally deficient at that
time. Progressive deterioration of the bridge had been noted. Since the incident, the
Pennsylvania Department of Transportation (PennDot) has been conducting emergency
inspections of all state-owned bridges of the same design. Then, on March 18, 2008 a
busy two-mile stretch of the elevated Interstate 95 in Philadelphia was subjected to
7
emergency closure to repair a crack in a reinforced concrete support pillar that had
grown from about a half-inch wide and four feet long in October 2007 to a staggering
two inches wide and eight feet long in March 2008 (Fig. 1.1), well before its scheduled
re-inspection. The highway, which ordinarily carries 190,000 vehicles a day, was closed
for at least two days while temporary supports were put in place. Later, the PennDOT
declared that it will take a fresh look at all bridges along the interstate, as a
precautionary measure (NBC 2008).
Figure 1.1: Notable structural failures in recent years: (Left) I-35W Mississippi River bridge (Source: wikipedia.org); (Middle) Lake View Drive Bridge (Source: CBS Broadcasting); (Right) cracked column holding up an I-95 overpass (Source: The Philadelphia
Inquirer)
These examples help to illustrate the gravity of the US infrastructure crisis and
while there is no argument that structures will deteriorate over their lifetime, what
remains a subject of debate is the best means to evaluate this infrastructure on more
regular intervals to identify the onset of damage as early as possible before failure or
the need for costly, expensive replacement ensues. This evaluation is critical to
developing an effective bridge management system to assign budgets for maintenance
and repair of deficient bridges based on rational decision criteria. This dissertation shall
8
suggest, consistent with the growing research and development focus at FHWA, that the
development of an online and near real-time sensing methodology is necessary to
achieve early detection of deficient bridges and enable timely intervention to prevent
catastrophic failure and mitigate expensive repair costs.
1.1.2 Current State-of-the-Art in Inspection
Currently most state-owned bridges are guarded by the National Bridge
Inspection Program (NBIP), established in 1970 by the FHWA. The program requires that
every bridge longer than 20 ft (6.1 m) be inspected at least once every two years.
According to this program, approximately 83% of bridges are inspected once every two
years, 12% are inspected annually, and 5% are inspected on a four-year cycle (Phares, et
al. 2004). A sample visual inspection checklist for a bridge of average length and
complexity is presented in Table 1.4 (USDOT 2004).
TABLE 1.4
SAMPLE INSPECTION STRUCTURAL ITEM LIST
Superstructure Element Beams and girders Floor beams and stringers Trusses Catenary and suspender cables Eye bar chains Arch ribs Frames Pins and hanger plates
Substructure Element Abutments Skewbacks (arches) Slope protection Piers Footings Piles Curtain walls
Typical Defects Corrosion Cracking Splitting Connection slippage Overstress Collision damage
Bridge inspection, in basic terms, is simply an educated assessment and
extrapolation of the condition of a bridge. Even though many highly sophisticated
9
evaluation tools may be available, the most common means of evaluating a bridge
currently is to simply assess the condition visually, which suffers from significant
drawbacks:
Due to high manpower demand such inspections cannot be performed
frequently, implying that damages may not be identified in their earliest stages
Due to the complexity of modern bridges, some critical components or types of
damage do not manifest themselves visually
The current national bridge inspection standards do not require inspectors to be
professional engineers for routine inspection. Furthermore, the inspectors may
not possess the knowledge of structural system, load path, and potential distress
indicators that are helpful to make a better decision
These shortcomings lead to subjective, qualitative and potentially inaccurate
evaluations of bridge safety and reliability. In the US more than 600,000 bridges that are
inspected at least once every two years have their condition rated on a scale from 9
(excellent) to 0 (failed) and this process has demonstrated significant variability.
Specifically, 95% of primary element condition ratings for individual bridge components
will vary within two rating points of the average and only 68% will vary within one point
(Phares, et al. 2004). A testament to the subjectivity of these inspections is the fact that
the I-35W Bridge had just received a comprehensive annual inspection approximately
one year before its collapse, with additional minor inspections in May 2007, just months
before the collapse.
10
While more scientific forms of non-destructive evaluation (NDE), e.g., acoustic or
ultrasonic techniques, magnetic field procedures, radiography, are available, they are
only invoked when visual inspection identifies a significant defect clearly, again noting
that the chance of detecting fatal defects depends entirely on the frequency and
sophistication of the initial inspections and the visibility of the defect.
1.2 The Need for a Paradigm Shift
Due to the inefficiency, subjectivity, infrequency and labor intensity of the
current manual inspection paradigm, it becomes practically and economically infeasible
to precisely detect subtle structural defects. This has spurred the development of
embedded sensor technologies that can autonomously assess the performance and
integrity of structures. The activities on this front have focused both on the hardware
necessary to capture data and relay that information, as well as the algorithms required
to evaluate that data and make appropriate decisions. This field has been generically
labeled Structural Health Monitoring (SHM).
1.2.1 Structural Health Monitoring
SHM aims to provide in-situ measurements of the response of a structure over
time using various sensing elements, which may target a number of observations
including strains, accelerations and deflections. Such deployments of monitoring
systems have been underway on bridges throughout the United States for over a
decade. For example, researchers at the Drexel’s Intelligent Infrastructure Center have
conducted several tests on the Commodore Barry Bridge connecting New Jersey and
11
Pennsylvania to test a supervisory control and data acquisition SHM system for long-
span bridges as early as 1998 (Aktan 2006). In 2001, several short-span bridges in state
of California were monitored to help understand the effects of seismic loading (Chen, et
al. 2006). In 2002, the new Benicia-Martinez Bridge was instrumented and monitored
from construction to completion by the California Department of Transportation
(CalTrans), as feedback for the design of future bridges in this highly seismic area
(Murugesh 2001). During 2003 to 2006, a 26-channel system was installed on the
Vincent Thomas Bridge in Los Angeles, vulnerable to possible earthquakes and
terrorism, and provided a wealth of data pertaining to the effects high traffic volume
(Masri, et al. 2004).
However, these deployments largely focus on proving the efficacy of monitoring
in a general sense or validating underlying design assumptions and behaviors and not to
explicitly alert for damage. However, the more pressing need for SHM in civil
infrastructure is “condition assessment,” which requires not only data acquisition and
processing, but also feature extraction/information condensation in order to permit
quantifiable damage assessment at four levels of precision (Rytter 1993):
Level 1: Existence. Is there damage in the system?
Level 2: Location. Where is the damage?
Level 3: Extent. How severe is the damage?
Level 4: Prognosis. How much useful life remains?
While this four level assessment can be done using a variety of approaches, there are
distinct advantages to performing this condition assessment using vibration-based
12
damage detection, which enables the identification of defects from dynamic response
quantities in an autonomous fashion using embedded sensing technology and the
naturally occurring loads without disrupting service. It is precisely this philosophy that
has received increasing attention in the last twenty years and is similarly adopted in this
dissertation. (Note that specific motivating factors for this approach to condition
assessment are provided in Chapter 2 with the introduction of the sensor architecture.)
Since the Los Alamos National laboratory’s seeded damage tests of the I-40 Bridge
(Farrar and Jauregui 1998 a; Farrar and Jauregui 1998 b), the palette of vibration-based
damage detection algorithms has grown considerably and can be grouped into several
classes, which will be discussed further in this dissertation:
Modal Parameters: These methods start with extracting modal parameters (such as the natural frequencies, mode shapes, and damping ratios) of the structure, and may involve using these values to calculate other physical properties, like stiffness matrix, modal curvature, and Rita vector. The damage indicators seek statistically significant differences between the intact and damaged structures (Doebling, et al. 1998); (Lee and Chung 2000); (Kim, et al. 2003).
Statistical Pattern Recognition: Pattern recognition uses statistical tools (such as response surface, F-statistics, control charts, and statistical energy, etc.) to quantity differences between data from the intact and damaged states (Iwasaki, et al. 2004); (Fugate, et al. 2001)
Time Series Prediction: These methods are typically based on using vibration measurements from a healthy structure to train a neural network to predict the system response. When damaged, there will be a change in the measured system response embodied by the error between the measured and predicted response (Masri, et al. 1996); (Xu, et al. 2003).
“Intelligent Diagnosis”: These methods use a variety of signal-processing and analysis techniques such as wavelets, artificial neural networks and genetic algorithms (Yan, et al. 2007) and provide considerable flexibility for application to wide ranging problems.
13
1.2.2 The Role of Wireless Networks
Because of size and the complexity of modern civil structures, the requisite
sensor densities to perform vibration-based condition assessment to localize minor
levels of damage can be significant, which has been the major economic obstacle to
widespread adoption of this concept. For example, the Wind and Structural Health
Monitoring system used by the Hong Kong Highway Department requires approximately
350 channels (more than 900 sensors) and costs $8 million to monitor the structural
behavior of Tsing Ma suspension bridge that runs between Tsing Yi and Ma Wan Islands
(Wong, et al. 2000). A noteworthy fraction of these costs are consumed in labor
associated with the instrumentation phase. In fact, installation labor costs can approach
well over 25% of the total system cost (Lynch, et al. 2001), and the installation process
itself can be cumbersome and time consuming. CalTrans reported that it costs over
$300,000 just to install a measurement system comprised of 60 to 90 accelerometers on
a bridge. This includes conduit to isolate the instrument cables from the bridge’s harsh
environment, at a price tag of $10 per linear foot (Hipley 2000). Further, the exposed
wires may experience accidental tearing or damage, rodent nibbling and measurement
corruption through signal noise, even when installed internal to the structure. Thus
there may be additional maintenance and repair costs for the instrumentation cables
themselves. Recognizing the recent advances in wireless communications, an alternative
to these expensive cable-based systems has been offered in the form of wireless sensor
networks (WSN) that provide flexible and scalable network architectures to remove
much of the upfront cost and burden associated with installation of a sensor network.
14
Perhaps a more powerful feature of this change in monitoring philosophy is the
presence of embedded computational capabilities within the network to permit a shift
from the hub-and-spoke model where sensors simply collect data and feed the raw
streams to a centralized acquisition system. Within this new paradigm, distributed
computational capabilities are now available at the sensor, permitting varying levels of
local processing and then the transmission of reduced data or extracted quantities to a
centralized server, which can aggregate this information and interface with the user for
final diagnosis, reporting, and alerting.
Thus, the advantages of a wireless approach to SHM, as discussed further by
Lynch, et al. (2004), Kim (2005), Harshvardan, et al. (2006), Kijewski-Correa, et al. (2006
a)and Nagayama and Spencer (2007), can be summarized as follows:
Low Cost: the hardware is relatively cheap and the wireless communications eliminates the need for laying costly, vulnerable cabling
Fast Deployment: since the sensor network does not require any fixed infrastructure or cabling and forms its own network (an ad-hoc network), it can be deployed very quickly
Scalability: The number and location of the sensing sites can be dynamically changed without any efforts to reconfigure the network
Readiness Level: Most of the hardware is “off-the-shelf” and the only significant development effort required is a signal conditioning circuit (and this is no different from traditional wired sensing)
Low Maintenance & Operating Cost: since sensor nodes consume very little power, are robust, and can be reprogrammed and calibrated wirelessly from a remote location, they require very little on-site maintenance
Even with these advantages, there are still acknowledged disadvantages that
require users to adjust their algorithms to operate within specified constraints on
15
communication bandwidth, communication reliability (packet loss), wireless signal
strength (requirement for repeater nodes), network synchronization, power, sensor
sensitivity, and local computational and memory resources. These constraints will all be
considered in crafting the framework in this dissertation. Additionally, even with the
advances in wireless communications and low cost distributed computational
capabilities, widespread adoption also requires affordable, high quality sensing
elements. For example, high precision, force balance accelerometers can cost on the
order of $1000, which reduces the economic feasibility of high density, high sensitivity
wireless sensing solutions for civil infrastructure. However, continuing advances in
micro-electromechanical systems (MEMS) devices that retail at a tenth of this cost
suggest that with time, high sensor density with the requisite sensitivity and low-
frequency detection capability can be achieved, opening the possibility of ubiquitous
monitoring to a wider cross section of our civil infrastructure.
1.2.3 Current State-of-the-Art in Wireless Sensor Networks
The movement toward wireless sensing began by leveraging commercially-
available wireless communications technologies: Pines and Lovell (1998, 1999) discussed
an approach using sensors and wireless communications to monitor the health of large
civil structures remotely using spread-spectrum wireless modems and conventional
strain sensors. Around the same time, Krantz et al. (1999) developed a micro-sensor
that could retrieve data from embedded strain gauges. Lemke (2000) described a
remote vibration monitoring system using Wi-Fi technology to leverage the internet in
16
data acquisition, while Oshima et al. (2000) presented a monitoring system that could
be interrogated via a mobile telephone.
Although a number of commercial technologies have been used, major
advancements in ubiquitous sensing have been enabled by WSNs using band limited
wireless radio transmitters on a compact computational platform or mote. While the
wireless mote was popularized in research circles by the Berkeley Mote (Hill, et al.
2000), some of the pioneering efforts in the application to SHM for Civil applications
came out of Stanford in the late 1990s (Straser, et al. 1998). Subsequently, the Berkeley
mote was commercialized (see Crossbow Technology’s MICA mote), and this and other
proprietary designs received increasing attention by a number of researchers, e.g.,
Mitchell et al. (2000), Lynch et al. (2002), Nagayama and Spencer (2007), and Bischoff et
al. (2007). While many of these systems have only been validated in laboratory settings,
field validations have been conducted as early as 1996 (Maser, et al. 1996). Later Straser
and Kiremidjian (1998) and Lynch et al. (2003) used the Alamosa Canyon Bridge in
southern New Mexico to validate the performance of their proprietary wireless sensor
network. The excursions into full-scale monitoring of bridges using WSNs have
continued, e.g., Meyer et al. (2006) on the Stork Bridge in Switzerland, Lynch et al.
(2006a) on the Geumdang Bridge in Korea and Liu et al. (2010) on the main span of the
ZhengDian Viaduct in Wuhan, China. Although these deployments were not permanent,
they made important contributions to the field validation of WSNs and underscored
many practical issues that require attention in full-scale deployment and operation.
However, recently Jo et al. (2011) deployed the world's largest hybrid WSN on the Jindo
17
Bridge, a cable-stayed bridge located in South Korea and experimentally verified several
key features through a long-term monitoring program.
1.3 Objectives of Proposed Research
These previous studies have demonstrated that several issues must still be
addressed to achieve viable WSNs for SHM on Civil Infrastructure: (1) development of
appropriate hardware, wireless communications protocols and network architectures,
(2) development of appropriate online damage detection algorithms that will operate
within the proposed wireless sensor networks to make an initial in-situ evaluation of
the structure using this hardware, and (3) development of offline damage localization
algorithms that will further refine the assessments on the collected data outside the
wireless sensor network. It is immediately acknowledged that the development of
hardware and wireless communications protocols shall not be the objective of this
research and requires partnerships with collaborators trained in Electrical Engineering.
Instead, the focus of this dissertation is the development of a wireless sensor network
concept well-suited to damage detection in Civil Infrastructure and the development of
vibration-based assessment methodologies that can detect, localize, and quantify
damage in its early stages within this framework. As such, the main objectives of this
research are:
Develop a wireless sensor network philosophy that provides reliable data for detection and localization of damage in complex Civil Infrastructure, while maximizing the performance and lifetime of the hardware
Develop an assessment framework suitable for damage detection and localization using data measured from a distributed wireless sensors and
18
suitable for operation within said network, i.e., recognizing the computational resources, communications constraints, and power available to the network
Analytically and experimentally verify, at various scales and levels of complexity, the proposed network philosophy and assessment framework.
The subsequent chapters of this dissertation will address specific research
activities associated with these objectives, each contributing to a different stage in the
health monitoring process, as shown in Figure 1.2. This figure will be re-introduced in
each chapter and progressively completed to reiterate each chapter’s contributions,
beginning first in Chapter 3 with the introduction of the network architecture, its unique
features and benefits for damage detection of civil infrastructure, and the constraints
this network imposes on any assessment algorithm. However, this must be first
preceded by an introduction to the test beds and hardware will be used to validate
various concepts (Chapter 2). The dissertation will then present online condition
assessment methodologies (Chapter 4), input activation strategies (Chapter 5), and
offline damage localization methodologies (Chapter 6), concluded by Chapter 7 with a
foreshadowing of future work.
19
Figure 1.2: Overview of key features of proposed wireless sensor network for structural health monitoring
APPROACH BENEFIT STAGE
DATA ACQUISITION
DATA REDUCTION
DETECTION
LOCALIZATION
20
CHAPTER 2:
TEST BEDS AND HARDWARE
Before presenting various damage detection schemes, it is necessary first to
introduce a variety of test beds, ranging from simulated responses to bench scale
experiments that will be used in their validation. The following sections will summarize
these test beds and the associated hardware for all experimental programs, as well as
simulated systems.
2.1 Bench scale Experimental Datasets
In order to validate the proposed methodologies against other published results,
publically available experimental datasets will be used, e.g., those from Los Alamos
National Laboratory (LANL), as appropriate. However, due to their sole use of
accelerometer data, additional experiments were also developed, as discussed herein.
These experimental datasets are now described.
2.1.1 LANL Vibrating Disc System
The LANL vibrating disc system is formed by eight translating masses
interconnected by springs (Farrar 1999). Each mass is a disc of aluminum 25.4 mm thick
and 76.2 mm in diameter with a center hole. The hole is lined with a Teflon bushing to
minimize frictional losses. There are small steel collars on each end of the discs. The
21
masses all slide on a highly polished steel rod that supports the masses and constrains
them to translate along the rod. The masses are fastened together with coil springs
epoxied to the collars that are, in turn, bolted to the masses to form the assembly in
Figure 2.1.
Figure 2.1: LANL eight degree-of-freedom system shaker with accelerometers mounted on each mass (Farrar 1999).
The undamaged configuration of the system is the state for which all springs
have identical linear springs. Various damage scenarios were generated by LANL
researchers (replacing springs at select locations with those with lesser spring constants),
as summarized in Table 2.1. The responses of the system are generated by exciting
Mass 1 using either an impact hammer or a 215-N (50 lb) peak force electro-dynamic
shaker (Figure 2.1). The acceleration responses of all the masses and the excitation force
were recorded.
22
TABLE 2.1
NOMINAL VALUES OF LANL EIGHT DISC SYSTEM
Mass (g) Spring Properties (Damage Between Mass 5 and Mass 6)
Mass 1 Mass 2-5 Undamaged Damaged 1 Damaged 2 Damaged 3
559.3 419.4 Constant 56.7 kN/M 52.6 kN/M 49.0 kN/M 43.0 kN/M
Reduction 0% 7% 14% 24%
2.1.2 LANL Bookshelf Structure
The three story frame structure shown in Figure 2.2 features aluminum plates
(floors) connected to the unistrut columns by a pair of bolts (Fasel, et al. 2003). All
bolted connections were tightened to a torque of 220 inch-pounds in the undamaged
state. Four Firestone air mount isolators, which allow the structure to move freely in
horizontal directions, were bolted to the bottom of the base plate. The structure was
instrumented with 24 piezoelectric accelerometers: 2 accelerometers were placed at
each joint with one accelerometer attached to the plate and the other accelerometer
attached to the unistrut column. A stringer was then used to connect a shaker to the
structure to simulate both translational and torsional motions. The input excitation via
shaker was band limited white noise (0 to 200 Hz).
23
Figure 2.2: Elevation (left) and plan (right) view of the LANL three story frame test structure (Fasel, et al. 2003).
In this experiment, damage was simulated in joints through the loosening of the
preload applied by the bolts at the joints of the structure. A “healthy” joint is held
together by bolts that are torqued to a value of 220 inch-pounds. Multiple damage
levels are then used so that the sensitivity of the damage detection method can be
tested. The first damage level is simulated by loosening the preload on the bolts at the
selected damaged joint to 15 inch-pounds. The next level has the preload being
loosened to 5 inch-pounds. Bolts on the selected joint are then completely removed to
simulate a crack in the joint for the final damage level. Various damage scenarios using
this methodology were implemented, as summarized in Table 2.2.
24
TABLE 2.2
SUMMARY OF DAMAGE CASES FOR LANL THREE STORY FRAME STRUCTURE
Location Applied Bolt Torque (Nm)
Undamaged - 220
Damage Case 1 Joint 2a 15
Damage Case 2 Joint 2a 5
Damage Case 3 Joint 2a 0 (bolts removed)
Damage Case 4 Joint 4b 15
Damage Case 5 Joint 4b 5
Damage Case 6 Joint 4b 0 (bolts removed)
Damage Case 7 Joint 2a and 4b 15
Damage Case 8 Joint 2a and 4b 5
Damage Case 9 Joint 2a and 4b 0 (bolts removed)
2.1.3 Thin Cantilever Beam
Unfortunately, the experimental datasets from LANL recorded only acceleration
responses. This research will explore the merits of heterogeneous sensing
(incorporation of multiple sensing modalities) in detecting minor levels of damage. As
such, a simple structure is first proposed to generate simultaneously recorded strain and
acceleration data. The system under consideration is a thin, cantilever beam made of
alloy aluminum 6063, with manufacturer-specified density of 2700 and Young’s
modulus experimentally identified as 69000 Mpa. Loading/initial displacements of the
beam for all experimental investigations were applied at point E in Figure 2.3a
(approximate free end), while all responses were measured at locations A-D. Seven
specimens were fabricated with identical properties and each specimen was firmly
clamped at one end using a notched steel plate assembly. On the top side of each beam,
at the four locations shown in Figure 2.3b, accelerometers were attached manufacturer-
3/ mkg
25
supplied mounts and epoxy. PCB piezotronics ceramic shear ICP® accelerometers
(model 333B52) were employed with sensitivities of approximately 1 V/g. Accelerations
were acquired through a pair of Spectral Dynamics SigLab 20-42 units, multiplexed to
provide up to 8 synchronized channels for testing over a 20 kHz bandwidth. These units
have 13 user-selectable sampling rates from 5.12 Hz to 51.2 kHz, 16-bit A/D conversion
and offer a number of built-in signal processing tools, including adjustable anti-aliasing
filters at each of the pre-defined sampling rates. On the underside side of the beam, at
each of the instrumentation points in Figure 2.3c, uniaxial strain gages were attached.
Strain gages used in this test are Vishay Micro-Measurements C2A-13-250LW-350, pre-
fabricated with 3 m of 330-DFV cable attached by integral solder tabs. Strains were
acquired by the StrainSmart System 5000 from Vishay (Model 5100b) with available
slots for up to four model 5110A Strain Gage Cards for up to 20 channels of strain data.
It has a usable digital resolution of approximately 15 bits with a 40 μs total conversion
time per sample. The device can scan all strain channels within 1ms; generally, 50
complete scans are recorded per second. Its normal measurement range is ±16 380με
with a resolution of 1με. The specifications of all sensors are summarized in Table 2.3.
26
Figure 2.3: Locations of response measurement and load application for cantilever beam specimen: (a) plan view schematic with dimensions; (b) topside with accelerometers; (c) underside
with strain gages.
AA BB CC DD E
12.125cm 12.125cm 12.125cm 12.125cm 1.25cm
(c)
(
a)
(
a)
Undamaged Beam
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
(
g)
Da
m
(b)
(a)
(
a)
27
TABLE 2.3
SPECIFICATIONS OF SENSORS USING BY THIN BEAM TEST BED
Sensor Manufacturer Model Gauge Factor Resistance Strain limit
Strain gage Vishay C2A-Series 2.13±0.5% 350.0±0.6% ±2%
Sensor Manufacturer Model Sensitivity Range Bias level
Accelerometer PCB 333B52 100mV/g 0.5-3000 Hz 11.5V
The free vibration response of each beam is recorded following an initial
displacement imparted using a thin nylon line and archer’s trigger, as shown in Figure
2.4, which provides a repeatable, nearly instantaneous release for the system. An HS25
linear variable displacement transducer (LVDT) from Measurements Group Inc. was
used to measure the initial displacement of the beam tip. All specimens were repeatedly
tested in their undamaged condition to form an undamaged reference database.
Damage is then introduced to the beam through a transverse cut, symmetrically
imparted midway between two measurement points. This is demonstrated in Figure 2.5
for the case of damage between points A and B. The transverse dimension of the cut is
specified as a percentage (for example p=20%) of the total width of the beam; the
longitudinal dimension of the cut is fixed at 5% of the total length of the beam for this
study (WD = 0.05L = 2.5 cm), as shown in Figure 2.5.
28
Figure 2.4: Archer’s trigger used to impart initial displacements to beam: deformed position (left) and released position (right).
Figure 2.5: Rendering of thin beam with damage.
Due to a lack of actuators, the dynamic testing of the beam specimens was
executed using random base excitations. To do so, each thin beam specimen was
vertically mounted to a small shaking table, as shown in Figure 2.6. Each specimen is
first excited in its undamaged state under simulated white noise base excitations to
form the reference database that will be used in subsequent statistical significance tests.
Samples were acquired at 50 Hz rate for strain and a 51.2 Hz rate for acceleration.
Acquired acceleration data was then digitally interpolated for a unified sampling rate of
50 Hz. Undamaged tests were run 20 times under independent random excitations to
29
form the undamaged reference pool. Cuts were then introduced to the beams to
simulate damage, as shown previously in Figure 2.5, and the damaged specimens were
excited by 10 independent random excitations to confirm the repeatability of the results.
Figure 2.6: Vertical cantilever beam test with inset photo of base mount.
2.1.4 Bridge Model
As discussed previously, although there are several experimental damage
detection datasets available in the literature, e.g., the Vibrating Disc System by LANL lab
(Farrar 1999) and the truss structure by UIUC (Nagayama and Spencer 2007), none of
30
those test beds are suitable to validate the enhancements offered by heterogeneous
sensing as they lack two basic requirements:
Synchronized high quality acceleration and strain time histories
Sufficient amounts of data (both undamaged and minor damaged)
As such, a new bridge test bed was constructed in Notre Dame’s DYNAMO Lab to
provide an appropriate experimental venue to validate heterogeneous sensing concepts.
The model truss bridge is shown schematically in Figure 2.7 and spans ten feet, with a
one foot width and one foot height. The bridge is comprised of two primary trusses with
single diagonal elements, interconnected at the top and bottom by a series of X-braces,
simply supported at its ends. The bridge was fabricated using 12L14 Carbon Steel Square
Bars (1/4"x1/4”) welding all joints.
Each bar comprising the primary trusses of the bridge is designed to have an
interchangeable segment, connected together by small bolts as shown in Figure 2.8.
Damage is simulated by these segments at various locations with a reduced cross-
section member, as shown in Figure 2.9, which shows the original (undamaged) member
(1/4"×1/4") and two reduced section members (1/4"×3/16", 1/4"×1/8"). The assembly
has the capability of simulating a variety of damage scenarios using this approach at
various locations.
31
Figure 2.7: Test Assembly Bridge: (a) 3-D view; (b) Side View; (c) Top view; (d) Bottom View; Note: Units shown are feet [circles
denote locations where impacts were imparted].
1
2 3
4 5
32
Figure 2.8: Detailed view of replaceable structural member (right side is before connecting; left side is after connecting).
Figure 2.9: Node connections and changeable parts (Dimensions in inches).
33
A finite element model of the undamaged bridge was constructed in SAP 2000,
with natural frequencies listed in Table 2.4 and mode shapes shown in Figure 2.10. Dead
load static and moving load dynamic analyses were then used to identify the response
levels of the model to determine optimal settings for the data acquisition system. Based
on these observations, a 2 kHz sampling rate was selected for the experiments to ensure
the first 5 modes of the response are captured.
TABLE 2.4
NATURAL FREQUENCIES OF SAP MODEL AND TEST BRIDGE ASSEMBLY
1st Mode 2nd Mode 3rd Mode 4th Mode 5th Mode
SAP Model 119.7 Hz 202.5 Hz 378.5 Hz 536.6 Hz 767.4 Hz
Actual Model 132.3 Hz 179.4 Hz 382.8 Hz 501.5 Hz 820.7 Hz
Figure 2.10: Mode shapes of the bridge model as predicted by SAP 2000.
34
Figure 2.11 shows the distribution of accelerometer and strain gages on the
model bridge, again using 333B52 shear type accelerometers by PCB and C2A-13-
250LW-350 strain gages by Vishay Micro-Measurements whose properties were
previously reported in Table 2.3. Figure 2.12 shows the final configuration of the bridge
model with the different sensors attached.
Figure 2.11: Distribution of sensors on bridge model (=accelerometers, = strain gages, number 1~5 are names of
the nodes).
Figure 2.12: Bridge model instrumented with accelerometer and strain gages.
Simultaneous collection of acceleration and strain at 2 kHz was achieved using
the National Instruments NI cRIO-9074 integrated system featuring a 400 MHz real-time
processor and an 8-slot chassis with an embedded, reconfigurable 2M gate FPGA chip.
Two National Instruments 9234 four-channel dynamic signal acquisition modules are
35
used for making high-accuracy measurements from Integral Electronics Piezoelectric
sensors. The first is the NI 9234 acceleration module and the second is the NI 9236, 8
channel C Series analog input module suitable for medium to high-channel-count strain
measurements with built-in voltage excitation for quarter-bridge sensors. Figure 2.13
shows the data acquisition hardware from National Instruments, Inc.
Figure 2.13: Data acquisition equipment for bridge testing: (a) NI cRIO-9074; (b) NI 9234; (c) NI 9236.
Two excitation schemes will be utilized for this system. The first is impulse
testing using a PCB, Inc. impact hammer shown in Figure 2.14 (model 086C03). Each
round of impulse testing is conducted using the same human operator with consistent
arm strength. Impulses are imparted independently at the intersection points of the
36
crossover bars on the top side of the bridge, as marked by solid circles on Figure 2.7.
The second form of excitation is white noise with constant amplitude in the range of 0
to 11k Hz imparted by a Modal Shop, Inc. electrodynamic vibration shaker also shown in
Figure 2.14 (model K2004E01). The properties of the shaker are shown in Table 2.5. The
bridge was excited by placing the shaker at the middle point of the bottom surface, as
shown in Figure 2.7 Figure 2.16 shows the examples of collected undamaged
acceleration and strain data for both the impulse response and random excitation
testing.
Figure 2.14: Excitation equipment for bridge testing: (left) 086C03 impact hammer; (right) K2004E01 electrodynamic vibration
shaker.
TABLE 2.5
PROPERTIES OF THE ELECTRODYNAMIC VIBRATION SHAKER
Output Force, (sinusoidal)
Output Force, (random)
Output Force, (shock)
Stroke Length Frequency Range
4.5 lbf 3 lbf 9 lbf 0.2 in DC-11 kH
37
(a)
(b)
Figure 2.15: Examples of acceleration and strain signals with power spectral densities for (a) hammer and (b) shaker tests.
0 0.2 0.4 0.6 0.8 1-0.2
-0.1
0
0.1
0.2
Time(s)
Accele
ration(m
/s2)
0 200 400 600 800 10000
5
10
Frequency(Hz)
Fre
quency C
onte
nt
0 0.2 0.4 0.6 0.8 17.8
8
8.2
8.4x 10
-5
Time(s)
Str
ain
0 200 400 600 800 10000
1
2
3
4x 10
-4
Frequency(Hz)
Fre
quency C
onte
nt
0 0.2 0.4 0.6 0.8 1-2
-1
0
1
2
Time(s)
Accele
ration(m
/s2)
0 200 400 600 800 10000
10
20
30
40
Frequency(Hz)
Fre
quency C
onte
nt
0 0.2 0.4 0.6 0.8 17
7.5
8x 10
-5
Time(s)
Str
ain
0 200 400 600 800 10000
1
2
3
4x 10
-4
Frequency(Hz)
Fre
quency C
onte
nt
38
By replacing some of the original 1/4"×1/4" section bars with 1/4"×1/8" bars on
one side of the bridge, as shown in Figure 2.17, the following four damage scenarios are
generated:
Damage scenario I: replacing 2 structural members (D1 in Figure 2.17)
Damage scenario II: replacing 4 structural members (D2 in Figure 2.17)
Damage scenario III: replacing 6 structural members (D3 in Figure 2.17)
Damage scenario IV: replacing 8 structural members (D4 in Figure 2.17)
Figure 2.16: Four damage scenarios independently simulated for truss bridge model, shaded circles indicate “damaged” members.
D2
D1
D3
D4
39
2.2 Simulated Responses
Many of the initial validations of the damage detection frameworks introduced
in this research are executed using simulated data. The underlying models used to do so
are now introduced.
2.2.1 Benchmark problem by the ASCE Task Group on Health Monitoring
To coordinate research activities in the area of damage detection, a benchmark
problem was proposed by the ASCE Task Group on Health Monitoring (Johnson, et al.
2000). The structure is a 4-story, 2-bay by 2-bay steel-frame (Figure. 2.18) quarter scale-
model that was erected in the Earthquake Engineering Research Laboratory at the
University of British Columbia (UBC) (Black and Ventura, 1998). It has a 2.5m×2.5m plan
and is 3.6m tall. The members are hot rolled grade 300W steel (nominal yield stress 300
MPa or 42.6 kpi). The sections are unusual, designed for a scale model, with properties
as given in Table 2.6. The columns are all oriented for strong axis bending in the y
direction. On each floor of each exterior face, there are two diagonal braces that may be
removed to emulate damage. To make the mass distribution more realistic, one floor
slab is placed in each bay: four 800 kg slabs at the first floor, four 600 slabs at the
second and third levels, and four 400 kg slabs on the fourth floor. The force input to the
structure is provided by a Ling Dynamic Systems electrodynamics shaker. The command
to the shaker is a band limited white noise with content between 4.6875–30Hz.
Accelerometers and displacement transducers were placed throughout the structure.
40
TABLE 2.6
PROPERTIES OF STRUCTURAL MEMBERS
Property Columns Beams Braces
Section type B100×9 S75×11 L25×25×3
Cross-section area A (m2) 1.133×10-3 1.43×10-3 0.141×10-3
Moment of inertia (strong direction) Iy (m4) 1.97×10-6 1.22×10-6 0
Moment of inertia (weak direction) Ix (m4) 0.664×10-6 0.249×10-6 0
Torsional constant J (m4) 8.01×10-9 38.2×10-9 0
Young’s modulus E (Pa) 2×1011 2×1011 2×1011
Shear modulus G (Pa) E/2.6 E/2.6 E/2.6
Mass per unit volume (kg/ m3) 7800 7800 7800
Figure 2.17: Schematic of ASCE benchmark building finite element model (Johnson, et al. 2000).
41
Based on these measurements, two finite element models were developed to
generate the simulated benchmark acceleration response data; however only the
12DOF shear-building model will be used in this research. This model constrains all
motion except the two horizontal translations and one rotation per floor. The columns
and floor beams are modeled as Euler-Bernoulli beams, and the braces are truss
elements with no bending stiffness. Structural damage was simulated by modifying the
stiffness of various elements in the finite element model. The six damage patterns
defined in the simulated benchmark data are shown in Figure 2.19, where dashed
elements indicate those that are affected, and summarized here:
Damage Pattern I: No stiffness in the braces of the first story (the braces still contribute mass, but provide no resistance within the structure);
Damage Pattern II: No stiffness in any of the braces of the first and third stories;
Damage Pattern III: No stiffness in one brace in the first story (north brace on the west face of the structure;
Damage Pattern IV: No stiffness in one brace in the first story (north brace on the west face) and in one brace in the third story (west brace on the north face);
Damage Pattern V: The same as Damage Pattern IV but with the north floor beam at the first level on the west face of the structure unscrewed from the northwest column; consequently, the beam–column connection there can only transmit forces and cannot sustain any bending moments;
Damage Pattern VI: Two thirds stiffness (a one-third stiffness loss) in one brace in the first story (the same brace damaged in Damage Pattern III)
42
Figure 2.18: Six damage patterns in ASCE Benchmark Building (Johnson, et al. 2000).
2.2.2 Thin Cantilever Beam Model
A finite element model of the thin beam introduced previously in Section 2.1.3
was also created to permit simulation of other loading conditions and damage levels
difficult to achieve experimentally, since the ASCE Benchmark Building presented in the
previous section does not include simulated strain responses. The model was calibrated
using the free vibration tests on the undamaged beam specimens. The beam is initially
modeled using a series of 100 finite elements interconnected only at the nodes, with
43
each node assumed to have only two degrees-of-freedom: a transverse displacement
and rotation. Each element has a length of 0.025 m; cross sectional area of 1.21×10-4 m2,
moment of inertia of 2.34×10-10 m4, modulus of elasticity of 6.87×1010 N/m2, and mass
density of 2700 kg/m3. To validate online damage detection, this model is compromised
in the same manner described in Section 2.1.3, with various degrees of damage (p = 0,
10, 30%) for a cut introduced at LD = 18.75 cm, midway between points A and B (see
Figure 2.5).
For the damage localization part in Chapter 6, identical similar FEM cantilever
beam is simulated, with a reduced number of elements (20) to ease the computational
burden, as shown in Figure 2.20a. To validate offline damage localization, six additional
damage scenarios are introduced, as shown in Figure 2.20b-f. In each of these cases,
damage is introduced through a transverse cut, symmetrically imparted through the
selected finite element(s). The transverse dimension of the cut is selected to achieve a
specified percentage reduction of the width of that element, while the longitudinal
dimension of the cut is fixed at the total length of that finite element. The affected
elements and percentage reductions are summarized in Table 2.7. Note that the choice
to reduce the cross section at element 4, which is closer to the fixed end, less than other
locations was a conscious effort. For all damaged and undamaged cases, acceleration
and surface strain time history pairs are simulated at the four locations (A-D) shown
previously in Figure 2.5, under the action of Gaussian white noise inputs applied at the
free end.
44
TABLE 2.7
SUMMARY OF DAMAGE LEVELS IMPARTED TO THIN BEAM MODEL FOR LOCALIZATION PROOF-
OF-CONCEPT
Damage Case 1 2 3 4 5 6
Affected Element(s) 4 8 12 16 4, 13 8, 13
Percentage Reduction in Each Element
12.5% 25% 25% 25% 12.5% 25%
45
Figure 2.19: Damage cases simulated on thin beam model for offline localization proof-of-concept: (a) undamaged beam with
element numbering convention, (b)-(g) damage patterns 1-6.
2.3 Summary
Table 2.8 summarizes the various test beds introduced in this chapter and the
type of validation that will be conducted with each of them. These test beds will be
(a) UNDAMAGED
(b) DAMAGE PATTERN 1
(c) DAMAGE PATTERN 2
(d) DAMAGE PATTERN 3
(e) DAMAGE PATTERN 4
(f) DAMAGE PATTERN 5
(g) DAMAGE PATTERN 6
1 20 5 10 15
46
referenced throughout this dissertation as various detection schemes are introduced
and validated.
TABLE 2.8
SUMMARY OF TESTBEDS TO BE USED IN THIS RESEARCH
Test Bed Type Purpose
ASCE Benchmark Simulation Verify online damage detection using acceleration only
Thin Beam Simulation Verify online damage detection using acceleration and strain; Verify offline
damage localization
LANL 8DOF System Experimental Verify online damage detection using acceleration only
LANL Bookshelf System Experimental Verify online damage detection using acceleration only
Thin Beam Experimental Verify online damage detection using acceleration and strain
Bridge model Experimental Verify online damage detection using acceleration and strain
47
CHAPTER 3:
OVERVIEW OF WIRELESS SENSOR NETWORK CONCEPT
3.1 Challenges to WSNs
Recent developments in wireless sensor networks (introduced in Chapter 1) have
demonstrated their potential to provide continuous, unattended data to assess
structural health at low cost, while facilitating rapid deployment and enhanced flexibility.
However, since research on wireless sensor networks for structural health monitoring is
in its infancy, there remain challenges that need to be addressed. Most of these are tied
to enhancing capacity by providing a high duty cycle, capability for high frequency, high-
fidelity sampling, and reliable collection/local storage/processing and transmission of
large amounts of data (Kim 2005), specifically:
Resolution: Given the low amplitude of ambient vibrations and the minor levels of damage to be detected, the platform must have ample sensitivity with minimal noise.
Synchronization: Signals must be accurately time stamped and sampled across the network, despite differences in drift of each node’s clock, to enable correlated analyses.
Tiered Communications: In the case of long span structures or in deployments with dense obstructions, it is impossible to cover the entire structure with a single tier of communication. Thus, multi-tier networks are necessary to provide requisite connectivity and subsequently to organize the analyses within the network.
Reliability: Data transfer needs to be reliable, minimizing (data) packet loss during wireless transmission.
48
Efficiency: On-board power must be conserved and ideally autonomously replenished to minimize maintenance efforts, particularly since the sensors may be embedded in inaccessible locations.
Robustness: The system must be able to monitor the responses caused by various inputs (wind, earthquake, traffic, etc.) and different environmental conditions (temperature, humidity, etc.).
This research focuses on means to address these challenges in both the
conceptualization of the network and the embedded algorithms. The discussion in this
chapter will first focus on the former, while the remainder of this dissertation will focus
on the latter.
3.2 Proposed WSN Monitoring Processes
In the author’s opinion, a comprehensive monitoring paradigm would begin with
the traditional aspects of structural health monitoring: a low-demand assessment of the
ongoing, in-service performance of structures using a variety of measurement
techniques (Aktan, et al. 1997) and, at frequent if not continuous intervals over the life
of a structure, reliable and timely condition assessment strategies for damage detection,
localization and quantization. As condition assessment is a more computationally (and
power) intensive operation, its use must be appropriately cycled. Thus the proposed
system should enable a baseline assessment followed by a period of real-time health
monitoring and an alarm-triggered condition assessment, as demonstrated in Figure 3.1.
49
Figure 3.1: One cycle of structural health monitoring and condition assessment.
It should be distinguished that the assessment aspect of Figure 3.1 should be
capable of rapidly delivering information to the end user, though not necessarily strictly
in real-time, while the monitoring aspect needs to be conducted in a real-time fashion.
This trade off is necessary since the condition assessment process again takes
considerably more memory and computational resources, thus being more practical to
execute offline. This division of labor furthermore reduces the demands on
computational and communications resources within the network, which will be
charged primarily with health monitoring. The discussion must now turn to how this
network is to be organized, activated and operated to address the challenges listed in
Section 3.1. First one should emphasize that many past research efforts focused only on
wireless hardware, wireless architectures or damage detection algorithms in isolation.
Herein, by forging appropriate collaborations, an integrated development of hardware,
wireless networking and damage detection is proposed to provide a truly end-to-end
Condition Assessment
Alert
Time
Co
mp
uta
tio
nal
Inte
nsi
ty
Continuous Monitoring
Baseline Assessment
High
Low
50
treatment of the problem at hand. The first aspect that will be addressed is the network
activation.
3.3 Restricted Input Network Activation Scheme (RINAS)
Ambient vibration testing or operational monitoring is generally preferred over
forced vibration testing as it is more economical and less obtrusive. A structure can be
adequately excited by wind, micro-tremors, traffic or human activities, and the resulting
motions can be readily measured with highly sensitive instruments. More importantly,
this does not require closure of the structure or risk the possibility of damaging the
structure through any form of actuation or controlled loading. Consequently, the overall
cost and effort of testing on a large structure ambiently is reduced (Gentile and Gallino
2008). Many applications of ambient vibration testing in the literature have shown that
it is an effective technique to determine the fundamental frequencies and mode shapes
of full-scale structures (Littler 1995), to find the changes in structural properties
(Mendoza, et al. 1991), and to contribute further to the development of structural
identification and health monitoring methods for bridges (Feng, et al. 2005); (He, et al.
2006); (Grimmelsman, et al. 2007). However, the low signal to noise ratio, the difficulty
in exciting higher modes, and the lack of measured input significantly complicates the
ensuing system identification (Kijewski-Correa and Cycon 2007). Particularly in the
instance of bridge monitoring, dynamic responses due to wind and traffic must be
simultaneously addressed, in addition to quasi-static response variations associated
with seasonal and daily environmental variations.
51
The detection of damage implicitly requires diagnostic algorithms to be able to
discern damage from the typical variations that are experienced as environmental and
loading conditions change. This discernment is generally made by comparing newly
acquired data to massive databases that catalog the bridge responses over a range of
environmental and loading conditions in its healthy or initial state. Generating, storing,
managing and manipulating such databases can be computationally intensive,
particularly in real time. Thus, while the input cannot be controlled or explicitly
measured in ambient vibration testing, this research seeks to instead improve the
performance of system identification and reduce the size of reference databases
through the introduction of a Restricted Input Network Activation Scheme (RINAS).
Through RINAS, the system will be triggered only by the detection of particular user-
defined traffic and environmental conditions; quantification of the traffic conditions can
be accomplished in a variety of ways, one of which is discussed later in Chapter 5.
Additional environmental information can be collected by a meteorological station to
further define the present condition (temperature, humidity, etc.). If the traffic and
environmental conditions, which are assessed at the network’s gateway node (shown in
Figure 3.2), are consistent with a user-specified scenario, the distributed wireless nodes
with their response sensors are activated and data is acquired as the vehicle(s) pass over
the bridge. The particular loading condition that will be sought in this research is the
passage of an isolated semi-tractor at night, thereby strictly specifying the dynamic
loading condition and removing the role of thermal expansions. The regulation of input
conditions in this way implies that the reference pool need only include data on the
52
response of the bridge in its healthy or initial condition under this loading scenario.
Again, once the system is trained for this loading condition, sensors will only be
activated when this condition is present for subsequent evaluations over the monitoring
cycle. In addition to the computational savings of this event triggering approach, this
network design also reduces the power demands on the system and extends network
lifetime, as sensors only operate under these specified conditions. Given the
intermittent monitoring using RINAS, Figure 3.1 is now modified as shown in Figure 3.3.
Figure 3.2: Two stage process of input selection and restricted activation for assessment. = gateway sensor, =
meteorological station, = wireless nodes, = response sensors.
53
Figure 3.3: Structural health monitoring and condition assessment period with RINAS (event triggered).
3.4 Multi-Scale Network Architecture
The size and the complexity of modern civil structures require a significant
number of sensors to perform vibration-based condition assessment. Creating scalable
networks that can effectively interact and be synchronized is challenging. In particular,
data transmission and power management can be difficult when a traditional hub and
spoke architecture (Fig. 3.4a) is employed, even in a wireless format (Fig. 3.4b), where
all the sensor units communicate directly with a central monitoring station called herein
a macro or M-node. Given the size of civil engineering structures, this architecture
implies that data will be transmitted over long distances, which is a major drain of
battery power and increases the risk of data loss and noise contamination. Since the
power demands of local computation are less than those associated with data
transmission, a more efficient wireless network utilizes the decentralized data
processing capability of the local resources within the network and then transmits only
Po
wer
Co
nsu
mp
tio
n
High
Low
Alert
Condition Assessment
Time
Baseline Assessment
Intermittent Monitoring
54
key parameters wirelessly up through the network and ultimately to the M-node for
transmission to the end user. The multi-scale design proposed herein adopts this
strategy, where the M-node serves the function of the aforementioned gateway sensor
to trigger the network under RINAS.
Figure 3.4: (a) Traditional wired hub and spoke architecture, (b) wireless hub and spoke architecture, (c) proposed multi-scale
wireless network.
55
In this research the traditional hub and spoke network is recast in a multi-scale
framework to satisfy key performance metrics such as maximizing network lifetime,
enhancing reliability, and facilitating scalability. The multi-scale WSN introduced by this
research divides the structure into a series of meso-networks (m-nets), as shown in
Figure 3.4c. Within this m-net, there are wireless motes with on-board accelerometers
tethered to multiple distributed strain gauges to monitor behavior of the structure at
critical locations, including the underside of the deck, joints, transverse beams near the
supports, and braces/ribs. Each accelerometer and their supporting strain gauges form a
micro-network or µ-net), where the initial diagnosis of damage is conducted. This
constitutes a heterogeneous approach to damage detection where different response
quantities are aggregated in the assessment scheme discussed later in Chapter 4. This
decentralized approach not only has power conservation benefits, but also escapes the
need for strict synchronization and provides resistance to latency that a centralized
approach to system identification would require. Thus lengthy time series are never
transmitted wirelessly, and the only information shared outside of the µ-net is the
binary damage diagnosis and/or the estimated damage sensitive feature (DSF), which is
a customized metric for rating damage. Note that the sensors within each µ-net are
locally tethered to the mote to concentrate power and processing. This helps to reduce
hardware costs and power consumption that a fully wireless system would entail,
particularly since the cable tethers between the mote and strain gages are relatively
56
short. Certainly the provision for a fully wireless system does exist and can be pursued if
desired.
Unlike many networks that rely on sentinels for triggering the network, this
system remains dormant until the signal to collect data is initiated by the gateway node
(M-node). Thus this system is cycled to perform regular inspections when approaching
traffic and environmental conditions meet specified criteria, as described in the previous
section. Again since ambient vibration monitoring is being employed, the input to the
bridge is never explicitly known or controlled. This complicates ensuing system
identification, and as detailed in Chapter 4, and would normally require rather
expansive reference databases to be populated to benchmark the undamaged condition
of the bridge under a wide range of environmental and operational modes; however,
the use of RINAS does allow the operational and environmental states to be restricted
to a specific subset for which a reliable reference pool has been generated. This reduces
the size of the reference pool, thereby easing computational burden and memory
demands at the M-Node. Furthermore, this form of triggering helps to increase network
lifetime since sentinel functions are not required, and the single M-node can generally
have access to renewable power supplies to support its receipt of information on
structural condition and potential damage locations wirelessly from the m-nets, through
multi-hop wireless communications, as discussed in a later section. The M-node then
interfaces with the end user to report the findings (Fig. 3.4c). Chapter 4 will address the
data reduction and assessment that is conducted locally within each µ-net using only
the on-board computational resources of the wireless mote.
57
Again it should be emphasized that under this architecture, shown in Figure 3.5,
the goal of lower-tiers of the network is to gather and assess data as effectively as
possible, while the upper-tier is designed to verify and transmit information to the end
user as efficiently as possible. Thus the structure is monitored locally within the lower
tier; then the decision for the global structure is made within the higher tier, as now
described.
Figure 3.5: Diagram of two-tiered wireless sensor networks.
3.5 Data Fusion within the Network
Another advantage of a multi-scale network is the ability to perform network-
level processing to conserve battery power. Chapter 4 will deal at length with the local
data processing in the µ-net to detect damage, which is by far the most challenging part
of this problem. Still the reliability of the network can be enhanced by performing
additional processing at the higher layers of the network. This is needed because,
Two-tiered Network Bridge End-User
Lower Level Wired or wireless
link
1. Data acquisition 2. Data Procession 3. Local damage detection
1. Data
acquisition
2. Data
Procession
3. Local
damage detection
Upper Level Wireless link
1. Data Fusion 2. Global damage detection 3. Localization
1. Data
Fusion
2.
Global damage
detection
-net M-net M-node
58
unfortunately, as the sensitivity of a damage detection algorithm increases, so too does
the tendency toward false positives, as subsequent examples in Chapter 4 will
demonstrate. One of the major advantages of the multi-scale network concept herein is
the ability to fuse data locally to enhance detection capabilities and reduce this
probability of false positives. As discussed previously, a damage diagnosis is conducted
locally within each µ-net, that then transmits a binary report (0 = no damage, 1 =
damage) and/or the corresponding DSF to the head node of the local m-net, where they
can be fused through a number of approaches (Kijewski-Correa, et al. 2006 b). The most
basic would be a local voting process involving the two or more of the nearest neighbors,
with the m-net signaling damage only when indicated by majority of the µ-nets. This
damage signal from the m-net would then be wirelessly transmitted to the M-node for
additional offline localization analyses and dissemination to the end user, as again
demonstrated by Figure 3.5. An added level of weighting can be introduced by including
the damage sensitive feature values in the voting scheme. As examples in Chapter 4
demonstrate, DSFs do show spatial sensitivity and thus this form of local data fusion can
be helpful in eliminating false positives due to an errant assessment at one µ-net
(Kijewski-Correa, et al. 2006 b).
59
Figure 3.6: Overview of key features of proposed wireless sensor network for structural health monitoring with addition of new
benefits introduced in Chapter 3.
3.6 Summary
This chapter introduced the overall concept for the wireless sensor network,
including how the network would be activated under restricted input conditions (RINAS)
and how information is processed and relayed within a multi-scale network concept.
Figure 3.6 has been updated to indicate the new approach offered by this dissertation,
and the projected benefit of this approach. It should be reemphasized that the design
of the hardware or the wireless communications protocols for this concept are beyond
the scope of this dissertation. Instead, focus in the next chapter will shift to a new
approach to online damage detection within the µ-nets of the network introduced in
this Chapter, designed with the constraints of this network in mind and the requirement
of damage detection under ambient conditions.
BENEFIT APPROACH STAGE
DATA ACQUISITION
DATA REDUCTION
DETECTION
LOCALIZATION
Multi-scale WSN
Low power, scalable
60
CHAPTER 4:
ONLINE DAMAGE DETECTION
Historically, vibration-based damage detection focused fundamentally on the
dynamic properties of the structure by examining changes in modal frequencies,
changes in mode shape (curvature) vectors, and changes in the flexibility (stiffness)
matrix (Farrar, et al. 1997); (Doebling, et al. 1998). Unfortunately, minor levels of
damage generally cannot produce statistically significant changes in the natural
frequencies, particularly in the lower modes of vibration most readily excited by
ambient vibration. While other dynamic properties may have greater sensitivity to
minor damage, they are computationally challenging to execute within wireless sensor
networks and cannot be estimated in a decentralized fashion without strict
synchronization requirements, violating fundamental requirements of the proposed
network architecture introduced in Chapter 3. Instead, the author and his collaborators
have focused on damage detection using decentralized system identification based on
time series models of response measured at the -net level. While Kijewski-Correa et al.
(2006 b) explored various damage sensitive features based on such models, showing
their performance in the presence of noise, this study will focus on the damage
detection technique based on coefficients of these time series models, which are
61
related to the dynamic properties of the system, as illustrated by Figure 4.1, and will be
shown to be sufficiently sensitive to damage.
Figure 4.1: Relationship between time series coefficients and dynamic properties.
Just like dynamic properties, coefficients of time series models will change with
changes to structural system (Nair, et al. 2006); (Nair and Kiremidjian 2007); (Su and
Kijewski-Correa 2007). In the following, a basic AR (AutoRegressive) model will be used
to show the relationship between coefficients of time series models and the dynamic
parameters. A typical AR model can be expressed as:
p
i
i ninAnA1
)()()(~
(4.1)
)(~
nA is the output prediction at time step n. )( inA is the previous outputs at time
step n-i, i is the ith AR coefficient, and )(n is the residual error.
Changes of Structural System
Changes of Dynamic Properties:f, ,f
Changes of Time series model Coefficients:
21,
62
Discrete-time signals can be transformed to the complex domain and described
by the complex variable, z, through the use of the Z-transform so that the z-transform of
a function )(nA , denoted by )(z , is defined as follows:
k
kzkAznAZ )()()]([ (4.2)
The physical meaning of the inverse discrete-time complex variable, iz , is simply a
discrete time unit delay. The z-transform of )( inA is described as follows:
)()]([ zzinAZ i (4.3)
Applying the z-transform to both sides of Equation (4.1):
)()(1 2
2
1
1 zzzzz p
p
(4.4)
where )(z is the z-transform of the residual error )(n . Ignoring the residual error of
the AR time-series model, the transfer function of the dynamic linear system can be
written in the discrete-time complex domain, )(zH :
p
p zzzzH
2
2
1
11
1)( (4.5)
The roots of the polynomial equations of the transfer function denominator are
termed the poles of the dynamic system.
]0[][ 2
2
1
1
p
p
pole
p
pole
p
pole zzz (4.6)
Using the theory of polynomial roots, the first three coefficients can be expressed as:
63
i
ipolez ,1 ( pi ,2,1 ) (4.7)
jpole
ji
ipole zz ,
,
,2 ( pi ,2,1 ; pj ,2,1 ; ji )
(4.8)
kpolejpole
ji
ipole zzz ,,
,
,3 ( pi ,2,1 ; pj ,2,1 ;
pk ,2,1 ; kji )
(4.9)
and the remaining coefficients would follow similarly.
Applying the Laplace transform
0
)()()( dttfetfsF st , demonstrates that
the poles of the transfer function directly yield the natural frequencies (n) and the
modal critical damping ratios ():
nnpole js 21 (4.10)
The bilateral z-transform is simply the two-sided Laplace transform of the ideal
sampled function, so if T is the sampling period, the relationship between z-transform
and Laplace transform is:
sTez (4.11)
and by substituting Equation (4.11) into Equation (4.10), the relationship between z-
transform poles and dynamic properties is revealed:
64
)1exp( 2 TjTz nnpole (4.12)
As it is widely established that damage to a structure will cause changes in the
system stiffness and damping, these changes can be quantitatively described by the
migration patterns of the transfer function poles. The relationship among AR
coefficients, z-transform poles, dynamic properties, and structural damages can be
illustrated by Figure 4.2.
Figure 4.2: The relationship among AR coefficients, z-transform poles, dynamic properties, and structural damage.
The advantages of a time series approach to damage detection can be defined as
follows:
Time series models provide a way to compactly and accurately represent signals.
The coefficients are easy to calculate, without the need for computationally intensive transforms.
Dynamic Properties
Time Series Coefficients
Transfer Function
Poles
Structural Damage
65
The coefficients contain multi-mode information, and as higher modes tend to show greater sensitivity to damage, this information can be conveyed through a single time series coefficient. As such, these coefficients will be subsequently shown to have enhanced sensitivity to damage.
Unfortunately, these coefficients will also vary depending on the loading and
environmental conditions, thus requiring a means to separate variations due to damage
in the system from those associated with changes in the forces (environmental and
loading) acting on the system.
Now, operating under the monitoring philosophy introduced in Chapter 3, the
objective of online structural health monitoring is to deliver a basic damage assessment
in real-time to the end user as soon as possible, while damage is in its early stages. Since
the damage is in its infancy, failure will not be immediate and time is available for
additional offline analysis to further determine the exact damage location and damage
extent. Thus the requirements for the online detection algorithm are: computational
simplicity, decentralized format, speed and accuracy. To achieve those requirements,
the Bivariate Regressive Adaptive INdex (BRAIN), built on statistical signal processing
techniques in the time domain, is proposed in this dissertation. The sections will
illustrate that BRAIN can satisfy the requirements for online detection, while minimizing
the energy consumed and the impact of variabilities in loading and environmental
conditions.
BRAIN belongs to a class of detection schemes using various forms of regressive
models to reconstruct acceleration responses and extract damage sensitive features
(DSFs). In this class of methods, DSFs can be defined using the model coefficients
66
themselves or the model residuals. The efforts to date have demonstrated the potential
of such regressive approaches to provide effective damage diagnosis and the capability
for embedment within wireless sensor networks (Lynch, et al. 2001; Lynch, et al. 2002);
(Kijewski-Correa, et al. 2006 a; Kijewski-Correa, et al. 2006 b). Unfortunately, the model
orders required for such autoregressive representations can strain the limited
computational capability of the local processors. Due to similar constraints, the DSFs
employed must also be simplified in nature, generally implying that they are specifically
tailored for the underlying time series model and/or application at hand, limiting their
robustness and ability to be extended to other applications. As a result, they are termed
“static” DSFs. BRAIN circumvents this limitation through a dynamic DSF adapting to the
most sensitive model coefficients in a given damage scenario, which vary with location,
loading condition and damage severity. The performance of this dynamic DSF in
comparison with comparable “static” DSFs will be presented in this chapter, using
several test beds.
Subsequently, it will be shown that this data-driven feature is capable of not only
incorporating a wide variety of potential regressive models, but also a diversity of sensor
data through bivariate autoregressive (BAR) models. This provision for heterogeneous
sensing, when coupled with the data-driven DSF concept, yields the formal BRAIN
concept and provides a dramatically enhanced detection capability, as demonstrated
later using the thin beam test bed with minor levels of damage.
The BRAIN framework will now be presented, but before doing so, it is important
to restate a number of terminologies that will be used in this chapter. Damage Sensitive
67
Features (DSFs) defined shortly will be termed “static” when they are defined a priori
based on specific coefficients of a specific time series model or application. They will
denoted by an “s.” These will be contrasted with “dynamic” DSFs that adaptively select
the time series coefficients used as their basis, denoted with a “d”. The underlying time
series models can be “homogeneous,” implying they use only one type of sensor
(usually accelerometers) and will be denoted with a “1”, in contrast to “heterogeneous”
models where multiple sensors are employed (in this dissertation local acceleration and
surface strain measurements), which will be denoted with a “2”. Thus BRAIN is
representative of a dynamic, heterogeneous approach, in contrast with methods in the
literature that are static, homogeneous approaches.
4.1 Time Series Models
The performance of time-series damage detection schemes is entirely reliant on
the underlying model used to represent the time series. Due to the limited
computational capability of the local processors, it is useful to pose two questions: 1)
can lower order models be used with heterogeneous detection? And 2) can a simple yet
adaptive DSF be developed that can accommodate various underlying models and even
significant changes in the application, while still providing reliable detection?
These questions will be pursued by first evaluating several regressive-type
models, including the formats commonly used in homogeneous sensing and those newly
proposed for heterogeneous detection to demonstrate their performance. Before doing
so, it is important to note that in general, measured signals are standardized before a
68
model is fit by demeaning and then normalizing by their standard deviation. It is
assumed this action is performed on all measured signals before they are fit by any of
the models proposed herein.
4.1.1 Homogeneous Representations
This section will introduce two time series representations that have been used
to detect damage in Civil Structures based solely on acceleration data (A). Both draw
their basis from the autoregressive model (AR) formulation, which is one of a group of
linear prediction formulas that attempt to predict an output )(~
nA of a system based on
the previous outputs )(~
inA and residual error )(n :
na
i
i ninAnA1
)()()(~
(4.13)
Similarly, an autogressive moving averages (ARMA) formulation can be used (Ljung
1999):
na
i
nb
j
ji njninAnA1 0
)()()()(~
(4.14)
where j is the jth MA coefficient. This approach has been used previously to model time
series for damage detection in Civil Structures (Nair et al., 2006).
The notation ARX (AutoRegressive with eXogenous inputs) refers to the model
with na autoregressive terms and nb exogenous input terms. This model contains the AR
model and a linear combination of the last nb terms of a known and external time series.
This has been used as part of a two-stage AR-ARX approach by Sohn et al. (2001), which
69
uses the residual error of an AR representation (,AR) as the exogenous input to a second
stage na+nb order ARX model:
na
i
nb
j
ARji njninAnA1 0
)()()()(~
(4.15)
where i and j are the ith and jth coefficients in the expansion and is the residual error,
whose statistics are utilized for detection of damage.
4.1.2 Heterogeneous Representations
Kijewski-Correa, et al. (2006a) later introduced an alternate formulation based
upon multiple vibration signals from different sensing elements, e.g., acceleration A and
strain S, realizing the unique information regarding damage that can be carried by each.
Various formulations that model the interrelation between these two measured
quantities have been offered to enhance damage detection (Law, et al. 2005), but prove
too computationally demanding for WSN platforms. Thus, this dissertation proposes a
bivariate autoregressive (BAR) model between strain and acceleration. In this
representation, a standardized strain and acceleration data pair (A, S) is fit by a na+nb
order model:
na
i
nb
j
ji njnSinAnA1 0
)()()()(~
(4.16)
where i is the ith AR acceleration coefficient and j is the jth AR strain coefficient and is
the residual error. Though BAR had previously been used in medical and financial
modeling as a means to use two-interrelated signals for a single signal model (Dacco and
70
Satchell 2001), (Soares and Cunha 2000), the concept is new to the field of structural
health monitoring.
4.1.3 Performance Assessment
Before assessing the ability to detect damage, it is first useful to note the
accuracy with which these various representations model a given signal. Figure 4.3
shows an example of an acceleration signal from a randomly excited undamaged thin
cantilever beam introduced in Chapter 2 that is fit with 20th order models: AR (na=20),
ARMA (na=13; nb=7) and ARX (na=13; nb=7) and this same acceleration signal and its
corresponding strain signal are also fit by a BAR model (na=13; nb=7). The
reconstructions by each model are also provided in Figure 4.3, along with the residual
errors and an inset summary of their statistics. The results in Figure 4.3 underscore the
superior performance of the BAR representation for same effective model order.
4.1.4 Computational Burden
While one of the primary merits of Sohn et al.’s (2001) AR-ARX approach is this
resistance to changes in the environmental and operational conditions of the system, as
noted by Lynch et al. (2006b), local computational/memory capabilities within WSNs are
often insufficient to execute the two stages of autoregressive-fitting, the database
search required to find the appropriate reference state, the signal reconstruction and
residual error estimation. In this study, an alternate approach circumvents this challenge
by minimizing environmental and operational variabilities through RINAS and permitting
the use of a one-stage autoregressive model, without the need for signal reconstruction
71
and residual error calculation. Furthermore, Sohn et al.’s (2001) two-stage AR-ARX
approach utilized the statistics of the model residual errors as its DSF, while others like
Nair et al. (2006) have used an ARMA time series representation and retained the AR
coefficient’s themselves as direct damage indicators. Such coefficient-based DSFs are
attractive for use in WSNs in that only the AR coefficients themselves need to be
retained and analyzed for detection, reducing the computational/memory burdens
associated with signal reconstruction and error estimation. This class of DSF will now be
explored in more detail.
72
Figure 4.3: Representation of (a) acceleration signal by (b) AR-ARX, (c) AR, (d) ARMA, and (e) BAR, with residual errors by (f) AR-
ARX, (g) AR, (h) ARMA, and (i) BAR.
4.1.5 Computational Demands
As discussed in Kijewski-Correa et al. (2006a), various prototypes in the literature
have varied computational capabilities (based on processor selection) and internal RAM
resources. Since the assembly of hardware is beyond the scope of this dissertation, this
section now addresses the larger concern of whether sufficient computational resources
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Simulated Signal(AR)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Residual error(AR)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Simulated Signal(ARMA)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Residual error(ARMA)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Simulated Signal(AR-ARX)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Residual error(AR-ARX)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Original Signal
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Simulated Signal(BAR)
1 1.02 1.04 1.06 1.08 1.1 1.12 1.14 1.16 1.18 1.2-5
0
5
Time(s)
Accele
ration(m
/s2)
Residual error(BAR)
Standard Deviation of Residual Error
AR-ARX AR ARMA BAR
0.4325 0.4077 0.3817 0.3319
(
a)
(
b)
(
c)
(
d)
(
f)
(
g)
(
h)
(
e)
(
i)
(a)
(b)
(c
(d)
(e)
(f)
(g)
(h)
(i)
73
are available on practical wireless platform designs. Though larger processors can be
used, they consume more power. In actuality, the critical issue generally reduces to
available internal RAM, as writing to the flash memory is costly, in terms of power;
therefore it becomes important to limit model orders to eliminate writing to the flash
memory. The most critical aspect of the implementation of the algorithms on a wireless
platform is definitely the matrix inversion operation necessary for the solution of the
Yule-Walker equations. The matrix to be inverted contains estimates of the discrete
time history autocorrelation function and thus is a Toeplitz matrix, which permits
relatively efficient inversion. The number of divisions and multiplications is proportional
to the model order cubed. The algorithm to estimate an 8th order bivariate regressive fit
to strain and acceleration data, as described by Equation (4.16), has been successfully
implemented on a wireless prototype, as discussed in Kijewski-Correa et al. (2006a),
using only 4 KB of internal RAM. It was observed that this was sufficient for not only this
order, but also model orders up to 15, as demonstrated in Table 4.1. Also note that this
was assuming the use of only 4 KB of RAM and the platforms described later in Chapter
7 will have additional memory available; therefore the proposed algorithms can be
easily embedded for decentralized signal processing.
74
TABLE 4.1
ESTIMATED AR COEFFICIENTS DEMONSTRATING THE PERFORMANCE OF EMBEDDED
ALGORITHM ON WIRELESS PLATFORM WITH 4 KB OF RAM
8th Order Model 15th Order Model Wireless Desktop Error Wireless Desktop Error
0.54432 0.56003 0.03 -6.2972 -6.4773 0.03 -4.2282 -4.24463 0 2.5644 2.7011 0.05 2.65337 2.55412 0.04 5.9705 6.0271 0.01 1.40866 1.55892 0.1 -3.4564 -3.5836 0.04 2.36874 2.30359 0.03 -2.2691 -2.2268 0.02 3.3254 3.39063 0.02 7.6161 7.7071 0.01
-3.72192 -3.7387 0 -3.6788 -3.7725 0.02 -1.85879 -1.89296 0.02 -2.6436 -2.6776 0.01
7.3872 7.5693 0.02 -2.8217 -2.9575 0.05 -2.9586 -2.9679 0.0 6.2588 6.4115 0.02 1.7927 1.7287 0.04 -6.6921 -6.7695 0.01 -0.3276 -0.3229 0.01
Source: (Kijewski-Correa, et al. 2006 a)
4.2 Online Damage Detection
This section will now demonstrate that the proposed autoregressive models,
when coupled with an adaptive DSF, can detect minor levels of damage using relatively
modest model orders to operate within the computational constraints of the wireless
platform.
4.2.1 Damage Sensitive Features for Homogeneous Representations
Through extensive investigation, Nair et al. (2006) found that the first AR
coefficient of an ARMA representation was the most sensitive to damage within their
75
homogeneous detection framework. However, as the application at hand changes, as
alternate underlying models, e.g., AR, are considered to further reduce computational
burdens in WSNs, or as heterogeneous detection frameworks using BAR models are
explored, the first AR coefficient may not be the most sensitive to damage. This type of
static DSF with a priori coefficient selection is replaced in this dissertation with a new
DSF that is more adaptive to changes in the AR coefficients. The premise for this DSF
further differs from past formulations in the literature in that it directly incorporates
information from the reference pool of undamaged states. Such reference pools are
used by all damage detection methods in this class and are populated by acquiring
multiple vibration signals under varying operational and environmental conditions from
the structure in its undamaged or initial state. Each of these reference signals should be
standardized and then fit by the desired model (AR, ARMA, BAR, etc.), with the model
coefficients stored in the reference database. Knowing such information is readily
available, this adaptive or dynamic DSF is then defined as the AR coefficient that has
changed most significantly when compared to the average values stored in the
reference database:
nai
kiref
kiref
ki
kstd
avg
dDSF
:1
][
][
max1
(4.17)
Here the notation ref refers to the mean (avg) and standard deviation (std) of the AR
coefficients in the reference database. Again the notation DSF1d implies that this is a
76
homogeneous, dynamic representation. The k denotes that this is the DSF defined at the
kth location on the structure. Two key features should be noted:
The original AR coefficients for each acceleration signal in the reference database need not be stored locally; only the mean and standard deviation of each coefficient are required. Thus only 2na reference values are finally stored at each sensor node after some training period for the WSN. Again keep in mind that na is relatively small (<20). This dramatically reduces not only the required on-board memory, but also any computation (and power drain) associated with the manipulation of a reference database.
The DSF is unaffected by the choice of underlying model (AR, ARMA, BAR, etc.), unlike other “static” DSFs that are tied to or have been validated with only a specific model type or application in mind. This also implies that if there is a location where higher order coefficients are more sensitive to damage, they will be exploited. Thus the DSF is data-driven and again involves minimal computational effort.
In most practical detection and health monitoring problems, the signals of
interest exhibit some variability not due to damage, but rather due to changes in the
environmental and operational conditions under which they are procured. Even with
the inclusion of selective triggering by RINAS, damage detection must be couched in a
probabilistic context capable of distinguishing these benign variabilities from more
serious indicators of damage. Thus, statistical significance must be established and can
be done so during the training period by evaluating the DSF in Equation (4.17) against
the values obtained when using signals from the reference pool. If the DSF in question
deviates significantly from the DSFs of the reference pool, damage is suspected. As
demonstrated by Figure 4.4, a Gaussian model can generally be applied as a
conservative representation of the positive tail region of the DSF values associated with
the reference pool, allowing the user to specify a desired percentile of statistical
77
significance, e.g., 95%. Damage is indicated with that percent certainty whenever a
future DSF value exceeds this threshold. Again this threshold would be established
during the network’s trailing period and stored in the -net for evaluation of all DSFs
during future condition assessments. For the discussions which immediately follow, the
damage pool will be comprised of 100 independent random simulations of the
undamaged thin beam model described in Chapter 2. Only acceleration data will be
considered in this section.
Figure 4.4: Normal distribution test on homogeneous dynamic DSF for reference pool of undamaged acceleration data.
78
To verify the performance of the data-driven or dynamic homogeneous DSF in
Equation (4.17), it is now compared to a “static” DSF also based on AR coefficients (Nair
et al., 2006):
2
3
2
2
2
1
11
kkk
k
ksDSF
(4.18)
Again the notation DSF1s implies this is a homogeneous, static DSF utilizing the
first three AR coefficients at the kth location on the structure. It should be noted that an
ARMA model was actually used by Nair, et al. (2006); however, for consistency, the
same AR model is used to represent the acceleration data and only the DSFs applied are
varied in the examples that follow. The validations in the subsequent section will be
concerned with the hypothesis: data-driven or dynamic DSFs are more robust and
reliable than their static counterparts in homogeneous sensor networks.
4.2.1.1 Validation Using Simulated Thin Beam Model
Various degrees of damage (percent area losses of 0, 10, 30%) are explored for
cuts introduced at LD = 18.75 cm, midway between points A and B (see Figure 2.5). Each
damage scenario is run 10 times to explore the repeatability of the results. A 97.5% one-
sided confidence interval is specified for distinguishing statistically significant damage in
Equation (4.17), while a two-sided confidence interval at 97.5% is used with Equation
(4.18). The results for both the static and dynamic DSFs are provided in Table 4.2, where
bold-faced values indicate that DSF falls outside the thresholds for statistical significance
implying that damage is detected. The percentage of cross sectional area lost due to
79
damage is specified for each case to demonstrate the minor level of damage being
considered and performance will be quantified by the damage detection rate (Det Rate).
Note the 0% damage case is provided to evaluate any tendency toward false positive
detection.
Several important conclusions can be drawn from the test results Table 4.2:
Neither DSF appears susceptible to false positives.
For the two cases of actual damage, the static DSF (Eq. 4.18) is generally unsuccessful. Only a 20% detection rate is recorded in one instance, at Point C and only for the larger of the two damage cases.
The dynamic DSF (Eq. 4.17) has no success detecting damage at point A, which is not surprising since this point is closest to the fixed end, thus producing the smallest of the simulated acceleration responses.
For the smallest level of damage, the dynamic DSF (Eq. 4.17) has its highest detection rate at point B, which is logical since damage is near this node. Beyond point B, the smaller of the two damage scenarios is scarcely detected.
For the larger of the two damage cases, the dynamic DSF (Eq. 4.17) has reasonable success at points B (60%), C (100%) and D (50%). The good performance at location C is due to it being ideally situated at a location where acceleration responses are larger, without being too far from the damage location.
The results also indicate that the dynamic DSF (Eq. 4.17) values and the detection rate both increase with damage level, so the method will not only be more reliable as damage increases beyond these modest levels, but the correlation of the DSF to damage level provides a means to quantify the extent of damage.
These findings clearly demonstrate the advantages of a dynamic DSF, even when
only acceleration responses are considered.
80
TABLE 4.2
DAMAGE DETECTION RESULTS FOR STATIC (EQ. 4.18) AND DYNAMIC (EQ. 4.17) DSF FOR SIMULATED THIN BEAM
POINT A POINT B
Static Dynamic Static Dynamic
Threshold (-0.49, -0.3) 1.54 (-1.17, -0.42) 1.49
Area Lost 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50%
Test 1 -0.42 -0.41 -0.39 1.01 1.04 1.06 -0.94 -0.97 -0.65 0.99 1.04 1.38 Test 2 -0.41 -0.4 -0.4 0.65 0.63 0.92 -0.91 -0.93 -0.86 1.48 1.9 2.01 Test 3 -0.34 -0.34 -0.34 0.73 0.72 0.7 -0.56 -0.47 -0.19 1.16 1.86 2.2 Test 4 -0.35 -0.35 -0.36 0.71 0.78 1.03 -0.99 -0.99 -0.89 0.8 1.17 1.48
Test 5 -0.37 -0.37 -0.38 0.57 0.62 0.76 -0.88 -0.84 -0.81 0.73 1.03 1.45
Test 6 -0.41 -0.4 -0.39 1.01 1.06 1.15 -0.69 -0.6 0.06 1.06 1.66 2.11 Test 7 -0.39 -0.39 -0.36 0.72 0.83 0.68 -0.91 -0.94 -0.78 0.49 0.91 1.4 Test 8 -0.36 -0.36 -0.37 1.03 1.05 0.98 -0.92 -0.87 -0.7 0.93 1.51 1.62 Test 9 -0.37 -0.39 -0.4 0.92 0.76 0.6 -0.68 -0.63 -0.77 1.13 1.74 1.87
Test 10 -0.43 -0.44 -0.45 0.89 0.71 0.73 -0.93 -0.67 -0.48 1.4 1.88 1.96
Det Rate 0% 0% 0% 0% 0% 0% 0% 0% 20% 0% 60% 60%
81
TABLE 4.2 (CONTINUED)
DAMAGE DETECTION RESULTS FOR STATIC (EQ. 4.18) AND DYNAMIC (EQ. 4.17) DSF FOR SIMULATED THIN BEAM
POINT C POINT D
Static Dynamic Static Dynamic
Threshold (-0.16, 1.43) 1.51 (-1.73, 1.37) 1.48
Area Lost 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50%
Test 1 0.05 -0.08 -0.31 0.96 1.3 1.83 -0.98 -0.96 -0.91 0.92 1.15 1.38 Test 2 0.92 0.92 0.68 1.26 1.37 2.24 -0.98 0.63 -0.67 1.21 1.29 2.12 Test 3 0.91 0.91 0.72 1.1 1.19 2.42 0.05 0.15 -0.41 0.93 0.87 1.61 Test 4 0.79 0.84 0.94 0.58 0.97 1.9 -0.99 -0.98 0.98 0.72 1.06 1.26 Test 5 0.84 0.87 0.83 0.73 1 1.74 0.35 0.22 -0.5 0.73 1.05 1.69 Test 6 0.82 0.87 0.88 0.95 1.79 2.44 0.23 -0.01 -0.51 0.82 1.28 1.92 Test 7 0.85 0.6 -0.03 0.62 1.43 2.27 -0.95 -0.94 -0.9 0.92 1.17 1.56 Test 8 0.96 0.97 0.84 1.08 1.25 1.98 -0.96 -0.95 -0.87 0.67 0.8 1.33
Test 9 0.85 0.8 0.54 0.99 1.3 1.66 0.97 0.97 0.97 1.02 0.79 0.98
Test 10 0.31 0.13 -0.27 1.28 1.87 3 -0.98 -0.98 -0.9 1.23 1.27 1.47
Det Rate 0% 0% 20% 0% 20% 100% 0% 0% 0% 0% 0% 50%
82
4.2.1.2 Validation Using Vibrating Disk Assembly
The Los Alamos National Laboratory 8DOF assembly will be used for the next
validation, simulating damage by changing the spring between masses 5 and 6 to one
having a 14% smaller spring constant. The system is excited in both its damaged and
undamaged states at mass 1 using an electro-dynamic shaker with different input
voltage levels (3V and 5V). The acceleration responses of all the masses were recorded
for repeated independent trials.
A 97.5% one-sided confidence interval is specified for distinguishing statistically
significant damage using Eq. (4.17), while a two-sided confidence interval at 97.5% is
used with Eq. (4.18). This is based on an undamaged reference pool consisting of 8
independent trials for the undamaged system. Note that the size of the undamaged
pool is limited by the amount of experimental data archived by LANL for public use. Each
time a DSF value falls outside of this confidence interval, damage is detected and is
signified in the tables that follow in bold. Four independent experimental trials of the
damaged state are available publicly from LANL and the damage detection rate (Det.
Rate) over these four trials is summarized in Tables 4.3 and 4.4 for the two different
input excitations. The following major observations can be drawn from these results:
The dynamic DSF (Eq. 4.17) is successful in detecting damage, with a perfect detection rate (100%) at all locations, while the static DSF (Eq. 4.18) is less successful, with an average detection rate of 53%.
Both DSFs showed a certain capability to locate damage. In fact, for the dynamic DSF, its values are largest near the point of damage: masses 4, 5, and 6 are the only 3 locations with values exceeding 10. This finding also supports the value of local data fusion in an m-net to avoid false positives
83
by using spatial affirmation of a report from a given node, as will be discussed at the close of this chapter.
For 3V input voltage level, the performance of static DSF is related to the strength of the signal. Detect rate is low at the fixed Mass 1 where the signal is the weakest and the detection rate is perfect at the free end. At the same time, detection rates of the intermediate locations are almost proportional to the distances to the fixed end. However, though the fixed end and free end still respectfully display the lowest and highest detection rates, detection rates at intermediate locations do not vary linearly for the 5 V case. The detection rates are either close to 100% or near to 0%, depending on whether the coefficients chosen a priori by the static DSF tend to be affected by the damage. This demonstrates how even variations in the excitation can influence the sensitivity of the coefficients and further advocates for the flexibility afforded by a data-driven DSF, whose detection is unaffected by the input variations.
84
TABLE 4.3
DAMAGE DETECTION RESULTS FOR STATIC (EQ. 4.18) AND DYNAMIC (EQ. 4.17) DSF FOR
8DOF SYSTEM UNDER 3V INPUT VOLTAGE LEVEL
Mass 1 Mass 2 Mass 3 Mass 4
Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Threshold (-0.245, 0.025) 1.544
(-0.454, 0.490) 1.589
(0.484, 0.540) 1.642
(0.522, 0.573) 1.503
Test 1 -0.214 5.983 0.482 3.881 0.516 3.246 0.590 13.743
Test 2 -0.215 6.371 0.473 5.384 0.473 5.076 0.575 15.837
Test 3 -0.193 8.278 0.484 3.835 0.506 4.832 0.53 6.548
Test 4 -0.236 8.039 0.486 4.926 0.510 1.839 0.546 7.643
Det Rate 0% 100% 0% 100% 25% 100% 50% 100%
Mass 5 Mass 6 Mass 7 Mass 8
Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Threshold (0.540, 0.585) 1.539
(0.545, 0.588) 1.571
(0.547, 0.619) 1.583
(0.647, 0.692) 1.544
Test 1 0.531 19.042 0.624 4.668 0.540 7.394 0.604 6.240
Test 2 0.573 13.179 0.628 4.086 0.548 5.161 0.643 5.135
Test 3 0.617 6.908 0.688 12.482 0.516 9.015 0.773 8.612
Test 4 0.631 9.890 0.695 14.155 0.539 8.781 0.771 8.745
Det Rate 75% 100% 100% 100% 75% 100% 100% 100%
85
TABLE 4.4
DAMAGE DETECTION RESULTS FOR STATIC (EQ. 4.18) AND DYNAMIC (EQ. 4.17) DSF FOR
8DOF SYSTEM UNDER 5V INPUT VOLTAGE LEVEL
Mass 1 Mass 2 Mass 3 Mass 4
Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Threshold (-0.241, 0.023)
1.561 (0.453, 0.489)
1.588 (0.483, 0.540)
1.642 (0.521, 0.572)
1.503
Test 1 -0.165 21.924 0.489 6.204 0.563 5.308 0.692 25.114
Test 2 -0.133 6.122 0.487 3.412 0.568 5.392 0.513 4.449
Test 3 -0.132 5.429 0.487 7.950 0.530 4.763 0.504 3.772
Test 4 -0.157 6.770 0.480 3.340 0.563 7.242 0.490 4.708
Det Rate 0% 100% 0% 100% 75% 100% 100% 100%
Mass 5 Mass 6 Mass 7 Mass 8
Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Threshold (0.540, 0.584)
1.5390 (0.525, 0.632)
1.6423 (0.547, 0.619)
1.583 (0.646, 0.691)
1.544
Test 1 0.569 8.780 0.585 9.274 0.501 6.607 0.624 4.387
Test 2 0.582 6.075 0.616 6.311 0.497 6.253 0.619 5.109
Test 3 0.547 12.241 0.577 9.979 0.499 6.750 0.646 3.595
Test 4 0.542 9.658 0.625 2.706 0.545 2.975 0.701 4.304
Det Rate 0% 100% 0% 100% 100% 100% 100% 100%
4.2.1.3 Validation Using LANL Bookshelf Structure
The Bookshelf Structure, another Los Alamos National laboratory assembly
introduced in Chapter 2, is used to further validate the merits of dynamic DSFs for
damage detection under four typical damage patterns. They are damage pattern 1 (the
preload torque of a bolt on the first floor is reduced by 93%), damage pattern 3 (a bolt
86
on the first floor is removed), damage pattern 4 (the preload torque of a bolt on the
second floor is reduced by 93%), and damage pattern 7 (the preload torques of a bolt on
the first floor and a bolt on the second floor at different corners are both reduced by
93%). The dataset consists of three undamaged time histories for each sensor location
and one time history for each damage pattern. Since this amount of data released to the
public by LANL is less than desirable, each time history is divided into 4 parts of equal
length to allow for a more robust reference pool and some evaluation of the
repeatability of damage detection.
A 97.5% one-sided confidence interval is specified for distinguishing statistically
significant damage using Eq. (4.17), while a two-sided confidence interval at 97.5% is
used with Eq. (4.18). This is based on an undamaged reference pool consisting of 12
undamaged time histories. Each time a DSF value falls outside of this confidence interval,
damage is detected and is signified in Table 4.5 in bold. The resulting damage detection
rate (Det. Rate) is also summarized in Table 4.5. Two major observations can be drawn
from these results:
The dynamic DSF (Eq. 4.17) is successful in detecting damage, with a perfect detection rate (100%) at all locations, while the performance static DSF (Eq. 4.18) is less successful. For sensors at the top floor, where the amplitudes of the response are largest, the average damage detection rate is 50%, but for sensors at lower floors the static DSF is incapable of detecting damage.
Both DSFs showed a certain capability to locate and even quantify the extent of damage. In fact, for the dynamic DSF, values at Floors A and B are larger than those of Floor C. Additionally the DSF values of the severe damage cases (3 and 7) are larger than those of the minor damage cases (1 and 4).
87
TABLE 4.5
DAMAGE DETECTION RESULTS FOR STATIC (EQ. 4.18) AND DYNAMIC (EQ. 4.17) DSF FOR
BOOKSHELF STRUCTURE
Floor A Damage Case 1 Damage Case 3 Damage Case 4 Damage Case 7 Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Thresh-old
(0.516,0.787)
1.777 (0.516, 0.787)
1.777 (0.516,0.787)
1.777 (0.516, 0.787)
1.777
Test 1 0.657 3.982 0.723 7.426 0.536 2.361 0.617 38.312 Test 2 0.741 2.737 0.611 26.587 0.609 1.845 0.744 16.118 Test 3 0.700 2.975 0.663 24.434 0.611 3.205 0.786 10.415 Test 4 0.708 1.788 0.688 19.767 0.502 3.589 0.772 14.102
Det Rate
0% 100% 0% 100% 25% 100% 0% 100%
Floor B Damage Case 1 Damage Case 3 Damage Case 4 Damage Case 7 Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Thresh-old
(0.480,0.782)
1.5659 (0.480,0.782)
1.5659 (0.480,0.782)
1.5659 (0.480,0.782)
1.5659
Test 1 0.670 4.572 0.709 6.512 0.725 1.780 0.568 9.183 Test 2 0.761 3.590 0.532 6.970 0.652 1.826 0.693 15.767 Test 3 0.760 3.234 0.616 6.826 0.552 3.471 0.743 14.590 Test 4 0.705 4.779 0.647 7.507 0.558 1.819 0.722 6.852
Det Rate
0% 100% 0% 100% 0% 100% 0% 100%
Floor C Damage Case 1 Damage Case 3 Damage Case 4 Damage Case 7 Static Dynamic Static Dynamic Static Dynamic Static Dynamic
Thresh-old
(0.596,0.677)
1.917 (0.596,0.677)
1.917 (0.596,0.677)
1.917 (0.596,0.677)
1.917
Test 1 0.644 7.456 0.627 7.839 0.664 1.989 0.862 6.443 Test 2 0.733 6.477 0.630 4.549 0.672 2.588 0.857 2.334
Test 3 0.696 5.695 0.680 3.379 0.672 2.892 0.818 5.921 Test 4 0.693 4.922 0.727 3.059 0.558 5.674 0.679 2.573
Det Rate
75% 100% 50% 100% 25% 100% 100% 100%
88
4.2.1.4 Validation Using Steel Truss Bridge Model
The merits of dynamic DSF are further established using the impact test results
for the model bridge experiment introduced in Chapter 2. The undamaged bridge was
impact tested 25 times at nodes 2, 4, and 5, for a total of 75 tests, so that the influence
of excitation proximity to the damage location and response level can be observed.
One-second acceleration time histories were acquired at all sensors. For each scenario,
only 20 of the 25 undamaged signals are stored in the reference database; the other 5
are used to test for false positives.
For each of these damage scenarios, the bridge was impacted at the same
locations, repeating the tests five times each. The normalized signals are fit by an 8th
order AR model in Equation (4.13). Then, static DSFs in Equation (4.18) and dynamic
DSFs in Equation (4.17) are calculated and evaluated against the reference database at a
97.5% level of statistical significance. Results that follow in Figure 4.5 are reported in
terms of detection rate. Considering all possible excitation and measurement scenarios,
over five repeated trials, a total of 150 DSFs for undamaged bridges were blind tested
using both the data-driven and static DSF. For each, only once was a false positive noted
– the detection of damage in a known undamaged structure. As a result, neither DSF
appears susceptible to false positives. The application of these DSFs to the data from
various damage scenarios, processed and presented in the exact same manner, reveals
that the data-driven DSF consistently performs at and often exceeds the detection rate
of the static DSF. Comparing the damage detection rates between data-driven DSFs and
static DSFs for various excitation and measurement combinations, detection rates
89
generally improve as the measurement point moves toward the mid-span. Keeping in
mind the present DSFs consider only acceleration responses, the larger amplitude of
these responses at the mid-span clearly contribute to the improved detection rates. The
exception is when the driving point is at the mid-span (point 5), which is a result of the
fact that even numbered modes cannot be excited, nor observed at this location,
dramatically reducing the amount of higher mode information available for damage
detection. The large amplitude responses and ability to observe and excite a full
spectrum of modes makes the results when exciting at point 4 most promising.
Above all, the performance of data-driven DSFs is influenced by four major
factors: modal participation, damage severity, damage proximity, and signal strength
(response amplitude), though the modal participation is the most important factor for
data-driven DSF in homogeneous detection. Again it should be emphasized that the
dynamic nature of this DSF allows it to adapt accordingly to maximize performance
despite these various influencing factors.
90
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
1 2 3 4 5
II III IV II III IVIII IV I II III IV IIm
pact
@ P
oint
2 I III II III IV I
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
Impa
ct @
Poi
nt 4
1 2 3 4 5
I II III IV I II III IV I II III IV I II III IV I II III IV
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
Impa
ct @
Poi
nt 5
1 2
I II III IV
3 4 5
IVI II IIIIIII IVII III IV I II IV I II III
Figure 4.5: Damage detection rate comparison between static DSF (Grey Bars) and dynamic DSF (Black Bars) based on only
acceleration responses of model bridge.
4.2.1.5 Validation Using Phase I IASC-ASCE Benchmark Problem
The Phase I IASC-ASCE Structural Health Monitoring Benchmark, previously
introduced in Chapter 2, is now used to explore the sensitivity of static and dynamic
DSFs to a range of damage severities. Each damage pattern (DP) listed in Table 4.6 is
independently simulated 10 times to explore the repeatability of the results. The
91
reference pool is comprised of 50 independent random simulations of the undamaged
structure. A 97.5% one-sided confidence interval is specified for distinguishing
statistically significant damage in Equation (4.17), while a two-sided confidence interval
at 97.5% is used with Equation (4.18). The damage detection rates at all floors are
provided in Figure 4.6 and are summarized in an averaged sense in Table 4.6 along with
a summary of the damage scenarios. Note that the DP0 damage case is provided to
evaluate any tendency toward false positives and should ideally have a 0% detection
rate.
TABLE 4.6
DAMAGE PATTERNS OF PHASE I IASC-ASCE BENCHMARK PROBLEM AND AVERAGE
DAMAGE DETECTION RATES
Pattern Description Damage
Level
Average Detection
Rate: Static DSF
Average Detection
Rate: Dynamic DSF
DP0
Undamaged None 0% 0%
DP1
All braces of 1st floor removed Severe 35% 100%
DP2 All braces of 1st and 3rd floor
removed Severe 100% 100%
DP3
One brace of 1st floor removed Moderate 17.5% 50%
DP4 One brace of 1st and 3rd floor
removed Moderate 17.5% 50%
DP5 Pattern 4 + floor beam partially
loosened Moderate 17.5% 50%
DP6 1/3 stiffness reduction, one
brace, 1st floor Minor 7.5% 12.5%
92
Figure 4.6: Damage detection results for IASC-ASCE benchmark building.
Several important conclusions can be drawn regarding overall damage detection
capability, i.e., ability to detect damage from any sensor output:
Neither DSF appears susceptible to false positives.
For the most severe level of damage (DP2), both the static and dynamic DSFs can detect damage consistently based on the response at any of the floors. For the other severe damage case (DP1), the static DSF detects damage only in the first floor consistently, closest to the point of damage, and has an average detection rate of 35%, while the dynamic DSF can again detect damage consistently at all 4 stories, for an average detection rate of 100%.
For the moderate and minor damage cases (DP3-6), the static DSF is not as successful: with detection rates as high as 40% at floor one, but as low as 0% at the top floor, for an average detection rate of 17.5% for modest damage levels (DP3-5) and 7.5% for minor damage levels (DP6). Detection capability is strongest at floors 1 and 3, where damage has been imposed. This indicates that when the damage severity is minor to modest, this static DSF is best suited for detection near the point of damage implying the need for high sensor density.
93
The dynamic DSF demonstrates the inverse capability for minor to modest damage levels. It can detect modest damage (DP3-5) with up to 80% repeatability at the fourth floor, though the capability progressively diminishes down the building. Minor damage (DP6) shows a similar trend, with 30% detection rate at the top floor, dropping to only 10% at the lower floors. This results in average detection rates of 50% under moderate damage (DP3-5) and 12.5% under minor damage (DP6). Since the acceleration response increases up the building, the findings here may indicate that the homogeneous dynamic DSF performs better as the response amplitude increases, consistent with the findings of Su and Kijewski-Correa (2007). This makes this class of DSF well-suited for applications where dense sensor networks are not feasible and measurements may only be taken at limited locations.
The dynamic DSF values have been shown to increase with the damage level, as shown by Table 2 in Su and Kijewski-Correa (2007), providing a means to directly quantify severity of damage.
While the capability to signify the presence of damage and even relative severity
is attractive, the ability to localize damage is also necessary. To assist in this, a damage
location index (DLI) is introduced:
T
un
T
un
T
un
DLI
2
(4.19)
where is the AR coefficients associated with the state being evaluated
na ,, 21 and un is the AR coefficients associated with the undamaged state
unnaun ,, 21 , for a given measurement location. For undamaged states, the
two vectors should be correlated and DLI should be unity. As damage levels
progressively increase, the correlation should reduce and DLI should tend toward zero.
Again, statistical significance can be verified by comparing the DLI to the confidence
interval derived from the undamaged reference pool. The DLI was applied to the IASC-
94
ASCE Benchmark Problem for DP1-2, with the results presented in Table 4.7. Note that
the localization of damage for both these damage patterns is successfully achieved.
TABLE 4.7
DAMAGE LOCALIZATION INDEX RESULTS FOR FIRST TWO DAMAGE PATTERNS OF IASC-
ASCE BENCHMARK PROBLEM
Floor 1 Floor 2 Floor 3 Floor 4
Reference Pool
Mean 0.992 0.989 0.988 0.974
Standard Deviation 0.007 0.012 0.011 0.019
Detection Threshold 0.978 0.965 0.968 0.937
DP 1 DLI 0.867 0.978 0.984 0.981
Damage Location? Yes No No No
DP 2 DLI 0.607 0.643 0.623 0.797
Damage Location? Yes Yes Yes Ys
4.2.2 Damage Sensitive Features for Heterogeneous Representations
Thus far, the utility of a data-driven or dynamic DSF has been demonstrated for
homogeneous representations (AR model of acceleration only) for a number of
experimental and simulated test beds. However, since it has been shown that the
combination of surface strain and acceleration data enhances damage detection in
comparison with the use of either strain or acceleration alone (Law, et al. 2005), the
dynamic DSF in Equation (4.17) is modified for a heterogeneous representation to
better exploit the most sensitive bivariate AR coefficients:
95
Nbj
kjref
kjref
kj
Nai
kiref
kiref
ki
kstd
avg
std
avg
dDSF
:0:1
][
][
,][
][
max2
(4.20)
where ref again indicates these statistics are calculated respectively over all the
acceleration () and strain () AR coefficients in the reference pool. The notation DSF2d
indicates again this is a heterogeneous, dynamic DSF calculated at the kth location on the
structure.
The validations in this section will then be concerned with the following
hypothesis: heterogeneous DSFs are more robust and reliable than their homogenous
counterparts.
4.2.2.1 Validation Using Simulated Thin Beam Model
To demonstrate the performance of the DSF in Equation (4.20), damage
detection results are compared between it and its homogeneous counterpart, Equation
(4.17), using the same simulated thin beam dataset from Section 4.2.1.1. The damage
detection results are shown in Table 4.8, where bold-faced values again indicate that
damage is detected. Results are compared to the homogeneous dynamic DSF, as
extracted from Table 4.2.
From the detailed results in Table 4.8 and the summary provided in Table 4.9,
several important conclusions can be drawn about the heterogeneous formulation:
96
Incidence of false positives for the heterogeneous approach (Equation (4.20)) is negligible in comparison with its detection rate.
The larger of the two damage scenarios can be identified reliably at all measurement locations for the heterogeneous approach. The reason for the stark contrast in performance between the heterogeneous and homogeneous approaches can be explained as follows: locations A and B, though being near the damage zone, are characterized by comparatively low acceleration responses; however, the surface strains at these points are comparatively higher than points C and D. Thus the vast improvement in detection capability in the vicinity of damage can be largely credited to the heterogeneous framework that recognizes the fact that structural response cannot be characterized by acceleration alone and a DSF that adapts to the response component most critical at that location.
Consistent with the homogeneous scheme, the heterogeneous DSF’s (Equation (4.20)) detection rate falls off further from the damage location for the smaller of the two damage scenarios. Still its reliability is 100% for the smaller of the two damage cases at points A and B.
Like their homogeneous counterparts, the heterogeneous DSF (Equation (4.20)) values increase with the damage level and proximity to the damage location. As expected, the dynamic DSF (Equation (4.20)) takes on its largest values at points A and B, demonstrating its localization capability.
97
TABLE 4.8
DETECTION RESULTS OF HOMOGENEOUS AND HETEROGENEOUS DYNAMIC DSF FOR SIMULATED THIN BEAM
POINT A POINT B
Homogeneous Heterogeneous Homogeneous Heterogeneous
Threshold 1.49 1.49 1.49 1.53
Area Lost 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50%
Test 1 1.01 1.04 1.06 1.91 32.44 136.2 0.99 1.04 1.38 1.27 2.68 5.78
Test 2 0.65 0.63 0.92 1.34 31.76 137.1 1.48 1.90 2.01 0.62 3.52 4.88
Test 3 0.73 0.72 0.70 1.22 33.12 139.3 1.16 1.86 2.20 1.14 2.62 4.71
Test 4 0.71 0.78 1.03 0.69 32.47 136.0 0.80 1.17 1.48 0.53 3.54 4.93
Test 5 0.57 0.62 0.76 0.41 32.70 136.5 0.73 1.03 1.45 0.31 2.53 4.62
Test 6 1.01 1.06 1.15 0.80 33.66 139.2 1.06 1.66 2.11 0.82 2.52 4.93
Test 7 0.72 0.83 0.68 0.68 32.77 137.2 0.49 0.91 1.40 0.97 3.18 4.69
Test 8 1.03 1.05 0.98 0.61 32.22 136.2 0.93 1.51 1.62 0.69 3.55 5.04
Test 9 0.92 0.76 0.60 1.15 32.40 137.0 1.13 1.74 1.87 1.02 2.48 4.70
Test 10 0.89 0.71 0.73 1.25 31.63 134.3 1.40 1.88 1.96 1.00 3.01 4.70
Det Rate 0% 0% 0% 10% 100% 100% 0% 60% 60% 0% 100% 100%
98
TABLE 4.8 (CONTINUED)
DETECTION RESULTS OF HOMOGENEOUS AND HETEROGENEOUS DYNAMIC DSF FOR SIMULATED THIN BEAM
POINT C POINT D
Homogeneous Heterogeneous Homogeneous Heterogeneous
Threshold 1.51 1.47 1.48 1.53
Area Lost 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50% 0% 0.50% 1.50%
Test 1 0.96 1.30 1.83 1.27 1.39 1.92 0.92 1.15 1.38 1.27 1.63 2.69
Test 2 1.26 1.37 2.24 0.62 0.93 1.69 1.21 1.29 2.12 0.62 0.88 1.51
Test 3 1.10 1.19 2.42 1.14 1.16 2.22 0.93 0.87 1.61 1.14 1.17 2.07
Test 4 0.58 0.97 1.90 0.53 0.89 1.78 0.72 1.06 1.26 0.53 1.25 2.03
Test 5 0.73 1.00 1.74 0.31 0.59 1.82 0.73 1.05 1.69 0.31 1.33 2.33
Test 6 0.95 1.79 2.44 0.82 1.15 1.89 0.82 1.28 1.92 0.82 1.28 2.72
Test 7 0.62 1.43 2.27 0.97 1.30 2.12 0.92 1.17 1.56 0.97 0.55 1.74
Test 8 1.08 1.25 1.98 0.69 0.75 1.68 0.67 0.80 1.33 0.69 1.12 2.04
Test 9 0.99 1.30 1.66 1.02 1.00 1.72 1.02 0.79 0.98 1.02 1.31 2.33
Test 10 1.28 1.87 3.00 1.00 0.98 1.84 1.23 1.27 1.47 1.00 1.38 1.92
Det Rate 0% 20% 100% 0% 0% 100% 0% 0% 50% 0% 10% 90%
99
TABLE 4.9
SUMMARY OF DETECTION RESULTS FOR SIMULATED THIN CANTILEVER BEAM:
COMPARISON OF STATIC HOMOGENEOUS AND HOMOGENEOUS/HETEROGENEOUS
DYNAMIC DAMAGE SENSITIVE FEATURES
Static DSF Dynamic DSF
Homogeneous Homogeneous Heterogeneous
Volume Lost 0% 0.5% 1.5% 0% 0.5% 1.5% 0% 0.5% 1.5%
Det. Rate 0% 0% 0% 0% 0% 0% 10% 100% 100%
Avg. DSF -0.38 -0.38 -0.38 0.82 0.82 0.86 1.01 32.52 136.9
Det. Rate 0% 0% 20% 0% 60% 60% 0% 100% 100%
Avg. DSF -0.84 -0.79 -0.61 1.02 1.47 1.75 0.84 2.96 4.90
Det. Rate 0% 0% 20% 0% 20% 100% 0% 0% 100%
Avg. DSF 0.73 0.68 0.48 0.95 1.35 2.15 0.83 1.01 1.87
Det. Rate 0% 0% 0% 0% 0% 50% 0% 10% 90%
Avg. DSF -0.42 -0.28 -0.37 0.92 1.07 1.53 0.84 1.19 2.14
Thus, damage detection capability within the heterogeneous framework is
dramatically improved in comparison to homogeneous methods, without a sizeable
increase in the rate of false positives, even for the very modest levels of damage
considered here. Thus, Equation (4.20) is adopted as the aforementioned Bivariate
A
B
C
D
100
Regressive Adaptive Index (BRAIN) for damage detection within a decentralized,
wireless sensor network.
To further explore the role of damage proximity/severity, observability and
signal strength, the same model is revisited with additional damage scenarios between
points B & C and C & D, respectively termed damage patterns 2 and 3. Acceleration and
surface strain time history pairs are repeatedly simulated at the four locations (A-D)
shown in Figure 2.5, under the action of Gaussian white noise inputs applied at the free
end. After generating a collection of simulated strain/acceleration pairs to form a
reference database, damage was subsequently introduced to the beam through a
transverse cut, symmetrically imparted midway between two of the measurement
points. The transverse dimension of the cut is specified as a percentage (0, 0.5%, 1%,
1.5%, 2.5%) of the total width of the beam; the longitudinal dimension of the cut is fixed
at 5% of the total beam length (WD = 0.05L = 2.5 cm), as shown in Figure 2.5. For each of
the damage patterns and severity levels, the response of the beam is generated through
ten independent simulations so that the damage detection rate (repeatability) of the
DSF can be reported. Damage is detected whenever the 97.5%, two-sided confidence
interval is surpassed for both heterogeneous detection (Equation (4.18)) and
homogeneous detection (Equation (4.17)).
101
Figure 4.7: First five normalized mode shapes of simulated thin beam with measurement points superimposed.
Several factors influence the sensitivity of damage detection spatially along the
beam and in some cases their trends offset one another. Thus before discussing results,
it is important to establish these factors. Obviously, considering the small levels of
damage simulated here, proximity to the damage site will be a considerable factor in
damage detection rates and is explored by moving the damage zone to three different
locations. Another issue, however, is observability. Particularly considering that
102
evidence of damage can be more pronounced in the higher modes, response
measurements at antinodes of these modes will obscure some of this valuable
information. Table 4.10 lists the percent reduction in stiffness for the first five modes for
each of the simulated damage patterns. Note that for all damage patterns, the first
mode is not the most affected mode, and the impact of the damage on this mode
diminishes as the damage moves toward the free end. For Damage Pattern 1, damage
between locations A and B, the second mode is most impacted. Damage Pattern 2,
damage between locations B and C similarly affects the second mode most, though the
fifth mode is nearly affected in equal proportion. Meanwhile, Damage Pattern 3,
damage between points C and D, most significantly affects the fifth and even fourth
modes. Considering the mode shapes and the proximity of their antinodes to the
measurement points, as shown in Figure 4.7, the modes most affected by Damage
Patterns 1 and 2 are hardly observed at point C, while the mode most affected by
Damage Pattern 3 is unobservable at Point B. The participation factors of the first five
modes of the undamaged system, from first to fifth, proportion as follows:
5.5:3.1:1.8:1.3:1 and this relative participation is affected differently by each damage
pattern, as also summarized in Table 4.10. Modal participation overall is most affected
by Damage Pattern 1, while the single most affected mode in each damage pattern is
Mode 5, Mode 3 and Mode 4, respectively, so clearly damage does become most
discernable through changes in the dynamics of the higher modes.
103
Figure 4.8: Average stiffness lost in the first five modes of simulated thin beam as a function of cross sectional area removed
and location of damage (damage pattern).
Another factor is signal strength, which the data-driven approach to damage
detection seeks to offset to some extent. Moving from the fixed end toward the free
end, acceleration responses increase, while surface strains decrease. Thus point D not
only cannot fully benefit from a heterogeneous sensing approach due to the low strain
levels at that location, but as explained shortly, may even suffer. An additional factor to
be considered is damage severity. While the same amount of cross sectional area is
compromised in Damage Patterns 1-3, the location where that damage is imparted has
varying effects on the stiffness lost. The average percent of stiffness lost in the first five
modes from Table 4.10 is visually displayed in Figure 4.9 to emphasize the greater losses
associated with Damage Pattern 1.
104
TABLE 4.10
PERCENT STIFFNESS LOST AND MODAL PARTICIPATION FACTOR CHANGE (ABSOLUTE) FOR EACH DAMAGE PATTERN FOR FIRST FIVE
MODES OF SIMULATED THIN BEAM
Damage Pattern 1: Area Loss Damage Pattern 2: Area Loss Damage Pattern 3: Area Loss
Mode 0.5% 1.0% 1.5% 2.5% 0.5% 1.0% 1.5% 2.5% 0.5% 1.0% 1.5% 2.5%
Percent Stiffness Lost
1 0.8% 1.8% 3.1% 7.0% 0.0% 0.1% 0.2% 0.4% 0.0% 0.0% 0.0% 0.0%
2 2.2% 4.8% 7.6% 15% 0.5% 1.1% 1.8% 4.2% 0.0% 0.1% 0.1% 0.3%
3 0.8% 1.7% 2.8% 6.2% 0.4% 0.8% 1.4% 3.0% 0.1% 0.3% 0.6% 1.4%
4 1.4% 3.0% 4.6% 8.5% 0.1% 0.1% 0.2% 0.5% 0.4% 0.8% 1.4% 3.2%
5 1.4% 3.1% 5.0% 10% 0.5% 1.1% 1.9% 3.9% 0.5% 1.2% 2.0% 4.4%
Avg. 1.3% 2.9% 4.6% 9.3% 0.3% 0.6% 1.1% 2.4% 0.2% 0.5% 0.8% 1.8%
Absolute Percent Change in Modal Participation Factor
1 0.1% 0.1% 0.2% 0.5% 0.0% 0.1% 0.1% 0.2% 0.0% 0.0% 0.0% 0.0%
2 0.3% 0.6% 1.1% 2.4% 0.1% 0.2% 0.4% 0.9% 0.0% 0.1% 0.1% 0.2%
3 0.6% 1.2% 2.1% 4.5% 0.6% 1.2% 2.1% 4.5% 0.1% 0.2% 0.3% 0.7%
4 0.1% 0.4% 0.7% 1.4% 0.3% 0.7% 1.1% 4.5% 0.1% 0.3% 0.4% 0.8%
5 0.9% 1.8% 3.0% 6.2% 0.0% 0.2% 0.4% 0.9% 0.0% 0.0% 0.0% 0.5%
Avg. 0.1% 0.3% 0.5% 1.0% 0.0% 0.1% 0.2% 0.4% 0.0% 0.1% 0.2% 0.2%
105
Figure 4.9: Matrix of damage detection rates on simulated thin beam under random excitation results (columns are damage
locations, rows are measurement locations).
With these important considerations in hand, the damage detection results
matrix is now presented in Figure 4.9, whose rows indicate the measurement locations
and whose columns signify the damage location. For example, the figure in the first row,
first column indicates the results for measurements at location A when damage is
between points A and B (Damage Pattern 1). At each position in the matrix, various
degrees of damage are simulated, as shown on the x-axis by the amount of area
D
C
B
A
RES
PO
NSE
MEA
SUR
EMEN
T LO
CA
TIO
N
× BAR (na=13, nb=7) o AR (na=13) AR (na =20)
DAMAGE LOCATION: DAMAGE PATTERN 1 3
106
removed. Note the 0% damage case is provided to evaluate any tendency toward false
positives (alerts of damage when none is present). In general, the damage detection
rate for the heterogeneous method exceeds that of the homogeneous method and
improves at measurement locations closest to the damage site. The only exception is for
the lowest damage cases approaching the free end, where the low amplitude of strain
responses may actually prove detrimental to the underlying BAR model at minor
damage levels. In addition, the issue of observability can be demonstrated by the
reduced detection rates at point C, as expected based on earlier discussions regarding
Damage Patterns 1 and 2, though still demonstrating that as the damage moves to
locations adjacent to point C and begins affecting modes that are observable at Point C,
damage detection rates improve. A similar observation can be made for point D. Again
keeping in mind the spatial sensitivity of area removal, the damage scenario in the third
column in particular has especially minor impacts on overall stiffness (see Figure. 4.8)
and is thus nearly five times harder to detect. Therefore it is not particularly shocking to
see the reduction in damage detection rate toward the free end of the beam. In
addition, as the strains are appreciably lower at points C and D, these locations receive
the least performance enhancement when heterogeneous detection is implemented,
and as one would expect, point A benefits the most as a region of high strain and low
acceleration. This is evident not only from the DSFs but also when inspecting the overall
quality of signal reconstruction by AR and BAR, which does of course underpin this DSF.
Let the quality of signal reconstruction be quantified by the standard deviation of the
residual error (). At point A, the error in the BAR reconstruction (na=13, nb=7) is
107
0.0041 m/s2, while it is orders of magnitude larger for the AR reconstruction (na=20) at
0.3279 m/s2. Meanwhile, at point D, while the BAR reconstruction is more accurate (
= 0.3384 m/s2), the improvement is not as marked (20th order AR reconstruction:
=0.4325 m/s2). Thus a further modification to this damage detection framework could
include an evaluation of relative signal strength with the provision for neglecting any
response types that were not sufficiently prominent relative to the noise floor and
modeling the remaining response quantity by a pure AR (homogenous) model. Finally
the vulnerability to false positives should be evaluated, as one of the most common
consequences of enhanced sensitivity and a severe contributor to the erosion of end
user confidence. Though neither DSF shows significant susceptibility to false positives,
the homogeneous method is slightly more pre-disposed to this behavior.
108
Another particularly relevant consideration for regressively modeled signals is
the determination of an appropriate model order. As the end goal of this research is the
embedment of these DSF algorithms in wireless platforms with stringent computational
and power constraints, low model orders are an important asset. In general, previous
comparisons between heterogeneous and homogenous detection generally preserved
the total model order (AR: na= BAR: na+nb), determined by seeking the order best
D
C
B
A
RES
PO
NSE
MEA
SUR
EMEN
T LO
CA
TIO
N
Figure 4.10: Matrix of standard deviation of residual error in underlying regressive model fit to simulated thin beam under random excitation (columns are damage locations, rows
are measurement locations).
AR (na=7) o AR (na=13) AR (na=20)
BAR (na=7, nb=3) x BAR (na=13, nb=7)
DAMAGE LOCATION: DAMAGE PATTERN 1 3
109
minimizing residual errors. As the damage detection rate in Figure 4.9 is tied to both the
performance of the DSF and the underlying regressive model, a similar performance
matrix is now presented in Figure 4.10 to isolate the influence of the underlying
regressive model. In all cases, the BAR model provides the lowest residual error and in
all cases na=13, nb=7 produces superior signal representation compared to na=7, ba=3,
which would be expected given the number of modes generally participating in the
response. This finding is again the motivation for using heterogeneous sensing in this
research and is a major contributing factor to its superior damage detection rates in
Figure 4.9. Further, the measured responses at location A demonstrate very minor
sensitivity to AR model order once a minimum order of na=13 is employed, which is
capable of representing the first five modes participating in the response. Responses
measured at point B show hardly any sensitivity to order of the AR model and only for
Damage Pattern 1 does the BAR model with na=7, nb=3 perform better than the AR
models, but keep in mind that at this point, two of the first five modes are unobservable,
thus allowing for a reduced model order. Similar to the observations at point A, for
responses measured at point C, once an AR model order of 13 is achieved, there is no
appreciable improvement in performance for higher model orders, though the poor
performance of AR models with na=7 is far more marked at this location, as the
resonant responses are more pronounced at this location, making their narrowband
response particularly sensitive to model order. This trend becomes increasingly evident
moving toward the free end of the beam, where residual errors associated with na=7 AR
model grows larger. However, at point D is the first evidence of BAR model sensitivity,
110
where the AR na=20 model performs better than the BAR na=7, nb=3, which actually
has the second highest residual error after AR na=7 model. This also is the location with
the highest residual errors overall. As the strain response diminishes as this location, it is
not capable of offsetting the low model order on the acceleration term in the BAR
representation, though it still should be noted that BAR na=13, nb=7 still performs
better than AR na=20 or AR na=13 (which as the same model order on its acceleration
term as the BAR model). Thus, it can be concluded that while BAR models of sufficient
order are more effective than their AR models of comparable order, this effectiveness is
maximized at locations where strain response is strongest. In total, these results also
clearly demonstrate that there is sensitivity to model order depending on the damage
scenario and location of response measurement, however the use of BAR model with
na=13, nb=7 and AR na=20 respectively produce the best performance across the board
for heterogeneous and homogeneous applications in this system and will serve as the
default cases for the remainder of this example.
Having now established the role of model order, the overall damage detection
rate through a data-driven DSF in Figure 4.9 can be revisited to shed additional light on
this issue. Here, detection rates for AR na=20 (same total order as comparison BAR
model) and 13 (same acceleration model order as comparison BAR model) are
presented along with the consistently superior comparison BAR model with na=13, nb=7
results. Again recall that na=20 and na=13 AR models showed little difference in residual
error for all measurement locations except D. This indicates that the underlying signal
representation is reconstructed with equivalent accuracy by these two AR models,
111
except at point D. Thus one may expect them to perform comparably at all locations but
D; however, this is not the case. While in some cases, the AR model order shows little
sensitivity (measurements at locations B and C in Damage Patterns 2 and 3), at
measurement location A, a pronounced improvement in performance is realized for all
three damage patterns when a lower model order AR is adopted, in direct contrast with
the results in Figure 4.10. This indicates that over-specifying the order of the AR model
likely leads to “mathematical modes” that can interfere with the performance of a data-
driven DSF. In DSFs that identify specific regressive model coefficients a priori (e.g., first
few AR coefficients), over specification of model order may not be a considerable
concern; however, in a data-driven DSF there is potential for one of these spurious
mathematical modes to appear erroneously as the most sensitive to damage, despite
having no physical basis, and thus be adopted as the metric for damage detection. This
is particularly an issue at point A due to the comparatively smaller acceleration levels at
this location. At other locations, this becomes less of a concern, as the acceleration
responses in the actual structural modes are likely affected significantly enough by the
damage that they drive the most sensitive regressive coefficient in the data-driven DSF
and overpower the influence of any mathematical modes. This more importantly
demonstrates that minimization of residual error itself may not be the sole or most
important factor in the selection of model orders or that AR model orders need be
adapted when using a data-driven DSF to account for the variability of modal
participation or acceleration signal strength at each measurement location. At locations
where acceleration responses are strongest and all modes are observable, less
112
sensitivity to AR model order is noted (see location D). Also recall that Damage Pattern 1
is the most severe, impacting modal participation and stiffness most significantly, and it
is this pattern that shows greatest sensitivity to AR model order along the length of the
beam. Still, with the exception of one scenario (lowest area losses in Damage Pattern 1
at location C), the damage detection rates using heterogeneous sensing consistently
perform at or above the rates of comparable order homogeneous representations.
4.2.2.2 Validation Using Experimental Thin Beam
The thin beam experimental assembly introduced in Section 2.1.3 was damaged
by a transverse cut, symmetrically imparted at mid-point of the beam as shown in Figure
4.11. The length of the cut (25mm) is 5% of the total length of the beam (500mm) and
the depth of the cut (1.25mm) is 10% of the total width of the beam (25mm). Thus the
total volume lost is 0.5%. The dynamic DSF is then applied to the recorded responses of
this beam under random base excitations, with the results presented in Table 4.11.
Figure 4.11: Definition of damage lengths on thin cantilever beam (plan view).
25
mm
1.25 mm
25 mm
113
TABLE 4.11
COMPARISON OF DYNAMIC DSF IN HOMOGENEOUS AND HETEROGENEOUS FORMATS
USING EXPERIMENTAL THIN CANTILEVER BEAM UNDER WHITE NOISE EXCITATION
POINT A POINT B
Volume
Lost
Homogeneous Heterogeneous Homogeneous Heterogeneous
0% 0.50% 0% 0.50% 0% 0.50% 0% 0.50%
Detection
Rate 0% 50% 0% 100% 0% 70% 0% 20%
POINT C POINT D
Area Lost
Homogeneous Heterogeneous Homogeneous Heterogeneous
0% 0.50% 0% 0.50% 0% 0.50% 0% 0.50%
Detection
Rate 0% 100% 0% 100% 0% 100% 10% 80%
From Table 4.11, several important conclusions can be drawn about the
heterogeneous formulation:
Incidence of false positives for the heterogeneous approach (Equation (4.20)) is negligible in comparison with its detection rate.
At point A (the nearest point to the fixed end), where strain is largest, the performance of heterogeneous DSF is much better than the homogeneous ones. At other locations, as acceleration responses increase, the performance of homogeneous DSF is comparable to the heterogeneous DSF.
Because the damage is the small, neither DSF showed significant capability for damage localization.
114
Note that at point C, the heterogeneous method performs poorly and contrary
to the expected performance in earlier simulations. This underscores particularly the
importance of high quality strain measurements in the BRAIN method. To insulate from
this vulnerability, the BRAIN algorithm can be modified with an override that will default
the method to a homogeneous form any time poor signal quality is detected from one
element of the heterogeneous array.
4.2.2.3 Validation Using Steel Truss Bridge Model
The validations herein will continue to compare the dynamic DSFs in their
homogeneous form to new results for the heterogeneous formats using the
experimental bridge assembly’s impact tests. For detection by heterogeneous sensing,
each acceleration signal is paired with the strain signal from the nearest horizontal bars.
These data are fit by an 8th order BAR model (na=5; nb=3), and Equation (4.20) is used
to generate heterogeneous DSFs for damage detection. Figure 4.12 shows the
comparison of damage detection rates between dynamic homogenous DSF and dynamic
heterogeneous DSFs, again for measurement and excitation points moving progressively
toward the bridge’s mid-span.
115
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
1 2 3 4 5
I II III IV I II III IV I II I II III IVIII IV I II IIIIm
pact
@ P
oint
2 IV
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
Impa
ct @
Poi
nt 4 IV I II III IVIII IV I II IIIII III IV I III II III IV I
1 2 3 4 5
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
Impa
ct @
Poi
nt 5 III IV I II IIIII III IV I II IV I II III IVI II III IV I
1 2 3 4 5
Figure 4.12: Damage detection rate comparison between homogenous dynamic DSF (Grey Bars) and heterogeneous
dynamic DSF (Black Bars) on experimental thin beam.
For heterogeneous detection, the performance will be determined by the
qualities of both acceleration and strain measurements. The quality of acceleration
signals are affected by two factors: the amplitude of responses and the modal
participation. By considering those two factors, excitations at point 4 were found to
produce the best damage detection, while point 5 was the worst. But strain signals are
dominated by the fundamental mode, so higher order modal participation has a minor
116
effect and the prevailing factor is the amplitude of the response. As a result, the best
strain signals are generated when excitations are provided at Point 5.
The benefits of such redundancy in detection capability, particularly as the two
signal types have offsetting behaviors, can be seen in the results when point 2 is excited.
Here, particularly for the smaller damage scenario, detection rates increased from 20%
to 100%. Further at the driving point (when both measuring and exciting point 2), the
detection rates are dramatically improved when strain is considered. The results when
exciting at point 5 clearly demonstrate that as strain response increases, the detection
capability is considerably enhanced, though again the inclusion of strain cannot fully
compensate for the unobservability of the higher modes in the acceleration response
that are the most sensitive to damage. This is clearly seen when point 5 is the driving
point. Overall, the results demonstrate that the expansion of the sensing array can help
to offset the limitations of acceleration-only measurements.
Then, one question is why the inclusion of strain actually worsened performance
in the case of excitations at point 4. This actually can be explained by the model orders
selected. For consistency, the total AR or BAR model order was kept fixed at a total of 8
coefficients. For homogeneous sensing, this implied that an 8th order model was
entirely devoted to acceleration, whereas, in the BAR model, the order of the fit to
accelerations was reduced to 5 to enable a 3rd order fit to the strain data. Therefore
some of the higher mode information, which often is most sensitive to damage, is not
retained. To demonstrate the results with higher mode information for both
117
homogeneous and heterogeneous DSFs, Figure 4.13 shows the detection rates for 8th
order AR model and 11th (na=8, nb=3) order BAR model.
Measuring Point
Damage Scenario
100%
80%
60%
40%
20%
1 2 3 4 5
I II III IV I II III IV I II III IVIII IV I II III IV I II
Figure 4.13: Damage detection rate comparison between 8th order homogenous dynamic DSF (Grey Bars) and 11th order (na=8, nb=3) heterogeneous dynamic DSF (Black Bars) for experimental
thin beam.
The results in Figure 4.13 show that when the acceleration model orders are
same for homogeneous and heterogeneous DSFs, the damage detection rates of 11th
order BAR model are no lower than those of 8th order AR model and in some cases are
superior.
4.3 Data fusion at the Meso-net
As discussed in Kijewski-Correa et al. (2006 b), one of the major advantages of
the multi-scale network concept is the ability to fuse data locally to enhance detection
capabilities and reduce the probability of false positives, which is a natural byproduct of
highly sensitive damage indicators. It is hypothesized that this higher level of
information exchange will be effective given the spatial sensitivity inherent in the DSFs
as demonstrated in numerous examples in this chapter. At each wireless sensor node,
118
the result of the online damage detection process described in this chapter would
generate a local binary report B (0 = no damage, 1 = damage). As proposed in Kijewski-
Correa et al. (2006 b), data fusion can be achieved through a number of approaches, the
most basic would be a local voting process involving the m nearest neighbors, signaling
damage only when indicated by majority, i.e.,
1)2
( m
floorBm
(4.21)
This approach was used in that study to improve the performance of DSF based on the
residual errors of the BAR model. To demonstrate the merits of data fusion within the
m-net for the dynamic, heterogeneous DSF introduced in this chapter, the simulated
thin beam model’s detection results in Table 4.8 are revisited. Each measuring point
considers the binary report from the neighboring sensors to either side to form a m-net
for decision making. Table 4.12 shows the damage detection rates before (taken from
Table 4.8) and after this data fusion process. This simplified fusion scheme improved
the damage detection in select homogeneous and heterogeneous assessments, as
shaded in grey. Note that in a number of cases, no improvement could be achieved
since the detection rate was already perfect. More importantly the only false positive
was eliminated by this fusion process. Note that this is a relatively simplified approach
to local decision making within the network provided to demonstrate the merits of this
added network-level processing. The addition of a weighting function as well as the
introduction of other more sophisticated approaches, provided they are not too
computationally intensive, will likely enhance performance even further.
119
TABLE 4.12
DAMAGE DETECTION RATE BEFORE AND AFTER LOCAL VOTING PROCESS
Point A Point B
Homogeneous Heterogeneous Homogeneous Heterogeneous
Area lost % 0% 10% 30% 0% 10% 30% 0% 10% 30% 0% 10% 30%
Before Fusion 0% 0% 0% 10% 100% 100% 0% 60% 60% 0% 100% 100%
After Fusion 0% 0% 10% 0% 100% 100% 0% 60% 80% 0% 100% 100%
Point C Point D
Homogeneous Heterogeneous Homogeneous Heterogeneous
Area lost % 0% 10% 30% 0% 10% 30% 0% 10% 30% 0% 10% 30%
Before Fusion 0% 20% 100% 0% 0% 100% 0% 0% 50% 0% 10% 90%
After Fusion 0% 30% 100% 0% 10% 100% 0% 0% 60% 0% 30% 100%
4.4 Summary
This chapter introduced and validated the online detection approach using
various simulated and experimental test beds introduced in Chapter 2 to confirm two
hypotheses:
Data-driven or dynamic DSFs are more robust and reliable than their static
counterparts in homogeneous sensor networks.
Heterogeneous DSFs are more robust and reliable than their homogenous
counterparts.
These results motivated the use of a time domain, bivariate autoregressive
damage detection method called BRAIN. Its novel feature is a dynamic DSF operating on
heterogeneous data pairs of strain and acceleration. In addition to its enhanced
detection capability, this format reduces the required on-board memory and
computational demands (and power drain) associated with the manipulation of a
120
reference database since only a few key statistics of the reference pool and a pre-
defined statistically significant threshold are required for damage detection. The
method showed no enhanced susceptibility to false positives and even the capability for
localization. An in depth exploration of various factors influencing detection rates were
also presented. A simplified approach to data fusion within the network was also
offered to demonstrate the additional enhancements in damage detection ability and
reduction of false positives. Figure 4.14 now shows the new benefits afforded by this
approach to online detection.
Figure 4.14: Overview of key features of proposed wireless sensor network for structural health monitoring, with addition of new
benefits introduced in Chapter 4.
BENEFIT APPROACH STAGE
DATA ACQUISITION
DATA REDUCTION
DETECTION
LOCALIZATION
Heterogeneous, Multi-scale WSN
Low power, scalable
Bivariate Autoregressive, Reference Database
Data-Driven DSF
More sensitive to damage, Easily embedded
More reliable, computational demand
reduced
121
CHAPTER 5:
RESTRICTED INPUT ACTIVATION STRATEGIES
In essence, the structural health monitoring process becomes an exercise in
statistical pattern recognition when actual bridges are considered and the ability to
account for operational and environmental variabilities becomes a key practical issue.
For bridge SHM strategies that rely on ambient vibration response characteristics, this
ability is essential not only for enhancing the reliability of detection but also to avoid
false-positives, as again the input to the system is never known explicitly. Instead, this
research offered a compromise between traditional operational monitoring and
controlled testing: the Restricted Input Network Activation Scheme (RINAS) discussed in
Chapter 3. This concept requires a diagnosis of the environmental conditions and traffic
loading on the bridge. The latter will be achieved using widely available traffic cameras,
knowing the former can then be readily quantified using meteorological stations. Both
the traffic camera and meteorological station would be installed at the gateway node,
as described in the Chapter 3. This gateway would be responsible for all processing of
the data to make a determination regarding whether to trigger the network provided
the user specified loading and environmental conditions are met and acquire bridge
response data. This concept was shown previously in Figure 3.2. To reiterate the
significance of this feature, although RINAS is not able to explicitly control or measure
the input, it does allow the operational and environmental states to be restricted to a
122
specific subset for which a reliable reference pool has been generated, e.g., the passage
of a semi-trailer at night under a particular weather condition. This not only enhances
damage detection reliability, but also reduces the size of the reference pool, thereby
easing computational burden and memory demands. Furthermore, this form of
triggering helps to increase network lifetime since wireless sentinel functions are not
required, and the network operates only when the target conditions are met. The
chapter introduces a camera-based implementation of RINAS using an embedded
demonstrative example of vehicle classification, followed by a validation of the RINAS
concept using the DSF introduced in the previous chapter.
5.1 Camera-Based Traffic Classification with Illustrative Example
This research adopts an intelligent video sensor (IVS) as a first choice for
identification of traffic conditions. The IVS combines video sensing with image
processing and data communication, so that from a captured video stream, high-level
traffic parameters are computed and then transmitted to the WSN MNode. IVS
represents a relatively new technology in comparison to traditional traffic sensors like
magnetic sensors or inductive loops, which can evaluate parameters such as the number
of passing vehicles, their speed and length. The most significant disadvantage of these
traditional sensors lies in the fact that they can survey only a limited region of the traffic
path. On the other hand, video-based systems can monitor and analyze a wider area to
provide a complete description of the oncoming traffic. Another advantage of IVS is its
high performance, reliability, and accuracy compared to other traffic sensors in the field
123
(Bramberger, et al. 2003). However, to reap these clear advantages, efficient and
accurate image classification techniques must be developed, as now described.
Image recognition is any form of signal processing for which the input is an
image, such as photographs or frames of video, and the output is either a modified
image or a set of characteristics or parameters related to the original image. Image
processing seeks to address three major problems concerned with pictures (Petrou and
Bosdogianni 1999):
Image digitization and coding
Image enhancement and restoration
Image segmentation and description
In recent decades, extensive research and development efforts have been
devoted to image processing techniques applied to traffic data collection and analysis
(Hoose 1991; Hoose 1992). These efforts can be divided into quantitative and qualitative
analyses. Quantitative analysis is the extraction of traffic parameters such as vehicle
counts, vehicle length, lane occupation, speed, etc., while the qualitative analysis simply
seeks to describe a traffic scene based on pre-defined categories. RINAS requires solely
the former as a means to trigger the WSN on a bridge.
Although image processing and recognition of moving objects pose a complex
mathematical, algorithmic and programming problem, a simplified image classification
algorithm must be developed to operate within the M-node’s local resources and enable
a rapid classification to trigger the network in near real time. Simply put, this requires
the system to autonomously distinguish large semi-trailers from passenger vehicles and
124
trigger only under the former condition, a problem that lends itself well to contour
extraction. Herein, the contour detection process will be divided into several
independent processing steps in order to solve the task logically. These steps are in the
following order of algorithmic processing: video stream conversion to single frames,
lane masking, background removal, noise filtration and contour extraction. Each step
has a specific processing algorithm, as now described. The approach here is based upon
the work of previous studies (Taktak, et al. 1996), but offering a new contour extraction
scheme to perform vehicle classification.
5.1.1 Video Conversion
The RGB model will be used to define a RGB value for each point in a scene2,
which will then be the basis for subsequent algorithms and calculations on the image.
5.1.2 Lane Masking
This operation is made to separate the image by traffic flow direction, to isolate
the part of the road with oncoming traffic and thus simplify the subsequent image
processing. The masking algorithm is given by formula:
)()()( pVpMpN (5.1)
2 The RGB color model is an additive color model in which red, green, and blue light are added
together in various ways to reproduce a broad array of colors. The name of the model comes from the initials of the three additive primary colors, red, green, and blue (Munch and Steingrimsson 2006), (Petrou and Bosdogianni 1999).
125
where )(pM is RGB value at an image point in primary frame, )(pN is a new image
point in the output image, )(pV is mask value for point p. 0)( pV if the corresponding
pixel is eliminated, otherwise 1)( pV . Figure 5.1 shows an example scene before lane
masking and after lane masking.
Figure 5.1: Example traffic scene (a) before masking and (b) after lane masking.
5.1.3 Background Removal
This algorithm removes all stationary objects from the lane observation zone
leaving only the items whose position has changed from frame to frame. This includes
vehicles as well as image noise and moving background images like swaying trees, flying
birds, shadows, precipitation, etc.
The background )(pB is calculated as an average value of the same image point
p in the pre-selected N background frames (frames without vehicles):
Nk
B
N
kPIpB
:1
),()(
(5.2)
126
where )(pB is the averaged background value for point p. ),( kpI B is the thk pre-
selected background frame. N is number of pre-selected background frames used in the
average.
The background removal frame for a random traffic scene image )( pI should be
the difference between new scene frame and the background frame.
)()()( pBpIpR (5.3)
where )(pR is the background removal frame and denotes the Euclidean distance
calculation. Figure 5.2 shows the same example scene as Figure 5.1, before background
elimination and after background elimination.
Figure 5.2: Simulation scene (a) before background elimination and (b) after background elimination.
5.1.4 Noise Filtration
By the time the image enters this stage it may have light background clouds that
are caused by differences between the individual background and the average
background in the previous stage. Noise in particular will surface as speckling in the
127
image. Filtration is an effective method to remove these effects. Here, filtration will be
conducted in two steps. The first step is threshold filtration to remove light background
colors, and the second step is median filtration to remove speckles.
The first step will be the removal of some light colors, easily accomplished using
a fixed threshold:
otherwise
thresholdpRforpRpF
0
)()()(
(5.4)
where )( pF is the RGB value of the point. The speckles are removed by a second
algorithm that relies on a nearest neighbor principle: if an image point p belongs
legitimately to an object then at least one of its eight adjacent neighboring points will
also be a part of this object. If its adjacent points do not belong to an object then the
point p does not either. Figure 5.3 shows the example scene from Figure 5.2 before
noise filtration and after noise filtration, note now that the three vehicles have been
successfully isolated.
Figure 5.3: Simulation scene (a) before noise filtration and (b) after noise filtration
128
5.1.5 Contour Extraction
The final stage identifies the remaining objects in the filtered image. For RINAS,
this classification need only be binary: is the vehicle a semi-trailer or a passenger
vehicle? Thus the contour extraction simplifies to a contour area recognition problem: if
the contour area in a scene is larger than a threshold representative of the surface area
of a semi-trailer, a vehicle of this class must be approaching. This would then indicate
one of the target operational conditions that may trigger the network. The following
steps would be necessary to extract and calculate the contour area.
First, define the four points 1p - 4p adjacent to pixel p , beginning from the pixel
directly above pixel p, moving clockwise, as shown in Figure 5.4. Each pixel is a unit
area. The basic idea here is to find out the number of consecutive pixels whose RGB
values are nonzero. This procedure starts by choosing a point close to the center of the
scene, whose RGB value is nonzero ( 0)( pF ), as starting point p. Then the next step is
to check the neighboring points encompassing p to form a bounding circle. For a given
pixel, if no RGB values in the bounding circle are zero, then the pixel is still in the
continuous area of the object and the bounding circle moves incrementally outward
toward the boundary of the object, repeating the analysis. When the procedure reaches
the boundary, RGB values of one or more points neighboring p will be zero, terminating
the bounding circle algorithm and thus determining the boundary of the image.
129
1p
4p p 2p
3p
Figure 5.4: Designation of adjacent pixels defining the bounding circle for contour area extraction.
The calculation of the area of an object can be conducted simultaneously by
introducing a target variable ‘Area’. The initial value of ‘Area’ is 0 and the final value of
‘Area’ will be used to indicate the size of the vehicle. For each pixel inside the boundary
of the object, the RGB value will be greater than zero and the value of ‘Area’ will be
increased by 1, provided that the pixel had not previously been checked. As the
bounding circle shifts outward, it will center at one the neighbors of the previous pixel
(p1-p4). The algorithm guarantees every pixel in the continuous nonzero RGB area will be
counted without repetition. Again, as the algorithm stops automatically when it gets the
border, the value of ‘Area’ at this point is an estimate of the vehicle’s size and can be
compared to pre-determined sizes for various vehicle classes based on its position
within the frame (foreground or background vehicle). This process is shown
schematically in Figure 5.5. For the illustrative example provided here, trucks occupy 20-
25% of the whole frame, depending on the spatial location, while smaller passenger cars
usually take less than 5% of the frame, again dependent on spatial location. Because the
algorithm is designed only to calculate the area and not analyze the details of the image,
the process described here takes about 0.1 second in MATLAB for a regular dual 3GHz
130
processor desktop PC with 2GB of RAM. Considering vehicles traveling at a speed of 70
mph, a gateway would need to be placed at least 10 feet up traffic of the instrumented
bridge to complete this calculation and make the decision to trigger the network. Given
the gateway’s computational resources are less robust; the gateway would need a larger
separation distance in field applications.
Figure 5.5: Logic tree for contour area calculation
5.2 RINAS Concept Verification
The previous section briefly introduced an efficient traffic classification
technique based on imaging approaches to verify that vehicles can be identified rapidly
from images. This section will now verify that the online detection capability introduced
in Chapter 4 is indeed enhanced when inputs are effectively restricted, i.e., to verify that
damage detection results using classified reference pools that would be generated by
131
RINAS are better than those of unclassified reference pools. Verification will start with
the simulated Vibrating Disk Assembly of LANL lab and will then be followed by the Steel
Truss Bridge in Notre Dame’s DYNAMO lab.
5.2.1 Validation Using Vibrating Disk Assembly
The Los Alamos National Laboratory 8DOF assembly will be used for the first
RINAS concept validation, simulating damage by changing the spring between masses 5
and 6 to reduce the spring constant by 14%. The system is excited in both its damaged
and undamaged states at mass 1 using an electro-dynamic shaker with three different
input voltage levels (3V, 4V and 5V). The acceleration responses of all the masses were
recorded for repeated independent trials. A 97.5% one-sided confidence interval is
specified for distinguishing statistically significant damage using the homogeneous DSF
defined in Eq. (4.17).
For the damage detection scheme without RINAS, the unclassified reference
pool consists of 24 independent trials encompassing all 3 input voltage levels (8 trials for
each level) of the undamaged system. This would embody a reference pool for a system
excited by three different “payloads.” This same data will be used to form three
different classified reference pools that contain 8 independent trials from the same
specified input voltage level. Any of these classified pools would embody a potential
restricted reference pool in the RINAS framework. Note that the size of the undamaged
pool is limited by the amount of experimental data archived by LANL for public use.
The responses of 4 independent damaged trials are next selected for each input
voltage level. To demonstrate a generic ambient vibration testing environment (no
132
RINAS methodology applied), these 12 damaged responses will be compared with the
unclassified reference pool with all 24 undamaged time histories at various excitation
levels. This will be termed traditional implementation. Conversely, the RINAS
implementation will only evaluate damaged responses against only the classified
reference pool tied to the input voltage level used to generate the damaged time
history. This will be termed the RINAS implementation.
Recall that each time a DSF value falls outside of this confidence interval,
damage is detected and is signified in Table 5.1 in bold. For the RINAS implementation,
each of the classified reference pools will have its own threshold value to establish
statistical significance. Two major observations can be drawn from these results:
The RINAS implementation is successful in detecting damage, with a perfect detection rate (100%) at all locations, while the performance of the traditional implementation is less successful. For example at Mass 1, the detection rate with the RINAS implementation is 100%, while the traditional implementation has a detection rate of only 33.3%.
Though both damage detection schemes showed a certain capability to find damage. The detection capability is enhanced in the RINAS implementation, as evidenced by the value of the DSFs, which exceed the threshold of statistical significance by a greater margin. This implies that the RINAS implementation allows for a greater sensitivity to minor damage levels.
133
TABLE 5.1
DAMAGE DETECTION RESULTS FOR TRADITIONAL AND RINAS IMPLEMENTATIONS
ONVIBRATING DISK ASSEMBLY
Mass 1 Mass2 Mass3 Mass4
RINAS Traditional RINAS Traditional RINAS Traditional RINAS Traditional
Threshold 1.543(3v) 1.634(4v) 1.561(5v)
2.774 1.589(3v) 1.435(4v) 1.587(5v)
2.312 1.642(3v) 1.567(4v) 1.641(5v)
2.194 1.503(3v) 1.764(4v) 1.503(5v)
2.141
Test 1(3v) 5.983 2.013 3.881 3.576 3.246 3.598 13.743 4.523
Test 2(3v) 6.371 2.438 5.384 4.416 5.076 5.616 15.837 2.749
Test 3(3v) 8.278 4.681 3.835 4.336 4.832 5.346 6.548 3.048
Test 4(3v) 8.039 2.577 4.926 4.783 1.839 2.169 7.643 5.004
Test 5(4v) 4.167 1.965 8.889 4.370 6.092 5.415 4.795 4.985
Test 6(4v) 3.004 3.133 8.633 4.578 4.438 4.889 5.934 4.716
Test 7(4v) 3.930 1.817 7.824 7.905 7.274 8.037 5.004 4.967
Test 8(4v) 7.760 2.280 9.097 6.383 7.443 4.533 6.386 4.970
Test 9(5v) 21.924 2.241 6.204 4.503 5.308 4.860 25.114 6.093
Test 10(5v) 6.122 3.968 3.412 2.726 5.392 5.964 4.449 4.505
Test 11(5v) 5.429 3.076 7.950 6.615 4.763 5.271 3.772 2.771
Test 12(5v) 6.770 1.698 3.340 3.008 7.242 2.409 4.708 4.073
Det Rate 100% 33.3% 100% 100% 100% 91.7% 100% 100%
134
TABLE 5.1 (CONTINUED)
DAMAGE DETECTION RESULTS FOR TRADITIONAL AND RINAS IMPLEMENTATIONS ON
VIBRATING DISK ASSEMBLY
Mass 5 Mass6 Mass7 Mass8
RINAS Traditional RINAS Traditional RINAS Traditional RINAS Traditional
Threshold 1.539(3v) 1.730(4v) 1.536(5v)
2.640 1.571(3v) 1.495(4v) 1.642(5v)
2.331 1.583(3v) 1.602(4v) 1.579(5v)
2.300 1.544(3v) 1.525(4v) 1.541(5v)
2.029
Test 1(3v) 19.042 11.276 4.668 3.329 7.394 6.785 6.240 5.487
Test 2(3v) 13.179 7.289 4.086 1.525 5.161 4.696 5.135 4.769
Test 3(3v) 6.908 5.050 12.482 4.176 9.015 8.303 8.612 7.314
Test 4(3v) 9.890 6.445 14.155 4.444 8.781 8.084 8.745 7.492
Test 5(4v) 7.758 10.137 7.302 4.895 6.372 3.332 26.200 11.074
Test 6(4v) 5.221 6.224 4.892 2.512 7.987 4.336 16.386 5.450
Test 7(4v) 5.370 5.763 4.333 2.667 12.279 6.836 13.500 7.002
Test 8(4v) 8.973 7.208 3.363 1.693 8.552 4.392 12.824 6.659
Test 9(5v) 8.780 7.331 9.274 3.021 6.607 6.049 4.387 4.007
Test 10(5v) 6.075 4.241 6.311 2.356 6.253 5.843 5.109 4.584
Test 11(5v) 12.241 7.714 9.979 5.041 6.750 4.851 3.595 3.197
Test 12(5v) 9.658 6.432 2.706 3.611 2.975 2.531 4.304 4.987
Det Rate 100% 100% 100% 83.3% 100% 100% 100% 100%
5.2.2 Validation Using Steel Truss Bridge Model
To further this validation, the controlled excitation tests via shaker are employed
on the Steel Truss Bridge Model to provide an experimental framework to validate the
RINAS concept. The test bed is the same Steel Truss Bridge Model introduced in Chapter
2. The shaker is placed at the midspan of the bridge deck, and five levels of white noise
inputs are applied by varying the input voltage levels to the shaker. In this study, five
input voltage levels are chosen for the shaker to ensure the bridge model is in its linear
stage during the tests. They are 25mv (LV1), 50mv (LV2), 75mv (LV3), 100mv (LV4) and
135
150mv (LV5). Two damage scenarios are evaluated, as introduced previously in Chapter
2: damage scenario I (minor damage scenario) and damage scenario III (major damage
scenario) are selected for the RINAS proof of concept, to resummarize:
Damage scenario I: replacing 2 structural members
Damage scenario III: replacing 6 structural members
Ten independent tests are conducted at each input voltage level for the
undamaged bridge. For the RINAS implementation, there are 5 independent reference
pools with ten undamaged responses in them. For each damage scenario, five
independent tests are conducted. In the RINAS implementation, the damaged responses
are compared only to the reference pool records generated at same input voltage level.
For the traditional implementation, all damaged responses are compared to the full
unclassified reference pool (all 50 records). Again, each time a DSF value falls outside of
the confidence interval, damage is detected and is signified in Tables 5.2 and 5.3 in bold.
For the RINAS implementation, each classified reference pool has a unique statistical
significance threshold. The node numbers referenced in Tables 5.2 and 5.3 indicate the
locations where responses were measured, shown previously in Figure 2.11.
136
TABLE 5.2
DAMAGE DETECTION RESULTS (DAMAGE SCENARIO I) FOR TRADITIONAL AND RINAS
IMPLEMENTATIONS ON STEEL TRUSS BRIDGE MODEL
Node # 1 2 3
Method RINAS Traditional RINAS Traditional RINAS Traditional
Threshold
1.927 2.044 2.540 2.380 2.148
2.540
2.512 2.399 2.298 2.594 2.612
3.065
2.116 2.090 2.261 2.315 2.164
2.459
`Test 1 (LV1) 1.781 1.078 1.407 1.059 3.862 0.940
Test 2 (LV1) 4.657 3.670 2.901 2.551 11.148 2.917
Test 3 (LV1) 2.829 2.272 7.161 7.912 10.685 2.552
Test 4 (LV1) 5.552 4.516 6.923 7.626 24.259 8.169
Test 5 (LV1) 1.251 1.070 1.514 0.842 7.579 1.661
Test 6 (LV2) 9.786 5.104 9.239 6.173 15.831 10.022
Test 7 (LV2) 6.408 3.475 2.590 3.277 8.676 1.214
Test 8 (LV2) 10.027 4.209 20.085 8.924 19.417 7.855
Test 9 (LV2) 9.110 0.777 1.511 0.703 10.106 0.731
Test 10 (LV2) 10.656 1.013 1.350 0.361 8.964 1.310
Test 11 (LV3) 1.779 1.719 4.234 0.739 3.308 2.103
Test 12 (LV3) 2.379 1.250 0.955 0.349 1.934 1.634
Test 13 (LV3) 9.913 4.209 9.903 8.924 35.235 7.855
Test 14 (LV3) 7.695 4.148 8.227 7.912 34.111 7.582
Test 15 (LV3) 3.095 0.886 0.936 0.342 8.900 1.692
Test 16 (LV4) 5.698 1.783 1.249 0.274 9.583 2.726
Test17 (LV4) 4.557 1.627 17.536 0.468 5.235 1.137
Test 18 (LV4) 4.985 1.572 12.557 4.898 3.827 1.584
Test 19 (LV4) 2.388 1.104 27.048 0.438 4.423 0.865
Test 20 (LV4) 2.039 0.700 1.211 0.379 4.219 1.812
Test 21 (LV5) 4.397 2.065 2.303 0.473 4.686 2.120
Test 22 (LV5) 18.155 4.626 21.172 3.354 12.003 2.787
Test 23 (LV5) 12.544 2.481 18.390 2.896 12.500 1.550
Test 24 (LV5) 21.257 2.510 43.335 6.380 30.103 1.744
Test 25 (LV5) 2.066 1.461 1.495 0.583 1.981 0.954
Det Rate 80% 32% 60% 40% 92% 36%
137
TABLE 5.3
DAMAGE DETECTION RESULTS (DAMAGE SCENARIO III) FOR TRADITIONAL AND RINAS
IMPLEMENTATIONS ON STEEL TRUSS BRIDGE MODEL
Node # 1 2 3
Method RINAS Traditional RINAS Traditional RINAS Traditional
Threshold
1.927 2.044 2.540 2.380 2.148
2.540
2.512 2.399 2.298 2.594 2.612
3.065
2.116 2.090 2.261 2.315 2.164
2.459
Test 1 (LV1) 2.619 3.105 3.321 3.196 5.711 2.170
Test 2 (LV1) 2.795 2.631 3.471 3.304 4.799 1.197
Test 3 (LV1) 1.316 0.979 3.686 3.374 5.841 1.081
Test 4 (LV1) 2.312 1.782 3.320 3.143 9.195 2.633
Test 5 (LV1) 2.791 2.296 2.660 2.441 11.649 3.437
Test 6 (LV2) 10.309 1.569 9.914 3.705 11.347 1.024
Test 7 (LV2) 9.755 0.829 5.085 3.076 11.509 1.067
Test 8 (LV2) 7.662 2.846 3.118 3.588 9.664 1.085
Test 9 (LV2) 9.208 4.108 3.627 2.816 4.332 1.975
Test 10 (LV2) 9.777 3.765 7.201 4.224 10.045 4.310
Test 11 (LV3) 6.377 3.724 4.085 3.752 9.187 2.075
Test 12 (LV3) 9.457 2.659 3.675 4.002 5.796 1.133
Test 13 (LV3) 6.482 3.614 3.007 3.549 15.182 2.665
Test 14 (LV3) 3.655 1.519 2.945 2.109 8.712 2.817
Test 15 (LV3) 4.727 2.079 3.610 2.353 14.392 4.180
Test 16 (LV4) 8.994 2.526 14.962 4.181 9.203 3.924
Test17 (LV4) 8.251 2.783 8.173 3.796 9.713 2.856
Test 18 (LV4) 10.564 4.495 6.778 3.042 11.085 2.213
Test 19 (LV4) 7.713 3.742 6.358 2.820 7.122 2.071
Test 20 (LV4) 5.887 1.856 5.918 2.361 7.944 2.419
Test 21 (LV5) 4.942 2.174 18.320 3.418 3.865 1.496
Test 22 (LV5) 3.893 1.611 18.944 3.302 7.010 2.231
Test 23 (LV5) 24.027 4.287 45.272 6.667 25.641 5.772
Test 24 (LV5) 19.607 3.115 27.856 4.092 23.239 3.520
Test 25 (LV5) 3.295 2.268 10.319 2.621 3.418 4.373
Det Rate 96% 52% 100% 68% 100% 44%
138
The following observations can be drawn from these experimental results:
As expected, the detection rates of the severe damage case (damage scenario III in Table 5.3) are higher than those of minor damage case in Table 5.2 for both schemes.
The detection rates of the RINAS implementation are higher than the traditional implementation. The average detection rate increases are by 41% for damage scenario I and 44% for damage scenario III.
5.3 Summary
The major objective of this chapter was to determine a non-intrusive yet
accurate technology for rapid sensing of traffic loading conditions approaching the
bridge. This technology requires various image processing algorithms, which were
overviewed in this chapter, and must be executed in near real-time to activate the
sensor network, including a new contour algorithm to determine the class of vehicle in
the image. Subsequently, the author demonstrated the advantage of restricted input
network activation on damage detection rates using two test beds (simulated and
experimental). The results verified substantial increases in detection rates when RINAS
was invoked. Moreover, because of the dramatic reductions in the size of the reference
pool, the RINAS concept has additional computational savings beneficial to the wireless
network overall. Figure 5.9 is reintroduced to chronicle the contributions of this chapter
to the overall assessment framework.
139
Figure 5.6: Overview of key features of proposed wireless sensor network for structural health monitoring with addition of new
benefits introduced in Chapter 5.
BENEFIT APPROACH STAGE
DATA ACQUISITION
DATA REDUCTION
DETECTION
LOCALIZATION
Multi-scale WSN, Restricted Event Triggering
Event synchronized, Minimized reference pool,
low power, scalable
Bivariate autogressive reference database
Data Driven DSF
More sensitive to damage, easily embedded
More reliable, computational demand
reduced
140
CHAPTER 6:
OFFLINE DAMAGE LOCALIZATION
Thus far, a novel heterogeneous multi-level wireless sensor network for
structural health monitoring has been introduced (Ch. 3) to overcome some of the
limitations posed by limited power and computational resources and synchronization.
Within this network, a new Bivariate Regressive Adaptive INdex (BRAIN) for damage
detection was introduced to provide efficient and reliable detection of even minor
damage under ambient vibration (Ch. 4). This detection capability is enhanced by
constraining the reference pool through a restricted activation scheme based on image
recognition (Ch. 5), which has a secondary benefit of further reducing energy
consumption. In total, these permit a decentralized and scalable approach to online
damage detection.
Although the BRAIN concept in Chapter 4 showed some ability to localize
damage and even quantify extent, these assessments must be executed with a high
degree of reliability as they are reported to end-users and if serious enough, will
warrant human intervention for more detailed inspection and non-destructive
evaluation on-site. As the intent of automated damage detection is to relieve the need
for human intervention in initial assessments, it is important that humans are only
notified when damage extent and location has been confirmed to avoid eroding their
141
confidence in the technology with false positives. This more refined assessment will be
the focus of this chapter. Again operating under the assumption that the initial
detection of damage by the schemes in Chapter 4 is executed when the damage is in its
minor stages, the final damage localization need not be real-time and can be finished
offline, so computational and power constraints are lifted, making a wider cross section
of signal processing and analysis tools viable.
Research on this type of vibration-based damage localization has been
expanding rapidly over the last decade, generally falling into two classes: Finite Element
Model Refinement Algorithms (FEMRA) (Chung, et al. 2003) and Theoretical Modal
Parameter Indicators (TMPI) (Li, et al. 2007). Unfortunately, many of these damage
localization methodologies require physical properties (like theoretical frequencies,
mode shapes) or even calibrated Finite Element Models and direct measurement of
input excitation for implementation; however, as this research has adopted the less
intrusive ambient vibration monitoring, all methods requiring measured input are
precluded.
As Chapter 4 demonstrated, by virtue of the higher mode information contained
in the time series coefficients, this approach had superior sensitivity to damage (Su et al.,
2007), which was further enhanced by the incorporation of data from multiple sensing
elements (strain and acceleration). Further, due to their compact representation, these
coefficients are easily stored and manipulated. As a result, this chapter will develop a
new damage localization method that intelligently integrates the information from
these sensors, introducing a new evidence theory approach to localize the damage. As
142
the data fusion method, evidence theory combines different information sources to
improve decision making and has been popularized in many fields, such as in medicine,
robotics, intelligent vehicles, and industrial engineering. Recently, information fusion
techniques have been extended to identify structural damage locations based on mode
shape or natural frequency data; its application to time series coefficients will now be
explored. Proof-of-concept is achieved in this chapter using the simulated thin
cantilever beam test bed introduced in Chapter 2, for scenarios with single and multiple
damage sites.
6.1 Revisiting Damage Localization using AR Model Coefficients
Before introducing the new damage localization method, it is important first to
emphasize the advantages of basing this approach on the BRAIN methodology
introduced in Chapter 4. First, unlike the traditional Multiple Damage Localization
Assurance Criterion (MDLAC) (Yan, et al. 2007) or Frequency Change Damage Detection
Method (FCDDM) (Guo and Zhang 2006), no theoretical dynamic parameters are
needed; the only data source for this damage localization method is ambient vibrations
from the target structure in its damaged and undamaged conditions under unrecorded
excitations. Secondly, the whole damage localization scheme is working in the time
domain, eliminating the signal processing concerns associated with frequency domain
operations. Finally, since the initial assessments are done in a decentralized framework,
in the event of confirmed damage detection, the local damaged time histories would be
fit by the required AR models locally and the coefficients of these models would be
143
transmitted to the gateway to better manage power and bandwidth constraints on the
network. Since the subsequent analysis is done in an offline manner, this transmission
does not have to occur simultaneously from all the nodes, reducing the demands on
bandwidth. This also reduces computational burden at the M-node (by distributing
more of the initial computations) and reduces the demands (bandwidth and power) on
the wireless transmissions within the network.
Recall the introduction of the AR model for representing acceleration time
histories in Equation (4.1). Validations in Chapter 4 demonstrated that when damage
occurs, the coefficients of this AR model increase with the proximity to the damage
location. As such, the fluctuations in each AR coefficients can serve as a damage
localization damage localization index (DLI), as previously introduced in Equation (4.19).
Unfortunately, the use of such time-series damage localization approaches is completely
reliant on the underlying model used to represent the time series and the selected
model order. The optimal model order is usually obtained using the Akaike Information
Criteria (AIC). The AIC consists of two terms: the first is a log-likelihood function and the
second is a penalty function for the number of terms in the AR model. According to the
bandwidth of the process and sampling rate, a certain range of model orders are
appropriate for analysis. Model orders outside this range can cause the problem of over
or under fitting the fundamental mode, as demonstrated in Figure 6.1.
144
Figure 6.1: Examples of over fitting, optimal fitting and under fitting (left to right).
The optimal order by AIC is generally driven by the dominant, generally
fundamental, modes, compromising the fit to higher modes, which tend to show greater
sensitivity to damage, emphasizing less the quality of fit in the modes most relevant to
damage detection. To solve this problem, the sensitivity to AR model order is explored
by evaluating multiple model orders in the evidence theory framework. As a result,
when commissioning this network on a bridge, the training phase will require the
undamaged time histories to be fit by all necessary AR models (with varying order) and
the coefficients stored at the M-node. Then only the statistics associated with the model
order used in the online damage assessment discussed in Chapter 4 need to be
uploaded to the respective -nets to locally calculate the DSF during the operation of
the system.
0 2 4 6 8-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1Over Fitting
Time
0 2 4 6 8-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1Good Fitting
Time
0 2 4 6 8-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1Under Fitting
Time
145
6.2 Introduction to Evidence Theory
Information fusion is the merging of information from disparate sources to
achieve improved accuracies and more specific inferences than could be achieved by the
use of a single source alone. Dempster-Shafer evidence theory is an important data
fusion theory, briefly summarized as follows. For a finite set of mutually exclusive and
exhaustive propositions , sometimes referred to as a frame of discernment (FOD), a
power set 2 is the set of all the subsets of including itself and a null set, . Each
subset is called a focal element. Evidence theory allows one to attach a probability value
between [0, 1] to any member of the power set of the frame of discernment. The value
0 indicates no belief in a proposition, the value 1 indicates total belief, and any values
between these two limits indicate partial beliefs. A portion of belief committed to one
focal element is also committed to any other implied focal elements and cannot be
further subdivided among the subsets.
Evidence theory allows mass or basic probability assignment (BPA) to individual
propositions and also to any subsets of the power set. The mass of the empty set is zero
and the masses of the remaining members of the power set add up to a total of 1, as
expressed in Equation (6.1):
0)( m , 1)(0 Am
and 1)()( A
Amm (6.1)
If the probability number for only a partial set of resources is unknown, then the
remaining complementary probability number is assigned to the ignorance m(), which
is the subset of all unknowns. From the mass assignments, two limited bounds of a
146
probability interval can be defined. The lower bound )(Abel for a set A is defined as the
sum of all the masses of subsets of the set of interest and the upper bound )(Apl is the
sum of all the masses of the sets B that intersect the set of interest A:
AB
BmAbel )()(
(6.2)
AB
BmAbelApl )()(1)(
(6.3)
The Dempster-Shafer rule strongly emphasizes the agreement between multiple
sources and ignores all the conflicting evidence through a normalization factor. The joint
mass is calculated from the two sets of masses m1 and m2 in the following manner:
CB
ACB
CmBm
CmBm
Am)()(1
)()(
)(21
21
2,1
(6.4)
The numerator represents the accumulated evidence for the sets B and C, which
supports the hypothesis A, and the denominator sum quantifies the amount of conflict
between the two sets. Equation (6.4) illustrates the combination between two
information sources. For cumulative data fusion with more information sources, the
procedure follows a tree structure and is illustrated in the following flowchart (Fig. 6.2).
147
Figure 6.2: The tree structure of Dempster-Shafer evidence theory data fusion.
It is commonly accepted that multiple evidences from different sources are not
equally important when they are combined according to Dempster-Shafer theory, but it
is not considered initially in this study because damage localization is treated as a pure
blind test. To guarantee the weights of all information sources are same value, the total
number of information sources must be a power of two.
6.3 Application of Evidence Theory and Proof-of-Concept
To demonstrate the role of evidence theory in localizing damage, the simulated
thin cantilever beam model introduced in Section 2.2.2 will be employed, again
subjected to Gaussian white noise input at the free end. A collection of strain time
histories are repeatedly simulated at the end of each element of the undamaged beam,
m()n
Fusion
m()n
m()n
Fusion
m()n
m()n+1
m()
n+1
Fusion
m()n+2
148
to form the reference database that will be used in subsequent blind tests, and for each
of the damage scenarios shown in Figure 2.20.
Suppose there are n elements in the structure, which are treated as subsets in
the Dempster-Shafer evidence theory. Damaged elements are preliminary localized
from m AR models of different model order, which are treated as information sources.
The following notation is then adopted for the damage localization index for the ith
element from jth source: mjniDLI j
i ,...,2,1;,...,2,1, .
The basic probability assignment for Dempster-Shafer evidence theory is
calculated through the belief measure of jth information source for the ith element:
n
i
j
i
j
ii
k
j
DLI
DLIem
1
)( (6.5)
k indicates the stage of data fusion discussed shortly.
6.3.1 Proof-of-Concept for Single Damage Site
Validation of the evidence theory approach is first achieved considering a single
damage site as specified by Damage Cases 1-4, summarized previously in Table 2.7.
Eight AR models are constructed from the time histories associated with each damage
case, with orders 8, 12, 16, 20, 24, 28, 32, and 36. To enhance the robustness of the
methodology, the average AR coefficients of 50 undamaged time histories is adopted as
unnaun ,, 21 and the average AR coefficients of 5 damaged time histories is
adopted as na ,, 21 in the damage quantification index in Equation (4.19). For
elements far away from damage, where the effect of damage is small, the two vectors
149
should be highly correlated, and DLI should be near unity. For elements progressively
closer to the damage site, the correlation should reduce and the DLI should tend toward
zero.
The procedure, depicted in Figure 6.3, can thus be summarized as follows:
1. Time histories at each location are fit with eight different AR models with varying orders (treated as eight “sources”) 3
2. The AR coefficients from these sources at each location are averaged over five damaged trials4
3. The DLI at each location is calculated from these averaged quantities for each of the eight sources according to Equation (4.19), where superscript will denote the element location and subscript will denote the source
4. The basic probability assignment for Dempster-Shafer evidence theory
)( i
k
j em is calculated according to Equation (6.5)
After each data fusion steps, the information source number will be reset; )(1
ij em and
)(2
ij em
do not refer to the same data source
3 The same action will have been previously executed during the training period
on the, in this case, fifty undamaged time histories at each location that form the reference pool
4 The same action will have been previously executed during the training period
on the reference pool AR coefficients for each location averaging over, in this case, fifty undamaged time histories
150
Figure 6.3: Schematic representation of Evidence Theory applied to single site damage detection in thin beam model.
)( 11 em )( 12 em )( 13 em )( 14 em )( 15 em )( 16 em )( 17 em )( 18 em
Fusion
)( 1
2
1 em
20
1
)(
i
j
i
j
i
ij
DLI
DLIem
)(
)(
)(
)(
)(
)(
)(
)(
)(
)(
58
51
48
41
38
31
28
21
18
11
em
em
em
em
em
em
em
em
em
em
)(
)(
)(
)(
)(
)(
)(
)(
)(
)(
208
201
198
191
188
181
178
171
168
161
em
em
em
em
em
em
em
em
em
em
Fusion Fusion Fusion
)( 1
1
1 em )( 1
1
2 em )( 1
1
3 em )( 1
1
4 em
Fusion Fusion
)( 1
2
2 em
Fusion
)( 1
3
1 em
Befor
e Fusion
1st
Level
2nd
Level
3rd
Level
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
19 20
. . .
151
The localization results ( )( 1emk
j to )( 20em k
j
) for single-site damage are presented in
Figures 6.4-6.7, where smaller probability assignment values are indicative of damage.
There are four images in each figure. The first one labeled “averaged result” is the
averaged probability assignments from all 8 information sources before data fusion. The
other 3 figures indicate the DLIs after each level of data fusion in Figure 6.3. From the
single-site damage cases, several conclusions can be summarized:
The proposed framework can localize damage successfully for all single-site damage cases, since the probability assignment value associated with the known damaged element in each case is smaller than the probability assignments associated with other elements.
As more information sources and levels of evidence theory are explored, the reduction in the probability assignment associated with the damaged element becomes more pronounced. In all cases, results of the 3rd level of evidence theory are the best and the results before any data fusion (averaged result) are the worst.
As expected, the larger the effective stiffness loss, the more pronounced the reduction in the probability assignment. For the same amount of cross sectional loss, the effective stiffness lost increases with proximity to the fixed end. As such, probability assignment reduction is expected to be most pronounced in damage case 2 and progressively less pronounced in damage cases 3 and 4.
152
Figure 6.4: Evidence Theory localization results for damage case 1 (actual damage location at element 4).
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Average Result
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 1
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 2
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 3
Element Number
DLI
153
Figure 6.5: Evidence Theory localization results for damage case 2 (actual damage location at element 8).
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Average Result
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 1
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 2
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 3
Element Number
DLI
154
Figure 6.6: Evidence Theory localization results for damage case 3 (actual damage location at element 12).
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Average Result
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 1
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 2
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 3
Element Number
DLI
155
Figure 6.7: Evidence Theory Localization results for damage case 4 (actual damage location at element 16).
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Average Result
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 1
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 2
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 3
Element Number
DLI
156
6.3.2 Proof-of-Concept for Multiple Damage Sites
The validation extends now to damage imparted at multiple sites, using the
same damage localization procedures and damage cases 5 and 6 summarized previously
in Table 2.7. Results are presented in Figures 6.8 and 6.9, in the same format as the
single damage site results.
157
Figure 6.8: Evidence Theory localization results for damage case 5 (actual damage locations at elements 4 and 13).
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Average Result
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 1
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 2
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 3
Element Number
DLI
158
Figure 6.9: Evidence theory localization results for damage case 6 (actual damage locations at elements 8 and 13).
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Average Result
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 1
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 2
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Level 3
Element Number
DLI
159
Similar conclusions can be inferred from these results, with the probability
assignment taking on reduced values for the known damaged elements; however the
reductions are not as pronounced as they were in the single damage site results. This is
due to the fact that more undamaged elements are affected when damage occurs at
multiple locations along the beam. This challenge is demonstrated by the 13th element
in damage case 5, which has a smaller effective stiffness loss than element 4 even
though the same cross sectional losses are applied. Additionally, when the damage is
larger (greater cross sectional area removed) and the damage elements are closer
together, the probability assignment values for other nearby elements are adversely
affected, e.g., elements near the free end of the bar (elements 18 to 20 in damage case
6). One solution to this problem is to classify the element members and treat them
differently according to their response intensity and location sensitivity. Currently, most
research treats all the structural members equally in the damage localization process,
despite the fact that response level varies spatially along the member. For example, the
cantilever beam used in this study, acceleration responses from free end are much
larger than those from fixed end. Similarly, elements closer to the fixed end, when
damaged, lead to greater losses in stiffness. Furthermore, all information sources were
considered equally. The question of how to distinguish useful information sources from
others needs therefore to be explored. The following section represents one of the
earliest attempts to conduct such assessment of the information sources.
160
6.4 Weighted Balance Evidence Theory
Multiple evidences from different sources with varying importance or reliability
should not be treated with equal importance when they are combined, but it is seldom
considered in the traditional Dempster–Shafer Theory. Recent studies have revised the
combination rule with two parameters: the correlation coefficient between evidences
and the reliability of the evidence (Wang 2008), while others have proposed a multi-
damage identification index to compare the identification results under different
weighting coefficients (Guo, et al. 2004). However, these adaptations to Evidence
Theory in damage detection require a calibrated FEM of the undamaged structure.
Therefore, this research now presents a weighted balance evidence theory approach
that does not require FEM and solely operates on ambient vibration responses.
To do so, the tree structure in Figure 6.2 is modified by including a weighting
coefficient wi to the mass of each source. The modified weighted tree structure is
presented in Figure 6.10.
161
Figure 6.10: The weighted tree structure of Dempster-Shafer evidence theory data fusion.
Then a set of rulescan be defined according to the assumed conditions
associated with the damage. The reliability of different information sources can then be
determined by the number of rules the information source satisfies. If an information
source A satisfies more rules than information source B, the reliability level is assumed
to be higher than that of information source B.
The values of weighting coefficient wi can be determined by the following rules:
wi = 1
Set the weighting coefficient of an information source that doesn’t satisfy any conditions as a basic value .
The weighting coefficients of higher reliability sources can be defined as
, where n is the number of conditions that the information source satisfies. For example, if an information source satisfied 2 conditions then the weighted coefficient is
;
Returning now to the cantilever beam damage localization problem, the newly
proposed Weighted Evidence Theory will be employed in an effort to refine the single-
m()n×w1
Fusion
m() n
×w2 m()
n×w3
Fusion
m() n
×w4
m()n+1 m()
n+1
Fusion
m()n+2
162
site damage cases. Because it is a pure black box damage localization problem, known
circumstances surrounding the damage are very limited, though it is known that there is
at least one damaged element among all the beam elements. So the condition can be
defined as one damage localization index is significantly smaller than others. An
example of how that rule can be expressed: if one and just one element’s probability
assignment is 10% less than the average value of all 20 probability assignments, then
the weight coefficient wi of the information source is 2 . Otherwise the weight
coefficient wi is 1 . When the number of data sources that can fit the rule is
determined, the value of can be calculated. For example, if five out of 8 data sources
satisfy the rule, the value is
. The weighting coefficient for a data
source that satisfies this rule is and the weighting
coefficient value for other data sources is . This rule will be used in the
following example.
Figure 6.11 shows the damage localization results of unweighted (Dempster) and
Weighted Evidence Theory of data fusion for damage case 1 and damage case 4. Those 2
damage cases are chosen because the localization results were not as obvious as others
(see Figures 6.4 and 6.7). In Figure 6.11, plots in the first row and second row indicate
the results of damage case 1 and damage case 4, respectively. The first column depicts
the unweighted (Dempster) results and the second column presents the weighted
results.
163
Figure 6.11: Damage localization results using unweighted/Dempster (Column 1) and weighted (Column 2) evidence theory. First row is results for damage case 1 and second row is results for damage case 4.
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Dempster
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Weighted
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Dempster
Element Number
DLI
0 2 4 6 8 10 12 14 16 18 200
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05Weighted
Element Number
DLI
164
The results in Figure 6.11 show that after three levels of data fusion, the
probability assignment values of damaged elements for weighted evidence theory are
reduced by 18%, on average, compared to those using unweighted (Dempster) evidence
theory, thus demonstrating the improvement in localization offered by weighted
evidence theory. Although only one condition could be defined a priori, it is suspected
that the performance of weighted evidence theory will improve as more conditions
(rules) are formulated.
6.5 Summary
This chapter firstly described the importance of accurate offline damage
localization in the structural health monitoring process and the limitations of current
localization methods ill-suited to the present application in ambient vibration
monitoring in the absence of a calibrated finite element model. To accommodate the
ambient vibration response localization, a new damage localization index (DLI),
previously introduced in Chapter 4 and based on the correlation of time series
coefficients, is employed. To enhance its accuracy, Dempster-Shafer evidence theory is
then invoked for data fusion. Results show that the proposed damage localization
scheme can successfully find the damage locations for a simulated cantilever beam. The
chapter concluded with a modification to evidence theory in the form of weightings.
This approach can further enhance localization capabilities if as little as one condition
can be imposed a priori on the damage circumstances. Figure 6.12 now presents the
completed hierarchy of approaches used to realize the wireless sensor network concept.
165
Figure 6.12: Overview of key features of proposed wireless sensor network for structural health monitoring, with addition of new
benefits introduced in Chapter 6.
STAGE BENEFITS APPROACH
DATA ACQUSITION
DATA REDUCTION
DETECTION
LOCALIZATION
Multi-scale WSN, Restricted Event Triggering
Event synchronized, minimized reference pool,
Low power, scalable
Bivariate Autoregressive, Reference Database
Data-Driven DSF
DLI with Evidence Theory
More sensitive to damage, easily embedded
Enhanced localization capability
More reliable, computational demand
reduced
166
CHAPTER 7:
CONCLUSIONS AND FUTURE DIRECTIONS
Civil Infrastructure worldwide is suffering from unseen levels of damage that, if
arrested in their infancy, can be repaired at lesser expense and more importantly avoid
the potential for collapse. Unfortunately, the current manual inspection paradigm is ill
suited for such a pro-active approach to maintaining Civil Infrastructure. This
dissertation focused on the use of a wireless structural health monitoring (SHM)
framework to assess a structure over its lifetime so that low levels of damage can be
detected continuously and automatically from only the measured responses and when
tied to a life-cycle assessment framework, can enable prioritization of rehabilitation and
maintenance efforts. This detection and localization of damage is accomplished by
monitoring the coefficients of time series regressive models..
7.1 Contributions of This Work
This work proposed a number of the new ideas and techniques to respond to the
issues confronted as wireless sensor networks are applied to Structural Health
Monitoring. These contributions and the benefits they provide were depicted in Figure
6.12. The specific contributions of this work are summarized herein.
167
7.1.1 Multi-scale Wireless Sensor Network
In this research the traditional hub and spoke network is recast in a multi-scale
framework to satisfy key performance metrics such as maximizing network lifetime,
enhancing reliability, and facilitating scalability. The multi-scale WSN introduced by this
research in Chapter 3 divides the structure into a series of meso-networks (m-nets).
Within this m-net, there are wireless motes with on-board accelerometers tethered to
multiple distributed strain gauges to monitor behavior at critical locations. Each
accelerometer and their supporting strain gauges form a micro-network (-net), where
the initial diagnosis of damage is conducted. This decentralized approach not only has
power conservation benefits, but also escapes the need for strict synchronization and
provides resistance to latency that a centralized approach to system identification
would require. Thus lengthy time series are never transmitted wirelessly, and the only
information shared outside of the -net is the binary damage diagnosis and/or the
estimated damage sensitive feature (DSF), which is a customized metric for rating
damage.
7.1.2 Restricted Input Network Activation Scheme (RINAS)
Ambient vibration testing or operational monitoring is generally preferred over
forced vibration testing as it is more economical and less obtrusive. However, the low
signal to noise ratio, the difficulty in exciting higher modes, and the lack of measured
input significantly complicates the ensuing system identification. While the input cannot
168
be controlled or explicitly measured in ambient vibration testing, this research sought to
instead improve the performance of system identification and reduce the size of
reference databases through the introduction in Chapter 3 of a Restricted Input
Network Activation Scheme (RINAS). Through RINAS, the system was triggered only by
the detection of particular traffic and environmental conditions. The regulation of input
conditions in this way implies that the reference pool need only include data on the
response of the bridge in its healthy or initial condition under this loading scenario. In
addition to the computational savings of this event triggering approach, this network
design also reduces the power demands on the system and extends network lifetime, as
sensors only operate under these specified conditions. Most importantly, damage
detection capability and reliability are increase by applying the Restricted Input Network
Activation Scheme. Chapter 5 introduced a new image processing technique that allows
classification of vehicles based on contour areas so that unobtrusive video sensors can
be used to classify oncoming traffic and activate the network when a solitary semi-
trailer approaches the bridge.
7.1.3 Data Reduction Using Time Series Models
Chapter 4 of this dissertation provided one of the most systematic explorations
and verifications of time series model coefficients for damage detection and localization
because of the following properties:
Time series models provide a way to compactly and accurately represent signals.
169
Their coefficients have shown sufficient sensitivity to damage to enable not only detection but also localization.
The coefficients are easy to calculate, without the need for computationally intensive transforms making them well-suited for embedment in wireless platforms.
The coefficients contain multi-mode information, and as higher modes tend to show greater sensitivity to damage, this information can be conveyed through a single time series coefficient.
7.1.4 Data-driven Bivariate Regressive Adaptive Index (BRAIN)
More importantly, this research was the first to employ heterogeneous sensing
(local acceleration and strain data) for the detection of damage in a decentralized
fashion within wireless sensor networks. Through the introduction of a Bivariate
Regressive Adaptive INdex (BRAIN), which was extensively vetted in Chapter 4 against
simulated and experimental test beds, enhanced damage detection was achieved in
comparision to that observed when solely acceleration data is employed. Another
unique feature of BRAIN was its novel introduction of a dynamic or data-driven damage-
sensitive feature, which extracts the most responsive model coefficients automatically
to enhance detection capability. Again extensive validations in Chapter 4 demonstrated
that this adaptive capability yielded more reliable outcomes than damage sensitive
features that a priori select the targeted model coefficients and was suitable to a wide
range of applications. These explorations affirmed the two major hypotheses posed at
the beginning of this dissertation:
170
Data-driven or dynamic DSFs are more robust and reliable than their static
counterparts in homogeneous sensor networks.
Heterogeneous DSFs are more robust and reliable than their homogenous
counterparts.
Additionally, Chapter 4 demonstrated the enhancement in detection reliability
possible when data fusion within the network is employed, even through basic voting
schemes, eliminating false positives that often occur when such sensitive detection
approaches are applied.
7.1.5 Novel Damage Localization Index and Evidence Theory
After introducing the wireless sensor network concept and developing methods
to trigger the network (RINAS) and classify damage in a decentralized format in real time
(BRAIN), the remaining task was to accurately identify the location of the damage to
notify end users/operators so a more detailed local inspection can be conducted with
non-destructive evaluation tools. Unfortunately, most existing damage localization
methodologies require the structural physical properties (like theoretical frequencies,
mode shapes) and direct measurement of input excitation for implementation.
However, in most cases, there is no easy way to measure those parameters. In order to
obtain more accurate and applicable methodology, this research developed a new
damage localization method to intelligently integrate the information from ambient
vibration signals from different types of sensors. A new damage localization index (DLI)
was proposed by monitoring the fluctuation of each AR coefficient. Since time series
models can be easily calculated, the approach provided a quick and convenient damage
171
localization technique. To achieve improved accuracy, evidence theory was used as data
fusion method for the first time in conjunction with such time series models. With the
additional incorporation of a weighted balancing step, the approach was shown to
successfully localize minor levels of damage in both simulated and experimental test
beds.
7.2 Future Directions
In general, all the efforts of this work were focused on the theories and
technologies in sensing, system identification and data analysis that are the
cornerstones of advanced health monitoring. However, the true nature of this work is to
provide the opportunity to extend many of the frameworks to realize efficient and
reliable health monitoring on real civil infrastructures. Some of the future work that
should be conducted based on the findings of this dissertation is now discussed.
7.2.1 Prototype Hardware
It was initially intended in this research to validate the proposed algorithms
within the actual wireless hardware; however, as the larger project this research was
tied to was not funded, the development of that hardware by commercial partners at
EmNet LLC and Columbia Research Labs could not be undertaken in the required time
frame. However, recently the prototype has been developed based off current
hardware from these commercial partners. The design is based upon the force balance
accelerometer line of Columbia Research Laboratories (SA-307 series), offering high
sensitivity, low noise triaxial sensing down to 0 Hz. Table 7.1 compares the specifications
172
of SA-307 series and accelerometers used in other full-scale monitoring projects. Note
that the accelerometers for this study are much more sensitive than others in the table
and have now been integrated with the wireless platform supplied by EmNet LLC with
appropriate signal conditioning, anti-aliasing filtering, local processing power and
wireless transmission capabilities. The final prototype is shown in Figure 7.1.
TABLE 7.1
ACCELEROMETER COMPARISON FOR DIFFERENT MONITORING PROJECTS
Sensor Manufacturer Project Ranges Sensitivity Excitation
SA-307 Columbia This Study ±0.5~±2g 7500mV/
g ±15VDC
393C PCB Henry Hudson
Bridge (NY) ±2.5g
1000mV/g
±18VDC
333B55
PCB I-40 Bridge
Albuquerque (NM) ±5.0g
1000mV/g
±18VDC
1221 Silicon Design Pedestrian bridge over passing I-80
Berkeley, CA ±2.0g
500~1000mV/g
±5 VDC
173
Figure 7.1: Prototype wireless unit to support -net for structural health monitoring (left) and deployed gateway node or M-node
(right) (Source: EmNet LLC).
The hardware is designed to support externally tethered sensors fabricated by
Columbia Research Labs through the access point at the base of the unit in Figure 7.1.
For the purposes of strain measurement, these will be the DT-3716 self-temperature
compensating strain gauges for straight mounting surfaces. Figure 7.2 shows the sensor,
with specifications in Table 7.1.
174
Figure 7.2: DT-3716 strain gauge: photo and schematic side and plan views (Source: Columbia Research Labs).
TABLE 7.2
DT-3716 STRAIN GAUGE SPECIFICATIONS
Manufacturer Linearity Range Sensitivity Excitation
Columbia 0.5% -3500 ~ +5000
μ
1.025 (±1%)
mV/V/1000μ ±10VDC
The sensors are supported by EmNet LLC wireless platform that powers the
sensors, conditions their signals, performs A/D conversion, processes the data, and
transmits the DSF. The design is based on EmNet’s Chasqui Inode or instrumentation
node, which has been used extensively for monitoring in-situ conditions in the City of
South Bend under combined sewer overflow events (Montestruque and Lemmon 2008).
The M-Node, which controls the higher level network decision making and
computation, including the interface with the RINAS sensing hardware, will be served by
175
EmNet’s gateway node or Gnode (shown in Fig. 7.1), which has greater computational
capabilities and provides secured end user access to damage reports. Thus the network
M-Node will not only collect information from Inodes serving as the cluster heads of the
m-net, but will also perform all RINAS algorithm functions described in Chapter 5 and
additional offline assessment described in Chapter 6. The M-Node can use cellular
connections, WiFi, wired ethernet, and other means to transmit information back to the
end user. Table 7.3 shows the specifications of the Gnode used in South Bend under
Combined Sewer Overflow events (Montestruque and Lemmon 2008)).
The gateway will interface with two additional sensors to allow identification of
environmental and loading conditions for RINAS. For environmental monitoring, it is
recommended to use the Vaisala Weather Transmitter WXT510 (Fig. 7.3). This compact
sensor has no moving parts and determines essential weather parameters like wind
speed and direction, liquid precipitation, barometric pressure, temperature and relative
humidity.
176
Figure 7.3: Vaisala Weather Transmitter WXT510, interfaced with EmNet gateway in field deployment in Chicago (left) with
elevation and plan view schematics (right) (Source: Vaisala Inc.).
177
TABLE 7.3
GNODE SPECIFICATIONS FOR SEWER OVERFLOW MONITORING (SOURCE: EMNET LLC)
Model Type: CSOnet™ GNode Telemetry System
Enclosure Size: 7 ½” x 9 ¾” x 4” Type: NEMA 4X
Communication 1. Sensor Network Side Frequency: 902 - 928 MHz Spread Spectrum: Frequency Hopping Spread Spectrum Modulation: Frequency Shift Keying Network Topologies: Mesh Ad-Hoc, Point to Point Security: 256 bit Advanced Encryption System Certification: FCC Part 15.247 OUR-9XTEND 2. Wide Area Network Side Options Ethernet: 10/100 Ethernet Port (Power over Ethernet available) Cellular: GSM (Global System for Mobile Com) Landline : 33.6K v.34 modem
Inputs Remote: Up to 10 Inodes 1. Analog Max Number: 4 analog sensors Range: Adj., voltage up to 24V, current up to 1A Excitation: Pulsed, 5VDC / 12VDC, adj. duration Resolution : 0.1% of full range 2. Digital Max Number: 1 digital sensor Format: RS-232 Modbus or similar Excitation : Pulsed, 5VDC / 12VDC, adj. duration Field length: 16 bits per field
Outputs Analog: 2 ports 4-to-20mA (12 bit precision) Digital: RS-232 Modbus or similar
Power Voltage: 110VAC Power: 10Watts Solar : Solar Panel optional
Electronics Processor: AMD Elan 520 (x86 compatible) Memory : 256 Mbytes Flash Operating System : Embedded Linux
For traffic monitoring, the SKY5303V Weatherproof Varifocal Bullet closed circuit
television (CCTV) camera is recommended. This camera can reduce the effect of bad
178
weather conditions like rain, snow and fog, thereby minimizing the amount of corrective
action required in the algorithms introduced in Chapter 5. The camera can operate in
infrared modes to continue acquiring images even in the absence of daylight. Figure 7.4
shows the SKY5303V camera and Table 7.4 provides its specifications.
Figure 7.4: Photo of SKY5303V CCTV Camera (Source: Skyway Security).
TABLE 7.4
SKY5303V CCTV CAMERA SPECIFICATIONS.
Manufacturer Resolution Output Lens Weatherproofing
Skyway 480 TV
lines 1.0 Vp-p, 75
ohm
4 to 9mm Varifocal
Lens IP 66 NEMA Rating
While the video identification using the hardware shown in Figure 7.4 offers an
excellent tool for vehicle classification, visual classification does not provide information
regarding weight: a large semi-trailer may have a payload that produces either larger or
179
smaller forces than the reference database suggests. Thus visual classification can still
offer significant uncertainty in the input conditions. While information on exact weight
is desirable, the placement of weigh-in-motion sensors generally requires the road
surface to be compromised. Thus one area of future work would be to explore a novel
non-destructive weighing technology that utilizes the existing highway weighing system,
an innovative data transmission strategy and a unique social science experiment.
The US Interstate system employs numerous weigh stations to minimize
overloaded vehicles. It therefore would be interesting to couple this mandatory
weighing process with Radio-frequency identification (RFID) tagging to transmit this
information to the gateway sensors operating as M-nodes at an instrumented bridge
down traffic of the weigh station.
Figure 7.5: Diagram of Passive RFID tag components.
Radio-frequency identification is an automatic identification method, relying on
the storage and remote retrieval of data using RFID tags. RFID is considered one of the
TAG
Antenna
RFID Reader
Computer Database
Radio frequency signal
Communication by internet or satellite
180
most fundamental technologies to enable wireless data transmission, used in a variety
of industries for product identification and tracking (Want 2006). Unlike other
technologies like bar-coding, RFID tags do not require line of the sight for data retrieval.
Most RFID tags contain at least two parts: an integrated circuit for a variety of functions,
including storing and processing information and modulating and demodulating a radio
frequency (RF) signal. The second is an antenna for receiving and transmitting the RF
signal. RFID tags come in three general varieties: passive, active, or semi-passive (also
known as battery-assisted). Passive tags require no internal power source, thus being
pure passive devices (they are only active when a reader is nearby to power them),
whereas semi-passive and active tags require a power source, usually a small battery
(Want 2006). In a system using passive tags, the tag is energized by the RF field from the
reader and transmits its ID to the reader. Other data transmission depends on the
protocol between reader and tag. Most passive tags signal by backscattering the carrier
wave from the reader. This means that the antenna has to be designed both to collect
power from the incoming signal and also to transmit the outbound backscatter signal.
Passive tags are most common used because they are cheap, can last indefinitely long as
there is no need for power supply, and they are small size what allows them easy to
integrate almost in every environment including cards and stickers. Figure 7.5 shows the
diagram of passive RFID tag that would be proposed for future research related to this
disseratation.
This approach would require RFID tags to be placed on the side of the trailer at
weigh stations. The RFID tag will be programmed with a unique vehicle identification
181
number. This identification number and time and date of the deployment will be logged
via Internet into a database along with the weight and type of vehicle (single, double or
triple trailer). This type of service would be implemented at the nearest weigh station
ahead of a monitored bridge. At the bridge, an RFID reader scans the RFID tags as they
pass by, the identification number is transmitted to the M-node, compared against the
web database and the vehicle information is retrieved. A decision is then made by the
M-node to determine whether the network should be activated. Figure 7.6 shows the
various stages in the RFID RINAS concept. Note that passive RFID tag may be read at a
range of over ten feet, though depending greatly on the operational frequency and
environment and most reading devices can scan tags moving as fast as 150 mph.
Obliviously a number of conditions must be satisfied for this new concept to be
realized. First, it will only be viable for bridges in close proximity to a weigh station.
Proximity is important to insure the cargo has not changed significantly between the
initial weighing and arrival at the bridge. Secondly, independent operators of the
vehicles as well as the highway officials must agree to support the technology.
Incentives will need to be exercised. For highway officials, the possibility of safer bridges
may be incentive enough. For the independent vehicle owners, incentives such as toll
reductions or fuel credits may be required. Eventually, if the concept proves viable, it
may be mandated, eliminating the need for fiscal incentive. The revenue cost-benefit
will have to be evaluated to determine if this technology is truly viable. To make this
framework more robust, Notre Dame researchers have also discussed the possibility of
self-reported payloads through a social science experiment with teamsters (Kijewski-
182
Correa, et al. 2010). This would eliminate the need to identify weigh stations up traffic
of the instrumented bridge and allow teamsters to upload a self-reported payload by
cell phone or internet, of course requiring the creation of an appropriate incentive
strategy.
Figure 7.6: Illustration of steps in RFID RINAS concept.
While weigh stations with data transmission by RFID tags may provide the most
quantitative information on vehicle loadings for RINAS, the acknowledged limitations
recognize that it is not appropriate for every bridge and failure of incentives for
teamsters may halt the application of these technologies. Therefore, more quantitative
M-Node
Vehicle Database
Vehicle Database
www 1. Vehicle enters weigh station
3. Tagged vehicle re-enters traffice
4. Vehicle is scanned by M-Node
5. Vehicle crosses instrumented bridge
2. Vehicle is weighed and taged
183
load classification for RINAS can be achieved using weigh-in-Motion (WIM) sensors.
WIM is a very traditional approach to record axle weights and gross weights as vehicles
drive over the sensor. Unlike older static weigh stations, WIM systems do not require
the vehicles to stop, making them much more efficient and less intrusive. There are
several types of WIN sensors currently available (Bushman 1998), as now summarized:
Bending Plate: Bending plate WIM systems use plates with strain gauges bonded to the underside. As a vehicle passes over the bending plate, the system records the strain measured by the strain gauge and calculates the dynamic load based on the plate properties.
Piezoelectric Sensor: Piezoelectric WIM systems use piezo sensors to detect a change in voltage caused by the pressure exerted on the sensor by the tire, thus allowing the axle weight to be determined.
Load Cell: Load cells use a single load cell with two scales to detect an axle and weigh both the right and left sides of the axle simultaneously. As a vehicle passes over the load cell, the system records the weights measured by each scale and sums them to obtain the axle weight.
184
Figure 7.7: Common configuration of different WIM systems.
These sensors are usually installed in a strip embedded in the pavement
perpendicular to the traffic direction, as shown in Figure 7.7. WIM systems record
instantaneous dynamic axle loads and spacings, the number of axles, and the speed of
the vehicle.
7.2.2 Full-scale Validation
The structural health monitoring framework developed in this research includes
several novel concepts, which were verified by numerous simulated and experimental
test beds. One obvious extension of this work is the full-scale validation to further testify
the efficiency of the proposed SHM framework and to prepare the final employment to
civil infrastructure systems. With the hardware prototypes now available, as described
in the previous section, future work would commence with three types of validations:
185
The collection of traffic data under a variety of conditions to verify the feasibility of RINAS and its ability to sufficiently restrict operational constraints to generate a reduced-size reference pool
The measurement of accelerations and strains in actual field conditions to confirm that the hardware and transmission can be conducted with sufficient accuracy, i.e., minimal noise interference
Field operation of the total wireless network for monitoring over a trial period to verify overall system performance in realistic conditions
The subsequent full-scale validations should be conducted at four levels. The first
level should collect traffic data to verify the ability of the RINAS system to isolate target
vehicles under a wide range of traffic patterns and weather conditions. In the event that
RINAS will utilize weigh station data, full-scale validation should be conducted on
bridges in close vicinity to state weigh stations. Table 7.5 shows the detailed information
of some weigh stations near South Bend. These are good candidates for possible test
beds for this variation on the RINAS concept.
186
TABLE 7.5
WEIGH STATIONS NEAR SOUTH BEND, IN (SOURCE: DIESELBOSS.COM)
STATE HWY EXIT DIRECTION AREA
IN I-465 171 WB
IN I-65 LOWELL SB S CHICAGO
IN I-65 SEYMOUR SB S INDIANAPOLIS
IN I-69 80 SB S FT WAYNE
IN I-70 RICHMOND WB OH BORDER
IN I-70 TERRE HAUTE EB E OF TERRE HAUTE
IN I-74 171 WB IN/OH BORDER
IN I-74 19 SB
IN I-74 VEEDERSBURG EB IN/IL BORDER
IN I-94 26 SB E CHICAGO
Note: WB = west bound, SB = south bound, EB = east bound
The second level of full-scale validation will then involve installation of the
wireless network on a highway bridge in parallel with a traditional wired system using
the same sensors and data loggers to compare the acquired signals to insure there are
no losses when a wireless format is employed. A Campbell CR3000 data logger with
specifications in Table 7.6 would be well suited to this task. The third level of validation
will then utilize the WSN in a training period to acquire the undamaged reference pool.
This will include the use of the RINAS concept. In the fourth and final level of validation,
the complete system will be set into operation to assess the bridge.
187
TABLE 7.6
SPECIFICATIONS OF CR3000 MICROLOGGER (SOURCE: CAMPBELL SCIENTIFIC)
Analog inputs: 28 single-ended or 14 differential, individually configured
Pulse counters: 4
Switched voltage excitations: 4
Switched current excitations: 3
Control/digital I/O ports: 8
Continuous analog outputs: 2
A/D bits: 16
Scan rate: 100Hz
7.2.3 Genetic Algorithm Methods for Damage Localization
Dampster-Shafter evidence theory was shown by this dissertation to be a
powerful method for combining accumulative evidence for the purposes of damage
localization; however, source importance is not considered when weighing multiple
evidences according to Dempster–Shafer Theory. The weighted balance Dempster–
Shafer theory in this dissertation provided some remedy to this problem. However,
future work should explore other methodologies, such as Genetic Algorithms (GA) for
damage localization. GA is a global probabilistic search algorithm which can identify
damage-driven effects from others. With coefficients of regressive models as the input
to such supervised learning approaches, there is the potential to enhance localization
capabilities.
188
APPENDIX A: PUBLICATIONS RELATED TO THIS RESEARCH
Kijewski-Correa, T., Montestruque, L., Su, S., and Savona, G. (2010) “A Rapidly Re-Deployable Wireless Sensor Network for Structural Assessment by Non-Expert End Users: The CITI-SENSE Concept,” Proceedings of 5th World Conference on Structural Control and Monitoring, July 12-14, Tokyo, Japan.
Kijewski-Correa, T. and Su, S. (2009) “BRAIN: A Bivariate Data-Driven Approach to Damage Detection in Multi-Scale Wireless Sensor Networks,” Smart Structures and Systems, 5(4): 415-426.
Su, S., Kijewski-Correa, T. and Pando Balandra, J. F., (2009), “Bivariate Regressive Adaptive INdex for Structural Health Monitoring: Performance Assessment and Experimental Verification, Proceedings of SPIE Smart Structures/NDE, March 9-12, San Diego.
Kijewski-Correa, T., Su, S. and Cycon, J. (2008) “System Identification in Wired, Wireless and Hybrid Architectures,” Proceedings of 5th International Engineering and Construction Conference (IECC’5), UC Irvine, August 27-29.
Su, S. and Kijewski-Correa, T. (2007) “On the Use of a Bivariate Regressive Adaptive INdex for Structural Health Monitoring,” Proceedings of SHM-II 2007, November, Vancouver, Canada.
Su, S. and Kijewski-Correa, T. (2007) “Performance Verification of Bivariate Regressive Adaptive Index for Structural Health Monitoring,” Proceedings of SPIE Smart Structures and Materials & Nondestructive Evaluation and Health Monitoring, March 18-22, San Diego, CA.
Kijewski-Correa, T., Su, S., Abittan, E., and Antsaklis, P. (2006) “On the Use of Heterogeneous, Wireless Sensor Networks for Damage Assessment in Bridges Under Unknown Excitations,” Fourth World Conference on Structural Control and Monitoring (4WCSCM), July 11-13, San Diego, CA.
Kijewski-Correa, T., Haenggi, M. and Antsaklis, P. (2006) “Wireless Sensor Networks for Structural Health Monitoring: A Multi-Scale Approach,” Proceedings of 2006 ASCE Structures Congress, 17th Analysis and Computation Specialty Conference, May 18-21, St. Louis.
189
Kijewski-Correa, T., Haenggi, M. and Antsaklis, P. (2005) “Multi-Scale Wireless Sensor Networks for Structural Health Monitoring,” Proceedings of SHM-II’05, Nov. 16-18, Shenzhen, China.
190
BIBLIOGRAPHY
Aktan, A. E. (2006) Commodore Barry Bridge USA,SAMCO Final Report 2006-F04 Case Studies, Philadelphia, PA.
Aktan, A. E., Helmicki, A. J., Hunt, V. J., Catbas, N., Lenett, M. and Levi, A. (1997) "Structural identification for condition assessment of civil infrastructure," Part 3 (of 6), June 4, 1997 - June 6, 1997.
ASCE. (2009) Report Card for American Infrastructure,American Society of Civil Engineers,
ASCE. (2005) "Report Card for American Infrastructure,"
Bischoff, R., Meyer, J., Feltrin, G. and Motavalli, M. (2007) Data Processing and Management Aspects of Wireless Sensor Networks for Structural Health Monitoring,
Bramberger, M., Roman, P., Rinner, B., Schwabach, H. and Strobl, B. (2003) "Intelligent Traffic Video Sensor: Architecture and Applications," Proceedings of the Workshop on Telecommunications and Mobile Computing (TCMC), San Diego, CA, March.
Bushman, R. P., A. (1998) "Weigh in Motion Technology: Economics and Performance," North American Travel Monitoring Exposition and Conference, Charlotte, North Carolina USA, May 11-15.
Chen, Y., Tan, C., Feng, M. Q. and Fukuda, Y. (2006) "A Video-Assisted Approach for Structural Health Monitoring of Highway Bridges Under Normal Traffic - Art. no. 61741V," Smart Structures and Materials 2006: Sensors and Smart Structures Technologies for Civil, Mechanical , and Aerospace Systems, Pts 1 and 2, 6174 :V1741.
Chung, S. W., Choi, Y. J. and Kim, S. J. (2003) "Computational Simulation of Localized Damage by Finite Element Remeshing Based on Bubble Packing Method," CMES - Computer Modeling in Engineering and Sciences, 4 (6):707.
Dacco, S. and Satchell, R. (2001) "Forward and Spot Exchange Rates in a Bivariate TAR Framework," European Journal of Finance, 7 (2):131.
Doebling, S. W., Farrar, C. R. and Prime, M. B. (1998) "Summary Review of Vibration-Based Damage Identification Methods," Shock and Vibration Digest, 30 (2):91.
191
Farrar, C. R. (1999) Test Report: Eight Degree of Freedom System,Los Alamos National Laboratory,
Farrar, C. R. and Jauregui, D. A. (1998 b) "Comparative Study of Damage Identification Algorithms Applied to a Bridge: II. Numerical Study," Smart Materials & Structures, 7 (5):720.
Farrar, C. R. and Jauregui, D. A. (1998 a) "Comparative Study of Damage Identification Algorithms Applied to a Bridge: I. Experiment," Smart Materials & Structures, 7 (5):704.
Farrar, C. R., Doebling, S. W., Los Alamos National Lab.,NM (United States) and Conference: DAMAS `97, Sheffield (United Kingdom),30 Jun - 2 Jul 1997. (1997) An Overview of Modal-Based Damage Identification Methods, United States.
Fasel, T. R., Sohn, H. and Farrar, C. R. (2003) "Application of Frequency Domain ARX Models and Extreme Value Statistics to Damage Detection," Smart Structures and Materials 2003: Smart Systems and Nondestructive Evaluation for Civil Infrastructures, 5057 :145.
Feng, M. Q., Chen, Y. B. and Tan, C. A. (2005) "Global Structural Condition Assessment of Highway Bridges by Ambient Vibration Monitoring," Nondestructive Detection and Measurement for Homeland Security III, 5769 :111.
Fugate, M. L., Sohn, H. and Farrar, C. R. (2001) "Vibration-Based Damage Detection using Statistical Process Control," Mechanical Systems and Signal Processing, 15 (4):707.
Gentile, C. and Gallino, N. (2008) "Ambient Vibration Testing and Structural Evaluation of an Historic Suspension Footbridge," Advances in Engineering Software, 39 (4):356.
Grimmelsman, K. A., Pan, Q. and Aktan, E. (2007) "Analysis of Data Quality for Ambient Vibration Testing of the Henry Hudson Bridge," Journal of Intelligent Material Systems and Structures, 18 (8):765.
Guo, H. Y., Zhang, L., Zhang, L. L. and Zhou, J. X. (2004) "Optimal Placement of Sensors for Structural Health Monitoring using Improved Genetic Algorithms," Smart Materials & Structures, 13 (3):528.
Guo, H. and Zhang, L. (2006) "A Weighted Balance Evidence Theory for Structural Multiple Damage Localization," Computer Methods in Applied Mechanics and Engineering, 195 (44-47):6225.
Harshvardan, P. J., Bogdan, B., Prabhu, M. M., et al. (2006) Wireless Structural Health Monitoring System Design, Implementation and Validation,Center for Advanced Computing and Communications (CACC), Raleigh, Raleigh, NC.
192
He, X., Moaveni, B., Conte, J. P. and Elgamal, A. (2006) "Comparative study of system identification techniques applied to New Carquinez Bridge," 3rd International Conference on Bridge Maintenance, Safety and Management - Bridge Maintenance, Safety, Management, Life-Cycle Performance and Cost, July 16, 2006 - July 19, 2006.
Hill, J., Szewczyk, R., Woo, A., Hollar, S., Culler, D. and Pister, K. (2000) "System architecture directions for networked sensors," Ninth International Conference on Architectural Support for Programming Languages and Operating Systems, 12.
Hipley, P. (2000) "Caltran’s Current State-of-Practice," Proceedings of the Invited Workshop of The Consortium of Organizations for Strong-Motion Observation Systems, Richmond, CA,
Holt, R., Hartmann, J. L. and Turner-Fairbank Highway Research Center. (2008) Adequacy of the U10 & L11 Gusset Plate Designs for the Minnesota Bridge no. 9340 (I-35W Over the Mississippi River), Federal Highway Administration, Turner-Fairbank Highway Research Center, [McLean, Va.].
Hoose, N. (1992) "IMPACTS: An Image Analysis Tool for Motorway Surveillance," Traffic Engineering and Control, 33 (3):140.
Hoose, N.,. (1991) Computer Image Processing in Traffic Engineering, Research Studies Press ; J. Wiley, Taunton, Somerset, England; New York.
Iwasaki, A., Todoroki, A., Shimamura, Y. and Kobayashi, H. (2004) "Unsupervised Structural Damage Diagnosis Based on Change of Response Surface using Statistical Tool (Application to Damage Detection of Composite Structure)," JSME International Journal, Series A: Solid Mechanics and Material Engineering, 47 (1):1.
Jo, H., Sung-Han Sim, Mechitov, K. A., et al. (2011) "Hybrid wireless smart sensor network for full-scale structural health monitoring of a cable-stayed bridge," Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2011, 7-10 March 2011.
Johnson, E. A., Lam, H. F., Katafygiotis, L. S. and Beck, J. L. (2000) "A benchmark problem for structural health monitoring and damage detection," Proceedings of the 14th ASCE engineering mechanics conference, Austin, TX, May 21–24.
Kijewski-Correa, T., Montestruque, L., Su, S. and Savona, G. (2010) "A Rapidly Re-Deployable Wireless Sensor Network for Structural Assessment by Non-Expert End Users: The CITI-SENSE Concept," Proceedings of 5th World Conference on Structural Control and Monitoring, Tokyo, Japan, July 12-14.
193
Kijewski-Correa, T., Su, S., Abittan, E. and Antsaklis, P. J. (2006 b) "On the Use of Heterogeneous, Wireless Sensor Networks for Damage Assessment in Bridges Under Unknown Excitations," 4th World Conference on Structural Control and Monitoring, San Diego, Ca, 11-13 July.
Kijewski-Correa, T., Haenggi, M. and Antsaklis, P. (2006 a) "Wireless sensor networks for structural health monitoring: A multi-scale approach," 17th Analysis and Computation Specialty Conference, May 18, 2006 - May 21, 2006.
Kijewski-Correa, T. L. and Cycon, J. (2007) System Identification of Constructed Buildings: Current State-of-the-Art and Future Directions,
Kim, J. T., Ryu, Y. S., Cho, H. M. and Stubbs, N. (2003) "Damage Identification in Beam-Type Structures: Frequency-Based Method Vs Mode-Shape-Based Method," Engineering Structures, 25 (1):57.
Kim, S. (2005) Wireless Sensor Networks for Structural Health Monitoring : Research Project,
Krantz, D., Belk, J., Biermann, P. J., et al. (1999) "Project Update: Applied Research on Remotely-Queried Embedded Microsensors," Smart Structures and Materials 1999: Smart Electronics and Mems, 3673 :157.
Law, S. S., Li, X. Y., Zhu, X. Q. and Chan, S. L. (2005) "Structural Damage Detection from Wavelet Packet Sensitivity," Engineering Structures, 27 (9):1339.
Lee, Y. S. and Chung, M. J. (2000) "A Study on Crack Detection using Eigenfrequency Test Data," Computers & Structures, 77 (3):327.
Lemke, J. (2000) "A Remote Vibration Monitoring System using Wireless Internet Data Transfer," Nondestructive Evaluation of Highways, Utilities, and Pipelines Iv, 3995 :436.
Li, H., Fang, H. and Sau-Lon, J. H. (2007) "Damage Localization and Severity Estimate for Three-Dimensional Frame Structures," Journal of Sound and Vibration, 301 (3-5):481.
Littler, J. D. (1995) "An Assessment of some of the Different Methods for Estimating Damping from Full-Scale Testing," Journal of Wind Engineering and Industrial Aerodynamics, 57 (2-3):179.
Liu, Z., Wang, B. and Yang, W. (2010) "Design and Deployment of Bridge Structural Health Monitoring System Based on Wireless Sensor Network," 2010 6th International Conference on Wireless Communications, Networking and Mobile Computing (WiCOM), 23-25 Sept. 2010.
194
Ljung, L. (1999) System Identification Theory for the User, Prentice Hall, Englewood Cliffs, NJ.
Lynch, J., Sundararajan, A., Law, K., et al. (2003) "Field Validation of a Wireless Structural Health Monitoring System on the Alamosa Canyon Bridge," SPIE’s 10th Annual International Symposium on Smart Structures and Materials, San Diego, CA, March 3–6.
Lynch, J. P. and Loh, K. J. (2006 b) "A Summary Review of Wireless Sensors and Sensor Networks for Structural Health Monitoring," Shock and Vibration Digest, 38 (2):91.
Lynch, J. P., Wang, Y., Loh, K. J., Jin-Hak Yi and Chung-Bang Yun. (2006 a) "Performance Monitoring of the Geumdang Bridge using a Dense Network of High-Resolution Wireless Sensors," Smart Materials and Structures, 15 (6):1561.
Lynch, J. P., Sundararajan, A., Law, K. H., Kiremidjian, A. S. and Carryer, E. (2004) "Embedding Damage Detection Algorithms in a Wireless Sensing Unit for Operational Power Efficiency," Smart Materials & Structures, 13 (4):800.
Lynch, J. P., Law, K. H., Kiremidjian, A. S., et al. (2002) "Validation of a Wireless Modular Monitoring System for Structures," Smart Structures and Materials 2002: Smart Systems for Bridges, Structures, and Highways, 4696 :124.
Lynch, J. P., Law, K., Straser, E., Kiremidjian, A. and Kenny, T. (2001) "The development of a wireless modular health monitoring system for civil structures," Proceedings of the Second MCEER Workshop on Mitigation of Earthquake Disaster by Advanced Technologies, Las Vegas, NV,
Maser, K., Egri, R., Lichtenstein, A. and Chase, S. (1996) Field Evaluation of a Wireless Global Bridge Evaluation and Monitoring System (WGBEMS),
Masri, S., Nakamura, M., Chassiakos, A. and Caughey, T. (1996) "Neural Network Approach to Detection of Changes in Structural Parameters," Journal of Engineering Mechanics, 122 (4):350.
Masri, S. F., Sheng, L. -., Caffrey, J. P., Nigbor, R. L., Wahbeh, M. and Abdel-Ghaffar, A. (2004) "Application of a Web-Enabled Real-Time Structural Health Monitoring System for Civil Infrastructure Systems," Smart Materials and Structures, 13 (6):1269.
Mendoza, L., Reyes, A. and Luco, J. E. (1991) "Ambient Vibration Tests of the Mexicali General Hospital," Earthquake Spectra, 7 (2):281.
Meyer, J., Bischoff, R., Feltrin, G. and Saukh, O. (2006) "A low power wireless sensor network for structural health monitoring," 3rd International Conference on Bridge
195
Maintenance, Safety and Management - Bridge Maintenance, Safety, Management, Life-Cycle Performance and Cost, Jul 16 - 19 2006.
Mitchell, K., Sana, S., Liu, P. X., Cingirikonda, K., Rao, V. S. and Pottinger, H. J. (2000) "Distributed Computing and Sensing for Structural Health Monitoring Systems," Smart Structures and Materials 2000: Smart Electronics and Mems, 3990 :156.
Montestruque, L. and Lemmon, M. D. (2008) "CSOnet: a metropolitan scale wireless sensor-actuator network," International Workshop on Mobile Device and Urban Sensing (MODUS), April 21.
Munch, B. and Steingrimsson, U. (2006) "Optimized RGB for image data encoding,"
Murugesh, G. (2001) "Health Monitoring of the New Benicia Martinez Bridge," Health Monitoring and Management of Civil Infrastructure Systems, 4337 :256.
Nagayama, T., and Spencer, B. F. (2007) Structural Health Monitoring using Smart Sensors,A Report on a Research Project Sponsored by the NSF,
Nair, K. K., Kiremidjian, A. S. and Law, K. H. (2006) "Time Series-Based Damage Detection and Localization Algorithm with Application to the ASCE Benchmark Structure," Journal of Sound and Vibration, 291 (1-2):349.
Nair, K. K. and Kiremidjian, A. S. (2007) "Time Series Based Structural Damage Detection Algorithm using Gaussian Mixtures Modeling," Journal of Dynamic Systems Measurement and Control-Transactions of the Asme, 129 (3):285.
NBC. (2008) PennDOT Plans to Re-Inspect all I-95 Bridge Structures,NBC 10,
Oshima, T., Rahman, M. S., Mikami, S., et al. (2000) "Application of Smart Materials and Systems to Long-Term Bridge Health Monitoring," Nondestructive Evaluation of Highways, Utilities, and Pipelines Iv, 3995 :253.
Petrou, M. and Bosdogianni, P. (1999) Image Processing : The Fundamentals, Wiley, Chichester [England]; New York.
Phares, B. M., Washer, G. A., Rolander, D. D., Graybeal, B. A. and Moore, M. (2004) "Routine Highway Bridge Inspection Condition Documentation Accuracy and Reliability," Journal of Bridge Engineering, 9 (4):403.
Pines, D. (2002) "Status of Structural Health Monitoring of long‐span Bridges in the United States," Progress in Structural Engineering and Materials, 4 (4):372.
196
Pines, D. J. and Lovell, P. A. (1998) "Conceptual Framework of a Remote Wireless Health Monitoring System for Large Civil Structures," Smart Materials & Structures, 7 (5):627.
Pines, D. J. (1999) "Monitoring the health of civil infrastructure," Proceedings of the 1999 Strutures Congress 'Structural Engineering in the 21st Century, April 18, 1999 - April 21, 1999.
Reid, R. L. (2008) "Special Report: The Infrastructure Crisis," Civil Engineering, 78 (1):40.
Rytter, A. (1993) Vibrational Based Inspection of Civil Engineering Structures.Instituttet for Bygningsteknik, Aalborg Universitetscenter, Aalborg.
Soares, C. G. and Cunha, C. (2000) "Bivariate Autoregressive Models for the Time Series of Significant Wave Height and Mean Period," Coastal Engineering, 40 (4):297.
Sohn, H., Farrar, C. R., Hunter, N. F. and Worden, K. (2001) "Structural Health Monitoring using Statistical Pattern Recognition Techniques," Journal of Dynamic Systems Measurement and Control-Transactions of the Asme, 123 (4):706.
Straser, E. G., Kiremidjian, A. S., Meng, T. H. and Redlefsen, L. (1998) "A Modular, Wireless Network Platform for Monitoring Structures," Imac - Proceedings of the 16th International Modal Analysis Conference, Vols 1 and 2, 3243 :450.
Su, S. and Kijewski-Correa, T. (2007) "Performance Verification of Bivariate Regressive Adaptive Index for Structural Health Monitoring," Sensors and Smart Structures Technologies for Civil, Mechanical, and Aerospace Systems 2007, Pts 1 and 2, 6529 :U575.
Taktak, R., Dufaut, M., Wolf, D. and Husson, R. (1996) "Analysis and Inspection of Road Traffic using Image Processing," Mathematics and Computers in Simulation, 41 (3-4):273.
USDOT. (2008) Fiscal Year 2009 Budget in Brief, U.S. Department of Transportation, Washington, DC.
USDOT. (2007) 2006 Status of the Nation's Highways, Bridges, and Transit Conditions & Performance : Report to Congress, U.S. Department of Transportation, [Washington, D.C.].
USDOT. (2006) 2006 Mn/DOT BRIDGE INSPECTION REPORT,U.S. Department of Transportation Federal Highway Administration,
USDOT. (2004) National Bridge Inspections Standards Regulation,U.S. Department of Transportation,
197
Wang, P. (2008) "On the realization of credible combination for D-S evidence theory," 7th World Congress on Intelligent Control and Automation, WCICA'08, June 25, 2008 - June 27, 2008.
Want, R. (2006) "An Introduction to RFID Technology," IEEE Pervasive Computing, 5 (1):25.
Wong, K. Y., Chan, W. Y. K., Man, K. L., Mak, W. P. N. and Lau, C. K. (2000) "Structural Health Monitoring Results on Tsing Ma, Kap Shui Mun and Ting Kau Bridges," Nondestructive Evaluation of Highways, Utilities, and Pipelines Iv, 3995 :288.
Xu, B., Wu, Z. and Yokoyama, K. (2003) "Response Time Series Based Structural Parametric Assessment Approach with Neural Networks,":601.
Yan, Y. J., Cheng, L., Wu, Z. Y. and Yam, L. H. (2007) "Development in Vibration-Based Structural Damage Detection Technique," Mechanical Systems and Signal Processing, 21 (5):2198.