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




Total Bridges




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.

http://en.wikipedia.org/wiki/Minnesota_Department_of_Transportation

http://en.wikipedia.org/wiki/June_15

http://en.wikipedia.org/wiki/2006

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:

http://en.wikipedia.org/wiki/Two-sided_Laplace_transform

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%



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


8DOF SYSTEM UNDER 5V INPUT VOLTAGE LEVEL



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%



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


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


DETECTION RESULTS OF HOMOGENEOUS AND HETEROGENEOUS DYNAMIC DSF FOR SIMULATED THIN BEAM

POINT C POINT D


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


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


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


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


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



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

http://en.wikipedia.org/wiki/Signal_processing

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).

http://en.wikipedia.org/wiki/Additive_color

http://en.wikipedia.org/wiki/Red

http://en.wikipedia.org/wiki/Green

http://en.wikipedia.org/wiki/Blue

http://en.wikipedia.org/wiki/Color

http://en.wikipedia.org/wiki/Primary_color#Additive_primaries

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


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



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


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


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


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


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


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

http://en.wikipedia.org/wiki/Automatic_identification_and_data_capture

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

http://en.wikipedia.org/wiki/Integrated_circuit

http://en.wikipedia.org/wiki/Modulation

http://en.wikipedia.org/wiki/Demodulation

http://en.wikipedia.org/wiki/Backscatter

http://en.wikipedia.org/wiki/Carrier_wave

http://en.wikipedia.org/wiki/Carrier_wave

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.

http://en.wikipedia.org/w/index.php?title=Axle_weight&action=edit&redlink=1

http://en.wikipedia.org/wiki/Weigh_station

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.

DECENTRALIZED DAMAGE DETECTION IN CIVIL ......Develop a wireless sensor network philosophy that...

Documents

Transcript of DECENTRALIZED DAMAGE DETECTION IN CIVIL ......Develop a wireless sensor network philosophy that...