structural health monitoring of cable-supported bridges...

ABSTRACT: This presentation summarizes the development of vibration-based structural health monitoring and damage detection methods and their applications to cable-supported bridges. It has been shown that because of very low modal sensitivity of the structures with respect to damage at a component level, only the methods which have high tolerance to incompleteness of measured data, measurement noise, modeling error and structural uncertainty are applicable to large-scale cable-supported bridges for vibration-based damage identification. The research focus has been given to eliminate the environmental effects in vibration-based damage detection. Both parametric and non-parametric procedures are proposed for eliminating the temperature effect in vibration-based damage detection, and several illustrative examples are provided.

KEY WORDS: Cable-supported bridge; Structural health monitoring; Vibration-based damage detection.

1 INTRODUCTION

Maintaining the safe and reliable operation of vital infrastructure systems that society depends upon is critical to securing the well-being of people, protecting significant capital investments, and supporting the vitality of the global economy. However, infrastructure systems cannot last forever; even after construction, these complex systems begin to deteriorate within the demanding operational environment in which they are placed. Given the costs associated with infrastructure repair and the high environmental impact of infrastructure construction, government authorities worldwide are increasingly seeking sensing and instrumentation systems to more objectively monitor their crucial infrastructure systems to ensure structural and operational safety. Structural health monitoring (SHM) technology can provide engineers and owners with early warnings on damage and structural deterioration prior to the need for costly repairs or situations that can lead to catastrophic structural collapses. The past two decades have witnessed a rapid increase in the applications of SHM technology to various civil engineering structures around the world [1-6].

The development of SHM technology in Hong Kong has evolved for over fifteen years since the implementation of the so-called “Wind And Structural Health Monitoring System (WASHMS)” on the suspension Tsing Ma Bridge in 1997. Five long-span cable-supported bridges in Hong Kong, namely the Tsing Ma (suspension) Bridge, the Kap Shui Mun (cable-stayed) Bridge, the Ting Kau (cable-stayed) Bridge, the Western Corridor (cable-stayed) Bridge, and the Stonecutters (cable-stayed) Bridge, have been instrumented by the Highways Department of the Hong Kong SAR Government with sophisticated long-term SHM systems [7-9]. The SHM systems also were periodically updated in order to effectively execute the functions of structural condition monitoring and structural degradation evaluation under in-service condition. A lot of investigations on using the monitoring data from the instrumented bridges for structural health and condition

assessment have been carried out [10-18]. This paper outlines the development of structural health monitoring and damage detection methods for cable-supported bridges based on long-term vibration monitoring data.

2 BRIDGE SHM SYSTEMS IN HONG KONG

The SHM systems for the five cable-supported bridges in Hong Kong were designed and implemented for real-time monitoring of four categories of parameters: (i) environments (wind, temperature, seismic, humidity, corrosion status, etc.), (ii) operational loads (highway traffic, railway traffic), (iii) bridge features (including static features such as influence coefficients and dynamic features such as modal parameters), and (iv) bridge responses (geometrical profile, cable force, displacement/deflection, strain/stress histories, cumulative fatigue damage, etc.). Figures 1 to 5 show the SHM systems deployed on the bridges. In accordance with a modular design concept [19], each system is composed of six modules, i.e., Module 1: Sensory System (SS), Module 2: Data Acquisition & Transmission System (DATS), Module 3: Data Processing & Control System (DPCS), Module 4: Data Management System (DMS), Module 5: Structural Health Evaluation System (SHES), and Module 6: Inspection & Maintenance System (IMS). The SS and DATS are sensors, on-structure data acquisition units (DAUs), and cabling networks for signal collection, processing and transmission. The DPCS is a computer system for the execution of system control, system operation display, and processing and analysis of data. The DMS refers to a database or data warehouse system for the storage and retrieval of monitoring data and analysis results. The SHES, which is the core of the SHM system, is a high-performance computer system equipped with appropriate software and advanced analysis tools for the execution of finite element analysis, sensitivity analysis and model updating, bridge feature and response analysis, diagnostic and prognostic analysis, and visualization of analyzed results. It includes an on-line structural condition evaluation system and

Structural health monitoring of cable-supported bridges based on vibration measurements

Y. Q. Ni

Department of Civil and Environnemental Engineering, The Hong Kong Polytechnic University, Hong Kong Email: [email protected]

Proceedings of the 9th International Conference on Structural Dynamics, EURODYN 2014Porto, Portugal, 30 June - 2 July 2014

A. Cunha, E. Caetano, P. Ribeiro, G. Müller (eds.)ISSN: 2311-9020; ISBN: 978-972-752-165-4

65

Fi

Fig

Figure 1. SH

igure 2. SHM

Figure 3. SH

gure 4. SHM

Figure 5. SH

HM system fo

system for K

HM system fo

system for W

HM system for

or Tsing Ma B

Kap Shui Mun

or Ting Kau B

estern Corrido

r Stonecutters

Bridge (TMB).

Bridge (KSM

Bridge (TKB).

or Bridge (WC

Bridge (SCB)

282

.

MB).

CB).

).

an ofIMSinspe

Thhavecableto deon mmonandthat (iv) evencollifor teffec(vi) bridgthe tSHMthan

3

The necefor ldetecmethfeatudegralgorthe realienvirwindenvirmodstrucare nor fahavesuccmethcorreeffecfromby em

Alpossdamastrucopercharstrucdetecof thfromfeatufalseenvir

ff-line structu is a laptopection and mahe SHM systee a better unde-supported bevelop and valmeasurement itoring and evlimiting the vinfluence bridto evaluate

nts such as sions, etc.; (vthe planning,ctive bridge ito minimize tge inspectiontraffic operat

M systems havten years.

VIBRATION‐B

implementatessity of develarge-scale ciction methodhods, which uures of a stradation, have rithms neglecdynamic charty, however,ronmental andd, traffic, ronmental ef

dal properties,ctural damagenot taken intoalse-negative e been identiessful applica

hods to largeelation betwects with the u

m the instrumemploying varilarming the pible stage is age detectionctural damage ration of a stracteristics) thcture are firsction techniquhe structure a

m the healthy sures shift from-negative damronmental effe

ral health andp-computer-aiaintenance of tems were dev

derstanding ofbridges under lidate the bridresults; (iii)

valuation modvariation rangdge performanstructural inttyphoons, str) to provide in, scheduling, inspection anthe lane-closu/maintenance ional period ve operated c

BASED DAMAG

tion of the eloping practicvil structures

ds available, use the measructure to inbeen extensiv

ct the effects racteristics of civil structd operational chumidity, a

ffects also ca which may

e. If the effectaccount in da

diagnosis mayified as one ations of vibre-scale civil een modal puse of long-tnted bridges iious statisticalresence of strthe most pr

n. The philoalarming is ra

ructure, damahat charactert extracted. Mue, subsequenare examined state. Ideally,

m the healthy smage may be

cts.

d safety assessided portablethe SHM systvised for the f the structura

their in-servidge evaluation to setup/caldels and criteges of the phnce under in-tegrity immetrong earthqunformation an

evaluating and maintenancure time requi activities, hof the bridge

continuously a

GE ALARMING

SHM systemcal damage d

s. Among varthe vibration

sured changesndicate strucvely investiga

of environmf the underlytures are subconditions suc

and solar rause changes

mask the chts of environm

damage detectiy occur. Envi

main pitfalration-based

structures. Mparameters anterm monitoriin Hong Kongl learning algoructural damarimary objectsophy of theather simple:

age-sensitive frize the healMaking use nt data from a

to see if thean alarm willstate. Howevealarmed in th

sment system. e system for em itself. objectives: (

al behavior ofice condition;

n techniques blibrate/update eria for descrihysical parameservice conditdiately after

uakes and vend analytical tand designingce strategies; red for exerci

hence maximies. The deploand steadily m

G

ms highlights detection methrieties of damn-based diagns in the dyna

ctural damageated. Most of t

mental changeying structurebject to varch as temperaradiation. T

in dynamic hanges causedmental condition, false-posironmental effl that limits damage detecModeling ofnd environmeing data obtag has been stuorithms [20-23age at the eartive of structe vibration-bduring the norfeatures (dynalthy state of of some novan unknown e features devbe signaled i

er, false-positivhe presence of

The the

i) to f the ; (ii) ased

the ibing eters tion; rare

essel tools g of and

ising izing oyed more

the hods

mage nosis amic e or these s on e. In rying ature, These

and d by tions sitive ffects

the ction f the ental ained udied 3]. rliest tural ased rmal amic

f the velty state viate f the ve or f the

Proceedings of the 9th International Conference on Structural Dynamics, EURODYN 2014

66

3.1

Sevenvbeencaudamrespdataenvutilistruof tnetwrealanamulthatuncAAThealarwhino falsdamthe In pimpAA

Tits ggencapmodstopresp(formeaTinmodTherespas ppos

Tpos

Non‐param

veral non-pavironmental en proposed [2sation pattern

mage-sensitiveponse. The paa of a stru

vironmental coized to exam

ucture to see ifthe non-paramwork (AANNlization (Figulysis for fealtivariate datat it doesn’t recertainty. Two

ANN-based dae first capabilrm shall not bich the damagdamage really

se-negative alamage independ

identificationpractice, the aportant issues,

ANN and the se

1X 2X 3X

1X̂ 2X̂ X̂

Figure 6. Au

The environmegeneralization

neralization tecacity in strucdels with thepping techniqpectively, are r the simulatiasured modalg Kau Bridgedel which is trey are referrpectively. Theper the comproitive alarm an

The capability itive alarm is

metric Approa

arametric meffects in stru24-28]. The pn between thee features iattern can be eucture in heonditions are nmine data frof the data are gmetric methodN) technique, ure 6) of the ature extractia [29]. The adequire a struco capabilities amage alarminlity is to avo

be signaled if tge-sensitive fey occurs. Thearm; that is, thdent of the envn of damage-sabove capabil i.e., the enviretup of the ala

3 … iX

3X … iX̂

uto-associativ

ent-tolerant can capability. Echnique to enctural damagee use of the que, and Bay

formulated uion of differel frequencies e, and are comrained by the red to as Neir environmeomise betwee

nd to avoid falof NN0, NN1examined us

ach

methods to uctural dama

principle behine environmens inherent extracted fromealthy state not measured;om an unkngenerated from

ds is the autowhich is th

nonlinear priion and novdvantages of

ctural model aare expected ng technique

oid false-positthe environme

eatures are mee second caphe AANN shavironmental csensitive featulities are closronment-tolerarming thresho

inp

out

dem

… nX

… nX̂

“bo

ma

ve neural netw

apacity of an AEfforts have benhance the ene alarming [1AIC and FPE

yesian regularusing the finent damage s(Figure 7) o

mpared with thconventional

NN1, NN2, nt-tolerant cap

en the capabililse-negative al1, NN2, and Ning a set of u

eliminate ge alarming nd them is thantal conditionsin the struc

m the measureeven when

; and then it cnown state om this pattern-associative n

he neural netincipal compo

velty detectiothis techniqu

and accommowhen applyinto real struct

tive alarm; thental conditioasured, changability is to a

all be able to dconditions at wures is perforsely related torant capacity oold.

put layer

tput layer

mapping layer

ottleneck” layer

apping layer

work (AANN).

AANN is relateen made to s

nvironment-tol15]. Three AE technique, rization technite element mscenarios) and

of the cable-she baseline Atraining algorNN3, and Npacity is evaluities to avoid flarm.

NN3 to avoid funseen testing

the have

at the s and ctural ement n the can be f the

n. One neural twork onent

on of ue are odates ng the tures.

hat is, ons, at ge but avoid detect which rmed. o two of the

.

ted to seek a lerant

AANN early

nique, model d the

stayed AANN rithm. NN0,

luated false-

false-data,

whicmeasfeeditheirindicindexwithline)trainregudeviphasthe nthe tundethe Ain sta

Anconf

F

Nlt

id

Nlt

id

ch are also obsured under ing the unseer outputs are ces in the testix sequences i

h the novelty i) obtained froning techniqueularization tecation in the no

ses for NN0, wnovelty indextest data are aer different enAANN wouldatistical sensen alarming thfidence interva

Figure 7. Meas

Figure 8. Nov

Figure 9. Nov

0 1000.15

0.20

0.25

0.30

0.35

0.40

Fre

quen

cy (

Hz)

0 1000

0.002

0.004

0.006

0.008

0.010

0.012

0.014

0.016

0.018

0.020

Nov

elty

ind

ex

TraValTesAla

0 1000

0.002

0.004

0.006

0.008

0.010

0.012

0.014

Nov

elty

ind

ex

TraValTesAla

btained from different en

en testing dataobtained and ing phase. Figin the testingindex sequencom NN0 (fore) and NN3 hnique), respovelty index bwhile no signi

between the also acquired nvironmental d not generate e between the threshold of thal and probab

sured modal fr

velty index of

velty index of

200 300Sam

200 300Samp

aining phaselidation phasesting phasearming threshold

200 300Samp

aining phaselidation phasesting phasearming threshold

the intact (henvironmental a into the fou

d used to consgures 8 and 9

g phase (dotteces in the trairmulated by (formulated

pectively. Thebetween the trificant deviatitwo phases f

from the healconditions), i

e deviation in training and t

he novelty inbility distributi

frequencies for

f healthy struct

f healthy struct

400 500 600mple order

400 500 600ple order


ealthy) bridgeconditions.

ur AANN modstruct the novshow the nov

ed line) compining phase (sthe conventiby the Baye

ere is an obvraining and tesion is observefor NN3. Beclthy structure it is expected the novelty inesting phases.dex based onion of the nov

r first 8 modes

ture for NN0.

ture for NN3.

0 700 800

700 800

700 800

e but By

dels, velty velty pared solid ional esian vious sting ed in cause

(but that

ndex . n the velty

s.


67

indeevaphathe outlratiothispresfor

Tin texcnumthesexathe signI econfromenvposlearthanenvpresandresugenformbec

Wpresthandamof aenvgenphaBayalsothe trainfalsandfals

Talarelimconnegdammod

ex sequence hluation of the

ases. When a sstructure (te

liers, which exo of the outlie

s set is subsescribed probacivil structure

Table 1 providthe testing peeding the al

mber of outlierse quantities, mple, when thconventional

naled becauseerror (4.409%ntradictory to tm the healthvironmental citive. It implirning the stochn the unde

vironmental csented with th

d FPE techniqult, the AIC neralization cmulated a momes 11.648%

When NN2 (fosented with thn the 2.000%

mage is alarmeavoiding the fvironmental neralizes to nease. By passinyesian regularo capable of afour AANN mning techniqu

se-positive alad the Bayesianse-positive ala

Table 1. Outl

AANNNNNNNNN

The capability rm is then exminate false-pnsidered here gative alarm. Tmage scenariodal frequenci

has been define deviation bset of novelty esting phase)xceed the alarers to the totalquently calcu

ability of typees), the occurrdes the statistiphase, where larming threshrs to the totalit is easy to tehe testing datal training tech r0 exceeds th

% vs 2.000the fact that thhy bridge tconditions. Ties that NN0 hastic variatioerlying causconditions anhe testing dataque) also issu

and FPE tecapability of more overfitte% now. ormulated by he testing da

% probability ed by NN2. Itfalse-positive effects on ew situations

ng the testing rization technavoiding the fmodels, it is c

ue and the AICarm, while bon regularizatioarm.

iers of novelty

NN N0 N0 1N1 5N2 N3

of the AANNxamined. Becpositive alarmto see if theyThe ‘measureds are generateies from a v

ned [15] to facetween the trindex from an

) is availablerming threshol number of thulated. If thise I error θ (it rence of damaics of the novN0 is the n

hold, and r0 l number of nell the state oa are fed to Nhnique), an alhe prescribed %). Howevehe testing datthough meas

Therefore, theoverfitted the

ons in the trainsation patternd modal fra, NN1 (form

ues a false-poechnique doe

the AANNed AANN m

the early stopata, r0 is 0.81

of type I errt evidences thalarm. It welthe modal that are unse

data to NN3 nique), it is ffalse-positive concluded tha

C and FPE tecoth the early son technique a

y index for he

(%) r10 453 12 01 0

N models to avcause NN0 am, only NNy are capable d’ modal freqed by superpovalidated finit

cilitate quantitraining and ten unknown stae, the numbeld, is counted

he novelty indratio exceed

can be set asge is alarmed.

velty index ounumber of ou

is the ratio oovelty index. f the structure

NN0 (formulatelarm of damaprobability of

er, this resua are also obtured at diffe alarm is f training data

ning samples rrn between requencies. W

mulated by thesitive alarm. esn’t enhance

N model. It model becaus

pping techniqu0%, which isror. Therefor

hat NN2 is capll characterize

frequencies een in the tra(formulated b

found that NNalarm. Comp

at the conventchnique may sstopping technare able to es

ealthy structur

r0 (%) 4.049 1.648 0.810 0.405

void false-negand NN1 faileN2 and NN3

of avoiding fquencies in vaosing the compte element m

tative esting ate of er of

d. The dex in ds the s 0.02 .

utliers utliers of the With

e. For ed by age is f type ult is tained ferent false-

a, i.e., rather

the When e AIC

As a e the even

se r0

ue) is s less re, no apable es the

and aining by the N3 is paring tional signal nique

schew

re.

gative ed to 3 are false-arious puted

model

(FEMextrashowcompphasdamaof a chanwere5.24novevaluthe pand the davoiddiffestoppenvirbest posit

F

F

T

3.2

The HongrespoWhe

Nlt

id

Nlt

id

M) of the bridacted from thw the noveltpared with th

se obtained frage is introdubearing elem

nge of modal e used to form0%) at maximelty index oues obtained frprescribed pro58.462% vs 2damage, and td false-negati

erent damage ping techniqronment-tolercompromise

tive alarm and

igure 10. Nov

igure 11. Nov

able 2. Outlie

AANNNN2NN3

Parametric

SHM systemg Kong provonses and temen both dynam

0 1000

0.005

0.010

0.015N

ovel

ty i

ndex

TraTesAla

0 1000

0.005

0.010

0.015

Nov

elty

ind

ex

TraTesAla

dge with the ehe real monitoty index seqhe novelty inom NN2 andced by losing

ment at the mafrequencies f

mulate the AANmum. Table 2tliers in the

from NN2 andobability of typ2.000%). Boththerefore the cive alarm is vscenarios, it ue is most

rant capacity obetween the

d to avoid fals

velty index of

velty index of

rs of novelty i

N N0 (2 283 26

c Approach

ms for the fivide simultanemperature at

mic properties

200 300Samp

aining phasesting phasearming threshold

200 300Samp

aining phasesting phasearming threshold

environmentaloring data. F

quences in thndex sequenced NN3, respecg torsional andain tower. Thefor the first eiNN models is2 provides thtesting phased NN3 are gr

ype I error (63h NN2 and Ncapability of validated. Afthas been shosuitable for

of the AANNe capabilities

se-negative ala

damaged stru

damaged stru

index for dam

(%) r0

87 6366 58

ve cable-suppeous measure

different strand temperat



l effects that wigures 10 andhe testing pes in the traictively, whered bending stiffe damage-induight modes w 8.710×10-3 H

he statistics ofe. In this casereatly higher .077% vs 2.0

NN3 clearly siNN2 and NNter compariso

own that the er enhancing as it achieves

s to avoid faarm.

ucture for NN2

ucture for NN3

maged structur

0 (%) 3.077 8.462

ported bridgement of dyna

ructural posititure are measu

700 800

700 800

were d 11

phase ining e the fness uced

which Hz (≈ f the e, r0 than 00% ignal

N3 to on in early

the s the

false-

2.

3.

e.

es in amic ions. ured,


68

a parametric method can be applied to eliminate the temperature effect in vibration-based structural damage detection. In this method, a correlation model between the modal frequencies and temperatures is first formulated by a statistical learning algorithm, such as the back-propagation neural network (BPNN) technique, with the use of long-term monitoring data of modal frequencies and temperatures from the healthy structure. The configuration and parameters of the BPNN model are optimized in effort to achieve good generalization (prediction) performance. With the formulated correlation model, the modal frequencies measured under different temperature conditions can be normalized to an identical reference status of temperature so as to eliminate the temperature effect. To this end, an arbitrary reference status of temperature is set. The reference temperature together with the temperatures at which the modal frequencies Mf were measured, is passed to the correlation model to produce a reference frequency Rf and the predicted frequencies Pf . Thus the temperature-caused change in the modal frequencies is obtained by subtracting the reference frequency from the predicted frequencies. Then the normalized modal features

Nf are obtained by subtracting the temperature-caused frequency change from the measured frequencies [14]

RPi

Mi

TMi

Ni ffffff (1)

where i represents the sample order in the sequence of the measured modal frequencies. After eliminating the temperature effect on the modal frequencies, damage is assumed to be the source responsible for the variation in modal frequencies. Then, structural damage alarming can be conducted by the aforementioned AANN technique with the use of the normalized modal frequencies.

The capability of the parametric method in eschewing false-positive alarm is examined using a set of unseen monitoring data obtained from the intact (healthy) bridge but measured under different environmental conditions. For comparison, two AANN models have been formulated: one (referred to as AANN-1) uses the normalized modal frequencies, and the other one (referred to as AANN-0) uses the originally measured modal frequencies. AANN-0 (trained with the originally measured modal frequencies) is first examined with the unseen testing data that were obtained from the same structure in healthy status. Figure 12 shows the novelty index sequence in the testing phase (dotted line) in comparison with the novelty index sequence in the training phase (solid line). As there is an obvious shift in the novelty index sequences between the testing and training phases (the quantitative evaluation is the same as before), AANN-0 issues an alarm on damage. It is contradictory to the fact that the testing data are also obtained from the healthy structure though measured at different environmental conditions. Therefore, the alarm is false-positive.

AANN-1 (trained with the normalized modal frequencies) is then examined with the same unseen testing data that were obtained from the structure in healthy status. Figure 13 shows the novelty index sequences in the testing phase (dotted line) and in the training phase (solid line) produced by AANN-1. As one might expect, no observable deviation appears in the novelty index sequences between the testing and training

phases. Therefore, the parametric method is shown to be able to eliminate the temperature effect and eschew false-positive damage alarm.

Figure 12. Novelty index of healthy structure for AANN-0.

Figure 13. Novelty index of healthy structure for AANN-1.

Figure 14. Novelty index of damaged structure for AANN-1.

The performance of the parametric method in detecting the presence of structural damage with use of monitoring data obtained under different environmental conditions is further examined. Again, the ‘measured’ modal frequencies in various damage scenarios are generated by superposing the computed modal frequencies from a validated FEM of the bridge with the environmental effects that are obtained from the real monitoring data. Figure 14 shows the novelty index sequences in the testing phase compared with the novelty index sequences in the training phase obtained from AANN-1, where the damage is introduced by losing torsional and bending stiffness of a bearing element at the main tower (same as the damage case before). The difference in the means of novelty indices between the testing and training phases for

0 100 200 300 400 500 6000

0.005

0.010

0.015

0.020

0.025

0.030

0.035

0.040

Sample order

Nov

elty

ind

ex

Training phaseTest phase

0 100 200 300 400 500 6000

0.005

0.010

0.015

0.020

0.025

Sample order

Nov

elty

ind

ex

Training phaseTesting phase

0 100 200 300 400 500 600 700 8000

0.005

0.010

0.015

0.020

0.025

0.030

Sample order

Nov

elty

ind

ex

Training phaseTesting phase


69

thiswithavo

4

4.1

Theexteof sthisa nutrannetwglobmeadamif itpha

Fmairegieachmodare fivevaluwithflexregiAAobtanovfor remspanno dthe phain ttestIndethe

4.2

TheBaycastappfromwhethe diretraincondataof tcha

Fto bcon

s damage cash the intact st

oid false-negat

VIBRATION‐

Multi‐nov

e AANN-basended to roughstructural mods purpose. In tumber of reginsducers at eacwork based nbal natural frasured from t

mage region ist exhibits a dease. or illustrationin span deck ions equally inh region five dal flexibility computed fro

e modes. Thues are corruph a variance xibility sequeion, are formu

ANN is trainedained for the

velty detectionthe damage

moving two opn deck sectiondamage is sigtraining phas

ase. For the strthe training ping phase, ineed, the true region III.

Probabilis

e probabilisticyesian decisiont into an ar

plied to damam the undamether a new mundamaged c

ectly casts thning samples

nvenient for dea when appliethe PNN is thracteristic as nigure 16 showbe used for dnsists of inpu

e has increastructure. Henctive alarm is v

‐BASED DAMA

velty indices fo

sed damage hly diagnose del. Multi-novthis approach,ions and it is ach region. Fornovelty detectrequencies anthe sensors ls signaled by teviation from

n, a simulationof the suspenn the longitudmodal compovalues [30] a

om a 3D FEMhe analyticallpted with a no

of 0.005 to ences. Five Aulated for damd using the fcorrespondin

n results by ucase in which

pposite longitun. For the strnaled becausese do not devructural regionphase significndicating damdamage occu

stic neural net

c neural netwon analysis witrtificial neuraage identificamaged and measurement class or the d

he probabilityin the networ

ealing with thed for damagehat it can expneuro-weightsws a three-laydamage localiut (distribut

sed by 105.16ce, the capabilvalidated.

AGE LOCALIZA

or identifying d

alarming tethe damage revelty indices , the structureassumed that r each structurtor is formul

nd the local mocated withinthe corresponthe training p

n study is mansion Tsing Mdinal directiononents are ‘mat the five nod

M of the bridgely evaluated ormally distribgenerate the AANN modemage region idfive modal fleng region. Figusing the mulh the damage

udinal bottom ructural regione the novelty iviate from thn III, the noveantly deviate

mage occurrenurs at the mid

twork for dam

ork (PNN) [31th the Parzen wal network f

ation, the PNNdamaged sysof unknown odamaged clasy density funrk, the netwo

he noisy and se identificationplicitly accoms in the trainedyer probabilisization in theion) layer, p

6% in compality of AANN

ATION

damage regio

echnique canegion withoutare formulate

e is partitionedthere are vibrral region, a nlated by usingmodal compon this region.ding novelty i

phase to the te

ade by dividinMa Bridge inton. It is assumeeasured’. Thedes for each ree by taking the

modal flexibuted random

‘measured’ mels, one for dentification. exibility sequegure 15 showlti-novelty inde is introducechords at the

ns I, II, IV anindex sequenc

hose in the teelty index sequs from that in

nce in this red-span deck w

mage localizati

1, 32] performwindows estimframework. WN uses exemstem to estaorigin comes ss. Since the nctions (PDFrk configuratieries measuren. A salient femmodate the d network. tic neural net

e present studpattern layer

arison N-1 to

on

n be t need ed for d into ration neural g the

onents . The index esting

ng the o five ed for en the egion e first ibility noise

modal each Each ences

ws the dices, ed by e mid-nd V, ces in esting uence in the egion. within

ion

ms the mator When

mplars ablish from PNN

Fs) of ion is ement eature noise

twork dy. It r and Fig

(

(

gure 15. Multi

(a) Novelty in

(b) Novelty in

(c) Novelty in

(d) Novelty in

(e) Novelty in

-novelty indic

ndex for Regi

ndex for Regio

ndex for Regio

ndex for Regio

ndex for Regio

ces based on m

ion I

on II

on III

on IV

on V

modal flexibili

ity.


70

x1 x2 xi xp

11z

O1

Ok Os

input layer

pattern layer

summation layer

11nz 1szssnz

Figure 16. Architecture of a three-layer PNN.

summation layer. An input vector X = {x1 x2 xi xp}T to be

classified is applied to the neurons of the distribution layer that just supply the same input values to all the pattern units. In our study, this input vector consists of p modal parameters (natural frequencies, mode shapes or their combination). For the purpose of damage localization, the so-called combined modal parameters [32] which are only dependent on damage location but independent of damage extent are a good choice for the input parameters. In the pattern layer, there are s pattern classes, each representing a possible damage location. The number of pattern classes depends on a specific structure. Each neuron in the pattern layer forms a dot product of the input vector X with a weight vector Wj of a given class, zj = XWj, and then performs a nonlinear operation on zj before output to the summation layer. The activation function used here is g(zj) = exp[(zj1)/2] where is a smoothing parameter. In the summation layer, each neuron receives all pattern layer outputs associated with a given class. For instance, the output of the summation layer neuron corresponding to the class k is

kk n

jkj

n

jkjk zf

1

2

1]/)1WXexp[()X( (2)

With Equation (2), the kernel density estimators for PDFs have been cast into the PNN by setting the network weight vectors as the corresponding training vectors. Such configured PNN outputs in the summation layer the PDF estimates for each pattern class at the test vector point. The pattern class with the largest PDF implies the class of the current test vector and thereby indicates the damage location with maximum likelihood.

For illustration, a simulation study of damage localization using the PNN is made on the Tsing Ma Bridge deck. The bridge main span comprises a total of 76 deck units. In the

present study, the main span deck is divided into 16 segments, each including 4 or 5 deck units. The damage to the deck members within the same segment is classified as one patter class. As a result, there are totally 16 pattern classes (s = 16). The following modal parameters are taken to constitute the input vector: the natural frequency change ratios of the first four modes, and the three translational components of the first mode vector at the 16 nodes of deck units 2, 7, 12, 17, 22, 27, 32, 36, 41, 45, 50, 55, 60, 65, 70 and 75 (one node for each selected unit). So the length of the PNN input vector is p = 4+163 = 52. In order to obtain the training vectors, for each pattern class two damage scenarios with the damage within the same segment but different units are introduced in the FEM of the bridge and the corresponding modal parameters are evaluated. Each set of the computed modal parameters are then added with a random sequence to form the training vectors.

50 sets of modal parameters are randomly produced for each damage scenario. After entering the noise-polluted training vectors of all pattern classes as weights between the distribution (input) and pattern layers, the PNN for damage localization is configured (trained). When presenting on it a new input vector (test vector) consisting of measured modal data of unknown source, the configured PNN outputs in the summation layer the PDF estimates for each pattern class at the test vector point, and the damaged deck segment is identified by the pattern class with the largest PDF.

The test vectors are produced in a similar way to obtaining the training samples. A total of 16 damage scenarios, with one for each deck segment (pattern class), are examined in the testing phase. The testing damage scenario for each pattern class is incurred at a deck unit different from the training damage scenarios. The analytical modal parameters when incurring damage at each deck segment in turn are calculated and then polluted with random noise to obtain the ‘measured’ test vectors. For each testing damage scenario, 500 sets of noise-corrupted test vectors are produced. Table 3 summarizes the damage location identification results under different noise levels. The integral numbers in the latter five rows of the table show the number (times) of correct identification, out of 500 tests for each damage scenario. The identification accuracy (IA) is defined as the ratio of the total number of correct identification to the total test number. The IA value is 73.75% when the noise level = 0.10%, 83.68% when = 0.08%, 90.34% when = 0.06%, 95.66% when = 0.04%, and 99.14% when = 0.02%.

Table 3. Summary of correct identification using PNN technique.

Pattern class No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 IA

Testing sample No. 500 500 500 500 500 500 500 500 500 500 500 500 500 500 500 500 ratio

= 0.10% 302 168 366 246 485 354 480 405 323 364 415 436 349 251 463 493 73.75%

Number = 0.08% 343 248 444 356 493 434 493 436 351 406 464 461 405 379 484 497 83.68%

of correct = 0.06% 483 339 489 467 464 486 498 472 350 432 477 491 463 395 442 479 90.34%

localization = 0.04% 431 372 498 476 500 498 500 492 426 480 500 500 492 488 500 500 95.66%

= 0.02% 478 470 500 500 500 500 500 500 485 500 500 500 500 498 500 500 99.14%

Note: IA (Identification Accuracy) is the ratio of total number of correct localization to total number of testing samples (50016).


71

5 CONCLUSIONS

This paper summarized several vibration-based methodologies developed for SHM and damage detection of cable-supported bridges. It has been shown that the auto-associative neural network (AANN) technique is promising for damage alarming in operation with noisy measurement data. Because no structural model is required and only modal frequencies are utilized, this method is applicable to structures of arbitrary complexity. Due to the environmental effects, however, false-positive and false-negative alarms may be issued in applying the AANN technique. Both non-parametric and parametric methods have been developed to eliminate the environmental effects. The AANN technique has been extended to formulate multi-novelty indices for damage region identification. While requiring no structural model, it requires vibration sensors deployed at each of the structural regions. When an accurate structural model is available, the probabilistic neural network (PNN) technique can provide more detailed identification of damage location. For the identification of local damage in large-scale bridges, however, the methods which use dynamic strain measurement are preferred [33].

ACKNOWLEDGMENTS

The work described in this paper was supported by a grant from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project No. PolyU 5224/13E).

REFERENCES

[1] Aktan, A.E., Chase, S., Inman, D., and Pines, D.D. (2001), “Monitoring and managing the health of infrastructure systems”, Health Monitoring and Management of Civil Infrastructure Systems, S.B. Chase, A.E. Aktan (eds.), SPIE Vol. 4337, SPIE, Bellingham, USA (CD-ROM).

[2] Ko, J.M., and Ni, Y.Q. (2005), “Technology developments in structural health monitoring of large-scale bridges”, Engineering Structures, 27: 1715-1725.

[3] Brownjohn, J.M.W. (2007), “Structural health monitoring of civil infrastructure”, Philosophical Transactions of the Royal Society A, 365: 589-622.

[4] Glisic, B., Inaudi, D., and Casanova, N. (2009), “SHM process lessons learned in 250 SHM projects”, Proceedings of the 4th International Conference on Structural Health Monitoring and Intelligent Infrastructure, Zurich, Switzerland (CD-ROM).

[5] Ni, Y.Q., Xia, Y., Liao, W.Y., and Ko, J.M. (2009), “Technology innovation in developing the structural health monitoring system for Guangzhou New TV Tower”, Structural Control and Health Monitoring, 16: 73-98.

[6] Wang, M.L., and Yim, J. (2010), “Sensor enriched infrastructure system”, Smart Structures and Systems, 6: 309-333.

[7] Wong, K.Y. (2004), “Instrumentation and health monitoring of cable-supported bridges”, Structural Control and Health Monitoring, 11: 91-124.

[8] Wong, K.Y., and Ni, Y.Q. (2009), “Structural health monitoring of cable-supported bridges in Hong Kong”, Structural Health Monitoring of Civil Infrastructure Systems, V.M. Karbhari and F. Ansari (eds.), Woodhead Publishing, Cambridge, UK, 371-411.

[9] Ni, Y.Q., Wong, K.Y., and Xia, Y. (2011), “Health checks through landmark bridges to sky-high structures”, Advances in Structural Engineering, 14: 103-119.

[10] Ni, Y.Q., Hua, X.G., Wong, K.Y., and Ko, J.M. (2007), “Assessment of bridge expansion joints using long-term displacement and temperature measurement”, Journal of Performance of Constructed Facilities, ASCE, 21: 143-151.

[11] Ko, J.M., Ni, Y.Q., Zhou, H.F., Wang, J.Y., and Zhou, X.T. (2009), “Investigation concerning structural health monitoring of an instrumented cable-stayed bridge”, Structure and Infrastructure Engineering, 6: 497-513.

[12] Ni, Y.Q., Ye, X.W., and Ko, J.M. (2010), “Monitoring-based fatigue reliability assessment of steel bridges: analytical model and

application”, Journal of Structural Engineering, ASCE, 136: 1563-1573.

[13] Wong, K.Y., and Ni, Y.Q. (2011), “Structural health monitoring of a suspension bridge”, Monitoring Technologies for Bridge Management, B. Bakht, A.A. Mufti and L.D. Wegner (eds.), Multi-Science Publishing, Essex, UK, 365-390.

[14] Zhou, H.F., Ni, Y.Q., and Ko, J.M. (2011), “Eliminating temperature effect in vibration-based structural damage detection”, Journal of Engineering Mechanics, ASCE, 137: 785-796.

[15] Zhou, H.F., Ni, Y.Q., and Ko, J.M. (2011), “Structural damage alarming using auto-associative neural network technique: exploration of environment-tolerant capacity and setup of alarming threshold”, Mechanical Systems and Signal Processing, 25: 1508-1526.

[16] Ni, Y.Q., Xia, H.W., Wong, K.Y., and Ko, J.M. (2012), “In-service condition assessment of bridge deck using long-term monitoring data of strain response”, Journal of Bridge Engineering, ASCE, 17: 876-885.

[17] Ye, X.W., Ni, Y.Q., Wong, K.Y., and Ko, J.M. (2012), “Statistical analysis of stress spectra for fatigue life assessment of steel bridges with structural health monitoring data”, Engineering Structures, 45: 166-176.

[18] Xia, H.W., Ni, Y.Q., Wong, K.Y., and Ko, J.M. (2012), “Reliability-based condition assessment of in-service bridges using mixture distribution models”, Computers and Structures, 106-107: 204-213.

[19] Wong, K.Y., and Ni, Y.Q. (2009), “Modular architecture of structural health monitoring system for cable-supported bridges”, Encyclopedia of Structural Health Monitoring, C. Boller, F.-K. Chang and Y. Fujino (eds.), John Wiley & Sons, Chichester, UK, Vol. 5, 2089-2105.

[20] Ni, Y.Q., Hua, X.G., Fan, K.Q., and Ko, J.M. (2005), “Correlating modal properties with temperature using long-term monitoring data and support vector machine technique”, Engineering Structures, 27: 1762-1773.

[21] Hua, X.G., Ni, Y.Q., Ko, J.M., and Wong, K.Y. (2007), “Modeling of temperature-frequency correlation using combined principal component analysis and support vector regression technique”, Journal of Computing in Civil Engineering, ASCE, 21: 122-135.

[22] Zhou, H.F., Ni, Y.Q., Ko, J.M., and Wong, K.Y. (2008), “Modeling of wind and temperature effects on modal frequencies and analysis of relative strength of effect”, Wind and Structures, 11: 35-50.

[23] Ni, Y.Q., Zhou, H.F., and Ko, J.M. (2009), “Generalization capability of neural network models for temperature-frequency correlation using monitoring data”, Journal of Structural Engineering, ASCE, 135: 1290-1300.

[24] Worden, K. (1997), “Structural fault detection using a novelty measure”, Journal of Sound and Vibration, 201: 85-101.

[25] Sohn, H., Worden, K., and Farrar, C.R. (2001), “Novelty detection under changing environmental conditions”, Smart Systems for Bridges, Structures, and Highways, S.C. Liu (ed.), SPIE Vol. 4330, Bellingham, USA, 108-118.

[26] Kullaa, J. (2001), “Elimination of environmental influences from damage-sensitive features in a structural health monitoring system”, Structural Health Monitoring - the Demands and Challenges, F.-K. Chang (ed.), CRC Press, New York, USA, 742-749.

[27] Yan, A.M., Kerschen, G., De Boe, P., and Golinval, J.C. (2005), “Structural damage diagnosis under varying environmental conditions. Part I: a linear analysis”, Mechanical Systems and Signal Processing, 19: 847-864.

[28] Yan, A.M., Kerschen, G., De Boe, P., and Golinval, J.C. (2005), “Structural damage diagnosis under varying environmental conditions. Part II: local PCA for non-linear cases”, Mechanical Systems and Signal Processing, 19: 865-880.

[29] Kramer, M.A. (1991), “Nonlinear principal component analysis using autoassociative neural networks”, AIChE Journal, 37: 233-242.

[30] Ni, Y.Q., Zhou, H.F., Chan, K.C., and Ko, J.M. (2008), “Modal flexibility analysis of cable-stayed Ting Kau Bridge for damage identification”, Computer-Aided Civil and Infrastructure Engineering, 23: 223-236.

[31] Specht, D.F. (1996), “Probabilistic neural networks and general regression neural networks,” Fuzzy Logic and Neural Network Handbook, C.H. Chen (ed.), McGraw-Hill, New York.

[32] Ni, Y.Q., Wang, B.S., and Ko, J.M. (2002), “Constructing input vectors to neural networks for structural damage identification”, Smart Materials and Structures, 11: 825-833.

[33] Ni, Y.Q., Xia, H.W., and Ye, X.W. (2012), “Neutral-axis position based damage detection of bridge deck using strain measurement: numerical and experimental verifications”, Structural Health Monitoring 2012: Proceedings of the 6th European Workshop on Structural Health Monitoring, C. Boller (ed.), Berlin, Germany, 1205-1211.


72

structural health monitoring of cable-supported bridges...

Documents