8/9/2019 Bayesian model updatingReview Paper
1/18
Ka-Veng Yuen
Sin-Chi Kuok
Department of Civil and Environmental
Engineering,
Faculty of Science and Technology,
University of Macau,
Macao, China
Bayesian Methods for UpdatingDynamic ModelsModel updating of dynamical systems has been attracting much attention because it has avery wide range of applications in aerospace, civil, and mechanical engineering, etc. Manymethods were developed and there has been substantial development in Bayesian methods
for this purpose in the recent decade. This article introduces some state-of-the-art work. Itconsists of two main streams of model updating, namely model updating using response timehistory and model updating using modal measurements. The former one utilizes directlyresponse time histories for the identification of uncertain parameters. In particular, theBayesian time-domain approach, Bayesian spectral density approach and Bayesian fast Fou-rier transform approach will be introduced. The latter stream utilizes modal measurements ofa dynamical system. The method introduced here does not require a mode matching processthat is common in other existing methods. Afterwards, discussion will be given about the rela-tionship among model complexity, data fitting capability and robustness. An application of a22-story building will be presented. Its acceleration response time histories were recordedduring a severe typhoon and they are utilized to identify the fundamental frequency of thebuilding. Furthermore, three methods are used for analysis on this same set of measurementsand comparison will be made.[DOI: 10.1115/1.4004479]
1 Introduction
The problem of parametric identification for mathematical modelsusing input-output or output-only dynamic measurements hasreceived much attention over the years [112]. One important specialcase is modal identification, in which the parameters for identificationare the small-amplitude modal frequencies, damping ratios, modeshapes and modal participation factors of the lower modes of the dy-namical system [1315]. In particular, model updating of mechanicalor structural systems using ambient vibration survey is important. Ithas attracted much interest because it offers a means of obtainingdynamic data in an economical and efficient manner, without requir-ing the setup of special dynamic experiments (e.g., actuators) whichare usually costly, time consuming, and often obtrusive [1618]. Inambient vibration survey, the naturally occurring vibration of thestructure is measured under wind, traffic, and micro-tremors, etc.Then, a system identification technique is used to identify the small-amplitude modal frequencies, damping ratios and mode shapes of thelower modes of the structure [19]. The assumption usually made isthat the input excitation is a broad-band stochastic process adequatelymodeled as stationary Gaussian white noise. A number of methodshave been developed to tackle this problem, e.g., the instrumentalvariable method [20], the eigensystem realization algorithm [21], therandom decrement technique [22], the novelty measure technique[23] and the natural excitation technique [24]. Several other methodsare based on fitting the correlation functions using least-squares typeof approach [25,26]. Different ARMA model based least-squaresmethods have also been proposed, e.g., the two-stage least-squaresmethod [2729]. Another important type of methods is the predictionerror methods [3032] that minimize the optimally selected one-step-ahead output prediction error. The possible usage of Kalman filter formodel identification was recognized in Ref. [33]. In Ref. [34], the
extended Kalman filter was investigated for the applications to esti-mate the dynamic properties, such as modal frequencies, modaldamping coefficients and participation factors, of linear multidegree-of-freedom systems. Since then, many methods were proposed as itsevolution for linear and nonlinear dynamical systems [3544]. Othernonlinear system identification methods based on advanced statisticaltools (such as wavelets, higher-order spectra, Lie series, and artificialneural networks) have also been investigated for the case of knowninput [4550] and for the case of unknown input [5154].
Results of modal/model identification are usually restricted to
the optimal values of the uncertain parameters. However, there isadditional information related to the uncertainty associated withthe parameter estimates and it is valuable for further processing.For example, in the case where the identified modal parametersare used to update the theoretical finite-element model of a struc-ture, the updating procedure involves the minimization of a posi-tive definite quadratic function of the differences between thetheoretical and the experimental modal parameters. The weightingmatrix in this goodness-of-fit function should reflect the uncer-tainty in the values of the identified modal parameters so it can bechosen as the inverse of the covariance matrix of these parame-ters. In practice, this covariance matrix is usually estimated bycomputing the statistics of the optimal estimates of the modal pa-rameters from several sets of ambient data. However, this estima-tion is unreliable unless the number of data sets is large. Recent
interest has been arisen to determine the uncertainty of the identi-fied parameters of mechanical/structural systems using Bayesianprobabilistic approach [5587]. The parametric uncertainty can bequantified in the form of probability distribution in Bayesian infer-ence [88,89]. The quantified uncertainty can be utilized for post-processing, such as probabilistic control [9092]. In this paper,some state-of-the-art Bayesian methods are introduced. First, anexact formulation with output-only data is presented and its com-putational difficulty will be discussed. Then, the general modelupdating problem will be formulated in Sec. 3. In parametric iden-tification or model updating, a given model class with uncertainparameters is prescribed and identification techniques are used toidentify these uncertain parameters. Two important problems instructural dynamics are introduced in Secs. 4 and 5. In Sec. 4,model updating using measured response of the underlying system
is considered. A method with input-output measurement will beintroduced. Its extension for explicit consideration of both inputand output noise is also presented. Then, in the second part of thissection, the Bayesian time-domain method, Bayesian spectral den-sity method and Bayesian fast Fourier transform method will bepresented for output-only measurement. In Sec.5, another impor-tant type of model updating problem using identified eigenvaluesand eigenvectors of the system is introduced. The method pre-sented here does not require model matching, which is necessaryin most existing methods but difficult in some practical situation.After introducing parametric identification techniques, Sec.6 dis-cusses some of the key issues of model updating, including datafitting capability, robustness and posterior uncertainty. Finally, anexample of structural health monitoring under severe typhoon is
Manuscript received October 26, 2010; final manuscript received May 31, 2011;published online September 26, 2011. Transmitted by Assoc. Editor: Chin An Tan.
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-1CopyrightVC 2011 by ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
2/18
demonstrated in Sec.7. The data will be analyzed by three meth-ods and comparison will be given.
2 Exact Bayesian Formulation and Its Computational
Difficulties
Consider a single-degree-of-freedom (SDOF) system withequation of motion:
x2fX _xX2x ft (1)
where X and f are the natural frequency and damping ratio of theoscillator, respectively. The inputfis modeled as a zero-mean sta-tionary Gaussian white noise with spectral intensitySf0. Assumingstationarity of the responsex, it is Gaussian with zero-mean andautocorrelation function
Rxs pSf0
2fX3expfXjsj cosXds
fffiffiffiffiffiffiffiffiffiffiffi1f2
p sinXdjsj" #
(2)
where Xd Xffiffiffiffiffiffiffiffiffiffiffiffiffi
1f2p
is the damped natural frequency of the os-cillator [93,94]
Discrete data is sampled with time stepDtand useyn to denotethe measured response at timet nDt. Due to measurement noiseand modeling error, there is a difference between the measuredresponse yn and the model responsexnDt, referred to hereafteras prediction error. It is assumed that the prediction error can be
adequately represented by a discrete white noise process e withzero mean and variance r2e :
yn xnDt en; n 1; 2; ;N (3)
so this processe satisfiesE enen0 r2ednn0 , where dnn0 denotes the
Kronecker delta. Furthermore, the stochastic responsex and pre-diction errore are assumed statistically independent.
Therefore, the set of measurement D includes the data points,y1;y2; ;yN, and define a column vector: Y y1;y2; ;yN
T. It
follows that the likelihood function for a given set of data D isgiven by
pDjh; C 2pN2 Chj j
12exp
1
2YTCh1Y
(4)
whereCis the prescribed class of models governed by Eq.(1)with
the parameterization h X; f; Sf0;reT for the dynamical modelto be identified. The notation jAj is used to denote the determinantof a matrix A. The likelihood function pDjh; C is an N-variateGaussian distribution of the measurement vectorY with zero meanand covariance matrix Ch with the (n,n0) element Cn;n
0hEynyn0 is given by
Cn;n0h Rxnn
0Dtjh r2ednn0 (5)
where Rx is the autocorrelation function of the system responsefor a given model parameter vectorh and it is given by Eq.(2).
However, for a large number of observed data points, repeatedevaluations of the likelihood functionpDjh; Cfor different valuesofh become computationally prohibitive. This is obvious from Eq.(4) that it requires the computation of the solution X of the alge-braic equation ChX Y and the determinant of theNNmatrixCh. This task is computationally very expensive for large numberof data points N. Repeated evaluations of the likelihood functionfor thousands times in the optimization process is computationallyprohibitive even for a linear single-degree-of-freedom system.Therefore, the exact Bayesian approach described above, based ondirect use of the measured dataD, becomes practically infeasible.
3 Formulation of Model Updating Using Response
Measurement
Consider a linear dynamical system withNddegrees of freedom(DOFs) and its equation of motion is given by
MxC _xKx T0Ft (6)
whereM ,C and K are the mass, damping and stiffness matrices,respectively; T0 2 R
NdNF is a force distributing matrix; and theinput Ft 2 RNF can be measured or unmeasured and thatdepends on the application.
Discrete data of a system response quantity Q, e.g., accelera-tion, is taken at No (Nd) measured DOFs at times t nDt,n 1; 2; ;N. Also, due to measurement noise and modelingerror, there is prediction error, i.e., a difference between the meas-ured response yn2 R
No and the model response at time t nDtcorresponding to the measured degrees of freedom. The latter is
given byQnD
t. Therefore, the measured responseynat timenD
tcan be expressed as follows:
yn QnDt en (7)
It is assumed that the prediction error can be adequately repre-sented by a discrete zero-mean Gaussian white noise e with thefollowingNoNo covariance matrix:
E eneTn0
Rednn0 (8)
wherednn0 is the Kronecker delta. Furthermore, the prediction errore and the stochastic response Q are assumed statisticallyindependent.
Use h to denote the uncertain parameter vector that determinesthe dynamical model within a prescribed class of modelsC. Theseparameters include: (i) the structural parameters that determine
the model matrices M,C and K; (ii) the forcing parameters defin-ing the stochastic model of the input if it is treated as a stochasticprocess; (iii) the parameters defining the stochastic properties ofthe measurement noise. Herein, the identification of the model pa-rameter vectorh given some measured dataD is concerned. Usingthe Bayes theorem, the posterior/updated probability densityfunction (PDF) of the model parameter vectorh is given by
phjD; C j1phjCpDjh; C (9)
where j1 is the normalizing constant such that integrating theright hand side over the parameter space yields unity. The priorPDFphjCreflects the prior information of the parameters with-out using the data. The likelihood functionpDjh; Cis the domi-nant factor when the number of data points is large. It reflects thecontribution of the measured data D in establishing the updatedPDF of the parameters. The relative plausibility between two val-ues of h does not depend on the normalizing constant j1. Itdepends only on the relative values of the product of the priorPDFphjC and likelihood function pDjh; C. In order to searchfor the most probable model parameter vector h, denoted by h,one minimizes the objective function: Jh lnphjCpDjh; C, which is the negative logarithm of the product of theprior PDF and the likelihood function.
The reason for minimizing the objective function Jh instead ofmaximizing the posterior PDF directly is that the former has bettercomputational condition. It is found that the updated PDF of the pa-rameter vectorh can be well approximated by a Gaussian distributionwith meanh and covariance matrixHh1, whereHhdenotesthe Hessian of the objective function Jhcalculated at h h.
4 Bayesian Model Updating Using Response
Measurements
4.1 Input-Output Data
4.1.1 Noise-Free Input Measurement. In this section, amethod is introduced to utilize input-output measurement foridentification purpose. The input measurement is considered noisefree so the uncertain parameter vector h includes the model pa-rametershm and the parameters that determine the elements of theupper right triangular part of the prediction-error covariance ma-trix Re (symmetry defines the lower triangular part of this matrix).The dynamic data D consists of the measured time histories atNdiscrete time steps of the excitation and system response. Assumeequal variances and stochastic independence for the prediction
010802-2 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
3/18
errors of different channels of measurements so the covariancematrix for the prediction errors is Re r2e INo , where INo is theNoNo identity matrix. Then, the updated PDF of the uncertainparameters in h hTm;r
2e
Tgiven the data D and model class C
can be expressed as
phjD; C j2phjC2pNNo
2 rNNoe exp NNo
2r2eJghm;D;C
(10)
wherej2 is a normalizing constant and phjCis the prior PDF of
the uncertain parameters in h, expressing the users judgmentabout the relative plausibility of the values of the uncertain param-eters before the data are used. The goodness-of-fit function isgiven by [58,95,96]
Jghm;D; C 1
NNo
XNn1
ynQnDt; hm; C 2 (11)
whereQnDt; hm; C is the model response based on the assumedclass of models and the model parameter vectorhm whileynis themeasured response at timenDt. Furthermore,jj.jj denotes the Eu-clidean norm (2-norm) of a vector. The most probable model pa-rameter vectorhm is obtained by maximizing the posterior PDF inEq.(10). For large Nor with improper prior, this is equivalent tominimizing the goodness-of-fit function Jghm;D; C in Eq. (11)
over all possible values ofh
m:h
m arg
hm
minJghm;D; C (12)
If Jghm;D; C is known only implicitly, numerical optimizationis needed to search for the optimal model parameters and this canbe done by the function fminsearch in MATLAB. The most proba-ble value of the prediction-error variance in h can be obtainedalso by maximizing the posterior PDF and the solution is availablein closed form:
r2e minhm
Jghm;D; C minhm
Jghm;D; C (13)
For dynamic testing, it is the number of observed degrees of free-dom, No, and their distribution that are the essential factors foridentifiability. On the other hand, increasing the number data
points Ndoes not increase the number of effective mathematicalconstraints, and hence the identifiability.
For globally model-identifiable cases with large N [97,98], itturns out that the posterior PDFphjD; C is approximately Gaus-sian. The mean is the optimal parameter vectorh and the covari-ance matrix is equal to the inverse of the Hessian matrix of theobjective functionJh lnpDjh;CphjCat h:
Hl;l0h
@2
@hl@hl0lnpDjh; CphjCj
hh (14)
Since the uncertainty of the estimation can be quantified, the opti-mal sensor locations can be obtained using the information en-tropy [99] as a measure of the uncertainty of multivariate randomvariables [100104].
In practice, the prior distribution may be used as a regularizer
[105,106] to improve the well-posedness of the inverse problem:
phjC 2pNh
2
YNhl1
rl
!exp
1
2
XNhl1
h2lr2l
! (15)
It is a Gaussian PDF with zero mean so it decays as the radial dis-tance to the origin increases in any direction. If there exist two ormore sets of parameters to give the same likelihood function values,using this radially decaying prior distribution helps trimming downthe set of the optimal parameters to the one with the smallest 2-norm.
4.1.2 Consideration With Input Measurement Noise. In theabove method, it is assumed that there is no measurement error inthe input. Here, a method is introduced to consider explicitly the
measurement noise in both the input and output data. This featureavoids the possible underestimation on the parametric uncertaintyin practice. Let Ndenote the total number of observed time steps.Using the Bayes theorem, the updated PDF of the parameters hgiven the measured input G1; G2; ; GNof the excitation F andthe measured responsey1;y2;; yNis given by
phjy1; y2; ; yN; G1; G2; ; GN; C
j3phjG1; G2; ; GN; C
py1; y2; ; yNjG1; G2; ; GN; h;C (16)
where j3 is a normalizing constant. The probability distributionphjG1; G2; ; GN; C represents the information from the meas-ured input only. It can be approximated by the prior PDFphjG1; G2; ; GN; C phjC since the measured input alonedoes not have much saying on the model parameters (though itcontains some information of the prediction-error variance, e.g.,the root-mean-square (rms) of the measurement noise should beless than the rms of the measurement). The likelihood functionpy1; y2; ; yNjG1; G2; ; GN; h; C reflects the contribution ofthe measured datay1; y2; ; yNand G1; G2; ; GNin establishingthe updated PDF of the model parameters. Since py1; y2; ; yNjG1;G2; ; GN;h; Cis jointly Gaussian, direct calculation of thisPDF encounters similar computational problems as in the exactformulation shown in Sec. 2. Therefore, an approximated likeli-hood expansion is introduced [107]:
py1; y2; ; yNjG1; G2; ; GN; h; C
py1; y2; ; yNp jG1; G2; ; GN; h; C
YN
nNp1
pynjynNp1; ; yn1; G1; G2; ; GN; h; C (17)
The conditional PDFs are conditional on the last Np time stepsonly. In order to obtain the reduced-order joint PDF and the condi-tional PDFs, the differential equation of the state space model wasconverted to a difference equation. By considering the input mea-surement noise, the usually observed problem of underestimatingthe parametric uncertainty can be resolved. The reduced-orderlikelihood function py1; y2; ; yNp jG1; G2; ; GN; h; C and theconditional PDFs can be computed with the state space method
described in Ref. [107].In Ref. [108], the Bayesian unified approach was introduced to
handle the more general case with incomplete input and incom-plete output noisy measurements. The Bayesian unified methodopens a wide range of applications, including the special cases ofambient vibration surveys, and measured input-output noisy data.An application was presented for a building subjected to wind andground excitation simultaneously. The unmeasured wind pressurewas modeled as a stochastic process with uncertain parametersbut the ground excitation was observed with measurement noise.The structural response was also observed at a limited number ofDOFs only. The unified method was applied successfully formodel updating and damage detection purpose.
4.2 Output-Only Methods
4.2.1 Bayesian Time-Domain Approach. Difficulty here is toconstruct the likelihood pDjh; C because the random compo-nents of the data points are correlated, which is in contrast to thecase with input-output data. By using the Bayes theorem, thelikelihood function can be expanded to a product of conditionalPDFs and a reduced-order joint PDF [109]:
pDjh;C py1;y2;;yNp jh;CYN
nNp1
pynjy1;y2;;yn1;h;C
(18)
This expansion is exact but it does not resolve the computationaldifficulties encountered in the exact formulation. This is because
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-3
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
4/18
the computation of each of the conditional PDFs pynjy1;y2; ; yn1; h; Cfor largen requires similar computational effortas in the computation of the likelihood function pDjh; C in thedirect exact formulation and the right hand side of the expansioninvolves many such conditional PDFs. In order to overcome thiscomputational obstacle, the following approximation for the like-lihood function is introduced [110]:
pDjh; C py1; y2; ; yNp jh; C
YN
nNp1
pynjynNp ; ynNp1; ; yn1; h; C (19)
In other words, the conditional PDFs depending on more than Npprevious data points are approximated by the conditional PDFsdepending on only the last Np data points. The sense of thisapproximation is that data points belonging too far in the past donot have significant information on the system response of thepresent point. Of course, one expects this to be legitimate, ifNp isso large that the correlation functions have decayed to negligiblevalues. However, it is sufficient to use a value ofNp to include thedata points within one fundamental period of the oscillator. Inother words, the value ofNp is chosen to cover roughly one funda-mental period of the system. In the case of multidegree-of-freedomsystems (i.e., multimode systems), such a selected value ofNpcov-ers more than one period of the higher modes so the approximationis even more accurate for the higher modes. Although the funda-mental frequency X is an unknown parameter, it can be roughlyestimated from the first peak of the response spectrum to obtainthe value ofNp prior to the identification. The identification resultis not sensitive to the selected value ofNp if it is sufficiently large.
UseYn;n0 to denote the vector comprised of the response measure-ments from timenDtton0Dt(n n0) in a time-descending order:
Yn;n0 yTn0 ; y
Tn01; ; y
Tn
T; n n0 (20)
Then, the joint PDF py1; y2; ; yNp jh; C follows an NoNp -vari-ate Gaussian distribution with zero mean and covariance matrixRY1;Np :
RY1;Np EY1;Np YT1;Np
CNp;Np sym
..
. . ..
C1;Np C1;1
264
375 (21)
where the submatrixCn;n0 has dimensionNoNoand it is given by
Cn;n0 E ynyTn0
RQnDt; n
0Dt Rednn0 (22)
wherednn0 is the Kronecker Delta;RQ denotes the autocorrelationmatrix function of the model response Q; and Re is the noise co-variance matrix defined in Eq. (8). Since the random vectorsy1; y2;; yNp are jointly Gaussian with zero mean, the reduced-order likelihood function p
y
1; y
2; ; y
Np jh; C
is given by
py1; y2; ; yNp jh; C 2p
NoNp
2 jRY1;Np j1
2
exp 1
2YT1;NpR
1Y1;Np
Y1;Np
(23)
To compute this likelihood function, it involves only the solutionof the linear algebraic equation RY1;Np X Y1;Np and the determi-nant of the matrix RY1;Np which isNoNpNoNp only.
Then, the general expression is given for the conditional proba-bility densities involving Np previous points pynjynNp ; ynNp 1;; yn1; h; C in Eq. (19), with n> Np 1. First, the randomvectorYnNp;nhas zero mean and covariance matrix RYnNp ;n :
RYnNp ;n E YnNp;nYTnNp ;n
h i
Cn;n sym
..
. . ..
CnNp;n CnNp;nNp
264
375(24)
where the submatrixCn;n0 is given by Eq.(22). Then, this matrixis partitioned as follows:
RYnNp ;n R11;n R12;n
R
T
12;n R22;n" # (25)
where the matrices R11;n, R12;n and R22;n have dimensionsNoNo, NoNoNp and NoNpNoNp, respectively. R11;n andR22;nare the unconditional covariance matrix of the prediction tar-get variables and the conditioning variables; and the matrix R12;nquantifies their correlation.
Since the measured response has zero mean, the optimal esti-matorynofynconditional onYnNp;n1is given by [111]:
yn EbynjynNp ; ynNp1; ; yn1; h; Cc R12;nR122;nYnNp;n1
(26)
and the covariance matrix Re;n of the prediction erroren ynynis given by
Re;n EbeneTn jynNp ; ynNp1; ; yn1; h; Cc
R11;n R12;nR122;nR
T12;n (27)
Therefore, the conditional probability density pynjynNp ;ynNp1;; yn1; h; C follows an No-variate Gaussian distribu-tion with meanyn and covariance matrix Re;n:
pynjynNp ;ynNp1; ;yn1; h; C
2pNo
2 jRe;nj1
2 exp 1
2 ynyn T
R1e;n ynyn
(28)
The most probable parameter vectorh is obtained by minimizingthe objective function which is the negative logarithm of the pos-
terior PDF (without including the terms that do not depend onh
):
Jh lnphjC 1
2ln jRY1;Np j
1
2YT1;NpR
1Y1;Np
Y1;Np
1
2
XNnNp 1
ln jRe;nj ynyn T
R1e;n ynyn h i
(29)
In the special case of stationary response, the matrices Re;n Reand R12;nR
122;n R12R
122 required in the computation of the opti-
mal estimatoryn do not depend on the time step n. Therefore, theobjective function in Eq.(29)can be simplified as follows:
Jh lnphjC 1
2ln jRY1;Np j
1
2YT1;NpR
1Y1;Np
Y1;Np
NNp2
ln jRej 12
XNnNp 1
ynyn
TR
1e ynyn
(30)
In this case, it requires only the computation of the inverse and de-terminant of the matrices RY1;Np (R22;n) and Re, which areNoNpNoNp, and NoNo, respectively. These sizes are muchsmaller than the NNoNNo matrix RY1;N appeared in the directformulation.
4.2.2 Bayesian Spectral Density Approach. The Bayesianspectral density approach is a frequency-domain approach. First,consider a discrete stochastic vector processy and a finite number
010802-4 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
5/18
of discrete data points Dfyn, n1,2,,Ng. Based on D a dis-crete estimator of the spectral density matrix of the stochastic pro-cess y is introduced:
Sy;Nxk YxkYxkH
(31)
where H denotes the adjoint or conjugate transpose of a complexvector/matrix. The vectorYxk 2 C
No denotes the scaled discreteFourier transform of the vector processy at frequency xk:
Yxk ffiffiffiffiffiffiffiffiffiDt2pN
r XN1
n0
ynexp inxkDt (32)
where xk kDx, k 0; 1; ;Nnqy with Nnqy INTN=2,Dx 2p=T, andT NDt.
Taking expectation of Eq.(31)yields
ESy;Nxkjh; C ESQ;Nxkjh; C ESe;Nxkjh; C (33)
The matricesSQ;Nxkand Se;Nxkare defined in similar mannerto Eq.(31)and Eq.(32)for the vector processesQ and e, respec-tively. One can show that the first term on the right hand side hasthe following form [112]:
ESl;l0Q;Nxkjh; C
Dt
2pN
NR
l;l0Q 0
XN1n1
Nn
Rl;l0
Q nDteinxkDt Rl
0;lQ nDte
inxkDth i
(34)
where RQ is the correlation matrix function of the quantity Q.This estimatorES
l;l0Q;Nxkjh; Ccan be computed efficiently using
the function fft in Ref. [113]. For the measurement noise, sinceits correlation function is nonzero only with zero time lag, theterm ESe;Nxkjh; C can be readily obtained in terms of the co-variance matrix Re:
ESe;Nxkjh; C Dt
2pRe Se0 (35)
Assume that Ns sets of independent and identically distributed
time histories are available: D1
;D2
; ;DNs
. As T! 1 andDt! 0, the corresponding discrete Fourier transformsYsxk,s 1; 2; ;Ns, are independent and follow an identical complexNo-variate normal distribution with zero mean. Then, the averagedspectral density matrix estimator:
Savy;Nxk 1
Ns
XNss1
Ss
y;Nxk (36)
follows the central complex Wishart distribution of dimensionNowith Ns degrees of freedom [114]. Its PDF is given by
p Savy;Nxkjh;C p
NoNo12 N
NoNsNos S
avy;Nxk
NsNo
QNos1 Nss!
E Sy;Nxkjh;C Ns
exp Nstr E Sy;Nxkjh;C 1
Savy;Nxkn o
(37)
wherejAj and trfAgdenote the determinant and trace of the ma-trix A, respectively. Note that this PDF exists if and only ifNs No. Furthermore, it can be shown that when T! 1 andDt! 0, the random matrices Savy;Nxk and S
avy;Nxk0 are inde-
pendently complex Wishart distributed fork6k0 [115].Although the Wishart distribution and independence properties
are exact only as Dt! 0, it can be verified by simulation thatthey are good approximations in a particular frequency range. It
was verified using simulation that such approximations are indeedaccurate if an appropriately chosen bandwidth is considered. Thereasons for the violation of these approximations are aliasing andleakage. Therefore, the range of frequency for accurate approxima-tions is the region with large spectral values since the aliasing andleakage effects have relatively minor contribution in such fre-quency range. As a result, in the case of displacement measure-ment such range of frequencies corresponds to the lower frequencyrange. This range increases for higher levels of prediction error.
Based on the above discussion regarding the statistical proper-
ties of the averaged spectral estimator, the Bayesian spectral den-sity approach for updating the uncertain parameter vector h isgiven as follows. With Ns (No) independent sets of observeddata D(s), s 1; 2; ;Ns, the corresponding observed spectraldensity matrix estimatorsS
sy;Nxk, s 1; 2; ;Ns, k2 K, can be
obtained using Eq.(31)and Eq.(32). Then, the averaged spectraldensity matrix estimatorsSavy;Nxkis readily obtained by Eq.(36)to form the averaged spectral set: Sav
K fSavy;NkDxjk2 Kg.
Here, the frequency index set Krepresents a range over the Wish-art distribution and independence approximations are satisfactory.Using the Bayes theorem in a similar fashion as Eq. (9), theupdated PDF of the model parameter vectorh given the averagedspectral setSav
Kis given by
phjSavK; C j4phjCpS
avK
jh; C (38)
where j4 is a normalizing constant. The likelihood functionpSav
Kjh; Cexpresses the contribution of the measured data. Based
on the Wishart distribution and independence approximation, itcan be calculated as follows:
p SavK
jh; C
j5Yk2K
1
jESy;Nxkjh; CjNs
expNstrfESy;Nxk1
Savy;Nxkg (39)
where
j5p
NoNo1Nx2
YNos1
Ns s!
!NxNNo Ns No Nxs
Yk2K
Savy;Nxk Ns No
is a constant that does not depend on the uncertain parameters; Nxis the number of frequency points to be considered and it is equal
to the number of the distinct elements in the set K. For a given setof data, the constant j5 does not depend on the model parametersso it does not affect the optimal parameters and their associateuncertainties. Finally, the most probable parameter vector h isobtained by minimizing the objective function:
Jh lnphj C NsXk2K
ln jESy;Nxkjh;Cj
trfESy;Nxk1
Savy;Nxkg (40)
In the case if a noninformative prior is used, the first term can besimply neglected. Furthermore, the updated PDF of the parametervectorh can be well approximated by a Gaussian distribution withmean h and covariance matrix Hh1, where Hh denotesthe Hessian matrix of the objective function J calculated ath h:
Hl:l0h Ns
Xk2K
@2
@hl@hl0ln ESy;Nxkjh; C
trfESy;Nxkjh; C1
Savy;Nxkg
hh
@2 lnphjC
@hl@hl0
hh
(41)
The two approximations used in the Bayesian spectral densityapproach are accurate in a particular frequency range. It is
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-5
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
6/18
recommended to select the frequency index set to include only therange around the peaks in the spectrum even though the Chi-square/Wishart distribution and independence approximations areaccurate over a wider range. This selection enhances the computa-tional efficiency without sacrificing substantially the informationfor the frequency structure of the dynamical system (though itinduces loss of information for the prediction-error variance).Another advantage is that the results by this choice will rely lesson the whiteness assumption since it requires a flat spectral den-sity function for each mode only around the corresponding peak
instead of over the whole frequency range. Moreover, the aliasingand leakage effects generally have less influence on this rangesince their spectral values are large.
Another important advantage of this cutoff frequency range isas follows. In most existing probabilistic methods, the uncertaintyof the model parameters will tend to zero if the sampling timeinterval tends to zero with a fixed observed duration (even if it isvery short) as long as it is globally identifiable. This is the conse-quence of the discrete white noise assumption. Note that this phe-nomenon occurs even for filtered white noise, such as movingaverage or autoregressive output of white noise. However, for theBayesian spectral density method with this proposed cutoff fre-quency range, the sampling time interval does not affect the fre-quency index set so the same number of frequencies is consideredregardless of the sampling time interval (if it is sufficiently small).Therefore, the uncertainty of the model parameters estimated by
the Bayesian spectral density method will stabilize as the sam-pling time step tends to zero. This feature of the methodology isappealing for practical purposes. Note that this method can beextended for nonlinear systems [116].
4.2.3 Bayesian Fast Fourier Transform Approach. TheBayesian fast Fourier transform approach uses the statistical prop-erties of discrete Fourier transforms, instead of the spectral den-sity estimators, to construct the likelihood function and theupdated PDF of the model parameters [117]:
p YKjh; C k6Yk2K
1
2pNoffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
jCZxkjp
exp 1
2Zxk
TCZxk
1Zxk (42)
where YKdenotes the set of discrete Fourier transform of the measure-ment defined in Eq. (32)in the frequency range described by the fre-quency index set K; the vectorZxk ReYxk
T; ImYxkTT
is real and it includes the real and imaginary parts of the discrete Fou-rier transform; and the matrix CZxkis the covariance matrix of thevectorZxkand its expression is given by Ref. [117].
This method does not rely on the approximation of the Wishartdistributed spectrum and the individual Gaussian likelihood func-tion given in Eq.(42)is exact. The only approximation was madeon the independency of the discrete Fourier transforms at differentfrequencies. Therefore, the Bayesian fast Fourier transformapproach is more accurate than the spectral density approach inthe sense that one of the two approximations in the latter isreleased. However, since the fast Fourier transform approach con-
siders the real and imaginary parts of the discrete Fourier trans-form, the corresponding covariance matrices are 2No2No,instead ofNoNo in the spectral density approach. Therefore, thespectral density approach is computationally more efficient thanthe fast Fourier transform approach.
4.2.4 Unknown Input Without Assuming its StochasticModel. In the literature and the aforementioned identificationmethods, the input is either measured or modeled as a prescribedparametric stochastic model (even though the parameters may beunknown). This seems to be a necessary condition for model iden-tification purpose. For example, consider a linear single-degree-of-freedom system. In frequency domain, the response X is equal tothe input F, magnified by the transfer function of the oscillatorH:
Xx HxFx (43)
Therefore, if the time-frequency model of the input is completelyunknown, the output measurement does not have any saying onthe transfer function of the oscillator since there exists a set ofinput to match with the measured output:
Fx Xx=Hx (44)
In other words, identification of the model parameters is impossi-ble in this case since all model parameters (as long as the associ-
ated transfer function is nonzero) give the same likelihood(perfect match) to the measurement.
However, if two or more DOFs are measured in a multistorybuilding subjected to ground motion, the ratio between theresponses of different DOFs will be constrained by the prescribedclass of structural models, such as shear building models. In Ref.[118], a frequency-domain method for unknown input was pro-posed and it does not assume any time-frequency model for theinput. The method takes the advantage that when the number ofthe measured channels is larger than the number of independentexternal excitations, there are mathematical constraints among theresponses of different degrees of freedom. For example, when thebuilding is subjected to earthquake ground motion, the number ofindependent input is one. If the responses of two or more degreesof freedom are measured for this building, there is information to
infer the structural properties even though no assumption is madeon the time-frequency content of the excitation. Specifically, thedata is partitioned into two parts:
DA yA0 ; yA1 ; ; y
AN1
; DB yB0 ; y
B1 ; ; y
BN1
(45)
whereyAn andyBn ,n 0; 1; ;N1, are the measurements corre-
sponding to the firstNF and lastNoNF DOFs, respectively. ThenumberNF denotes the dimension of the inputF. Then, the likeli-hood function can be expanded by the Bayes theorem:
pDjh; C pDA;DBjh; C pDBjh;DA; CpDAjh; C (46)
It turns out that the likelihood function pDAjh; C does notdepend on h for this partitioning arrangement and the conditionallikelihood function pDBjh;DA; C can be constructed in the fre-
quency domain. The idea is that since the number of the independ-ent excitation components is smaller than the number of observeddegrees of freedom, one can develop easily the linear relationshipbetween the Fourier transform of QA (measured quantity corre-sponding toyA) andQB (measured quantity corresponding toyB):
QBx HBAxQAx (47)
where QAxand QBxare the discrete Fourier transform of the
measured quantify QA and QB; HBAx is the transfer functionfrom QA to QB and it can be obtained easily using the excitation-response transfer function. For details, please refer to Ref. [118].By using the relationship in Eq.(47), the likelihood function canbe obtained by treating the measurement in DA as the input:
p DAjh;DB; C
j7Yk2K
12pN0NF
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffijCxkj
pexp
1
2
ReYBxk lRImYBxk lI
" #T
Cxk1 ReY
Bxk lRImYBxk lI
" #! (48)
wherelR and lIare the means and Cxkis the covariance matrixand their expressions are given in Ref. [118]. However, cautionmust be made on the ill-posedness of the inverse problem sincethis relaxation of the assumption on the input stochastic model
010802-6 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
7/18
breaks the bonding between different frequencies and induces ahigher degree of ill conditioning of the problem. Therefore, theBayesian framework is important to indicate if it is globally iden-tifiable to avoid misleading results.
If the mathematical model for the concerned system has toomany uncertain parameters, the measurement will not provide suf-ficient mathematical constraints/equations to uniquely identify theuncertain parameters. However, experienced engineers can iden-tify the critical substructures for monitoring. Then, a free bodydiagram can be drawn to focus on these critical substructures
only. Note that the internal forces on the boundary of the substruc-tures are unknown and difficult to measure, so they are treated asuncertain input to the substructure [57,119]. Furthermore, theseinternal forces share the dominant frequencies of the structure sothey cannot be modeled arbitrarily as white noise or other pre-scribed colored noise. However, with the same idea as in[118,120,121], these interface forces can be treated as unknowninputs without assuming their time-frequency content [122]. Thisenables a large number of possible applications in structuralhealth monitoring and also enhances the computational efficiencysince one does not need to consider the whole system.
5 Bayesian Model Updating Using Modal
Measurements
This section introduces another type of Bayesian model updat-
ing techniques using modal measurements, i.e., modal frequenciesand mode shapes. For structural/mechanical dynamics problems,the eigenvalues are the squared modal frequencies: k m Xm2,m 1; 2;;Nd, and the eigenvectors,/
1;/2; ;/Nd, are themode shapes.
There are nonmodel based methods [123129] and model basedmethods [130134]. The method introduced here falls into thesecond category and most existing global structural health moni-toring methods use dynamical model updating to determine localloss of stiffness by minimizing a measure of the differencebetween the modal frequencies and mode shapes measured indynamic tests and those calculated from a finite-element model ofthe structure. The measured modal parameters are those estimatedfrom dynamic test data using some modal identification proce-dure. A generic form of the goodness-of-fit function to be mini-
mized is
Jgh XNmm1
wmkmh km2
XNmm1
w0m /mh /m
2(49)
where km and /m are the measured eigenvalue and eigenvectorof the mth mode; kmh and /mh are the eigenvalue andeigenvector of themth mode from the dynamical model with pa-rameter vector h that determines the stiffness and mass matrix;and wm and w
0m, m 1; 2; ;Nm, are chosen weightings that
depend on the specific method. One major difficulty is that modematching is required, i.e., it is necessary to determine whichmodel mode matches which measured mode. If only measure-ments of partial mode shapes are available, this will not be a triv-
ial task. Another major difficulty is that the Nm observed modes indynamic tests might not necessarily be the Nm lowest-frequencymodes in practice. In other words, some lower modes might notbe detected. For example, some torsional modes are not excited.Furthermore, in the case where there is damage in the structure,the order of the modes might switch because the local loss of stiff-ness from damage may affect some modal frequencies more thanothers, making the mode matching even more challenging [ 135].
Recently, methods have been proposed for solving this modelupdating problem which avoid mode matching [136,137]. This isaccomplished by employing the concept of system mode shapesthat are used to represent the actual mode shapes of the structuralsystem at all degrees of freedom corresponding to those of the dy-namical model, but they are distinct from the mode shapes of the
dynamical model, as will be seen more clearly later. Bayesianprobabilistic methods are then used to update the dynamicalmodel parameters and the system mode shapes based on the avail-able modal data. Furthermore, Rayleigh quotient frequencies,which are based on the dynamical model and the system modeshapes, are employed instead of the modal frequencies of the dy-namical model, so the eigenvalue problem needs never be solved.
As shown in previous research [136138], the realistic assump-tion is made that only the modal frequencies and partial modeshapes of some modes are measured; system mode shapes are also
introduced, which avoids mode matching between the measuredmodes and those of the dynamical model. The novel feature inthis work is that system frequencies are also introduced as param-eters to be identified in order to represent actual modal frequen-cies of the dynamical system (assuming that the dynamicalbehavior is well approximated by linear dynamics; otherwise theyshould be interpreted in the equivalent linear sense). The eigenequations of the dynamical model are used only in the prior proba-bility distribution to provide soft constraints. Furthermore, to cal-culate the most probable values of the model parameters based onthe modal data, an efficient iterative procedure is used thatinvolves a series of coupled linear optimization problems, ratherthan directly solving the challenging nonlinear optimization prob-lem by some general algorithms that may give convergence diffi-culties in the high-dimensional parameter space.
A class of dynamical models C is considered. It has a known
mass matrixM 2 RNdNd (which is assumed to be established withsufficient accuracy from the engineering drawings of the struc-ture) and the stiffness matrix K2 RNdNd is parameterized byh h1; h2; ; hNh
T 2RNh as follows:
Kh K0XNhl1
hlKl (50)
where the subsystem stiffness matrices Kl, l 0; 1; ;Nh, arespecified; e.g., by a finite-element model of the structure. Thescaling parameters in h allow the nominal model matrix given byh hg in Eq.(50)to be updated based on dynamic test data fromthe system.
Assume thatNm (Nd) modes of the system are measured (notnecessarily the first Nm lowest frequency modes), which have
eigenvalues km, m 1; 2;;Nm, and real eigenvector compo-nents/m 2RNd,m 1; 2; ;Nm. It is assumed that these modalparameters do not necessarily satisfy exactly the eigen equationwith any given dynamical model M; Kh because there arealways modeling approximations and errors. The quantities km
and/m are referred to as the system eigenvalue and eigenvectorof the mth mode to distinguish them from the correspondingmodal parameters given by any dynamical model specified by h.Given a parameter vector h, the model can be defined in modelclass C. The prior probability density function for k k1;k2;; kNmT and U /1T;/2T; ;/NmTT is chosen as
pk;Ujh; C j8exp 1
2Jg k;U;h
(51)
where j7 is a normalizing constant and the goodness-of-fit func-tion is given by
Jg k;/;h
Khk1M/1
Khk2M/2
..
.
KhkNmM/Nm
2666664
3777775
T
R1eq
Khk1M/1
Khk2M/2
..
.
KhkNmM/Nm
2666664
3777775
(52)
The uncertainty in the equation errors for each mode are modeledas independent and identically distributed, so Req r
2eqINdNm ; r
2eq
is a prescribed equation-error variance. The usage of this variance
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-7
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
8/18
parameter allows for explicit treatment of modeling error as theparametric models for the stiffness matrix, and hence, the eigenequation, is never exact in practice. If this error level can be esti-mated, the mathematical constraint given by the eigen equationwill become a soft constraint instead of a rigid constraint. In otherwords, errors of the eigen equation in the level corresponding toreq is allowed. The value ofr
2eq may be chosen to be very small
so that the eigen equations are nearly satisfied. This means thatthe system modal frequencies and mode shapes will correspondclosely to modal parameters of the identified dynamical model.
For modal data from a real structure, this would be a reasonablestrategy to start with. If the measured modal parameters did notagree well with those corresponding to the identified (most proba-ble) dynamical model, implying considerable modeling errors,thenr2eqcould be increased. This procedure allows explicit controlof the inherent trade-off between how well the measured modalparameters are matched and how well the eigen equations of theidentified dynamical model are satisfied. This additional modelingflexibility is an appealing feature of this method.
The prior PDF pk;Ujh; C implies that, given a class of dy-namical models and before using the dynamic test data, the mostprobable values ofk and U are those that minimize the Euclideannorm (2-norm) of the error in the eigen equation for the dynamicalmodel. This implies that the prior most probable values ofk andU are the squared modal frequencies and mode shapes of a dy-namical model, but these values are never explicitly required.
This prior PDF will have multiple peaks because there is noimplied ordering of the modes here.
The prior PDF for all the unknown parameters is given by
pk;U; hjC pk;Ujh; CphjC (53)
where the prior PDF phjC can be taken as a Gaussian distribu-tion with mean hg representing the nominal values of the modelparameters and with covariance matrix Rh. For example, the priorcovariance matrix Rh can be taken as diagonal with large varian-ces, giving virtually a noninformative prior.
To construct the likelihood function, the measurement errore isintroduced:
k
W" #
k
LoU e (54)
and a Gaussian probability model is chosen fore 2 RNmNo 1 withzero mean and covariance matrix Re, which can be obtained byBayesian modal identification methods (such as the ones intro-duced in the previous sections); W u1T; u2T;; uNmTT
and k k1;k2; ;kNmT, where um 2RNo gives theobserved components of the system eigenvector of themth modeand km gives the corresponding observed system eigenvaluefrom dynamic test data. Finally,Lo is an NmNoNdNm observa-tion matrix of 1s or 0s that picks the components ofU corre-sponding to the No measured degrees of freedom. The likelihoodfunction is therefore
pk; Wjk;U; h; C pk; Wjk;U (55)
is a Gaussian distribution with meankT; LoU TT and covariance
matrix Re.The posterior PDF for the unknown parameters is given by the
Bayes theorem:
pk;U; hjk; W; C j9pk; Wjk;U; h; Cpk;Ujh; CphjC
j9pk; Wjk;Upk;Ujh; CphjC (56)
The most probable values of the unknown parameters can befound by maximizing this PDF. To proceed, the objective functionis defined as Ref. [139]:
Jk;U; h 1
2 hhg TR1h hh
g
1
2r2eq
XNmm1
kKh kmM/mk2
1
2
kk
W LoU
" #TR
1e
kk
W LoU
" # (57)
which is the negative logarithm of the posterior PDF without
including the constant that does not depend on the uncertain pa-rameters. Here, jj.jj denotes the Euclidean norm. Then, the func-tion Jk;U; h is minimized instead of maximizing the posteriorPDF. This objective function is not quadratic for the uncertain pa-rameters. However, this function is quadratic for any of the uncer-tain parameter vector of k, U or h if the other two are fixed.Therefore, the original nonlinear optimization problems can bedone iteratively through a sequence of linear optimizationproblems.
The mode shapes are usually measured/identified with incom-plete components, i.e., with missing DOFs but the modal frequen-cies are measured with relatively high accuracy. Therefore, thesequence of optimization starts from computing the missing com-ponents of the mode shapes. First set the updated model parame-ters at their nominal values (h hg) and the eigenvalues at their
measured values (k
k). Then, perform a sequence of iterationscomprised of the following linear optimization problems:
U argmin
U
Jk;U; h; k argmink
Jk;U; h;
h argmin
h
Jk;U; h
until the prescribed convergence criteria is satisfied and thesethree optimization problems is explained in more detail in thefollowing.
By minimizing the objective function Jk;U; h in Eq. (57)with respect to U, the optimal vectorU can be obtained:
U r2eq
k1MK2 0
. ..
0 kNmMK2
2664
3775 LTo R1e 22Lo
2664
3775
1
LTo R1e
21
kk
R1e
22W
h i (58)
where R1e
21 and R1e
22
are the NmNoNm left bottom andNmNoNmNo right bottom sub-matrices ofR
1e ; and the updated
stiffness matrixK Kh.
By minimizing the objective function Jk;U; h in Eq.(57) with respect to k, the updated parameter vector k isgiven by
010802-8 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
9/18
k r2eq
/1TM2/1 0
. ..
0 /NmTM2/Nm
2664
3775 R1e 11
2664
3775
1
r2eq
/1TMK/1
/2TMK/2
..
.
/NmTMK/Nm
2666664
3777775 R
1e
11k R1e
12
WLo/
0BBBBB@
1CCCCCA
(59)
where R
1
e
11 and R
1
e
12 are the NmNm left top andNmNmNoright top sub-matrices ofR1e .
By minimizing Eq. (57) with respect to h, the updated modelparameter vectorh is given by
h r2eqG
ThGh R
1h
1
r2eqGTh
k1MK0/1
..
.
kNmMK0/Nm
2664
3775 R1h hg
0BB@
1CCA (60)
where the matrixGh is given by
Gh
K1/1 K2/
1 KNh/1
K1/2 K2/2 KNh/2
..
. ... . .
. ...
K1/Nm K2/
Nm KNh/Nm
2666437775
NdNmNh
(61)
The iterative procedure consists of the following steps:
(1) Take the initial values of the model parameters as the nomi-nal values: h hg, and the eigenvalues as the measuredvalues:k k. Then,K Kh.
(2) Update the estimates of the system eigenvectors /m,m 1; 2; ;Nm, using Eq.(58).
(3) Update the estimates of the system eigenvalues (squaredmodal frequencies)km,m 1; 2; ;Nm, using Eq.(59).
(4) Update the estimates of the model parameters h by using
Eq.(60)(5) Iterate Steps 2, 3 and 4 until the model parameters inh sat-isfy some convergence criterion, thereby giving the mostprobable values of the model parameters based on themodal data.
The posterior PDF in Eq.(56) can be well approximated by aGaussian distribution centered at the optimal (most probable) pa-rametersk;U; h and with covariance matrix Rk;U; h equalto the inverse of the Hessian of the objective functionJk;U; h lnpk;U; hjk; W; C calculated at the optimal pa-rameters. For details, the expression can be found in Ref. [87].
6 Model Complexity, Sensitivity, Data Fitting
Capability and Robustness
This section discusses the relationship among complexity, out-put sensitivity, data capability and robustness of identificationmodel class. This is a very important issue because it plays amajor role on the quality of identification results. Influenced bythe mind of forward modeling, it is easily directed to adopt com-plicated model classes in order to capture various physical behav-ior/mechanism. However, the more complex the model class, themore uncertain parameters are usually introduced unless extramathematical constraints of these parameters are imposed. In theformer case, the model output may not necessarily be accurateeven if the model well characterizes the underlying phenomenon/system since the combination of the many small errors from eachuncertain parameter can induce large output error. Furthermore, it
may lead to an unidentifiable model so the uncertain parameterscannot be uniquely determined. In the latter case, it is inevitablethat the extra constraints/approximations induce substantial errors.Therefore, it is important to use a proper model class for paramet-ric identification purpose. For example, consider a detailed identi-fication model with thousands of degrees of freedom for ahighrise building. For such a model, there will be a large numberof structural elements (beams, columns, plates, ) associatedwith uncertain mass and stiffness parameters. It is obvious that themodel is unidentifiable unless a huge number of sensors are usedto give sufficient constraints to the identification problem. Oneway to resolve the identifiability problem is to enforce extra con-straints so as to reduce the number of uncertain parameters, e.g.,by assuming rigid floor or grouping the stiffness parameters witha prescribed proportion. However, such simplification/approxima-tion will induce modeling error to the identification model.
It is clear that a suitable model class for identification should becapable to fit the data well and to provide sufficient robustness tomodeling error and measurement noise [140]. Now, an example isused to illustrate that there is no definite relationship between pos-terior uncertainty and data fitting capability. Consider an exampleof two single-parameter model classes with likelihood functionssatisfying the following relationship for the same set of measure-ment: pDjh1 h; C1 2pDjh2 h; C2 so lnpDjh1 h; C1lnpDjh2 h; C2 ln 2. Here,D denotes the measurement;h1and h2 are the uncertain parameter for model class C1 and C2,respectively. With the same prior distribution, the posterior distri-butions and uncertainty of the parameters in these two modelclasses are identical since the difference between the log-likeli-hood functions is a constant (Eq.(14)). However, model class C1fits the data better as its maximum likelihood value is twice of
that forC 2: pDjh1; C1 2pDjh2;C2, where h1 and h2 are theoptimal parameters for the two model classes. Therefore, the samelevel of posterior uncertainty is associated with different levels ofdata fitting in this case. This example demonstrates that the degreeof data fitting has no direct relationship with the posterior uncer-tainty of the parameters.
Posterior uncertainty is a measure of the spread of the posteriordistribution, which is proportional to the product of the prior dis-tribution and the likelihood function. According to Ref. [87],small posterior uncertainty is possible to associate with poor datafitting. On the other hand, sensitivity can be defined as the changeof the model output due to unit parameter perturbation. The slopeof the posterior PDF depends on the sensitivity, which controlsthe rate of the change of the model output due to perturbation ofthe parameters. On the other hand, posterior uncertainty of the pa-
rameters is controlled by the decaying rate (slope) of the posteriorPDF in the neighborhood around the optimal point. Therefore, itis particularly important to investigate the sensitivity of a modelaround the optimal parameters. If a model class is utilized forfuture prediction, it is desirable to obtain a robust model class thatfits the data well even with error in the parameters. In this case,the maximum likelihood value has to be large and the likelihoodvalue remains large in a sufficiently large region. In other words,the topology of the likelihood function around the maximum isreasonably flat (i.e., the output sensitivity is low).
Figure1 shows schematically two likelihood functions to dem-onstrate this relationship. The two likelihood functions areassumed to be obtained with the same set of measurement butwith different model classes. Model classC1 obviously has larger
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-9
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
10/18
maximum likelihood value than C2, indicating that the optimalmodel in C 1 fits the data better than that ofC2. Furthermore, theposterior uncertainty of the parameter inC1 is smaller than that of
C2 if the same prior distribution is used. However, if the optimalmodels are used for future prediction, model classC2 is more reli-able since it has a larger parameter range that provides satisfactoryfitting to the data. Therefore, mode class C2 is more robust. Onthe other hand, the maximum likelihood margin of C1 may beinsufficient to compensate if the model parameter has tiny error.Therefore, a reliable model class should have reasonable balancebetween data fitting capability and robustness to modeling error.This example also demonstrates that even for the same number ofparameters, a model class with larger maximum likelihood valueis not necessarily more suitable for parametric identification.
Table1 shows the four combinations of data fitting capabilityand posterior uncertainty. It is desirable to use a model class withhigh maximum likelihood value and large posterior uncertainty sothat efficient and robust identification performance can beexpected. However, this is difficult to achieve. Although data fit-ting capability can be enhanced by adding free parameters, thiswill inevitably degrade the robustness of a model class. Therefore,a suitable model class should possess reasonable balance betweenthese two properties. It should be noted that the complexity of amodel class depends not only on the number of its adjustable pa-rameters but also the specific structure of the solution space.
One possible way to rank among prescribed model classes isthe Bayesian model class selection approach, which utilizes themeasurement, D, from the underlying system. Since probabilitymay be interpreted as a measure of plausibility based on specifiedinformation [141], the posterior plausibility of a model classCjisgiven by [142144]:
PCjjD pDjCjPCj
PNCj1pDjCjPCj
; j 1; 2; ;NC (62)
where pDjCj is the evidence for model class Cj, PCj is theprior plausibility ofCjand the denominator is a normalizing con-stant such that the total posterior plausibility is normalized to be
unity:PNC
j1PCjjD 1. In most applications, uniform priorplausibility can be taken so the model class selection will relysolely on the measurement, i.e.,PCj 1=NC,j 1; ;NC.
Usehj2 RNj to denote the uncertain parameter vector in the pa-
rameter space Hj for model class Cj, where Nj is the number ofuncertain parameters. By the theorem of total probability, the evi-dence ofCjis given by:
pDjCj
Hj
pDjhj;CjphjjCjdhj (63)
wherepDjhj; Cjis the likelihood function of model class CjandphjjCj is the prior distribution of the uncertain parameters. For alarge number of data points, the integrand in the evidence integralin Eq. (63) can be well approximated as an unnormalized Gaus-sian distribution so this evidence integral can be approximated bythe Laplaces asymptotic approximation [145]:
pDjCj pDjh
j; CjFj (64)
wherepDjhj; Cjis the likelihood function of model class CjandphjjCj is the prior distribution of the uncertain parameters; h
j is
the optimal parameter vector that maximizes the integrand in Eq.(63)and the Ockham factorFjrepresents the penalty against com-plicated parameterization [143,146]:
Fj phjjCj2pNj=2jHjhjj1=2 (65)
The matrix Hjh
j is the Hessian matrix of objective function lnpDjhj; CjphjjCj evaluated at hj h
j . By Eq. (64), the
evidence of a model class represents the competition between datafitting capability (measured by the maximum likelihood valuepDjhj; Cj) and the robustness of the model class (measured bythe Ockham factor). Hence, the most plausible model class is theone that possesses the best trade-off between the data fitting effi-ciency and robustness to modeling error. This approach has beenapplied to a number of engineering problems, including architec-ture of neural networks [105,147], damage detection [148], esti-mation of fault rupture extent [149]; identification for hystereticstructural dynamical models [150], air quality modeling [151],soil compressibility relationship [152], biomedical engineering
[153], environmental effects to structural modal parameters [154],structural integrity monitoring [155], crack detection [156,157],and seismic attenuation relationship [158]. For more general casesin which the Gaussian approximation is inappropriate, one mayuse the transitional Markov chain Monte Carlo simulation method[159] to evaluate directly the evidence integral in Eq. (63).
Next, we examine the Bayesian model class selection approachaccordingly. By Eq. (63), it is clear that the evidence integral isthe standard inner product of the prior distribution and the likeli-hood function. In order to possess a high value of the evidence, agood model class should have large likelihood value over a rea-sonably large region, which overlaps with the significant region ofthe prior distribution. Furthermore, the prior distribution valuesare important even if a uniform distribution is used and this is incontrast to parametric identification. In parametric identification,
noninformative prior distribution is often used to let the identifica-tion rely solely on the likelihood of the data. As long as it is suffi-ciently flat and covers the significant region of the likelihoodfunction, the identification results will virtually be independent tothe choice of the prior distribution. Therefore, it is popular to evenabsorb the prior distribution into the normalizing constant toobtain the maximum likelihood solution. However, this is not thecase for model class selection. For example, doubling the range ofa sufficiently wide uniform prior distribution will reduce its proba-bility density value by half. This will not affect the parametricidentification results but the evidence will be half. If the user has alot of experience of a particular model class, the prior informationallows the use of a more concentrated prior distribution of theuncertain parameters. This leads to a larger value of the evidence
Table 1 Likelihood, posterior uncertainty and robustness
Maximum likelihood value High Low
Large posterior uncertainty Efficient and robust Poor modelingSmall posterior uncertainty Efficient but fragile Poor modeling
Fig. 1 Schematic plot of two likelihood functions
010802-10 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
11/18
8/9/2019 Bayesian model updatingReview Paper
12/18
hence the structural response, is considered as a nonstationary sto-chastic process, especially when the typhoon is close to the city[166,167]. As shown on the right tails of the figure, the typhoon-effect was mitigated after the landfall was made and the majorwind direction was recovered to the background prevalent off-shore wind direction.
Acceleration time histories of the East Asia Hall were recordedon the 18th story and its direction is shown in Fig.2. The durationof the measurement was 43hs which covered the whole durationof Hagupit, with sampling frequency 500 Hz. Fig. 5 shows theroot-mean-square acceleration and the ten minute-average windspeed observed in Macao by the Meteorological and Geophysical
Bureau of the Macao government. The rectangular window enclo-ses the region in which the wind speed exceeded 40 km/hr and thetwo arrows indicate the peaks of the wind speed and structuralresponse. It is observed that the structural acceleration magnitudeon the right boundary of the high-wind-speed window was about
twice of the left boundary, but they are associated with the samewind speed (40 km/hr). Moreover, the arrows indicate that thepeaks of the RMS acceleration and wind speed did not occur atthe same time. This is caused by the wind attacking angle of thetyphoon. From Fig.4, the wind direction changed gradually fromnorth to east and the drag coefficient was increasing in this pro-cess. Under the same wind speed, larger resultant force was cre-ated in the later stage so the corresponding structural responsewas larger. It turned out that the peak of the structural responseoccurred 7 hs after the peak of the wind speed. When the maxi-mum response was achieved, the wind speed was 20% lower thanits maximum value.
In the present study, the modal parameters of the structure andthe excitation are identified using the Bayesian spectral densityapproach for each of the five-minute records in order to trace theirfluctuation. The spectral density estimators are used up to 1.6 Hz,which is roughly the middle of the first two peaks in the spectrum.Figure 6 shows the identification results and there are threecurves. The middle one shows the identified modal frequencyusing each of the five-minute acceleration record. The identifiedmodal frequency is referred to the equivalent modal frequencysince the building may not behave perfectly linear, especiallyunder the severe wind pressure. The intervals between the othertwo curves are the confidence intervals with probability 99%. It isstatistically evident that there was notable reduction of the modalfrequency in the high wind speed region and the maximum esti-mated reduction was 6%. Note that this can be concluded because
the uncertainty of the identified values can be quantified using theBayesian method. Such reduction was recovered almost immedi-ately after the typhoon was dissipated and the ambient conditions(temperature and humidity) went back to normal situation [168].
The Bayesian fast Fourier transfer approach is also applied tothe same set of measurements. It is not surprising that the results,including the identified values and the uncertainty estimates, arevirtually the same as those by the Bayesian spectral densityapproach. It is because these two frequency-domain methods uti-lize the same approximation of independency in the same fre-quency range. The only difference is that the Bayesian fastFourier transform approach considers the exact covariance matri-ces of the fast Fourier transforms while the Bayesian spectral den-sity approach further approximates the structure of these matrices
Fig. 4 Wind speed and wind direction
Fig. 5 RMS acceleration and mean wind speed
010802-12 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
13/18
to obtain the Wishart distribution. However, this approximation issufficiently accurate, especially in the frequency range considered,so the results are practically the same. On the other hand, it shouldbe noted that the computational effort required by the Bayesianfast Fourier transform approach is significantly higher than theBayesian spectral density approach because the size of the matri-ces and vectors involved in the former is twice of the latter.
Furthermore, the Bayesian time-domain approach is alsoapplied to analyze the same set of measurements. As in the Bayes-ian spectral density approach, the measurement is used up to 1.6Hz and the response in each five-minute time intervals is assumedstationary. The truncation variableNpin Eq.(19)is taken as 50 sothat the conditioning data points cover slightly more than one fun-damental period of this structure. Figure 7 shows the identified
Fig. 6 Identified values of the modal frequency using Bayesian spectral densityapproach
Fig. 7 Identified values of the modal frequency using Bayesian time-domain approach
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-13
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
14/18
modal frequencies and the associated confidence intervals in thesame way as Fig.6. It turns out that there are some discrepancies ofthe identified values from the 28th to 32nd hour between the time-domain and frequency-domain approaches. It is believed that theresults obtained by the latter are more reliable as these frequency-domain methods are more robust to modeling error, such as linearityof structures, probability distribution of the input and measurementnoise, stationarity of the response, etc. Moreover, the identified val-ues by the Bayesian time-domain approach are substantially more
fluctuating in consecutive five-minute intervals during the aforemen-tioned five hours. On the other hand, the confidence intervalsobtained by the Bayesian time-domain approach are significantlynarrower than the frequency-domain approaches because the formerutilizes all information of the data while the frequency-domainapproaches use only a much narrower frequency band. However,one should note that these uncertainty estimates do not take intoaccount the bias induced by the violation of assumptions and thiswill be an important subject for further study.
Fig. 8 Identified values of the modal spectral intensity using Bayesian spectral densityapproach
Fig. 9 Identified values of the modal frequency versus spectral intensity
010802-14 / Vol. 64, JANUARY 2011 Transactions of the ASME
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
8/9/2019 Bayesian model updatingReview Paper
15/18
Next, we investigate the modal force spectral intensity and itsrelationship with the fundamental frequency of the structure. Theresults presented below are based on those obtained by the Bayes-ian spectral density approach but similar conclusion can be madeusing the Bayesian time-domain approach. Figure 8 shows thecorresponding spectral intensity of the modal force. Comparedwith the wind speed shown in Fig. 5, it is observed that the spec-tral intensity had a similar trend with the root-mean-square of thestructural response but not the wind speed. In Fig.5, the peak ofthe wind speed occurred at approximately the 19th hour but the
peak of the structural acceleration occurred approximately 7 hslater. This confirms that not only the wind speed but also the winddirection affects the magnitude of the structural response. It canbe further verified by the fact that the peak of structural response(Fig.5) and the spectral intensity (Fig. 8) occurred almost at thesame time (approximately the 26th hour).
Figure9shows the modal frequency versus the spectral inten-sity of the modal force in the semilogarithmic scale. The modalfrequency decreased as the spectral intensity of the modal forceincreased. In other words, the structural stiffness was degradedwhen the excitation was at high level. However, there is no evi-dence for structural damage since such loss was recovered afterthe typhoon was dissipated. One possible explanation is that thestructure went through nonlinear or even hysteretic behavior sothe identified modal frequency of the equivalent linear system wasreduced [169172].
8 Conclusion
In this paper, some recently developed Bayesian model updat-ing methods were introduced. Bayesian inference is useful foruncertainty quantification of identification problems. The intro-duced methods can be categorized into two types: using responsemeasurements and using modal measurements. The Bayesiantime-domain approach utilizes an approximated probability den-sity expansion and it can handle both stationary and nonstationaryresponse measurement. The Bayesian spectral density approachand Bayesian fast Fourier transform approach are frequency-do-main methods. They make use of the statistical properties of dis-crete Fourier transform and they can handle stationary response oflinear and nonlinear systems. Another model updating methodthat uses modal measurements was also introduced. An importantfeature of this approach is that it does not proceed with the com-monly required mode matching. This is important as the measuredmodes are not necessarily the lowest ones and their order may beunknown. An application of a 22-story building subjected tosevere typhoon was presented. Reduction of the modal frequencywas observed when the wind speed was over 40 km/hr. The com-parison could be made because the uncertainty level of the estima-tion could be quantified. The Bayesian methodology is necessaryin this type of application which requires comparison of the iden-tified parameters. Finally, comments were made on the suitabilityof model class for identification.
Acknowledgment
The authors would like to gratefully acknowledge the generoussupport by the Research Committee of University of Macau underGrant No. RG059/09-10S/YKV/FST.
References[1] Eykhoff, P., 1974, System IdentificationParameter and State Estimation,
Wiley, New York.[2] Kozin, F., and Natke, H. G., 1986, System Identification Techniques,Struc-
tural Safety,3, pp. 269316.[3] Natke, H. G., 1988, Updating Computational Models in the Frequency Do-
main Based on Measured Data: A Survey, Probab. Eng. Mech., 3(1), pp. 2835.
[4] Mottershead, J. E., and Friswell, M. I., 1993, Model Updating in StructuralDynamics: A Survey,J. Sound Vib.,167(2), pp. 347375.
[5] Ghanem, R., and Shinozuka, M., 1995, Structural-System Identification. I:Theory,J. Eng. Mech.,121(2), pp. 255264.
[6] Peeters, B., and De Roeck, G., 2001, Stochastic System Identification forOperational Modal Analysis: A Review, J. Dyn. Syst., Meas., Control,123(4), pp. 659667.
[7] Hemez, F. M., and Doebling, S. W., 2001, Review and Assessment of ModelUpdating for Nonlinear, Transient Dynamics, Mech. Syst. Signal Process.,15(1), pp. 4574.
[8] Chang, P. C., Flatau, A., and Liu, S. C., 2003, Review Paper: Health Monitor-ing of Civil Infrastructure,Struct. Health Monit.,2(3), pp. 257267.
[9] Sohn, H., Farrar, C. R., Hemez, F. M., Shunk, D. D., Stinemates, D. W., andNadler, B. R., 2003, A Review of Structural Health Monitoring Literature:1996-2001, Los Alamos National Laboratory Report No. LA-13976-MS.
[10] Johnson, E. A., Lam, H. F., Katafygiotis, L. S., and Beck, J. L., 2004, Phase IIASC-ASCE Structural Health Monitoring Benchmark Problem Using Simu-lated Data,J. Eng. Mech.,130(1), pp. 315.
[11] Kerschen, G., Worden, K., Vakakis, A. F., and Golinval, J. C., 2006, Past,Present and Future of Nonlinear System Identification in Structural Dynami-cs,Mech. Syst. Signal Process.,20(3), pp. 505592.
[12] Lopez, I., and Sarigul-Klijn, N., 2010, A Review of Uncertainty in Flight Ve-hicle Structural Damage Monitoring, Diagnosis and Control: Challenges andOpportunities,Prog. Aerosp. Sci.,46(7), pp. 247273.
[13] Brown, D. L, Allemang, R. J., Zimmerman, R., and Mergeay, M., 1979,Parameters Estimation Techniques for Modal Analysis, Society of Automo-tive Engineers, Technical Paper No. 790,221.
[14] Doebling, S.W., 1997, An Overview of Modal-Based Damage IdentificationMethods, Los Alamos National Laboratory, Los Alamos, NM.
[15] Ren, W. X., and De Roeck, G., 2002, Structural Damage Identification UsingModal Data. II: Test Verification,J. Struct. Eng.,128(1), pp. 96104.
[16] Ivanovc, S. S., Trifunac, M. D., and Todorovska, M. I., 2000, AmbientVibration Tests of Structures - A Review, ISET J. Earthquake Technol.,37(4), pp. 165197.
[17] Farrar, C. R., Doebling, S. W., and Nix, D. A., 2001, Vibration-Based Struc-tural Damage Identification, Philos. Trans. R. Soc. London, Ser. A,359(1778), pp. 131149.
[18] Brownjohn, J. M. W., 2003, Ambient Vibration Studies for System Identifica-tion of Tall Buildings,Earthquake Eng. Struct. Dyn.,32(1), pp. 7195.
[19] Koakowski, P., 2007, Structural Health MonitoringA Review With the Em-phasis on Low-Frequency Methods, Eng. Trans.,55(3), pp. 137.
[20] Young, P. C., 1970, An Instrumental Variable Method for Real-Time Identifi-cation of a Noisy Process,Automatica,6(2), pp. 271287.
[21] Juang, J. N., and Pappa, R. S., 1985, An Eigensystem Realization Algorithmfor Modal Parameter Identification and Model Reduction,J. Guid. ControlDyn. (AIAA),8, pp. 620627.
[22] Asmussen, J. C., Ibrahim, S. R., and Brincker, R, 1997, Application of VectorTriggering Random Decrement, Proceedings of 15th International ModalAnalysis Conference, Orlando, FL, Vol. 2, pp. 11651171.
[23] Worden, K., 1997, Structural Damage Detection Using a Novelty Measure,J. Sound Vib.,201(1), pp. 85101.
[24] Caicedo, J. M., Dyke, S. J., and Johnson, E. A., 2004, Natural ExcitationTechnique and Eigensystem Realization Algorithm for Phase I of the IASC-
ASCE Benchmark Problem: Simulated Data, J. Eng. Mech., 130(1), pp.4960.
[25] Beck, J. L., May, B. S., and Polidori, D. C., 1994, Determination of ModalParameters From Ambient Vibration Data for Structural Health Monitoring,Proceedings of 1
stWorld Conference on Structural Control, Pasadena, CA, pp.
TA3: 312.[26] Kim, H. M., Vanhorn, D. A., and Doiron, H. H., 1994, Free-Decay Time-Do-
main Modal Identification for Large Space Structures, J. Guid. Control Dyn.(AIAA),17(3), pp. 513519.
[27] Gersch, W., Taoka, G. T., and Liu, R., 1976, Structural System ParameterEstimation by Two-Stage Least Squares Method, J. Eng. Mech. , 102(5), pp.883899.
[28] Conte, J. P., and Kumar, S., 1996, Statistical System Identification of Struc-tures Using ARMA Models, Proceedings of 7th ASCE Specialty Conferenceon Probabilistic Mechanics and Structural Reliability, Worcester, MA, pp.142145.
[29] Moore, S. M., Lai, J. C. S, and Shankar, K., 2007, ARMAX Modal ParameterIdentification in the Presence of Unmeasured ExcitationI: Theoretical Back-ground,Mech. Syst. Signal Process.,21(4), pp. 16011615.
[30] Goodwin, G. C., and Sin, K. S., 1984,Adaptive Filtering Prediction and Con-
trol, Prentice-Hall, Englewood Cliffs, NJ.[31] Ljung, L., 1987, System Identification: Theory for the User (Prentice-Hall,
Englewood Cliffs, NJ, 1987).[32] Soderstrom, T., and Stoica, P., 1989,System Identification, Prentice-Hall, Eng-
lewood Cliffs, NJ.[33] Beck, J. L., 1978, Determining Models of Structures from Earthquake
Records, Earthquake Engineering Research Laboratory, California Instituteof Technology, Pasadena, CA, Technical Report No. EERL 78-01.
[34] Hoshiya, M., and Saito, E., 1984, Structural Identification by Extended Kal-
man Filter,J. Eng. Mech.,110(12), pp. 17571770.[35] Hoshiya, M., 1988, Application on Extended Kalman Filter-WGI Method in
Dynamic System Identification, Stochastic Structural Dynamics (Progress inTheory and Application), Elsevier, New York, pp. 103124.
[36] Juang, J. N., and Suzuki, H., 1988, An Eigensystem Realization Algorithm inFrequency Domain for Modal Parameter Identification, ASME J. Vib.,Acoust., Stress, Reliab. Des.,110(1), pp, 2429.
Applied Mechanics Reviews JANUARY 2011, Vol. 64 / 010802-15
Downloaded 29 Sep 2011 to 161.64.70.118. Redistribution subject to ASME license or copyright; see http://www.asme.org/terms/Terms_Use.cfm
http://dx.doi.org/10.1016/0167-4730(86)90006-8http://dx.doi.org/10.1016/0167-4730(86)90006-8http://dx.doi.org/10.1016/0266-8920(88)90005-7http://dx.doi.org/10.1006/jsvi.1993.1340http://dx.doi.org/10.1061/(ASCE)0733-9399(1995)121:2(255)http://dx.doi.org/10.1115/1.1410370http://dx.doi.org/10.1006/mssp.2000.1351http://dx.doi.org/10.1177/1475921703036169http://dx.doi.org/10.1177/1475921703036169http://dx.doi.org/10.1061/(ASCE)0733-9399(2004)130:1(3)http://dx.doi.org/10.1016/j.ymssp.2005.04.008http://dx.doi.org/10.1016/j.paerosci.2010.03.003http://dx.doi.org/10.1061/(ASCE)0733-9445(2002)128:1(96)http://dx.doi.org/10.1098/rsta.2000.0717http://dx.doi.org/10.1002/eqe.v32:1http://dx.doi.org/10.1016/0005-1098(70)90098-1http://dx.doi.org/10.1016/0005-1098(70)90098-1http://dx.doi.org/10.2514/3.20031http://dx.doi.org/10.2514/3.20031http://dx.doi.org/10.1006/jsvi.1996.0747http://dx.doi.org/10.1061/(ASCE)0733-9399(2004)130:1(49)http://dx.doi.org/10.2514/3.21228http://dx.doi.org/10.2514/3.21228http://dx.doi.org/10.1016/j.ymssp.2006.07.003http://dx.doi.org/10.1061/(ASCE)0733-9399(1984)110:12(1757)http://dx.doi.org/10.1115/1.3269475.http://dx.doi.org/10.1115/1.3269475.http://dx.doi.org/10.1115/1.3269475.http://dx.doi.org/10.1115/1.3269475.http://dx.doi.org/10.1115/1.3269475.http://dx.doi.org/10.1061/(ASCE)0733-9399(1984)110:12(1757)http://dx.doi.org/10.1016/j.ymssp.2006.07.003http://dx.doi.org/10.2514/3.21228http://dx.doi.org/10.2514/3.21228http://dx.doi.org/10.1061/(ASCE)0733-9399(2004)130:1(49)http://dx.doi.org/10.1006/jsvi.1996.0747http://dx.doi.org/10.2514/3.20031http://dx.doi.org/10.2514/3.20031http://dx.doi.org/10.1016/0005-1098(70)90098-1http://dx.doi.org/10.1002/eqe.v32:1http://dx.doi.org/10.1098/rsta.2000.0717http://dx.doi.org/10.1061/(ASCE)0733-9445(2002)128:1(96)http://dx.doi.org/10.1016/j.paerosci.2010.03.003http://dx.doi.org/10.1016/j.ymssp.2005.04.008http://dx.doi.org/10.1061/(ASCE)0733-9399(2004)130:1(3)http://dx.doi.org/10.1177/1475921703036169http://dx.doi.org/10.1006/mssp.2000.1351http://dx.doi.org/10.1115/1.1410370http://dx.doi.org/10.1061/(ASCE)0733-9399(1995)121:2(255)http://dx.doi.org/10.1006/jsvi.1993.1340http://dx.doi.org/10.1016/0266-8920(88)90005-7http://dx.doi.org/10.1016/0167-4730(86)90006-8http://dx.doi.org/10.1016/0167-4730(86)90006-88/9/2019 Bayesian model updatingReview Paper
16/18
[37] Lin, J. S., and Zhang, Y., 1994, Nonlinear Structural Identification UsingExtended Kalman Filter,Computers and Structures,52(4), pp. 757764.
[38] Koh, C. G., See, L. M., and Balendra, T., 1999, Determination of Storey Stiff-ness of Three-Dimensional Frame Buildings,Eng. Struct.,17(3), pp. 179186.
[39] Lus, H., Betti, R., and Longman, R. W., 1999, Identification of Linear Struc-tural Systems Using Earthquake-Induced Vibration Data, Earthquake Eng.Struct. Dyn.,28(11), pp. 14491467.
[40] Quek, S. T., Wang, W., and Koh, C. G., 1999, System Identification of LinearMDOF Structures Under Ambient Excitation,Earthquake Eng. Struct. Dyn.,28(1), pp. 6177.
[41] Shi, T., Jones, N. P., and Ellis, J. H., 2000, Simultaneous Estimation of Sys-tem and Input Parameters From Output Measurements, J. Eng. Mech.,126(7), pp. 746753.
[42] Lus, H., Betti, R., and Longman, R. W., 2002, Obtaining Refined First-OrderPredictive Models of Linear Structural Systems, Earthquake Eng. Struct.Dyn.,31(7), pp. 14131440.
[43] Lin, J. W., and Betti, R., 2004, Online Identification and Damage Detectionin Nonlinear Structural Systems Using a Variable Forgetting FactorApproach,Earthquake Eng. Struct. Dyn.,33(4), pp. 419444.
[44] Papadimitriou, C., Fritzen, C. P., Kraemer, P., and Ntotsios, E., 2011, FatiguePredictions in Entire Body of Metallic Structures from a Limited Number ofVibration Sensors Using Kalman Filtering,
Top Related