Bayseian Time Domain

Reliability Engineering and System Safety 160 (2017) 174–190
Contents lists available at ScienceDirect
Reliability Engineering and System Safety

journal homepage: www.elsevier.com/locate/ress
Implementation of an adaptive meta-model for Bayesian finite element MARK

model updating in time domain
⁎
H.A. Jensena, , C. Essea, V. Arayaa, C. Papadimitrioub
a
Department of Civil Engineering, Santa Maria University, Valparaiso, Chile
b
Department of Mechanical Engineering, University of Thessaly, GR-38334 Volos, Greece
A R T I C L E I N F O A BS T RAC T
Keywords: This work explores the feasibility of integrating an adaptive meta-model into a finite element model updating
Advanced simulation methods formulation using dynamic response data. A Bayesian model updating approach based on a stochastic
Bayesian model updating simulation method is considered in the present formulation. Such approach is combined with a surrogate
Kriging approximation technique and an efficient model reduction technique. In particular, an adaptive surrogate model based on
Reduced-order models
kriging interpolants and a model reduction technique based on substructure coupling are implemented. The
Transitional Markov chain Monte Carlo
integration of these techniques into the updating process reduces the computational effort to manageable levels
allowing the solution of complex problems. The effectiveness of the proposed strategy is demonstrated with
three finite element model updating applications.
1. Introduction requires a large number of system simulations. Thus, the computa-

tional demands may become excessive in cases of computationally
Finite element model updating using measured dynamic system expensive model evaluations. To address this difficulty two approaches
responses has a broad range of applications in a number of engineering are implemented in the present work without undermining the
fields [1–5]. In this context, Bayesian model updating approaches efficiency of the TMCMC method. First, some of the quantities involved
provide a robust and rigorous framework for model updating [6–20]. in the model updating process and related to model evaluations are
In fact, Bayesian inference tools supply a consistent structure for approximated by using a meta-model. In particular, a kriging approx-
reducing uncertainties in existing models by means of new informa- imation technique is selected [26,29]. Some of the advantages of this
tion. This information is used to update the initial belief about the surrogate model are that it does not require a regular grid of support
model parameters (prior probability density function) to a posterior points for constructing the approximations and provides an error
belief (posterior probability density function). For practical problems estimate of the approximations. The surrogate model is employed to
samples from the posterior probability density function have to be replace a significant number of time consuming full system simulations
generated numerically. In this setting, Markov chain Monte Carlo by fast approximate estimates. In order to maintain the kriging
algorithms are generally accepted as the most attractive Bayesian estimates quality the support points need to be in the vicinity of each
inference tools [3,21,22]. These algorithms are not only used for candidate sample during the updating process. Thus, a local scheme
sampling the target or posterior probability density function of the which relies on a set adaptive support points is considered here. In this
model parameters but for evaluating probability integrals over high- regard, some of the ideas introduced in [26] are adopted and
dimensional spaces of the uncertain model parameters as well. Of implemented in the context of finite element model updating using
particular importance among these algorithms is the transitional dynamic response data. In such work an adaptive kriging for Bayesian
Markov chain Monte Carlo (TMCMC) method which belongs to the inverse modeling was introduced and applied to several test problems
class of sequential particle filter methods [23–25]. The method which as well as to a finite element model updating problem based on modal
efficiently generates samples asymptotically distributed as the posterior frequencies and mode shapes. The basic idea is based on the fact that
probability density function can be applied to a wide range of cases the support points at a given stage or previous stages of the TMCMC
including multimodal distributions, peaked probability density func- algorithm can be used as support points for making surrogate estimates
tions, and probability density functions with flat regions [16,17,26– at new samples generated at the current stage. In this manner the full
28]. Although the TMCM method is quite flexible and general it system simulations available at previous stages allow the enhancement,
⁎
Corresponding author.
E-mail address: hector.jensen@usm.cl (H.A. Jensen).
http://dx.doi.org/10.1016/j.ress.2016.12.005
Received 29 July 2016; Received in revised form 5 December 2016; Accepted 10 December 2016
Available online 23 December 2016
0951-8320/ © 2016 Elsevier Ltd. All rights reserved.
H.A. Jensen et al. Reliability Engineering and System Safety 160 (2017) 174–190
with additional support points, of the region of interest where dimensional integral
candidate samples are generated. Second, since the construction of
the kriging interpolants is based on a number of support points a p (D|M ) = ∫Θ l (D|θ, M ) p (θ|M ) dθ (3)
potentially large number of full simulation runs would be still required
for this purpose. To reduce the computational effort involved in the full over the space of uncertain parameters Θ .
simulation runs (dynamic finite element analyses) a model reduction
technique is introduced into the updating process. In particular, a 2.2. Input-output measurements case
method based on substructure coupling is implemented [16,30,31].
The idea is to replace the original finite element model by a reduced- In the context of the present work it is assumed that D contains
order model with a greatly reduced number of degrees of freedom. In input dynamic data and output responses from measurements on the
summary, the goal of this contribution is to explore the feasibility of structural system. Let un (t j , θ|M ) denotes the response at time tj at the
extending and generalizing some of the concepts presented in [26]. In nth observed degree of freedom predicted by the proposed structural
this regard new aspects of this work include: use of a Bayesian model, and un* (t j ) denotes the corresponding measured output. The
formulation based on transient response data; implementation of a prediction and measurement errors en (t j , θ ) = un* (t j ) − un (t j , θ|M ) for
general (nonlinear) reduced-order model parametrization scheme; n = 1, …, nd , and j = 1, …, nt , where nd denotes the number of observed
parameter estimation in nonlinear structural dynamic applications; degrees of freedom and nt denotes the length of the discrete time
and efficient implementation of the TMCMC method in an underlying history data, are modeled as independent and identically distributed
normal space that includes acceptance rate control of the samples Gaussian variables with zero mean and variance σe2 [33]. This
during the updating process. assumption implies stochastic independence of the prediction errors
The organization of this contribution is as follows. Essential aspects for different channels of measurements and for different time in-
of Bayesian finite element model updating using dynamic response stances. In this regard, it is noted that different prediction error model
data are first presented in Section 2 together with a brief review of the classes can be used as well [34]. Using the above probability model for
TMCMC method. Next, approximation techniques such as a kriging the prediction error it can be shown that the likelihood function
meta model and a model reduction technique are discussed in Section l (D|θ, M ) can be expressed in terms of the quantity
3. The adaptive kriging technique and its integration within the exp(−1/2J (θ|M , D )) where
Bayesian model updating approach are presented in Section 4. The nd nt
1
effectiveness of the proposed scheme in terms of computational J (θ|M , D ) = ∑ ∑ [un* (t j ) − un (t j, θ|M )]2
nt nd σe2 n =1 j =1 (4)
efficiency and accuracy is demonstrated in Section 5 with three
example problems. The work closes with some conclusions and final is a measure-of-fit function between the measured response and the
remarks. model response at the measured degrees of freedom [7,8,33]. In order
to estimate the model parameters a simulation-based Bayesian model
2. Finite element model updating problem updating technique is used in this work. In particular a class of
sequential particle filter methods called the transitional Markov chain
2.1. Bayesian formulation Monte Carlo (TMCMC) method is adopted in the present study [23,24].
Such technique, which efficiently generates samples asymptotically
Within the framework of Bayesian model updating it is assumed distributed as the posterior probability density function is briefly
that a finite element model class M of a structural system is reviewed in the following section.
parameterized by a set of model parameters θ ∈ Θ ⊂ R np , which are
considered uncertain in the analysis. It is assumed that the finite 2.3. TMCMC method
element model class satisfies the equation of motion
M (θ ) u¨ (t , θ ) + C (θ ) u˙ (t , θ ) + K (θ ) u (t , θ ) = fNL (u (t , θ ), u˙ (t , θ ), φ (t , θ )) + f (t , θ ) Validation calculations have shown the effectiveness of the TMCMC
approach in a series of practical Bayesian model updating problems.
(1)
The method can be applied to a wide range of cases as indicated before.
where u(t , θ ) denotes the displacement vector of dimension n, u̇(t , θ ) the The method relies on the construction of a series of non-normalized
velocity vector, ü(t , θ ) the acceleration vector, fNL (u (t , θ ), u˙ (t , θ ), φ (t , θ )) intermediate distributions
the vector of non-linear restoring forces, φ (t , θ ) the vector of a set of
pj (θ|M , D ) = l (D|θ, M )αj p (θ|M ) j = 0, …, nst , 0 = α0 < α1 < … < αnst = 1
variables which describes the state of the nonlinear components, and
f(t , θ ) the external force vector. The matrices M(θ ), C(θ ), and K(θ ) (5)
describe the mass, damping, and stiffness, respectively. The evolution of where nst is the number of stages of the updating process. The
the set of variables φ (t , θ ) is described by a first-order differential parameter αj can be interpreted as the percentage of the total
equation φ˙ (t , θ ) = κ (u (t , θ ), u˙ (t , θ ), φ (t , θ )) , where κ represents a information provided by the dynamic data which is incorporated in
non-linear vector function. The equation of motion for the displacement the jth iteration of the updating procedure. The first step (j=0)
vector u(t , θ ) and the equation for the evolution of the set of variables corresponds to the prior distribution and in the last stage ( j = nst )
φ (t , θ ) constitute a system of coupled non-linear differential equations. It the samples are generated from the posterior distribution. The idea is
is noted that the above characterization of the dynamical system is quite to choose the value of exponent αj+1, given αj, in such a way that the
general since it allows to model different types of nonlinearities [32]. The change of the shape between two adjacent intermediate distributions
objective of model updating is to compute the posterior probability be small. To this end, the exponent αj+1 is chosen such that the
density function of the model parameters p (θ|M , D ) using available data coefficient of variation of the plausibility weights corresponding to
D. This is achieved by using the Bayes' Theorem stage j equal to a prescribed tolerance indicating the population
l (D|θ, M ) p (θ|M ) elitism diversification. In this context, the plausibility weights are
p (θ|M , D ) = defined as w (θjk ) = pj +1 (θjk |M , D )/ pj (θjk |M , D ) = l (D|θjk , M )αj +1− αj , k = 0,
p (D|M ) (2)
1, …, nj , where the upper index k = 1, …, nj denotes the sample number
where l (D|θ, M ) is the likelihood of observing the data D from the in the jth iteration step (θjk , k = 1, …, nj ). Once the parameter αj+1 has
model class M, p (θ|M ) is the prior probability density function been determined, the samples for stage j + 1 are obtained by generat-
representing the initial belief or information about the distribution of ing Markov chains where the lead samples are selected from the
θ , and p (D|M ) is the evidence of the model class M given by the multi- samples of the previous stage, i.e. θjk , k = 1, …, nj according to their
175
n
probability equal to the normalized weight w (θjk ) = w (θjk )/ ∑l =1 j
3. Approximation techniques
w (θjl ), k = 1, …, nj . Each sample of the chain is generated by applying
the Metropolis-Hastings algorithm [35,36], where the proposal prob- The proposed Bayesian model updating technique is in general
ability density function is a Gaussian distribution centered at the computationally very demanding due to the large number of dynamic
preceding sample of the chain with a covariance matrix Σp equal to a finite element analyses required during the updating process. In fact,
scaled version of the sample covariance matrix of the current inter- the model updating technique based on the TMCMC method involves
mediate distribution pj (θ|M , D ), that is, Σ p = ηΣj where drawing a large number of samples (of the order of thousands) for
populating the importance region in the uncertain parameter space. As
nj nj
previously pointed out the computation of each sample implies a
Σj = ∑ w (θjk )[(θjk − θj )(θjk − θj )T ], θj = ∑ w (θjl ) θjl dynamic finite element analysis (evaluation of the measure-of-fit
k =1 l =1 (6)
function needed for the estimation of the likelihood function).
It is noted that the computation of a sample in any chain requires Consequently, the computational demands may become excessive
the estimate of the likelihood l (D|·,M ). This is because the intermediate when the computational time for performing a dynamic finite element
probability density function pj+1 (·|M , D ) which is involved in the analysis is significant. To cope with this difficulty two approximation
Metropolis-Hasting step is defined in terms of the likelihood function. techniques are implemented in the present formulation: a surrogate
This is turn implies a dynamic finite element analysis. The procedure is model for the measure-of-fit function; and a model reduction technique
repeated until the parameter αj is equal to 1. At the last stage the for defining reduced-order models of the structural system. Among the
samples (θnkst , k = 1, …, n nst ) are asymptotically distributed as p (θ|M , D ). different surrogate models kriging approximation is selected in the
For a detailed implementation of the TMCMC method the reader is present implementation [29,41]. Some of the advantages of kriging
referred to [23,24]. interpolants are that they do not require regular grid of support points
and provide an error estimate of the approximations. On the other
hand, a model reduction technique based on substructure coupling for
dynamic analysis is considered here [30]. For completeness the
2.4. Implementation aspects essential aspects of kriging approximation and substructure coupling
are reviewed in this section.
As indicated in the previous section the candidate normal proposal
involved in the Metropolis-Hasting step to rejuvenate the samples is a 3.1. Kriging meta-model
multivariate Normal distribution with covariance matrix Σ p = ηΣj . The
scaling factor η has a significant impact on the performance of the 3.1.1. Basic formulation
updating process. In fact, if the scaling factor is too small the samples Kriging starts with the assumption that the function to approximate
of the Markov chain are close to each other producing a chain with a is a sample path of a Gaussian process [29]. In this context, consider m
large correlation. Besides, the samples will not properly propagate into support points zi , i = 1…m in the underlying normal space Z where
the relevant domain of the uncertain parameters. On the other hand, if the measure-of-fit function is known. That is, the quantities Jz (zi|M , D ),
the scaling factor is too large the probability to accept a new candidate i = 1…m are assumed to be known where the measure-of-fit function
point is very low. This results in a chain remaining still for long periods defined in the space Z is given by Jz (zi|M , D ) = J (θ (zi)|M , D ), with
of time with large correlation. Note that the larger the correlation in a θ = θ (z) being the transformation between the Z space and Θ space.
Markov chain, the less efficient is the sampling procedure since the The approximation of the measure-of-fit function at a new point z takes
effective number of independent samples in the chain is reduced. the form Jz (z|M , D ) = f (z)T β + β (z), where f (z)T = < f1 (z), …, fp (z)> are
Because of the above difficulties it is convenient to monitor the basis functions, β = < β1, …, βp >T are regression coefficients and the
acceptance rate during the updating process. To this end, the proposal term β (z) is a zero mean stochastic process. The corresponding
distribution which is set-up originally in the Θ space is first represented correlation function of the process is Cβ = σ 2R (ψ , z1, z2), where σ2 is
in terms of an underlying vector z ∈ Z ⊂ R np of independent standard the variance, ψ a set of parameters, and z1 and z2 are points in the
Normal variables. Thus, the updating problem is solved in terms of the underlying normal space Z . A common choice of correlation function is
parameters z instead of θ and consequently the covariance matrix of the anisotropic squared exponential given by R (ψ , z1, z2) = exp
np
the proposal distribution is a scaled version of the sample covariance [− ∑l =1 ψl (zl1 − zl2 )2 ] , where ψl ≥ 0, l = 1, …, np , and zl1 and zl2 are the
matrix in terms of z . Note that the transformation between the Θ and Z l − th components of the vector z1 and z2 , respectively [41]. For a given
spaces (or vice versa) can be obtained by means of several techniques value of ψ and σ the best linear unbiased predictor of Jz (z|M , D ) is
[37–39]. Performing the updating process in the underlying standard obtained as
normal space has some numerical advantages. For example, the
uncertainty in the vector z is normalized whereas the uncertainty in Jlz (z|M , D ) = f (z)T βl + r (z)T R−1(j − Fβl ) (7)
θ is not. Also, the support of θ may be bounded while on the contrary with variance
the support of z is not bounded. These properties tend to increase, on
average, the performance of the updating process [24]. Once the σJ2z (z) = σ 2 [1 − r (z)T R−1r (z) + b (z)T (FT R−1F)−1b (z)],
problem is set on the space of independent standard Normal variables b (z) = FT R−1r (z) − f (z) (8)
an initial scaling factor is set for the first Nm Markov chain steps within
the updating process. Based only on these chains the corresponding where r (z)T = < R (ψ , z , z1), …, R (ψ ,
z, z m)>
is the cross-correlation
acceptance rate is calculated. If the acceptance rate is lower than a vector, R = [R (ψ , zi , z j)], i , j = 1, …, m is the correlation matrix,
specified target acceptance-rate ρtarget , the scaling factor is decreased to F = [ f j (zi)], i = 1, …, m, j = 1 ,.., p is the regression matrix, jT = <
a lower value (with respect to the previous scaling factor) for the next Jz (z1|M , D ), …, Jz (z m|M , D )> is the vector of available values of the
Nm Markov chain steps. If the acceptance rate is greater than ρtarget , the measure-of-fit function in the Z space and βl = (FT R−1F)−1FT R−1j is the
scaling factor is increased (with respect to the previous scaling factor) generalized least-squares solution of the linear regression problem.
to a greater value for the next Nm Markov chain steps. The procedure
proceeds like this until all Markov chain steps have been performed 3.1.2. Parameter Estimation
during the updating process. A strategy based on the heuristic schemes It is noted that the correlation matrix R , the cross-correlation
suggested in [24,40] for updating the scaling factor is considered in the vector r (z), the regression coefficient βl , and the vector b (z) are
present implementation. functions of the set of parameters ψ , i.e. R = R (ψ ), r (z) = r (z, ψ ),
176
βl = βl (ψ ), b (z) = b (z, ψ ) respectively. At the same time it is observed u (t )T = < ui (t )T , ub (t )T >. Next, the global displacement vector u(t ) is
that the variance of the predictor depends on the variance σ2 of the written as
stochastic process β (z). Thus, the accuracy of the predictor Jlz (z) ⎡[Φ1 , …, Φ Ns] [Ψ1 , …, Ψ Ns] ∼ ⎤
⎧ v (t ) ⎫
depends on the choice of ψ and σ . One of the most popular u (t ) = T ⎨ i ⎬, T = ⎢ ii ii ib ib T ⎥
choices for determining the best values of these parameters is based ⎩ u b (t ) ⎭ ⎢⎣ [0, …, 0] [I, …, I] ⎥⎦ (12)
on maximum likelihood estimation. In this context, the maxi- where vi (t )T = < v1i (t )T , …, v iNs (t )T
>∈ R ni is the vector of independent
mum likelihood estimate of σ is given by σl (ψ )2 = 1/ m
generalized coordinates of all substructures, Φ iis is the complete set of
(j − Fβl (ψ ))T R−1(ψ )(j − Fβl (ψ )) [42], where the maximum likelihood nis fixed-interface normal modes, Ψib s
are the interface constraint
estimate of ψ is the one that minimizes the quantity ∼
modes, 0 is a null matrix, I is an identity matrix, T is a transformation
m log(σl (ψ )2) + log(det(R (ψ ))). The predictor Jlz (z|M , D ) corresponding
matrix that map the vector ub (t ) to the vector ubγ (t ), that is
to the optimum ψ l , that is, ∼
ubγ (t ) = Tub (t ), and [·, …, ·] indicates a block diagonal matrix having
Jlz (z|M , D ) = f (z)T βl (ψ
l ) + r (z , ψ l )−1 (j − Fβl (ψ
l )T R (ψ l )), as diagonal blocks the matrices inside the square bracket. The fixed-
l T −1 l −1 T −1 l interface normal modes are defined by the substructure eigenvalue
β (ψ ) = (F R (ψ ) F) F R (ψ ) j (9)
problems Kiis Φ iis − Miis Φ iis Λiis = 0, s = 1, …, Ns , where Λiis is the matrix
with variance containing the eigenvalues [30]. These modes are normalized with
respect to the mass matrix Miis satisfying Φ iisT Miis Φ iis = Iiis , and
l )2 [1 − r (z, ψ
σ Jl2 (z) = σl (ψ l )T R (ψ
l )−1r (z, ψ
l) s s
z Φ iisT Kiis Φ iis = Λiis , where Iiis ∈ R ni × ni is the identity matrix. On the other
l )T (FT R (ψ
+ b (z , ψ l )−1F)−1b (z, ψ
l )] , hand, the interface constraint modes are given by
s s s s
l ) = F T R (ψ
l )−1r (z, ψ ) − f (z) Ψib = −Kiis−1Kib ∈ R ni × nb [30]. From Eq. (12) the vector of physical
b (z , ψ (10)
coordinates at the internal degrees of freedom of all substructures ui (t )
is called the empirical best linear unbiased predictor [29,43]. It is noted can be decomposed into the dominant and residual modes
that the measure-of-fit function value at the different support points is ∼
obtained by performing a full model simulation to avoid error ui (t ) = [Φ1id , …, Φ idNs] vid (t ) + [Φ1ir , …, Φ irNs] vir (t ) + [Ψ1ib, …, Ψ ibNs] Tub (t )
propagation and subsequent deterioration of the surrogate quality. (13)
s s s
where Φ id ∈ R nd and Φ irs ∈ R nr , nds + nrs = nis are the dominant and
3.2. Model reduction technique residual fixed-interface normal modes of substructure s, respectively,
and vid (t ) and vir (t ) are the corresponding dominant and resi-
In order to control the accuracy of the kriging estimates the support dual generalized coordinates of all substructures, that is
points need to be in the vicinity of each candidate point during the vi (t )T = < vid (t )T , vir (t )T >
updating process. Thus, a scheme which relies on a set of adaptive
support points needs to be constructed. Then, it is seen that even 3.2.2. Reduced-order model
though kriging interpolants are used to replace time-consuming full In the standard formulation of substructure coupling a small
simulations by fast approximate estimates a potentially large number number of dominant modes, i.e. nds ⪡nis , is used to define the reduced
of full simulation runs is still required. To reduce the corresponding order model. In this case the global displacement vector u(t ) is
computational effort to manageable levels a model reduction technique approximated by the dominant fixed-interface normal modes as
is implemented in the present formulation. The objective is to evaluate
⎧ v (t ) ⎫ ⎡[Φ1 , …, Φ Ns] [Ψ1 , …, Ψ Ns] ∼ ⎤
the measure-of-fit function at the different support points by using u (t ) ≈ TR q (t ), q (t ) = ⎨ id ⎬, TR = ⎢ id id ib ib T ⎥
reduced-order models instead of the original unreduced finite element ⎩ u b (t ) ⎭ ⎢⎣ [0, …, 0] [I, …, I] ⎥⎦
model. (14)
where TR is the reduced transformation matrix known as the Craig-
3.2.1. Background
A model reduction technique based on substructure coupling for Bampton transformation matrix [30,31]. Using this transformation
matrix it can be shown that the reduced mass matrix MmR and the
dynamic analysis is considered here. To simplify the notation, the
lR take the form [44]
reduced stiffness matrix K
dependence of the different quantities on the set of model parameters θ
is not shown in this section. The original structural model is partitioned ⎡ m ibNs] ∼
m1ib, …, M ⎤
⎢ [I, …, I] [M T ⎥
by a number of substructures (linear or nonlinear). For each linear m l
MR = ⎢ ∼T ⎡ 1T T ⎤ T 1 Ns ∼ ⎥ , KR
substructure s, s = 1, …, Ns the following partitioned form of the mass ⎢ T ⎢M m ibNs ⎥ ∼
m ib , …, M T [Mm bb, …, M
m bb ] T⎥
s s s s
matrix Ms ∈ R n × n and stiffness matrix Ks ∈ R n × n is considered ⎣ ⎣ ⎦ ⎦
⎡ M s Mib
s ⎤ ⎡ Kiis Kib
s ⎤ ⎡[Λ1 , …, ΛNs ] ∼
[0, …, 0] T ⎤
= ⎢ ∼T ⎥
dd dd
Ms = ⎢ sii s ⎥, K = ⎢ s
s
s ⎥
⎣ Mbi Mbb ⎦ ⎣ Kbi Kbb ⎦ ⎢ T [0, …, 0] ∼ T
l1bb, …, K Ns ∼ ⎥
l bb
(11) ⎣ T [K ] T⎦ (15)
where the indices i and b are sets containing the internal and boundary with
degrees of freedom of substructure s, respectively. The boundary
degrees of freedom ubγ s s
(t ) ∈ R nbγ include only those that are common m ib
M
s T
= Φ iks Miis Ψib
s T
+ Φ iks Mib
s l bb
, K
s sT s
= Kib s
Ψib + Kbb ,
with the boundary degrees of freedom of adjacent substructures, while m bb
s sT sT sT
s M = (Ψ ib Miis + Mib s
) Ψib + Ψ ib s
Mib s
+ Mbb s = 1, …, Ns (16)
the internal degrees of freedom uis (t ) ∈ R ni are not shared with any
adjacent substructure. With this partition the following vectors are and where the matrices Λsdd , s = 1, …, Ns contain the nds kept dominant
defined: the vector of physical coordinates at the internal degrees of fixed-interface normal modes for each substructure s, s = 1, …, Ns .
freedom of all substructures ui (t )T = < u1i (t )T , …, uiNs (t )T > ∈ R ni , Based on the previous reduced matrices the equation of motion of
Ns
ni = ∑s =1 nis ; the vector of interface coordinates of all substructures the system can be written in terms of the generalized coordinates q(t)
Ns
ubγ (t )T = < u1bγ (t )T , …, ubγ
N
s
(t )T > ∈ R nbγ , nbγ = ∑s =1 nbγs ; the vector of [44]. The dimension of the matrices involved in this equation can be
physical coordinates at the Nb independent interfaces substantially smaller than the dimension of the unreduced matrices,
Ns
Nb
ub (t )T = < u1b (t )T , …, ubNb (t )T > ∈ R nb , nb = ∑l =1 nbl , where nbl is the e.g. nq ⪡n , where nq = ∑s =1 nds + nb . The equation of motion of the
number of degrees of freedom at the interface l , l = 1, …, Nb ; and the reduced-order model together with the first-order nonlinear differen-
displacement vector of physical coordinates of all substructures tial equation that describes evolution of the set of variables φ (t ) can be
177
integrated efficiently by an appropriate step-by-step integration σ Jl z (z new)

≤ ϵ,
scheme [45]. Once the vector q(t ) has been obtained the vector of Jlz (z new|M , D ) (17)
independent physical coordinates of the original unreduced system u(t )
The tolerance value ϵ , which affect the convergence speed and
can be computed directly as u (t ) = TTR q (t ), where T ∈ R n × nu
accuracy of the updating process, is problem dependent.
(n u = ni + nb ) is a constant matrix that map the vector of physical
If one of the above criteria is not satisfied a direct model evaluation,
coordinates of all substructures u(t ) to the independent physical
actually a reduced-order model evaluation, is performed in order to
coordinates of the original unreduced system u(t ).
evaluate the measure-of-fit function at the sample z new .
4.2. Reduced-order model parametrization

4. Numerical implementation
It is clear that dividing the structure into substructures and
4.1. Adaptive kriging
reducing the number of physical coordinates to a much smaller number
of generalized coordinates certainly alleviates part of the computational
As previously pointed out a local approximation scheme which
effort in evaluating directly the measure-of-fit function (cases when
relies on a set of adaptive support points is needed to control the
kriging estimates are not performed). However, the repetitive genera-
accuracy of the kriging estimates during the updating process. Due to
tion of reduced-order models for different values of the model
the annealing property of the transitional Markov chain Monte Carlo
parameters θ , required during the updating process, can be computa-
method the support points at stage j or previous stages can be used as
tionally very expensive due to the substantial computational overhead
potential support points for constructing the kriging estimates at new
that arises at substructure level. In this regard, a very efficient
samples generated at stage j + 1. These support points are required to
implementation of the model reduction technique is considered in
be in the neighborhood of the sample z new for which the measure-of-fit
the present work. In the proposed implementation the division of the
function need to be estimated. Based on some of the ideas introduced
original structure is guided by a parametrization scheme so that the
in [26] a kriging estimate is performed and accepted if the following
substructure matrices for each one of the introduced linear substruc-
criteria are satisfied simultaneously.
tures depend on only one of the model parameters. In this setting, the
Criterion 1:
stiffness and mass matrix of a substructure s, s = 1, …, Ns depend on
The estimate is based on a number of support points zi , i = 1, …, nsu
only one of the system parameters. It is noted that such parametriza-
which are in the vicinity of the point z new with respect to some measure.
tion is often encountered in a number of practical applications
For example the support points can be selected according to the
[31,32,44]. Specifically, let Sj be the set of substructures that depend
Euclidean norm, the Mahalanobis distance defined in terms of the
on the model parameter θj . It is assumed that the stiffness and mass
sample covariance matrix at the stage j (Σj ), or the nsu closest points to
matrices take the form Ks = K sgKj (θj ) and Ms = M sgMj (θj ), where the
z new . The minimum number of support points depends on the dimen-
matrices K s and M s are independent of θj , and gKj (θj ) and gMj (θj ) are
sion of the uncertain parameter space np and the order of the basis
linear or nonlinear functions of the model parameter θj . Based on this
functions used in kriging. For example, nsu ≥ np + 1 and
parametrization it is easy to show that the eigenvalues and eigenvectors
nsu ≥ (np + 1)(np + 2)/2 for kriging based on first and second order
associated with the dominant fixed-interface normal modes can be
polynomial basis, respectively. The actual number of support points is
expressed as [32].
taken as a multiple of the minimum number of support points. To
preserve the quality of the kriging estimates the support points within a s gKj (θj ) s 1
chain, in the context of the transitional Markov chain Monte Carlo Λsdd = Λ dd s
, Φ id = Φid
gMj (θj ) gMj (θj ) (18)
method, are kept fixed and they correspond to the leader point in the
s s
chain. In this context the number of support points should be sufficient where the matrices and
Λ dd are the solution of the eigenproblem
Φid
s s s s s s s
to contain all the Metropolis-Hasting steps performed within a chain. K ii Φid − Mii Φid Λ dd = 0 , where the matrices Φid and Λ dd are indepen-
Additionally, in order to consider only interpolation estimates, the dent of θj . On the other hand, the interface constraint modes are also
point z new should belong to the np-dimensional convex hull of its independent of θj since Ψib s −1 s s s
= −K iis−1 gKj (θj ) K ib gKj (θj )= −K iis−1K ib = Ψ ib
support points [26,46]. . Then it is seen that a single eigen-analysis of each substructure is
Criterion 2: required to provide the exact estimate of the normal and constraint
The surrogate estimate of the corresponding likelihood function is modes for any value of the model parameter θj . In other words, the
checked to verify whether its approximate value is within the lower r % computationally intensive re-analyses for estimating the fixed-interface
(quantile qr /100 ) of the last nli likelihood function values accounted for constrained modes and the interface constraint modes of each sub-
with full model simulation. This is done in order to avoid estimates structure for different values of θj required during the updating process
with very high plausibility, in the framework of the transitional Markov are completely avoided. In addition, the previous result allows expres-
chain Monte Carlo method, which may be erroneous. This in turn could sing the reduced-order model matrices explicitly in terms of the model
lead to the collapse of the identification process. The value of r should parameters. In fact, following the partition and parametrization of the
be relatively high in order to prevent the elimination of potential new stiffness matrix Ks and the mass matrix Ms , the parametrization of the
good sample points. On the other hand the value of nli should reflect eigenvalues Λsdd and eigenvectors Φ id s
associated with the dominant
only the recent values of the likelihood function obtained with full fixed-interface normal modes, the parametrization of the interface
model simulation. Note that the values obtained during earliest stages constraint modes Ψib s
, and the definition of the transformation matrix
are usually smaller than the values obtained at latest stages (samples TR , it can be demonstrated that the mass and stiffness matrices of the
far away from the final samples of the updating process). Thus, if nli is reduced-order model can be expressed as [32,44]
too large the bound of the quantile qr /100 would decrease. This in turn np
would increase the probability of eliminating potential good sample mR = M
M m0 + m +M
∑ {M m m g j (θ )}
gMj (θj ) + M
1j 2j 3j M j
points. The following values of r and nli are used in the present j =1 (19)
implementation: r=95 (quantile 0.95); and nli = 500 . These values have
been shown to be adequate for the problems considered in this study.
np ⎧ gKj (θj ) ⎫
lR = K
K l0 + l
∑ ⎨K
⎪
l g j (θ ) ⎬
+K
⎪
1j 2j K j
Criterion 3: ⎩
⎪
gMj (θj ) ⎭
⎪
(20)
j =1
The surrogate estimate is accepted if the coefficient of variation of
the estimate is smaller than a specified value ϵ, that is, m ,M
m0, M
where the matrices M m ,Mm ,K l ,K
l0, K l are constant matrices
1j 2j 3j 1j 2j
178
Fig. 1. Isometric view of the structural model. Example 1.
and given in Appendix A. For a detailed derivation of the different 9) Evaluate the likelihood function l (D|z new , M ) corresponding to
matrices involved in the reduced-order models the reader is referred to Jz (z new|M , D ). Continue with the acceptance/rejection rule of the
the previous references. Based on the above parametrization scheme it sample z new in the context of the TMCMC method
is seen that all matrices involved in the formulation of the model
reduction technique are computed and assembled once before the This procedure is repeated during the different stages of the
updating process takes place. Thus, the characterization of such TMCMC method.
matrices corresponds to off-line calculations. The result is that the
computational effort of the on-line calculations during the updating 4.4. Computational efficiency
process is reduced significantly since the update of the reduced-order
matrices for different values of the model parameters is immediate (see The computational efficiency of the proposed approach can be
Numerical Examples Section). It is emphasized that the efficiency of increased additionally by considering high performance computing
the above parametrization scheme is based on the assumption that the techniques at the computer hardware level. In fact, the transitional
substructure matrices of the introduced linear substructures depend on Markov chain Monte Carlo method contains a large number of chains
only one of the model parameters. Some guidelines on how to proceed which are perfectly parallel. Thus a number of computer workers can
for more general cases can be found in [31,32]. handle the generation of samples corresponding to the different chains.
The load balance in the computer workers can be based, in principle,
4.3. Likelihood function evaluation on a static or dynamic job-scheduling scheme [47,48]. In addition,
parallelization techniques are also possible at the model reduction
The numerical implementation of the proposed approximation technique level. In fact, the definition of all substructure matrices in
techniques into the context of the TMCMC method is next considered. generalized coordinates can be carried out in parallel, reducing the
As previously pointed out the proposed surrogate approximation and computational time of the proposed implementation even further.
model reduction technique allow the evaluation of the measure-of-fit
function and consequently the likelihood function in an efficient 5. Numerical examples
manner. The actual numerical evaluation of these functions is as
follows. Given a sample z new from one of the TMCMC stages the Three numerical examples are presented in this Section. The
following pseudo-code illustrates how the likelihood function objective of the first example is to evaluate quantitatively and qualita-
l (D|z new , M ) is evaluated. tively the effect of using kriging approximations in an updating
problem that considers a relatively simple linear system. On the other
1) Define the number of support points nsu and error tolerance level ϵ hand the goal of the second and third example is to evaluate the
2) Find the closest nsu support points with respect to the candidate accuracy and computational efficiency of the proposed scheme in an
sample z new involved finite element model updating problem. In all cases the
3) Check if z new belongs to the np-dimensional convex hull of the identification process is based on simulated response data.
support points. If not, go to step 8
4) Estimate Jlz (z new|M , D ) and σ Jl z (z new) using Eqs. (9) and (10) 5.1. Example 1: Simple linear model
5) Check if the estimate of the corresponding likelihood function
l ̂ (D|z new , M ) is within the lower 95 % quantile, q0.95, of the last 5.1.1. Structural system
nli=500 likelihood function values accounted for with full reduced- The two-story reinforced concrete structure shown in Fig. 1 is
order model simulation. If not, go to step 8 considered in this first example problem. Forty eight columns of square
σ l (z new)
6) Check if l Jnewz
≤ ϵ . If not, go to step 8 cross section support each floor. The dimension of the columns are
Jz (z | M , D)
7) Kriging estimate of the measure-of-fit function is accepted for the equal to 0.57 m and 0.53 m for the first and second floor, respectively.
candidate sample z new . Set Jz (z new|M , D ) = Jlz (z new|M , D ). Go to step Both floors have a constant height of 3.0 m. The mass of each floor is
9 equal to 1260 ton. Material properties are taken as: Young's modulus
8) Evaluate Jz (z new|M , D ) directly from the physical reduced-order equal to E = 2.5 × 1010 N/m 2; and Poisson ratio ν = 0.3. For the
model. Store the values of z new and Jz (z new|M , D ) in the database dynamic analysis it is assumed that each floor may be represented as
of support points rigid within the plane when compared with the flexibility of the other
structural components. Then, the degrees of freedom of the finite
179
Fig. 2. Posterior histograms of model parameters θ1, θ2, θ3, θ4 and θ5 obtained from the original version of the TMCMC method. Mean estimates: θ1 = 0.80 , θ 2 = 0.90 , θ 3 = 0.91,
θ4 = 0.94 , θ5 = 1.21.
element model are linked to three degrees of freedom per floor (two example problems considered in this Section.
translational displacements and one rotational displacement) by using
condensation techniques. Finally, a 3 % of critical damping for the 5.1.3. Results
modal damping ratios is introduced to the model. Due to the simplicity Fig. 2, 3 and 4 show the posterior histograms of the model
of the model, a reduced-order model is not considered in this example parameters θ1, θ2, θ3, θ4 and θ5 obtained by using the original and
problem. the surrogate versions of the TMCMC method, respectively. In addition
the mean estimate of the parameters is also indicated in the figures. For
5.1.2. Identification problem the surrogate versions of the TMCMC method ϵ-tolerance levels equal
The structural system described in the previous section is taken as to ϵ = 0.01 and ϵ = 0.1 are considered. The surrogate estimates are
the model class for identification purposes. The rigidities correspond- based on a second-order kriging approximation considering 100
ing to the columns of the model class are parameterized as follows: support points (about five times the required minimum number).
EI1x = θ1 EI1x , EI2x = θ2 EI2x , EI1y = θ3 EI1y , and EI2y = θ4 EI2y , where Iix and These points correspond to the closest points (with respect to Euclidian
Iiy, i = 1, 2 represent the nominal moment of inertia of the columns of norm) of the sample where the measure-of-fit function is estimated.
the i floor in the x and y direction, respectively. These nominal values The results of the figures correspond to one particular run of the
correspond to the columns of the structural system previously de- updating process.
scribed. On the other hand, the damping ratios for the vibrational It is observed that the histograms and the mean estimates of the
modes are identical and parameterized as ζ = θ5 ζ , where ζ is the parameters obtained by the surrogate versions of the TMCMC method
nominal value and equal to 3 %. The identification process is based on are similar to the predictions obtained by the original version of the
simulation data. For illustration purposes the actual structure used to TMCMC method. It is also observed that the actual system and the
generate the measured data is defined in terms of the following models characterized by the posterior samples are very close for the
properties: I1x = 0.80I1x , I2x = 0.90I2x , I1y = 0.90I1y , I2y = 0.95I2y and original and the surrogate versions of the method. Actually, the
ζ = 1.2ζ . This system may be interpreted as the actual structural samples of the parameters related to the stiffness are distributed
system where changes in the stiffness and damping properties have around their actual values θ1 = 0.80, θ2 = 0.90, θ3 = 0.90, θ4 = 0.95,
occurred due to for example deterioration from environmental effects. while the samples of the parameter associated with the damping ratios
The input ground acceleration history to generate the data corresponds are distributed around the actual value θ5 = 1.20 .
to the Santa Lucia ground-motion record recorded during the 2010 Information about the use of kriging during the different stages of
Chilean earthquake. It is applied at 45° with respect to the x axis as the identification process for the different error tolerance levels ϵ is
shown in Fig. 1. The measured response is simulated by first calculat- shown in Fig. 5 and 6. In the present formulation an initial database of
ing the absolute acceleration response of the actual structure at the first support points is constructed during the first stage of the identification
and second floor (in the x and y direction) and then adding 10 % rms process. Thus, full model simulations are performed during this stage
Gaussian white noise. The responses are computed at the center of and the use of kriging is considered only from the second stage. Of
mass of each floor. Thirty seconds of data with sampling interval course, the use of kriging can also be considering during the initial
Δt = 0.02 s are used, given a total of nt=1501 data points. Independent stage of the updating process. In the case of small ϵ -tolerance value
uniform prior distributions are assumed for the parameters θ1, θ2, θ3, (ϵ = 0.01) the support points do not provide a good surrogate approx-
θ4 and θ5 over the ranges [0.3, 2.0] for the parameters related to the imation at the initial stages. So the condition that prevails for rejecting
moments of inertia and [0.8, 1.4] for the parameter associated with the the surrogate estimates is failure to meet the ϵ -tolerance criterion
damping ratios, respectively. Recall that the updating problem is (Criterion 3). At later stages an increase of the accepted surrogate
performed in terms of an underlying vector of independent standard points is observed. It is seen that the rejection of the surrogate
Normal variables. The transitional Markov chain Monte Carlo method estimates due to overshooting the 95 % quantile range (criterion 2)
with nj = 1000, j = 1, …, nst , ρtarget = 40% and Nm=100 (see Section 2.4) and due to the absence of the required number of support points
is implemented for the identification process. This level of acceptance (criterion 1) is very small. The overall surrogate acceptance ratio
rate ρtarget and number of intermediate Markov chain steps Nm for during the updating process is approximately 30 % for this small error
monitoring the acceptance rate are found to be adequate for the tolerance level.
180
Fig. 3. Posterior histograms of model parameters θ1, θ2, θ3, θ4 and θ5 obtained from the surrogate version of the TMCMC method. Error tolerance value ϵ = 0.01. Mean estimates:
θ1 = 0.80 , θ 2 = 0.91, θ 3 = 0.90 , θ4 = 0.95, θ5 = 1.21.
For a larger ϵ value (ϵ = 0.1) a large percentage of the surrogate Finally, in order to get insight into the updating process Fig. 7
points are accepted especially at the last stages. In fact, the acceptance shows the samples in the (θ1, θ2 ) space and (θ3, θ4 ) space at the last stage
ratio is more than 90 % at the last five stages of the updating process. In of the process. These parameters are related to the stiffness in the x and
this case the accumulation of support points, from previous stages, y direction at the first and second floor, respectively. It is observed that
results in a high concentration of points in a relatively small area of the the data is strongly correlated along a certain direction in the
parameters space. The measure-of-fit function is smoothly varying over parameter space. Such correlation shows that an increase in the
this area and thus the surrogate estimate is more accurate meeting the stiffness of the columns of the first floor is compensated by a decrease
ϵ -tolerance criterion for most of the samples. The overall surrogate in the stiffness of the columns of the second floor during the
acceptance ratio is approximately 60% for this ϵ value case. As in the identification process, which is reasonable from a structural point of
previous case (ϵ = 0.01) the reason that prevails for rejecting the view. Thus, all points along that direction correspond to structural
surrogate estimates is the unmet ϵ -tolerance criterion. In other words, models that have almost the same response at the measured degrees of
criteria 1 and 2 are relatively inactive throughout the identification freedom. The results correspond to the surrogate version of the
process in this case. That is, the rejection due to these criteria is TMCMC method with an error tolerance value equal to ϵ = 0.1. The
unlikely. From the previous result is expected that an important distribution of these samples is very similar to the one obtained by the
reduction in the number of function evaluations (dynamic finite original version of the TMCMC method as previously pointed out.
element analyses) will be achieved. Actually, a 60 % reduction in the
number of function evaluations with respect to the required by the
original version of the TMCMC method is obtained for the case in 5.2. Example 2: Linear finite element model
which ϵ = 0.1. This in turn implies an important reduction in computa-
tional cost. Such reduction may be significant in complex finite element 5.2.1. Model description and reduced-order model
model updating problems (see next examples). The second example consists of the ten floors three-dimensional
reinforced concrete building model shown in Fig. 8. The model, which
Fig. 4. Posterior histograms of model parameters θ1, θ2, θ3, θ4 and θ5 obtained from the surrogate version of the TMCMC method. Error tolerance value ϵ = 0.1. Mean estimates:
θ1 = 0.80 , θ 2 = 0.90 , θ 3 = 0.90 , θ4 = 0.96 , θ5 = 1.21.
181
modeled with shell elements with a thickness of 0.3 m and beam

elements of rectangular cross section of dimension 0.3 m× 0.6 m from
floors 1–5 and 0.25 m × 0.5 m from floors 6–10. Each floor is
supported by 48 columns of rectangular cross section of dimension
0.8 m× 0.9 m. The corresponding finite element model has approxi-
mately 40,000 degrees of freedom. A classical damping is assumed for
the model with a 3 % of critical damping for the first modes.
For illustration purposes it is assumed that a stiffness reduction of
20 % of the nominal stiffness value of the columns of the first floor has
occurred. In the context of this example problem it is desired to
monitor the stiffness of the columns corresponding to the first, second
and third floor. Note that the stiffness of the columns of the second and
third floor is unchanged. In order to perform the model updating a
model class containing three parameters (θ1, θ2, θ3) associated with the
stiffness (modulus of elasticity) of the columns of the first, second and
third floor, respectively, is introduced. The structure is subdivided into
six substructures as shown in Fig. 9. Substructures 1, 3 and 5 are
Fig. 5. Use of kriging during the different stages of the identification process. Error
tolerance level ϵ = 0.01. composed by the column elements of the first, second and third floor,
respectively. On the other hand, substructures 2 and 4 correspond to
the slabs and beam elements of the first and second floor, respectively,
and substructure 6 contains the upper floors structural components
(columns, beams and slabs). Then, substructures 1, 3 and 5 depend on
the model parameters θ1, θ2, and θ3, respectively, while substructures
2, 4 and 6 do not depend on the model parameters. The parametriza-
tion functions of the model parameters are given by gKj (θj ) = θj and
gMj (θj ) = 1 in this case (see Section 4.2). The reference structure
(unchanged) is characterized in terms of the model parameters with
values equal to one, i.e. θi = 1, i = 1, 2, 3.
Due to the relatively large dimension of the finite element model a
reduced-order model is considered. It is characterized by a total of 300
generalized coordinates corresponding to the fixed-interface normal
modes of all substructures. In particular 50 fixed-interface normal
modes are retained for each one of the substructures 1, 2, 3, 4, 5, and 6.
On the other hand, the number of interface degrees of freedom is equal
to 864 for the proposed reduced-order model. Thus, a reduced-order
model of 1164 degrees of freedom is obtained. This number of degrees
of freedom represents an almost 97 % reduction with respect to the
unreduced finite element model. Validation calculations show that this
Fig. 6. Use of kriging during the different stages of the identification process. Error
tolerance level ϵ = 0.1.
reduced-order model is adequate in the context of the present applica-
tion. In fact, the relative error between the first 15 modal frequencies
obtained by using the unreduced finite element model and the modal
is assumed to be linear, has been borrowed from [16] where it was used frequencies obtained by the reduced-order model is less than 0.4 %.
for model updating in the context of model reduction techniques. The Additionally, MAC (modal assurance criterion) values between the
material properties of the reinforced concrete structure are defined as modal vectors computed from the full finite element model and from
follows: Young's modulus E = 2.34 × 1010 N/m 2; Poisson ratio ν = 0.3; the reduced-order model indicate that they are consistent. The
and mass density ρ = 2500 kg /m3. The height of each floor is 3.5 m comparison with the lowest 15 modes is adequate since validation
leading to a total height of 35.0 m for the structure. The floors are calculations show that the contribution of the higher order modes in
Fig. 7. Samples of the posterior probability density function in the (θ1, θ 2 ) space and (θ 3, θ4 ) space generated at the last step of the surrogate version of the TMCMC method. Error
182
Nm=100 for monitoring the acceptance rate. The updating problem

is performed in terms of an underlying vector of independent standard
Normal variables as indicated in Section 2.4. For the surrogate versions
of the TMCMC method three error values ϵ = 0.01, 0.1 and 0.5 are
considered in this case. The surrogate estimates are based on a second-
order kriging approximation considering 20 support points (about
three times the required minimum number). The accuracy of the
surrogate versions of the TMCMC estimates is shown in Table 1.
This table shows the sample mean and sample standard deviation of
the marginal distributions of the model parameters θ1, θ2, and θ3
obtained by the original and surrogate versions of the TMCMC method.
The different versions of the method are implemented by using the
reduced-order model defined in the previous section. The statistics are
obtained from 50 independent runs. In general for a small value of the
error tolerance level ϵ (i.e. ϵ = 0.01) is expected to yield estimates
similar to the ones obtained without surrogates. This is confirmed by
the results shown in the table. As the value of ϵ increases the accuracy
of the estimates is expected to decrease. However, it is observed that in
this case the prediction of the sample mean of the model parameters is
excellent even for a high value of ϵ (i.e. ϵ = 0.5). In addition, it is seen
that the sample mean of the parameters is very close to their actual
values θ1 = 0.80 , θ2 = 1.00 , and θ3 = 1.00 . Note that the corresponding
COV of the estimates is very small (less than 1.4 %) for all cases. In
Fig. 8. Isometric view of the finite element model. Example 2. other words, the surrogate versions of the TMCMC method are able to
identify the stiffness reduction of the columns of the first floor and the
the dynamic response of the model is negligible. fact that the stiffness of the columns of the second and third floor
remains unchanged.
5.2.2. Simulated response data An insight into the effect of the error tolerance level ϵ on the mean
As in the previous example the identification process is based on surrogate acceptance ratio during the model updating process is
simulated data. The original unreduced finite element model is excited presented in Fig. 10. As in the previous example problem the first
horizontally in the x direction by the Concepcion ground-motion stage of the identification process is performed with full simulation
recorded during the 2010 Chilean earthquake. The measured responses runs (reduced-order model) in order to construct an initial database of
are simulated by first calculating the acceleration response at floors 1, support points. It is seen that for the case of small ϵ value (ϵ = 0.01)
2, 3 and 4 (in the x and y direction) of the original unreduced model most of the surrogate points are not accepted during the second and
with a stiffness reduction of 20 % of the first floor columns. Then, a third stages. After the third stage a steep increase of the accepted
Gaussian discrete white noise sequence with standard deviation equal surrogate points is observed. The overall mean surrogate acceptance
to 5 % of the root-mean square value of the corresponding acceleration ratio is approximately 60 % (considering that no surrogate estimates
time histories is added. The responses are computed at the center of are computed during the first stage). In the case of intermediate ϵ value
mass of the floors. In this case sixty seconds of data with sampling (ϵ = 0.1) and large ϵ value (ϵ = 0.5) a very large percentage of surrogate
interval Δt = 0.05 s are used, given a total of 1201 data points to be points are accepted throughout the updating process. The total mean
used during the identification process. surrogate acceptance ratio is approximately 81 % for the intermediate ϵ
value case and 85% for the large ϵ value case. In these two cases the
5.2.3. Model updating results tolerance criterion is relatively high and it is met by most surrogate
A uniform prior distribution is used with bounds [0.5,1.5] for the points. As a result only a low percentage of full simulation runs are
model parameters θ1, θ2, and θ3. The TMCMC method set-up includes, performed in these two cases. This in turn has significant consequences
as in the first example, 1000 samples per stage, an acceptance rate on the computational cost of the model updating process as shown in
ρtarget = 40%, and a number of intermediate Markov chain steps the next section.
Fig. 9. Substructures of the finite element model used for model updating.
183
Table 1
Accuracy of the surrogate estimates for different ϵ values. Example 2.
TMCMC Mean Standard COV

Method Value Deviation %
Version Model parameters Model parameters Model parameters
θ1 θ2 θ3 θ1 θ2 θ3 θ1 θ2 θ3
Original 0.80 0.99 1.00 0.0025 0.012 0.0092 0.31 1.22 0.92
Surrogate (ϵ = 0.01) 0.80 0.98 1.00 0.0021 0.012 0.0074 0.27 1.26 0.74
Surrogate (ϵ = 0.1) 0.80 1.01 0.98 0.0024 0.014 0.0084 0.30 1.38 0.86
Surrogate (ϵ = 0.5) 0.80 1.01 0.99 0.0026 0.018 0.011 0.33 1.8 1.14
TMCMC method. In this context, the speedup is the ratio of the execution
time by using the original version of the TMCMC method with the full
finite element model and the execution time by using alternative
implementations of the TMCMC method. The computational times
reported in Table 2 are based on the implementation of the updating
process in a four core computer unit (Intel Core i7 processor). The actual
procedure was carried out by using a homemade code based on a Matlab-
C++ platform.
First, it is noted that a speedup equal to 145 is obtained by using the
original version of the TMCMC method with the proposed reduced-
order model formulation. This large value is due to the fact that the
different matrices involved in the reduced-order model are computed
and assembled once during the updating process. This in turn implies
that the actualization of the reduced-order model matrices is very
efficient. On the contrary, the matrices that define the full finite
element model need to be computed and assembled for each new
value of the model parameters required during the updating process.
Fig. 10. Use of kriging during the different stages of the identification process for The speedup is even more significant when surrogate approximations
different error tolerance levels ϵ. Example 2.
are considered. Actually, a speedup of 500 is achieved for the error
tolerance level ϵ = 0.5 case. This value is more than 3 times the
speedup corresponding to the reduced-order case without surrogate
approximations. This is because fast approximate estimates of the
5.2.4. Computational cost measure-of-fit function replace reduced-order model simulations for
The corresponding cost is reported in Table 2. The table shows the evaluating such function. The speedup values obtained represent a
total time (in minutes) for one particular run of the updating process by reduction of more than two orders of magnitude. Thus, the combina-
considering the following implementations of the TMCMC method: the tion of kriging approximations with reduced-order models is quite
original version of TMCMC with the full finite element model; the original beneficial in term of computational cost. Note that this significant
version of TMCMC with the reduced-order model; and the surrogate reduction in computational effort is obtained without compromising
versions of TMCMC with the reduced-order model considering the three the accuracy of the updating process.
error tolerance levels ϵ . The total cost showed in the table includes all
tasks required during the updating process including the assemblage of 5.2.5. Updated samples of model parameters
the finite element model, the construction of the reduced-order model, the To get further information about the updating process the interac-
solution of the equation of motions, the implementation of the TMCMC tion between the model parameters is shown in Fig. 11. This figure
algorithm, etc. The last column of the table indicates the speedup (round shows the samples of the posterior probability density function in the
to the nearest integer) achieved by the different implementations of the (θ1, θ2 ), (θ2, θ3), and (θ1, θ3) spaces.
The results correspond to the ones obtained from the surrogate
version of the TMCMC method with an error tolerance level ϵ = 0.5. It
is observed that the data is correlated along certain directions in the
Table 2 (θ1, θ2 ) space and (θ2, θ3) space. Such correlation shows that an increase
Computational cost involved during the updating process by considering different in the stiffness of the columns of the first floor is compensated by a
implementations of the TMCMC method. Example 2.
decrease in the stiffness of the columns of the second floor. Similarly,
Method Total time (m) Speedup an increase in the stiffness of the columns of the second floor is
compensated by a decrease in the stiffness of the columns of the third
Original version of TMCMC with full FE model 5197.0 1 floor. The interaction is particularly strong in the (θ1, θ2 ) space which is
Original version of TMCMC with reduced-order FE 35.8 145
reasonable from a structural view point. On the other hand there is
model
Surrogate version of TMCMC (ϵ = 0.01) with 18.2 285 almost no interaction between the stiffness of the columns of the first
reduced-order FE model floor and the stiffness of the columns of the third floor ((θ1, θ3) space)
Surrogate version of TMCMC (ϵ = 0.1) with reduced- 11.3 460 which is also in accordance with the physical interpretation of these
order FE model parameters. As a consequence all points along the different directions
Surrogate version of TMCMC (ϵ = 0.5) with reduced- 10.4 500
shown in Fig. 11 correspond to structural models that have almost the
order FE model
same response at the measured degrees of freedom. The results
corresponding to error tolerance levels ϵ = 0.01 and ϵ = 0.1 are shown
184
Fig. 11. Samples of the posterior probability density function in the (θ1, θ 2 ), (θ 2, θ 3) , and (θ1, θ 3) spaces generated at the last step of the surrogate version of the TMCMC method. Error
in Figs. 12 and 13, respectively. It is seen that the distribution of the ⎧ u (t , θ ) ⎫ ⎡ M e (θ ) Me (θ ) Ge ⎤

u (t , θ ) = ⎨ e ⎬ , M (θ ) = ⎢ T ⎥,
samples of the posterior probability density functions are very similar ⎩ u b (t , θ ) ⎭ ⎣ Ge Me (θ ) Mb (θ ) + GTe Me (θ ) Ge ⎦
for the three error tolerance levels. So, the distribution of the samples ⎡ C (θ ) 0 ⎤
at the last stage of the updating process remains almost invariant with C (θ ) = ⎢ e ⎥,
⎣ 0 Cb (θ )⎦
respect to the error tolerance values considered in this case.
⎡ K (θ ) 0 ⎤ ⎡ Me (θ ) Ge ⎤
K (θ ) = ⎢ e ⎥ f (t , θ ) = − ⎢ ⎥ u¨ g (t ),
5.3. Example 3: Non-linear finite element model ⎣ 0 K b (θ ) ⎦ T
⎣ Mb (θ ) + Ge Me (θ ) Ge ⎦
⎧ 0 ⎫
5.3.1. Structural model fNL (t , θ ) = −⎨ ⎬
⎩ frb (t , θ )⎭ (21)
For the third example problem the linear structural model de-
scribed in the previous example is isolated in its base by means of 48 where u e (t , θ ) is the vector of relative displacements of the super-
nonlinear rubber bearings. The rubber bearings consist of layers of structure with respect to the base, Me (θ ), Ce (θ ), K e (θ ) are the
rubber and steel, with the rubber being vulcanized to the steel plates. corresponding mass, damping and stiffness matrices, respectively, Ge
Fig. 14 shows a schematic representation of a rubber bearing, where De is the matrix that couples the excitation components of the vector üg (t )
represents the external diameter of the isolator, Di indicates the to the degrees of freedom of the superstructure, ub (t , θ ) is the vector of
internal diameter, Hr = tr nr is the total height of rubber in the device, base displacements relative to the ground, Mb (θ ) is the mass matrix of
where tr is the layer thickness and nr is the number of rubber layers. the rigid base, Cb (θ ) is the resultant damping matrix of viscous
The nominal values of the isolator parameters are set equal to De = 0.85 isolation components, Kb (θ ) is the resultant stiffness matrix of linear
m, Hr = 0.14 m, and Di = 0.10 m. The base platform where the rubber elastic isolation components, and frb (t , θ ) is the vector containing the
bearings are located has a total mass of 6.0 × 10 5 kg . The force nonlinear forces activated on the isolators. In connection with section
displacement characteristics of the isolator elements are modeled by 2.1, a detailed description of the underlying set of variables φ (t , θ )
a biaxial hysteretic behavior. An analytical model based on a series of which characterizes the nonlinear behavior of the rubber bearings can
experimental tests conducted for real-sized rubber bearings is used in be found in [51].
the present application. For a detailed description of the analytical
model that describes the nonlinear behavior of the bearings the reader
is referred to [49–51]. For illustration purposes Fig. 12 also shows a 5.3.2. Updating problem
typical displacement-restoring force curve of one of the isolators under The objective in this example problem is to monitor some of the
ground motion. parameters that characterize the nonlinear behavior of the isolators. In
The combined equation of motion of the base-isolated structural particular, the external diameter of the isolators De and the total height
system (superstructure and base isolation system) can be written as in of rubber in the devices Hr are selected for identification purposes.
Eq. 1 with [16] Information about the specific effect of these parameters on the
185
nonlinear behavior of the isolators is provided in [50,51]. The para- 5.3.3. Results of identification process
meters are parameterized as De = θ1 De and Hr = θ2 Hr , where De = 0.85 The sample mean and the corresponding sample coefficient of
m and Hr = 0.14 m are the corresponding nominal values. The super- variation of the marginal distributions of the model parameters θ1 and
structure is subdivided into the six linear substructures shown in Fig. 9 θ2 obtained by the original and surrogate versions of the TMCMC
plus one nonlinear substructure composed by the isolation devices and method are shown in Table 3. The statistics are obtained from 50
the base platform. The reduced-order model defined in Example 2 is independent runs and the different versions of the TMCMC are
used for the linear superstructure. Note that the linear substructures do implemented with the reduced-order model of the superstructure
not depend on the model parameters since the parameters θ1 and θ2 defined in the previous example. It is seen that the prediction of the
are related to the nonlinear substructure. Thus, the linear substruc- sample mean of the model parameters is excellent even for a high value
tures do not need to be updated during the identification process. The of ϵ. Further, the sample mean of the parameters is almost equal to
simulated data is generated with the same ground acceleration used in their actual values θ1 = 0.92 (De = 0.782) and θ2 = 1.2 (Hr = 0.168).
the previous example. In this case twenty seconds of data with From Table 3 it is also inferred that the sample standard deviations
sampling interval Δt = 0.05 s are used, given a total of Nt=401 data of the marginal distributions of the model parameters θ1 and θ2
points to be used during the identification process. The actual base- obtained for the ϵ = 0.5 case are about twice of the ones obtained for
isolated system used to generate the simulated measured data is the ϵ = 0.01 case. However, the corresponding sample coefficients of
characterized by isolators with parameters De=0.782 m and variation are quite small. Thus, the surrogate versions of the TMCMC
Hr=0.168 m. Note that this model is more flexible than the nominal method are able to identify the external diameter of the isolators and
system since the isolators have smaller external diameters and larger the total height of rubber in the devices. The corresponding histograms
heights. The measured responses are simulated by first calculating the of the parameters θ1 and θ2 obtained from one particular run of the
acceleration response in the x direction at the base platform where the surrogate versions of the TMCMC method are shown in Figs. 15 and
rubber bearings are located. Then, a Gaussian discrete white noise 16. They correspond to the error tolerance levels ϵ = 0.01 and ϵ = 0.5,
sequence with standard deviation equal to 5% of the root-mean square respectively. It is observed that the histograms are very similar and that
value of the corresponding acceleration time history is added. A the large prior uncertainty about these parameters (uniformly distrib-
uniform prior distributions are assumed for the parameters θ1 and θ2 uted over the range [0.5,1.4]) is significantly reduced during the
over the range [0.5, 1.4]. The transitional Markov chain Monte Carlo updating process.
method is implemented as in the previous example. Two error To complement the previous information, the actual distribution of
tolerance values, ϵ = 0.01 and ϵ = 0.5, are considered for the surrogate the samples from the posterior probability density function in the
versions of the TMCMC method in this application. The surrogate (θ1, θ2 ) space is shown in Fig. 17 for the two surrogate versions of the
estimates are based on a second-order kriging approximation con- TMCMC method. Note that the distribution is very similar in both
sidering 20 support points as in the previous example. cases. It is also noted that the parameters θ1 and θ2 are correlated along
Fig. 14. Schematic representation of a rubber bearing and a typical displacement-restoring force curve of the rubber bearing.
186
Table 3 value (ϵ = 0.01) the overall acceptance ratio is approximately 68 %. On

Accuracy of the surrogate estimates for different ϵ values. Example 3. the other hand for the case of high ϵ value (ϵ = 0.5) the corresponding
acceptance ratio is almost 96 %. Thus, a very low percentage of full
TMCMC Mean COV
Method Value % simulation runs are performed in this case. In summary, the identifica-
Version Model parameters Model parameters tion process gives excellent results for small and high error tolerance
levels. In the case of a high value of ϵ an increase in the coefficient of
θ1 θ2 θ1 θ2 variation of the estimates is observed. However, the value of the
Original 0.922 1.208 0.28 0.59
coefficient of variation is very small for both parameters.
Surrogate (ϵ = 0.01) 0.923 1.209 0.26 0.60
Surrogate (ϵ = 0.5) 0.923 1.209 0.41 1.22 5.3.4. Numerical efficiency
The number of full simulation runs for different implementations of
the TMCMC method is given in Table 4. In particular the following
implementations are considered: the original version of TMCMC with
the reduced-order model; and the surrogate versions of TMCMC with
the reduced-order model. The last column of the table indicates the
efficiency measure achieved by the different implementations of the
TMCMC method. The definition of the efficiency measure is similar to
the speedup measure previously defined. That is, the ratio (round to the
nearest integer) of the number of full simulation runs by using the
original version of the TMCMC method and the number of full
simulation runs by using a surrogate version of the TMCMC method.
It is seen that the effect of using kriging is substantial. In fact an
efficiency measure of almost 25 is achieved for the error tolerance level
ϵ = 0.5 case. In terms of computational time a speedup of 7 is obtained
for the high tolerance value level case.
As in the previous examples the drastic decrease in computational
demands is achieved without compromising the predictive capabilities
of the proposed updating procedure. Note that the comparison between
the full and reduced-order model of the superstructure is not con-
sidered here since the substructures of the superstructure are inde-
pendent of the model parameters in this application. Thus, the full
Fig. 15. Posterior histograms of model parameters θ1 and θ2. Surrogate version of the
TMCMC method with error tolerance level ϵ = 0.01. finite element model of the superstructure as well as the matrices
involved in the reduced-order model of the superstructure are com-
puted and assembled once during the updating process. Only the
a particular direction. The correlation structure is consistent with the nonlinear substructure (rubber bearings) needs to be updated during
fact that the base isolation system becomes stiffer as the rubber the process. In other words, the performance of the full and reduced-
diameter is increased. Contrarily, the isolation system becomes more order model is relatively similar in the context of this application
flexible as the height of the rubber is increased [51]. Therefore, an problem.
increase in the rubber diameter is compensated by an increase in the
height of the rubber during the updating process. Thus all points along 6. Conclusions
that direction correspond to isolation system models that have similar
base drift responses. The feasibility of integrating an adaptive meta-model into a finite
Information about the use of kriging during the different stages of element model updating formulation using dynamic response data has
the identification process is given in Fig. 18. For the case of small ϵ been explored. A simulation-based Bayesian model updating technique
is adopted in the present work. In particular the transitional Markov
chain Monte Carlo method is implemented in an underlying space of
standard Normal variables. The updating technique is combined with
an adaptive surrogate model based on kriging interpolants of the
measure-of-fit function. The information needed for the construction of
the kriging estimates, that is, the measure-of-fit function values at the
database of support points, is obtained by a reduced-order model. To
this end a model reduction technique based on substructure coupling is
considered here. In this setting, the proposed finite element model
updating technique uses the distribution of the spatial structure of the
transitional Markov chain Monte Carlo samples to replace a large
number of model simulations by approximate, fast and efficient
interpolating estimates. Numerical results have demonstrated that an
important reduction in the number of full model runs is achieved. In
fact, speedup values of more than two orders of magnitude have been
obtained in involved finite element model updating problems.
Moreover, this reduction is accomplished without compromising the
predictive capability of the identification process. Future research
efforts are aimed at considering reduced-order models of even smaller
dimension in order to further reduce the solution time. This can be
Fig. 16. Posterior histograms of model parameters θ1 and θ2. Surrogate version of the achieved by using generalized coordinates not only in terms of the
TMCMC method with error tolerance level ϵ = 0.5. fixed-interface normal modes but in terms of the interface constraint
187
Fig. 17. Samples of the posterior probability density function in the (θ1, θ 2 ) space. A: Surrogate version of the TMCMC method with error tolerance level ϵ = 0.01. B: Surrogate version of
the TMCMC method with error tolerance level ϵ = 0.5 .
Table 4
Number of full simulation runs for different implementations of the TMCMC method.
Example 3.
Method Number Efficiency
Original version of TMCMC with reduced-order FE model 6040 1

Surrogate version of TMCMC (ϵ = 0.01) with reduced- 1971 3
order FE model
Surrogate version of TMCMC (ϵ = 0.5) with reduced-order 256 24
FE model
vance. In this case, the removal of potential high frequency compo-

nents in the structural response needs to be considered among other
things.
Acknowledgments
Fig. 18. Use of kriging during the different stages of the identification process for small
and high error tolerance levels. Example 3 .
The research reported here was supported in part by CONICYT
(National Commission for Scientific and Technological Research)
modes as well. Also, the consideration of the substructure residual under grant number 1150009 which is gratefully acknowledged by
fixed-interface normal modes can be considered in this regard. In the authors. Also this research has been implemented under the
addition, attention to problems with a relatively large parameter space "ARISTEIA" Action of the "Operational Programme Education and
dimension is also an important topic for future research. Finally, the Lifelong Learning" and was co-funded by the European Social Fund
consideration of measured response data, such as earthquake vibration (ESF) and Greek National Resources under Grant no. 53-UQ-dy-
data, for identification purposes is another aspect of practical rele- namics.
Appendix A. Appendix A
m ,M
m0, M
The matrices M m ,Mm ,K l ,K
l0, K l that define the mass and stiffness matrices of the reduced-order model are given by
1j 2j 3j 1j 2j
⎡ ⎡ m1 ⎤∼ ⎤
⎢ m Ns
[Iδ10, …, Iδ Ns 0] ⎢Mib δ10, …, Mib δ Ns 0 ⎥ T ⎥
⎢ ⎣ ⎦ ⎥
m0 =
M ⎢ ⎡ 1T NsT ⎤ ∼T ⎡ 1 ⎤ ∼⎥
∼T m m δ ⎥ T M
⎢ T ⎢Mib δ10, …, M m m Ns
ib Ns 0 ⎢ bb δ10, …, M bb δ Ns 0 ⎥ T ⎥
⎢⎣ ⎣ ⎦ ⎣ ⎦ ⎥⎦ (22)
⎡ ∼ ⎤
m = ⎢[Iδ1j, …, Iδ Ns j ] [0, …, 0] T ⎥
M1j
⎢⎣ T [0, …, 0] T [0, …, 0] ∼
∼T ∼T
T ⎥⎦ (23)
188
⎡ ⎡ m1 ⎤ ∼⎤
⎢ m Ns
[0, …, 0] ⎢Mib δ1j, …, Mib δ Ns j ⎥ T ⎥
⎣ ⎦ ⎥
m =⎢
M2j ⎢ ⎡ 1T NsT ⎤ ⎥
∼T m m δ ⎥
⎢ T ⎢Mib δ1j, …, M ∼T ∼ ⎥
ib Ns j T [0, …, 0] T
⎢⎣ ⎣ ⎦ ⎥⎦ (24)
⎡ [0, …, 0] ∼
[0, …, 0] T ⎤
m = ⎢⎢ ⎡ ⎤
⎥
M3j ∼T ∼T m 1
m s
N ∼⎥
⎢ T [0, …, 0] T ⎢⎣M bb δ1j, …, M bb δ Ns j ⎥⎦ T ⎥
⎣ ⎦ (25)
⎡ 1 Ns ∼ ⎤
⎢[Λ kk δ10, …, Λ kk δ Ns 0] [0, …, 0] T ⎥
l
K0 = ⎢ ⎤ ∼⎥
∼T ∼T ⎡l1 l Ns
⎢ T [ 0 , …, 0 ] T ⎢ bb 10
K δ , …, K bb Ns 0 ⎥ ⎥
δ T
⎣ ⎣ ⎦ ⎦ (26)
⎡ 1 Ns ∼ ⎤
l = ⎢[Λ kk δ1j, …, Λ kk δ Ns j ] [0, …, 0] T ⎥
K1j
⎢⎣ ∼T
T [0, …, 0] T [0, …, 0] T ⎥⎦
∼T ∼
(27)
⎡ [0, …, 0] ∼
[0, …, 0] T ⎤
l ⎢ ⎥
K2j = ⎢ ∼T ∼T ⎡l1 l Ns ⎤ ∼ ⎥
⎢ T [0, …, 0] T ⎢⎣K bb δ1j, …, K bb δ Ns j ⎥⎦ T ⎥
⎣ ⎦ (28)
where δs0 = 1 if the substructure s does not depend on the vector model parameters and δs0 = 0 otherwise, and δsj = 1 if s ∈ Sj and δsj = 0 otherwise.
See [44] for a detailed derivation of these matrices.
References 2011;89:1323–32.
[23] Ching J, Chen YC. Transitional markov chain monte carlo method for bayesian
updating, model class selection, and model averaging. J Eng Mech
[1] Sindel R, Rackwitz R. Problems and solution strategies in reliability updating. J 2007;133:816–32.
Offshore Mech Arct Eng 1998;120(2):109–14. [24] Betz W, Papaioannou I, Straub D. Transitional markov chain Monte Carlo:
[2] Papadimitriou C, Beck JL, Katafygiotis L. Updating robust reliability using observations and improvements. J Eng Mech 2016;142(5).
structural test data. Probabilistic Eng Mech 2001;16:103–13. [25] Chopin N. A sequential particle filter method for static models. Biometrika, vol.
[3] Cheung SH, Beck JL. Bayesian model updating using hybrid monte carlo simulation 89(3). p. 539–52.
with application to structural dynamic models with many uncertain parameters. J [26] Angelikopoulus P, Papadimitriou C, Koumoutsakos P. X-TMCMC: adaptive kriging
Eng Mech 2009;1135(4):243–55. for bayesian inverse modeling. Comput Methods Appl Mech Eng
[4] Jensen HA, Vergara C, Papadimitriou C, Millas E. The use of updated robust 2015;289:409–28.
reliability measures in stochastic dynamical systems. Comput Methods Appl Mech [27] Hadjidoukas P, Angelikopoulus P, Papadimitriou C, Koumoutsakos P. Π 4U: a high
Eng 2013;267:293–317. performance computing framework for Bayesian uncertainty quantification of
[5] Zheng W, Cheng TY. Novel probabilistic approach to assessing barge-bridge complex models. J Comput Phys 2015;284:1–21.
collision damage based on vibration measurements through transitional Markov [28] Ortiz GA, Alvarez DA, Bedoya-Ruiz D. Identification of bouc wen type models using
chain Monte Carlo sampling. J Struct Health Monit 2014;14(2):119–31. the transitional markov chain monte carlo method. Comput Struct
[6] Deodatis G, Asada H, Ito S. Reliability of aircraft structures under nonperiodic 2015;146:252–69.
inspection – a Bayesian approach. J Eng Fract Mech 1996;53(5):789–805. [29] Santner T, Williams B, Notz W. The design and analysis of computer experiments.
[7] Beck J, Katafygiotis L. Updating models and their uncertainties. I: bayesian New York: Springer; 2003.
statistical framework. J Eng Mech 1998;124(4):455–61. [30] Craig RR. Jr., Structural dynamics, an introduction to computer methods. New
[8] Katafygiotis L, Beck J. Updating models and their uncertainties. II: model York: John Wiley & Sons; 1981.
identifiability. J Eng Mech 1998;124(4):463–7. [31] Papadimitriou C, Ch. Papadioti D. Component mode synthesis techniques for finite
[9] Shoji H, Shinozuka M, Sampath S. A bayesian evaluation of simulation models of element model updating. Comput Struct 2013;126:15–28.
multiple-site fatigue crack. Probabilistic Eng Mech 2001;16:355–61. [32] Jensen HA, Muñoz A, Papadimitriou C, Millas E. Model-reduction techniques for
[10] Beck JL, Au SK. Bayesian updating of structural models and reliability using reliability-based design problems of complex structural systems. Reliab Eng Syst
markov chain monte carlo simulation. J Eng Mech 2002;128(4):380–91. Saf 2016;149:204–17.
[11] Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis. Boca Raton, [33] Beck JL, Yuen KV. Model selection using response measurements: Bayesian
Florida: CRC Press; 2003. probabilistic approach. J Eng Mech 2004;130(2):192–203.
[12] Beck JL. Bayesian system identification based on probability logic. Struct Control [34] Simoen E, Papadimitriou C, Lombaert G. On prediction error correlation in
Health Monit 2010;17(7):825–47. bayesian model updating. J Sound Vib 2013;332(18):4136–52.
[13] Yuen KV. Bayesian methods for structural dynamics and civil engineering. [35] Metropolis N, Resenbluth A, Resenbluth M, Teller A, Teller E. Equations of state
Singapore: John Wiley & Sons; 2010. calculations by fast computing machines. J Chem Phys 1953;21(6):1087–92.
[14] Bolstad WM. Understanding computational Bayesian statistics. John Wiley & [36] Hastings W. Monte carlo sampling methods using markov chains and their
Sons. vol. 644; 2004. applications. Biometrika 1970;57(1):97–109.
[15] Yuen KV, Kuok SC. Bayesian methods for updating dynamic models. Appl Mech [37] Hohenbichler M, Rackwitz R. Non-normal dependent vectors in structural safety. J
Rev 2011;64(1), [010802-1-010802-18]. Eng Mech Div 1981;107(6):1227–38.
[16] Jensen HA, Millas E, Kusanovic D, Papadimitriou C. Model reduction techniques [38] Der Kiureghian A, Liu PL. Structural reliability under incomplete probability
for Bayesian finite element model updating using dynamic response data. Comput information. J Eng Mech 1986;112(1):85–104.
Methods Appl Mech Eng 2014;279:301–24. [39] Ditlevsen O, Madsen OH. Structural reliability methods. Chichester: John Wiley
[17] Straub D, Papaioannou I. Bayesian updating with structural reliability methods. J and Sons; 1996.
Eng Mech 2015;141(3):04014134. [40] Zuev KM, Beck JL, Au SK, Katafygiotis L. Bayesian post-processor and other
[18] Au SK. Insights on the Bayesian spectral density method for operational modal enhancements of Subset Simulation for estimating failure probabilities in high
analysis. Mech Syst Signal Process 2016;66–67:1–12. dimensions. Comput Struct 2012;92–92:283–96.
[19] Au SK, Zhang FL. Fundamental two-stage formulation for Bayesian system [41] Sacks J, Welch WJ, Mitchell TJ, Wynn HP. Designs and analysis of computer
identification, Part I: general theory. Mech Syst Signal Process 2016;66–67:31–42. experiments. Stat Sci 1989;4(4):409–35.
[20] Zhang FL, Au SK. Fundamental two-stage formulation for bayesian system [42] Rusmussenand C, Williams C. Gaussian processes for machine learning. Adaptive
identification, Part II: application to ambient vibration data. Mech Syst Signal computation and machine learning. Cambridge, Massachussetts: MIT Press; 2006.
Process 2016;66–67:43–61. [43] Lophaven SN, Nielsen HB, J. Søndergaard. DACE: A MATLAB Kriging Toolbox.
[21] Gilks W, Richardson S, Spiegelhalter D. Markov Chain Monte Carlo in practice. DTU. Lyngby, Denmark; 2002.
CRC Press; 1996. [44] Jensen HA, Muñoz A, Papadimitriou C, Vergara C. An enhanced substructure
[22] Nichols JM, Moore EZ, Murphy KD. Bayesian identification of a cracked plate using coupling technique for dynamic re-analyses: application to simulation-based
a population-based markov chain monte carlo method. Comput Struct problems. Comput Methods Appl Mech Eng 2016;307:215–34.
189
[45] Bathe J. Finite element procedures. Upper Saddle River, New Jersey: Prentice Hall; [49] Minewaki S, Yamamoto M, Higashino M, Hamaguchi H, Kyuke H, Sone T, et al.
2006. Performance tests of full size isolators for super high-rise isolated buildings. J
[46] Barber CB, Dobkin DP, Huhdanpaa H. The quickhull algorithm for convex hulls. Struct Eng AIJ 2009;55:469–77, [B].
ACM Trans Math Softw 1996;22:469–83. [50] Yamamoto M, Minewaki S, Yoneda H, Higashimo M. Nonlinear behavior of high-
[47] Angelikopoulos P, Papadimitriou C, Koumoutsakos P. Bayesian uncertainty quan- damping rubber bearings under horizontal bidirectional loading: full-scale tests
tification and propagation in molecular dynamics simulations: a high performance and analytical modeling. Earthq Eng Struct Dyn 2012;41(13):1845–60.
computing framework. J Chem Phys 2012;137, [1441103-1-144103-19]. [51] Jensen HA, Kusanovic DS. On the effects of near-field excitations on the reliability-
[48] Pellissetti MF. Parallel processing in structural reliability. J Struct Eng Mech based performance and design of base-isolated structures. Probabilistic Eng Mech
2009;32(1):95–126. 2014;36:28–44.
190

Bayseian Time Domain

Uploaded by

Copyright:

Available Formats

Bayseian Time Domain

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bayseian Time Domain

Uploaded by

Copyright:

Available Formats

Reliability Engineering and System Safety 160 (2017) 174–190

Contents lists available at ScienceDirect

Reliability Engineering and System Safety

Implementation of an adaptive meta-model for Bayesian ﬁnite element MARK

1. Introduction requires a large number of system simulations. Thus, the computa-

integrated eﬃciently by an appropriate step-by-step integration σ Jl z (z new)

4.2. Reduced-order model parametrization

Fig. 1. Isometric view of the structural model. Example 1.

modeled with shell elements with a thickness of 0.3 m and beam

Nm=100 for monitoring the acceptance rate. The updating problem

TMCMC Mean Standard COV

in Figs. 12 and 13, respectively. It is seen that the distribution of the ⎧ u (t , θ ) ⎫ ⎡ M e (θ ) Me (θ ) Ge ⎤

Table 3 value (ϵ = 0.01) the overall acceptance ratio is approximately 68 %. On

Method Number Efficiency

Original version of TMCMC with reduced-order FE model 6040 1

vance. In this case, the removal of potential high frequency compo-

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.