Mechanical Systems and Signal Processing 107 (2018) 502–514

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Uncertainty quantification and propagation in dynamic models

using ambient vibration measurements, application to a
10-story building
Iman Behmanesh a,⇑, Seyedsina Yousefianmoghadam b, Amin Nozari c, Babak Moaveni c,
Andreas Stavridis b
a
Building Structures, WSP USA, New York, NY, USA
b
University at Buffalo, New York, NY, USA
c
Tufts University, Medford, MA, USA

a r t i c l e i n f o a b s t r a c t

Article history: This paper investigates the application of Hierarchical Bayesian model updating for uncer-
Received 26 May 2017 tainty quantification and response prediction of civil structures. In this updating frame-
Received in revised form 6 December 2017 work, structural parameters of an initial finite element (FE) model (e.g., stiffness or
Accepted 22 January 2018
mass) are calibrated by minimizing error functions between the identified modal parame-
ters and the corresponding parameters of the model. These error functions are assumed to
have Gaussian probability distributions with unknown parameters to be determined. The
Keywords:
estimated parameters of error functions represent the uncertainty of the calibrated model
Hierarchical Bayesian modeling
Model updating
in predicting building’s response (modal parameters here). The focus of this paper is to
Uncertainty quantification answer whether the quantified model uncertainties using dynamic measurement at build-
Modeling errors ing’s reference/calibration state can be used to improve the model prediction accuracies at
Prediction bias a different structural state, e.g., damaged structure. Also, the effects of prediction error bias
Uncertainty propagation on the uncertainty of the predicted values is studied. The test structure considered here is a
ten-story concrete building located in Utica, NY. The modal parameters of the building at
its reference state are identified from ambient vibration data and used to calibrate param-
eters of the initial FE model as well as the error functions. Before demolishing the building,
six of its exterior walls were removed and ambient vibration measurements were also
collected from the structure after the wall removal. These data are not used to calibrate
the model; they are only used to assess the predicted results. The model updating
framework proposed in this paper is applied to estimate the modal parameters of the
building at its reference state as well as two damaged states: moderate damage (removal
of four walls) and severe damage (removal of six walls). Good agreement is observed
between the model-predicted modal parameters and those identified from vibration tests.
Moreover, it is shown that including prediction error bias in the updating process instead
of commonly-used zero-mean error function can significantly reduce the prediction
uncertainties.
Ó 2018 Elsevier Ltd. All rights reserved.

⇑ Corresponding author.
E-mail address: iman.behmanesh@wsp.com (I. Behmanesh).

https://doi.org/10.1016/j.ymssp.2018.01.033
0888-3270/Ó 2018 Elsevier Ltd. All rights reserved.
 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514 503

1. Introduction

The dynamic response of civil structures to external loadings is often predicted using finite element (FE) models. Exam-
ples include prediction of structural response to wind or seismic loads. It is common to validate linear FE models by com-
paring their response or features extracted from response such as modal parameters with the ones identified from vibration
data [1–6]. The FE models of civil structures are often associated with many sources of uncertainty and modeling errors;
Haukaas and Gardoni [7] provide a discussion on different sources of uncertainty in FE structural models. These uncertainty
sources affect the design and performance assessment of civil structures when these FE models are used for predicting a
quantity of interest [2,8–17]. Calibration/updating of the FE models [18,19] using static and/or dynamic data can reduce
some of these uncertainties [20–22]. In the model updating process, selected parameters of the model are adjusted so that
the model-predicted response time histories, modal parameters, or frequency response functions best match the corre-
sponding quantities obtained from the test data. In most cases, the updated FE models still have bias in predicting the data
that are used for model calibration [4,20,23–26]. This problem is not limited to the deterministic model updating frame-
works; the accuracy of probabilistic updating frameworks, as presented in [12,27,28], has been discussed by a number of
researchers. In [29], the effects of weighting factors (or prediction error variances) on the model updating results in the pres-
ence of modeling errors are studied, and Pareto optimal models are used to provide more robust identifications in the pres-
ence of modeling errors. Ching and Beck [30] report modeling errors to be the major source of inaccurate damage
identification results in their case study. In [31–34], the authors propose model falsification instead of model identification
to overcome the challenges of accounting for modeling errors in model identification. One major argument in the last three
studies is that the estimated structural parameters from Bayesian updating techniques are biased from their true values in
the presence of modeling errors. They also discuss that models cannot be fully validated, and therefore, they utilize the con-
cept of model falsification as an alternative approach to handle the uncertainty of FE models. Prediction error bias is one of
the major drawbacks of Bayesian identifications. Goller and Schueller [23] investigate this problem through Bayesian model
updating using zero-mean Gaussian error functions and conclude that the prediction bias of the updated FE models should
be considered when the model is used for predictions.
This paper presents an improved model identification technique that aims to quantify the uncertainty of a model in pre-
dicting a specific quantity of interest. The estimated uncertainty of the model is then propagated in model-predicted quan-
tities. To this end, the Hierarchical Bayesian Modeling [35–37] is used to calibrate the initial FE model of a 10-story concrete
building using identified natural frequencies and mode shapes of the building extracted from ambient vibration measure-
ments. A probability distribution with unknown mean and covariance matrix is considered for the error function between
the identifiable modal parameters and those of the model. Parameters of the probability distribution are considered as
updating parameters. The effects of prediction error bias are studied by considering an error function once as zero-mean
and once as non-zero mean random variable. The calibrated FE model and the prediction error model are used to predict
the natural frequencies of the building at different states, i.e., after removing four and six of its exterior walls, respectively.
It is shown that considering the mean of error functions as updating parameters reduces the prediction error variances and
the commonly-used zero-mean error function may result in large prediction uncertainty. This paper is the extension of the
work presented by the authors in conference paper [38].

2. Model calibration in the presence of modeling errors

Model calibration aims to reduce the discrepancy between model-predicted features (here natural frequencies and mode
shapes) and their experimentally-identified counterparts by tuning a set of model parameters. As it has been discussed in
past studies, tuning structural parameters alone may not fully eliminate the misfit between the data and model predictions
(i.e., bias) because of modeling error effects [7,23,27,31]. The probabilistic FE model calibration method used in this study
includes the structural updating parameters as well as modeling error parameters to capture the remaining misfit between
identified and model-predicted response features. The following error functions are defined for the natural frequency and
model shape of mode m:

x2m ðht Þ  x
~ 2tm
extm ¼  m ¼ 1 : Nm ð1Þ
x
~m 2

CUm ðht Þ ~ tm
U
eUtm ¼ cosðutm Þ  ð2Þ
~ tm k
kCUm ðht Þk kU

In Eqs. (1) and (2) x ~ tm are the circular natural frequency and mode shape of mode m identified using data set t,
~ tm and U
and x 
~ m is the average of identified natural frequencies over all data sets. xm ðht Þ and Um ðht Þ are the model-calculated natural
frequency and mode shape of mode m given ht, where ht is the vector of structural parameters corresponding to data set t. Nm
is the number of identified modes, matrix C picks the measured mode shape components, and utm is the angle between
identified and model-calculated mode shapes which can be estimated using Eq. (3).
504 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514

~ T  ðCUm ðht ÞÞ
U
cosðutm Þ ¼ tm
ð3Þ
kCUm ðht ÞkkU~ tm k

As described in [39], the error functions (1) and (2) ‘‘provide a bridge” between the measured quantities of interest (nat-
ural frequencies and mode shapes) and the FE model output. The mode shape error function of Eq. (2) is similar to the func-
tion originally used in [37]. The eigenvalue (i.e., squared natural frequency) error function of Eq. (1), however, is slightly
different from [37]. The eigenvalue error function of this study is more suitable for implementation in response prediction
and model extrapolation. Note that different mode shape error functions can be defined and used instead of Eq. (2). However,
the vector error function of Eq. (2) provides the ability to estimate statistics of different mode shape components which are
not available when using a scalar error function such as eUtm ¼ 1  MACðU ~ T  ðCUm ðht ÞÞÞ. It also allows evaluation of error in
tm
model-predicted mode shapes at each measured degree of freedom.
The error functions are modeled as Gaussian distributions with mean le and covariance matrix Re [36,39,40]:


ext
et ¼  Nðle ; Re Þ ð4Þ
eUt
In the absence of any prior knowledge about the distribution, the Gaussian distribution would be the optimal choice for
error functions in order to maximize the potential uncertainty (principle of maximum entropy). In this study the error func-
tion distributions are not limited to zero-mean (unbiased) distributions. Furthermore, considering Gaussian distribution for
the error functions will simplify the conditional probability terms. The number of independent updating parameters in le
and Re can be defined by the analyst. Without introducing any constrains, le should be considered as a full vector and Re
as a full symmetric matrix with all their components considered as updating parameters except the symmetric components
of Re . However, this approach increases the computational cost and complicates the response prediction process. Excluding
the correlations between the error functions reduces N m ðN s þ 1ÞðN m ðN s þ 1Þ  1Þ=2 updating parameters, which decreases
the computational cost. Ns is the total number of identified mode shape components which is equal to the number of sensors.
If the correlations between the error functions are not considered, the error functions can be written as:

extm  Nðlxm ; r2xm Þ ð5Þ

eUtm  NðlUm ; diagðr2/m1 ; . . . ; r2/ms ; . . .ÞÞ s ¼ 1 : Ns ð6Þ

where r2/ms is the variance of sth component of the mth mode shape error.
The posterior joint probability distribution of the updating model parameters can be written as:
Y
Nt Y
Nm
pðH; lh ; Rh ; le ; Re jx ~ /
~ 2 ; UÞ pðx ~ tm jht ; l ; Re Þpðht jl ; Rh Þpðl ÞpðRh Þpðl ÞpðRe Þ
~ 2tm ; U ð7Þ
e h h e
t¼1 m¼1

where H ¼ fh1 ; . . . ; ht ; . . . ; hNt g, and Nt is the total number of available data sets. Note that the non-diagonal terms in covari-
ance matrices Rh and Re are zero, thus Rh includes Np (number of structural parameters) variance terms r2hp , and Re includes
Nm variance terms for eigenvalue errors r2 and Ns Nm mode shape error variances r2 . In Eq. (7), pðx ~2 ;U~ tm jht ; l ; Re Þ is
xm /ms tm e

the likelihood function, assessing the probability of how plausible model parameters ht , le and Re represent the data x ~ 2tm and
~ tm . The prior pðht jl ; Rh Þ is considered to be a Gaussian distribution Nðht jl ; Rh Þ. Uninformative (i.e., uniform) priors are
U h h
assumed for lh and le , i.e., pðlh Þpðle Þ / 1, and an Inverse-Gamma distribution is considered for the variance terms,
pðr2hp Þ ¼ Inv erseGammaða; 1=bÞ, pðr2xm Þ ¼ Inv erseGammaðae ; 1=be Þ and pðr2Ums Þ ¼ Inv erseGammaðae ; 1=be Þ. The considered
priors in this study are selected as proposed in [36,40]. Gelman [40] suggests the use of uninformative priors as a starting
point (‘‘reference model”) and then re-evaluate this choice after observing the resulted posterior distributions. In this paper,
no information is available for le and thus uninformative prior is the only option. Uninformative lh is also selected to allow
the updating process exploring the parameter space with no prior constraint. The inverse gamma distribution as the prior for
variance terms is a conjugate prior, which makes the conditional posterior distribution of the variance terms as gamma dis-
tribution and improves the computational efficiency of sampling process.
It can be shown that the maximum a-posteriori (MAP) estimates of the updating parameters can be obtained from [37]:
^ht ¼ ArgminðJðht ; l ^ e Þ þ ðht  l
^ e; R ^ 1 ðht  l
^ h ÞT R ^ h ÞÞ
h
ht
ð8Þ
where Jðht ; l ^ e Þ ¼ ðet  l
^ e; R ^ 1 ðet  l
^ e ÞT R ^eÞ
e

1X Nt
^ht
l^ h ¼ ð9Þ
Nt t¼1

PNt ^ 2
t¼1 ðhtp  lhp Þ
2
þ ^
r^ 2hp ¼ b p ¼ 1 : Np ð10Þ
N t  2 þ 2a
 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514 505

1X Nt
l^ e ¼ ^t
e ð11Þ
Nt t¼1

PNt
^ l
^ xm Þ2
2
þ t¼1 ðextm
r^ 2xm ¼ be m ¼ 1 : Nm
Nt  2 þ 2ae
P t ð12Þ
2
þ Nt¼1 ð^e/tms  l
^ /ms Þ2
r
^ 2/ ¼
be
m ¼ 1 : Nm s ¼ 1 : Ns
ms
Nt  2 þ 2ae
The MAP estimates for the parameters in Eqs. (8)–(12) are achieved through multiple iterations. Note that one can make
an analogy of Eqs. (8)–(12) to the estimated parameters from the frequentist approach. Eq. (8) is a combination of the func-
tion J which is the weighted error squares, and a regularization term ðht  l ^ 1 ðht  l
^ h ÞT R ^ h Þ. Eq. (10) and (12) become the
h
error variances if a ¼ ae ¼ 0:5 and parameters b and be are large enough. Eq. (12) indicates that the prediction error variance
is a function of the prior knowledge (ae and be) and the discrepancies between model-predicted and identified eigenvalues
P t
and model shapes. Therefore, even if the model updating process results in a perfect fit, i.e., Nt¼1 et  l
ð^ ^ Þ2 ! 0, the prediction
2=be
error variances will converge to a non-zero value of Nt 22 ae . This value can be regarded as the prior uncertainty of the model
imposed by the analyst.
If the prediction errors are modeled as a zero-mean jointly Gaussian distribution, the error variances of Eq. (12) can be
written as:

PNt
2
þ ^2
t¼1 extm
rxm ¼
^2 be
m ¼ 1 : Nm
Nt  2 þ 2ae
P t 2 ð13Þ
2
þ Nt¼1 ^e/
r
^ 2/ ¼
be tms
m ¼ 1 : Nm s ¼ 1 : Ns
ms
N t  2 þ 2ae

By comparing Eqs. (12) and (13), it can be inferred that excluding the error means from the updating process results in
higher error variances r2xm and r2/ms . Hence, by setting the error mean values to zero, the error variances will increase to
compensate for the bias resulting in more conservative model uncertainty estimates.
To reduce the number of iterations required for the convergence of updating parameters, Eq. (8) is replaced by the first
two non-zero terms of its Taylor series expansion [36,41]:

Jðht ; l ^ e Þ  Jð^ht ; l
^e; R ^ e Þ þ 1 ðht  ^ht ÞT Hð^ht Þðht  ^ht Þ
^ e; R ð14Þ
2

where Hð^ ht Þ is the Hessian of Jðht ; l ^t . Computational details for calculating the derivatives of
^ e Þ at the optimum value h
^ e; R
eigenvalues and mode shapes over the structural parameter ht are provided in [41,42]. This simplification can be avoided
by using advanced sampling techniques [43,44].
As discussed earlier, calibrated FE models are often used to predict quantities of interest such as natural frequencies and
mode shapes. By rewriting Eq. (1), the mth eigenvalue of the structure at a future state FS can be estimated as:

^ ^ ~ 2
m Þ ¼ xm ðhjq; lh ; Rh Þ  xm  ðexm jlxm ; rxm Þ
2
ðxFS 2 ^ ^2 m ¼ 1 : Nm ð15Þ

where the vector of known changes in structural mass and/or stiffness properties in the state FS is shown by q. In Eq. (15),
hjq; l ^ h represents random simulations of h from Nðl
^ h; R ^ h Þ after changes q are applied to the updated FE model. Also in this
^ h; R
equation, exm jl ^ xm ; r
^ xm represents random simulations of exm from Nðl
2 ^ xm ; r
^ 2xm Þ. The parameters l ^x, l
^ x, R ^ h are esti-
^ h , and R
mated in the calibration process and assumed to be the same at the future state. Similarly, the structural mode shapes at a
future state FS can be predicted as:

CUm ðhjq; l ^hÞ

^h; R
m ¼ cosðum Þ
UFS  ðeUm jlUm ; diagðr2/m1 ; . . . ; r2/ms ; . . .ÞÞ ð16Þ
kCUm ðhjq; l ^ h Þk
^h; R

The predicted xFS and UFS values of Eqs. (15) and (16) can be used to predict displacement, velocity, or acceleration
response of the structure to external vibration loads.
Note that if a mode is not identified from the vibration data or is not included in the updating process, no posterior values
for lxm and r2xm would be available for that mode. Therefore, the lxm and r2xm of such mode rely on the prior information and
engineering judgement. For example, the maximum values of fr2x1 ; r2x2 ; . . . ; r2xNm g of Nm estimated eigenvalue error
variances can be considered as the eigenvalue error variance of an unidentified mode.
506 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514

3. Case study: 10 story concrete building

3.1. Building description and instrumentation

The structure, shown in Fig. 1 was a 10-story, eight-bay by four-bay concrete building with a slab-column structural sys-
tem and reinforced concrete infill walls in the exterior bays. The building, constructed in 1914, was located in Utica, NY and
was used as a warehouse before it was slated to be demolished to allow the construction of a new highway bridge. The plan
view dimensions of the structure were 24.4 m by 48.8 m (80 ft by 160 ft) while the story height was 2.59 m (80 600 ). Fig. 2 pre-
sents the plan view of a typical story. In this building, the bays had the same length in both directions. The concrete columns
were circular in the interior, and rectangular along the perimeter, except for the corner columns which had L-shape cross
sections. Due to the lack of beams, all columns had corbels at their connections to the slab above. Flat slabs, having a thick-
ness of 23 cm (9 in.), with drop panels at the location of columns were used in the structure. The RC walls in the perimeter
bays had a thickness of 20 cm (8 in.). Some of the walls initially had window openings. The majority of these openings were
later infilled with concrete masonry units, while in some of the walls, the window had been partially filled with clay
masonry. More information about the building can be found in [45].
The acceleration along the vertical and the two horizontal directions at two locations of every slab except the roof were
recorded using force-balance accelerometers. The accelerometers were installed near the North-West (NW) and South-East
(SE) corners, as indicated in Fig. 2, to monitor both the translational and torsional motions. The recorded acceleration time
histories are divided into 5-min intervals and structural modal parameters are extracted from each 5-min ambient vibration
data. Overall, 343 sets of identified modal parameters are extracted, 50 of which are used for model calibration. The Natural
Excitation Technique combined with Eigensystem Realization Algorithm (NExT-ERA) [46–48] is used to extract the modal
parameters. Table 1 presents the statistics of the identified natural frequencies, and Fig. 3 shows the identified mode shapes
of the first five modes. In Fig. 3, the dark and grey lines show model-predicted mode shapes at SW and NE corners of floors,
respectively, the red stars are identified mode shapes at SW corner and the green stars correspond to the NE corners of floor
plan.

3.2. Initial FE model

An FE model of the structure was created using the SAP2000 [49] structural analysis software, and then transferred to the
MATLAB-based structural analysis tool, FEDEASLAB [50,51]. This model included 2914 nodes, 2555 frame elements and 2593

Fig. 1. Building view from south west corner.

Fig. 2. Plan view of second story showing accelerometers position in each story.
 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514 507

Table 1
The identified and model-calculated natural frequencies at structural reference state.

Model 1 Mode 2 Mode 3 Mode 4 Mode 5

Mean Standard Mean Standard Mean Standard Mean Standard Mean Standard
Deviation Deviation Deviation Deviation Deviation
Identified natural frequencies 2.24 0.008 3.34 0.013 4.71 0.034 7.35 0.070 12.38 0.141
MAC between identified and initial 0.99 0.010 0.99 0.003 0.97 0.009 0.98 0.009 0.92 0.039
FE model mode shapes
Natural frequencies of initial FE 2.23 3.34 4.71 7.41 11.11
model

Mode 1 Mode 2 Mode 3 Mode 4 Mode 5

10 10 10 10 10
9 9 9 9 9
8 8 8 8 8
7 7 7 7 7
Story

6 6 6 6 6
5 5 5 5 5
4 4 4 4 4
3 3 3 3 3
2 2 2 2 2
1 1 1 1 1

−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4

Mode 1 Mode 2 Mode 3 Mode 4 Mode 5

10 10 10 10 10
9 9 9 9 9
8 8 8 8 8
7 7 7 7 7
Story

6 6 6 6 6
5 5 5 5 5
4 4 4 4 4
3 3 3 3 3
2 2 2 2 2
1 1 1 1 1

−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4

Fig. 3. Building’s first five mode shapes; black line: FE SW corner, grey line: FE NE corner, red dots: identified SW corner; green dots: identified NE corner;
top row: X (north-south) components, bottom row: Y (east-west) components. (For interpretation of the references to color in this figure legend, the reader
is referred to the web version of this article.)

shell elements. The model was computationally expensive to be used in the FE model updating process where several iter-
ations are required for each updating. There is a rich literature on reducing the computational cost of FE model updating
applications of civil structures, which include the use of surrogate models [52,53], parallel computing for stochastic simu-
lations [44,54,55], model reduction [43,56], or model simplification [28]. In this study, four simplifying assumptions com-
mon for building structures are considered to reduce the size of the model and increase the computational speed. These
assumptions include (1) modeling the floor slabs as beam elements [57], (2) constraining the floors to have rigid in-plane
rotation at each level (rigid diaphragm assumption), (3) assigning the building mass as lumped mass at floor levels, and
(4) neglecting the nodal masses in the vertical direction. These simplifying assumptions result in the initial model with
30 condensed (non-zero mass) degrees of freedoms (DOFs), including the motions in global X and Y directions, and the
in-plane rotation of each diaphragm. This simplified FE model is referred to as the initial FE model in this paper. The natural
frequencies of the initial model are reported for the first five modes in Table 1. The statistics of the MAC values over the 50
508 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514

sets of the identified mode shapes and the mode shapes of the initial model are also reported in Table 1. As it can be seen, the
natural frequencies and mode shapes of the initial model are in relatively good agreement with their identified counterparts
except for the fifth mode. Note that only five modes are reliably identified from the recorded data.

4. Application to the 10-story concrete building

In this section, the initial FE model of the considered building is first calibrated using the discussed Hierarchical Bayesian
framework, and its uncertainty in predicting the first five frequencies and mode shapes are estimated. Next, the calibrated
model is used to probabilistically predict the building first five natural frequencies with the estimated model/parameter
uncertainties propagated in the predicted natural frequencies. The predictions are compared with measured data. The effects
of different modeling assumptions of the error functions are investigated on the calibration results and the resulted model
prediction uncertainties. More specifically, the effects of prediction error bias and variances are discussed in detail.

4.1. Model updating considering zero-mean prediction errors

The initial FE model of the building is calibrated using 50 randomly selected sets of identified modal parameters out of the
343 available data sets. The remaining data sets are used for validation purposes in the prediction step. Two structural updat-
ing parameters are considered for model calibration in this section, namely, the effective Young’s modulus of exterior walls
(h1) and the effective Young’s modulus of the slab beams (h2). In this section, the error functions are modeled as zero-mean
Gaussian distributions. The updating parameters include Nt = 50 ht values, mean lh ¼ ½ lh1 lh2  and variances ½ r2h1 r2h2  of
the structural updating parameters (diagonal terms of Rh ), variances of the five eigenvalue error functions
½ r2x1 r2x2 r2x3 r2x4 r2x5  (diagonal terms of Rx ), and 135 (5 modes and 27 mode shape components for each mode) vari-
ances of mode shape error functions. The chosen values for a, b, ae and be in Eq. (13) are 0.5, 200, 0.5 and 1000, respectively.
Table 2 reports the maximum a-posteriori (MAP) updating parameters from the process explained in Section 2. As it can be
seen from this table, the calibrated model can predict the first and third natural frequencies more accurately with error stan-
dard deviations of 0.010 and 0.007, respectively, while the frequency of the fifth mode has the largest error standard devi-
ation of 0.188. The mode shapes of modes 1 and 2 are fit more closely than mode shapes of modes 3–5, and the 5th mode
shape is the least accurate mode shape predicted by the calibrated FE model. Note that the minimum standard deviation rx
is set as 0.006 given the considered priors.
To show the contribution of prediction error variances in the model prediction process (second term of Eq. (15)), the nat-
ural frequencies are predicted by the calibrated model once without including the prediction error variances (lower triangle
subplots of Fig. 4), and once when accounting for the modeling errors (upper triangle subplots of Fig. 4). The results are plot-
ted together with all the 343 sets of identified natural frequencies. The model-predicted frequencies are shown with grey
circles while the identified frequencies are shown with black stars. The subplots in the upper triangle include the variability
of structural updating parameters, Rh , and modeling errors parameters, Rx ¼ ½ r2x1 r2x2 r2x3 r2x4 r2x5  while the lower
triangular subplots compare the identified and the FE model-predicted frequencies considering the variability of structural
model parameters Rh only. This plot clearly shows the importance of including the modeling error when a probabilistically
calibrated model is used for predicting a quantity of interest. By including the modeling error parameters, all the identified
frequencies fall in the high probability region of the model predictions. Moreover, the strong correlations in the predicted

Table 2
updating results for the three model updating cases.

2 structural parameters; zero mean 2 structural parameters; non-zero mean error

error functions (Section 4.1) functions for 5th eigenvalue (Section 4.2)
Mean and standard deviation of l^ h1 0.995 0.995
structural parameters r^ h1 0.021 0.021
l^ h2 1.132 1.129
r^ h2 0.015 0.014

Eigenvalue error standard deviations r^ x1 0.010 0.011

r^ x2 0.014 0.014
r^ x3 0.007 0.007
r^ x4 0.034 0.034
r^ x5 0.188 0.021

Error Bias lx5 0 0.185

Mode shape error standard TraceðRU1 Þ 0.371 0.371

deviations TraceðRU2 Þ 0.401 0.401
TraceðRU3 Þ 0.693 0.693
TraceðRU4 Þ 0.642 0.642
TraceðRU5 Þ 1.100 1.100
 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514 509

2.30

f1 [Hz]
2.25

2.20

2.15
3.20 3.30 3.40 3.50
3.50 f2 [Hz] 3.50
f2 [Hz]

f2 [Hz]
3.40 3.40

3.30 3.30

3.20 3.20
4.50 4.65 4.80 4.95
4.95 f3 [Hz] 4.95
f3 [Hz]

f3 [Hz]
4.80 4.80

4.65 4.65

4.50 4.50
6.80 7.20 7.60 8.00
8.00 f4 [Hz] 8.00
f4 [Hz]

f4 [Hz]
7.60 7.60

7.20 7.20

6.80 6.80
6.00 9.00 12.00 15.00
15.00 f5 [Hz]
f5 [Hz]

12.00

9.00

6.00
2.15 2.20 2.25 2.30 3.20 3.30 3.40 3.50 4.50 4.65 4.80 4.95 6.80 7.20 7.60 8.00
f1 [Hz] f2 [Hz] f3 [Hz] f4 [Hz]

Fig. 4. Identified vs. model-calculated natural frequencies using zero-mean Gaussian prediction errors. Black stars: identified natural frequencies, grey
crosses: model predicted natural frequencies without modeling error, grey circles: model-predicted natural frequencies considering modeling errors.

frequencies of FE-model is diminished after adding the prediction error variances, i.e., the distribution of model predicted
frequencies becomes closer to that of identified frequencies after including prediction error variances.
It should be noted that in some cases including the error variances may result in significantly large uncertainties for
model predictions. For example, the model-predicted frequency of the fifth mode varies from around 7 Hz to 15 Hz. This
is because the FE model, even after calibration, underestimates the frequency of the fifth mode, as indicated in the
bottom row of Fig. 4. This bias makes the prediction error standard deviation of the fifth frequency too large, and con-
sequently results in high prediction uncertainty of this frequency (see 5th column of Fig. 4). Therefore, modeling the error
functions as zero-mean Gaussian distribution may not be justifiable is some cases. As discussed in [39] and presented in
Eqs. (12) and (13), the zero-mean model of error function results in higher standard deviation and higher (or more con-
servative [39]) overall prediction uncertainty. Thus, when tuning the structural parameters alone cannot reduce the bias,
the prediction error mean should be included in the updating process, i.e., the error function should not be modeled with
zero mean.

4.2. Model updating considering prediction error bias

In this section, the prediction error bias of the fifth mode is considered as an additional updating parameter. The most
probable updating parameters, including lx5 , are shown in Table 2. The estimated MAP values are very close to those
estimated in the previous section, except for the prediction error standard deviation of the fifth mode that is dropped
from 0.188 to 0.021. The model-predicted natural frequencies versus their identified counterparts are shown in Fig. 5.
Compared to Fig. 4, the uncertainty of the fifth frequency prediction is drastically reduced and the identified frequencies
fall in the high probability region of the model predictions. Note that, the prediction uncertainty of the fourth mode
510 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514

2.30

f1 [Hz]
2.25

2.20

2.15
3.20 3.30 3.40 3.50
f2 [Hz] 3.50

f2 [Hz]
3.40

3.30

3.20
4.50 4.65 4.80 4.95
f3 [Hz] 4.95

f3 [Hz]
4.80

4.65

4.50
6.80 7.20 7.60 8.00
f4 [Hz] 8.00

f4 [Hz]
7.60

7.20

6.80
11.5 12.0 12.5 13.0
f5 [Hz]

Fig. 5. Identified vs. model-calculated frequencies using non-zero-mean Gaussian prediction errors for fifth eigenvalue. Black stars: identified natural
frequencies, grey circles: model-predicted natural frequencies considering modeling errors.

might be relatively high and if one wishes to reduce it, the prediction error mean of this frequency can also be considered
as an updating parameter.
The model updating process reduces the discrepancies between measured and model-predicted quantities of interest
(error functions) but cannot fully eliminate the model prediction errors. Therefore, it is important to try to characterize
the model prediction errors by estimating their mean (bias) and standard deviations. Considering non-zero mean error func-
tions and updating the mean values will provide a more accurate estimate of the error standard deviations which will in turn
yield tighter confidence intervals on response predictions. On the other hand, the addition of more structural parameters in
the updating process can potentially reduce the error function bias. However, increasing the number of structural updating
parameter will add to the computational time and can cause the inverse problem to become non-unique. Furthermore, if the
data is not sensitive to certain updating parameters, their prior and posterior probability distributions would be similar, i.e.,
no information would be gained about them through the updating process. Therefore, it is recommended to consider a smal-
ler number of structural updating parameters and to propagate the modeling error uncertainties when the model is used for
prediction rather than calibrating a large set of structural updating parameters especially when the response features are less
sensitive to some of those parameters.

5. Uncertainty propagation in predictions at future states

The calibrated FE models at the reference structural state (where data is available and used for calibration) are often
implemented to predict the structural responses at a future state for which the FE model is not calibrated. The future states
can refer to the structural condition after a damaging event such as an earthquake [58,59] or enhancement to an existing
structure such as retrofit [3,58,60]. In this section, the natural frequencies of the building are estimated for two damaged
states, namely with moderate and severe damage. The moderate damage state refers to the state of building after removing
four exterior walls on the second story as indicated in Fig. 6 while the severe damage state refers to the state after removing
two additional walls in the first story. The calibrated model of Section 4.2 (considering non-zero bias) that is obtained for the
reference state building is used for prediction of natural frequencies (Eq. (15)). The known damage at each state (removal of
4 or 6 walls) is accounted for by parameter q in the equation. As previously mentioned, the considered two states of damage
 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514 511

Fig. 6. Building in the future state; the removed walls are shown by black.

were actually introduced on the real structure and vibration data at each state were measured. Overall, 12 and 22 sets of
5-min ambient vibration data were collected the moderate and damage states of building, respectively, resulting in 12
and 22 sets of identified modal parameters at these states. Note that these data sets are not used for model calibration
and only considered for evaluating the performance of calibrated model in future damage states, i.e, model validation.
Fig. 7 compares the model-predicted natural frequencies of the first five modes with their identified counterparts. From
Fig. 7, it can be seen that the model predicts the natural frequencies of the building at its moderately damage state reason-
ably well as the identified frequencies fall in the high probability region of model predictions. However, small bias in pre-
diction of second mode frequency can be observed. Fig. 8 shows a comparison between model predictions (with and
without considering modeling errors) and the identified natural frequencies at the severe damage state. From this figure,
it can be observed that the model predictions for the first two modal frequencies are not in good agreement with the iden-
tified values. Although the initial FE model and the calibrated model of the building could predict the first four frequencies of

2.29
f1 [Hz]

2.21

2.13

2.05
3.15 3.25 3.35 3.45
f2 [Hz] 3.45
f2 [Hz]

3.35

3.25

3.15
4.45 4.60 4.75 4.90
f3 [Hz] 4.90
f3 [Hz]

4.75

4.60

4.45
6.80 7.20 7.60 8.00
f4 [Hz] 8.00
f4 [Hz]

7.60

7.20

6.80
10.5 11.5 12.5 13.5
f5 [Hz]

Fig. 7. Identified versus model-calculated frequencies of the building in the moderate damage state (removal of four walls on second story). Black stars are
identified and grey circles are model-predicted natural frequencies considering modeling errors.
512 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514

2.14

f1 [Hz]
2.06

1.98

1.90
2.85 2.95 3.05 3.15
3.15 f2 [Hz] 3.15
f2 [Hz]

f2 [Hz]
3.05 3.05

2.95 2.95

2.85 2.85
4.40 4.55 4.70 4.85
4.85 f3 [Hz] 4.85
f3 [Hz]

f3 [Hz]
4.70 4.70

4.55 4.55

4.40 4.40
6.50 6.90 7.30 7.70
7.70 f4 [Hz] 7.70
f4 [Hz]

f4 [Hz]
7.30 7.30

6.90 6.90

6.50 6.50
10.0 11.0 12.0 13.0
13.0 f5 [Hz]
f5 [Hz]

12.0

11.0

10.0
1.90 1.98 2.06 2.14 2.85 2.95 3.05 3.15 4.40 4.55 4.70 4.85 6.50 6.90 7.30 7.70
f1 [Hz] f2 [Hz] f3 [Hz] f4 [Hz]

Fig. 8. Identified versus model-calculated frequencies of the building in the damaged condition. Black stars: identified natural frequencies, grey crosses:
model predicted natural frequencies without modeling error, grey circles: model-predicted natural frequencies considering modeling errors.

the building with negligible bias at its reference state, the same is not true at the severe damage state, i.e., data fit should not
be mistaken with model correctness, especially when the model is used outside its calibration range. This indicates that the
updated error function parameters at the reference state do not accurately represent those at the severe damage state.
Therefore, the performance of the presented uncertainty quantification/propagation framework for model extrapolation
depends on the model prediction bias at the future state which can be different from the bias estimated at the reference/
calibration state. This different bias at the two states can cause inaccurate prediction at severe damage cases [61].

6. Summary and conclusion

In this paper, a Hierarchical Bayesian framework is proposed for model calibration where the statistical properties (mean
and covariance) of error functions are also considered as updating parameters. The updated mean and covariance of error
functions are used for probabilistic response prediction from the calibrated model. The framework is suited for probabilistic
response prediction of complex structural systems where such predictions often include bias and uncertainty/variability. The
framework is applied for model calibration of a ten-story concrete building using its identified modal parameters from ambi-
ent vibration measurements. The modeling errors are represented by normally (Gaussian) distributed error functions, and
defined as the difference between the model-calculated and identified natural frequencies and mode shapes. Two cases of
model calibrations are performed: (a) assuming zero-mean error functions (i.e., not updating bias), and (b) updating a
non-zero mean (bias) for the error function of the fifth modal frequency where a clear bias was observed in case (a). For both
cases, the calibrated models are used to predict the building response once with and once without including the modeling
error parameters. The predictions without accounting the modeling errors are not in good agreement with identified values
(validation data sets). The identified data, however, fall into the high probability region of model predictions when modeling
errors are considered in the predictions. The first calibrated model (with zero mean error functions) provides very large
 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514 513

estimation uncertainties for the fifth modal frequency. The large estimation uncertainty in this case compensates for the
existing bias for the fifth mode that is not considered in the updating process. The second calibrated model (with non-
zero mean error function for the fifth mode) provides a much tighter prediction for the fifth mode and thus more accurate
predictions.
Finally, the performance of the second calibrated model (i.e., non-zero bias) is studied for prediction of building natural
frequencies for two damaged states of the building for which the model is not calibrated. The considered damaged states
include the moderate damage state, i.e., after removing four exterior walls of the building at the second story, and the severe
damage state, i.e., after removing two additional exterior walls at the first story. The model-predicted results are compared
to the natural frequencies identified from the ambient vibration measurements of the building after the wall removals. The
model predictions in the moderate damage state provide close estimations for the observed natural frequencies, however,
higher discrepancies are observed for the severe damage case, where the probabilistic predictions for the first two modal
frequencies do not coincide with the measured data. Although the difference is not large, the observations fall outside the
high probability regions of predictions for the first two modes. Therefore, it is important to note that performance of the pre-
sented uncertainty quantification/propagation framework depends on evolution of modeling errors from the calibrated state
to the future state. Large discrepancies between the modeling errors (especially bias) at the two states can cause inaccurate
response prediction and performance assessment. However, the proposed approach provides accurate confidence bounds on
response predictions when the future structural state contains small changes compared to its reference/calibration state
(e.g., moderate damage level).

Acknowledgments

The study presented here has been partially by the National Science Foundation – United States Grants 1430180 and
1254338. The collaboration of NEES@UCLA during the planning and execution stages of the experiments is acknowledged.
The authors would also like to thank the New York State Department of transportation (NYSDOT) personnel and their direc-
tor Andrew Roberts for allowing the execution of these tests and for their remarkable cooperation in every part of the exper-
iment. The support of Tufts Technology Services is highly appreciated. The opinions expressed in this paper are those of the
authors and do not necessarily represent those of the sponsor or the collaborators.

References

[1] Q.S. Li, J.R. Wu, Correlation of dynamic characteristics of a super-tall building from full-scale measurements and numerical analysis with various finite
element models, Earthquake Eng. Struct. Dyn. 33 (2004) 1311–1336.
[2] T. Kijewski-Correa, J. Kilpatrick, A. Kareem, D.-K. Kwon, R. Bashor, M. Kochly, B.S. Young, A. Abdelrazaq, J. Galsworthy, N. Isyumov, Validating wind-
induced response of tall buildings: synopsis of the Chicago full-scale monitoring program, J. Struct. Eng. 132 (2006) 1509–1523.
[3] S. Soyoz, E. Taciroglu, K. Orakcal, R. Nigbor, D. Skolnik, H. Lus, E. Safak, Ambient and forced vibration testing of a reinforced concrete building before and
after its seismic retrofitting, J. Struct. Eng. 139 (2012) 1741–1752.
[4] F.N. Catbas, S.K. Ciloglu, O. Hasancebi, K. Grimmelsman, A.E. Aktan, Limitations in structural identification of large constructed structures, J. Struct. Eng.
133 (2007) 1051–1066.
[5] Y.-L. Xu, S. Zhan, H. Xia, Y. Xia, N. Zhang, Field measuremnets of the new CCTV tower in Beijing, Int. J. High-Rise Build. 2 (2013) 171–178.
[6] D.H. Kim, J.Y. Kim, Assessment on natural frequencies of structures using field measurement and FE analysis, Int. J. High-Rise Build. 3 (2014) 305–310.
[7] T. Haukaas, P. Gardoni, Model uncertainty in finite-element analysis: Bayesian finite elements, J. Eng. Mech. 137 (2011) 519–526.
[8] E.N. Chatzi, C. Papadimitriou, J. Beck, Special issue on uncertainty quantification and propagation in structural systems, Am. Soc. Civ. Eng. (2016).
[9] C. Papadimitriou, L. Katafygiotis, S.K. Au, Effects of structural uncertainties on TMD design: a reliability-based approach, J. Struct. Control 4 (1997) 65–
88.
[10] K. Lievens, G. Lombaert, G. De Roeck, P. Van den Broeck, Robust design of a TMD for the vibration serviceability of a footbridge, Eng. Struct. 123 (2016)
408–418.
[11] T. Kijewski-Correa, J.D. Pirnia, Dynamic behavior of tall buildings under wind: insights from full-scale monitoring, Struct. Des. Tall Spec. Build. 16
(2007) 471–486.
[12] M.W. Vanik, J.L. Beck, S.K. Au, Bayesian probabilistic approach to structural health monitoring, J. Eng. Mech.-ASCE 126 (2000) 738–745.
[13] B. Moaveni, A. Stavridis, G. Lombaert, J.P. Conte, P.B. Shing, Finite-element model updating for assessment of progressive damage in a 3-story infilled
RC frame, J. Struct. Eng. 139 (2013) 1665–1674.
[14] H. Sun, A. Mordret, G.A. Prieto, M.N. Toksöz, O. Büyüköztürk, Bayesian characterization of buildings using seismic interferometry on ambient
vibrations, Mech. Syst. Sig. Process. 85 (2017) 468–486.
[15] S. Mukhopadhyay, H. Lusß, R. Betti, Probabilistic structural health assessment with identified physical parameters from incomplete measurements,
ASCE-ASME J. Risk Uncertain. Eng. Syst. Part A: Civ. Eng. (2015) B4015003.
[16] P. Gardoni, A. Der Kiureghian, K.M. Mosalam, Probabilistic capacity models and fragility estimates for reinforced concrete columns based on
experimental observations, J. Eng. Mech. (2002).
[17] B. Moaveni, J.P. Conte, F.M. Hemez, Uncertainty and sensitivity analysis of damage identification results obtained using finite element model updating,
Comput.-Aided Civ. Infrastruct. Eng. 24 (2009) 320–334.
[18] M.I. Friswell, J.E. Mottershead, Finite Element Model Updating in Structural Dynamics, Kluwer Academic Publishers, Boston, Dordrecht, 1995.
[19] J.E. Mottershead, M. Link, M.I. Friswell, The sensitivity method in finite element model updating: a tutorial, Mech. Syst. Sig. Process. 25 (2011) 2275–
2296.
[20] S. Živanović, A. Pavic, P. Reynolds, Finite element modelling and updating of a lively footbridge: the complete process, J. Sound Vib. 301 (2007) 126–
145.
[21] E. Ntotsios, C. Papadimitriou, P. Panetsos, G. Karaiskos, K. Perros, P.C. Perdikaris, Bridge health monitoring system based on vibration measurements,
Bull. Earthq. Eng. 7 (2009) 469–483.
[22] M. Sanayei, G.R. Imbaro, J.A.S. McClain, L.C. Brown, Structural model updating using experimental static measurements, J. Struct. Eng.-ASCE 123 (1997)
792–798.
[23] B. Goller, G. Schueller, Investigation of model uncertainties in Bayesian structural model updating, J. Sound Vib. 330 (2011) 6122–6136.
[24] J. Wu, Q. Li, Finite element model updating for a high-rise structure based on ambient vibration measurements, Eng. Struct. 26 (2004) 979–990.
514 I. Behmanesh et al. / Mechanical Systems and Signal Processing 107 (2018) 502–514

[25] E. Yu, J.W. Wallace, E. Taciroglu, Parameter identification of framed structures using an improved finite element model-updating method-Part II:
application to experimental data, Earthq. Eng. Struct. Dyn. 36 (2007) 641–660.
[26] M. Sanayei, B. Arya, E.M. Santini, S. Wadia-Fascetti, Significance of modeling error in structural parameter estimation, Comput.-Aided Civ. Infrastruct.
Eng. 16 (2001) 12–27.
[27] J.L. Beck, L.S. Katafygiotis, Updating models and their uncertainties. I: Bayesian statistical framework, J. Eng. Mech.-ASCE 124 (1998) 455–461.
[28] H. Sohn, K.H. Law, A Bayesian probabilistic approach for structure damage detection, Earthquake Eng. Struct. Dyn. 26 (1997) 1259–1281.
[29] K. Christodoulou, E. Ntotsios, C. Papadimitriou, P. Panetsos, Structural model updating and prediction variability using Pareto optimal models, Comput.
Meth. Appl. Mech. Eng. 198 (2008) 138–149.
[30] J. Ching, J.L. Beck, Bayesian analysis of the Phase II IASC-ASCE structural health monitoring experimental benchmark data, J. Eng. Mech.-ASCE 130
(2004) 1233–1244.
[31] J.A. Goulet, I.F.C. Smith, Structural identification with systematic errors and unknown uncertainty dependencies, Comput. Struct. 128 (2013) 251–258.
[32] R. Pasquier, I.F. Smith, Robust system identification and model predictions in the presence of systematic uncertainty, Adv. Eng. Inf. (2015).
[33] J.A. Goulet, I.F. Smith, Predicting the usefulness of monitoring for identifying the behavior of structures, J. Struct. Eng. 139 (2012) 1716–1727.
[34] S.G.S. Pai, I.F. Smith, Comparing three methodologies for system identification and prediction, in: 14th International Probabilistic Workshop, Springer,
2017, pp. 81–95.
[35] W.R. Gilks, Encyclopedia of Biostatistics; Markov Chain Monte Carlo, Wiley Online Library, 2005.
[36] W.R. Gilks, S. Richardson, D.J. Spiegelhalter, Markov Chain Monte Carlo in Practice, Chapman & Hall, Boca Raton, FL, 1998.
[37] I. Behmanesh, B. Moaveni, G. Lombaert, C. Papadimitriou, Hierarchical Bayesian model updating for structural identification, Mech. Syst. Sig. Process.
64–65 (2015) 360–376.
[38] I. Behmanesh, S. Yousefianmoghadam, A. Nozari, B. Moaveni, A. Stavridis, Effects of Prediction Error Bias on Model Calibration and Response Prediction
of a 10-Story Building 34th International Modal Analysis Conference, Springer, Orlando, FL, USA, 2016.
[39] J.L. Beck, Bayesian system identification based on probability logic, Struct. Control Health Monit. 17 (2010) 825–847.
[40] A. Gelman, Prior distributions for variance parameters in hierarchical models (comment on article by Browne and Draper), Bayesian Anal. 1 (2006)
515–534.
[41] I. Behmanesh, B. Moaveni, Accounting for environmental variability, modeling errors, and parameter estimation uncertainties in structural
identification, J. Sound Vib. 374 (2016) 92–110.
[42] R.L. Fox, M.P. Kapoor, Rates of change of eigenvalues and eigenvectors, AIAA J. 6 (1968) 2426.
[43] H. Jensen, A. Muñoz, C. Papadimitriou, C. Vergara, An enhanced substructure coupling technique for dynamic re-analyses: application to simulation-
based problems, Comput. Meth. Appl. Mech. Eng. 307 (2016) 215–234.
[44] P.E. Hadjidoukas, P. Angelikopoulos, C. Papadimitriou, P. Koumoutsakos, P4U: a high performance computing framework for Bayesian uncertainty
quantification of complex models, J. Comput. Phys. 284 (2015) 1–21.
[45] S. Yousefianmoghadam, I. Behmanesh, A. Stavridis, B. Moaveni, A. Nozari, A. Sacco, System identification and modeling of a dynamically tested and
gradually damaged 10-story reinforced concrete building, Earthquake Eng. Struct. Dyn. (2018), https://doi.org/10.1002/eqe.2935.
[46] J.-N. Juang, R.S. Pappa, An eigensystem realization algorithm for modal parameter identification and model reduction, J. Guid. Control Dyn. 8 (1985)
620–627.
[47] J. Caicedo, Practical guidelines for the natural excitation technique (NExT) and the eigensystem realization algorithm (ERA) for modal identification
using ambient vibration, Exp. Tech. 35 (2011) 52–58.
[48] J.M. Caicedo, S.J. Dyke, E.A. Johnson, Natural excitation technique and eigensystem realization algorithm for phase I of the IASC-ASCE benchmark
problem: simulated data, J. Eng. Mech. 130 (2004) 49–60.
[49] CSI, Integrated Finite Element Analysis and Design of Structures Basic Analysis Reference Manual, Berkeley, CA, USA, 2015.
[50] F.C. Filippou, M.F. Constantinides, Getting Started Guide and Simulation Examples, Technical Report NEESgrid-2004-22, Civil and Environmental Eng.
Dept. University of California at Berkeley, Berkeley (CA), 2004.
[51] MathWorks, MATLAB User’s Guide, MathWorks Inc., Natick, MA, 2017.
[52] S.G. Shahidi, S.N. Pakzad, Effect of measurement noise and excitation on generalized response surface model updating, Eng. Struct. 75 (2014) 51–62.
[53] R. Madarshahian, J.M. Caicedo, Surrogate-Based Approach to Calculate the Bayes Factor, Model Validation and Uncertainty Quantification, vol. 3,
Springer, 2017, pp. 277–281.
[54] P. Angelikopoulos, C. Papadimitriou, P. Koumoutsakos, X-TMCMC: adaptive kriging for Bayesian inverse modeling, Comput. Meth. Appl. Mech. Eng. 289
(2015) 409–428.
[55] V. Yaghoubi, M.K. Vakilzadeh, T. Abrahamsson, A Parallel Solution Method for Structural Dynamic Response Analysis, Dynamics of Coupled Structures,
vol. 4, Springer International Publishing, 2015, pp. 149–161.
[56] H. Jensen, E. Millas, D. Kusanovic, C. Papadimitriou, Model-reduction techniques for Bayesian finite element model updating using dynamic response
data, Comput. Meth. Appl. Mech. Eng. 279 (2014) 301–324.
[57] ACI-318, Building Code Requirements for Structural Concrete and Commentary, ACI committee 318, 2011.
[58] T.S. Fu, A.J. Garcia-Palencia, E.S. Bell, T. Adams, A. Wells, R. Zhang, Analyzing prerepair and postrepair vibration data from the Sarah mildred long bridge
after ship collision, J. Bridge Eng. 21 (2015) 05015002.
[59] H. Yun, R. Nayeri, F. Tasbihgoo, M. Wahbeh, J. Caffrey, R. Wolfe, R. Nigbor, S. Masri, A. Abdel-Ghaffar, L.H. Sheng, Monitoring the collision of a cargo ship
with the Vincent Thomas Bridge, Struct. Control. Health Monit. 15 (2008) 183–206.
[60] R. Nayeri, S. Masri, A. Chassiakos, Application of structural health monitoring techniques to track structural changes in a retrofitted building based on
ambient vibration, J. Eng. Mech. 133 (2007) 1311–1325.
[61] I. Behmanesh, B. Moaveni, C. Papadimitriou, Probabilistic damage identification of a designed 9-story building using modal data in the presence of
modeling errors, Eng. Struct. 131 (2017) 542–552.