Proceedings of the IMAC-XXVII

February 9-12, 2009 Orlando, Florida USA

©2009 Society for Experimental Mechanics Inc.

Damage Identification of a Seven-Story Reinforced Concrete

Shear Wall Building Using Bayesian Model Updating
G. Lombaert1 , B. Moaveni2 , X. He3 and J.P. Conte4

1 Structural Mechanics Division, K.U.Leuven, Kasteelpark Arenberg 40, B-3001 Leuven, Belgium
2 Dept. of Civil & Environmental Engineering, Tufts University, 200 College Avenue, Medford, MA, USA
3 DMJM Harris, AECOM, 999 Town & Country Road, Orange, CA, USA
4 Dept. of Structural Engineering, University of California at San Diego, 9500 Gilman Drive, La Jolla, CA, USA

Abstract
Localizing and quantifying potential damage in large and complex structures is one of the most challenging
problems in developing a robust and reliable structural health monitoring system. Vibration-based finite element (FE)
model updating has proven to be a powerful methodology to identify (i.e., detect, localize and quantify) structural
damage. However, due to the presence of uncertainties on the identified modal parameters used in model updating,
probabilistic damage identification methods are preferable. Such methods must allow accounting for all pertinent
sources of uncertainty and must express the damage identification results in probabilistic terms. In this paper,
Bayesian FE model updating is applied for the damage identification of a full-scale seven-story reinforced concrete
building slice tested on the UCSD-NEES shake table. The shake table tests were designed so as to damage the
building progressively through a sequence of historical earthquake records reproduced on the shake table. In this
study, modal parameters identified based on ambient vibration data acquired at various levels of damage are used
to identify the existing damage by means of FE model updating. For this purpose, the FE model of the structure is
divided into a number of substructures. The updating parameters are the stiffness of these substructures. Damage
is therefore identified and localised as a decrease of the stiffness in these substructures. The Bayesian inference
scheme is used to compute the posterior joint probability density of the stiffness in the substructures. The maximum
a posteriori (MAP) estimate of the stiffness has a low value for the substructures where the actual damage was
observed in the test structure. The variance of the identified values tends to be lower for those substructures where
the most severe decrease in stiffness is identified.

1 Introduction

In recent years, structural health monitoring has received increasing attention in the civil engineering research
community with the objective to develop methods through which structural damage can be identified at the ear-
liest possible stage and the remaining useful life of structures evaluated (damage prognosis). Vibration-based,
non-destructive damage identification makes use of changes in dynamic characteristics (e.g., modal parameters)
to identify structural damage. Experimental modal analysis (EMA) has been used as a technology for identifying
modal parameters of a structure based on low amplitude vibration data. It should be emphasized that the success
of damage identification based on EMA depends strongly on the accuracy and completeness of the identified struc-
tural dynamic properties. Extensive literature reviews on vibration-based damage identification were provided by
Doebling et al. [1] and Sohn et al. [2]. Damage identification consists of: (1) detecting the occurrence of damage,
(2) localizing the damage zones, and (3) estimating the extent of damage in the various damage zones. Numerous
vibration-based methods have been proposed to achieve these goals.
Within the frame of the present paper, vibration-based damage assessment by means of finite element model
updating is considered. The model updating problem is generally formulated as a constrained optimization problem
where the model parameters are determined by minimizing the cost or misfit function that measures the discrep-
ancy between the simulated and experimental data. One of the main challenges in model updating is that the
inverse problem is often ill-posed. Therefore, the existence, uniqueness, and stability of the solution with respect
to measurement errors are not guaranteed. This issue is dealt with by means of regularization techniques or an
appropriate parametrization [3]. Alternatively, a Bayesian inference scheme can be used to update our knowledge
about the model parameters based on the experimental observations. In Bayesian inference, probability theory is
used to represent epistemic uncertainty, which is due to the lack of knowledge. Probability density functions repre-
sent the plausibility or the degree of belief that is attributed to uncertain parameters. The prior probability distribution
represents our knowledge prior to the observations and is transformed into the posterior probability distribution that
accounts for both the uncertain prior information and the uncertain experimental data. In this way, all conceivable
solutions are considered, and characterized with probability density functions that reflect how likely the solutions are
in view of the uncertain prior information and the uncertain experimental data.
Beck and Katafygiotis [4] were among the first to propose a Bayesian framework for model updating using
vibration data in the time domain. A similar methodology based on modal data has been proposed by Vanik et
al. [5] and Papadimitriou [6]. Based on the posterior probability density function, the optimal model parameters
can be chosen as those corresponding to the point of maximum a posteriori probability (MAP) or the maximum
likelihood estimate (MLE). These solutions will be similar in the case where the prior probability density function is
relatively flat in the region of interest and if the experimental data are sufficiently informative. A linearization of the
inverse problem is often used to quantify the uncertainty of the identified parameters by the posterior covariance.
For large uncertainties or strongly non-linear inverse problems, however, this approach is no longer sufficient for the
characterization of the posterior probability [7, 8]. In this case, a Monte Carlo simulation of the posterior probabilty
density can be used to generate an ensemble of models that are consistent with the prior information and the
observed data. Mosegaard and Tarantola [9] apply such a strategy to an inverse problem of gravimetry. Beck and
Au [10] use a similar strategy for finite element model updating in structural dynamics. Schevenels et al. [11] have
used this methodology for a probabilistic assessment of the resolution of the Spectral Analysis of Surface Waves
(SASW) test.
Recently, unique tests on a full-scale seven-story reinforced concrete test structure have been performed on
the UCSD-NEES shake table at the Englekirk Structural Engineering Center of the University of California, San
Diego (UCSD) [12]. The objective of this test program was to verify the seismic performance of mid-rise reinforced
concrete shear wall buildings designed for lateral forces obtained from a displacement-based design methodology,
which are significantly smaller than those dictated by current force-based seismic design provisions in the United
States [12]. The shake table tests were designed so as to damage the building progressively through several
historical seismic motions reproduced on the shake table. At various levels of damage, several low amplitude white
noise base excitations were applied through the shake table to the building which responded as a quasi-linear
system with dynamic parameters depending on the level of structural damage. In addition to white noise base
excitation tests, ambient vibration tests were also performed on the building specimen at different damage levels.
Different state-of-the-art system identification methods were applied to acceleration response measurements in
order to estimate modal parameters (natural frequencies, damping ratios and mode shapes) of the building in its
undamaged (baseline) and various damage states [13].
In this study, a Bayesian model updating strategy is applied to identify damage from the modal characteristics
as determined from the ambient vibration tests at various damage states. The outline of the paper is as follows.
First, the test structure is presented, with an overview of the dynamic tests that have been performed. The modal
characteristics of the first three longitudinal modes of the structure are given for four damage states of the structure.
Next, the finite element model of the test structure is presented. The updating parameters in the Bayesian inference
scheme consist of the stiffness in a number of substructures. Damage is therefore identified and localised as a
decrease of the stiffness in these substructures. Finally, a choice is made for the prior probability density distribution
of the updating parameters and the posterior probability density function is sampled using Markov Chain Monte
Carlo techniques.

2 The seven-story test structure

2.1 The test structure and the dynamic tests

The test structure (figure 1a) represents a slice of a full-scale reinforced concrete shear wall building consisting of a
main shear wall (web wall), a back wall perpendicular to the main wall (flange wall) for transversal stability, a concrete
slab at each floor level, an auxiliary post-tensioned column to provide torsional stability, and four gravity columns to
transfer the weight of the slabs to the shake table. Pinned-pinned slotted slab connections capable of transferring
in-plane diaphragm forces are placed between the web and flange walls at floor levels in order to minimize the
moment transfer and coupling between the two walls. Figures 1a and 1b show the test structure mounted on the
shake table and an elevation view with its general dimensions, respectively. Construction drawings and details about
the material test data are available in Panagiotou et al. [12].
A sequence of dynamic tests was applied to the test structure during the period October 2005 - January 2006
including ambient vibration, free vibration, and forced vibration tests (white noise and seismic base excitations) using
the UCSD-NEES shake table. The seven-story test structure was damaged progressively through a sequence of
four historical ground motion records used as table input motions and the modal parameters of the test structure
were identified at different damage states [13]. The four earthquake records applied to the test structure consist
of (http://peer.berkeley.edu/smcat): (1) longitudinal component of the 1971 San Fernando earthquake (M = 6.6)
recorded at the Van Nuys station (EQ1), (2) transversal component of the 1971 San Fernando earthquake recorded
 (a) (b)

Figure 1: The seven-story test structure: (a) view of the structure mounted on the shake table and (b) elevation
view of the test structure (units: m).

at the Van Nuys station (EQ2), (3) longitudinal component of the 1994 Northridge earthquake (M = 6.7) recorded
at the Oxnard Boulevard station in Woodland Hill (EQ3), and (4) 360 degree component of the 1994 Northridge
earthquake recorded at the Sylmar station (EQ4). In the following, the damage state S0 is defined as the baseline
state of the structure before its exposure to the first seismic excitation (EQ1), while damage states S1, S2, S3 and
S4 correspond to the state of the structure after exposure to the first (EQ1), second (EQ2), third (EQ3), and fourth
(EQ4) seismic excitations, respectively. Damage state S0 does not correspond to the uncracked state of the test
structure, since the latter had already been subjected to low amplitude white noise base excitations (0.02-0.03g
RMS) for the purposes of checking the instrumentation and data acquisition system and tuning the shaking table
controller.

2.2 The modal characteristics

Modal parameters of the test structure were identified using state-of-the-art system identification methods based
on measured data from low amplitude dynamic tests (i.e., ambient vibration tests and white noise base excitation
tests) performed at various damage states [13]. In this study, the natural frequencies and mode shapes of the first
three longitudinal vibration modes are used for damage identification of the test structure at damage states S0,
S2, S3, and S4. These data have been determined from the measured ambient vibration data of 14 longitudinal
accelerometers on the web wall (at each floor level and at mid-height of each story) by means of the Data-Driven
Stochastic Subspace Identification (SSI-DATA) algorithm [14]. Figure 2 shows the real part of the complex-valued
mode shapes of the first three longitudinal vibration modes (1st-L, 2nd-L, 3rd-L) of the test structure based on
ambient vibration data from damage state S0. The real part has been determined by projecting all mode shape
components onto their major principal axis in the complex plane.
Table 1 reports the natural frequencies, damping ratios, and modal assurance criterion (MAC) values of these
three longitudinal vibration modes (1st-L, 2nd-L, 3rd- L) identified from ambient vibration excitation by means of
the SSI-DATA algorithm at all damage states considered. The MAC values are calculated between normalized
mode shapes identified at various damage states with their counterparts identified at the undamaged state of the
building. It is observed that: (1) the identified natural frequencies decrease with increasing level of damage, while
the identified damping ratios do not show a clear trend as a function of increasing structural damage. (2) The
calculated MAC values exhibit a general decreasing trend with increasing level of damage.
 Figure 2: Real part of the complex-valued mode shapes of the first three longitudinal vibration modes (1st-L,
2nd-L, 3rd-L) of the test structure at damage state S0.

Damage state Natural Frequency [Hz] Damping Ratio [%] MAC

1st-L 2nd-L 3rd-L 1st-L 2nd-L 3rd-L 1st-L 2nd-L 3rd-L
S0 1.91 10.51 24.51 2.3 2.4 0.5 1 1 1
S1 1.88 10.21 24.31 2.9 2.7 0.6 1.00 1.00 1.00
S2 1.67 10.16 22.60 1.3 1.4 0.9 1.00 0.97 0.96
S3 1.44 9.23 21.82 2.7 1.3 1.4 1.00 0.96 0.92
S4 1.02 5.67 15.09 1.0 1.7 1.0 0.99 0.87 0.81

Table 1: Natural frequencies, damping ratios and MAC values identified based on acceleration data from
ambient vibration tests.

3 The finite element model

In this study, a probabilistic Bayesian updating strategy is used to identify (detect, localize and quantify) the damage
in the test structure based on the identified natural frequencies and mode shapes for the first thee longitudinal
modes of the shear wall building slice. Damage in the structure is identified as a change in material stiffness
(effective modulus of elasticity) of the finite elements in the different substructures of the FE model used for damage
identification. The FE model of the structure is a three-dimensional linear elastic model (figure 3) developed using

Figure 3: The FEDEASLab finite element model of the seven-story test structure.

the general-purpose FE structural analysis program FEDEASLab [15]. An elaborate description of the model can
be found in Moaveni [13].
Initially, the material properties of the structure were based on concrete cylinder tests at various heights along
the structure. These values have been used as the starting values for a deterministic updating of the finite element
model by Moaveni [13]. This updated model is taken as the reference model in the following. In order to identify
damage, the structure is subdivided into a number of substructures. As it can be anticipated that most of the
damage will occur in the web wall, the substructures consist of every half of the part of the web wall for each of
the first three stories, and the entire part of the web wall for each of the four remaining stories. This amounts
to a total of 10 substructures that are considered in the damage identification. The Young’s moduli of the 10
substructures in the reference model are scaled by the parameters θM k that are collected in the vector θ M . This
defines a model class MM of mechanical models that maps the parameter space ΘM ⊆ Rn to a set of models
M∗M = {MM (θ M )|θ M ∈ ΘM }, where MM (θ M ) represents a single model of the class parametrized by θ M . In the
probabilistic Bayesian updating procedure, the parameters θ M are considered to be random. The analytical modal
characteristics of the structure {ω(θ M ), φ(θ M )}T for a given vector θ M are found by the solution of the generalized
eigenvalue problem:
" 10
#
0
X 2
K + Kk θM k φm = ωm Mφm (1)
k=1

0
where K is the stiffness matrix of the structure without the contribution of the 10 substructures and Kk is the stiff-
ness matrix corresponding to substructure k. The eigenvalue problem is reduced by means of a modal superposition
[16] based on a subset φ0 of the eigenvectors for a fixed value of the vector θ M :
" 10 
#
X 
φT
0 Kφ 0 + φ T
0 K φ
k 0 θ M k α̂m = ω̂m 2
φT
0 Mφ 0 αm (2)
k=1

where ω̂m and α̂m are the natural frequency and eigenvector, respectively, of the reduced eigenvalue problem.
The natural frequency and the mode shape of the original eigenvalue problem in equation (1) are approximated as
ωm ≈ ω̂m and φm ≈ φ0 αm .

4 Bayesian model updating

4.1 General formulation of Bayes theorem

In Bayesian model updating, the model parameters θ are represented by random variables. The corresponding
joint probability distribution represents the degree of belief that we attribute to different sets of model parameters
and, therefore, the model parameter uncertainty. According to Bayes theorem, the posterior probability p(θ|d) that
accounts for the experimental data d is obtained from the prior distribution p(θ) as follows:

p(d|θ)p(θ)
p(θ|d) = (3)
p(d)

where p(d|θ) is the probability of observing the data d given the parameters θ. Bayes theorem is often re-written
as follows:

p(θ|d) = cp(d|θ)p(θ) (4)

where the normalization constant c ensures the posterior PDF integrates to one, i.e.
Z
c−1 = p(d|θ)p(θ) dθ (5)
ΘM

In many cases, however, c does not need to be calculated explicitly, and it is sufficient to know the posterior proba-
bility density function p(θ|d) up to a multiplicative constant.
Once the observations d are made, and the corresponding numerical values d̃ are inserted in the probability
distribution p(d̃|θ), this function no longer explicitly depends on the variable d. The function p(d̃|θ) is defined as the
likelihood function L(θ|d̃):

L(θ|d̃) ≡ p(d̃|θ) (6)

The likelihood function L(θ|d̃) expresses how likely the parameters θ are in view of the experimental data d̃.
4.2 Bayesian updating based on the modal characteristics
In the following, the Bayesian inference scheme is elaborated for the case of vibration-based model updating by
means of the observed modal parameters d̃ = {ω̃, φ̃}T . The calculation of the likelihood function L(θ|d̃) is based
on the following equations for the observed prediction error for the natural frequency and the mode shape of each
mode m [5, 6]:

ω̃m = ωm (θ M ) + eωm (7)

γm φ̃m = Lφm (θ M ) + eφm (8)

where eωm and eφm are the observed prediction errors on the natural frequency and the mode shape, respectively,
of mode m. The matrix L is a No × NDOF matrix that defines a mapping from the NDOF degrees of freedom
(DOF) of the finite element model to the No observed modal displacements. The scalar γm is a scaling factor which
is obtained through a least-squares fit [6] of the experimental mode shapes φ̃m and the computed mode shapes
Lφm (θ M ).
The observed prediction errors eωm and eφm in equation (8) are due to the measurement error and the prediction
error. In the case where the data consists of features that are extracted from the raw measurement data, the mea-
surement error also contains the estimation or identification error. The prediction error can be further decomposed
into the modeling error and the model parameter error. The modeling error is due to the fact that the model class
under consideration represents an idealization of the true system behavior, through linearization of the problem,
assumptions about the boundary conditions, . . . . The model parameter error is due to the fact that within the model
class, the model parameters do not have their optimal values. Due to the inherent uncertainty with respect to the
observed prediction error, a probabilistic decription is used. The errors eωm and eφm for all modes are assumed
to be mutually independent and modelled as Gaussian distributed random variables with the following probability
density functions:

1 e2
 

−1/2
peωm (eωm |σωm ) = 2πσωm2
exp − ωm (9)
2
2 σωm
!
 −1/2 1 keφm k2
peφm (eφm |σφm ) = (2π)No σφm
2No
exp − 2
(10)
2 σφm

where σωm is the standard deviation of the error in natural frequency, σφm is the standard deviation of the error
in the modal displacements and No is the number of observed modal displacements. It is now assumed that the
ratio between the standard deviation σωm of the error in natural frequency and the measured natural frequency ω̃m
has the same value σω for all modes, so that σωm = σω ω̃m . The standard deviation σ√ φm of the error in the modal
displacements is assumed
√ to be proportional to the root mean square value kγ m φ̃ m k/ No for each mode, so that
σφm = σφ kγm φ̃m k/ No . The unknown dimensionless standard deviations σω and σφ are collected in the vector
θ e = {σω , σφ }T , where the subscript e refers to the probabilistic error model.
The choice of the probability distributions peωm (eωm |σωm ) and peφm (eφm |σφm ) corresponds to a choice of the
model class Me for the observed prediction errors e = {. . . , eωm , eφm , . . .}. In some cases, the experimental data d̃
can be used to estimate the parameters θ e of the error model, so that the vector of model parameters θ considered
in the inference scheme is equal to {θ M , θ e }T . The corresponding model class M consists of the joint model class
MM × Me for the physical model and the observed prediction error. The likelihood function L(θ M , θ e |d̃) is defined
as the probability p(d̃|θ) of observing the experimental data d̃ for a given set of parameters θ = {θ M , θ e }T and is
computed from equations (9) and (10) as follows:

L(θ M , θ e |d̃) ≡ p(d̃|θ) = ΠN

m peωm (eωm |σωm )peφm (eφm |σφm ) (11)

The expressions (7) and (8) for the observed prediction errors are now introduced in equation (11):
 
1
L(θ M , θ e |d̃) ∝ σω−N σφ−NNo exp − J(θ M , θ e , d̃) (12)
2

with J(θ M , θ e , d̃) the following measure of fit:

N N
X (ωm (θ M ) − ω̃m )2 X ||Lφ m (θ M ) − γm φ̃m ||2
J(θ M , θ e , d̃) = 2 2
+ No (13)
m=1
σω ω̃m m=1
2 σ 2 kφ̃ k2
γm φ m

which is very similar to the least-squares cost function in deterministic model updating [17].
5 The seven-story test structure

5.1 The prior probability density function

In the following, the uncertain experimental data d̃ are used to identify the parameters θ M of the mechanical model
according to the Bayesian inference scheme. Since only the modal characteristics of the three longitudinal modes
of the structure are used in the identification, no attempt is made to identify the standard deviation σω of the relative
error on the natural frequency from the observed data d̃, and a fixed value of σω = 0.005 is used instead. The
vector θ in the Bayesian inference scheme therefore contains the parameters θ M of the mechanical model and the
dimensionless standard deviation θe = σφ .
The prior probability density function p(θ M , θe ) is defined, assuming that θ M and θe are statistically independent,
so that the joint prior PDF p(θ M , θe ) = p(θ M )p(θe ). The Beta distribution pBeta (x; α, β), parametrized by α and β,
and defined for x ∈ [0, 1] is used to define the joint prior density p(θ M ) of the scaling factors of the stiffness of the
10 substructures of the web wall:
  
1 θM,m
p(θ M ) = ΠM m=1 p Beta ; αm , βm (14)
ηm ηm

using mutually independent scaled Beta distributions, where the scaling parameter ηm = 2 defines the upper bound
of the considered range [0, ηm ]. The stiffness of the substructures can therefore vary in a range between 0 and 2
times the value of the reference model that has been obtained by the deterministic updating of the FE model in the
damage state S0. The parameters αm and βm of the Beta distribution are chosen as αm = 6 and βm = 6 leading
to a mean value of 1 and a coefficient of variation of 28%. This choice of the prior distribution results in a relatively
smooth prior PDF with a large range, while excluding values of the stiffness that are negative or unrealistically high.
The prior PDF p(θe ) of the parameter σφ is chosen inversely proportional to the variable, according to the
”noninformative” prior distribution proposed by Jeffreys [18] for a continuous positive parameter. The prior joint PDF
p(θ M , θe ) now becomes:
  
1 θM,m
p(θ M , θe ) = p(θ M )p(θe ) ∝ ΠM m=1 p β ; αm , βm σφ−1 (15)
ηm ηm

The same prior joint PDF is used for every damage state of the seven-story test structure. Information available
from earlier states is therefore not accounted for in the identification.

5.2 The posterior probability distribution

A Markov Chain Monte Carlo (MCMC) algorithm [19] is now used to sample the posterior joint probability density
function p(θ M , θe |d̃). MCMC algorithms generate a sequence of states θ = {θ M , θe }T such that the complete
chain converges to a prescribed probability distribution, without requiring a closed-form expression of the latter. The
transition from one state to the next state occurs by a random perturbation of the current state, that is determined by
the so-called proposal distribution. The candidate state is either accepted or rejected, in which case the next state is
the same as the previous one. A good choice of the proposal density is important to ensure an adequate sampling of
the probability distribution. These algorithms are very useful in Bayesian inference when no closed-form expression
of the posterior PDF is available. In the literature, a wide variety of MCMC algorithms exists. In the following, the
Metropolis-Hastings algorithm [19, 20] is used to sample the posterior distribution. The sampling is performed in two
stages. In a first stage, 4 × 105 samples of θ are generated with a proposal distribution that does not give preference
to a particular direction in the parameter space Θ. These samples are used to estimate the covariance matrix Cθ .
In the second stage, the eigenvalues and eigenvectors of the covariance matrix Cθ are used to determine the step
directions and step sizes, respectively, and a total number of between 4 × 105 and 8 × 105 samples are generated.
Based on the samples in the second stage of the MCMC procedure, the maximum a posteriori (MAP) and max-
imum likelihood estimate (MLE) have been determined for the four damages states considered in the identification.
Figure 4 shows both estimates of the scaling factors θ M , multiplied with the Young’s moduli of the substructures
in the reference model. The first six values in the figures therefore correspond to the effective Young’s modulus of
the lower and upper half of the web wall at each of the first three stories, while the four remaining values represent
the effective Young’s moduli of the web wall at the four top stories. When both estimates are compared for the four
damage states, it is observed that the MLE of the stiffness in a substructure is generally situated more towards the
lower or higher limit value of the range considered in the identification.
For the first two substructures, that correspond to the lower and upper half of the web wall at the fist story, a clear
trend can be observed throughout the damage states. The MAP estimates of the effective Young’s moduli gradually
decrease from values around 20000 MPa in state S0 to much lower values below 5000 MPa in state S4. The low
value in the reference state S0 could be due to the damage caused by the white noise base excitation that has been
used to check the equipment prior to the seismic excitation. For the other substructures, no clear trend is observed.
 4 4 4 4
x 10 x 10 x 10 x 10
6 6 6 6

Stiffness [MPa]

Stiffness [MPa]

Stiffness [MPa]

Stiffness [MPa]
4 4 4 4

2 2 2 2

0 0 0 0
0 5 10 0 5 10 0 5 10 0 5 10
Substructure Substructure Substructure Substructure

(a) S0 (b) S2 (c) S3 (d) S4

Figure 4: Maximum a posteriori (circles) and maximum likelihood estimate (crosses) of the effective Young’s
modulus in the damage states (a) S0, (b) S2, (c) S3, and (d) S4.

Within a single damage state, the values first decrease with increasing height and then decrease. In a similar way,
the value for a single substructure can decrease from one damage state to another and increase subsequently. The
average value of the stiffness in the substructures 3 to 10, however, seems to decrease as well from state S0 to S4.

−4 −4 −4 −4
x 10 x 10 x 10 x 10

4 4 4 4
PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

PDF [m2/N]
3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Stiffness1 4 Stiffness1 4 Stiffness1 4 Stiffness1 4
x 10 x 10 x 10 x 10
−4 −4 −4 −4
x 10 x 10 x 10 x 10

4 4 4 4
PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

PDF [m2/N]
3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Stiffness2 4 Stiffness2 4 Stiffness2 4 Stiffness2 4
x 10 x 10 x 10 x 10
−4 −4 −4 −4
x 10 x 10 x 10 x 10

4 4 4 4
PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Stiffness3 4 Stiffness3 4 Stiffness3 4 Stiffness3 4
x 10 x 10 x 10 x 10
−4 −4 −4 −4
x 10 x 10 x 10 x 10

4 4 4 4
PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Stiffness4 4 Stiffness4 4 Stiffness4 4 Stiffness4 4
x 10 x 10 x 10 x 10
−4 −4 −4 −4
x 10 x 10 x 10 x 10

4 4 4 4
PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Stiffness5 4 Stiffness5 4 Stiffness5 4 Stiffness5 4
x 10 x 10 x 10 x 10
−4 −4 −4 −4
x 10 x 10 x 10 x 10

4 4 4 4
PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

PDF [m2/N]

3 3 3 3
2 2 2 2
1 1 1 1
0 0 0 0
0 2 4 6 0 2 4 6 0 2 4 6 0 2 4 6
Stiffness6 4 Stiffness6 4 Stiffness6 4 Stiffness6 4
x 10 x 10 x 10 x 10

(a) S0 (b) S2 (c) S3 (d) S4

Figure 5: The posterior (blue) and prior (red) marginal probability density of the effective Young’s modulus of
the substructures 1 up to 6 in the damage states (a) S0, (b) S2, (c) S3, and (d) S4.
 Based on all samples of the posterior probability density in the second stage of the MCMC solution, the posterior
marginal probability density of the scaling factor is estimated based on a normal kernel function. The estimated
marginal PDF of the scaling factors is transformed into a marginal PDF of the effective Young’s moduli of the sub-
structures, using the values in the reference model. Figure 5 compares the posterior and prior marginal probability
density of the effective Young’s modulus for the substructures 1 up to 6. This is where most of the damage has been
observed during the tests.
For the two first substructures, it is observed that as the damage progresses (from S0 to S4), the posterior
marginal PDF shifts progressively to lower values of the stiffness and becomes more narrow (i.e., smaller coefficient
of variation). Given the fact that the posterior PDF for all damage states is calculated from the same prior PDF,
this means that with increasing level of damage the data become more informative about the stiffness at the lower
stories, where a very large decrease of the stiffness is observed. For the substructures 3 and 4, that correspond
to the lower and upper half of the web wall at the second story, the posterior marginal PDF shifts to lower values
from damage states S0 to S3. This trend is not continued to damage state S4, where the posterior marginal PDF is
very close to the prior marginal PDF, so that it seems that the data in state S4 do not contain sufficient information
to identify the (decreased) stiffness at these locations. For substructures 5 to 6, that correspond to the lower and
upper half of the web wall at the third story, a clear difference between the prior and posterior marginal PDFs is only
observed at damage state S4, where the posterior marginal PDF has shifted to lower values.

(a) (b)

Figure 6: Observed cracks at (a) the first and (b) the second story of the web wall at damage state S4.

The probabilistic damage identification results are now compared to the actual damage that has been observed
at damage state S4. Figures 6a and 6b show the damage at the bottom two stories of the web wall at state
S4. From these figures, it can be observed that horizontal flexural cracks as well as inclined/diagonal cracks have
developed. Furthermore, during the seismic test EQ4, a lap-splice failure (i.e., debonding between longitudinal steel
reinforcement bars and the surrounding concrete) occurred in the web wall at the bottom of the second story on the
west side as shown in Figure 7.
These results confirm the dramatic decrease of the wall stiffness in the lower and upper half of the first story
(substructures 1 and 2), as identified from the Bayesian updating procedure at damage state S4. The cracks
observed in the web wall at the second story in Figure 6 also explain the shift of the marginal PDF towards lower
values of the stiffness in substructures 3 and 4 at damage state S3.

6 Conclusion

In this paper, Bayesian model updating is used to identify damage in a seven-story reinforced concrete test struc-
ture that has been tested on the UCSD-NEES shake table. This test structure has been submitted to a sequence
of historical earthquake records that were designed so as to damage the building progressively. After each seismic
excitation, vibration data from both white noise and ambient excitation were used to determine the modal charac-
teristics of the structure. The modal characteritiscs of the first three longitudinal modes of the building are used
for a probabilistic assessment of the damage by means of Bayesian FE model updating. The updating parameters
are the stiffness of 10 substructures in the FE model. The Bayesian inference scheme is used to compute the
posterior joint probability density of the stiffness in the substructures. The maximum a posteriori (MAP) estimate
of the stiffness has a low value for the substructures where the actual damage was observed in the test structure.
The variance of the identified values tends to be lower for those substructures where the most severe decrease in
stiffness occurs. A good agreement is obtained between the results of the Bayesian model updating and the actual
damage observed in the test structure.
 (a) (b)

Figure 7: Splitting crack due to lap-splice failure at the bottom of the second story of the web wall on the west
side at damage state S4.

7 Acknowledgements

Partial supports of this research by Lawrence Livermore National Laboratory with Dr. David McCallen as Program
Leader and by the Englekirk Center Industry Advisory Board are gratefully acknowledged. Any opinions, findings,
and conclusions or recommendations expressed in this material are those of the authors and do not necessarily
reflect those of the sponsors.
The work on the Bayesian model updating was supported in part by a travel grant awarded to Geert Lombaert
by the Research Foundation-Flanders (FWO-Vlaanderen) for a research stay at the University of California, San
Diego. The financial support of the FWO-Vlaanderen is kindly acknowledged.

References

[1] S.W. Doebling, C.R. Farrar, Prime M.B., and D.W. Shevitz. Damage identification and health monitoring of
structural and mechanical systems from changes in their vibration characteristics: a literature review. Report
LA-13070-MS, Los Alamos National Laboratory, Los Alamos, NM, 1996.
[2] H. Sohn, C.R. Farrar, F.M. Hemez, D.D. Shunk, D.W. Stinemates, and B.R. Nadler. A review of structural health
monitoring literature: 1996-2001. Report LA-13976-MS, Los Alamos National Laboratory, Los Alamos, NM,
2003.
[3] M.I. Friswell, J.E. Mottershead, and H. Ahmadian. Finite-element model updating using experimental test data:
parametrization and regularization. Philosophical Transactions of the Royal Society, 359:169–186, 2001.
[4] J.L. Beck and L.S. Katafygiotis. Updating models and their uncertainties. I: Bayesian statistical framework.
ASCE Journal of Engineering Mechanics, 124(4):455–461, 1998.
[5] M.W. Vanik, J.L. Beck, and S.K. Au. Bayesian probabilistic approach to structural health monitoring. ASCE
Journal of Engineering Mechanics, 126(7):738–745, 2000.
[6] C. Papadimitriou. Bayesian inference applied to structural model updating and damage detection. In 9th ASCE
Specialty Conference on Probabilistic Mechanics and Structural Reliability, Albuquerque, USA, July 2004.
[7] M. Sambridge and K. Mosegaard. Monte Carlo methods in geophysical inverse problems. Reviews of Geo-
physics, 40(3):1–29, 2002.
[8] A. Tarantola. Inverse problem theory and methods for model parameter estimation. SIAM, Philadelphia, USA,
2005.
[9] K. Mosegaard and A. Tarantola. Monte Carlo sampling of solutions to inverse problems. Journal of Geophysical
Research, 100:12431–12447, 1995.
[10] J.L. Beck and S.-K. Au. Bayesian updating of structural models and reliability using Markov Chain Monte Carlo
simulation. ASCE Journal of Engineering Mechanics, 128(4):380–391, 2002.
[11] M. Schevenels, G. Lombaert, G. Degrande, and S. François. A probabilistic assessment of resolution in
the SASW test and its impact on the prediction of ground vibrations. Geophysical Journal International,
172(1):262–275, 2008.
[12] M. Panagiotou, J.I. Restrepo, and J.P. Conte. Shake table test of a 7-story full scale reinforced concrete struc-
tural wall building slice phase I: rectangular wall section. Report draft SSRP-07-07, Department of Structural
Engineering, University of California, San Diego, La Jolla, CA, 2007.
[13] B. Moaveni. System and Damage Identification of Civil Structures. PhD thesis, Department of Structural
Engineering, University of California, San Diego, 2007.
[14] P. Van Overschee and B. De Moor. Subspace identification for linear systems. Kluwer Academic Publishers,
Dordrecht, The Netherlands, 1996.
[15] F.C. Filippou and M. Constantinides. Fedeaslab getting started guide and simulation examples. Technical
report NEESgrid-2004-22, 2004.
[16] R.W. Clough and J. Penzien. Dynamics of structures. McGraw-Hill, New York, third edition, 1995.
[17] A. Teughels, J. Maeck, and G. De Roeck. Damage assessment by FE model updating using damage functions.
Computers and Structures, 80(25):1869–1879, 2002.
[18] E.T. Jaynes. Probability Theory. The Logic of Science. Cambridge University Press, Cambridge, UK, 2003.
[19] C.P. Robert and G. Casella. Monte Carlo statistical methods. Springer, New York, 2nd edition, 2004.
[20] W.K. Hastings. Monte Carlo sampling methods using markov chains and their applications. Biometrika,
57(1):97–109, 1970.
 1 1 1 1
2 2 2 2
Substructure

Substructure

Substructure

Substructure
0.5 0.5 0.5 0.5
4 4 4 4
0 0 0 0
6 6 6 6
8 −0.5 8 −0.5 8 −0.5 8 −0.5

10 −1
10 −1
10 −1
10 −1
2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10
Substructure Substructure Substructure Substructure
(a) S0 (b) S2 (c) S3 (d) S4

Figure 8: Posterior correlation coefficient of the scaling factor of the stiffness in the substructures 1 up to 7 in
the damage states (a) S0, (b) S2, (c) S3, and (d) S4.