ARTICLE IN PRESS

JOURNAL OF
SOUND AND
VIBRATION
Journal of Sound and Vibration 305 (2007) 211–225
www.elsevier.com/locate/jsvi

Sensitivity-based finite element model updating using

constrained optimization with a trust region algorithm
Pelin Gundes Bakira,b,, Edwin Reyndersb, Guido De Roeckb
a
Department of Civil Engineering, Istanbul Technical University, 34496 Maslak-Istanbul, Turkey
b
Department of Civil Engineering, K.U. Leuven, Kasteelpark Arenberg 40 B-3001 Leuven, Belgium
Received 26 July 2006; received in revised form 28 March 2007; accepted 28 March 2007

Abstract

The use of changes in dynamic system characteristics to detect damage has received considerable attention during the
last years. Within this context, FE model updating technique, which belongs to the class of inverse problems in classical
mechanics, is used to detect, locate and quantify damage. In this study, a sensitivity-based finite element (FE) model
updating scheme using a trust region algorithm is developed and implemented in a complex structure. A damage scenario is
applied on the structure in which the stiffness values of the beam elements close to the beam–column joints are decreased
by stiffness reduction factors. A worst case and complex damage pattern is assumed such that the stiffnesses of adjacent
elements are decreased by substantially different stiffness reduction factors. The objective of the model updating is to
minimize the differences between the eigenfrequency and eigenmodes residuals. The updating parameters of the structure
are the stiffness reduction factors. The changes of these parameters are determined iteratively by solving a nonlinear
constrained optimization problem. The FE model updating algorithm is also tested in the presence of two levels of noise in
simulated measurements. In all three cases, the updated MAC values are above 99% and the relative eigenfrequency
differences improve substantially after model updating. In cases without noise and with moderate levels of noise;
detection, localization and quantification of damage are successfully accomplished. In the case with substantially
noisy measurements, detection and localization of damage are successfully realized. Damage quantification is also
promising in the presence of high noise as the algorithm can still predict 18 out of 24 damage parameters relatively
accurately in that case.
r 2007 Elsevier Ltd. All rights reserved.

1. Introduction

Most of the bridges in USA and Europe are built in the 1960s and the majority of these have reached their
critical age. It has been reported that about 125,000 out of the total 585,000 bridges in USA are structurally
deficient [1,2]. It is expected that the budget demands for maintenance will peak in 2010 [3]. The earthquakes
in Turkey and Greece in 1999 [4,5], the Umbria-Marche earthquake in 1997 in Italy and the recent
earthquakes in Portugal and Iceland have damaged many buildings, in which the damage is not visible by

Corresponding author. Department of Civil Engineering, Istanbul Technical University, 34496 Maslak-Istanbul, Turkey.
Tel.: +90 216 359 74 53; fax: +90 216 373 44 27.
E-mail address: [email protected] (P.G. Bakir).

0022-460X/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jsv.2007.03.044
 ARTICLE IN PRESS
212 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

visual inspections. Vibration monitoring is a useful instrument in evaluating the condition of these bridges and
buildings and in making maintenance schedules. Vibration-based damage detection and finite element (FE)
model updating have emerged in the 1990s as topics of prominent importance to the design, construction and
maintenance of civil engineering structures. An excellent review on all techniques can be found in the book of
Friswell and Mottershead [6] and in two reports of the Los Alamos National Laboratory [7,8]. Other
comprehensive reviews on the topic can be found in Refs. [9–12].
Structural damage in civil engineering structures results in changes in the modal parameters such as natural
frequencies, mode shapes and modal damping values. Modal parameters can be easily obtained from vibration
testing. Four levels of damage identification are possible as proposed by Rytter [13]:

(1) Level 1—detection: Is the structure damaged or not?

(2) Level 2—localization: What is the location of the damage in the structure?
(3) Level 3—quantification: What is the extent of damage?
(4) Level 4—prediction: What is the remaining service life of the structure?

Many model-based damage detection methods attempt to detect changes in the natural frequencies of a
structure. Salawu [14] reviewed different methods of structural damage detection through changes in natural
frequencies. The author stresses that the use of natural frequencies as diagnostic parameters in structural
assessment procedures using vibration monitoring is an inexpensive structural assessment technique.
However, there are limitations that could prevent successful application. The changes in natural frequencies
can not or hardly provide the spatial information about structural damage. Therefore, mode shape
information is also required to uniquely localize damage. Since this study is on FE model updating by ambient
vibration measurements, the identified experimental modes will be unscaled.
This paper describes the sensitivity-analysis-based FE model updating method. Sensitivity-based techniques
have been used for FE model updating in Refs. [15–20] but most of these applications are on beam-like
structures (e.g. bridges) and plates but not on building type structures where damage is typically caused by an
earthquake like action. For building systems designed according to the capacity-based philosophy, the damage
resulting from an earthquake is globally distributed throughout the structure, in regions specifically designed
to dissipate energy (plastic hinges). The original contribution of the paper is the application of the method on
the structure condition assessment with particular reference to buildings. This procedure is applicable to
buildings which have been damaged by earthquakes, but that behave in the linear domain during an ambient
vibration measurement after the earthquake. Therefore, structures that have been heavily damaged beyond
repair are outside the scope of this study. This is an ongoing research and this part of the study is a feasibility
investigation of the sensitivity-based FE model updating method using a trust region algorithm on a complex
structure with a complicated damage pattern.

2. Theoretical background

Sensitivity-based FE model updating method is the most frequently used updating method [21–23]. This
method directly compares eigenfrequencies and mode shapes. The compared quantities can be collected in a
vector such as
vðaÞ ¼ ½l1 ; . . . ; lmf ; fT1 ; . . . ; fTmm T , (1)
where, lj ¼ o2j are the eigenvalues with lj ¼ ð2pnj Þ2 and eigenfrequency nj , fj refers to the eigenvectors of the
model, mf denotes the number of identified eigenfrequencies that are used in the updating process and ms is
equal to the product of the number of identified mode shapes mm and the number of dofs used N for mode fj .
The residual vector r is defined as the difference between the measured quantities and quantities calculated
from the FE model as shown:
" #
rf ðaÞ
rðaÞ ¼ ; rf : R n ! R mf ; rs : R n ! R ms . (2)
rs ðaÞ
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 213

As the quantities are of different magnitudes, a weighted residual has to be taken into consideration for the FE
model updating:
2 3
l~1  l1
6
6 .. 7
7
6
6 . 7
7
6 l~  l 7
6 mf mf 7
rðaÞ ¼ Wð~v  vðaÞÞ ¼ W66 f~  f 7,
7 (3)
6 1 1 7
6
6 .. 7
7
6
4 . 7
5
f~  f mm mm

where l~ are the measured eigenvalues, f~ are the measured eigenmodes, l are the calculated eigenvalues, f are
the calculated eigenmodes, W is a diagonal weighting matrix that scales the eigenvalues and the maximum
value of the mode shapes to the order of unity:
W ¼ diagð1=l~1 ; . . . ; 1=lm
~ ; of ; . . . ; of Þ.
f 1 mm
(4)
Comparison of Eq. (4) with Eq. (3) shows that weighted eigenfrequency residuals can be expressed as
lj ðaÞ  l~j
rf ðaÞ ¼ with lj ¼ ð2pnj Þ2 . (5)
l~j
In FE calculations, the mode shapes are normalized with respect to the mass matrix. When ambient vibrations
are used to extract modal parameters on the other hand, the experimental mode shapes remain unscaled.
Therefore, the weighted mode shape residuals rs ðaÞ are defined as
~
flj ðaÞ flj
~ l Þ,
rs ðaÞ ¼ r  r ¼ olj ðflj  f (6)
j
fj ðaÞ f~j

where l and r denote an arbitrary and a reference dof of mode shape fj (or f~ j ), respectively. If Eq. (6) is
substituted in Eq. (3), the diagonal entries for the weighting matrix corresponding to mode shapes of are
obtained as
~
ðflj ðaÞf~rj  flj frj ðaÞÞ
olj ¼ . (7)
~
frj ðaÞf~rj ðflj ðaÞ  flj Þ
This factor may be changed to give less importance to the mode shapes with respect to the frequencies. As
stated above, the experimental eigenfrequencies are in general the most accurate experimental data that can be
obtained and because of their high sensitivity to the physical properties of the structure, they improve the
condition of the problem. Experimental mode shapes on the other hand, are much more noisy, can reduce the
stability of the minimization problem, but provide more local information. An appropriate weighting is
therefore generally necessary. It should be stated that the weight factors influence the result only in the case of
over-determined systems when the number of residuals is higher than the number of design variables. The
model update is carried out by minimizing the residual:
1
min 12rðaÞT WrðaÞ ¼ 12kW2 rðaÞk2 . (8)
The value of each uncertain variable X e is determined from the multiplication of the initial value of the
variable X eo and the dimensionless updating parameter ae as follows:
X e ¼ X eo ð1  ae Þ. (9)
The above-mentioned uncertain physical properties of the numerical model are the stiffness values in the FE
model updating problem. Damage results in a stiffness reduction in civil engineering structures. The updated
 ARTICLE IN PRESS
214 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

stiffness of an element of the model can be expressed as

Ke ¼ Keo ð1  ae Þ. (10)
The global stiffness matrix will then be:
n
X
K ¼ Ku þ Keo ð1  ae Þ, (11)
e¼1

where Keo and Ke are the initial and updated element stiffness matrix, respectively; K is the global stiffness
matrix and Ku is the stiffness matrix of the elements whose properties remain unchanged, n is the number of
elements that are updated.
The optimization algorithm used to minimize the objective function is a trust region Newton method
which is a sensitivity-based iterative method [24]. The quadratic model mðzÞ is defined by a truncated Taylor
series of f ðaÞ:
min mðzÞ ¼ f s þ ½rf s T z þ 12zT ½r2 f s z, (12)
z

such that kzkpDs where z denotes a step vector from a and where f s ; rf s , and rf 2s are the values of the
function, the gradient and the Hessian of f, respectively. After an iterative process, the minimum a of f ðaÞ is
reached where rf ða Þ  0. The gradient and the Hessian of f ðaÞ are:
k
X
rf ðaÞ ¼ rj ðaÞrrj ðaÞ ¼ Ja ðaÞT rðaÞ, (13)
j¼1

k
X
r2 f ðaÞ ¼ Ja ðaÞT Ja ðaÞ þ rj ðaÞr2 rj ðaÞ  Ja ðaÞT Ja ðaÞ, (14)
j¼1

where Ja is the Jacobian matrix (or sensitivity matrix), which contains the partial derivatives of the residuals rj
with respect to a:
n
X drj
Drj ¼ Dae (15)
e¼1
dae

or in full form:
dr1 dr1 dr1 dr1
2 3
...
6 da1 da2 da3 dap 7
6 7
6 dr dr2 dr2 dr2 7
6 2 ... 7
6 da1 da2 da3 dap 7
6 7
½Jmp ¼ 6 dr3 , (16)
6 7
dr3 dr3 dr3 7
6
6 da ... 7
6 1 da2 da3 dap 7
7
6 ... ... ... ... ... 7
4 drm drm drm drm 5
6 7
...
da1 da2 da3 dap

where m ¼ mf þ ms and p is the number of the updating parameters.

In this paper, Gauss–Newton method is applied with a trust region algorithm to improve the convergence.
In the trust region approach, the algorithm constructs a model function mk , whose behaviour is similar to that
of the actual objective function f near the current point ak , and determines a region surrounding ak where this
model can be trusted. A candidate for the new iterate is then computed by approximately minimizing mk inside
the trust region. In case a sufficient decrease is not obtained in f, the subproblem is again solved with a smaller
trust region. Otherwise, the candidate is accepted as a new iterate from which the process reiterates [25]. The
trust region is a sphere with trust region radius D, which restricts the design variables kakpD. The radius Dk is
adjusted between iterations such that good agreement is obtained between predicted and actual reduction in
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 215

the function as measured by the ratio ak :

f ðak Þ  f ðak þ zk Þ
rk ¼ . (17)
f ðak Þ  mk ðzk Þ
When there is good agreement, rk approaches 1 and Dk is increased. If the agreement is poor, rk takes either
small or negative values. In that case Dk is decreased. Otherwise Dk remains unchanged.
The modal sensitivities with respect to the correction factors ae can be calculated using the finite differences
[18,26]. The first partial derivative of each frequency residual rf and mode shape residual rs with respect
to a are:
drf 1 dlj
¼ , (18)
dae l~ j dae

l
drs 1 dfj flj dfrj
¼  . (19)
dae frj dae ðfrj Þ2 dae

3. FE model updating of a frame type structure

The natural frequencies and mode shapes are directly related to the stiffness of a structure. Therefore,
a drop in natural frequencies or mode shapes will indicate a loss of stiffness which is a consequence
of damage in certain elements of the structure. As the number of elements to be updated increases, the ill-
posedness of the sensitivity matrix increases. In this case, some kind of regularization is necessary. Damage
functions have been recommended for decreasing the number of parameters to be updated in the FE model
updating of a reinforced concrete beam structure by Teughels et al. [27] which can be interpreted as
a kind of regularization method. The principal idea in this methodology is that in order to prevent an
ill-conditioned Jacobian matrix due to a high number of updating variables, the distribution of the
unknown physical property is approximated by combining a limited set of damage functions. The updating
parameters are then selected as the factors by which each of the damage functions has to be multiplied
before combining them. By this methodology, not only the number of the updating parameters are reduced,
but also the ill posedness of the optimization problem is prevented. However, in more complicated structures,
the damage mechanisms depend on engineering judgement and the use of damage functions may not be
appropriate. As mentioned above, in this study, the FE model updating problem is applied on a reinforced
concrete frame structure in which the use of the damage functions is not appropriate. In the following
subsections, the reinforced concrete building type selected is explained first. The damage mechanisms of
the reinforced concrete frames are then discussed followed by the description of the damage scenario applied
to the building.

3.1. Criteria for the choice of the building type considered and the description of the building

North-west Turkey has been hit by two destructive earthquakes in 1999. A general review of these
earthquakes can be found in other papers of the first author [4,5]. In a recent study [28], just over 750,000
buildings in the region to the north of the Sea of Marmara in Turkey are investigated for deriving an
earthquake loss model for the particular area due to the high levels of hazard and exposure in this region. The
buildings have been separated into ‘good’ and ‘poor’ classes. This classification was based on the
considerations of the construction year, evidence of poor construction quality derived from construction type,
and the presence of a weak ground floor.
The building type that is going to be studied in the present study is a ‘good’ quality typical existing building
in the region to the north of the Marmara Sea with a beam side-sway failure mechanism. The building is
chosen such that its geometrical, material and limit state properties are within the confidence intervals of the
‘good’ building classification for the above-mentioned study which is based on 750,000 buildings from the
Marmara region.
 ARTICLE IN PRESS
216 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

The building considered is located in Bolu and is a moment resisting reinforced concrete frame system that
represents a typical residential building in the Marmara region isolated from other buildings at both sides. The
drawings of the building are taken from the curated depository of Turkish building data on the Kocaeli
Golcuk and Duzce Bolu earthquakes of 1999 maintained by Purdue University, University of Michigan and
University of Texas at Austin [29]. The building is a four storey structure with three bays. The floor system is
flat slab with beams. The building has a typical storey height of 2.85 m and considered as regular in elevation.
The dimensions of the columns are typically 0:6 m  0:25 m or 0:25 m  0:6 m and the dimensions of the beams
are 0:25 m  0:5 m for beams. The concrete design strength is 20 MPa. The building does not have a basement
and is fixed at the foundations. The building was under construction when the 1999 M w ¼ 7:4 Kocaeli and
M w ¼ 7:1 Duzce earthquakes hit the region. After the 1999 earthquakes, it is reported that the reinforced
concrete frame was moderately damaged.

3.2. FE model of the reinforced concrete frame

The structure is modeled by FEs and is idealized as a two-dimensional structure taking into account the
weights of the slabs in the transverse direction in the FE model. Eight beam elements per beams and 6 beam
elements for columns are generated for the FE model. The measurement locations are shown in Fig. 1.
In the first stage, in order to obtain the natural frequencies and mode shapes of the system, modal analysis is
carried out by using the Block Lanczos extraction method with a sparse matrix solver. The mode shapes are
normalized to unity. This stage of the analysis gives the undamaged structure’s modal properties. The first two
mode shapes of the frame are shown in Figs. 2 and 3.

3.3. Damage scenario for the reinforced concrete frame

Typically, in modeling the response of reinforced concrete structures to earthquake loading, it is assumed
that the joint regions are rigid and the damage is limited to flexural yielding of beams, columns, slabs and
walls. If the beam–column joint regions are adequately designed, and if the reinforced concrete frame is
designed appropriately according to the weak beam–strong column philosophy, then the plastic hinges will
form in beams close to the face of the columns. On the other hand, the joints are anticipated to fail by bond

Fig. 1. The location of the response points of the reinforced concrete frame.
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 217

Fig. 2. The first vibration mode of the reinforced concrete frame.

Fig. 3. The second vibration mode of the reinforced concrete frame.

slip or joint shear failure modes if the beam–column joint region does not have transverse reinforcement, the
development lengths of beam bars or column depth to beam bar diameter ratio are less than the recommended
values in codes and smooth bars are used for the beam longitudinal reinforcement. In this study, a damage
scenario is considered such that the joint region is assumed to be adequately detailed with closely spaced
stirrups and the frame is assumed to be designed according to the weak beam strong column philosophy.
Consequently, the joints are assumed to remain rigid during seismic excitations and the plastic hinges are
expected to occur in the beams close to the beam column joint regions spreading towards the point of contra
flexure in the beams. Research [30] has shown that the damaged region in this case can not be represented by
two springs because cracking spreads over a finite region at the ends of the reinforced concrete girders. For
this type of scenario, the stiffnesses of the beam elements in the FE model close to the beam–column joint
regions are decreased. It is reasonable to assume that the stiffness at this cracked region is constant. This is
primarily due to the fact that the reinforcement layout will not change along the cracked zone length provided
this zone does not extend beyond the quarter span point. Fig. 4 shows the elements of the reinforced concrete
frame for which damage is simulated by reducing their stiffnesses. The stiffness reduction factors of these
elements are given in Table 1.
 ARTICLE IN PRESS
218 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

5 9 13 17 21 25

4 8 12 16 20 24

3 7 11 15 19 23

2 6 10 14 18 22

Fig. 4. The material numbers for the damaged elements in the FE model of the reinforced concrete frame. Materials numbers 2 to 25
represent the damaged elements whose stiffnesses will be updated. Material number 1 corresponds to the undamaged elements.

Table 1
Stiffness reduction factors for the 24 damaged elements

Material no. Damage scenario a FEM update a Detected damage (yes/no) Actual damage (yes/no)

1 0 NA NA NA
2 0.65 0.65 Yes Yes
3 0.50 0.50 Yes Yes
4 0.40 0.40 Yes Yes
5 0.30 0.30 Yes Yes
6 0.80 0.80 Yes Yes
7 0.90 0.90 Yes Yes
8 0.20 0.20 Yes Yes
9 0 0 No No
10 0.50 0.50 Yes Yes
11 0.55 0.55 Yes Yes
12 0 0 No No
13 0.10 0.10 Yes Yes
14 0.85 0.85 Yes Yes
15 0.70 0.70 Yes Yes
16 0 0 No No
17 0 0 No No
18 0.60 0.60 Yes Yes
19 0.40 0.40 Yes Yes
20 0.10 0.10 Yes Yes
21 0.15 0.15 Yes Yes
22 0.90 0.90 Yes Yes
23 0.60 0.60 Yes Yes
24 0.40 0.40 Yes Yes
25 0.30 0.30 Yes Yes
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 219

In the rest of the paper, the model of the damaged frame will be referred to as the model of the simulated
experiment.

3.4. Justification for the selection of the number of modes and the trust region algorithm

In updating the FE model of the reinforced concrete frame, several parametric studies are carried out for
selecting the optimum number of modes to include in the model updating. Within this context, model quality
indices are introduced in which percentage average error in natural frequencies (IAENF) and percentage
average error in mode shapes (IAEMS) are calculated using the following expressions:
mf  
100 X vu  v~
IAENF ¼ abs , (20)
mf i¼1 v~

j
!
mm X N
100 X fi ju  f~ i
IAEMS ¼ abs j
, (21)
mm N i¼1 j¼1 f~ i
where nu and fu are the updated natural frequencies and mode shapes, respectively. The results of the
parametric studies are summarized in Table 2 in which the error indices are given and in Table 3 in which the
condition numbers, number of iterations, CPU time for a computer with an Intel-Core 2 Duo processor
are compared.
It is apparent from the tables that the best condition number for the Jacobian is obtained for the case of
four modes. Four modes also give the lowest IAEMS error index after updating. For this reason, four modes
are selected for inclusion in to the updating process.
For the case when a single mode is included in the updating process, no convergence was achieved. In this
case, the trust region method could not be implemented, as it requires at least as many equations as design

Table 2
Error indices before and after model updating for different number of modes

Mode no. IAENF (BU) (%) IAENF (AU) (%) IAEMS (BU) (%) IAEMS (AU) (%)
a
1 16.57 6.28 77.16 39.87
2 11.37 0.02 139.07 0.34
3 8.96 0.0015 109.52 0.23
4 7.25 0.0034 102.86 0.16
5 7.08 0.013 120.28 0.77
6 7.08 3.51 154.23 82.98
7 6.77 3.48 171.24 186.56
a
The case of no convergence.

Table 3
Number of iterations, CPU time and the condition number of the Jacobian for the model updating algorithm for different number of
modes

Mode no. No. of iterations Condition number of the Jacobian CPU time (s)

1a 4 2.43E7 350.594
2 9 1986 172.63
3 9 339.65 166.59
4 22 191.17 397.81
5 20 162.30 356.61
6 32 1029.8 605.39
7 23 6446 405.84
a
The case of no convergence.
 ARTICLE IN PRESS
220 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

variables. Consequently, a line-search algorithm using Levenberg–Marquardt method is implemented for

which bound-constraints are not permitted. In model updating problems, due to the ill-conditioned nature of
the Jacobian matrix, the calculated search directions are very sensitive to approximations and errors.
Consequently, if the design variables are not bounded in the line searches along the incorrect directions,
extreme step lengths are obtained that go beyond unrealistic limits. As a result, the optimization process did
not converge for the case of one mode as shown in Tables 2 and 3.
In this study, a large-scale trust region method is selected for solving the nonlinear optimization problem in
FE model updating. There are two main reasons behind this selection. First, other eigensensitivity techniques
based on first-order approximations of FE model properties that use line search methods such as the
Gauss–Newton and the Levenberg–Marquardt do not allow bound-constrained problems. As stated above,
even very small errors in the line searches result in extreme step lengths, hence divergence of the algorithm.
Secondly, the trust region strategy improves the condition number of the Jacobian and acts as a regularization
technique. There are also two reasons for this. First, the radius of the trust region is updated in each iteration
according to the accuracy of the approximating model function. This strategy, reduces oscillations in the
design variables and results in a more robust optimization method. Second, in addition to the application of a
trust region, explicit bound constraints are introduced into the optimization procedure which prevents
overshooting and improves stability. It is true that the convergence speed of the large-scale trust method,
where the subproblems are solved with an iterative conjugate gradient algorithm, will be lower than the other
algorithms. However, since robustness is the major issue to be considered in ill-conditioned FE model
updating problems, the convergence speed may be considered as rather secondary.
A last remark to be noted is that, in this study, the modes are mainly presented in increasing number.
However, throughout the updating algorithm, possible eigenvalue crossing troubles, which may occasionally
(though not necessarily) occur are also taken into account. This phenomenon occurs when the order of modes
change in the damaged structure. In order to solve it, it is important to look to the modeshapes between two
successive damage steps (or iterations in the updating process) to be sure that the same vibration modes are
compared. This is achieved via the MAC matrix as described in the following subsection.

3.5. The MAC matrix and the analysis of results

The index MAC (modal assurance criterion) indicates the correlation between two sets of mode shapes [31].
MAC produces a matrix of inner products between the mode shape vectors as
jfi f~ j j2
MACðfi ; f~ j Þ ¼ T
. (22)
ðfTi fi Þðf~ j f~ j Þ
MAC matrix values change between 0 and 1. A MAC value close to 1 indicates a good correlation, and a
MAC value close to 0 indicates a poor correlation. All the analytical modes are correlated with all the
measured modes and the results are placed in a matrix. After the updating algorithm is completed, the MAC
matrix is calculated in order to see the improvements in the eigenfrequencies and the eigenmodes of the
building. In the first step, the correlation between the initial FE model and the damaged model is investigated.
In the second step, the reference FE model is updated and the correlation between the updated FE model and
the damaged model is again calculated. As a total four eigenfrequencies and four mode shapes are used in the
FE model updating problem. Sixteen measurement points are used in the numerical simulation. Table 1 shows
the actual and detected damage states by the model as well as the actual and predicted stiffness reduction
factors. The table shows that all the damage states and the stiffness reduction factors are predicted accurately
by the model. The FE model updating scheme used can accomplish the first three levels of damage
identification, namely; detection, localization and quantification, successfully. Tables 4 and 5 also show that
the relative eigenfrequency differences and the MAC values are considerably improved after FE model
updating and the updated model gives accurate prediction of these modal parameters.
Next, the FE model updating scheme is tested in the presence of noise. Two noise levels are considered to
simulate measurement errors. The first is moderate noise level in which random noise with normal distribution
and 0:5% standard deviation is applied to the eigenvalues. For the mode shapes, noise with 1% standard
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 221

Table 4
Relative eigenfrequency differences and MAC values between modes obtained from the damaged and the initial reference undamaged
FEM

Mode no. Damaged v~ (Hz) Undamaged v (Hz) ðv  v~Þ MAC (%)

(%)
v~

1 1.45 1.70 16.57 99.82

2 5.18 5.50 6.17 99.28
3 9.82 10.23 4.13 98.18
4 13.11 13.39 2.13 87.50

Table 5
Relative eigenfrequency differences and MAC values between modes obtained from the damaged and the updated FEM

Mode no. Damaged v~ (Hz) Updated v (Hz) ðv  v~Þ MAC (%)

(%)
v~

1 1.45 1.45 1e3 100

2 5.18 5.18 1.88e4 100
3 9.82 9.82 3.76e5 100
4 13.11 13.11 3.95e5 100

Table 6
Stiffness reduction factors predicted for the 24 damaged elements in the presence of random noise with normal distribution and 0.5%
standard deviation applied to the eigenvalues; 1% standard deviation relative to the maximum amplitude applied to the mode shapes

Material no. Damage þ noise a FEM update a Detected damage (yes/no) Actual damage (yes/no)

1 0 NA NA NA
2 0.65 0.65 Yes Yes
3 0.50 0.51 Yes Yes
4 0.40 0.42 Yes Yes
5 0.30 0.29 Yes Yes
6 0.80 0.76 Yes Yes
7 0.90 0.91 Yes Yes
8 0.20 0.23 Yes Yes
9 0 0 No No
10 0.50 0.63 Yes Yes
11 0.55 0.55 Yes Yes
12 0 0 No No
13 0.10 0.14 Yes Yes
14 0.85 0.83 Yes Yes
15 0.70 0.70 Yes Yes
16 0 0 No No
17 0 0.06 Yes No
18 0.60 0.13 Yes Yes
19 0.40 0.44 Yes Yes
20 0.10 0.13 Yes Yes
21 0.15 0.098 Yes Yes
22 0.90 0.92 Yes Yes
23 0.60 0.60 Yes Yes
24 0.40 0.42 Yes Yes
25 0.30 0.25 Yes Yes

deviation relative to the maximum amplitude is applied. The second noise level simulates substantially noisy
measurements. Random noise with normal distribution and 1:5% standard deviation is applied to the
eigenvalues in this case. For the mode shapes, noise with 3% standard deviation relative to the maximum
amplitude is applied.
 ARTICLE IN PRESS
222 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

Table 6 shows the detected and actual damage as well as the predicted stiffness reduction factors in the
presence of noise. The stiffness reduction factors are predicted very closely except for the damaged element 18.
This may be most probably due to the fact that a relatively complicated damage pattern is assumed for the

Table 7
Relative eigenfrequency differences and MAC values between modes obtained from the initial FEM and the damaged in the presence of
moderate noise

Mode no. Damaged þ noise v~ (Hz) Undamaged v (Hz) ðv  v~Þ MAC (%)
(%)
v~

1 1.457 1.696 16.42 99.75

2 5.171 5.496 6.28 99.16
3 9.825 10.230 4.13 98.02
4 13.108 13.387 2.13 87.19

Table 8
Relative eigenfrequency differences and MAC values between modes obtained from the updated FEM and the damaged in the presence of
moderate noise

Mode no. Damaged þ noise v~ (Hz) Updated v (Hz) ðv  v~Þ MAC (%)
(%)
v~

1 1.457 1.456 0.05 99.97

2 5.171 5.177 0.10 99.97
3 9.825 9.812 0.13 99.99
4 13.108 13.094 0.10 99.97

Table 9
Stiffness reduction factors predicted for the 24 damaged elements in the presence of substantial noise

Material no. Damage þ noise a FEM update a Detected damage (yes/no) Actual damage (yes/no)

1 0 NA NA NA
2 0.65 0.22 Yes Yes
3 0.50 0.49 Yes Yes
4 0.40 0.47 Yes Yes
5 0.30 0.31 Yes Yes
6 0.80 0.70 Yes Yes
7 0.90 0.93 Yes Yes
8 0.20 0.33 Yes Yes
9 0 0.07 Yes No
10 0.50 0.70 Yes Yes
11 0.55 0.66 Yes Yes
12 0 0 No No
13 0.10 0.21 Yes Yes
14 0.85 0.82 Yes Yes
15 0.70 0.67 Yes Yes
16 0 0 No No
17 0 0.25 Yes No
18 0.60 0.75 Yes Yes
19 0.40 0.49 Yes Yes
20 0.10 0.42 Yes Yes
21 0.15 0.37 Yes Yes
22 0.90 0.94 Yes Yes
23 0.60 0.52 Yes Yes
24 0.40 0.57 Yes Yes
25 0.30 0.064 Yes Yes
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 223

Table 10
Relative eigenfrequency differences and MAC values between modes obtained from the initial FEM and the simulated experiment in the
presence of substantial noise

Mode no. Damaged þ noise v~ (Hz) Undamaged v (Hz) ðv  v~Þ MAC (%)
(%)
v~

1 1.437 1.696 17.99 99.30

2 5.17 5.496 6.31 99.01
3 9.82 10.230 4.17 97.71
4 13.09 13.387 2.27 85.36

Table 11
Relative eigenfrequency differences and MAC values between modes obtained from the updated FEM and the simulated experiment in the
presence of substantial noise

Mode no. Damaged þ noise v~ (Hz) Updated v (Hz) ðv  v~Þ MAC (%)
(%)
v~

1 1.437 1.451 0.94 99.51

2 5.170 5.138 0.62 99.60
3 9.82 9.772 0.50 99.84
4 13.09 12.95 1.07 99.78

damage scenario. Most damaged elements are adjacent to each other and the adjacent elements’ stiffness
reduction factors are assumed to be substantially different. However, overall results are again quite accurate.
Tables 7 and 8 show the relative eigenfrequency differences and the MAC values between the numerical and
the simulated experimental modes in the presence of moderate noise. The results show that the relative
differences in the eigenfrequencies and the MAC values are considerably improved after the FE model
updating. The results show that the first three levels of damaged identification are again possible with the FE
model updating scheme in the presence of moderate noise.
The second case considered is the substantial noise in measurements. Table 9 shows the stiffness reduction
factors predicted. The results show that in the presence of substantial amount of noise, the model can under or
overestimate the stiffness reduction factors, but 18 out of 24 damage parameters are predicted relatively
accurately again even for the complex damage pattern adopted. Tables 10 and 11 also show that the MAC
values as well as the relative eigenfrequency differences are improved substantially after FE model updating.

4. Conclusions

This study investigates the feasibility of the FE model updating method on a multistorey complex structure
with a complex damage pattern in the absence and presence of noise. A sensitivity-based FE model updating
scheme using constrained optimization with a trust region algorithm is applied on a numerical model of an
actual residential building from Turkey that had been subjected to the 1999 Kocaeli and Duzce earthquakes.
A damage scenario is applied on the reinforced concrete frame such that the stiffnesses of the elements close to
the beam–column joints are deliberately decreased to simulate damage. A complex worst case damage scenario
in which adjacent elements have substantially different stiffness levels is considered. The initial FE model is
then updated to tune the initial modal parameters with the modal parameters from the numerical model in
which the damage is simulated. The results of the updating showed that the first three levels of damage
identification as proposed by Rytter; namely, damage detection, localization and quantification are
successfully accomplished by the FE model updating algorithm used. The relative eigenfrequency differences
and the MAC values are substantially improved after updating. The FE model updating algorithm is also
tested in the presence of two noise levels which simulate moderate and substantial noise in measurements,
respectively. In the presence of moderate noise levels, damage is successfully detected and located in all
elements. With the exception of one element, all the 23 stiffness reduction factors are predicted quite
 ARTICLE IN PRESS
224 P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225

accurately. The relative eigenfrequency differences and the MAC values are considerably improved after
model updating. In the presence of high amount of noise, damage is detected and located correctly in all the
damaged elements except for one element. The majority of the stiffness reduction factors (18 out of 24) are
predicted quite accurately. The MAC values and the relative eigenfrequency differences are substantially
improved after model updating. The results demonstrate that a sensitivity-based FE model updating using a
trust region algorithm is promising for the detection of damaged elements in actual multistorey buildings.

Acknowledgements

The present research is carried out in the framework of OMAX (Operational Modal Analysis in presence of
eXogeneous inputs) project, supported by the Flemish National Fund for Scientific Research. Pelin Gundes
Bakir was a postdoctoral fellow of the Flemish National Fund for Scientific Research- Belgium (FWO-
Flanders) at the Katholieke Universiteit Leuven between years 2003 and 2006, during which this research was
conducted.

References

[1] D.L. Brown, A.J. Dalal, A.J. Helmicki, V.J. Hunt, A.E. Aktan, D.F. Farhey, S.J. Shelley, Condition assessment for bridge
assessment, Journal of Infrastructure Systems, ASCE 2 (3) (1996) 108–117.
[2] H. Hao, J.M.W. Brownjohn, P.Q. Xia, Y. Xia, Civil structure condition assessment by FE model updating: methodology and case
studies, Journal of Sound and Vibration 37 (2001) 761–775.
[3] B. Peeters, System Identification and Damage Detection in Civil Engineering, PhD Thesis, K.U. Leuven, Belgium, 2000.
[4] P.G. Bakir, H.M. Boduroglu, Earthquake risk and hazard mitigation in Turkey, Earthquake Spectra 18 (3) (2002) 427–447.
[5] P.G. Bakir, Proposal for a national mitigation strategy against earthquakes in Turkey, Natural Hazards 33 (3) (2004) 405–425.
[6] M.I. Friswell, J.E. Mottershead, Finite Element Model Updating in Structural Dynamics, Kluwer Academic Publishers, Dordrecht,
The Netherlands, 1995.
[7] S.W. Doebling, C.R. Farrar, M.B. Prime, D.W. Shevitz, Damage identification and health monitoring of structural and mechanical
systems from changes in their vibration characteristics, Technical Report LA-13070-MS, Los Alamos National Laboratory,
May 1996.
[8] H. Sohn, C.R. Farrar, F.M. Hemez, D.D. Shunk, D.W. Stinemates, B.R. Nadler, A review of structural health monitoring literature:
1996–2001, Technical Report Annex to SAMCO Summer Academy, Los Alamos National Laboratory, Cambridge, 2003.
[9] A. Alvandi, C. Cremona, Assessment of vibration-based damage identification techniques, Journal of Sound and Vibration 292 (2006)
179–202.
[10] C.R. Farrar, N.A.J. Lieven, Damage prognosis: the future of structural health monitoring, Philosophical Transactions of the Royal
Society A 365 (2007) 623–632.
[11] S.V. Modak, T.K. Kundra, B.C. Nakra, Comparative study of model updating methods using simulated experimental data,
Computers and Structures 80 (2002) 437–447.
[12] S. Kanev, F. Weber, M. Verhaegen, Experimental validation of a finite-element model updating procedure, Journal of Sound and
Vibration 300 (2007) 394–413.
[13] A. Rytter, Vibration Based Inspection of Civil Engineering Structures, PhD Thesis, Department of Building Technology and
Structural Engineering, Aalborg University, Denmark, 1993.
[14] O.S. Salawu, Detection of structural damage through changes in frequency: a review, Engineering Structures 19 (9) (1997) 718–723.
[15] J.M.W. Brownjohn, P.Q. Xia, H. Hao, Y. Xia, Civil structure condition assessment by FE model updating: methodology and case
studies, Finite Elements in Analysis and Design 37 (2001) 761–775.
[16] J.R. Wu, Q.S. Li, Finite element model updating for a high-rise structure based on ambient vibration measurements, Engineering
Structures 26 (2004) 979–990.
[17] D.A. Rade, G. Lallement, A strategy for the enrichment of the experimental data as applied to an inverse eigensensitivity-based fe
model updating method, Mechanical Systems and Signal Processing 12 (2) (1998) 293–307.
[18] B. Jaishi, W.X. Ren, Structural finite element model updating using ambient vibration test results, Journal of Structural Engineering,
ASCE 1231 (4) (2005) 617–628.
[19] B. Jaishi, W.X. Ren, Damage detection by finite element model updating using modal flexibility residual, Journal of Sound and
Vibration 290 (2006) 369–387.
[20] J.F. Unger, A. Teughels, G. De Roeck, System identification and damage detection of a prestressed concrete beam, Journal of
Structural Engineering, ASCE 132 (11) (2006) 1691–1698.
[21] A. Teughels, G. De Roeck, Damage detection and parameter identification by FE model updating, Archives of Computational
Methods in Engineering 12 (2) (2005) 123–164.
[22] A. Teughels, G. De Roeck, Damage assessment of the Z24 bridge by FE model updating, Key Engineering Materials 245 (2)
(2003) 19–26.
 ARTICLE IN PRESS
P.G. Bakir et al. / Journal of Sound and Vibration 305 (2007) 211–225 225

[23] A. Teughels, G. De Roeck, Structural damage identification of the highway bridge Z24 by FE model updating, Journal of Sound and
Vibration 278 (3) (2004) 589–610.
[24] J. Nocedal, J. Wright, Numerical Optimization, Springer, New York, USA, 1999.
[25] A. Teughels, Inverse Modeling of Civil Engineering Structures Based on Operational Modal Data, PhD Thesis, K.U. Leuven,
Belgium, 2003.
[26] E. Reynders, G. De Roeck, P. Gundes Bakir, C. Sauvage, Damage identification on the Tilff Bridge by vibration monitoring using
optical fiber strain sensors, Journal of Engineering Mechanics, ASCE 133 (2) (2007) 185–193.
[27] J. Maeck, A. Teughels, G. De Roeck, Damage assessment by FE model updating using damage functions, Computers and Structures
80 (2002) 1869–1879.
[28] H. Crowley, J.J. Bommer, R. Pinho, J. Bird, The impact of epistemic uncertainty on an earthquake loss model, Earthquake
Engineering and Structural Dynamics 14 (2005) 1653–1685.
[29] AnatolianQuake.Org. http://bridge.ecn.purdue.edu/ anatolia/.
[30] A. D’Ambrisi, F.C. Filippou, A. Issa, Nonlinear static and dynamic analysis of reinforced concrete subassemblages, Report No.
UCB/EERC-92/08, University of California, Berkeley, 1992.
[31] R.J. Allemang, D.L. Brown, A correlation coefficient for modal vector analysis, Proceedings of IMAC I: 1st International Modal
Analysis Conference, Orlando, Florida, 1982, pp. 110–116.