Vous êtes sur la page 1sur 7

Engineering Structures 30 (2008) 3724–3730

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Finite element model updating: Multiple alternatives


Boris A. Zárate, Juan M. Caicedo ∗
Department of Civil and Environmental Engineering, University of South Carolina, 300 Main, Columbia, SC 29208, United States

article info a b s t r a c t

Article history: This paper focuses on the model updating of complex structural systems. Traditional model updating
Received 22 January 2008 techniques optimize an objective function to calculate one single optimal model that behaves similarly
Received in revised form to the real structure and represents the physical characteristics of the structure. One can argue that
17 June 2008
due to numerical and identification errors, and the limited number of sensors in structures, a local
Accepted 18 June 2008
Available online 26 July 2008
minimum rather than the global minimum could be a better representation of the physical properties
of the structure. The methodology proposed in this paper identifies physically different local minima,
Keywords:
giving the analyst the power to decide what model would better describe the system base on his/her
Model updating experience and engineering judgment. Two examples are used in the paper to explore the capabilities
Modeling to generate alternatives of the technique. First, a simple numerical example is used to demonstrate the existence and correct
Modal identification identification of local minima in a model updating objective function. The second problem identifies
model updating alternatives for a finite element model of the Bill Emerson Memorial Bridge. Acceleration
records from the bridge’s permanent instrumentation are used to update the model.
© 2008 Elsevier Ltd. All rights reserved.

1. Introduction frequencies and Modal Assurance Criteria (MAC) values of 73% [5].
These differences can be caused by assumptions made in the de-
Accuratefinite element models are needed for a large num- velopment of the finite element model and uncertainties in ge-
ber of applications such as validating innovative structural ometry, material properties or boundary conditions. In order to
designs, evaluating the effects of earthquake or strong winds on reduce these discrepancies, the numerical model should be cali-
in-service structures, and the implementation of structural con- brated based on information from the real structure [6].
trol and structural health monitoring strategies. The accurate rep- Many model updating techniques have been tested and
resentation of the structure depends on the type of finite ele- published. Most techniques are based on the minimization of
ment model used to represent the structural members, and on the structural parameters to minimize an error function between
properties (e.g. elasticity modulus, moment of inertia) assigned to the measured and numerical responses. The success of these
these elements. Significant differences between the dynamic be- methodologies depends on having a meaningful error function,
havior of a finite element model before updating and the corre- an appropriated numerical model, an accurate identification of
sponding real structure are common. Zhang [1] reported differ- the experimental parameters and using an effective optimization
ences of up to 17.4% between the experimental natural frequen- algorithm to find the global minimum of this function. This paper
cies and those of an initial finite element model of the Kap Shui focuses on the optimization algorithm. Given that the number of
Mun cable-stayed bridge. Similarly, Brownjohn et al. [2] found dif- variables used for the updating process is usually larger than the
ferences of up to 23% between the experimental dynamic charac- number of equations available, and that there are several uncertain
teristics and those of a finite element model of the Pioneer Bridge parameters that are not considered in the updating process, more
in Western Singapore. Brownjohn and Xia [3] showed significant than one good solution may exist. The aim of the proposed method
is to create a set of models that hold similar dynamic characteristics
differences between the experimental and numerical modal char-
but are physically different. Depending upon the final use the
acteristics of a curved cable-stayed bridge before updating, while
analyst then makes a decision, based on his/her experience, to
Jaishi and Ren [4] reported large differences on a steel arch bridge.
select one or many models for subsequent analysis.
Modeling of relatively simpler structures, such as the case of a plate
The proposed methodology is based on Modeling to Generate
test structure, also produced differences between the experimen-
Alternatives (MGA). MGA was developed with the goal of provid-
tal values and a raw finite element model of up to 12% in the natural
ing solutions to complex, incomplete problems by coupling the
computational power of computers and human intelligence [7,8].
MGA creates several possible good solutions for a problem by elim-
∗ Corresponding author. Tel.: +1 803 777 1925; fax: +1 803 777 0670. inating bad alternatives using a mathematical model. These so-
E-mail address: caicedo@engr.sc.edu (J.M. Caicedo). lutions are designed to be as different physically as possible but
0141-0296/$ – see front matter © 2008 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2008.06.012
B.A. Zárate, J.M. Caicedo / Engineering Structures 30 (2008) 3724–3730 3725

provide a similar outcome to the problem. MGA has been previ- Bayesian spectral-density approach, and Yang et al. [29] applied a
ously applied to forest level planning [9,10], the seismic design and technique to decompose the signal into the frequency domain, and
evaluation supports for pipes [11], structural optimization [12], air then calculated the natural frequencies and damping ratios.
quality management problems [13], airline route design [14] and Identification algorithms such as the Eigensystem Realization
the design of wastewater-treatment plants [15]. In addition, a sim- Algorithm (ERA) [30,31], and the Prediction Error Method Through
ilar methodology was used in [16], where genetic algorithms were Least Squares [32] are based on common assumptions: (i)
used to create different alternatives for policy design. Here, MGA the structure or system behaves linearly, (ii) the structure is
is used to provide solutions for the model updating of the cable- time invariant, and (iii) the input loads are a realization of a
stayed bridge. In particular a nonlinear variation of the Hop, Skip Gaussian white noise stochastic process uncorrelated with the
and Jump method (HSJ) proposed by Brill et al. [7] and used for a system response [33]. Subspace identification algorithms have
land use planning problem is used in this paper. the same assumption and obtain a state space representation
Both a numerical and an experimental example are included
of the system as the outcome of the identification procedure.
here to explore the capabilities of the methodology. A simple
Natural frequencies, mode shapes and damping rations can
numerical example, where two variables are updated, is used to
then be calculated from the identified system matrices. In this
validate the successful identification of local and global minima in
paper stochastic subspace identification (SSI) is used. A common
the objective function, while acceleration records from the cable-
mathematical background for subspace algorithms for linear
stayed Bill Emerson Memorial Bridge are used to investigate the
systems is presented by Overschee and De Moor [34]. The
capabilities of the proposed methodology in a real structure. A
algorithm is called Stochastic Subspace Identification (SSI) when
numerical model of the Bill Emerson Memorial Bridge previously
used for structural control [17,18] and later used for structural the input of the system is stochastic and cannot be measured.
health monitoring studies [19] is updated based on acceleration This methodology has gained popularity among researchers in
records from the bridge’s permanent instrumentation system [20]. the last few years due to its relatively simple application. Giraldo
et al. [35] compared the SSI with other output only identification
methodologies, finding that the SSI performs as well as or better
2. Modal identification
than other well known algorithms and requires fewer parameters
In order to obtain a dynamically realistic numerical model of a and less experience to set up. Consider a time-invariant system
structure, it is necessary to: (i) identify the dynamic characteristics described by the equations
of the real structure (e.g. natural frequencies, mode shapes and z (k + 1) = Az (k) + Bu (k) (1)
damping ratios) and (ii) develop a numerical model that can
emulate that dynamic behavior [21]. The first issue is part of what y (k) = Cz (k) + Du (k) (2)
is known as modal identification.
where the matrices A, B, C and D are the system matrices, y is the
Modal identification can be performed based on different types
vector of outputs, u is the vector of inputs and z is the state vector
of test, including: (i) forced vibration, (ii) free vibration, and
at the k-th step. The SSI identifies the system matrices A and C
(iii) ambient vibration tests. Typically, forced vibration and free
from which modal parameters can be calculated. This methodology
vibration tests are difficult to perform in full-scale civil structures
requires the creation of the block Hankel matrix
under normal operation because they are large systems where
external loads are difficult to control and measure. Furthermore,  y (0) y (1) ··· y (p − 1) 
the use of known excitation methods requires the temporary
y (1) y (2) ··· y (p)
closing of the structure while tests are performed, significantly .. ..
..
 
.
 
increasing the cost related with each test. The current state of  . . 
Y0:q−1
 
 y (q − 1) y (i) y (q + p − 2) 
 
the art in ambient vibration tests allows the identification of ···
= (3)
modal parameters with good agreement with those calculated Yi:2q−1
 y (q) y (q + 1) y (q + p − 1) 
 
from forced or free vibration tests. Boroschek et al. [22] determined
···
 .. .. .. 
the modal characteristics of a dock at the harbor of Ventanas . . .
 
(Chile) by performing both free and ambient vibration tests. A y (2q − 1) y (2q) y (2q + j − 2)
static force of 50 tons was applied to the structure for the free
vibration test using a specially designed steel fuse. This force where y (k) is the vector of measured responses at the k-th step.
created an initial displacement on the structure, and allowed free Then, the orthogonal projection L of the row space of the upper
vibration after braking. The Stochastic Subspace Identification (SSI) part of the Hankel matrix Yi:2i−1 (future) onto the row space of the
methodology was used for ambient vibration tests in the same lower part of the Hankel matrix Y0:i−1 (past) is calculated using
structure. Ambient vibration was primarily due to the sea weaves
striking the dock. Comparison of the results shows good agreement
−1
L = Yq:2q−1 YT0:q−1 Y0:q−1 YT0:q−1 Y0:q−1 . (4)
between both types of test. Basseville et al. [23] investigated
theoretical and experimental issues in output only methodologies Using singular value decomposition, the matrix L is decom-
such as robustness related to no stationary loads and handling posed as
of measured data from multiple sensor setups. Mevel et al. [24]
found that output only identification has similar results when L = U S VT (5)
compared with input–output identification for in-flight airplane
structures. Chang et al. [25] identified the main natural frequencies where the matrices U and V are square unitary matrices that
and damping ratios of the Kap Shui Mun cable-stayed bridge in represent the output and input basis vector directions for L, and the
Hong Kong, China, using a Peak Peaking method and an ARMA matrix S is diagonal positive. The predicted observability matrix Ô,
model. Additionally the SHM ASCE group has performed system representing a set of estimated free response for each measured
identification in the benchmark problem, using a large variety of output, is calculated as
output only identification methodologies. For instance, Caicedo
et al. [26] used the Eigensystem Realization Algorithm, Yuen Ô = U S1/2 . (6)
et al. [27] utilized the MODE-ID method, Lam et al. [28] used a
3726 B.A. Zárate, J.M. Caicedo / Engineering Structures 30 (2008) 3724–3730

The matrix Ô can be rewritten in terms of the estimated


matrices  and Ĉ as
   
Ô0 Ĉ
 Ô1   Ĉ Â 
Ô =  .  =  .. (7)
   
. .

 .   
q −1
Ôq−1 Ĉ Â

where Ôi corresponds to the i-th block submatrix of Ô. Therefore,


the estimated matrices  and Ĉ are calculated using the equations

T
 T
−1
 = Ô1:q−1 Ô0:q−2 Ô0:q−2 Ô0:q−2 (8)
Fig. 1. Feasible region of a two-variable nonlinear programming problem.
Ĉ = Ô0 . (9)
The natural frequencies and damping ratios of the structure solution and previous solutions are orthogonal, and low if the
can be calculated from the eigenvalues of  [36], and the output two vectors are parallel. Fig. 1 provides a graphical representation
shapes can be calculated as Φid = ĈΦ, where Φ is the matrix of this concept, showing two solutions for a problem with two
of eigenvectors of  and Φid is the matrix of identified mode variables, x1 and x2 . Given that any orthogonal vector does not
shapes. The matrix Φid will have as many rows as sensors are in provide a good fit for the finite element model, a constraint is used
the structure and as many columns as modes are identified. The to assure that the new solution has a similar performance to the
coordinate of the mode shape is identified at the location of the first solution found. This constraint is described by the equation
sensor. f (pl+1 ) + 1 ≤ α (f (p1 ) + 1) (12)

3. Model updating where f (p1 ) is the value of the objective function shown in Eq.
(6) for the first solution and α is a constant assigned by the
The objective of the finite element model updating process is analyst depending on the structure being updated. Large complex
to adjust the parameters of the finite element model such that structures with a limited number of sensors are expected to
the error between the identified dynamic parameters and those have multiple solutions with similar performance of the objective
of the numerical model is minimized. Classical methods focus function and require a lower value of α than problems with a large
on identifying the global minimum of this objective function, number of sensors. The analyst can run the analysis with different
not identifying other minima with similar performance. One values of α and determine the most appropriated value for his/her
can argue that one of these local minima might represent the respective problem based on the performance and number of
optimal solution for the problem, if the number of non-quantifiable alternative solutions. The analyst should also determine an upper
variables that were not enlisted in the original objective function, value of α depending on the final use of the updated model because
or the uncertainties in the model, are now considered. In this paper the value of α limits acceptable values of the objective function
we present a methodology that determines a family of possible for the alternative solutions. Higher values of α will increase the
solutions to the model updated process. A new method, based on feasible region in which the solutions are searched. A value of α =
the Hop, Skip and Jump method (HSJ) [7] which was originally 1.3 was used in this paper. The Sequential Quadratic Programming
used for a land use planning problem is proposed to generate (SQP) algorithm for non-linear programming [38,39] was used to
the multiple alternatives. The first part of the proposed updating minimize the objective functions shown in Eqs. (6) and (7). The
process consists of finding an initial solution by minimizing the result from this second step provides physically different alternate
objective function: solutions but with a similar performance.
n 
ωid,i − ωfe,i (p)
X 
f (p) = 1 − MAC φid,i , φfe,i (p) 4. Numerical verification
 
+ (10)
i=1
ω id,i

Numerical verification of the methodology was performed on
where kk denotes absolute value, n is the number of identified
the rectangular 3D frame structure shown in Fig. 2. Rectangular
modes, MAC(φid,i , φfe,i (p)) is the modal assurance criteria [37]
columns of 0.3 × 0.3 m and 0.4 × 0.4 m, having a lateral
between the i-th identified mode shape (φid,i ) and the i-th mode
stiffness of 1555 kN/m and 49 152 kN/m, respectively, are used
shape of the finite element model φfe,i (p), ωid,i is the i-th identified
in the model. Each floor was modeled rigid with a total lumped
natural frequency, ωfe,i (p) is the i-th natural frequency of the
mass of 15 480 kg located in the middle of each floor. Only two
finite element model and p is the vector of parameters to be
parameters, x1 and x2 , are considered in this example to facilitate
optimized. Constrains are applied to the procedure to assure that
the verification of the methodology by plotting the objective
the variations of the parameters are not higher than reasonable
function and visually identifying local minima. The parameters to
limits.
be updated correspond to factors multiplying the original stiffness
The next step consists in performing a search in a perpendicular
of columns 1 and 7. In equation form this is
direction of the original solution by minimizing a new objective
function k1 = x1 k1 and k7 = x2 k7 (13)
l
p ·p where the values of k1 and k7 are the original values of the stiffness
j l +1
X
g (pl+1 ) = (11)
pj pl+1 and k1 and k7 are the updated stiffness values. The exact solution
j =1
will lead to values of x1 = 1 and x2 = 1. Modeling errors and errors
where pj is the j-th solution, (·) denotes dot product and || denotes in the identification of the modal parameters are not considered
norm. The value of this function is high if the vector of the current in this example. Therefore, the global minimum is expected to
B.A. Zárate, J.M. Caicedo / Engineering Structures 30 (2008) 3724–3730 3727

to permit the longitudinal expansion of the deck caused by changes


in temperature. Moreover, earthquake restrainers are used in the
transverse direction at the connections between the deck and
the towers. The deck is restrained in the vertical direction at the
towers.
The cables are made of high-strength, low-relaxation steel
(ASTM A882 grade 270) and covered with a polyethylene piping to
prevent corrosion. The area of the cables varies between 28.5 cm2
and 76.3 cm2 . The towers are H-shaped with a height of 102.4 m at
pier 2 and 108.5 m at pier 3. The towers are of reinforced concrete
with a resistance fc’ of 37.9 MPa.
Fig. 2. Numerical model for validation.
The finite element model shown in Fig. 5 is composed of 575
nodes, 156 beam elements, 512 cable elements and 418 rigid links.
The movements of the deck in the lateral and vertical directions
as well as the rotation with respect to the X axis at bent 1 and
piers 2, 3 and 4 are restricted by applying constraint equations.
At pier 1 only displacement in the X axis and rotations about the
Y and Z axis are allowed. The model of the structure does not
include the Illinois approach as shown in Fig. 5, because neither the
displacements nor the rotations about the X axis are restrained at
pier 4, disconnecting the dynamics of the two systems. Cables are
modeled using the Elastic Catenary models [40] subdivided into
four elements per cable. This cable model was chosen to better
represent the cable–deck dynamics. The soil structure effects are
considered negligible in this study and the structure is assumed
directly fixed to the foundation. The finite element model of the
bridge was first developed in ABAQUS for the benchmark problem
of control of cable-stayed bridges under earthquake excitation [17,
18] and later converted to Matlab for studies in structural health
Fig. 3. Objective function. monitoring.
The Emerson Bridge is permanently instrumented with ac-
Table 1
celeration sensors distributed along the structure and surround-
Numerical example results
ing soil [20]. A total of 84 accelerations channels of Kinemetrics
Solution x1 x2 f (pi ) Eq. (6) EpiSensor have been installed in the bridge. Additionally the bridge
p1 1.000 1.000 0.000 is equipped with Q330 digitizers and a data concentrator, and mass
p2 0.130 3.308 0.184 storage devices (Balers) with wireless units. The acquisition system
has the ability to either acquire data in case of a special event as
coincide with the exact solution. This is not necessarily the case an earthquake using a trigger algorithm, or to acquire long records
when considering modeling and measurement errors. of low frequency ambient vibrations in a scheduled way. The ana-
The first and sixth modal parameters corresponding to frequen- log signal from the accelerometers is digitalized at the bridge
cies of 7.45 Hz and 52.78 Hz are used to update the model. Fig. 3 and then sent by wireless communication to a Central Recording
shows the objective function for this problem. Darker colors in- System (CRS). Once at the CRS, the data are transmitted to the In-
dicate high values of the objective function while lighter colors corporated Research Institutions for Seismology (IRIS), and finally
correspond to lower values of the objective function. Two local broadcasted through the Internet.
minima can be observed at (1, 1) and (0.13, 3.3). The solutions of
the proposed methodology are shown in Table 1. Here is important 5.1. Modal identification
to mention that the SQP algorithm minimizing the objective func-
tion shown in Eq. (6) obtains either (1, 1) or (0.13, 3.3) depending A modal identification of the Emerson Bridge was performed
on the initial values for x1 and x2 . The methodology is successful using the SSI. A total of 25 acceleration channels were used;
finding the two local minima independently of which minimum is 17 of these acceleration channels are located on the deck in
obtained first. the vertical direction and 8 acceleration channels are located
on the towers oriented in the longitudinal direction. The sensor
5. Experimental implementation on the Bill Emerson Memorial locations were chosen to identify the dynamic characteristics of
Bridge the structure corresponding to the vertical mode shapes. Song
et al. [41] previously performed identification of the bridge using
The Bill Emerson Memorial Bridge opened to traffic in ARMAV models. In his paper the first 5 natural frequencies and
December 2003. The bridge crosses the Mississippi River on Cape vertical mode shapes were identified, using 16 vertical channels
Girardeau, Missouri, USA, and it is composed of two towers, 128 of acceleration corresponding to the sensors on the deck.
stay cables and 4 main piers. The bridge has a main span of 350.6 m Six hours of acceleration data were used for the system
and two side spans of 142.7 m, for a total length of 1205.8 m, as identification. The first three hours of data were obtained on 2
shown in Fig. 4. The deck has a width of 29.3 m and is composed November 2005 and the last three were obtained on 5 December
by girders of ASTM A709 grade 50W steel and prestressed concrete 2005. Data were originally collected at 200 Hz, and resampled to
slabs with an fc’ of 41.4 MPa. Additionally the bridge deck has 16 2 Hz to capture the fundamental modes of the structure which are
shock transmission devices in the connection with the towers and lower than 1 Hz. The data were divided in 36 windows of 10 min
bent. The purpose of those devices is to restrain the longitudinal for a total of 1200 data points per channel per window. SSI was
movement during seismic excitation, becoming extremely stiff, but performed in each window, so that a set of 36 natural frequencies,
3728 B.A. Zárate, J.M. Caicedo / Engineering Structures 30 (2008) 3724–3730

Fig. 4. Bill Emerson Memorial Bridge.

Fig. 5. Finite element model of the Bill Emerson Memorial Bridge.


Fig. 6. Identified natural frequencies.

mode shapes and damping rations are obtained for the complete
record. The Hankel matrices used had a total of 750 rows and 1170
columns.
In cable-stayed bridges it is challenging to differentiate non-
physical poles from the physical poles because the mode shapes are
closely spaced. Stabilization diagrams were used for identification
for this purpose [42]. One set of poles was calculated for each Fig. 7. Master masses.

data window and plotted in Fig. 6. True frequencies are detected


based on the consistency through all the windows. An automated types of parameters were considered for updating: (i) mass of
recognition system was used to detect the true poles from the deck, (ii) the rotational stiffness of the connection between
those created by noise and numerical errors [35,43]. True poles the deck and the towers, and (iii) moment of inertia of the spine
are recognized by identifying parameters within some specific beam. These parameters were chosen because of their influence in
characteristics. Here, acceptable modes of vibration should be the overall dynamic behavior of the structure. Other parameters
identified at least five times in different data windows with a such as rotational springs located at the connections of deck and
maximum variation of the natural frequency of 30%, have damping towers were also considered but they had a very small effect in
value lower than 5%, and MAC values between the identified modes the dynamic characteristics of the structure. All the parameters
higher than 0.98. The circles in Fig. 6 show the raw identified considered in this study were lumped into a total of six variables
modes and the stars indicate the acceptable modal parameters. The divided into three distinctive groups.
standard deviation of each remaining natural frequencies (stars in A total of 66 translational lumped masses are used in the bridge
Fig. 6) and the mean MAC value of the corresponding mode shapes deck’s model. Assuming that the mass distribution is symmetric
are also shown in this table. due to the symmetry of the bridge, this can be reduced to 33
Table 2 summarizes the identified modal parameters of the parameters. Because a sudden change in the mass along the deck
Emerson Bridge. The natural frequencies and damping ratios are is unlikely the number of parameters was reduced to three master
calculated as the average of the identified parameters for each masses at three locations along the deck. The other 30 masses were
mode in all the data sets. The standard deviations for each calculated using a spline function. The masses at the mid span
parameter are also shown. The first four mode shapes targeted between bent 1 and pier 2, and one quarter and one half the main
have a low standard deviation in terms of the natural frequencies, span were selected as the locations for the three master masses
damping ratios and mode shapes, indicating a small variation of the (Fig. 7). The masses were constrained to be in the range of −5% to
parameters for all the data records. This indicates that the system 5% of the mass calculated from the blueprints.
is behaving linearly for ambient vibrations. The degree of connection between the deck and the towers is
highly uncertain because of the shock absorbers and the restrainers
5.2. Model updating installed at these locations. The stiffnesses of the four springs used
to model these connections constitute the second set of parameters
The properties of the numerical model were updated such that to update. One parameter was used for bent I and pier IV and a
its modal characteristics would match as close as possible the second parameter was used for the connection at the towers due to
identified natural frequencies and mode shapes. Three different the symmetry of the bridge. The springs’ stiffnesses are constrained
B.A. Zárate, J.M. Caicedo / Engineering Structures 30 (2008) 3724–3730 3729

Table 2
Dynamic characteristics identified
Mode Hits Frequency (Hz) Std. dev. Frequency (Hz) Damping (%) Std. dev. damping (%) MAC Std. dev. MAC

1 31 0.324 0.0036 1.05 0.37 0.9977 0.0019


2 16 0.413 0.0026 0.64 0.09 0.9967 0.0023
3 25 0.635 0.0041 0.67 0.20 0.9983 0.0009
4 21 0.706 0.0032 0.73 0.51 0.9962 0.0046

Table 3
Model updating alternatives
Solution Mass (%) Inertia (%) Stiffness (%) f (p)
Loc1 Loc2 Mid-span Bent Tower Eq. (6)
Original 0 0 0 0 0 0 0.524

p1 0.00 −5.00 −5.00 5.00 100.00 27.36 0.375


p2 0.00 0.00 0.00 0.00 99.73 0.00 0.451
p3 −4.74 0.00 0.00 0.00 0.00 0.00 0.492
p4 0.00 0.00 0.00 0.00 0.00 100.00 0.495

Table 4
Dynamic characteristics of the model updating alternatives
Solution ω1 (Hz) ω2 (Hz) ω3 (Hz) ω4 (Hz) 100*(1-MAC)
(φid,1 , φfe,1 ) (φid,2 , φfe,2 ) (φid,3 , φfe,3 ) (φid,4 , φfe,4 )
Exp. 0.32 0.41 0.63 0.71 – – – –
Original 0.29 0.39 0.60 0.63 4.9 5.4 3.1 10.4
1 0.31 0.41 0.61 0.63 4.3 5.9 4.1 7.5
2 0.31 0.40 0.61 0.63 4.3 6.1 3.9 10.8
3 0.30 0.40 0.62 0.63 4.4 6.0 5.1 14.6
4 0.30 0.40 0.61 0.63 4.7 5.6 3.5 10.1

to be in the range of −100% and 100% of the highest stiffness found table also shows the residual MAC values between the identified
in the stiffness matrix. mode shapes and the numerical modes in percentage. The
The variation of the moment of inertia is assumed to be constant frequencies improved in all the solutions found, when compared
along the deck, so that only one parameter was used to update the to the original finite element model, while keeping the MAC values
deck’s stiffness. The moment of inertia is constrained to be in the in good agreement with the measured data.
range of −5% to 5% of the moment of inertia calculated from the The selection of the final model using the proposed technique
blueprints. is dependent on its use and the experience of the analyst.
Furthermore, the analyst could choose more than one model for
5.3. Results subsequent analysis. For example, if a seismic reliability analysis
were to be performed with the updated model, the analyst could
The starting point for this first optimization was the original decide to perform the reliability analysis with more than one
numerical model as calculated from design drawings. The value alternative and draw conclusions about the reliability of the
of Eq. (6) for the original model is (f (p) = 0.524). Table 3 structure from this subsequent analysis. For this particular bridge
shows the parameters for all the four solutions found. As described it is unlikely that the mass, moment of inertia and stiffness of the
previously in the paper, the first solution was obtained using the connection between deck and tower reach their maximum limits
SQP algorithm to minimize Eq. (6). The solution found suggests that simultaneously. For this reason the second solution could be a
the mass between bent 1 and location 1 should be unchanged while more physically meaningful solution than the first one, and it is
the mass at the main span should be decreased. In addition, the selected as the updated model.
moment of inertia of the deck and the stiffness of the connection
between the deck and bent 1 are increased, while the stiffness
6. Conclusions
at the connection between the deck and the main towers is
decreased. These changes create an improvement of 28.6% in the
objective function. The alternative models are obtained by running This paper discusses a new methodology for model updating
the second optimization process until no additional solutions are of complex structural systems. Traditional model updating tech-
found. The process was repeated with different starting points niques minimize an objective function to minimize the error be-
of the optimization algorithm to explore the completeness of tween a numerical model and the actual structure. Most model
the solutions found. All analyses resulted in the same alternative updating techniques assume that the global minimum of this ob-
solutions. All solutions found are physically different to this jective function will provide the best physical representation of
first solution. The second solution increases the stiffness at the the structure. In contrast, the proposed methodology recognizes
connection between the bent and the deck, while the other that a local minimum could be a better physical representation
parameters remain constant. The third solution decreases the mass of the structure due to incomplete data and identification and
between bent 1 and the mid-span of the first span, while the modeling errors. The proposed method is based on Modeling to
other parameters remain constant. Finally, solution 4 increases the Generate Alternatives (MGA), and it is used to calculate physically
stiffness of the connection between the towers and the deck. different solutions, while obtaining a relative good performance in
Table 4 shows the first four natural frequencies identified the objective function. Two examples are included to show the ca-
experimentally in addition to those of the updated models. This pabilities of the technique. The first example is a simple numerical
3730 B.A. Zárate, J.M. Caicedo / Engineering Structures 30 (2008) 3724–3730

example used to numerically validate the methodology. The sec- [18] Caicedo JM, Dyke SJ, Moon SJ, Bergman LA, Turan G, Hague S. Phase II
ond example updates a model of the Bill Emerson Memorial Bridge benchmark control problem for seismic response of cable-stayed bridges.
J Struct Control Monit 2003;10:137.
with acceleration data used from its permanent instrumentation. [19] Caicedo JM. Structural health monitoring of flexible civil structures. Civil
Final results show four physically different solutions for the model engineering, vol. D.Sc.. Saint Louis, Missouri: Washington University in Saint
updating problem. Louis; 2003. p. 160.
[20] Çelebi M. Real-time seismic monitoring of the new Cape Girardeau bridge and
preliminary analyses of recorded data: An overview. Earthq Spectra 2006;22:
Acknowledgments 609.
[21] Zhang QW, Chang CC, Chang TYP. Finite element model updating for structures
with parametric constraints. Earthq Eng Struct Dyn 2000;29:927.
The first author would like to thank the University of South [22] Boroschek RL, Baesler H, Valdés PT, Vega C. Determinación Experimental De
Carolina Graduate Research Fellowship for the partial financial Propiedades Modales Del Puerto De Ventanas. Congreso Chileno de Sismología
e Ingeniería Antisísmica IX Jornadas. Concepción, Chile; 2005.
support for this paper. The writers would like to thank Mr. [23] Basseville M, Benveniste A, Goursat M, Hermans L, Mevel L, Van der
Bryan Hartnagel of the Missouri Department of Transportation Auweraer H. Output-only subspace-based structural identification: From
theory to industrial testing practice. J Dyn Syst Meas Control-Trans Asme 2001;
for pertinent information and data records on the Bill Emerson
123:668.
Memorial Bridge and Dr. Atanu Dutta for his help in including [24] Mevel L, Benveniste A, Basseville M, Goursat M, Peeters B, Van der Auweraer H,
linear springs in the Emerson Bridge’s finite element model. Vecchio A. Input/output versus output-only data processing for structural
identification — Application to in-flight data analysis. J Sound Vibration 2006;
295:531.
References [25] Chang CC, Chang TYP, Zhang QW. Ambient vibration of a long-span cable-
stayed bridge. Bridge Eng ASCE 2001;6:46.
[26] Caicedo JM, Dyke SJ, Johnson EA. Natural excitation technique and eigensys-
[1] Zhang QW, Chang TYP, Chang CC. Finite-element model updating for the Kap tem realization algorithm for phase I of the IASC-ASCE benchmark problem:
Shui Mun cable-stayed bridge. J Bridge Eng 2001;6:285. Simulated data. J Eng Mech 2004;130:49.
[2] Brownjohn JMW, Moyo P, Omenzetter P, Lu Y. Assessment of highway bridge [27] Yuen KV, Au SK, Beck JL. Two-stage structural health monitoring approach for
upgrading by dynamic testing and finite-element model updating. J Bridge Eng phase I benchmark studies. J Eng Mech-Asce 2004;130:16.
2003;8:162. [28] Lam HF, Katafygiotis LS, Mickleborough NC. Application of a statistical model
[3] Brownjohn JMW, Xia P-Q. Dynamic assessment of curved cable-stayed bridge updating approach on phase I of the IASC-ASCE structural health monitoring
by model updating. J Struct Eng 2000;126:252. benchmark study. J Eng Mech-Asce 2004;130:34.
[4] Jaishi B, Ren W-X. Structural finite element model updating using ambient [29] Yang JN, Lei Y, Lin S, Huang N. Hilbert-Huang based approach for structural
vibration test results. J Struct Eng 2005;131:617. damage detection. J Eng Mech-Asce 2004;130:85.
[5] Bohle K, Fritzen C-P. Results obtained by minimizing natural frequency and [30] Juang J-N, Pappa RS. An eigensystem realization algorithm for modal
MAC-value errors of a plate model. Mech Syst Signal Process 2003;17:55. parameter identification and model reduction. J Guid 1985;8:620.
[6] Mottershead JE, Friswell MI. Model updating in structural dynamics: A survey. [31] Juang J-N, Pappa RS. Effects of noise on modal parameters identified by the
J Sound Vibration 1993;167:347. eigensystem realization algorithm. Journal of Guidance 1986;9:294.
[7] Brill ED, Chang S-Y, Hopkins LD. Modeling to generate alternatives: The HSJ [32] Andersen P. Identification of civil engineering structures using ARMA models.,
approach and an illustration using a problem in land use planning. Manage Sci vol. Doctoral: Aalborg University; 1997.
1982;28:221. [33] Andersen PBR. Estimation of modal parameters and their uncertainties. In:
[8] Brill ED, Flach JM, Hopkins LD, Ranjithan S. MGA: A decision support system for Proceedings of the 17th international modal analysis conference. 1999. p. 323.
[34] Van Overschee P. Moor BLRd. Subspace identification for linear systems:
complex, incompletely defined problems. IEEE Trans Syst Man Cybern 1990;
Theory, implementation, applications. Boston: Kluwer Academic Publishers;
20:745.
1996.
[9] Campbell GE, Mendoza GA. Adapting modeling to generate alternatives (Mga)
[35] Giraldo D, Caicedo JM, Song W, Mogan B, Dyke SJ. Modal identification through
techniques to forest level planning. J Environ Manag 1988;26:151.
ambient vibration: A comparative study. In: 24th international modal analysis
[10] Sprouse WL, Mendoza GA. Modeling to generate alternatives — a Shawnee-
conference. 2006.
National-Forest example. Comput Environ Urban Syst 1990;14:203.
[36] Arici Y, Mosalam KM. Modal identification of bridge systems using state-space
[11] Gupta A, Kripakaran P, Mahinthakumar Jr GK. JWB. Genetic algorithm-based models. Struct Control Health Monit 2005;12:381.
decision support for optimizing seismic response of piping systems. J Struct [37] Allemang RJ. The modal assurance criterion — Twenty years of use and abuse.
Eng 2005;131:389. J Sound Vibration 2003;37:14.
[12] Baugh JW, Caldwell SC, Brill ED. A mathematical programming approach for [38] Gill PE, Murray W, Wright MH. Practical optimization. London: Academic
generating alternatives in discrete structural optimization. Eng Optim 1997; Press; 1981.
28:1. [39] Coleman TF, Li Y. On the convergence of reflective Newton methods for large-
[13] Loughlin DH, Ranjithan SR, Brill ED, Baugh JW. Genetic algorithm approaches scale nonlinear minimization subject to bounds. Math Programm 1994;67:
for addressing unmodeled objectives in optimization problems. Eng Optim 189.
2001;33:549. [40] Chang SP, Park YY. Dynamic analysis of the cablesuspended roof system by
[14] Zechman EM, Ranjithan SR. An evolutionary algorithm to generate alternatives using the three-dimensional cable element. In: Proceedings of the second
(EAGA) for engineering optimization problems. Engi Optim 2004;36:539. Japan–Korea joint seminar on steel bridges. 1992. p. 203–16.
[15] Uber JG, Brill ED, Pfeffer JT. Use of mathematical-programming methods for [41] Song W, Giraldo D, Clayton EH, Dyke SJ, Caicedo JM. Application of ARMAV for
complex-systems. J Water Resour Planning Manag-Asce 1992;118:281. modal identification of the Emerson Bridge. In: Third international conference
[16] Rubenstein-Montano B, Anandalingam G, Zandi I. A genetic algorithm on bridge maintenance, safety and management. 2006.
approach to policy design for consequence minimization. European J Oper Res [42] Brincker Andersen. Understanding stochastic subspace identification. In:
2000;124:43. IMAC. 2006.
[17] Dyke SJ, Caicedo JM, Turan G, Bergman LA, Hague S. Phase I benchmark control [43] Giraldo D. A structural health monitoring framework for civil structures.
problem for seismic response of cable-stayed bridges. J Struct Eng 2003;129: Department of Civil Engineering, vol. D.Sc. Saint Louis, Missouri: Washington
857. University in St. Louis; 2006. p. 157.

Vous aimerez peut-être aussi