Engineering Structures 120 (2016) 166–182

Contents lists available at ScienceDirect

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

Model updating using genetic algorithms with sequential niche

technique
Faisal Shabbir a, Piotr Omenzetter b,⇑
a
Department of Civil Engineering, The University of Engineering and Technology, Taxila, Pakistan
b
The LRF Centre for Safety and Reliability Engineering, The University of Aberdeen, Aberdeen AB24 3UE, UK

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

Article history: Structural model updating is an optimisation problem where parameters that minimise the errors
Received 13 April 2015 between the model and the actual structure are sought. However, multiple solutions may be present.
Revised 30 January 2016 Global optimisation algorithms are efficient optimisation tools but are not fully immune to missing
Accepted 11 April 2016
the global minimum. To increase the chance of finding the global minimum, a combination of genetic
Available online 3 May 2016
algorithm with sequential niche technique is proposed. The method performs systematic search to find
multiple minima and facilitates detecting the minimum that best describes the system. The technique
Keywords:
is applied to experimental data from a simple laboratory structure and a full-scale pedestrian cable-
Cable-stayed bridge
Deceptive problem
stayed bridge, and also tested on a deceptive problem using the numerical model of a space frame.
Global optimisation algorithms Ó 2016 Elsevier Ltd. All rights reserved.
Inverse problem
Model updating
Multiple minima
Genetic algorithm
Sequential niche technique
Structural optimisation

1. Introduction by refining the mathematical model of an actual structure so that

its predictions agree better with test measurements. Typically,
Methods for assessment of the performance, condition and experimental dynamic characteristics, such as resonant frequen-
health of civil, mechanical and aerospace structural systems cies and mode shapes, are used for benchmarking and updating
combining physics-based computer models and in-situ testing structural models. Model updating can be understood and formu-
and monitoring are increasing in importance. They are especially lated as an optimisation problem in which an optimal solution is
useful for civil engineering structures as incorrect understanding sought by perturbing the uncertain model parameters so that the
of their performance and condition may have significant life safety numerical prediction errors taken into account in the objective
and economic implications should such structures underperform function are minimised.
or fail. Important components of infrastructure should maintain It is usually not possible to measure all experimental dynamic
long service life and be able to perform up to the expected responses required for correction of all uncertain parameters in a
standards under everyday as well as extreme loads, while mainte- structural model. This is because of the difficulties in exciting
nance costs need to be kept low [6,10,11,19,42]. and measuring responses at higher frequencies and coarser higher
However, when trying to combine information from experi- mode mapping due to limited numbers of sensors used. A large
ments and analytical or numerical models of as-built systems, number of uncertain parameters in the structural model compared
significant differences between the experiment and the model to experimentally available information may lead to the existence
have been noted in numerous studies [44,8,16]. These differences of multiple minima for the optimisation problem [25]. Adding to
can be mainly attributed to simplifications inherent in modelling that, the assumptions made in the development of FE model and
of complex structures and uncertainties associated with materials, uncertainties associated with material properties, boundary condi-
geometry, and boundary and connectivity conditions [30]. Model tions and geometry may result in significant differences in the
calibration, or updating, is a process of reconciling such differences natural frequencies and mode shapes predicted by the initial FE
model and their experimental counterparts [32]. The algorithm
searching for the global minimum of the objective function may
⇑ Corresponding author. Tel.: +44 1224 272529.
then be lured into those local minima in problems that Goldberg
E-mail address: piotr.omenzetter@abdn.ac.uk (P. Omenzetter).

http://dx.doi.org/10.1016/j.engstruct.2016.04.028
0141-0296/Ó 2016 Elsevier Ltd. All rights reserved.
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 167

et al. [18] call ‘deceptive’. This unsatisfactory behaviour is often global minimum can be discovered with increased confidence. This
encountered in model updating, for example when the popular systematic search is performed by combining the basic GOA, GA in
sensitivity method, based on the iterative steepest-gradient this case, with sequential niche technique (SNT) [5]. SNT is simple
approach, is adopted [14]. Many previous model updating studies and does not require modifications to the search algorithm itself. It
report a single solution [8,24,50] but acknowledge the possible only modifies the objective function after any local (or global)
existence of other solutions. The updating algorithm is sometimes solution has been reached in such a way that the optimisation
run several times with perturbed initial parameter values, how- algorithm is forced to search for new, yet undiscovered, minima.
ever, this is not a systematic search over the solution domain. Only The outline of the paper is as follows. In the first section, the
limited studies have reported multiple model updating solutions theory of GA and SNT are briefly reviewed. The objective function
and consequently the problem and its remedies have not been used in the case studies considered is also introduced. Then vibra-
sufficiently explored [49]. tion testing and modal system identification of a simple laboratory
Artificial intelligence techniques are popular and proven alter- structure is described. Next, a detailed study of updating of an ana-
natives to the more classical optimisation techniques and have a lytical model of the laboratory structure using the proposed new
history of applications to different sectors of engineering, military, updating method combining GA and SNT is presented. The discus-
medicine, economics, etc. Global optimisation algorithms (GOAs), a sion than moves to a more challenging case of a full-scale cable-
branch of artificial intelligence, are numerical techniques that stayed pedestrian bridge, where bridge testing, modelling and
explore the search space widely in an attempt to increase the model updating using the proposed method are presented and
chance of discovering the global minimum [38,39,45]. discussed. The third example uses a numerically simulated steel
One of the efficient GOAs, which is used in this study, are space frame where an artificially introduced error in a target
genetic algorithms (GAs). GAs already have a track record of appli- frequency leads to a truly deceptive problem. Conclusions and
cations to condition assessment and damage detection via model recommendations wrap up the paper.
updating in civil engineering structures. For example, a GA was The main contribution of this study is a novel approach to
used for damage detection in a simulated beam structure and an model updating that combines GA and SNT and facilitates a
experimental beam structure by Perera and Torres [37]. Multiple systematic exploration of the optimisation domain for increased
damage scenarios were studied along with the effect of different confidence in finding the most acceptable solution. The new
noise levels on the simulated beam structure. Later, the GA was method performance is comprehensively evaluated using vibration
applied to a laboratory beam structure to verify its effectiveness data from a laboratory structure, a full-scale bridge and a complex
for damage detection and its assessment. Raich and Liszkai [40] numerical model of a space frame.
presented an advanced GA and applied it to a simulated beam
and frame structures for improving the performance of a damage
2. Theory
detection method. Different GA-based multi-objective optimisa-
tion schemes for damage detection and assessment were applied
GOAs make few assumptions about the problem to be opti-
to numerically simulated and experimental beam problems by
mised and can handle large or complex search spaces often
Perera et al. [36].
encountered in multidimensional updating problems. A methodol-
The application of GAs for detection of multiple minima in
ogy based on combining the stochastic search GA with SNT is
updating problems has recently attracted more interest. In a study
proposed and investigated in this paper to improve the perfor-
by Zárate and Caicedo [49], multiple admissible solutions to a
mance of GOA-based model updating for systematic search of solu-
model updating problem were identified using a modelling-to-gen
tions. In the subsequent sections, the theory of GAs and SNT are
erate-alternatives algorithm. A full-scale model of the Bill Emerson
explained. The objective function used in the case studies that
Memorial Bridge was updated and different plausible solutions
includes regularisation is also introduced.
were detected and discussed. The authors selected a solution
which had a better physical justification but a higher objective
function value instead of the global minimum. An evolutionary 2.1. Genetic algorithms
algorithm using novel GA operators and capable of identifying both
local and global optimal solutions was also proposed by Caicedo GAs [20] mimic the process of natural evolution according to
and Yun [9]. The algorithm was used on the simulated numerical Darwin’s theory at the chromosome level. GAs work on a popula-
example of the American Society of Civil Engineers (ASCE) Struc- tion of potential outcomes, or individuals, based on the principle
tural Health Monitoring (SHM) Benchmark Structure. A random of survival of the fittest to produce better and better solutions to
white noise was added to the acceleration records which resulted the optimisation problem at hand. Each individual is encoded as
in the creation of two minima and it was found that the local, a string of binary values referred to as the chromosome [27]. A
but physically correct, solution had a higher objective function population of candidate solutions is first randomly generated and
value than the other deceptive minimum. then three randomized operators, namely selection, cross-over
The majority of recent studies involving the use of GAs for and mutation, are applied. Selection [14] is a process of choosing
model updating [40,34,35] are related to the damage detection parents from the initial population that will create offspring for
problems on simulated or laboratory structures where the ‘correct the subsequent generation in such a way so as to favour the
answer’ (in the form of damage location and severity) is normally better-fitted individuals. There are different techniques available
known a priori to the analyst for the direct verification of updated for the selection such as the roulette wheel selection, rank selec-
results. However, when updating models of actual, full-scale struc- tion, tournament selection and others. In this research, the tourna-
tures, the ‘correct’ answer is typically unknown, which leads to ment selection is used, where several tournaments are held
considerable challenge in interpreting the obtained results. While, between a few solutions taken from the entire population at ran-
as explained earlier, GAs in their basic form attempt to locate the dom. The winners are placed in the pool for subsequent mating
global solution to an optimisation problem, they cannot fully guar- or cross-over. The cross-over takes two random parent chromo-
antee a search will always be successful. These challenges moti- somes from an enriched population obtained via previous selection
vated the approach proposed and explored in this study, which is and combines parts of their genes to form an offspring. In this
based on a systematic search over the solution domain for the research, the two-point cross-over is used as it improves explo-
potentially multiple minima of the objective function so that the ration capability compared to a single-point cross-over. The
168 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

cross-over ratio limits the population undergoing cross-over, The niche radius r is an important parameter as it is used to
because leaving some individuals unaltered may also be beneficial. define the size of the part of the search domain in the neighbour-
Selection and cross-over lead to irreversible loss of diversity, and to hood of a minimum where the objective function is modified.
counteract this mutation is introduced. Mutation creates new Smaller values of niche radii produce more concavity in the derat-
genetic solutions which help the algorithm to avoid being trapped ing function and limit the spatial extent of its effect but may intro-
in local minima. The mutation ratio specifies the fraction of each duce spurious minima with low objective function values, while
population that will undergo mutation and is usually kept small larger niche radii can affect the other true minima in the search
to not turn the algorithm into a random search. In mutation, parts space and even hide them completely from the search algorithm.
of the parent chromosome are randomly changed at locations The niche radius r has been determined in this study with the help
selected by chance. After the application of these three operators, of a method used by Deb [13] that suggests the value
the fitness value of each individual is evaluated. The process pffiffiffi
continues to form new populations and since the best-fitted k
r¼ pffiffiffi ð3Þ
individuals are favoured for selection and breeding, this leads to 2 kp
a population of individuals that are better suited for minimising
where k represents the dimension of the problem (the number of
the objective function. The GA search is terminated when the
parameters) and p is the expected number of minima. The niche
stopping criteria, such as the minimum fitness value, number of
radius of Eq. (3) assumes that all minima are equally distributed
generations or minimum change to the objective function in
throughout the search domain. However, this is not generally the
several consecutive generations, are met.
case and also the number of minima will not be known a priori.
Therefore, the value determined from Eq. (3) was used only as an
2.2. Sequential niche technique initial indicator of the order of niche radius, and a value equal to
50% of that calculated from Eq. (3) was used in this study. Each
The principle of SNT is to carry over knowledge gained during parameter has to be normalised between 0 and 1 for the use of
subsequent iterations of an optimisation algorithm so that differ- radius formula of Eq. (3).
ent minima are discovered one by one [5]. The basic idea behind
SNT is that when a minimum is found in the search domain, the
2.3. Objective function for model updating
surrounding area, referred to as the niche, is ‘filled in’ and no
longer attracts the individuals in subsequent iterations. This forces
The objective function quantifies the deviation of the analytical
the optimisation algorithm to converge to another, yet unvisited,
predictions of modal parameters from those obtained experimen-
niche. The process continues until the criteria such as the maxi-
tally. The following objective function is used in this study:
mum number of iterations, maximum number of discovered
!2 pffiffiffiffiffiffiffiffiffiffiffiffi 2
minima and the upper threshold value of the objective function Xn
f a;i  f e;i X
m
ð1  MACi Þ
at a minimum have been met. P ¼ a1 þ a2 þ a3 Pr ð4Þ
f e;i MACi
Note that while GA is used in this study as the basic search i¼1 i¼1

algorithm, SNT can be combined with many other algorithms with The first term is the total relative difference between the
remarkable ease as it only alters the objective function but not the experimental and analytical frequencies, where f represents the
algorithm itself. frequency, subscripts a and e refer to analytical and experimental,
Initial iterations in search of the first minimum are made with respectively, and n is the total number of frequencies considered.
the basic search algorithm without SNT by using the raw objective The second term measures the difference in mode shapes in terms
function. Once the first minimum has been found, the objective of the modal assurance criterion (MAC) [29], where m is the total
function values of the individuals in the vicinity of the minimum number of mode shapes considered. The third term, Pr, is related
are modified, and the search for the next minimum commences. to regularisation. Regularisation in model updating is often intro-
The modifications to the original objective function are introduced duced for ill-posed problems that may not have a unique solution
by multiplying it by a derating function using the following recur- [2,16,21,31,47]. Regularisation, following the original idea by
sive formula: Tikhonov [46], augments the objective function with new condi-
tions, that depend on updating parameters rather than the mea-
Pnþ1 ðxÞ ¼ Pn ðxÞ  Gðx; sn Þ ð1Þ
sured responses, in order to steer optimisation into the regions of
where Pnþ1 ðxÞ is the modified objective function to be used for search space where it is assumed or believed their values belong.
searching for the n + 1-th minimum, Pn ðxÞ is the previous objective The regularisation term for time being is left in a general form in
function used for searching for the n-th minimum, Gðx; sn Þ is the Eq. (4) to be specified later to suit the particular requirements of
derating function, x is the vector of updating parameters, and sn is the case studies considered. Finally, a1 through a3 are weighting
the n-th found minimum. factors allowing for relative promotion and demotion of the error
The following exponential derating function is used in this terms.
study [5]:
(
 3. Testing, modelling and model updating of a laboratory
nÞ
exp log m  rdðx;s if dðx; sn Þ < r structure
Gðx; sn Þ ¼ r
ð2Þ
1 otherwise
In this study, the testing, analytical or numerical modelling and
where m is the derating value used to control concavity of the model updating exercises of two experimental structures are
derating function, r is the niche radius, and d(x, sn) is the distance discussed. The experimental structures are a simple three degree
between the current point x and best individual sn. Beasley et al. of freedom (3DOF) laboratory structure and a full-scale cable-
[5] studied various derating curves, and found that the shape of stayed pedestrian bridge. Additionally, a numerical model of a steel
the curve is not influential as long as the curve fills the niche and space frame is investigated using simulated results. The laboratory
repels the optimisation algorithm from its neighbourhood. For that, based investigations, presented first in this section, aimed at eval-
the derating value m has to be high enough to sufficiently increase uating the proposed updating method in a controlled laboratory
the objective function values within the niche. environment and on a simple model to build confidence in its
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 169

capabilities and understand limitations before its application to the has a height of 700 mm. The steel plate is 650 mm  650 mm and
more challenging case of the full-scale bridge. This section provides has a thickness of 4 mm. The aluminium angles have dimensions of
a description of the laboratory structure, its dynamic testing and 30  30 mm with a thickness of 3 mm. The columns are attached
system identification, analytical modelling and, finally, model to the steel plate at the top and to a plywood sheet at the bottom
updating using the proposed combination of GA and SNT. The with L-shaped aluminium brackets having a width of 30 mm,
concepts and techniques used are also demonstrated and explained thickness of 4.5 mm and length of 75 mm. A close up of a
in detail so that the subsequent full-scale bridge and space frame bracket–column–plywood sheet connection is shown in Fig. 1c. A
exercises can be treated more succinctly in this respect. total of four M8 bolts are used to connect each bracket with the
steel plate or plywood sheet and an aluminium angle. It is impor-
tant to note that because of the bracket orientation the column
3.1. Description of the laboratory structure stiffness in the x-direction is considerably higher than in the
y-direction (being closer to a fixed end for the former and a pinned
The laboratory structure considered in this study is shown in end for the latter, respectively).
Fig. 1a (mounted on a shake table and equipped with four
accelerometers), whereas Fig. 1b shows the coordinate system 3.2. Testing and system identification of the laboratory structure
adopted for describing its motions, whose origin is in the geomet-
ric centre of the top plate. The structure is made of a stainless steel The structure was tested on a small uniaxial shake table that
top plate supported on four columns made of aluminium angles. It uses hydraulic pressure generated by a pump and a proportional
integral derivative controller for its motion. As the shake table
was able to excite the structure in one direction only, the structure
was fitted at an angle of 20° to the direction of shaking as shown in
Fig. 1a. This was done to excite effectively the structure in both
x- and y-direction, although the excitation level in y-direction
would be stronger than in x-direction. A small mass eccentricity
was also introduced by accelerometers (see Fig. 1a and b) that
resulted in some, albeit small, torsional excitation.
A total of five uniaxial accelerometers were used to measure the
response. One accelerometer measured the shake table motion,
one was fitted on the plate for measuring the accelerations in
y-direction, whereas three accelerometers were used to measure
the acceleration in x-direction, enabling also determination of
torsional accelerations (Fig. 1a and b). (Note that obviously this
setup is redundant as three accelerometers would suffice to
capture the 3DOF motion of the plate.) A PC and Matlab-based data
acquisition system was used to acquire the data from the
(a) accelerometers at a sampling rate of 200 Hz. A total of three
frequency sweep tests were conducted by exciting the structure
within a range from 0 to 15 Hz with the sweep duration of 155 s.
Spectral analysis was carried out on the data to obtain the
frequency response functions (FRFs) and subsequently determine
modal frequencies by peak picking. FRF is a measure of the system
response to the input signal at each frequency. FRFs for each degree
of freedom (translations in x- and y-direction and torsion) were
calculated from the auto-spectrum of the excitation (shake table
acceleration) and cross-spectra between the responses and excita-
tion [16]. To do so, the data was first filtered to remove any high
frequency content above 50 Hz and then Welch’s method using 5
data segments, 50% overlap and the Hamming window was applied
(b) [15]. FRFs from the three sweeping tests were then averaged to
produce the final estimate. To assess the quality of the FRFs and
distinguish between real and spurious peaks, coherence can be
used [15]. Coherence can be calculated using the aforementioned
spectra and the auto-spectrum of the response. High coherence
values, close to one, indicate that the response at a given frequency
is caused by the measured input rather than other sources of exci-
tation, whereas low coherence points out to possible spurious
results introduced by noise. The FRF absolute value for the single
accelerometer in the y-direction along with its phase and coher-
ence is shown in Fig. 2. A clear peak at 3.56 Hz in the FRF modulus
indicates a translational mode in the y-direction (Mode 1). A phase
change of 180° can also be observed at 3.56 Hz further confirming
the modal frequency. The coherence between the input and
(c) response at 3.56 Hz is higher than 0.85 also confirming the pres-
Fig. 1. Laboratory structure: (a) general view when fitted on the shake table and the
ence of a mode. Similar observations of the FRFs from the
direction of shaking, (b) coordinate system adopted, and (c) detail of column– accelerometers measuring in x-direction indicated a clear modal
plywood base connection. frequency (Mode 2) at 5.02 Hz. However, the identification of the
170 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

FRF modulus (m/s /N)

0.4

2
0.3
0.2
0.1
0
1 2 3 4 5 6 7 8 9
Frequency (Hz)
FRF phase (Degrees)

100

-100

1 2 3 4 5 6 7 8 9
Frequency (Hz)
1
Coherence

0.5

0
1 2 3 4 5 6 7 8 9
Frequency (Hz)

Fig. 2. FRF and coherence for acceleration in y-direction for laboratory structure.

Fig. 3. Stability diagram for acceleration in y-direction for laboratory structure.

torsional mode (Mode 3) from FRFs was not conclusive due to poor has been increased from 6 to 40. These are at 3.58 Hz, 5.01 Hz and
excitation and low signal to noise ratio and this mode was later 7.28 Hz, respectively.
estimated based on time domain system identification. The natural frequencies of the laboratory structure identified
As it is good practice to confirm system identification results by using the FRFs and N4SID method are shown in Table 1. It can be
another method, modal analysis was also carried out in the time
domain using the data driven subspace system identification
method (N4SID algorithm) [48] with stability diagrams [7] to Table 1
discern superfluous modes from true structural modes. With the Natural frequencies of laboratory structure using FRFs and N4SID method.
increase in the assumed model order in the state-space formula- Mode no. Mode type Exp. freq. (Hz) Damping ratios (%)
tion employed by N4SID, the identified modal properties should N4SID
Peak N4SID
remain stable in the stability diagram. In this research, for a mode picking
to be declared stable the relative change in the frequency had to be
1 Dominant translation in 3.56 3.58 1.2
less than 1% and MAC values greater than 0.90 between two y-direction
consecutive model orders. Stable modes are shown by black dots 2 Dominant translation in 5.02 5.01 0.8
in Fig. 3. It can be noted from Fig. 3 that there are three frequencies x-direction
which are stable for the laboratory structure when the model order 3 Dominant torsion – 7.28 0.8
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 171

seen that the frequencies identified by both of these methods The key to the symbols used in Eq. (6) is as follows: kij (i, j = x, y)
match very well for Mode 1 and 2 (where both estimates exist), is the translation stiffness coefficient for the elastic restoring force
and a single consistent frequency estimate is available for Mode acting in i-direction mobilised by the displacement in j-direction.
3 from the N4SID algorithm. Damping ratios identified by the The coordinates of the centre of stiffness are ex and ey. Finally,
N4SID algorithm are also shown. They range between 0.8% and kqq is the rotational stiffness of the system. Initial values for
1.2% indicating a lightly damped system. Mode shapes identified stiffness coefficients kxx and kyy in the mass-spring model were
from the N4SID method are shown in Table 2. For a more meaning- calculated from a finite element (FE) model of a single bracket–
ful comparison, the angular displacement values for the torsional column–bracket assembly formulated in SAP2000 [43] using 3D
DOF are multiplied by half of the plate breadth b = 0.325 m. It is 4-node shell elements which include the effects of membrane
noted that while y-direction translation strongly dominates Mode and plate bending behaviour. The stiffnesses were found by apply-
1, x-direction translation Mode 2, and torsion Mode 3, respectively, ing unit loads in turn in x- and y-direction and finding the resulting
there is small but noticeable presence of the other DOFs in each displacements at the top. The model had all displacements and
mode. This is because of the obvious mass eccentricity brought rotations restrained at the bottom where the bracket meets the
by the accelerometers off-centre locations, but can also be due to plywood. At the top, the bracket was restrained against rotations.
asymmetric stiffness. Also, when the force was applied in x-direction, the upper bracket
was not allowed to move in y-direction simulating the stiff
3.3. Mass-spring model for the laboratory structure diaphragm effect of the steel plate; an analogous condition was
assumed for the force in y-direction. Modulus of elasticity of
The steel plate of the laboratory structure was assumed to form aluminium was assumed as 70 GPa. The total stiffness values kxx
a rigid diaphragm and the structure was modelled as a 3DOF and kyy were found by adding contributions from the four columns
system. The coordinate system is shown in Fig. 1b. The DOFs used as 20,294 N/m and 8904 N/m, respectively. The difference in these
in the analytical model were the translations in x- and y-directions two stiffness values corresponds to the different stiffness of the
and the torsional angle multiplied by half of the steel plate breadth brackets in x- and y-directions as discussed earlier. The rotational
b, giving the DOF vector [x y b/]T, where superscript T denotes stiffness kqq of the whole structure was calculated as 3068 N m
transposition. In such a model, contributions to mass and stiffness using the column stiffnesses in x- and y-directions and multiplying
from the different system components are lumped into the corre- them by the appropriate squared coordinates of column locations.
sponding entries of the mass and stiffness matrices. The model As it was assumed that the structure was nominally symmetric, the
takes into account the obvious eccentricities in the mass distribu- values of kxy, ex and ey were all assumed to be zero. Table 3 shows
tion introduced by the accelerometers (see Fig. 3a and b). Also, the comparison of the natural frequencies and mode shapes
while from the point of view of stiffness the structure is nominally (MACs) between the experiment and initial analytical model. It
symmetric with respect to axes x and y, nevertheless possible can be seen that the frequencies differ by as much as 8.2%. MACs
differences in fastening of each column and misalignments may values are, however, very high, all in excess of 0.97.
have led to some violations of these symmetries and the centre
of stiffness not coinciding with the geometric centre of the steel 3.4. Model updating of the laboratory structure
plate. Those imperfections may give rise to off-diagonal terms in
the stiffness matrix and are also accounted for. In model updating, dynamic measurements such as natural
The mass matrix is frequencies and mode shapes are correlated with their mathemat-
2 3 ical model counterparts to calibrate the structural model. There is a
m 0 mym =b degree of uncertainty in the assessment of the actual properties of
6 m mxm =b 7 the materials used in the full-scale structure, geometry, as well as
M¼4 5 ð5Þ
2 the most realistic representation of the element stiffness, supports
sym: J=b
and connections between structural parts in the initial model. The
and the stiffness matrix is challenge of finding a set of suitable parameters having physical
2 3 justification necessitates the need to use physically significant
kxx kxy ðkxx ey þ kxy ex Þ=b updating parameters and suitable optimisation tools.
6 ðkxy ey þ kyy ex Þ=b 7
k¼4 kyy 5 ð6Þ
2 3.4.1. Selection of updating parameters for the laboratory structure
sym: ðkxx e2y  2kxy ex ey þ kyy e2x þ kqq Þ=b
Parameter selection is a crucial step for the success of model
The symbols used in Eq. (5) are as follows: m is the total mass, J updating. Too many updating parameters compared to the number
is the total second mass moment of inertia, and xm and ym are the of experimentally measured responses (natural frequencies and
coordinates of the centre of mass. In establishing these inertia- mode shapes (MACs)) may result in an ill-posed problem without
related properties, the steel plate together with the attached clear distinctive minima. On the other hand, if the number of
accelerometers, upper brackets and bolts and upper halves of the parameters is too small, the updating process may never converge
columns were taken into account. The total mass was measured to a solution close enough to the experimental results [47].
directly and found to be m = 17.5 kg. The total second mass Furthermore, it can only be justifiable to change the values of
moment of inertia was calculated to be J = 1.265 kg-m2 and mass
eccentricities to be xm = 0.010 m and ym = 0.006 m based on the
directly measured masses, dimensions and locations. Table 3
Comparison between frequencies and mode shapes (MACs) between initial analytical
model and experiment for laboratory structure.
Table 2 Mode no. Frequency MAC
Normalised mode shapes of laboratory structure obtained using N4SID.
Experiment by N4SID Initial FE model Error
Mode no. 1 2 3 (Hz) (Hz) (%)
Translation in x-direction 0.004 1.00 0.02 1 3.58 3.59 0.3 0.989
Translation in y-direction 1.00 0.07 0.07 2 5.01 5.42 8.2 0.974
Torsion 0.01 0.03 1.00 3 7.28 7.85 7.8 0.999
172 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

parameters that are significantly uncertain in the first place. premature convergence to local minima observed, it turned out
Finally, updating parameters should be influential, or in other that the smallest numbers of evaluations were for 20 and 30 chro-
words model responses sensitive to them, i.e. relatively small mosomes. Furthermore, combining GA with SNT aims at alleviating
changes in the parameter values should give rise to significant the risk of potential premature convergence – even if in the first
changes in the model responses. If this is not the case, during iteration only a local minimum is achieved the GA will still have
updating the parameters will vary widely and may take unrealistic an additional chance (or even more chances) to find the global
values. minimum. Tournament was used as the selection operator, and
In the analytical model derived in the previous section (Eqs. (5) tournament size was set to four. A parametric study for tourna-
and (6)), the possible parameters to be updated in the stiffness ment sizes larger than four revealed there was little improvement
matrix are kxx, kyy, kxy, kqq, ex and ey, whereas in the mass matrix in terms of convergence speed. Two point cross-over with a prob-
possible parameters to be updated are m, J, xm and ym. There are ability of 70% and flipping as mutation operator with a mutation
thus a total of 10 potential updating parameters. There are, on probability of 1% were adopted. The maximum number of genera-
the other hand, a total of six known experimental responses tions was set to 500, the threshold of the objective function to stop
(i.e. three frequencies and three MACs). The total mass of the struc- iterations to 105, and the threshold for the minimum change in
ture was directly measured in the laboratory and mass eccentrici- the objective function in five consecutive iterations to 107.
ties and second moment of inertia were determined accurately The objective function of Eq. (4) was used. The values of a1 and
from the actual measured dimensions and location of the masses. a2 were taken as one thus weighting both the frequency and MAC
Therefore none of the inertia parameters was considered for updat- related residuals equally. The regularisation parameter a3 was
ing. However, all stiffness-related parameters were selected for taken as zero as based on the sensitivity diagram ill-posing was
updating as they were considered to be more uncertain. not expected. (The relation of sensitivity to ill-posing is discussed
Sensitivity analysis was carried out for these parameters to later while updating the cable-stayed bridge, where regularisation
assess how influential they are. The relative sensitivity is the ratio was adopted.) The upper and lower bounds of the updating param-
of the relative change in the response value caused by a relative eters, within which they were allowed to vary during updating,
change in the parameter value. The use of relative rather than were selected based on judgment as ±20% for kxx, kyy and kqq.
absolute sensitivities enables treating all the parameters on equal The bounds for the remaining three parameters, i.e. kxy, ex and ey,
footing when comparing their influence. In this study, the relative were taken as ±500 N/m, ±0.050 m, ±0.050 m, respectively. The
sensitivities were calculated using a finite difference method by experimental results from the N4SID algorithm were taken as
changing the parameters by 0.1% with respect to their initial values updating targets.
for kxx, kyy and kqq. For parameters kxy, ex and ey, their initial values Model updating was first performed using GA only (i.e. without
are zero and the relative sensitivities were calculated for changes SNT) for 10 times with different randomly selected initial chromo-
between 0 and 10 N/m, and between 0 and 0.001 m, respectively. somes. The best updating solution that gave the lowest objective
The relative sensitivities of frequencies are graphically shown in function value is reported in Table 4 along with the standard
Fig. 4. It can be noted that parameters kxx, kyy and kqq strongly deviations for the 10 runs (in parentheses) and initial values of
influence Mode 2, 1 and 3, respectively. On the other hand, kxy, ex the model stiffness parameters. From Table 4, it can be noted that
and ey are much less influential and it can therefore be expected the maximum standard deviations of the updated parameters were
that in updating their values will vary more and they will have less small (except for standard deviation of kxy due to its low sensitivity
certain final results. It is, however, advisable to retain them as as discussed earlier), which assures that all the 10 solutions are in
otherwise it would be impossible to quantitatively explain the close vicinity of the same minimum.
observed small coupling between the three DOFs in mode shapes Although GA alone convincingly converged to the same mini-
(see Table 2). mum during all trials with random starting points, it is worthwhile
to explore systematically the search space for multiple minima,
3.4.2. Application of model updating to the laboratory structure and to that end GA with SNT was applied. The parameter m in
Model updating of the laboratory structure was attempted the derating function (Eq. (2)) was assumed as 1000 after several
using a GA implementation available in MATLAB [28]. A population preliminary simulations indicating that the value is sufficient to
of 30 chromosomes was used by the GA algorithm. This population repel the GA from the previously discovered niches. The niche
size was chosen by evaluating different candidate sizes between 10 radius for SNT was calculated according to Eq. (3) for four minima
and 320 and observing the numbers of objective function evalua- as 0.97, but to account for possible closeness of some of these min-
tions required to achieve convergence to the first minimum. While ima 50% of this value was adopted. After a run with this value, the
the same minimum was achieved in all the cases, i.e. there was no niche radius was further reduced by 50% and simulations rerun to

0.6

0.5
Relave sensivity of frequencies

0.4
kxx
0.3
kyy
0.2 k_rho
0.1 kxy
ex
0
1 2 3 ey
-0.1

-0.2
Mode No.

Fig. 4. Sensitivity of modal frequencies to structural parameters for laboratory structure.

 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 173

Table 4
Initial and updated parameter and final objective function values for GA alone for laboratory structure.

Model Parameter values (standard deviation) Final value of objective function

kxx (N/m) kyy (N/m) kxy (N/m) kqq (N m) ex (m) ey (m)
Initial model 20,294 8904 0 3068 0.000 0.000 0.013
Updated model 17,332 8873 305 2637 0.009 0.004 1.3  106
(121) (21) (84) (13) (0.001) (0.001)

make sure the larger radius did not hide any minima. At the end of Table 6
each simulation, a check was performed to make sure the new Experimental and updated frequencies (first minimum/lowest value of objective
function) and MACs using GA with SNT for laboratory structure.
minimum found was located outwith all the previously introduced
niches and therefore was not distorted. The runs with different Mode Experimental Updated FE model Error in MAC
radii produced equivalent results and filling of niches did not affect no. frequencies by N4SID frequencies (Hz) frequencies (–)
(Hz) (%)
subsequently discovered minima. GA with SNT was iterated four
times with random starting points and the results are shown in 1 3.58 3.58 0.001 0.998
2 5.01 5.01 0.0005 0.999
Table 5. Again, the best solution reached for each minimum is
3 7.28 7.28 0.0004 0.999
reported. It can be seen that the first minimum found was the same
one as found earlier by GA alone. In further iterations, different
minima with increased objective function values as compared to
observed which may be attributed to minor structural misalign-
the first minimum were found. Also, the updated parameter values
ments, as can be stiffness kxy found to be 305 N/m. The drop in
for those minima are in many cases markedly different than for the
torsional stiffness of the model, kqq, was to 2637 N/m, or 14.0%.
first minimum. This is because SNT forbids the search algorithm to
Since no algorithm can give absolute guarantee that the global
converge again to the same niche. The systematic search using GA
minimum was indeed found it is very important to subject results
with SNT gave an increased confidence in finding the global
of any inverse problem to critical scrutiny for credible and coher-
minimum.
ent quantitative and qualitative explanation of the outcomes. This
The updated model frequencies corresponding to the first min-
was done in our case in the previous paragraph and justifies the
imum (i.e. the one with the lowest value of the objective function)
correctness of the final solution we chose. It is of note that even
found by GA with SNT are shown in Table 6 along with the exper-
for a relatively simple case of the laboratory structure, boundary
imental frequencies and relative differences. The MAC values
conditions can be a major source of error in modelling – the reality
between the experimental and updated analytical mode shapes
often encountered in analytical simulations that challenges the
are also shown in Table 6. It has been found that GA has given
analyst.
improved results as compared to the initial analytical model
(see Table 3) and has decreased the difference between experimen-
tal and analytical frequencies from the largest error of 8.2% to prac- 4. Testing, modelling and model updating of a full-scale bridge
tically nil. All MAC values are higher than 0.997. These results
represent a very good agreement with the experiment. 4.1. Description of the bridge
The updated parameter values are discussed now to assess their
plausibility and explain the possible causes for changes. From The full-scale structure under study is a 59,500 mm long cable-
Table 4, the updated stiffness in y-direction, kyy, was found to be stayed footbridge with two symmetrical spans supported on
8873 N/m, which is very close to the initially estimated value of abutments, a central A-shaped pylon and six pairs of stays as shown
8904 N/m, a drop of just 0.3%. However, it has been found that in Fig. 5. Fig. 6 shows the deck cross-section, which comprises a
the updated stiffness in x-direction, kxx, was 17,332 N/m, which trapezoidal steel girder with overhangs of a total width of
is markedly less (14.6%) than the initially estimated stiffness value 2500 mm and depth of 470 mm, made of 16 mm thick plates, and
of 20,294 N/m. It is noted that while in y-direction the brackets will a non-composite concrete slab of thickness 130 mm. The concrete
bend relatively easily with respect to their weak axis (see Fig. 1c), slab was assumed in design to provide no contribution to the deck
in x-direction bracket bending deformations will be much more stiffness, although its mass and weight were accounted for. Closed
constrained. The arrangement of bolts fixing the bracket to the steel rectangular pipes having a cross-section of 250  150  9 mm
plywood and steel plate also provides for a much longer lever also run on both sides of the bridge deck and enclose two 100 mm
arm in y-direction. These factors will lead to larger forces exerted ducts for service pipes, with surrounding void spaces filled with
on the surface of plywood, bracket ‘digging’ into the plywood grout. Railing was provided on both sides of the bridge and it has
and consequently some amount of rocking at the base (and possi- a total height of 1400 mm. The sections of railings were discon-
bly also at the top) for x-direction motion lowering the correspond- nected from each other at every 8000 mm.
ing stiffness. Small values for stiffness eccentricities ex and ey, not The girder is continuous over the entire bridge length. It is sup-
exceeding 10 mm (or less than 1.4% of the plate breadth), are also ported on two elastomeric pad-type bearings of dimensions

Table 5
Updated model parameter and final objective function values for GA with SNT for laboratory structure.

Minimum no. Updated parameters Value of objective function

kxx (N/m) kyy (N/m) kxy (N/m) kqq (N m) ex (m) ey (m)
1 17,332 8873 305 2637 0.009 0.004 1.3  106
2 17,367 8957 233 2650 0.008 0.005 6.2  105
3 17,602 9392 499 2634 0.004 0.004 0.0009
4 17,045 9817 498 2953 0.014 0.002 0.0061
174 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

Fig. 5. Full-scale cable-stayed footbridge.

Railing 2 x 12mm
thick stanchions
1400
250 x 150 x 9 rectangular
pipe hollow section

Concrete slab
130
Steel girder
Box beam made from 470
16 mm thick steel plates

2500

Fig. 6. Cross-section of bridge deck (all dimensions in mm).

1600
SHAKERS
1000
1000
ACCELEROMETERS
N)

CA

CA
C A 3 (T e
on= N)

N)
75 k
k

5k B
B-
B-
95

-1
=5
2
n=

n (T
(T

i o e
io

ns ns 14100
en
s

io n
nsi

Te
en

sio
nsi

1( =5
(T

(Te

n=

B- 5k
on=
2

95

N)
B-

CA
B-3

kN
CA

75 k

)
CA

N)

4700

5709 8000 7993 8048 8048 7993 8000 5709

Fig. 7. Basic bridge dimensions, cable post-tension forces and location of shakers and accelerometers in the experiment (all dimensions in mm).

90  180  12 mm at the central pylon. At each abutment, two have a diameter of 32 mm. Different post-tension forces, ranging
150  150  12 mm elastomeric pad-type bearings are also from 55 kN to 95 kN in each cable, were specified in design. The
provided, but these have a special arrangement that allows for cables were connected to the top of the 22,400 mm high centre
longitudinal sliding while constraining any lateral horizontal pylon, which is composed of two steel I-sections joined with cross
displacements. The sliding bearings were provided to accommo- bracing that supports the deck. The size of the pylon I-sections is
date creep, shrinkage and temperature deformations, and to allow 400WC328 [4].
the bridge to move longitudinally in the event of a strong seismic
excitation. The distance between bearing axes is 450 mm. The 4.2. Testing and system identification of the bridge
abutments are supported by two concrete piles, and 10 concrete
piles and a pile cap are used at the central pylon. Experimental work has been carried out using uni-axial Honey-
The six pairs of stay cables are fixed to the deck at distances of well QA 750 accelerometers to measure structural response,
about 8000 mm centre-to-centre as shown in Fig. 7. All the cables uni-axial Crossbow CXL series MEMS accelerometers to measure
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 175

shaker input force and a desktop computer fitted with an NI DAQ one torsional mode shape identified from modal tests using the
9203 data acquisition card. Data was collected at a sampling rate N4SID method are shown in Fig. 10. Note Mode 1 and 2 have
of 200 Hz. Three APS ElectroSeis Model 400 shakers [3], capable relatively close frequencies and very similar shapes over the length
of providing a combined dynamic force of up to 1.2 kN, were used of the deck. However, they are distinct modes as was later
in a synchronised mode to impart excitation to the structure. confirmed by FE analysis. While experimental responses of cables
The bridge has been tested using sweep sine excitation ranging were not measured, the FE model (described in the following
from 1 Hz to 15 Hz with a total duration of 391 s applied by the section) revealed a different pattern of cable vibrations in these
three shakers located away from the centre line of the deck to two modes.
excite both the vertical and torsional modes. To excite horizontal
modes, the shakers were tilted at 90°. Accelerometers were placed 4.3. Finite element model of the bridge
on both sides of the deck to capture vertical and torsional
responses. The locations of the shakers and accelerometers are A good representation of bridge deck for box girder sections can
shown in Fig. 7. The vertical and horizontal tests were repeated be achieved by using beam elements with rigid links joining the
twice to ensure good quality data. cable elements with the deck elements [12,41]. In this research,
For system identification in the frequency domain, peak picking the bridge was modelled in SAP2000 [43] and the FE model is
from FRFs was used. An example of a FRF obtained during a vertical shown in Fig. 11. The deck and pylon were modelled using 3D
shaker test is shown in Fig. 8, where FRF modulus or magnitude, two-node frame elements which include the effects of biaxial
phase and coherence are shown. It can be noticed that the magni- bending, biaxial shear deformations, axial deformations and
tude has peaks at 1.64 Hz, 1.90 Hz, 3.66 Hz, 6.32 Hz, 7.42 Hz and torsion. The deck was discretized into 48 elements, whereas the
8.33 Hz. All but the last peak at 8.33 Hz, which is a torsional mode, pylon was discretized into 40 elements. The cables were modelled
correspond to vertical modes. The phase of the FRF shows a change using non-linear catenary elements capable of modelling slender
by 180° close to all these frequencies and coherence between exci- cables under self-weight and strain loading provided in SAP2000
tation and response has values of more than 0.8, further confirming and were discretized into four elements for each cable. These num-
these are modal frequencies. Some other peaks, e.g. just above bers of elements were selected as they enabled the replication of
10 Hz, can also be seen but the corresponding coherence values all the modes observed experimentally and further discretization
are low. Similarly, two resonance frequencies were identified using did not appreciably affect the results of numerical modal analysis
horizontal shaker excitation at 4.85 Hz and 5.36 Hz, respectively. of the first eight modes and only resulted in an increased compu-
For cross-checking the results of peak picking, the N4SID technique tational cost.
was used. For stability diagrams, the model orders ranged from 10 The modulus of elasticity for steel was taken as 200 GPa, for
to 80 and a threshold of 1% for frequency variation and a value cables as 165 GPa and for concrete as 28 GPa following available
above 0.8 for MAC between two subsequent model orders were design specifications. The cast in situ concrete slab was assumed
used. The stability diagram for a vertical shaking test is shown in to be fully composite with the steel girder resulting in a combined
Fig. 9. It can be seen that the six previously observed modes, five equivalent steel cross-sectional second moment of inertia of
vertical and one torsional, are stable and can be identified from 0.06140 m4 for horizontal bending, 0.00439 m4 for vertical bending
the vertical tests as shown by the black dots in the figure. Some and torsional constant of 0.00810 m4. (Note that this contradicts
spurious modes, that did not meet the stipulated stability criteria, the assumption made in design that there is no contribution from
were also detected as shown by the white dots in the figure. In a the slab to the stiffness of the deck. However, it was anticipated
similar way, two modes previously seen in the FRFs were identified that partial composite action between the steel and concrete did
from the horizontal tests. exist, as is often the case in real structures, and its actual extent
Table 7 summarises the natural frequencies identified from the will be quantified via model updating later.)
peak picking and N4SID method. It can be seen from the results The stiffness of the bearings in shear and compression was
that the frequencies identified by both methods match very well. calculated using the procedure proposed by Gent [17]. The formu-
The damping ratios identified by the N4SID method are also shown las take into account the effective load area, thickness of bearing
in Table 7. It is observed that damping in the bridge is small, rang- and its shape, and Young’s and shear moduli of the elastomer. Fur-
ing between 0.2% and 1.4%. The five vertical, two horizontal and thermore, as explained before the distance between bearing axes is

Fig. 8. FRF measured during vertical shaker test on cable-stayed bridge.

176 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

Fig. 9. Stability diagram for a vertical shaker test on cable-stayed bridge (black dots indicate stable modes).

Table 7 total shear, vertical and torsional stiffness for the pylon bearings
Experimentally identified natural frequencies and damping ratios for cable-stayed were found to be 2.14  106 N/m, 7.70  107 N/m and 3.90 
bridge.
106 N m/rad, respectively. The total vertical and torsional stiffness
Mode no. Mode type Experimental Damping ratios (%) of the abutment bearings were found to be 1.60  108 N/m and
frequencies (Hz) N4SID 8.86  108 N m/rad, respectively. No shear deformations were
Peak picking N4SID allowed at the abutment in accordance with the support design
1 1st vertical 1.64 1.64 0.2 explained earlier. The freedom of the abutment bearings to slide
2 2nd vertical 1.90 1.90 0.9 in the longitudinal direction was ignored; this was not expected
3 3rd vertical 3.66 3.69 0.5 to have any strong effects on the model accuracy as neither was
4 1st horizontal 4.85 4.86 0.8
the bridge excited in the longitudinal direction during dynamic
5 2nd horizontal 5.36 5.33 0.6
6 4th vertical 6.32 6.31 0.5 tests, nor were the corresponding modes identified or considered
7 5th vertical 7.42 7.42 1.0 in the analysis. The SAP2000 linear link elements were used to
8 1st torsional 8.33 8.32 1.4 model the bearings, where each element consists of six separate
springs responsible for six deformational degrees of freedom,
although as explained above some degrees of freedom were
450 mm and the torsional restraint provided by the bearings was constrained.
also taken into account [25]. The Young’s and shear moduli were An initial non-linear static analysis was performed to account
assumed after Gent [17] as 3.2 GPa and 0.8 GPa, respectively. The for the geometric non-linearity caused by the cable sag and this

Mode 1 (1st vertical): Frequency 1.64 Hz Mode 2 (2nd vertical): Frequency 1.89 Hz
1 1

-1 -1

Mode 3 (3rd vertical): Frequency 3.69 Hz Mode 4 (1st horizontal): Frequency 4.86 Hz
1 1

-1 -1

Mode 5 (1st horizontal): Frequency 5.33 Hz Mode 6 (4th vertical): Frequency 6.31 Hz
1 1

-1 -1

Mode 7 (5th vertical): Frequency 7.42 Hz Mode 8 (1st torsional): Frequency 8.32 Hz

1 1

-1 0 20 40 60 -1 0 20 40 60
Length (m) Length (m)

Fig. 10. Normalised vertical, horizontal and torsional mode shapes identified using N4SID method for cable-stayed bridge.
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 177

ted to many factors such as application of different tensioning

forces than those specified in design, relaxation of steel stresses
with time, and slippage in anchorages and between cable strands.
Stiffness of the deck depends on Young’s modulus of both steel and
concrete; especially the latter shows considerable variability. The
connection between steel girder and concrete slab will typically
y be designed to allow for either composite action or lack thereof.
x z However, real bridges will always exhibit a certain degree of
composite action which is difficult to predict. Furthermore, non-
structural elements, such as pavement, railings, and services, also
make a contribution to stiffness that is difficult to quantify and
model precisely. Also, stiffness of the bearings was assumed from
literature as the exact specifications were not known, and is thus
prone to uncertainty further exacerbated by the inherent variabil-
ity of elastomer properties.
There are three pairs of stay cables on each side of the central
pylon. The four identical cables closest to the abutments are
referred to as Cab-1, the four cables in the middle as Cab-2, and
the four cables nearest to the pylon as Cab-3 (see Fig. 7). The cables
Fig. 11. FE model of the bridge using SAP2000. were post-tensioned, as per design documentation, with different
forces indicated in Fig. 7. The effective axial stiffness of a cable
depends on its projected length, self-weight, axial stiffness EA
was followed by a linear dynamic analysis to obtain the natural fre-
(where E is Young’s modulus and A is cross-sectional area) and
quencies and mode shapes. A linear analysis that uses stiffness
tension force in the cable [33]. For taut cables with small sag, the
from the end of the non-linear static analysis for cable-stayed
influence of axial stiffness EA on the effective stiffness is more pro-
structures has been demonstrated to provide good results in previ-
nounced than that of the tension force. A simple hand calculations
ous studies [1]. Table 8 summarises the errors between the exper-
using the Ernst formula for cable stiffness [33] showed that the
imental frequencies and mode shapes (MACs) and those identified
effect of tension force on stiffness is much more important in
by the initial FE model. It has been found that the frequencies
cables Cab-1 compared to the remaining cables. This was later con-
obtained from the initial FE model differ from the experimental
firmed by the sensitivity analysis on the FE model, and therefore
frequencies by up to 8.6% and MAC values are between 0.980
only tension TCab-1 was included in the updating parameters.
and 0.999. The systematic attempts to improve the agreement
Sensitivity analysis using the FE model was conducted to con-
between the experimental and numerical predictions via GA and
firm the selected updating parameters can influence the analytical
SNT-based model updating are discussed in the next section.
responses. The selected parameters based on sensitivity analysis
and engineering insight into their uncertainty were deck flexural
4.4. Model updating of the bridge
stiffness for vertical (Ky,deck) and horizontal (Kx,deck) bending, deck
torsional stiffness (Kt,deck), axial stiffness of all cables (Kcable), cable
4.4.1. Selection of updating parameters for the bridge
tension for Cab-1 (TCab-1), and stiffness of bearings (Kbearing). The
In this study, only a relatively small number of parameters were
bearing stiffness, Kbearing, is to be understood as a single parameter
selected based on a priori knowledge of their potential variability,
whose changes affect proportionally the stiffness of the bearing
and a sensitivity analysis was carried out to confirm they influence
springs in the horizontal, vertical and torsional direction; this
the responses. The various inertia parameters of the structure were
was done to keep the number of bearing related updating param-
not included as these are typically less uncertain than stiffness
eters to a minimum.
parameters. The bridge was also supported at clearly defined
The sensitivities of modal frequencies to the updating parame-
points using specialised bearings that permitted making good
ters are shown in Fig. 12. It can be noticed that, as expected,
judgment about the appropriate modelling of boundary conditions,
parameters Ky,deck, Kcable and TCab-1 influence appreciably, albeit to
except the numerical values of bearing spring stiffness. Thus, can-
a varying degrees, the frequencies of vertical Modes 1, 2, 3, 6 and
didate parameters considered for calibration in this study were
7. Additionally, Kcable influences the torsional Mode 8. Parameter
cable tensions, cable axial stiffness, bending and torsional stiffness
Kx,deck influences the horizontal Modes 4 and 5. Two of the param-
of the deck and stiffness of the bearings.
eters that influence the torsional Mode 8, Kbearing and Kt,deck, require
The likely uncertainty of the parameters characterising cable
additional consideration. Ignoring a very small influence Kbearing
stiffness, i.e. cable axial stiffness and tension force, can be attribu-
has on other modes, the two parameters practically only influence
Mode 8. It can thus be expected, and indeed it was confirmed in
Table 8
preliminary simulations, that without constraining the two param-
Comparison between frequencies and mode shapes (MACs) between initial FE model
and experiment for cable-stayed bridge. eter variations attempts to update Mode 8 will create an ill-posed
problem with no unique solution for those parameters. To over-
Mode no. Frequency MAC
come the problem, a regularisation constraint was applied to keep
Experiment by N4SID Initial FE model Error the ratio Kt,deck/Kbearing approximately constant during updating.
(Hz) (Hz) (%)
The objective function, given in its general form in Eq. (4), now
1 1.64 1.66 1.2 0.999 becomes as follows:
2 1.90 1.88 1.1 0.995
!2 pffiffiffiffiffiffiffiffiffiffiffiffi 2
3 3.69 3.88 5.2 0.999 Xn
f a;i  f e;i X m
ð1  MACi Þ
4 4.86 5.28 8.6 0.999 P¼ þ þ 0:0005
5 5.33 5.45 2.3 0.993 i¼1
f e;i i¼1
MACi
6 6.31 6.79 7.6 0.990  
 K t;deck;j K t;deck;0 
7 7.42 7.76 4.6 0.980    ð7Þ
8 8.32 8.66 4.1 0.993 K bearing;j K bearing;0 
178 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

0.45 K y,deck
0.4
K x,deck

Relave sensivity of frequencies

0.35
K t,deck
0.3
K cable
0.25
T cab-1
0.2
K bearing
0.15

0.1

0.05

0
1 2 3 4 5 6 7 8
Mode No.

Fig. 12. Sensitivity of modal frequencies to selected updating parameters for cable-stayed-bridge.

where subscript j in the regularisation part denotes the values of remained undistorted. The different radii produced equivalent
parameters in the j-th iteration step and subscript 0 represents their results and the niches did not affect other minima.
initial values. The value of the regularisation term weighting factor Model updating by GA alone, i.e. without SNT, was attempted
a3 = 0.0005 (Eq. (4)) was adjusted by trial and error so that the ratio first. Ten independent runs were tested with different, randomly
of the two parameters did not change more than ±20%. It is selected initial populations of chromosomes. The best solution
acknowledged here that the way the two parameters were linked (i.e. the one giving the smallest value of the objective function)
can be considered arbitrary, but it can also be argued to be physi- from the 10 runs is shown in Table 9 in the form of the ratios of
cally plausible. Furthermore, the examination of the potential the updated to the initial stiffness values together with the stan-
ill-posing of the problem should in future be performed in a more dard deviations of these stiffness ratios from the 10 runs shown
systematic and rigorous way. An alternative to introducing regular- in parentheses. It can be seen that the maximum standard devia-
isation would be to declare the two stiffness parameters ‘unobserv- tion of the updated parameter ratios is only 0.0059, giving confi-
able’ and exclude them from updating altogether. However, that dence that all the solutions correspond to the same point in the
would also discard the information that experimental identification search space.
of Mode 8 has provided, and we decided not to proceed in this way. GA with SNT was now iterated five times with random starting
From the point of view of formulating and solving the optimisation points and the best solutions reached for each minimum (or SNT
problem, multiple objective optimisation [22,35] can be an alterna- iteration) are reported in Table 10. It can be seen that the first min-
tive to the single objective optimisation with regularisation in imum found is the same minimum as the one found earlier by GA
Eq. (7). In those approaches, several, possibly conflicting objectives, alone. In further iterations, different minima with increased objec-
e.g. the frequency and mode shape errors and the regularisation tive function values were found. The systematic search using GA
penalty, can be optimised separately to provide a selection of solu- with SNT gives a better understanding of the topology of the solu-
tions that trade off between the different objectives. However, this tion space and an increased confidence in finding the global
alternative will only be investigated in a future study. Herein, only minimum.
the single objective approach is used. The reasons for opting for this The updated frequencies corresponding to the first minimum
approach include that it is arguably simpler and that a multiple (i.e. the one with the lowest value of the objective function found
objective approach would be computationally more expensive, by GA with SNT), their errors compared to the experimental
especially if it were to be conducted at each iteration of SNT. results, and MACs between the updated and experimental mode
From the frequency errors in Table 7, it can be concluded that shapes are shown in Table 11. All frequency errors are generally
the initial FE model generally overestimates the stiffness, therefore less than 3% after updating. The largest error dropped from 8.6%
the lower bound has been selected as 40% and the upper bound to 2.9%, and in fact corresponds to a small error increase for the
was selected as +30% for all the parameters. first vertical mode. This indicates that it is possible to improve
the FE model considerably via adjusting the particular set of updat-
ing parameters considered, but some trade-off are inevitable. On
4.4.2. Application of model updating to the bridge structure
the other hand, MAC values did not change appreciably, with some
This section applies the proposed method to updating of the
small positive and negative changes in different modes, and the
pedestrian bridge FE model in order to explore the performance
of the approach. All the eight experimentally identified modal
frequencies and mode shapes were used. The parameters govern-
ing the working of GA and SNT were the same as in the case of Table 9
the laboratory structure but the maximum number of generations Ratios of updated to initial stiffness and final objective function values for GA alone
was set to 200 and the threshold of the objective function to stop for cable-stayed bridge.

iterations to 0.002. Similar to the laboratory structure, the niche Ratio of updated to initial stiffness (standard deviation) Final
radius for SNT was adopted as 50% of the value indicated by Ky,deck Kx,deck Kt,deck Kcable TCab-1 Kbearing
value of
Eq. (3), but this time for five minima, and was further reduced in objective
function
the subsequent rerun to 25% to make sure the larger radius did
not obscure any minima. A check was performed if all the minima 0.843 0.830 0.929 0.925 1.161 0.928 0.0023
(0.0048) (0.0018) (0.0053) (0.0008) (0.0021) (0.0059)
found were located outwith the others’ niches and therefore
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 179

Table 10 in the modelling of structural connectivity. Uncertainties in mate-

Ratios of updated to initial stiffness and final objective function values for GA with rial properties and slippage between cable strands can be other
SNT for cable-stayed bridge.
influencing factors. In general, the updated model represents the
Minimum Ratio of updated to initial stiffness Final value optimal solution for the optimisation problem of Eq. (7) that is also
no. of objective justified by engineering judgment, but hinges on the validity of the
Ky,deck Kx,deck Kt,deck Kcable TCab-1 Kbearing
function
initial model topology, discretization and parameterization. We
1 0.843 0.830 0.929 0.925 1.161 0.928 0.0023 argue that these are adequate.
2 0.855 0.933 0.913 0.850 1.006 0.855 0.0047
3 0.800 0.985 0.901 0.888 0.963 0.713 0.0066
Examining the results reported in Table 10, it can be seen that
4 0.816 1.138 0.737 0.730 1.017 0.736 0.0160 the other minima found are associated with higher objective func-
5 0.812 1.025 1.149 0.778 0.986 1.031 0.0073 tion values. The second minimum has an overall similar parameter
change pattern but twice as large the objective function value,
making it a less obvious choice. Subsequent minima result also
Table 11 in less realistic updated mechanical parameters. For example,
Experimental and updated frequencies (first minimum/lowest value of objective increase and decrease of stiffness related parameters, and signifi-
function) and MACs using GA with SNT for cable-stayed bridge. cant drops in cable related parameters with corresponding
Mode Experimental Updated FE model Error in MAC increases in bearing stiffness has been noted in other solutions
no. frequencies by N4SID frequencies (Hz) frequencies (–) obtained with SNT. These reasons provide ground to argue that
(Hz) (%) our chosen solution is the closest to the true dynamics of the
1 1.64 1.69 2.9 0.999 bridge.
2 1.90 1.86 2.1 0.996 The application of GA combined with SNT increases the confi-
3 3.69 3.71 0.5 0.998
dence in the obtained results as most of the solution space has
4 4.86 4.97 2.3 0.991
5 5.33 5.28 0.9 0.986 been searched sequentially. The user, based on engineering judg-
6 6.31 6.41 1.6 1.000 ment, can select the best solution from a list of different available
7 7.42 7.29 1.8 0.991 solutions. The results demonstrate how a multi-dimensional
8 8.32 8.40 1.0 0.993
search space can be systematically explored and how the applica-
bility of updating techniques can be extended to more challenging
problems.
minimum value is 0.986. MAC values, however, were already very
good in the initial FE model.
5. Updating of the numerical model of a space frame
The updated parameters should be physically meaningful;
otherwise it is difficult to justify the updating results. The vertical
The previous two examples were concerned with experimental
bending stiffness of the bridge deck has decreased to 84.3%, hori-
structures for which the ‘correct’ solution was not known a priori
zontal stiffness to 83.0% and torsional stiffness to 92.9% of their
and the solution finally adopted had to be chosen from amongst
respective initial values for the solution with the smallest objective
several provided by the GA with SNT and then justified based on
function value. This could be mainly attributed to the fact that the
engineering insight. Finally, the solutions adopted were the same
initial model takes the cast in-situ concrete slab as fully composite
as the first solution found by the GA alone and they also had the
with the steel girder, whereas no concrete contribution to deck
lowest objective function values. Nevertheless, by using the GA
stiffness was assumed in design and, consequently, no special
with SNT this decision was better informed by considering com-
shear connectors were provided. (For comparison, when one
peting alternatives, and therefore was made with a better insight
ignores the concrete slab, the deck stiffness is 84.6%, 73.6% and
into the updating problems and increased confidence.
60.9% for vertical bending, horizontal bending and torsional
However, the previous experimental examples did not yet
stiffness, respectively, compared to the fully composite case.) The
demonstrate fully the benefit of using SNT, i.e. safeguarding the
updated results reveal, contrary to the design assumption, that
analyst from accepting wrong solutions for truly deceptive prob-
there may be some, albeit at best only partial, composite action
lems. To demonstrate this advantage of using SNT, another exam-
between the slab and the steel girder contributing to the stiffness
ple is studied in this section. Since it is difficult to identify or
of the whole deck. The consistent decrease in all the parameters
purposefully create a deceptive problem using a complex experi-
related to the deck stiffness supports this conclusion. Nevertheless,
mental case, a numerical approach is adopted. We use a numerical
different than assumed stiffness of concrete and steel girder
model of the ASCE SHM Benchmark Structure [26] and our
(e.g. due to stiffeners), and non-structural components can also
approach to creating a deceptive updating problem with multiple
be responsible. However, with only the limited number of mea-
minima follows the idea of Caicedo and Yun [9]. However, the
sured modes available, further granularity in girder stiffness mod-
details of our implementation differ and when they do this is
elling would be difficult to be conclusively explored and is a
explained.
limitation of this updating exercise.
The bearing stiffness has decreased to 92.8%, which is a plausi-
ble reduction given that the initial value was assumed from litera- 5.1. Description of the space frame, FE modelling and simulated system
ture. Due to the regularisation term included in Eq. (7), the values identification results
of the updated bearing stiffness and torsional stiffness of the deck
change in a similar relative way. The increase in cable tension TCab-1 The ASCE SHM benchmark structure is a two-bay by two-bay
to 116.1%, shows that these post-tension forces are more than the four-storey steel space frame as shown schematically in Fig. 13.
designed value of 55 kN, indicating some possible overstressing of Each bay is 1.25 m wide and each storey is 0.9 m high. The struc-
the cables. On the other hand, the cable axial stiffness shows a drop ture comprises a total of nine columns (B199  9 sections), 48
to 92.5%. The latter result can be attributed to many factors. The FE beams (S75  100) and 32 diagonal bracing elements
model uses a rather coarse parameterization. As a result, potential (L25  15  3). The bracings are located on the outside periphery
localised stiffness changes may be lumped into those parameters. of the structure and only along one diagonal in each floor-bay
For example, the identified drop in the cable axial stiffness may panel. The mass is 3200 kg for the first storey, 2400 kg for the
well be because of slippage in the cable anchorages, i.e. uncertainty second and third storey and 1600 kg for the fourth storey,
180 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

Legend
Damaged beam
Damaged bracing

3.60 m
Group 2 of
damaged
members
2.
St

50
ro
ng

m
di
re
c

Group 1 of damaged members

on

2.50 m
on
direc
Weak

Fig. 13. Schematic view of the space frame model with introduced damage.

respectively. A data generation program called ‘datagen.m’ devel- Table 12

oped by the International Association for Structural Control – Correct and modified frequencies of numerical space frame model with damage.
American Society of Civil Engineers Structural Health Monitoring Mode Correct Frequencies with Introduced
Task Group [23] was used in this research. no. frequencies introduced error (Hz) frequency error (%)
A lumped mass FE model with 120 DOFs was used. The model (Hz)
constraints all the floor nodes to have the same horizontal transla- 1 7.17 7.17 –
tion and in-plane rotation. The columns and floor beams were 2 9.18 9.18 –
modelled as Euler–Bernoulli beams and the braces as pinned– 3 13.66 13.66 –
4 21.24 21.24 –
pinned bars.
5 25.95 25.95 –
Damage has been introduced into the FE model along the weak 6 35.71 34.00 4.8
direction of the frame by reducing the stiffness of one group of 7 39.25 39.25 –
structural members (Group 1: one ground level beam and two 8 42.21 42.21 –
adjacent ground level bracings) by 20% and the stiffness of another 9 46.81 46.81 –
10 56.74 56.74 –
group of members (Group 2: two ground level bracings located on
the opposite side) by 90%. The damaged elements are shown in
Fig. 13 using dashed lines. Such a damage can plausibly occur as
a result of a strong seismic event. The introduced damage state and comparisons cannot be conducted, and/or an automatic
can be described using a two-element parameter vector h0 = system identification algorithm is used without being subjected
[h1,0 h2,0] = [0.80 0.10], where the first parameter corresponds to to scrutiny by an experienced modal analysist. Model updating will
the remaining stiffness of Group 1 damaged members, the second now be attempted using these erroneous ‘experimental’
parameter to the remaining stiffness of Group 2 damaged mem- frequencies.
bers, and subscript 0 signifies the actual assumed damage. Note
that in all the subsequent simulations the degree of damage for 5.2. Model updating for the space frame model with ‘experimental’
all the members in Group 1 is assumed the same and a similar errors
assumption is applied to Group 2 members. The first 10 natural
frequencies of the frame from a numerical eigenvalue analysis of The model updating will try to identify only the stiffness loss in
the FE model corresponding to the actually introduced damage the affected members described by parameters [h1 h2]. This
state h0 are listed in Table 12. assumes that damage location is known. Again, it is not unrealistic
It is now assumed that all but one natural frequencies of the to imagine a scenario in which a visual inspection has detected the
frame have been identified with negligible error for example from locations of damaged members but could not assess the severity
a recorded seismic response. However, for one frequency, the and updating is then used for the latter purpose. For simplicity,
frequency of Mode 6, it is assumed that a 4.8% error in identifica- the objective function for updating considers only the frequency
tion occurred (see Table 12). This can be a plausible scenario where errors (i.e. a1 = 1 and a2 = a3 = 0 in Eq. (4)). Fig. 14 shows the values
noise in data can cause errors, especially if a limited number of of the objective function over a range of parameters [h1 h2]. This
short records, perhaps only one, are available, and so averaging function has two clear minima: the first at h1 = [0.00 0.93] with
 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182 181

Fig. 14. Objective function values for space frame model.

Table 13 problems, including full-scale structures and deceptive problems.

Updated parameters and final objective function values for GA with SNT for space The paper also emphasised that no updating exercise must accept
frame model.
numerical results without their careful interpretation and physical
Minimum no. h1 h1 Objective function value justification.
1 0.00 0.93 0.0024
2 0.90 0.05 0.0025
Acknowledgments

The authors would like to express their gratitude to organiza-

the objective function value of 0.0024, and the second at h2 = [0.89
tions and people that supported this research. Piotr Omenzetter’s
0.05] with the objective function value of 0.0025, respectively. The
work within the Lloyd’s Register Foundation Centre for Safety
second minimum corresponds to the actually introduced damage
and Reliability Engineering at the University of Aberdeen is
h0 = [0.80 0.10] (note it is slightly shifted because of the error in
supported by Lloyd’s Register Foundation. The Foundation helps
the frequency of Mode 6). The first minimum is a deceptive one:
to protect life and property by supporting engineering-related
it does not correspond to the actual state of the ‘physical’ system
education, public engagement and the application of research.
but has the smallest objective function value. Indeed, five indepen-
Ben Ryder of Aurecon and Graeme Cummings of HEB Construction
dent runs of a GA alone all converged to this minimum.
assisted in obtaining access to the bridge and information for mod-
(The parameters controlling the GA working were assumed the
elling. Luke Williams and Graham Bougen, undergraduate research
same as for the laboratory structure studied earlier.)
students, assisted with testing.
Updating using GA with SNT was subsequently attempted with
the expected number of minima set to two and the niche radius
taken as 50% of the value suggested by Eq. (3), and then further References
reduced to 25% in additional runs performed to check if no harmful
distortion was introduced. The results are shown in Table 13. It can [1] Abdel-Ghaffar AM, Khalifa MA. Importance of cable vibration in dynamics of
cable-stayed bridges. J Eng Mech, ASCE 1991;117(11):2571–89.
be clearly seen that only the second minimum found is the correct [2] Ahmadian H, Mottershead JE, Friswell MI. Regularization methods for finite
one. Had SNT not been used, only the wrong minimum would have element model updating. Mech Syst Signal Process 1998;12(1):47–64.
been found by GA alone. This illustrates the distinctive and [3] APSDynamics. Dynamic ElectroSeis shaker Model 400; 2012. <http://www.
apsdynamics.com/>.
dramatic improvement in updating results due to the use of SNT [4] AS/NZS. AS/NZS 3679 Part-1 Structural steel hot rolled bars and sections:
when a deceptive problem has been encountered. Jointly published by Standards Australia, Sydney and Standards New Zealand,
Wellington; 1996.
[5] Beasley D, Bull DR, Martin RR. A sequential niche technique for multimodal
6. Conclusions function optimization. Evol Comput 1993;1(2):101–25.
[6] Beck JL, Jennings PC. Structural identification using linear models and
earthquake records. Earthquake Eng Struct Dynam 1980;8(2):145–60.
A combination of GA and SNT has been proposed in this study to [7] Bodeux JB, Golinval JC. Application of ARMAV models to the identification and
systematically explore the search domain in updating problems damage detection of mechanical and civil engineering structures. Smart Mater
Struct 2001;10:479–89.
with multiple minima. SNT works by ‘filling in’ the objective func- [8] Brownjohn JMW, Xia PQ, Hao H, et al. Civil structure condition assessment by
tion niches, corresponding to the previously discovered minima, FE model updating: methodology and case studies. Finite Elements Anal Des
and forces GA to expand its region of search, thereby increasing 2001;37(10):761–75.
[9] Caicedo JM, Yun G. A novel evolutionary algorithm for identifying multiple
the chance of finding the minimum best describing the actual alternative solutions in model updating. Struct Health Monit 2011;10
structural characteristics. The performance of GA augmented with (5):491–501.
SNT has been explored using experimental modal analysis results [10] Catbas FN, Aktan AE. Condition and damage assessment: issues and some
promising indices. J Struct Eng, ASCE 2002;128(8):1026–36.
from a simple laboratory structure and a full-scale cable-stayed
[11] Catbas FN, Ciloglu SK, Hasancebi O, et al. Limitations in structural
pedestrian bridge. An additional deceptive numerical example identification of large constructed structures. J Struct Eng, ASCE 2007;133
using a space frame model has also been investigated. The results (8):1051–66.
show that the methodology proposed herein gives the analyst [12] Chang CC, Chang TYP, Zhang QW. Ambient vibration of long-span cable-stayed
bridge. J Bridge Eng, ASCE 2001;6(1):46–53.
more confidence in the model updating results and that it can [13] Deb K. Genetic algorithms in multimodal function optimization. Tuscaloosa
successfully be applied to a range of challenging updating (AL): The University of Alabama; 1989.
182 F. Shabbir, P. Omenzetter / Engineering Structures 120 (2016) 166–182

[14] Deb K. Optimization for engineering design: algorithms and examples. New [33] Nazmy AS, Abdel-Ghaffar AM. Three-dimensional nonlinear static analysis of
Delhi: Prentice-Hall of India; 1998. cable-stayed bridges. Comput Struct 1990;34(2):257–71.
[15] Ewins DJ. Modal testing: theory, practice, and application. Baldock, [34] Perera R, Fang S-E, Huerta C. Structural crack detection without updated
Hertfordshire, England: Research Studies Press; 2000. baseline model by single and multiobjective optimization. Mech Syst Signal
[16] Friswell MI, Mottershead JE. Finite element model updating in structural Process 2009;23(3):752–68.
dynamics. Netherlands: Kluwer Academic Publishers; 1995. [35] Perera R, Fang SE, Ruiz A. Application of particle swarm optimization and
[17] Gent AN. Engineering with rubber: how to design rubber components. third genetic algorithms to multiobjective damage identification inverse problems
ed. Munich: Carl Hanser Verlag; 2012. with modelling errors. Meccanica 2010;45(5):723–34.
[18] Goldberg DE, Deb K, Horn J. Massive multimodality, deception, and genetic [36] Perera R, Ruiz A, Manzano C. Performance assessment of multicriteria damage
algorithms. Parallel Problem Solv Nature 1992;2(2):37–46. identification genetic algorithms. Comput Struct 2009;87(1–2):120–7.
[19] Grimmelsman KA, Pan Q, Aktan AE. Analysis of data quality for ambient [37] Perera R, Torres R. Structural damage detection via modal data with genetic
vibration testing of the Henry Hudson bridge. J Intell Mater Syst Struct algorithms. J Struct Eng, ASCE 2006;132(9):1491–501.
2007;18(8):765–75. [38] Pintér J. Global optimization in action: continuous and Lipschitz optimization-
[20] Holland J. Adaptation in natural and artificial systems. Cambridge, Mass: MIT algorithms, implementations, and applications. Boston: Kluwer Academic
Press; 1975. Publishers; 1996.
[21] Hua XG, Ni YQ, Ko JM. Adaptive regularization parameter optimization in [39] Price K, Storn RM, Lampinen JA. Differential evolution: a practical approach to
output-error-based finite element model updating. Mech Syst Signal Process global optimization. New York: Springer-Verlag; 2005.
2009;23(3):563–79. [40] Raich AM, Liszkai TR. Improving the performance of structural damage
[22] Hui L, Qingfu Z. Multiobjective optimization problems with complicated detection methods using advanced genetic algorithms. J Struct Eng, ASCE
Pareto sets, MOEA/D and NSGA-II. IEEE Trans Evol Comput 2009;13 2007;133(3):449–61.
(2):284–302. [41] Ren WX, Peng XL. Baseline finite element modeling of a large span cable-
[23] International association for structural control – american society of civil stayed bridge through field ambient vibration tests. Comput Struct 2005;83
engineers structural health monitoring task group; 1999. <http://wusceel.cive. (8–9):536–50.
wustl.edu/asce.shm/benchmarks.htm> [accessed January 2011]. [42] Salawu OS, Williams C. Review of full-scale dynamic testing of bridge
[24] Jaishi B, Ren WX. Structural finite element model updating using ambient structures. Eng Struct 1995;17(2):113–21.
vibration test results. J Struct Eng, ASCE 2005;131(4):617–28. [43] SAP2000. Structural analysis program, computers and structures, CA; 2009.
[25] Jaishi B, Ren WX. Finite element model updating based on eigenvalue and [44] Schlune H, Plos M, Gylltoft K. Improved bridge evaluation through finite
strain energy residuals using multiobjective optimisation technique. Mech element model updating using static and dynamic measurements. Eng Struct
Syst Signal Process 2007;21(5):2295–317. 2009;31(7):1477–85.
[26] Johnson E, Lam H, Katafygiotis L, Beck J. Phase I IASC-ASCE structural health [45] Storn R, Price K. Differential evolution – a simple and efficient heuristic for
monitoring benchmark problem using simulated data. J Eng Mech ASCE global optimization over continuous spaces. J Global Optimiz 1997;11
2004;130(1):3–15. (4):341–59.
[27] Knowles J, Corne D, Deb K. Multiobjective problem solving from nature: from [46] Tikhonov AN. Regularization of incorrectly posed problems. Soviet Math
concepts to applications. New York: Springer-Verlag; 2008. 1963;4(1):1624–7.
[28] MathWorks. Global Optimization Toolbox for use with MATLAB; 2010. [47] Titurus B, Friswell MI. Regularization in model updating. Int J Numer Methods
[29] Möller PW, Friberg O. Updating large finite element models in structural Eng 2008;75(4):440–78.
dynamics. AIAA J 1998;36(10):1861–8. [48] Van Overschee P, De Moor B. N4SID: subspace algorithms for the identification
[30] Moon FL, Aktan AE. Impacts of epistemic (bias) uncertainty on structural of combined deterministic–stochastic systems. Automatica 1994;30(1):75–93.
identification of constructed (civil) systems. Shock Vibrat Digest 2006;38 [49] Zárate BA, Caicedo JM. Finite element model updating: multiple alternatives.
(5):399–420. Eng Struct 2008;30(12):3724–30.
[31] Mottershead JE, Foster CD. On the treatment of ill-conditioning in spatial [50] Zivanovic S, Pavic A, Reynolds P. Finite element modelling and updating of a
parameter estimation from measured vibration data. Mech Syst Signal Process lively footbridge: the complete process. J Sound Vibrat 2007;301(1–
1991;5(2):139–54. 2):126–45.
[32] Mottershead JE, Friswell MI. Model updating in structural dynamics: a survey.
J Sound Vibrat 1993;167(2):347–75.