See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261986821

Collapse fragility curve development using Monte Carlo simulation and

artificial neural network

Article  in  Proceedings of the Institution of Mechanical Engineers Part O Journal of Risk and Reliability · May 2014
DOI: 10.1177/1748006X13518524

CITATIONS READS

9 1,564

4 authors:

S. Bahram Beheshti-Aval Ehsan Khojastehfar

Khaje Nasir Toosi University of Technology Vali-e-Asr University Of Rafsanjan
38 PUBLICATIONS   456 CITATIONS    10 PUBLICATIONS   14 CITATIONS   

SEE PROFILE SEE PROFILE

Mohammad R Zolfaghari Kourosh Nasrollahzadeh

Khaje Nasir Toosi University of Technology K. N. Toosi University of Technology
51 PUBLICATIONS   265 CITATIONS    24 PUBLICATIONS   167 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Analysis of Uncertainties in the Development of Seismic Fragility Curves for Equipment Using Virtual Shake-table View project

Master Thesis View project

All content following this page was uploaded by S. Bahram Beheshti-Aval on 30 April 2014.

The user has requested enhancement of the downloaded file.

 Proceedings of the Institution of Mechanical
Engineers, Part O: Journal of Risk and
Reliability
http://pio.sagepub.com/

Collapse fragility curve development using Monte Carlo simulation and artificial neural network
Ehsan Khojastehfar, Seyed Bahram Beheshti-Aval, Mohammad Reza Zolfaghari and Kourosh Nasrollahzadeh
Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability published online 16
January 2014
DOI: 10.1177/1748006X13518524

The online version of this article can be found at:

http://pio.sagepub.com/content/early/2014/01/16/1748006X13518524

Published by:

http://www.sagepublications.com

On behalf of:

Institution of Mechanical Engineers

Additional services and information for Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability can
be found at:

Email Alerts: http://pio.sagepub.com/cgi/alerts

Subscriptions: http://pio.sagepub.com/subscriptions

Reprints: http://www.sagepub.com/journalsReprints.nav

Permissions: http://www.sagepub.com/journalsPermissions.nav

>> OnlineFirst Version of Record - Jan 16, 2014

What is This?

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

Original Article

Proc IMechE Part O:

J Risk and Reliability
0(0) 1–12
Collapse fragility curve development Ó IMechE 2014
Reprints and permissions:
using Monte Carlo simulation and sagepub.co.uk/journalsPermissions.nav
DOI: 10.1177/1748006X13518524

artificial neural network pio.sagepub.com

Ehsan Khojastehfar, Seyed Bahram Beheshti-Aval,

Mohammad Reza Zolfaghari and Kourosh Nasrollahzade

Abstract
Seismic fragility curves represent likelihood of structures meeting various damage stages. Epistemic as well as aleatory
uncertainties associated with seismic loads and structural behaviors are usually taken into account in order to analytically
develop such curves. Such structural analyses are time-consuming, demanding extensive computational efforts. In this
study, in order to reduce this endeavor, artificial neural network method is applied to develop structural seismic fragility
curves under collapse damage state, considering effects of record-to-record variability and modeling parameter uncer-
tainties. Structural analyses are performed for a limited number of scenarios of structures under a limited number of
recorded strong ground motion records. Probability distribution for each modeling parameter was used to simulate each
structure scenario. Incremental dynamic analysis was used to assess spectral acceleration associated with collapse limit
state for each structure scenario. The results of the analyses were used to train and validate a three-layered artificial
neural network, and Monte Carlo simulation is implemented based on trained neural network for a sample moment-
resisting steel frame in order to derive collapse fragility curve. Application of the proposed method enhances accuracy
of identical computational run time compared with response surface–based method.

Keywords
Epistemic uncertainty, collapse fragility curves, moment-resisting steel frames, artificial neural network, Monte Carlo
simulation

Date received: 4 June 2013; accepted: 3 December 2013

Introduction exceedance probability of structures being subjected to

collapse stage. Collapse fragility analysis is one of the
Structural collapse has been one of the main causes of main components of consequence-based earthquake
economic and human losses during past earthquakes.1,2 engineering.8 Furthermore, seismic loss estimation is
Sideway collapse, which is defined as lateral instability decomposed into two major limit states (namely, col-
of structures excited by seismic strong ground motions, lapse and non-collapse limit states). In investigation of
has gained the attention of many researches recently.3–5 both states, collapse fragility curve plays an important
Furthermore, preventing and controlling structures role.
from such damage state during earthquakes has been Two distinct sources of uncertainties affect median
the major concern in force-based, performance-based and dispersion values of probabilistic distribution
and recently proposed consequence-based seismic eva- representing collapse fragility curves. Aleatory
luation procedures.6–8 Incorporation of uncertainties
associated with seismic load as well as building beha-
vior in structural analyses has been addressed by the Faculty of Civil Engineering, K. N. Toosi University of Technology Tehran,
consequence-based seismic evaluation procedures. Iran
Uncertainty analyses in conjunction with more reliable
prediction schemes result in better evaluation of struc- Corresponding author:
Ehsan Khojastehfar, PhD Candidate, Faculty of Civil Engineering, K.N.
tural collapse capacity. Representation of such results Toosi University of Technology, Mirdamad Ave. West, Tehran, I.R. of Iran
in probabilistic term provides collapse fragility curves, Post Code: 19697 64499 Tel: 98 21 8888 2991-3 Fax: 98 21 8879 7469
which represent strong ground motion levels versus Email: [email protected]

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

2 Proc IMechE Part O: J Risk and Reliability 0(0)

uncertainties associated with random nature of seismic and resultant output layers are considered as the spec-
strong ground motions are usually taken into account tral accelerations (SAs) corresponding to various limit
by applying a set of recorded or simulated strong states.
ground motion time histories. This would allow the While efficiency of ANN method has been shown to
implementation of uncertainties from seismic source derive fragility curves of limit states other than collapse
effects as well as duration and frequency content of (i.e. immediate occupancy, life safety and collapse pre-
strong ground motion records. There are also epistemic vention) by Mitropoulo and Papadrakakis,16 the main
uncertainties associated with lack of information as objective of this article is to demonstrate efficiency of
well as inability of analytical models to mimic all ANN method in deriving collapse fragility curves,
aspects of seismic behavior of structures.9 while modeling parameter uncertainty effects are
Implementation of such uncertainties in analysis taken into consideration. As an extension to the work
process requires simulation and nonlinear time history by Mitropoulo and Papadrakakis,16 similar ANN
analysis of many combinations of seismic load and approach is applied to implement both effects of
structural scenarios. Simple method such as the first- strong ground motions’ variability (aleatory uncer-
order second-moment (FOSM) to more elaborate tainties) and modeling parameter uncertainties (epis-
method such as crude Monte Carlo method have been temic uncertainties) in development of fragility curves
utilized to incorporate such uncertainties. Crude Monte for collapse limit state of structures. A three-layered
Carlo simulation method requires a large number of network is trained and validated according to limited
simulations in order to properly convolute full range of numbers of simulations derived from modeling para-
probabilistic distributions associated with each source meters of structure and strong ground motion
of uncertainty which could be too time- and resource- records. The responses of structure simulated by mod-
consuming.4,5 In order to overcome this problem, the eling parameters under ground motion excitation are
response surface in conjunction with Monte Carlo captured through application of incremental dynamic
simulation method has been proposed recently which analysis (IDA) method. Analytically derived SA for-
reduces computational efforts.4 On the other hand, the cing structure into the collapse stage (SAcollapse) is
response surface method could be substituted with arti- considered as the target for the network. Monte Carlo
ficial neural network (ANN) method to involve effects simulation based on trained ANN results in collapse
of uncertainties in reliability problems. The main lim- fragility curve, in which effects of both strong ground
itation of response surface–based Monte Carlo method motion uncertainties and modeling parameter uncer-
is assumption of fixed functional form to calculate tainties are involved. To justify the efficiency of the
mean and standard deviation of collapse fragility proposed method, a three-story moment-resisting
curves. Furthermore, utilizing the higher order of steel frame is modeled as the case study in this work.
response functions requires more data to determine Results of quadratic and logarithmic response
coefficients. It was shown that ANNs can be used to surface–based and proposed method are compared in
approximate any form of functions. Cardaliaguet and view of collapse fragility curves.
Euvrand12 used an ANN algorithm to approximate a
function and its derivatives in control theory. Chapman
and Crossland13 presented an example of ANN appli- Fragility curve formulation
cation for prediction of the failure probability of pipe
work under varying operation conditions. Li14 showed IDA is a regular method in estimating fragility curves
that any multivariate function and all its existing deri- for various limit states of structures affected by earth-
vatives can be simultaneously approximated by a radial quakes.17 Each IDA curve is prepared by several non-
basis ANN, where the assumptions on the functions are linear dynamic analyses of structure, while it is affected
relatively mild. by increasing intensities of strong ground motions.
ANN approaches have been used for developing fra- These curves represent structural response parameter,
gility curves in a limited number of studies. Lagaros entitled as engineering demand parameter (EDP), ver-
and Fragiadakis15 applied ANN for the rapid evalua- sus characteristics of affected strong ground motion,
tion of the probability of exceedance of the limit state entitled as intensity measure (IM). As a result of IDA
for a specific hazard level. Papadrakakis et al.26 pro- curves, one may derive fragility curves for different
posed Monte Carlo simulation based on ANN for the limit states by horizontal or vertical statistics
vulnerability analysis of large concrete dams. Lagaros methods.16
and Fragiadakis15 applied Monte Carlo simulation IM-based collapse limit state is defined as the IM of
based on ANN while randomness associated with strong ground motion while the excited structure under-
material properties and geometry as well as imposed goes dynamic instability. In other words, IMcollapse is
seismic loading was considered. ANN method is defined as the IM of last-converged point on an IDA
applied by Mitropoulo and Papadrakakis16 to develop curve. A point exists on each IDA curve which is repre-
fragility curves for various limit states of concrete sentative of IMcollapse, and the resultant collapse fragi-
structures. They proposed that input layers fed in lity curve is the probability distribution fitted to
ANN consist of strong ground motion parameters, IMcollapse values.5 Applying this method, the

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

Khojastehfar et al. 3

mathematical statement of collapse fragility curve is based method was loss of accuracy in approximated
described as equation (1) collapse fragility curves. In this research, to increase
the accuracy of collapse fragility curve, prediction of
P(CollapsejIM = imi ) = mean and standard deviation values is to be done by
P(imi . IMcollapse ) = FIMcollapse (imi ) ð1Þ models with fewer approximation errors. To demon-
strate the efficiency of proposed method to predict
Log-normal probability distribution function is collapse fragility curve fully, quadratic and logarith-
shown as the optimum probability distribution to rep- mic response surfaces which have been applied in pre-
resent collapse fragility curves of structures.12 Applying vious researches are implemented.5,23
log-normal probability distribution as cumulative
probability distribution of equation (1) results in equa-
tion (2) Epistemic uncertainty treatment
  In this article, ANN is implemented to predict SA,
Ln(imi )  Ln(hc )
P(CollapsejIM = imi ) = F ð2Þ as strong ground motion IM, corresponding to
bc collapse limit state. ANNs aim to mimic the elemen-
In equation (2), F represents Gaussian distribution tary functions of biological neurons through applica-
tion of artificial neurons and present a computational
function, hc is the median and bc is the dispersion of
device to calculate approximate output given an
collapse probability function.
input set.
The input is organized as a layer of neurons, each
Involvement of epistemic uncertainties corresponding to one of the input variables, and the
output is contained in an output layer. Intermediate,
Mean estimate, confidence interval5,18–20 and Monte hidden layers contain a number of neurons that
Carlo simulation methods21 are three regular receive information from the input layer and pass it
approaches to evaluate and combine effects of aleatory on to subsequent layers. The complicated relationship
and epistemic uncertainties. In mean estimate method, between a set of input data, set as modeling para-
it is assumed that only variance of fragility curves is meters and characteristics of strong ground motions,
changed affecting epistemic uncertainties; on the other and a set of output data, set as IM corresponding
hand, in confidence interval method, mean values are to collapse, is predicted by trained ANN. The neu-
affected involving epistemic uncertainties and variance rons in each layer are connected to the neurons in
remains unchanged. Unlike these simplifying assump- next layer by links which are characterized by a
tions, it is shown that epistemic uncertainty causes a weight W. The input data to a neuron are processed
shift in both mean and standard deviation values of by a transfer function h before being sent forward to
collapse fragility curves.21 In Monte Carlo method, the neurons in the next layer. For a network with a
thousands of simulations for modeling parameter val- single hidden layer, output can be presented by equa-
ues based on their statistic distributions are implemen- tion (3)
ted and then the structure is analyzed assuming !!
simulated values as modeling parameters of the struc- XJ X
N

ture. The thousands of the probability of collapse ver- Output = h wkj h wji (X)i ð3Þ
k=0 i=0
sus IM values denoted as collapse fragility curves
involving effects of epistemic uncertainties resulted In equation (3), h() is the transfer function, w repre-
from this rigorous analyses. This method is very ela- sents the weights and (X)i are input variables plus one
borative in practice due to the time needed for several bias term. N and J are the number of neurons in input
nonlinear dynamic analyses of structure for each simu- and intermediate layers, respectively.
lated value of modeling parameter. The training procedure of ANN consists of
Monte Carlo simulation applying a predefined definition architecture of the network, adjusting
regressed function, as response surface, has been pro- weight values and selection of transfer functions.
posed as an alternative to direct time history dynamic Normally, a fraction of the available data is used for
analysis to reduce the computational effort in the con- training (e.g. 80%), while the rest are used for valida-
text of the previous researches.22 In this method, first, tion of the neural network predictions. The number
fixed formats of functions were interpolated to the of neurons in the hidden layer must be set to make a
limited number of simulations of modeling variables, prediction of output values with minimum errors, but
as inputs which lead to resultant means and standard not so large that the network may produce inaccurate
deviations of collapse fragility curves and as outputs values for input combinations which do not exist in
of the function. In the next step, means and standard training data. Weight matrices and bias vectors are
deviations of collapse fragility curves for a large num- adjusted to make the prediction error minimum.
ber of simulations of modeling parameters are calcu- Mean square function is normally used as an indica-
lated applying derived analytical functions. The price tor of neural network error which is presented in
of efficiency in analysis time in the response surface– equation (4)

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

4 Proc IMechE Part O: J Risk and Reliability 0(0)

step of Monte Carlo method, large numbers of

realizations for input vectors are simulated. These
realizations are based on probability distributions of
modeling parameters and characteristic parameters of
strong ground motions. In the second step, SA corre-
sponding to collapse (SAcollapse) is predicted by
trained neural network. The probability distribution
fitted to values of SAcollapse presents collapse fragility
curve which involves effects of both aleatory and epis-
temic uncertainties.

Sample structure and analytical

assumptions
The proposed method is applied to a three-story, three-
bay moment-resisting steel frame (Figure 1). The grav-
ity loads are assumed according to Table 1. The three-
dimensional (3D) frame was designed based on UBC-
97 code (considering R = 8.5).6 All floors are assumed
Figure 1. 3D view of sample structure. to be rigid diaphragm.
The fundamental period of the frame is 1.075 s. The
nominal yield strength is 240 MPa which is used for all
Table 1. Gravity load assumptions. members in the design. The section properties of mem-
bers are shown on two-dimensional (2D) analytical
Dead load of regular story 500 kg/m2 model. All beam sections are from IPE 300 and all col-
Partition load 100 kg/m2
umn sections are from tube type. Exterior columns
Live load of regular story 200 kg/m2
Dead load of roof 520 kg/m2 are TUBE 180 3 180 3 16 and interior ones are from
Live load of roof 150 kg/m2 TUBE 200 3 200 3 16. The assumed structure is sym-
Perimeter wall load 580 kg/m metric in plan and elevation, which allows 2D structural
analyses. The 2D frame shown in Figure 2 represents
X1 one interior frame of a series of identical frames.
E= ðY(xm )  tm Þ2 ð4Þ The proposed method is applied to an interior frame.
m
m
Moment-resisting connections are considered as rota-
In equation (4), m is the number of training pairs (x, t); tional springs whose behaviors are based on modified
xm and tm are the mth input and target data, respec- Ibarra–Medina–Krawinkler model.24 M2-WO panel
tively; and Y is the value predicted by neural network zone model is considered since yielding in the beams, col-
while xm is assumed as the input of the network. The umns and panel zones is represented well by this model.25
optimum weights of the network, which minimize net- The parameters of moment rotation are considered
work error, are evaluated by solving the minimization as epistemic uncertainty, while strong ground motions
problem. A common minimization algorithm, which is applied in nonlinear time history analysis of the struc-
also applied in this study, is back propagation algo- ture are considered as aleatory uncertainty. IM-based
rithm, in which updating of the weights is done by sev- formulation is applied to derive collapse fragility curves
eral iterations and along the direction of steepest from IDA of the sample structure.
descent. The weights of the network in iteration t + 1 The backbone curve of presumed moment rotation
are calculated by equation (5) model, referred to as modified Ibarra–Medina–
Krawinkler model,20 is shown in Figure 3.
w(t + 1) = wt + Dwt ð5Þ Definitions of modeling parameters, shown in
t Figure 3, are as follows:24 uc is the cap rotation, My is
In equation (5), w is the weight matrix in iteration the effective yield moment, uy is the effective yield rota-
t and updating part Dwt is calculated by equation (6) tion, uu is the ultimate rotation capacity, up is the plastic
Dwt = a  Dwt1 + h  dt ð6Þ rotation capacity and upc is the post-capping rotation
capacity.
where dt specifies the direction of search and consists of The hysteretic behavior of the connection is defined
partial derivatives of error function with respect to based on deterioration rules according to hysteretic
weights, a is the corresponding step size and h is the energy dissipated in each hysteretic cycle. The deteriora-
momentum term defined in [0,1). tion of basic strength, post-capping strength, unloading
After training the network, Monte Carlo simulation stiffness and reloading stiffness can be considered in
is implemented based on trained network. In the first this model.24

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

Khojastehfar et al. 5

Figure 2. The analytical model of three-story, three-bay moment-resisting frame under consideration.

small distance (LMSR) according to Medina.10 These

records are ordinary strong ground motions recorded
in California and do not involve pulse-type near-field
characteristics. Stiff soil is assumed (corresponding to
type D according to National Earthquake Hazards
Reduction Program (NEHRP)). Source to site distance
is between 13 and 40 km and moment magnitude ranges
from 6.5 to 6.9.
Arias intensity (IA), characteristic intensity (IC) and
absolute velocity (CAV) of strong ground motions are
considered as strong ground motion parameters and
are defined by equations (8)–(10). They are selected to
reflect amplitude, frequency content and duration of
Figure 3. Backbone curve of moment rotation model based on applied strong ground motions. Furthermore, they have
modified Ibarra–Medina–Krawinkler model.24 a good correlation with seismic-induced damages.16 In
equations (8)–(10), a(t) is the strong ground motion his-
tory, Td is the strong ground motion duration and arms
Capacity of energy dissipation of the component, by is the root mean square of acceleration history. These
which deterioration rules are formulated, is defined as parameters and deterioration model parameters are
follows24 used as predictors of (SAcollapse) through application of
neural network
Et = L  My ð7Þ
  ð‘
p
In equation (7), L is the rate of cyclic deterioration and IA = ½a(t)2 dt ð8Þ
is evaluated according to calibration of experimental 2g
0
results. Considering deterioration model is a key ele-
ment in evaluation of collapse capacity of structures. IC = arms Td 1:51:5
ð9Þ
The modified Ibarra–Medina–Krawinkler deterioration ðd
T
model, applied in this example, is defined based on CAV = ja(t)jdt ð10Þ
backbone deformation parameters (i.e. up , upc ) and cyc-
0
lic deterioration parameter (i.e. L). Experimental cali-
bration of these parameters results in log-normal To obtain input data to evaluate response surface
probability distribution functions to present uncertain- and train ANN, five realizations for each modeling
ties due to up , upc , L. The statistical parameters of these random variables (up , upc , L) are considered, which cor-
modeling variables are shown in Table 2. responds to mean, mean minus and plus 1 standard
In this study, the structure is excited by a set of 40 deviation and mean minus and plus 2 standard devia-
strong ground motions, entitled as large magnitude tions (totally 5 3 5 3 5 simulations). For each

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

6 Proc IMechE Part O: J Risk and Reliability 0(0)

Table 2. Modeling parameters’ mean and dispersion and correlation calibration based on experimental results.3

Median up (rad) sup (rad) Median upc (rad) supc (rad) Median L sL rup, upc rup, L rupc, L

0.025 0.43 0.16 0.41 1.00 0.43 0.69 0.44 0.67

Figure 4. Sample IDA curves.

IDA: incremental dynamic analysis.

 
realization of input variables, IDA is implemented and upc
Ln(m) = C0 + C1 Ln(up ) + C2 Ln + C3 Ln(L)
SAcollapse is derived for each record. These values are up
considered as target data in proposed neural network. ð11Þ
Therefore, 40 3 125 vectors are applied to train and  
test the proposed network. upc
Ln(s) = C00 + C01 Ln(up ) + C02 Ln + C03 Ln(L)
Sample IDA curves are shown in Figure 4, and the up
architecture of proposed neural network is presented in ð12Þ
Figure 5.
Response functions, applied to estimate mean and X
N X
N
mLnðSAc Þ = C0 + Ci xi + Cij xi xj
standard deviation of collapse fragility curves, are
i=1 i\j
introduced in equations (11)–(14). Since the response
X
N
surfaces aim to predict mean and standard deviation of + Ci xi 2 = C0 + C1 up + C2 upc
collapse fragility curve given modeling parameters, 125 i=1
data points are available, corresponding to 125 cases + C3 L + C4 up upc + C5 up L + C6 upc L
for modeling variables (since for each case, IDA of the
structure produces one collapse fragility curve with + C7 up 2 + C8 upc 2 + C9 L2
known mean and standard deviation). Of these data ð13Þ
points, 65% are selected to interpolate assumed func-
tions since the same amount of data (81 3 40) are uti- X
N X
N
sLnðSAc Þ = C00 + C0i xi + C0ij xi xj
lized to train the ANN model, while the ANN model i=1 i\j
predicts SAcollapse, and then, collapse fragility curve is X
N
derived through fitting a log-normal probability + C0i xi 2 = C00 + C01 up + C02 upc
distribution. i=1
Evaluation of constant coefficients of equations + C03 L + C04 up upc + C05 up L + C06 upc L
(11)–(14) is implemented by nonlinear regression analy-
+ C07 up 2 + C08 upc 2 + C09 L2
sis. Estimated coefficients are listed in Tables 3 and 4
ð14Þ

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

Khojastehfar et al. 7

Figure 5. Proposed neural network.

SA: spectral acceleration.

Table 3. Constant coefficients of response surface functions for mean and standard deviation (equations (13) and (14)).

C0 C1 C2 C3 C4 C5 C6 C7 C8 C9 Mean square
error

mLn(SAc ) 20.179 234.647 3.638 0.892 228.36 3.916 0.472 845.984 25.278 20.277 0.0729
C00 C10 C20 C30 C40 C50 C60 C70 C80 C90 Mean square
error
sLn(SAc ) 20.505 72.941 20.536 0.104 14.088 21.282 0.106 21538.47 0.385 20.018 0.0417

Table 4. Constant coefficients of response surface functions for mean and standard deviation (equations (11) and (12)).

C0 C1 C2 C3 Mean square error

mLn(SAc ) 20.9 0.1559 0.3438 0.5149 0.0164

C00 C10 C20 C30 Mean square error
sLn(SAc ) 20.4579 0.1579 0.1 0.0993 0.0025

Mean and standard deviation values, evaluated by CV stands for calculated values based on IDA and N is
analytical equations (i.e. equations (11)–(14)) consider- the number of samples
ing five realizations of modeling parameters and values  
calculated based on IDA of structure, are compared in jEV  CVj
Error = 3 100 ð15Þ
Figures 6–9. In these figures, horizontal axis shows val- CV
ues that are evaluated by direct IDA and vertical axis vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N  
those which are estimated by regressed analytical func- u 1 X EVk  CVk 2
sError = t ð16Þ
tions. The solid black line shows the position where N  1 k=1 CVk
approximate values are equal to IDA-based values,
while the dots show the achieved data and dashed lines In order to achieve the best performances, different
show the bound which includes 68.7% of achieved data. ANN configurations with variable number of neurons
Estimation error (deviation of data from solid line) is in each layer are considered. Each of the ANN config-
presented by average of the ratio represented by equa- urations was trained and the performance error was
tion (15) and standard deviation represented by equa- evaluated by equation (4). Using such a trial-and-error
tion (16). In these equations, EV stands for estimated approach, the best ANN model corresponding to the
values based on neural network or response surfaces, least error measure is determined. As a result, a 6-45-

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

8 Proc IMechE Part O: J Risk and Reliability 0(0)

Figure 6. Comparison of calculated mean values based on IDA versus mean values based on equation (11).
IDA: incremental dynamic analysis.

Figure 7. Comparison of calculated standard deviation values based on IDA versus estimated standard deviation values based on
equation (12).
IDA: incremental dynamic analysis.

45-1 neural network containing two hidden layers with Collapse fragility curve of structure by Monte
45 neurons for each is finally selected. Besides, more Carlo simulation is achieved based on trained ANN
care was taken to assess the effect of different training and regressed response surfaces (Table 6). Ten thou-
algorithms (such as Levenberg–Marquardt, Broyden– sand values for input variables are simulated and
Fletcher–Goldfarb–Shanno (BFGS), Quasi-Newton SAcollapse is estimated by trained ANN for each simu-
and Scaled Conjugate Gradient). Again, by a trial-and- lation. Probability distribution fitted to values of
error approach, Levenberg–Marquardt training algo- SAcollapse presents collapse fragility curve. On the
rithm is chosen. other hand, applying response surfaces, mean and
Predicted SA corresponds to collapse limit state of standard deviation values of collapse fragility curve
sample structure versus IDA-based derived SAcollapse as are estimated for 10,000 simulations of modeling vari-
shown in Figure 10. The estimation error, applying ables. Collapse fragility curve of the frame according
trained ANN in prediction of SAcollapse values, is 8%. to proposed method is compared with other methods
Prediction of collapse fragility curve by fitting log- in Figure 11, and collapse fragility curve neglecting
normal distribution to SAcollapse values (which are eval- effects of modeling uncertainties (while modeling
uated through trained ANN) shows error of 0.5% and parameters are set as their mean values) is shown in
3.7% for mean and standard deviation, respectively, Figure 12.
which are considerably less than those of response The main limitation of response surface–based meth-
surface–based methods shown in Table 5. ods to involve epistemic uncertainties is application of

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

Khojastehfar et al. 9

Figure 8. Comparison of calculated mean values based on IDA versus estimated mean values based on equation (13).
IDA: incremental dynamic analysis.

Figure 9. Comparison of calculated standard deviation values based on IDA versus standard deviation values based on equation (14).
IDA: incremental dynamic analysis.

Figure 10. Comparison of calculated SAcollapse based on IDA and ANN.

ANN: artificial neural network; SA: spectral acceleration; IDA: incremental dynamic analysis.
Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014
10 Proc IMechE Part O: J Risk and Reliability 0(0)

Table 5. Errors of estimation of means and standard deviations of collapse fragility curves in various methods.

Predicted parameter Response surface–based Response surface–based ANN-based method (%)

method (quadratic response, method (logarithmic response,
equations (11) and (12)) (%) equations (13) and (14)) (%)

Mean values 27 50 0.5

Standard deviation values 6 7 3.7
SAcollapse – – 8

ANN: artificial neural network.

Table 6. Mean and standard deviation of collapse fragility curves applying Monte Carlo based on various methods (10,000
simulations of input random variables).

Error values Response surface–based Response surface–based ANN-based Neglecting epistemic

method (quadratic response, method (logarithmic response, method uncertainties
equations (11) and (12)) equations (13) and (14))

Mean values 0.4866 0.4377 0.4386 0.6292

Standard deviation values 0.4762 0.4480 0.5125 0.3894

ANN: artificial neural network.

because of the presence of high nonlinearity behavior

of structures. On the other hand, the proposed ANN
model is shown to be an efficient method to predict
SAcollapse through application of several hidden layers
as well as number of neurons in each layer given repre-
sentative characteristics of strong ground motions and
modeling parameters’ values. Then, the collapse fragi-
lity curve is derived through fitting a log-normal distri-
bution function to SAcollapse values.
The generality of the proposed ANN procedure is
not limited to the chosen limit state, and thereby, the
ANN method can be readily extended to cover other
limit states given the data provided for the specific limit
Figure 11. Collapse fragility curves, comparison of various state become available. For instance, when the life
methods. safety limit state is considered, the SA data points cor-
ANN: artificial neural network; SA: spectral acceleration.
responding to a drift related to life safety (instead of
collapse) should be collected and then used for training
the ANN.

Conclusion
In this article, ANN method is applied to predict col-
lapse fragility curve of structures while uncertainties
associated with modeling parameters are incorporated.
As a case study, a three-story moment-resisting steel
frame is considered, while parameters proposed by
modified Ibarra–Medina–Krawinkler moment rotation
model are considered as modeling parameters. Neural
Figure 12. Comparison of involvement and neglecting network training and validating are implemented by
epistemic uncertainty on collapse fragility curve. limited numbers of collapse fragility curves. These
SA: spectral acceleration. curves are derived through structural analyses using
IDA, while several realizations of values for modeling
parameters and strong ground motion parameters are
fixed format of functions to represent mean and stan- assumed. To implement Monte Carlo simulation, large
dard deviation values of fragility curves. This limitation number of realizations for modeling variables and
is pronounced when collapse limit state is investigated strong ground motion parameters are simulated, and

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

Khojastehfar et al. 11

then, SAcollapse values are evaluated by trained ANN. References

Fitting log-normal probability distribution to SAcollapse 1. Wyllie LA and Filson JR. Special supplement Armenia
values results in ANN-based collapse fragility curve. Earthquake reconnaissance report. Earthquake Spectra,
Applied method is compared with response surface 1989; 1–175.
technique. 2. Khurana A. Report on earthquake of 8th October in
Involving sources of epistemic uncertainty in devel- some parts of northern India, 2005, http://www.ndmin
opment of collapse fragility curves results in reduction dia.nic.in
of mean and increment of standard deviation values 3. Lignos D. Sideway collapse of deteriorating structural sys-
comparing with cases where such uncertainties are tems under seismic excitations. PhD Dissertation, Depart-
ment of Civil Engineering, Stanford University,
ignored. As shown in Figure 11, neglecting epistemic
Stanford, CA, 2008.
uncertainty causes underestimation of collapse fragility
4. Liel AB, Haselton CB, Deierlein GG, et al. Incorporat-
curves. Since the prediction error by ANN method is ing modeling uncertainties in the assessment of seismic
less than quadratic and logarithmic response surface– collapse risk of buildings. Struct Saf 2009; 31: 197–211.
based method, it can be concluded that prediction of 5. Zareian F, Krawinkler H, Ibarra L, et al. Basic concepts
mean and standard deviation by Monte Carlo simula- and performance measures in prediction of collapse of
tion based on ANN method would achieve more reli- buildings under earthquake ground motions. Struct Des
able outcomes. Obtained results show that ANN-based Tall Spec. Build. 2010; 19: 167–181.
method presents less reduction of mean values com- 6. International Conference of Building Officials. Uniform
pared with quadratic and logarithmic response surface Building Code (UBC). Whittier, California, USA, 1997.
methods, while more increment of standard deviation is 7. Federal Emergency Management Agency (FEMA). Rec-
ommended seismic design criteria for new steel moment
achieved. The superiority of ANN approach in predic-
frame buildings. Report no. FEMA-350, June 2000.
tion of collapse fragility curve of assumed structure, while Washington, DC: SAC Joint Venture.
considering aleatory and epistemic uncertainty sources, is 8. Federal Emergency Management Agency (FEMA).
shown through considerable reduction of estimation Seismic performance assessment of buildings, volume 1—
errors. Furthermore, the neural network is trained to pre- methodology. Report no. FEMA P-58-1, 2012, https://
dict SAcollapse albeit response surfaces aim to predict www.atcouncil.org/Projects/atc-58-project.html
mean and standard deviation values directly. Application 9. Kramer SL. Geotechnical earthquake engineering. New
of large amount of data given by IDA curves (5000 in Delhi, India: Pearson Education, 1996.
this example) allows us to train a relatively complicated 10. Medina R. Seismic demands for non-deteriorating frame
neural network in view of a number of hidden layers and structures and their dependence on ground motions. PhD
a number of neurons in each layer. This fact makes pre- Dissertation, Department of Civil Engineering, Stanford
University, Stanford, CA, 2002.
diction of collapse fragility curve with much fewer errors
11. Krawinkler H, Zareian F, and Lignos DG and Ibarra LF
while identical computational effort. Prediction of collapse of structures under earthquake
The procedure underlying the proposed ANN excitations. In: Proceedings of ECCOMAS thematic con-
method to capture the effects of epistemic uncertainties ference on computational methods in structural dynamics
for collapse limit state is not limited to the case study and earthquake engineering, (2009, June), Rhodes, Greece,
considered herein. The ANN method is a universal CD-ROM paper, paper no. CD449 (pp. 22–24).
approximator which is able to accurately predict the 12. Cardaliaguet P and Euvrand G. Approximation of a
target values provided that the ANN is well trained function and its derivatives with a neural network. Neural
using appropriate data points. This article demon- Networks 1992; 5: 207–220.
strates the efficiency of the ANN method against the 13. Chapman OJV and Crossland AD. Neural networks in
response surface method through a specific example probabilistic structural mechanics. In: Sundararajan C
(ed.) Probabilistic structural mechanics handbook: theory
(i.e. three-story, three-bay 2D frame). Similar proce-
and industrial application. New York: Chapman & Hall,
dure can be employed to train a proper ANN model 1995, pp.317–330.
for any other structure. It is evident that the ANN 14. Li X. Simultaneous approximations of multivariate func-
model developed for a specific structure is to be com- tions and derivatives by neural networks with one hidden
pared with the corresponding response surface model layer. Neurocomputing 1996; 12: 327–343.
produced for the same structure. 15. Lagaros ND and Fragiadakis M. Fragility assessment of
steel frames using neural networks. Earthq Spectra 2007;
23: 735–752.
Declaration of conflicting interests
16. Mitropoulou CC and Papadrakakis M. Developing fragi-
The authors declare that there is no conflict of interest. lity curves based on neural network IDA predictions. Eng
Struct 2011; 33(12): 3409–3421.
17. Vamvatsikos D and Cornell CA. Incremental dynamic
Funding
analysis. Earthq Eng Struct D 2002; 31(3): 491–514.
This research received no specific grant from any fund- 18. Zareian F and Krawinkler H. Assessment of probability
ing agency in the public, commercial or not-for-profit of collapse and design for collapse safety. Earthq Eng
sectors. Struct D 2007; 36(13): 1901–1944.

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

12 Proc IMechE Part O: J Risk and Reliability 0(0)

19. Benjamin JR and Cornell CA. Probability, statistics and frame and shear wall structural systems. Eng Struct 2011;
decision for civil engineers. New York: McGraw-Hill, 33: 1107–1116.
1997. 24. Ibarra LF, Medina RA and Krawinkler H. Hysteretic
20. Ellingwood B. Quantifying and communicating uncer- models that incorporate strength and stiffness deteriora-
tainty in seismic risk assessment. Struct Saf 2009; 31: tion. Earthq Eng Struct D 2005; 34(12): 1489–1511.
179–187. 25. Foutch DA and Yun SY. Modeling of steel moment
21. Rubinstein RY. Simulation and the Monte Carlo method. frames for seismic loads. J Constr Steel Res 2002; 58:
New York: John Wiley & Sons, 1981. 529–564.
22. Pinto P, Giannini R and Franchin P. Seismic reliability 26. Papadrakakis M, Papadopoulos V, Lagaros ND, Vulner-
analysis of structures. Pavia: IUSS Press, 2005. ability analysis of large concrete dams using the conti-
23. Shafei B, Zareian F and Lignos DG. A simplified method nuum strong discontinuity approach and neural
for collapse capacity assessment of moment-resisting networks. Struct Saf 2008; 30(3): 217–235.

Downloaded from pio.sagepub.com at Istanbul Universitesi on January 18, 2014

View publication stats