Structural Safety xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Structural Safety
journal homepage: www.elsevier.com/locate/strusafe

A general framework for the estimation of analytical fragility functions

based on multivariate probability distributions
Irmela Zentner ⇑,1
IMSIA, Institute for Mechanical Sciences and Industrial Applications, UMR 9219 CNRS-EDF-CEA-ENSTA ParisTech, France
EDF R&D Lab Saclay, France

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

Article history: A fragility curve is a function that expresses the probability of failure of a structure or component as a
Received 18 March 2015 function of the intensity of external aggression. This paper proposes a general framework for the devel-
Received in revised form 20 September opment of analytical fragility functions from data based on the copula approach. Such a model allows for
2016
any kinds of marginal distributions and dependence structures so that it can be applied to various types
Accepted 22 September 2016
Available online xxxx
of fragility data, analytical or empirical. The fragility function is then derived from the joint distribution of
intensity and damage measures. The Bayesian information criterion is used to select the most plausible
model among the candidate joint distributions, given the data. The practical implementation of the
Keywords:
Fragility curve
methodology is illustrated by an analytical test case and by the evaluation of seismic fragility curves
Copula for a reinforced concrete building. Several candidate marginal distributions, in agreement with the nature
Performance based earthquake engineering and the physical properties of the variables (e.g. common intensity and damage measures take only pos-
Conditional probability itive values) are evaluated. In particular, seismic intensity measures are lognormal random variables
Time history analysis according to seismological models. This paper is focused on bivariate distributions but the case of vector
valued intensity measures can be treated accordingly.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction nature and number of available data. Epistemic uncertainties are

paid particular attention to in Liel et al. [5] and Celik & Ellingwood
A fragility curve is a function that expresses the probability of [6]. In the nuclear sector, the seismic Probabilistic Risk Assessment
failure of a structure or component as a function of the intensity (PRA) dates back to the 70s and is now widely used for the evalu-
of external aggression. The intensity of the aggression is character- ation of plant safety [7].
ized by a ground motion Intensity measure (IM). The most popular Classical methods for the evaluation of fragility curves, when
IM for the evaluation of fragility functions are the Peak Ground input–output samples are available, are [8]:
Acceleration (PGA) and the Pseudo-Spectral Acceleration (PSA) at
a fundamental frequency of the structure. In the recent years, var-
Method of moments e.g. [9,3]. It can be used when a sample of
ious methods to determine fragility curves have been proposed in capacities has been observed: this is the case in incremental
the literature. The increased interest is, among others, motivated dynamic analysis [10] or when qualification tests until failure
by the ongoing implementation of performance based earthquake are available.
engineering (PBEE) procedures in civil engineering design. In the
Maximum likelihood method, for example [11,12]. This method
US, the Applied Technology Council (ATC) [1] implemented the requires demand data before and beyond failure.
PBEE methodology to be used for civil structures, in agreement
Fitting of a linear seismic demand model in the log-scale by
with the concepts developed at the Pacific Earthquake Engineering means of linear regression, e.g. [13,14]. This method is applica-
Research (PEER) Center [2]. A comprehensive set of procedures to ble for any kind of data but it means extrapolation of the behav-
estimate fragility functions from data is proposed in references ior beyond yield if no failure is observed.
[3,1,4]. In particular, the fragility analysis methods depend on the
Both in PBEE and seismic PRA, the seismic fragility curves are
modeled as lognormal cumulative distribution functions.
⇑ Address: IMSIA, Institute for Mechanical Sciences and Industrial Applications,
UMR 9219 CNRS-EDF-CEA-ENSTA ParisTech, France.
In a recent contribution [15], Gaussian kernel smoothing has
E-mail address: [email protected] been proposed to evaluate non parametric fragility functions. This
1
7, bvd Gaspard Monge, Palaiseau, France. allows to develop functional forms without any assumption on the

http://dx.doi.org/10.1016/j.strusafe.2016.09.003
0167-4730/Ó 2016 Elsevier Ltd. All rights reserved.

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
2 I. Zentner / Structural Safety xxx (2016) xxx–xxx

probability distribution. One botttleneck of this methodology is the When a lognormal model is adopted, then the fragility curve
dependence of the estimations on the choice of the bandwidth. The can be expressed as:
trade-off between robustness, over-fitting and accuracy is indeed a

lnða=Am Þ
major issue in the framework of Gaussian kernel smoothing. Gen- Pf ðaÞ ¼ U ; ð3Þ
b
erally speaking, a non parametric approach allows to overcome
model assumptions (such as lognormal probability distribution) where U denotes standard normal (Gaussian) cumulative distribu-
but at the expense of robustness. Parametric models posses the tion function (cdf) with standard deviation b and mean value ln
advantage of reduced numerical cost and improved robustness. If Am . The variable Am is the median of the lognormal distribution
the parametric model is adequate, then the precision is higher at and is also called median capacity.
the same cost (in terms of data). For example, it is well known that
extreme value distributions are good candidates for modeling peak
responses and common seismic motion intensity measures are log- 2.1. Maximum likelihood estimator
normal according to Ground Motion Prediction Equations (GMPE)
[16]. Any multivariate joint distribution can be written in terms The maximum likelihood method is of particular interest when
the damage (or failure) is not defined by a continuous but a dis-
of the marginal distribution functions and a copula which
crete state variable. In this framework, the dataset consists of N
describes the dependence structure between the variables. This
approach is pursued in this paper: the fragility function is derived pairs fai ; xi g. If the failure or damage criterion is reached, then
xi ¼ 1 and otherwise (no failure) it equals zero: xi ¼ 0. These events
according to its intrinsic definition as a conditional probability
from the joint distribution of the IM and the damage measure arrive with probability P f ðaÞ and respectively 1  Pf ðaÞ. For the log-
(DM), where the latter is defined by the marginal distributions normal model, the latter can be calculated by expression (3). The
and a copula. Copula models have been applied for reliability anal- likelihood function for this problem reads:
ysis and hazard assessment by various authors in the recent years Y
N
 xi  1xi
[17–24]. In a recent contribution, Goda [17] evaluates the joint dis- L¼ Pf ðai Þ 1  Pf ðai Þ ð4Þ
tribution of peak and residual displacement seismic demands of a i¼1

simple structure. Moreover, a comprehensive introduction to the The estimators of parameters Am and b are solution of the fol-
copula technique is provided by that author. Tang et al. [19,20] lowing optimization problem:
investigate the impact of different copula models when construct-
^ m ; bÞ
ðA ^ ¼ arg minð ln LÞ: ð5Þ
ing bivariate distributions for parallel systems and slope reliability Am ;b
analysis.
This paper proposes a new, general framework for the develop- As pointed out in [4], the maximum likelihood estimator is
ment of analytical fragility functions from data (continuous or dis- equivalent to generalized linear regression with a Probit link func-
crete, simulated or observed) based on the copula approach. In tion. Prior information on median structural capacity Am or stan-
what follows, we first give a short overview over common lognor- dard deviation b can be introduced by Bayesian updating, e.g.
mal fragility function fitting with a focus on analytical approaches [25,26]. This is the case, if either an expert opinion or previous
that do not require the scaling of accelerograms. Secondly, the cop- analysis are available. The profiled likelihood allows to evaluate
ula based fragility function estimation is presented. The approach confidence intervals [27,28]. However, when the damage measure
is illustrated by an analytical test case and the application to seis- is a continuous design variable, then information is lost by the
mic analysis of a reinforced concrete building. For this purpose, a transformation to the discrete damage indicator. Indeed, only the
Gaussian copula model is chosen. information on damage state membership is used while the dis-
tance to the threshold is disregarded. Moreover, this methodology
can not be applied when only few or no failures (or occurrences of
2. Computation of analytical lognormal fragility curves the damage state) are observed.

A fragility curve is a function that expresses the probability that

a structure or component submitted to an excitation of intensity a, 2.2. Linear regression
fails or reaches a previously defined damage state. As it is common
use, in what follows, random variables will be denoted by capital The linear regression approach [13,14] assumes a continuous
letters while their realization are lower case. More precisely, if damage measure of the form
damage or failure is assumed to occur when the design variable D ¼ bac g ð6Þ
or damage measure D exceeds a threshold d0 (which can be itself
where g is a lognormal random variable with unit median and log-
a random variable), then the fragility curve is given by the follow-
standard deviation equal to b. The linear regression is performed in
ing conditional probability:
log space:
Pf ðaÞ ¼ PðD > d0 j IM ¼ aÞ: ð1Þ ln D ¼ ln b þ c ln a þ ; ð7Þ

When discrete damage states are considered, then the damage where  is a centered normal random variable with standard
measure boils down to a Bernoulli random variable X, taking the deviation r . The regression parameters ln b; c and r read:
value x ¼ 1 when the damage state is reached and otherwise c ¼ r rln D =rln IM and ln b ¼ lln D  r rln D =rln IM lln IM , where lln D ; rln D
x ¼ 0. This is the case, for example, when failure is defined as the and lln IM ; rln IM are the mean and standard deviations of variables
buckling of a structural element or for damage data from ln D and ln IM and r is the linear correlation coefficient. They have
post-earthquake field observations, where only the membership to be approximated by their statistical estimator. Denoting by ei
to a certain damage state (e.g. ‘‘no damage”, ‘‘slight damage”, . . ., the regression residual, an unbiased estimation of r2 is given by
‘‘failure”) can be assigned. In the latter framework, the fragility the formula [29]:
function formally reads:
1 X 2
s2 ¼ e :
Pf ðaÞ ¼ Pðx ¼ 1 j IM ¼ aÞ: ð2Þ n2 n i

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
 I. Zentner / Structural Safety xxx (2016) xxx–xxx 3

By virtue of expression (7), the conditional distribution for the of a multivariate normal distribution. The rank correlation assesses
damage measure ln D reads: how well the relationship between two variables can be described
  by a monotonic (and possibly non linear) function. The Spearman’s
ðln D j ln IM ¼ ln aÞ  N ln b þ c ln a; r2e : ð8Þ rank correlation coefficient denoted by q is used here. It is evalu-
In consequence, the linear regression approach yields a lognor- ated simply as the linear correlation coefficient of the rank trans-
mal fragility curve: formed data, where the rank transformation consist in replacing

 the data values by their ranks. The linear correlation coefficient
lnðbac =D0 Þ
Pf ðaÞ ¼ U : ð9Þ is the ratio between the covariance and the product of the
re standard-deviations of the two variable: r ¼ COVðX 1 ;X 2 Þ
rX 1 rX 2 .
c
Substituting D0 ¼ bAm in expression (9), one obtains the classi- The second interesting property of copula models is the possi-
cal expression of the lognormal fragility curve given by Eq. (3) with bility to determine the marginal distributions independently from
median Am ¼ exp lnðDc0 =bÞ and log-standard deviation b ¼ re =c. This the model for dependency. For a more detailed introduction to cop-
approach proves to be quite robust and can be applied to any kind ula models, the reader is referred to the literature and the material
of continuous damage measure or demand parameter. However, it presented in Refs. [17,24].
might not be well adapted to any kind of data. This is further dis- For the computation of fragility functions, we consider the two
cussed in the following sections. random variables (IM, D) with marginal distributions
fF 1 ðaÞ; F 2 ðdÞg. The copula approach then allows to construct the
joint distribution of the two variables and to evaluate the fragility
3. Development of generalized multivariate fragility functions
curve by Eq. (11).
For such a bivariate distribution, one has to first find two ade-
We consider the two random variables ðIM; DÞ defined respec-
quate marginal distributions and a copula and then determine the
tively as the ground motion intensity and the damage measures.
model parameters. In practice, the choice of marginal distribu-
The idea of the approach presented in this section is to construct
tions for intensity and damage measures will be guided by the
their joint distribution f ðd; aÞ and to evaluate the conditional prob-
nature of the variables considered. For example, the IM are gen-
ability distribution by means of the Bayes theorem and integration.
erally positive. Moreover, seismological models exist for most
Indeed, the conditional probability of d knowing a reads:
intensity measures, in particular Peak Ground Acceleration
f ðd; aÞ (PGA) and Pseudo-Spectral Accelerations (PSA) are modeled by
f ðd j IM ¼ aÞ ¼ ; ð10Þ
f ðaÞ lognormal random variables. In consequence, when the ground
motion time histories used for the calculation of the structural
so that the fragility function of Eq. (1) can be evaluated for any
response are selected or simulated to be in agreement with the
given damage level threshold D0 as:
target scenario, then the IM is log-normally distributed. This is
Z 1
1 an important information that should be introduced in the fragi-
Pf ðaÞ ¼ Pðd > D0 j IM ¼ aÞ ¼ f ðd; aÞ dd: ð11Þ
f ðaÞ D0
lity function. For discrete damage states or structural failure,
binomial or multinomial distributions are adequate probability
The joint distribution of a multivariate random variable can be distributions. The model choices should be corroborated by statis-
described by the marginal distributions and a copula. Such a model tical tests. In this paper, the best bivariate model is selected by
is very flexible since any known probability distribution can be a means of Bayesian and Akaike’s information criteria as described
candidate for the marginals. This is described in what follows with in Section 3.2.
more detail for bivariate random variables. The proposed approach can be easily extended to the case of
vector valued intensity measures and multivariate random vari-
3.1. Construction of fragility functions from marginal densities and ables. This allows to address the assessment of vulnerabilities with
copula respect to multi-hazards, such as flooding, tsunami or hurricane,
e.g. [31,32], in a very flexible way.
Copula allow for a representation of dependence between mul-
tivariate data. The joint distribution of any bivariate random vari- 3.2. Model selection
able fX 1 ; X 2 g can be described by the two univariate marginal
cumulative distributions fF 1 ðx1 Þ; F 2 ðx2 Þg and a copula. Sklar’s theo- In statistics, Akaike’s Information Criterion (AIC) and the Baye-
rem [30] states that the joint cumulative probability distribution sian Information Criterion are commonly used for model selection.
Hðx1 ; xk Þ ¼ PðX 1 < x1 ; X 2 6 x2 Þ of the two variables fx1 ; x2 g can be The Bayesian Information Criterion (BIC) [29,33] expresses the
expressed by a copula function as plausibility of observed data given a model M with parameters
Hðx1 ; x2 Þ ¼ CðF 1 ðx1 Þ; F 2 ðx2 ÞÞ ¼ Cðu1 ; u2 Þ ð12Þ hM . It depends on the likelihood LM , the number of model param-
eters np and the sample size N s :
where ui ¼ F i ðxi Þ are standard uniform random variables. The sim-
plest copula is the Gaussian copula which depends only on the cor- BIC ¼ 2 ln LM ðhM Þ þ lnðNs Þnp : ð14Þ
relation coefficient q. It reads:
The BIC gives a preference to the simpler model, featuring less

Cq ðu1 ; u2 Þ ¼ Uq
N 1 1
U ðu1 Þ; U ðu2 Þ : ð13Þ parameters in order to avoid over-fitting. The most plausible model
is the one with the smallest BIC score. The AIC has a very similar
Other common copula models are the Clayton copula, the Gum- expression:
bel copula, the Frank copula and the extreme value copula. One AIC ¼ 2 ln LM ðhM Þ þ 2np ; ð15Þ
interesting property is that the copula C is invariant under strictly
monotonic transformation of the random variables. This is also the but it does not introduce the sample size and gives less importance
case for Spearman’s and the Kendall’s rank correlation coefficients to the penalty term.
which are used as an intrinsic measure for dependence in conjunc- Given the sample of N s damage measures x ¼ x1 ; . . . ; xn , associ-
tion with copulas [24]. The Pearson’s linear correlation coefficient ated to N s intensity measures a1 ; . . . ; an , the likelihood function
is almost identical to Spearman’s rank correlation only in the case appearing in expression (14) is evaluated as:

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
4 I. Zentner / Structural Safety xxx (2016) xxx–xxx

Y Y 1
LC ðhC Þ  LC ðx j a; hC Þ ¼ f ðxi j ai ; hC Þ ¼ f hC ðxi ; ai Þf ðai Þ ð16Þ time history analysis, then this is not an issue because of the
i i proper definition of the fragility curve: it expresses a conditional
where hC are the parameters of the two marginal distributions and probability, i.e. the failure probability given a particular value of
the correlation model described by the copula C. For simplicity’s ground motion IM. In consequence, the epistemic uncertainty on
sake, the dependence of the marginal distribution f ðai Þ on the the IM is not counted twice. Another interesting feature of the
parameters is not explicitly mentioned in Eq. (16). Expression (16) numerical approach presented in this paper is the possibility to
shows that the evaluation of the BIC involves a weighting by f ðai Þ, directly introduce the known distribution of the IM. This is partic-
which is an interesting property since it gives more weight to the ularly useful when the ground motion time histories used for the
more rare events. structural analysis have been selected or simulated to be in agree-
ment with a given scenario (e.g. [34,35]). In the latter case, the
3.3. Lognormal bivariate distribution (lognormal) distribution of the IM is completely determined by
the hazard curves. In the other cases, it has to be inferred from
We now illustrate the computation of fragility curves for the data. The approach is not dependent on the choice of the particular
special case where both the intensity and the damage measures distribution of the IM, as highlighted by Eq. (10). This results from
are lognormal with Gaussian copula. Then the problem boils down the mathematical definition as a conditional probability, given the
to the estimation of the parameters of the bivariate log-normal dis- IM. The fragility curves are however site-specific as they do depend
tribution or, more conveniently in the log-space, the bivariate nor- on the particular spectral shape and other features linked to the
mal (Gaussian) distribution. The bivariate distribution of scenario (such as duration and other IM) through the set of
Z ¼ fZ 1 ; Z 2 g ¼ fln IM; ln Dg is characterized by 5 parameters: accelerograms used for the structural analysis.
lZ1 ; lZ2 ; rZ1 ; rZ2 ; q, where q is the correlation coefficient. Classi-
cal statistics can be used to estimate these parameters. For the cor-
4. Application
relation, Spearman’s rank correlation coefficient is used. More
precisely, the bivariate distribution of Z reads:

 In the following applications, the adequateness of lognormal,
1 1 Weibull and Generalized Extreme Value (GEV) distribution to rep-
f ðz1 ; z2 Þ ¼ exp  ðz  lÞT R1 ðz  lÞ ; ð17Þ
2pz1 z2 det ðRÞ1=2 2 resent seismic damage measures is analyzed. The IM is a lognormal
random variable and a Gaussian copula is assumed.
where The Generalized Extreme Value distribution [27,36] has 3
" # !
r2Z1 q rZ1 rZ2 lZ1 parameter, namely shape, scale and location parameter. The shape
R¼ ; l¼ parameter determines the type of GEV (type I, II, III), where type I is
q rZ1 rZ2 r2Z2 lZ2
the Gumbel distribution, type II is the Fréchet distribution and type
Then, the conditional distribution of Z 2 given Z 1 ¼ z1 yields III is the reversed Weibull distribution (distribution of minima). By

 the extreme value theorem, the GEV distribution is the only possi-
rZ 2
ðZ 2 j z1 Þ  N lZ2 þ qðz  lZ1 Þ; ð1  q2 Þr2Z2 : ð18Þ ble limit distribution of properly normalized maxima of a sequence
rZ 1 1
of independent and identically distributed random variables. The
Eventually, the fragility function, expressed with respect to the lognormal distribution has 2 parameters, median and log std. The
variables IM; D, reads Weibull distribution has 2 parameters, shape and scale. The corre-

 lation model, that is the copula, has one parameter, namely the
mðaÞ  ln DS
Pf ðaÞ ¼ Pðd > DS j IM ¼ aÞ ¼ U ; ð19Þ rank correlation coefficient q. Thus, the number of model parame-
b ters is 6 if the damage measure follows a GEV distribution while
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where mðaÞ ¼ lln D þ rrln D qðln a  lln IM Þ and b ¼ ð1  q2 Þrln D . only 5 model parameters are needed in the case of lognormal or
ln IM

It is interesting to acknowledge that the bivariate log-normal Weibull distributions. The three extreme value distributions have
distribution leads to the same fragility function as the linear regres- indeed different ranges of definition: Gumbel is unlimited, Fréchet
sion approach of Section 2.2, but with a slightly different expression has a lower limit and the reversed Weibull has an upper limit. In
for standard-deviation. Indeed, Eq. (8) can be rewritten as consequence, the Fréchet (type II) GEV distribution is expected to

 be the most adequate to represent damage measures.
rln D
ðln D j ln IM ¼ ln aÞ  N lln D þ r ðln a  lln IM Þ; r2 ð20Þ The statistical analysis are performed with Matlab. The model
rln IM parameters that have to be estimated for the bivariate distribu-
which can be compared to expression (18). In particular, if the data tions are the rank correlation coefficient and the parameters of
is actually lognormal, then the linear correlation coefficient r eval- the marginal distributions. The latter are estimated by the maxi-
uated from the log-transformed sample is almost equal to the rank mum likelihood estimators available with the Statistics toolbox
correlation q used in the bivariate model. of Matlab. For the regression analysis, the polyfit function was used.

3.4. Uncertainty and distribution of the intensity measure

4.1. Analytical test case
The uncertainty related to the intensity measures is a particular
issue when empirical fragility data is used. Indeed, the uncertainty In order to illustrate the methodology and to evaluate the
related to the proper evaluation of the ground motion parameters, effects of the different distributions, we start with an analytical
such as magnitude and distance, can be significant. When perform- test case. Two cases are examined:
ing seismic risk assessment, then the annual frequency of failure is
evaluated by convolution of the seismic hazard curve with the fra- 1. The damage measure follows a GEV distribution of Fréchet type
gility curve (i.e. by calculating the total probability). Thus, when with shape, scale and location parameters ð1:0; 0:0024; 0:0024Þ,
the epistemic uncertainty related to the IM is accounted for in respectively.
the hazard curve, then it should not be included in the fragility 2. The damage measure is lognormal with median and logarithmic
since this would mean ‘‘double counting” that uncertainty. When standard deviation ð0:006; 0:3Þ, respectively.
fragility curves are evaluated by numerical simulation based on

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
 I. Zentner / Structural Safety xxx (2016) xxx–xxx 5

FRAGILITY CURVE FRAGILITY CURVE

1 1

0.9 0.9

0.8
0.8
0.7
0.7
0.6 Lognormal Lognormal
0.6
Pf (IM)

Weibull Weibull

Pf (IM)
0.5
GEV 0.5 GEV
0.4 Regression Regression
0.4
0.3
0.3
0.2
0.2
0.1

0 0.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0
IM 0 0.1 0.2 0.3 0.4 0.5 0.6

Fig. 1. Fragility curves evaluated for GEV distributed DM and lognormal IM. The IM
dash-dotted curve is the fragility curve obtained by linear regression in log space.
Fig. 2. Fragility curves evaluated for lognormal DM and lognormal IM. The dash-
Table 1 dotted curve is the fragility curve obtained by linear regression in log space.
BIC for the 3 models with q ¼ 0:95, the smallest score is in bold font.
is a reduced scaled model (1/4th scale) representative of a typical,
Case 1 Case 2
simplified half part of an electrical nuclear building. This structure
BIC AIC BIC AIC has been the object of numerous experimental and numerical stud-
Lognormal 441.9 425.4 314.3 297.8 ies in the framework of the SMART2008 [37] and more recently the
Weibull 673.8 657.3 393.9 377.4 SMART2013 benchmark. It was designed, built and tested in order
GEV II (Frechet) 323.1 303.3 321.2 301.4 to improve knowledge on the mechanical behavior of complex 3D
structures exhibiting coupled torsional and flexural response under
In both cases, the IM is assumed to be lognormal with median seismic loading. The mock up is shown in Fig. 3. The finite element
0.25, logarithmic standard deviation (log std) 0.6 and a rank corre- model used in this case study contains shell and beam elements.
lation coefficient q ¼ 0:95. The failure threshold was taken as The model constructed for the 2013 benchmark includes the
D0 ¼ 0:006. For the two test cases, we simulate a sample of shaking table since interaction was found to be important. For
N = 200 pairs of intensity measure and damage measure (IM,
DM). It has been verified, by repeating the simulations, that this
sample size allows for a robust estimation of the parameters and
stable results. Fig. 2 shows the fragility curves obtained for test
case one, together with the linear regression fragility curve. The
solid line represent the ‘‘correct” fragility curve, obtained here with
the bivariate model with GEV distributed DM. As expected, the lin-
ear regression in log-space (dash red) and the bivariate lognormal
model (dash blue) lead to very similar fragility curves. In particular,
the median values are the same, but the dispersion (given by the b-
value) is slightly different. As explained in Section 3.3, this differ-
ence is due to slightly different formulas for the estimators of beta
in linear regression and the bivariate lognormal model. Fig. 1
shows the fragility curves obtained for test case two, together with
the linear regression fragility curve. Again, the solid line represent
the ‘‘correct” fragility curve, obtained here with the bivariate log-
normal model. All of the fragility curves are quite close, except
for the Weibull distribution. It is now interesting to determine
the BIC associated to the different models and to check whether
the ‘‘correct” model, given the data, can be detected.
The BIC and AIC values are shown in Table 1 for the two test
cases. The GEV is correctly detected as the most plausible DM dis-
tribution for test case 1 and the bivariate lognormal fragility curve
is detected as the most plausible model for test case 2. For test case
2, also the fitted GEV type II model (dash magenta) performs well,
since it is close to the bivariate lognormal fragility curve, as can be
seen in Fig. 2.

4.2. SMART 2013 benchmark case study

In this section, the bivariate fragility model is applied to assess

the seismic fragility of a reinforced concrete building. The structure Fig. 3. SMART mock up.

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
6 I. Zentner / Structural Safety xxx (2016) xxx–xxx

the computation of the fragility curves, the structure is supported

by equivalent springs. The properties of these springs were pro-
vided by the benchmark specifications. More information can be
found on the dedicated website: http://smart2013.eu. The numer-
ical transient analyses are performed with the opensource multi-
purpose Finite Element (FE) software Code_Aster (http://
www.code-aster.org). Nonlinear structural behavior is accounted
for by the homogenized nonlinear constitutive model for rein-
forced concrete GLRC_DM available with Code_Aster and Rayleigh
damping is applied. Uncertainty of the structural model parame-
ters is accounted for. The damping and stiffness of the supporting
springs and the structural damping ratio are modeled by lognormal
random variables with median equal to the best-estimate value.
The coefficients of variation are 1% and 2% for the spring stiffness
and damping, respectively, and 33% for the equivalent structural
damping ratio. The spring stiffness and damping values are corre-
lated with correlation coefficient 0.8. These choices were made
according to the benchmark specifications. Latin Hypercube sam-
pling is used to generate sets of 50 model parameters [38]. This
results in a mean value of 5.74 Hz for the first eigenfrequency of Fig. 5. Response spectra of the 50 accelerograms in Y-direction, together with
the undamaged structure (determined by modal analysis). The median and 1r values. The target spectra are in red. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of
damage measure considered here is the peak InterStory Drift
this article.)
(ISD). The ISD ratio threshold for a medium damage level, as spec-
ified for the benchmark, is ld =200, where ld ¼ 1:2 m is the inter- FRAGILITY CURVE
storey height. This corresponds to slight damage of concrete 1
walls and is associated to immediate occupancy level (S-1) in 0.9
FEMA356 [39]. The structure is submitted to 50 pairs of horizontal
(XY directions) ground motion time histories that were provided to 0.8
the benchmark participants. The ground motion time histories are
0.7
artificial accelerograms obtained with a stochastic ground motion
Lognormal
Pf (PSAgm)

model. They were simulated to match the target spectrum in med- 0.6
ian and 1r values [40]. The target spectra (red curves) are shown Weibull
0.5
on Figs. 4 and 5 together with the respective values of the simu- GEV
lated accelerograms (magenta curves) in x-direction (Fig. 4) and 0.4 Regression
y-direction (Fig. 5). The target spectrum is a design spectrum asso-
0.3
ciated to a magnitude M = 6.5 and distance D = 9 km event for a
rock site (French regulation RFS-2001-01). The geometrical mean 0.2
of the two horizontal pseudo-spectral acceleration values at the
0.1
first fundamental frequency of the structure, PSAgm, and the hori-
zontal geometrical mean of peak ground acceleration, PGAgm, are 0
the IM considered here. Both IM are assumed to have a lognormal 0 0.5 1 1.5 2 2.5 3 3.5 4
PSAgm (g)
Fig. 6. Fragility curves for PSAgm considering different marginal distributions for the
damage measure; regression curve in red. (For interpretation of the references to
color in this figure legend, the reader is referred to the web version of this article.)

marginal distributions, as predicted by the GMPE [16]. Again, 3 dif-

ferent candidate marginal distributions are considered for the
damage measure: lognormal, Weibull and GEV distributions. The
dependence is modeled by a Gaussian copula with correlation coef-
ficient q.
Figs. 7 and 6 show the fragility curves obtained for the two IM
and considering different distributions for the peak ISD. The rank
correlation coefficients between intensity an damage measures
are q ¼ 0:96 for PSAgm and q ¼ 0:95 for PGAgm. The regression
analysis yields a median capacity Am ¼ 1:27g and log-std
b ¼ 0:24 (s ¼ 0:32) when the intensity measure is PSAgm and
Am ¼ 0:50; b ¼ 0:23 (s ¼ 0:34) for PGAgm. For both IM, the regres-
sion curves are, as expected, very close to the bivariate lognormal
fragility function. Differences are due to the slightly different
expressions for the estimators as discussed in Section 3.3.
The results of the linear regression approach are illustrated in
Fig. 8 for PSAgm and in Fig. 9 for PGAgm. The figures show the
Fig. 4. Response spectra of the 50 accelerograms in X-direction, together with
median and 1r values. The target spectra are in red. (For interpretation of the
expected value of the DM as a function of the IM (according to
references to color in this figure legend, the reader is referred to the web version of expression (6)) together with the considered damage threshold
this article.) (blue) and the scatterplots of the data. The BIC and AIC values

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
 I. Zentner / Structural Safety xxx (2016) xxx–xxx 7

FRAGILITY CURVE Table 2

1 BIC for 3 different distributions for ISD (lognormal, Weibull and GEV) for the two IM
PSAgm and PGAgm .
0.9
PSAgm PGAgm
0.8
BIC AIC BIC AIC
0.7 LN 62.70 53.1 69.79 60.2
Weib 97.60 88.0 92.65 83.1
Pf (PGAgm)

0.6
GEV II 58.76 47.3 69.31 57.8
0.5 Lognormal
Weibull 0.12
0.4
GEV
data
0.3 Regression
0.1 simulated LN
0.2 simulated Weib
simulated GEV

peak ISD (m)

0.1 0.08

0
0 0.5 1 1.5
PGAgm (g) 0.06

Fig. 7. Fragility curves for PGAgm considering different marginal distributions for
damage measure; regression curve in red. (For interpretation of the references to 0.04
color in this figure legend, the reader is referred to the web version of this article.)

0.02
0.06
Data
0
0.05
Regression curve 0 2 4 6 8 10
Damage threshold PSAgm (g)
0.04 Fig. 10. Scatterplot for data (red) and samples of simulated couples ðPSAgm ; peakISDÞ
peak ISD (m)

using the 3 models: Lognormal (LN), Weibull (Weib) and GEV. (For interpretation of
the references to color in this figure legend, the reader is referred to the web version
0.03 of this article.)

0.12
0.02

0.1 data
0.01
simulated LN
simulated Weib
0.08
simulated GEV
peak ISD (m)

0
0 1 2 3 4 5 6 7
PSAgm (g)
0.06
Fig. 8. Scatterplot, damage threshold and regression curve in physical space for
PSAgm .
0.04

0.06
0.02
Data
0.05 Regression curve
damage threshold 0
0 0.5 1 1.5 2 2.5 3
0.04 PGAgm (g)
peak ISD (m)

Fig. 11. Scatterplot for data (red) and samples of simulated couples
0.03 ðPGAgm ; peakISDÞ using the 3 models: Lognormal (LN), Weibull (Weib) and GEV.
(For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)
0.02
are shown in Table 2. For both IM, the most plausible model is the
one with (type II) GEV distributed peak ISD. The exact fragility
0.01
curves are not known here. The solid lines in Figs. 7 and 6 represent
the fragility models detected as the most plausible by the BIC and
0 AIC. The bivariate probability model allows to generate new real-
0 0.5 1 1.5 2 izations of input–output samples. For illustration, Fig. 10 and
PGAgm (g)
Fig. 11 show scatterplots of the initial data (red) together with a
Fig. 9. Scatterplot, damage threshold and regression curve in physical space for sample of simulated couples (IM, DM), using the 3 different distri-
PGAgm . butions for peak ISD and lognormal IM.

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003
8 I. Zentner / Structural Safety xxx (2016) xxx–xxx

5. Summary and conclusion [7] Wakefield D, Ravindra M, Merz K, Hardy G. Seismic probabilistic risk
assessment implementation guide, Final report 1002989. Palo Alto: EPRI;
2003.
In this paper, a general framework for the construction of mul- [8] Zentner I, Gündel M, Bonfils N. Fragility analysis methods: Review of existing
tivariate fragility functions is proposed. The methodology is con- approaches and application. Nucl Eng Des 2016 (submitted for publication).
[9] Reed J, Kennedy R. Methodology for developing seismic fragilities, Final report
ceptually simple while allowing for various kinds of marginal
TR-103959. Palo Alto: EPRI; 1994.
distributions and dependence structures via the copula approach. [10] Vamvatsikos D, Cornell A. Incremental dynamic analysis. Earthquake Eng
In particular, it allows to develop fragility functions for both con- Struct Dyn 2002;31:491–514.
[11] Shinozuka M, Feng Q, Lee J, Naganuma T. Statistical analysis of fragility curves.
tinuous damage measures or discrete damage states. Being a para-
ASCE J Eng Mech 2000;126:1224–31.
metric model, it is more robust than purely empirical fragility [12] Zentner I. Numerical computation of fragility curves for NPP equipment. Nucl
functions. The lognormal distribution has proven to be a good Eng Des 2010;240:1614–21.
choice for seismic fragility analysis in several studies. Nevertheless, [13] Ellingwood BR, Kinali K. Quantifying and communicating uncertainty in
seismic risk assessment. Struct Saf 2009;31:179–87.
a methodology that provides tools to evaluate the adequateness of [14] Zentner I, Humbert N, Ravet S, Viallet E. Numerical methods for seismic
this choice and possibly corroborate the lognormal fragility curve fragility analysis of structures and components in nuclear industry -
is useful for the engineer, The proposed methodology allows to test application to a reactor coolant system. Georisk 2011;5:99–109.
[15] Noh HY, Lallemant D, Kiremidjian AS. Development of empirical and analytical
different distributions for IM and DM without considerable supple- fragility functions using kernel smoothing methods. Earthquake Eng Struct
mentary effort Interestingly, the choice of a bivariate lognormal Dyn 2015;44(8):1163–80.
distribution with Gaussian correlation model leads to the same [16] Campbell K, Bozorgnia Y. NGA ground motion model for the geometric mean
horizontal component of PGA, PGV, PGD and 5% damped linear elastic
result as the standard linear regression approach in log space, pro- response spectra for periods ranging from 0.01s to 10s. Earthquake Spectra
posed in the literature. The choice of the marginal distributions 2008;24:139–71.
depends on the nature and the physical properties of the variables [17] Goda K. Statistical modeling of joint probability distribution using copula:
Application to peak and permanent displacement seismic demands. Struct Saf
(e.g. common intensity and damage measures take only positive
2010;32:112–23.
values). In earthquake engineering, seismological models assume [18] Goda K, Tesfamariam S. Multi-variate seismic demand modelling using
lognormal distributions for ground motion intensity measures. copulas: Application to non-ductile reinforced concrete frame in victoria,
canada. Struct Saf 2015;56:39–51.
Such information is useful for the choice of the candidate marginal
[19] Tang X-S, Li D-Q, Zhou C-B, Phoon K-K, Zhang L-M. Impact of copulas for
distributions. Classical information criteria (such as BIC and AIC) modeling bivariate distributions on system reliability. Struct Saf
are then used for the selection of the most plausible model among 2013;44:80–90.
the candidates, given the data. [20] Tang X-S, Li D-Q, Zhou C-B, Phoon K-K. Copula-based approaches for
evaluating slope reliability under incomplete probability information. Struct
The approach is applied to the seismic fragility analysis of a Saf 2015;52(Part A):90–9.
reinforced concrete building where the damage is measured by [21] Michele CD, Salvadori G, Passoni G, Vezzoli R. A multivariate model of sea
the peak interstory drift. The copula model is constructed for log- storms using copulas. Coast Eng 2007;54:734–51.
[22] Masina M, Lamberti A, Archetti R. Coastal flooding: A copula based approach
normal intensity measures (geometrical mean of PGA and PSA) for estimating the joint probability of water levels and waves. Coast Eng
and different candidate distributions for the damage measure. 2015;97:37–52.
For the SMART2013 fragility analysis, the Fréchet type GEV distri- [23] Huard D, Evin G, Favre A-C. Bayesian copula selection. Comput Stat Data Anal
2006;51:809–22.
bution is the most appropriate marginal distribution for the dam- [24] Genest C, Favre A-C. Everything you always wanted to know about copula
age measure. modeling but were afraid to ask. ASCE J Hydrolog Eng 2007;12:347–68.
This paper is focused on bivariate distributions but the case of [25] Yamaguchi A. Usage of information entropy in updating seismic fragilities. In:
Proceedings of the OECD specialist meeting on the probabilistic safety
vector valued IM can be treated accordingly. In particular, the
assessment of nuclear facilities. Korea: OECD; 2006.
approach could be easily extended to multi-hazards vulnerability [26] Koutsourelakis P. Assessing structural vulnerability against earthquakes using
assessment, implying different kinds of external aggression. multi-dimensional fragility surfaces: A bayesian framework. Probab Eng Mech
2010;25:49–60.
Besides, only Gaussian copula models have been explored in the
[27] Coles S. An introduction to statistical modeling of extreme values, Series in
applications. The construction of fragility surfaces and the analysis statistics. New York: Springer; 2001.
of other copula models will be investigated in further studies. [28] Zentner I, Poirion F. Enrichment of seismic ground motion databases using
Karhunen-Loéve expansion. Earthquake Eng Struct Dyn 2012;41:1945–57.
[29] Saporta G. Probabilités, Analyse de Données et Statistiques. 2nd
Acknowledgements ed. Paris: TECHNIP; 2006.
[30] Sklar M. Fonctions de repartition a n dimensions et leurs marges. Publ Inst Stat
Univ Paris 1959;8:229–31.
This research has obtained financial support from the French [31] Wang Z, Padgett JE, Duenas-Osorio L. Risk-consistent calibration of load factors
Research Foundation ANR via the national research project for the design of reinforced concrete bridges under the combined effects of
earthquake and scour hazards. Eng Struct 2014;79:86–95.
SINAPS@ (Earthquake and Nuclear Installations: Ensuring and Sus-
[32] Lee KH, Rosowsky DV. Fragility analysis of woodframe buildings considering
taining Safety), which is gratefully acknowledged. The author also combined snow and earthquake loading. Struct Saf 2006;28:289–303.
wants to thank Nicolas Bonfils (from Electricité de France, Nuclear [33] Schwarz G. Estimating the dimension of a model. Ann Stat 1978;6:461–4.
Engineering Division) for providing the nonlinear structural model. [34] Jayaram N, Lin T, Baker JW. A computationally efficient ground-motion
selection algorithm for matching a target response spectrum mean and
variance. Earthquake Spectra 2011;27:797–815.
[35] Zentner I. A procedure for simulating synthetic accelerograms compatible with
References
correlated and conditional probabilistic response spectra. Soil Dyn Earthquake
Eng 2014;63:226–33.
[1] Seismic performance assessment of buildings, https://www.atcouncil.org/ [36] Leadbetter M, Lindgren G, Rootzén H. Extremes and related properties of
Projects/atc-58-project.html; 2012 random sequences and processes. Springer-Verlag; 1983.
[2] Porter K. An overview of PEERs performance-based earthquake engineering [37] Richard B, Martinelli P, Voldoire F, Chaudat T, Abouri S, Bonfils N. SMART 2008:
methodology. In: Civil Engineering Risk and Reliability Association (CERRA), Overview, synthesis and lessons learned from the international benchmark.
ICASP9, San Francisco, CA. Eng Struct 2016;106:166–78.
[3] Porter K, Kennedy R, Bachman R. Creating fragility functions for performance [38] Helton J, Davis F. Latin hypercube sampling and the propagation of uncertainty
based earthquake engineering. Earthquake Spectra 2007;23:471–89. in analyses of complex systems. Reliab Eng Syst Saf 2003;81:23–69.
[4] Baker J. Efficient analytical fragility function fitting using dynamic structural [39] Federal Emergency Management Agency. FEMA356: Prestandard and
analysis. Earthquake Spectra 2015;31:579–99. commentary for the seismic rehabilitation of buildings. American Society of
[5] Liel AB, Haselton CB, Deierlein GG, Baker JW. Incorporating modeling Civil Engineers, ASCE; 2000.
uncertainties in the assessment of seismic collapse risk of buildings. Struct [40] Zentner I, Allain F, Humbert N, Caudron M. Generation of spectrum compatible
Saf 2009;31:197–211. ground motion and its use in regulatory and performance-based seismic
[6] Celik OC, Ellingwood BR. Seismic fragilities for non-ductile reinforced concrete analysis. In: Proceedings of Eurodyn, Porto, Portugal.
frames role of aleatoric and epistemic uncertainties. Struct Saf 2010;32:1–12.

Please cite this article in press as: Zentner I. A general framework for the estimation of analytical fragility functions based on multivariate probability dis-
tributions. Struct Saf (2016), http://dx.doi.org/10.1016/j.strusafe.2016.09.003