Professional Documents
Culture Documents
∗ Corresponding author.
E-mail address: schak@civil.iiests.ac.in (S. Chakraborty).
https://doi.org/10.1016/j.probengmech.2020.103081
Received 15 February 2020; Received in revised form 6 May 2020; Accepted 18 June 2020
Available online 23 June 2020
0266-8920/© 2020 Elsevier Ltd. All rights reserved.
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
intensity levels reduces the difficulty of ground motion selection as intensity parameter (spectral acceleration, 𝑆𝑎 at the fundamental period
required for the cloud method. However, scaling of ground motion of the structure) is included as an added dimension to the proposed
at multiple intensity levels and performing NLTHA for each of those BLR demand prediction models. Thus, the response approximation is
requires considerable computational time. In the likelihood method, not conditional to a specific intensity of earthquake. For each set
the fragility parameters are estimated by maximizing the likelihood of of random structural parameters and intensity parameter, a random
observing the failure at multiple intensity levels, given the response ground motion is selected from the ground motion bin and scaled to
data. In essence, the likelihood method is based on incremental dy- the corresponding intensity value. The ground motions are uniformly
namic analysis method and thus involves similar computational effort. scaled over an intensity range, but at each intensity level only one
Thereby, a compromise between accuracy and computational time can ground motion is scaled. Thereby, repeated scaling of ground motions is
be observed in the popular analytical SFA approaches. Moreover, due to avoided which makes the approach efficient for fragility computation.
coupling of different sources of uncertainty, without a measure of the The NLTHA is performed for each such structure-intensity combination
uncertainties involved in the fragility estimates, sufficient confidence and BLR is performed among the observed demand values and the
in such estimate cannot be achieved. To circumvent these limitations, input variables for both the BLS and RVM models. The ensembles of
an efficient Bayesian approach for SFA of structures is explored in the model parameters are obtained to furnish the underlying conditional
present study using Bayesian linear regression (BLR) demand models distribution of the fragility parameters. Based on the estimated model
for posterior simulation of the fragility parameters. parameters, the fragility is estimated (in robust sense) along with a
Application of linear regression demand models in SFA of struc- confidence bound representing the degree of uncertainty associated
tures is notable [26–28]. The Bayesian approach of updating linear with estimated fragility. The efficacies of the proposed BLR models
regression parameters are also attempted [23,29–31]. In this regard, are elucidated numerically by considering a nonlinear single-degree-
a class of literatures on well-established results in Bayesian parame- of-freedom (SDOF) system and a five-storey reinforced concrete (RC)
ter estimation approach for SFA of structures by constructing poste- building frame.
rior joint probability distribution of parameters of a regression-based
fragility model is worth mentioning [32,33]. The Bayesian approach 2. Proposed Bayesian framework of seismic fragility analysis
combines Bayesian updating of analytical fragility curves with seismic
response data obtained by NLTHA of a considered structure to pro- The proposed SFA approach employs Bayesian framework to obtain
the conditional distributions of the model parameters. Accordingly,
vide confidence bounds on the fragility curve [6,31,34–36]. Gokkaya
seismic fragility is estimated along with a confidence bound on the
et al. [37] presented a Bayesian method to integrate prior information
estimated fragility. To present the proposed Bayesian framework of SFA
on collapse response using response data obtained from NLTHAs to
in an effective way, the overview of both the BLS and RVM models
provide a more informed estimate of collapse risk of structures. Pu-
and subsequent estimation of fragility in the Bayesian framework are
jari et al. [38] presented a Bayesian approach using Markov Chain
presented in the following subsections.
Monte Carlo (MCMC) simulation scheme for updating fragility param-
eters combined with several alternative reliability methods to estimate
2.1. Seismic fragility analysis
robust fragility considering both record-to-record variability and struc-
tural parameter uncertainty. Xu and Gardoni [39] proposed bivariate
Seismic fragility is defined as the probability of failure of a structure
fragility estimates of reinforced concrete buildings considering demand
for a prescribed limit state when subjected to an earthquake ground
and capacity models using MCMC simulation scheme to obtain the
motion of some specified intensity. Usually, fragility is modelled by a
posterior distribution of unknown parameters of the demand models.
lognormal cumulative distribution function [43]. Based on this and as-
Jalayer et al. [40] proposed a Bayesian cloud analysis procedure by
suming that structural demand and capacity variables are uncorrelated,
constructing a two parameter regression model. The joint probabil-
following the first order reliability method [44], the fragility can be
ity distribution of regression parameters are updated using standard estimated for a particular threshold limit state, 𝑑 𝐿 as,
Monte Carlo procedure. Jeon et al. [30] developed a multi-parameter ( )
demand model for different components of curved multi-frame concrete 𝜇ln 𝐘 − ln(𝑑 𝐿 )
𝑃𝑓 = 𝛷 (1)
box-girder bridges using Bayesian regression analysis. It is generally 𝜎ln 𝐘
noted from the previous studies that a Bayesian approach can esti- Where, 𝜇ln 𝐘 and 𝜎ln 𝐘 are the mean and standard deviation of the
mate fragility along with a measure of its uncertainty by furnishing logarithm of demand variable, Y. In the present study, the fragility
a confidence bound on the fragility curves. However, it is important is estimated based on two different BLR demand models, i.e., the BLS
to address the issues of accuracy and computational time at the same and the RVM model. In the following sections, the development of the
time. These are not well reflected in the existing studies. Thus, a generic BLS model is presented first, followed by the development of the RVM
Bayesian framework for SFA of structures accommodating different model.
sources of uncertainty with due importance to accuracy and efficiency
is felt important. 2.2. Proposed BLR model based on Bayesian least squares regression
Two different BLR models are explored in the present study for SFA
of structures utilizing limited number of NLTHA results. The proposed In SFA of structure, the quantity of primary interest is the condi-
BLR models incorporate uncertainty due to both the record-to-record tional distribution of structural demand variable 𝐘 for a given seismic
variation of ground motions and structural parameter uncertainty. The intensity 𝑆𝑎 i.e., 𝑃 (𝐘|𝑆𝑎 ). Further, to take into account the structural
first one is the classical Bayesian least squares (BLS) regression where parameter uncertainty, the above conditional distribution can be pa-
the posterior samples of the fragility parameters are generated using rameterized as 𝑃 (𝐘|𝜽, 𝑆𝑎 ) where, 𝜽 = (𝜃1 , 𝜃2 , … , 𝜃𝑛 ) is the vector of
MCMC simulation [33]. In the second model, the sparse Bayesian ‘n’ random structural parameters. The distribution of demand variable
learning is explored using Relevance Vector Machine (RVM) [41]. The (𝐘) is conventionally assumed to be lognormal and transformed into
dispersion in response due to record-to-record variations is considered a normal distribution by taking logarithm of 𝐘 to construct the BLS
by selecting a bin of earthquake records based on the hazard level model. The expected value of ln 𝐘 is represented as,
as identified from the probabilistic seismic hazard analysis of the ∑
𝑛
considered site [42]. For regression analysis, the structural demand 𝐸(ln 𝐘|𝜽, 𝑆𝑎 ) = 𝑤0 + 𝑤𝑗 𝜃𝑖𝑗 + 𝑤𝑆𝑎 ln 𝑆𝑎,𝑖 , (𝑖 = 1, 2, … , 𝑁) (2)
data are obtained from NLTHA based on random samples generated by 𝑗=1
Latin Hypercube sampling following the probability density function of Where, 𝑤0 is the bias, 𝑤𝑗 is the weight associated with the 𝑗th random
the random parameters. It is to be noted here that the ground motion structural parameter 𝜃𝑗 , 𝑤𝑆𝑎 is the weight associated with 𝑆𝑎 values
2
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
and 𝐸(.) indicates the statistical expectation. The above equation can Then, the posterior density in Eq. (8) can be computed by marginal
be represented in a more compact form as, 2
and conditional simulation of 𝜎ln and 𝐰|𝜎ln 2 , respectively. For known
𝐘 𝐘
2 , the posterior distribution, 𝑃 (𝐰|𝜎 2 , ln 𝐘) can be obtained as [32],
𝜎ln
𝐸(ln 𝐘|𝜽, 𝑆𝑎 ) = 𝐗𝐰 (3) 𝐘 ln 𝐘
following section.
2.2.1. Posterior distribution of demand parameters
Following Bayesian updating concept, the joint posterior distribu- 2.2.2. Estimation of BLR model parameters by MCMC simulation
tions of the model parameters can be represented as the product of the The MCMC simulation technique can generate random samples
joint likelihood and the prior density as, directly from the posterior distribution of the model parameters where
the posterior distribution can only be known up to a normalizing
2 2 2
𝑃 (𝐰, 𝜎ln 𝐘
|ln 𝐘 ) ∝ 𝑃 (ln 𝐘|𝐰, 𝜎ln 𝐘
)𝑃 (𝐰, 𝜎ln 𝐘
) (5) constant. The most popular MCMC sampling techniques used are the
Where, 𝑃 (ln 𝐘|𝐰, 𝜎ln 2 ) is the likelihood, 𝑃 (𝐰, 𝜎 2 ) represents the prior Metropolis–Hastings algorithm [23,38], Gibbs sampling [45] and ad-
𝐘 ln 𝐘
distribution. The symbol∝ represents that the posterior distribution is vanced samplers like Hamiltonian Monte Carlo [46] and NUTS (No-
proportional to the product of the likelihood and the prior. The left U-Turn sampler) [47]. The present study adopts the Gibbs sampler
hand side of Eq. (5) is not exactly equals to the right hand sides, but in- in which the posterior samples are generated by sweeping through
volve a normalizing constant. The normalizing constant is analytically each variable from its conditional distribution while keeping remaining
variables fixed at their current values. The posterior samples of 𝐰|𝜎ln 2
intractable in many cases and rarely needs to be evaluated for Bayesian 𝐘
2
and 𝜎ln 𝐘 can be drawn easily from the joint posterior distribution of
updating using MCMC technique [32,33]. A normal likelihood function
and a noninformative prior distribution is adopted for posterior simu- 𝑃 (𝐰, 𝜎ln 𝐘 | ln 𝐘) by simulating 𝜎ln 𝐘 from 𝑃 (𝜎ln 𝐘 | ln 𝐘) using Eq. (12)
2 2 2
lation of the model parameters. For known values of 𝜎ln 2 , the normal with an initial estimate of the coefficient vector 𝐰̂ from Eq. (10) and
𝐘
then simulating the coefficient vector 𝐰̂ from 𝑃 (𝐰|𝜎ln 2 , ln 𝐘) using
likelihood function can be represented as, 𝐘
( ) Eq. (9). However, the initial samples may not be true representative
2 1
𝑃 (ln 𝐘|𝐰, 𝜎ln 𝐘
) ∝ exp − (𝐰 − 𝐰) ̂ 𝑇 𝜮 −1 (𝐰 − 𝐰)
̂ (6) of the samples from the actual posterior distribution as the algorithm
2
2 is initiated with random seeds. Hence, it is desirable to reject some of
Where, 𝜮 = [𝐗𝐓 𝐗]−𝟏 𝜎ln 𝐘
is the covariance matrix.
the initial samples as burn-in.
In Bayesian inference problem, choice of prior distribution plays an 2
Once the posterior statistics of 𝐰 and 𝜎ln 𝐘
are obtained, the mean
important role. If no specific information is available on the prior dis-
values of ln 𝐘, i.e. 𝜇ln 𝐘 can be estimated for varying levels of seismic
tribution of the model parameters, the assumption of a non-informative
intensity. Now, for a chosen intensity value, the seismic fragility can
prior distribution can be a reasonable choice. A useful choice of non- 2 .
2 ) be obtained from Eq. (1) using the simulated values of 𝜇ln 𝐘 and 𝜎ln
informative prior for normal regression model is uniform on (𝐰, ln 𝜎ln 𝐘
𝐘
Finally, the mean and the standard deviation of fragility, 𝑃𝑓 can be
[33] i.e.
estimated as,
2
𝑃 (𝐰, ln 𝜎ln 2
) ∝ 1 or, equivalently, 𝑃 (𝐰, 𝜎ln −2
) ∝ 𝜎ln (7) ( )
𝐘 𝐘 𝐘 ∑𝑚 𝜇ln 𝐘,𝑖 −ln(𝑑 𝐿 )
𝑖=1 𝛷 𝜎ln 𝐘,𝑖
Based on this non-informative prior distribution, the joint distribution
2 | ln 𝐘) can be represented as, 𝑃𝑓 = (14)
𝑃 ( 𝐰, 𝜎ln |
𝐘|
𝑚
[ ( ) ]2
( ) ∑𝑚
𝜇ln 𝐘,𝑖 − ln(𝑑 𝐿 )
1
2 −𝑁∕2 1 𝜎𝑃2 = 𝛷 − 𝑃𝑓 (15)
𝑃 (𝐰, 𝜎ln 𝐘 |ln 𝐘 ) ∝ 𝜎ln 𝐘 exp − 𝑇
(ln 𝐘 − 𝐗𝐰) (ln 𝐘 − 𝐗𝐰) (8) 𝑓 (𝑚 − 1) 𝑖=1 𝜎ln 𝐘,𝑖
2
2𝜎ln 𝐘
3
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Where, 𝑚 is the total number of generated posterior samples excluding 2 . For this, the posterior distribution is conveniently
respect to 𝜶 and 𝜎ln 𝐘
the burn-in samples. From the knowledge of 𝑃 𝑓 and 𝜎𝑃𝑓 for different decomposed as,
values of 𝑆𝑎 , the mean fragility curve and its confidence bounds can be 2 2 2
𝑃 (𝜶, 𝜎ln | ln 𝐘) ∝ 𝑃 (ln 𝐘|𝜶, 𝜎ln )𝑃 (𝜶)𝑃 (𝜎ln ) (24)
constructed. 𝐘 𝐘 𝐘
and 𝐀 = diag(𝛼1 , 𝛼2 , … , 𝛼𝑁 ). The effectiveness of the proposed BLR model for SFA of structure is
Now, to compute the second term in Eq. (20), the hyper parameter demonstrated numerically by considering two examples. The first one
2 | ln 𝐘) is approximated by obtaining their posterior
posterior 𝑃 (𝜶, 𝜎ln is a nonlinear SDOF system for which a large number of non-linear
𝐘
2
mode (𝜶 MP , 𝜎ln ) that maximizes the posterior distribution with seismic responses by MCS can be readily obtained with reasonable time
𝐘,MP
4
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Fig. 1a. Step-by-step procedure for estimating fragility by the BLS model.
Fig. 1b. Step-by-step procedure for estimating fragility by the RVM model.
for proper comparative study. The second example is a more realistic are several methods and computational tools available for this. The
one i.e. a five-storey RC building frame. present study adopts the spectral matching technique [49] for artificial
The choice of ground motions plays an important role to obtain
ground motion generation. Eight artificial accelerograms compatible
structural demand parameters through NLTHA for SFA of structure. The
most acceptable form for this is the use of recorded accelerograms. The to the acceleration response spectra for rock and hard soil for 5%
Guwahati city of northeast India, one of the seismically most active damping [50] are generated. Further, eight synthetic accelerograms are
regions of the world, is considered for numerical demonstration of generated following the stochastic point source model as proposed by
both the examples. However, due to scarcity of recorded accelerogram
specific to the desired hazard level for the considered location, the Boore [51]. These augment the natural ground motion database and
choice of natural ground motions in the present study is limited to result in a bin of twenty four accelerograms used for NLTHA. Further
eight numbers. These are selected from the past earthquake data in the details on the ground motion generation for the region can be found in
region covering a surface wave magnitude (𝑀𝑗 ) ranging from 6.0 to 8.0
Ghosh and Chakraborty [24,52]. In this regard, it may be noted that the
and epicentral distance (𝑅𝑖 ) within 300 km for rock site. To augment
the ground motion database for a statistically meaningful study, the Broadband Platform, an open access software for simulation of ground
accelerograms are further generated synthetically and artificially. There motions may also be readily used [53].
5
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Fig. 2. (a) The SDOF system and (b) the force deformation behaviour of the nonlinear spring.
The first example is a nonlinear SDOF oscillator characterize with Random structural parameters Distribution type Mean cov
bilinear force deformation behaviour as shown in Fig. 2. The hysteretic Natural frequency (𝜃1 ) in rad/s Normal 6.28 0.2
behaviour of the nonlinear spring is modelled in the OpenSees [54] Damping ratio (𝜃2 ) Normal 0.02 0.25
Yield strength (𝜃3 ) in KN Normal 8.0 0.2
with multilinear plastic material where the unloading branch is par-
Post yield slope (𝛼) (𝜃4 ) Normal 0.05 0.25
allel to the loading branch. Four uncorrelated random variables 𝜽 =
(𝜃1 , 𝜃2 , 𝜃3 , 𝜃4 ) representing the uncertain system parameters as depicted
in Table 1 are considered [55]. The spectral acceleration, 𝑆𝑎 is also
considered as a random variable uniformly distributed over accelera- technique and initial 1000 samples are rejected as burn-in samples.
tion range of 0.1g to 1.0g. Twenty-four random samples of 𝜽 and 𝑆𝑎 are Fig. 3 shows the marginal posterior distributions of the weights as
generated and combined using Latin Hypercube sampling with reduced obtained from the MCMC simulation, along with the fitted normal
correlation [56] to form the input matrix X for NLTHA. Thus, each row distributions curve with their mean and standard deviation values.
X𝑖 = (𝜃𝑖1 , … , 𝜃𝑖4 , 𝑆𝑎,𝑖 ), 𝑖 = 1, 2, … , 24 of the input matrix X represents The typical joint posterior samples of 𝜇ln 𝐘 and 𝜎ln 𝐘 as obtained from
a random realization of the structures and a particular 𝑆𝑎 value. For MCMC simulation for 𝑆𝑎 = 0.5g are shown in Fig. 4(a). The conditional
each X𝑖 , a random ground motion is selected from the bin and scaled to distribution of 𝜇ln 𝐘 and the marginal distribution of 𝜎ln 𝐘 are also
the corresponding 𝑆𝑎 value. The NLTHA is performed for each of such plotted in the corresponding axes. To compare the distribution of the
realization of the system, providing a vector of twenty-four observed generated samples with respect to the actual value, the most accurate
demand values 𝐘. values of the parameters are estimated by direct MCS technique. For
With the knowledge of 𝐘 and 𝐗, the SFA is first carried out by the this, 30,000 random realizations of the system are generated by LHS
BLS model as presented in Section 2.2. In doing so, 5000 posterior sam- according to the distributions of the parameters as described in Table 1.
ples of the parameters (𝐰, 𝜎ln 2 ) are simulated using MCMC simulation Now, each realization of the system is associated with a randomly
𝐘
6
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Fig. 4. (a) The posterior samples of 𝜇ln 𝐘 and 𝜎ln2 𝐘 for 𝑆𝑎 = 0.5g obtained from the BLS model and (b) the predicted mean demand line and its ± 2𝜎ln 𝐘 confidence bounds obtained
from the BLS and RVM models.
selected ground motion from the bin scaled to a particular 𝑆𝑎 level prepared for NLTHA and linear regression is performed in the loga-
and NLTHA is carried out accordingly. The procedure is repeated for rithmic scale with the obtained demand values and the corresponding
different 𝑆𝑎 levels. The mean and standard deviation of the demand 𝑆𝑎 values. The median of the demand values and their dispersion are
values are estimated for each of such 𝑆𝑎 levels. The estimated mean estimated from the regression and fragility is obtained using Eq. (1).
and standard deviation of the demand values as obtained from the More detail on the cloud method can be found in [18,57]. For the
direct MCS technique are shown in the same plot (star symbol) for likelihood method, each ground motion is scaled to varying levels of 𝑆𝑎
comparative study. It can be noted that the distribution of the generated and for each of such levels; the numbers of ground motions exceeding
posterior samples obtained by the MCMC method are almost centred 𝑑 𝐿 is obtained. The fragility parameters are estimated by maximizing
around the most accurate estimate obtained from the direct MCS. the likelihood of observing the failure for different 𝑆𝑎 levels [22,58]
In the RVM model, a linear kernel function centred at each data and fragility curves are constructed accordingly. Fig. 5 compares the
point is used. With the knowledge of 𝐘 and 𝐗, the BLR is performed fragility results obtained by the different approaches. Additionally,
using the RVM model as explained in Section 2.3. A uniform hyper prior to investigate the influence of structural parameter uncertainty, the
is used by setting the hyper parameters in Eq. (19) to zero (i.e., 𝑎 = fragility curves obtained from the BLS approach without consideration
𝑏 = 𝑐 = 𝑑 = 0). After automatic relevance determination by the of the structural parameter uncertainty is also shown in the same plot
RVM algorithm, the relevant non-zero weights and their corresponding (marked as ‘BLS0’). Regarding computational efficiency, it is important
basis functions are retained. To estimate the mean demand for varying to note that the proposed BLR models require same twenty four num-
levels of IMs, the random structural parameters as given in Table 1 bers of NLTHAs as necessary by the cloud method of fragility analysis.
is set at their mean values and the kernel function is recalculated as, Whereas, total 240 numbers of NLTHAs are necessary for the likelihood
𝐾(𝐗, 𝐗𝐦𝐞𝐚𝐧 ) = ⟨𝐗.𝐗𝐦𝐞𝐚𝐧 ⟩, i.e., centred at the mean structural parameter method due to scaling of ground motions at ten different 𝑆𝑎 levels (0.1g
vector where ⟨.⟩ denotes the inner product. The actual demand values to 1.0g, @0.1g). Moreover, the cloud and the likelihood methods are
obtained from the NLTHA along with the fitted mean and its ±2𝜎ln 𝐘 performed considering record-to-record variation of ground motions
confidence bounds as obtained from the BLS and RVM models are only to restrict the comparative study to a small number of NLTHAs.
shown in Fig. 4(b). It can be readily observed from the plots that Thus, to incorporate structural uncertainty, the number of NLTHAs
the estimated mean demand line and its ±2𝜎ln 𝐘 confidence interval necessary by the likelihood method will be much more depending
obtained from both the BLS and the RVM models are almost same. on the considered number of random realizations of the structure.
However, the RVM estimates the mean demand using only five relevant Also, for the cloud method a separate capacity analyses is involved
vectors out of total twenty four as shown in the plot. Overall, it is to obtain the statistics of the capacity parameters. Thus, with respect
observed that the estimated mean demand lines are practically centred to the computational time and accuracy, the proposed BLR approach
about the data points and almost all the data points are contained shows significant improvement over the conventional analytical SFA
within the ±2𝜎ln 𝐘 confidence interval for both the approaches. approaches.
For demonstrating the fragility assessment by the proposed Bayesian The estimated fragilities by the direct MCS technique are considered
framework, two performance limit state thresholds (𝑑 𝐿 ) are considered, as the benchmark solution for comparison purpose. It can be readily
namely LS1 and LS2, which corresponds to the maximum displacement noted from the plot that the proposed BLR models estimate the fragility
of 200 mm and 300 mm, respectively. In the BLS model, for each of with reasonable accuracy and the estimates are almost equal for both
the simulated pairs of (𝐰, 𝜎ln 2 ), the 𝑃 is estimated for varying level
𝐘 𝑓 the approaches. When compared to the conventional analytical SFA
of 𝑆𝑎 . The mean fragility (𝑃 𝑓 ) and its ±2𝜎𝑃𝑓 confidence bounds are approaches, it is quite apparent from the plots that the BLR models
obtained by using Eq. (14) and Eq. (15). In the RVM model, 5000 are more accurate than the conventional approaches. The effect of
random samples are generated from the posterior distribution of the structural parameter uncertainty in the fragility estimate is quite notice-
relevant weights following Section 2.3.2. The value of 𝜎ln 2 is set at
𝐘 able when the results of the BLR approach is compared with the BLR
its posterior mode 𝜎ln2 . With the generated posterior samples, the
𝐘,MP approach without consideration of uncertainty (i.e. BLS0). The effect
median fragility (𝑃 𝑓 ) and its ±2𝜎𝑃𝑓 confidence bounds are obtained of structural parameter uncertainty makes the fragility curve flatter,
using Eq. (14) and Eq. (15). i.e. an increase in the dispersion of the fragility due to inclusion of
To study the effectiveness of the proposed BLR based SFA approach, additional random parameters in the model. The confidence bounds are
a comparative study is made with the fragility estimate obtained by the shown in the same plot for both the BLS and the RVM models. It is
commonly used cloud and likelihood methods. For the cloud method, observed that the proposed RVM model provides a relatively narrower
the unscaled ground motions are used from the ground motion bin ±2𝜎ln 𝐘 confidence bound. However, it is important to note that the
7
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Fig. 5. Comparison of fragility curves generated by different methods for (a) LS1 and (b) LS2.
Table 2
The random structural parameters considered for SFA.
Random parameters Distribution type Mean cov
Concrete compressive strength (=𝜃1 ) Normal 25 MPa 0.21
Yield strength of steel (=𝜃2 ) Normal 500 MPa 0.1
Elastic modulus of concrete (=𝜃3 ) Log Normal 25000 MPa 0.15
Elastic modulus of steel (=𝜃4 ) Log Normal 2 × 105 MPa 0.05
Damping ratio (=𝜃5 ) Normal 0.05 0.4
8
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Fig. 7. (a) The elevation of the building frame considered for analysis and the fibre discretization of the (b) beams and the (c) columns..
mean demand line and its ±2𝜎ln 𝐘 confidence bounds are estimated bounds are estimated from the 5000 posterior samples of the random
using both the BLS and the RVM model similar to the previous example. parameters(𝐰, 𝜎ln2 ) after rejecting 1000 burn-in samples. For RVM
𝐘
Fig. 9(b) shows the obtained response values from the NLTHA and the model, with the knowledge of the posterior statistics (𝝁𝐰 , 𝜮) of the
fitted mean demand line and its ±2𝜎ln 𝐘 confidence bounds estimated weights, 5000 posterior samples of the relevant weights are generated
from both the BLR models. The estimated mean demand line and its and the mean fragility and its confidence bounds are estimated as
confidence bounds almost coincides for both the approaches and almost
explained in the previous example. The estimated fragilities by the
all the data points are contained within the obtained confidence band
proposed BLR models are compared with the fragility estimates by the
for this example also. Thus, it is quite obvious that both the BLR models
cloud and the likelihood methods based on the fragility results obtained
are efficient in estimating the mean of the responses. However, the
RVM model requires only five relevant vectors for this estimation which by the direct MCS technique as the benchmark. Fig. 10 shows the esti-
in turn results in a very sparse data that is easier and faster to compute mated fragility results by the BLR models along with the fragility results
on later. obtained by the cloud and the likelihood methods. The plot clearly
Now SFA of the frame for all the three limit states are taken reveals that both the BLS and the RVM models of the proposed BLR
up. For BLS model, the mean fragility (𝑃 𝑓 ) and its ±2𝜎𝑃𝑓 confidence based SFA approach estimates the fragility with improved accuracy.
9
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
Fig. 9. (a) The posterior samples of 𝜇ln 𝐘 and 𝜎ln2 𝐘 for 𝑆𝑎 = 0.7g obtained from the BLS model and (b) the predicted mean demand line and its ± 2𝜎ln 𝐘 confidence bounds obtained
from the BLS and RVM models.
Fig. 10. The fragility results obtained by different methods for a (a) IO, (b) LS and (c) CP limit states.
Though the fragility estimates by both the models are almost equal for The BLR models require twenty four numbers of NLTHA which is
the IO limit state, the RVM approach shows better accuracy for the LS same as the cloud method. But, the likelihood method requires 240
limit state. However, the differences are very small. With increase in NLTHA. However, using only twenty four NLTHA, the proposed BLR
the limit state threshold, increased deviation of the fragility estimates models efficiently integrate the structural parameter uncertainty with
is observed for the cloud and the likelihood methods with respect to the the record to record variation of ground motions; whereas the cloud
MCS based fragility estimate. But the proposed BLR approach is consis- method and the likelihood method require additional analysis effort.
tently accurate for all the three limit states. To investigate the influence
of the structural parameter uncertainty, the fragility estimates without
4. Summary and conclusion
considering the structural parameter uncertainty is additionally shown
in the plot. It can be noticed that the influence of structural parameter
uncertainty is comparatively less here than the first example. Thus, Two computationally efficient BLR demand prediction models are
the deviations of the fragility estimates by the other three approaches proposed for SFA of structures by integrating both the record-to-record
(i.e. the cloud method, the likelihood method and the BLS0) are less variation of seismic motions and structural model parameters uncer-
from the direct MCS based fragility estimates compared to the first tainties. The effectiveness of the proposed BLR models for SFA of
example. However, in both the example the proposed BLR models show structures is demonstrated through two numerical examples. The im-
better accuracy. Regarding computational efficiency, it can be noted portant observations and conclusions based on the presented study are
that similar computational effort is involved as in the previous example. summarized as following:
10
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
• The estimated mean demand lines by the proposed BLR models [9] K.R. Collins, Y.K. Wen, D.A. Foutch, Dual-level seismic design: A reliability-based
are almost centred around the data points and all the data points methodology, Earthq. Eng. Struct. Dyn. 25 (1996) 1433–1467, http://dx.doi.org/
10.1002/(SICI)1096-9845(199612)25:12<1433::AID-EQE629>3.0.CO;2-M.
are contained within the confidence bands. Thus, the proposed
[10] S.W. Han, Y.K. Wen, Method of reliability-based seismic design. II: Calibration
BLS models and also the RVM model can reasonably estimate the of code parameters, J. Struct. Eng. 123 (1997) 264–270.
mean demand quite well. [11] P. Bazzurro, C.A. Cornell, Seismic hazard analysis of nonlinear structures I:
• It is clearly noted that both the BLR models can estimate fragility Methodology, J. Struct. Eng. (1994) http://dx.doi.org/10.1061/(ASCE)0733-
with improved accuracy using limited number of NLTHA results 9445(1994)120:11(3320).
compared to the commonly used cloud and the likelihood meth- [12] A.H.M. Muntasir Billah, M. Shahria Alam, Seismic fragility assessment of highway
bridges: a state-of-the-art review, Struct. Infrastruct. Eng. 11 (2015) 804–832,
ods considering the direct MCS based fragility results as the http://dx.doi.org/10.1080/15732479.2014.912243.
benchmark. [13] M. Aghababaei, M. Mahsuli, Detailed seismic risk analysis of buildings using
• Structural parameter uncertainty has considerable influence in structural reliability methods, Probab. Eng. Mech. 53 (2018) 23–38, http://dx.
the estimated fragility and the proposed BLR models incorpo- doi.org/10.1016/j.probengmech.2018.04.001.
rate that effect quite efficiently. It may be noted that the usual [14] Y.S. Choun, A.S. Elnashai, A simplified framework for probabilistic earthquake
loss estimation, Probab. Eng. Mech. 25 (2010) 355–364, http://dx.doi.org/10.
cloud and likelihood methods require additional analysis effort 1016/j.probengmech.2010.04.001.
to consider the effect of structural parameter uncertainty. [15] O.S. Kwon, A. Elnashai, The effect of material and ground motion uncertainty on
• As the Bayesian statistics is sensitive to new evidences, uncer- the seismic vulnerability curves of RC structure, Eng. Struct. 28 (2006) 289–303,
tainty in the estimated fragility can further be reduced on avail- http://dx.doi.org/10.1016/j.engstruct.2005.07.010.
ability of sufficient number of ground motions specific to a con- [16] G.C. Marano, R. Greco, E. Morrone, Analytical evaluation of essential facilities
fragility curves by using a stochastic approach, Eng. Struct. 33 (2011) 191–201,
sidered hazard level. http://dx.doi.org/10.1016/j.engstruct.2010.10.005.
• It is important to note that the proposed BLS model can provide [17] X. Romão, A.A. Costa, E. Paupério, H. Rodrigues, R. Vicente, H. Varum, A.
full posterior distribution of the fragility parameters (𝜇ln 𝐘 , 𝜎ln 𝐘 ). Costa, Field observations and interpretation of the structural performance of
But, the RVM approach cannot provide the full posterior distri- constructions after the 11 2011 lorca earthquake, Eng. Fail. Anal. (2013) http:
bution of the fragility parameters as the posterior distribution //dx.doi.org/10.1016/j.engfailanal.2013.01.040.
[18] F. Jalayer, Direct Probabilistic Seismic Anaysis: Implementing Non-Linear
information of standard deviation of response quantity (𝜎ln 𝐘 ) is
Dynamic Assessments, (Ph.D. thesis), Stanford University, Stanford, 2003.
not available by this approach. [19] B.R. Ellingwood, O.C. Celik, K. Kinali, Fragility assessment of building structural
systems in mid-america, Earthq. Eng. Struct. Dyn. 36 (2007) 1935–1952, http:
In general, it can be concluded that the proposed Bayesian approach //dx.doi.org/10.1002/eqe.693.
of SFA efficiently addresses three important aspects i.e. accuracy, com- [20] D. Vamvatsikos, C.A. Cornell, Incremental dynamic analysis, Earthq. Eng. Struct.
putational efficiency and measure of uncertainty in fragility estimate Dyn. 31 (2002) 491–514, http://dx.doi.org/10.1002/eqe.141.
required for analytical SFA of structures. The applications of the pro- [21] D. Lallemant, A. Kiremidjian, H. Burton, Statistical procedures for develop-
ing earthquake damage fragility curves, Earthq. Eng. Struct. Dyn. 44 (2015)
posed BLR models in this study are focused on specific examples and
1373–1389, http://dx.doi.org/10.1002/eqe.2522.
particular study area. However, it is generic enough to be readily ap- [22] M. Shinozuka, M.Q. Feng, J. Lee, T. Naganuma, Statistical analysis of fragility
plied for other type of structures and for other locations. Moreover, the curves, J. Eng. Mech. 126 (2000) 1224–1231, http://dx.doi.org/10.1061/(asce)
algorithms can be easily implemented in any programming platform. 0733-9399(2000)126:12(1224).
One only needs to replace the mechanical model of the considered [23] F. Jalayer, H. Ebrahimian, A. Miano, G. Manfredi, H. Sezen, Analytical fragility
assessment using unscaled ground motion records, Earthq. Eng. Struct. Dyn. 46
structure and generation of ground motion data of a specific site of
(2017) 2639–2663, http://dx.doi.org/10.1002/eqe.2922.
interest. Thus, the proposed BLR models can be applied generically for [24] S. Ghosh, S. Chakraborty, Seismic performance of reinforced concrete building
improved SFA of structures. in guwahati city, northeast India, Sci. Iran. A 24 (2017a) 1821–1833.
[25] D. Lu, X. Yu, M. Jia, G. Wang, Seismic risk assessment for a reinforced concrete
Declaration of competing interest frame designed according to chinese codes, Struct. Infrastruct. Eng. 10 (2014)
1295–1310, http://dx.doi.org/10.1080/15732479.2013.791326.
[26] Q. Huang, P. Gardoni, S. Hurlebaus, Probabilistic seismic demand models and
The authors declare that they have no known competing finan- fragility estimates for reinforced concrete highway bridges with one single-
cial interests or personal relationships that could have appeared to column bent, J. Eng. Mech. 136 (2010) 1340–1353, http://dx.doi.org/10.1061/
influence the work reported in this paper. (asce)em.1943-7889.0000186.
[27] K.R. Mackie, B. Stojadinović, Performance-based seismic bridge design for
References damage and loss limit states, Earthq. Eng. Struct. Dyn. 36 (2007) 1953–1971,
http://dx.doi.org/10.1002/eqe.699.
[28] I. Zentner, A general framework for the estimation of analytical fragility functions
[1] J.R. Gaxiola-Camacho, H. Azizsoltani, F.J. Villegas-Mercado, A. Haldar, A novel
based on multivariate probability distributions, Struct. Saf. 64 (2017) 54–61,
reliability technique for implementation of performance-based seismic design
http://dx.doi.org/10.1016/j.strusafe.2016.09.003.
of structures, Eng. Struct. 142 (2017) 137–147, http://dx.doi.org/10.1016/j.
[29] F. Berahman, F. Behnamfar, Probabilistic seismic demand model and fragility
engstruct.2017.03.076.
estimates for critical failure modes of un-anchored steel storage tanks in
[2] S. Günay, K.M. Mosalam, PEER Performance-based earthquake engineering
petroleum complexes, Probab. Eng. Mech. 24 (2009) 527–536, http://dx.doi.
methodology, revisited, J. Earthq. Eng. 17 (2013) 829–858, http://dx.doi.org/
org/10.1016/j.probengmech.2009.03.005.
10.1080/13632469.2013.787377.
[3] D. Tolentino, S. Márquez-Domínguez, J.R. Gaxiola-Camacho, Fragility assessment [30] J.S. Jeon, S. Mangalathu, J. Song, R. Desroches, Parameterized seismic fragility
of bridges considering cumulative damage caused by seismic loading, KSCE J. curves for curved multi-frame concrete box-girder bridges using Bayesian pa-
Civ. Eng. (2020) 1–10. rameter estimation, J. Earthq. Eng. 23 (2017) 1–26, http://dx.doi.org/10.1080/
[4] M. Fragiadakis, D. Vamvatsikos, M. Karlaftis, N. Lagaros, M. Papadrakakis, 13632469.2017.1342291.
Seismic assessment of structures and lifelines, J. Sound Vib. 334 (2015) 29–56. [31] S.K. Ramamoorthy, P. Gardoni, J.M. Bracci, Probabilistic demand models and
[5] N.D. Lagaros, Y. Tsompanakis, P.N. Psarropoulos, E.C. Georgopoulos, Computa- fragility curves for reinforced concrete frames, J. Struct. Eng. 132 (2006)
tionally efficient seismic fragility analysis of geostructures, Comput. Struct. 87 1563–1572, http://dx.doi.org/10.1061/(ASCE)0733-9445(2006)132:10(1563).
(2009) 1195–1203, http://dx.doi.org/10.1016/j.compstruc.2008.12.001. [32] G.E.P. Box, G.C. Tiao, Bayesian Inference in Statistical Analysis, John Wiley &
[6] P. Gardoni, K.M. Mosalam, A.D. Kiureghian, Probabilistic seismic demand models Sons, 40, 1992.
and fragility estimates for RC bridges, J. Earthq. Eng. 7 (spec 01) (2003) 79–106, [33] A. Gelman, J.B. Carlin, H.S. Stern, D.B. Dunson, A. Vehtari, D.B. Rubin, Bayesian
http://dx.doi.org/10.1080/13632460309350474. Data Analysis, third ed, Chapman and Hall/CRC, 2013.
[7] G.C. Marano, R. Greco, M. Mezzina, Stochastic approach for analytical fragility [34] P.S. Koutsourelakis, Assessing structural vulnerability against earthquakes using
curves, KSCE J. Civ. Eng. 12 (2008) 305–312, http://dx.doi.org/10.1007/s12205- multi-dimensional fragility surfaces: A Bayesian framework, Probab. Eng. Mech.
008-0305-8. 25 (2010) 49–60, http://dx.doi.org/10.1016/j.probengmech.2009.05.005.
[8] N. Shome, C.A. Cornell, P. Bazzurro, J.E. Carballo, Earthquakes, records, and [35] A. Singhal, A.S. Kiremidjian, Bayesian Updating of fragilities with application
nonlinear responses, Earthq. Spectra 14 (1998) 469–500, http://dx.doi.org/10. to RC frames, J. Struct. Eng. 124 (2002) 922–929, http://dx.doi.org/10.1061/
1193/1.1586011. (asce)0733-9445(1998)124:8(922).
11
S. Ghosh and S. Chakraborty Probabilistic Engineering Mechanics 61 (2020) 103081
[36] D. Straub, A. Der Kiureghian, Improved seismic fragility modeling from empirical [48] M.E. Tipping, A.C. Faul, Fast marginal likelihood maximisation for sparse
data, Struct. Saf. 30 (2008) 320–336, http://dx.doi.org/10.1016/j.strusafe.2007. Bayesian models, in: The Ninth International Workshop on Artificial Intelligence
05.004. and Statistics, Key West, Fiorida, 2003..
[37] B.U. Gokkaya, J.W. Baker, G. Deierlein, Illustrating a Bayesian approach to seis- [49] D. Gasparini, E.H. Vanmarcke, SIMQKE: A Program for Artificial Motion Gener-
mic collapse risk assessment, in: 12th International Conference on Applications ation, Department of Civil Engineering, Massachusetts Institute of Technology,
of Statistics and Probability in Civil Engineering, ICASP12, July (2015) 12-15, Cambridge, MA, 1976.
2015, http://dx.doi.org/10.1007/s10689-009-9240-1. [50] IS1893, Indian Standard Criteria for Earthquake Resistant Design of Structures,
[38] N.N. Pujari, S. Ghosh, S. Lala, Bayesian Approach for the seismic fragility Part 1 - General Provisions and Buildings, Bureau of Indian Standards, New
estimation of a containment shell based on the formation of through-wall cracks, Delhi, India, 2016.
ASCE-ASME J. Risk Uncertain. Eng. Syst. A 2 (2015) B4015004, http://dx.doi. [51] D.M. Boore, Simulation of ground motion using the stochastic method, Pure Appl.
org/10.1061/ajrua6.0000840. Geophys. 160 (2003) 635–676.
[39] H. Xu, P. Gardoni, Probabilistic capacity and seismic demand models and fragility [52] S. Ghosh, S. Chakraborty, Probabilistic seismic hazard analysis and synthetic
estimates for reinforced concrete buildings based on three-dimensional analyses, ground motion generation for seismic risk assessment of structures in the
Eng. Struct. 112 (2016) 200–214, http://dx.doi.org/10.1016/j.engstruct.2016.01. northeast India, Int. J. Geotech. Earthq. Eng. (IJGEE) 8 (2017) 39–59.
005. [53] J.R. Gaxiola-Camacho, A. Haldar, A. Reyes-Salazar, F. Valenzuela-Beltran, G.E.
[40] F. Jalayer, R. De Risi, G. Manfredi, Bayesian Cloud analysis: Efficient struc- Vazquez-Becerra, A.O. Vazquez-Hernandez, Alternative reliability-based method-
tural fragility assessment using linear regression, Bull. Earthq. Eng. 13 (2015) ology for evaluation of structures excited by earthquakes, Earthq. Struct. 14
1183–1203, http://dx.doi.org/10.1007/s10518-014-9692-z. (2018) 361–377.
[41] M.E. Tipping, Sparse Bayesian learning and the relevance vector machine, J. [54] F. McKenna, Opensees: A framework for earthquake engineering simulation,
Mach. Learn. Res. 1 (2001) 211–244. Comput. Sci. Eng. 13 (2011) 58–66, http://dx.doi.org/10.1109/MCSE.2011.66.
[42] B.Ö. Ay, S. Akkar, A procedure on ground motion selection and scaling for [55] S. Ghosh, A. Roy, S. Chakraborty, Support vector regression based metamodeling
nonlinear response of simple structural systems, Earthq. Eng. Struct. Dyn. 41 for seismic reliability analysis of structures, Appl. Math. Model. 64 (2018)
(2012) 1693–1707, http://dx.doi.org/10.1002/eqe.1198. 584–602.
[43] I. Zentner, Numerical computation of fragility curves for NPP equipment, Nucl. [56] A.M.J. Olsson, G.E. Sandberg, Latin hypercube sampling for stochastic finite
Eng. Des. (2010) http://dx.doi.org/10.1016/j.nucengdes.2010.02.030. element analysis, J. Eng. Mech. 128 (2002) 121–125, http://dx.doi.org/10.1061/
[44] A. Haldar, S. Mahadevan, Probability, Reliability, and Statistical Methods in (ASCE)0733-9399(2002)128:1(121).
Engineering Design, John Wiley and Sons, New York, 2000. [57] K. Kinali, B.R. Ellingwood, Seismic fragility assessment of steel frames for
[45] J. Alam, D. Kim, B. Choi, Uncertainty reduction of seismic fragility of intake consequence-based engineering: A case study for memphis, TN, Eng. Struct. 29
tower using Bayesian inference and Markov chain Monte Carlo simulation, Struct. (2007) 1115–1127, http://dx.doi.org/10.1016/j.engstruct.2006.08.017.
Eng. Mech. 63 (2017) 47–53, http://dx.doi.org/10.12989/sem.2017.63.1.047. [58] S. Ghosh, S. Ghosh, S. Chakraborty, Seismic fragility analysis in the probabilistic
[46] R.M. Neal, S. Brooks, A. Gelman, G. Jones, X.L. Meng, MCMC Using hamiltonian performance-based earthquake engineering framework: an overview, Int. J. Adv.
dynamics, in: Handbook of Markov Chain Monte Carlo, first ed, Chapman & Hall Eng. Sci. Appl. Math. (2017) http://dx.doi.org/10.1007/s12572-017-0200-y.
/ CRC Press, 2011. [59] A. Bhosale, R. Davis, P. Sarkar, Sensitivity and reliability analysis of masonry
[47] M.D. Hoffman, A. Gelman, The no-u-turn sampler: Adaptively setting path lengths infilled frames, Int. J. Civ. Environ. Struct. Constr. Arch. Eng. 10 (2016)
in hamiltonian Monte Carlo, J. Mach. Learn. Res. 15 (2014) 1593–1623, http: 1452–1456.
//dx.doi.org/10.1190/1.3627885. [60] FEMA356, Prestandard and Commentary for the Seismic Rehabilitation of
Buildings, Rep. (356), Federal Emergency Management Agency, Washington, DC,
2000.
12