Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Contents lists available at ScienceDirect

Soil Dynamics and Earthquake Engineering

journal homepage: www.elsevier.com/locate/soildyn

Support vector machine based reliability analysis of concrete dams MARK

a,⁎ b
Mohammad Amin Hariri-Ardebili , Farhad Pourkamali-Anaraki
a
Department of Civil, Environmental and Architectural Engineering, University of Colorado, Boulder, CO, USA
b
Department of Applied Mathematics, University of Colorado, Boulder, CO, USA

A R T I C L E I N F O A B S T R A C T

Keywords: This paper presents possible combination of structural responses of concrete dams with machine learning
Concrete damage techniques. Support vector machine (SVM) method is adopted and two broad applications are presented: one for
Dams a simplified flood reliability assessment of gravity dams and the other for detailed nonlinear seismic finite
Reliability element method (FEM) based analysis. Up to seventeen random variables are considered in the former example
Classification
and the results of SVM contrasted with classical reliability analyses techniques (i.e., first- and second-order
Randomness
reliability methods, Monte Carlo simulation, Latin Hypercube and importance sampling techniques). For the
Support vector machine
latter example, a FEM-SVM based hybrid methodology is proposed for reduction of number of nonlinear ana-
lyses. A discussion is provided on the relation between the optimal earthquake intensity measures, the damage
states and the accuracy of prediction. It is found that the family of SVM (i.e. standard, least squares, multi-class
and regression) is an useful and effective tool for classification, response prediction and reliability analysis of the
concrete dams with reasonable accuracy.

1. Introduction FS = f (T , W , U , φ , c, A, α, FSφ, FSc ) (2)

1.1. Transition from deterministic to probabilistic methods
Traditionally in structural and geotechnical engineering, the struc-
tural safety has been evaluated based on deterministic factor of safety
(FS ), defined as the ratio between the average resistance, R (i.e., ca-
pacity), the maximum load under which a system can perform its in-
tended function, and the average stress, S (i.e., applied load or demand)
[109]:
R by the larger uncertainties in the design [112]. This method uses the
FS =
S (1)
If FS > 1, a margin of safety exists. Here, the fundamental concept is
to design the structure with appropriate safety margin, so that any
source of embedded uncertainty either in the demand or the capacity do
not threaten to cause failure [33]. Based on Ruggeri [101], the most
well-consolidated traditional method for deterministic safety assess-
ment of concrete dams is limit equilibrium method (LEM) in which the
dam is assumed to be a rigid body and the sliding is only allowed along
the critical surfaces (i.e., concrete-rock interface and concrete lift
joints). This method which is followed by many regulators/countries is
mainly based on experiences and engineering judgment [114]. Gen-
erally, many of these factors of safety have the following form:
where T is shear force, W is the weight, U is uplift force, φ and c are
angle of friction and cohesion at the considered plane respectively, A is
the area of rapture, α is the inclination of the sliding with respect to
horizon and finally, FSφ ≥ 1 and FSc ≥ 1 are the partial factor of safeties
with respect to friction and cohesion, respectively.
The advantage of this method is that the expression for (sliding)
safety factor is straightforward; however, they have not been calibrated
against a useful safety level [61]. A larger factor of safety does not
necessarily imply a smaller risk, because its effect can be canceled out
identical threshold factor of safety value for a given failure mode
without accounting for the degree of uncertainty involved in the cal-
culation [101].
Then, the margin of safety, Z, can be defined (it is the difference
between resistance and stress as oppose to FS which was the ratio be-
tween those two):
Z=R−S (3)
If capacity exceeds demand, Z > 0 , the system is in a survival state.
If demand exceeds capacity, Z < 0 , the system is in a failure state. The
condition Z = 0 is the limiting state. This function, Eq. (3), is known as
limit state (LS) equation or performance function. For example, the
following simple LSs can be defined for a typical gravity dam:

⁎
Corresponding author.
E-mail address: mohammad.haririardebili@colorado.edu (M.A. Hariri-Ardebili).

http://dx.doi.org/10.1016/j.soildyn.2017.09.016
Received 15 May 2017; Received in revised form 31 July 2017; Accepted 21 September 2017
0267-7261/ © 2017 Elsevier Ltd. All rights reserved.
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Z1 = (W − U ). tan φ + c. A − T
n m
Z2 = ∑i = 1 (FR. dR)i − ∑ j = 1 (FS . dS ) j (4)
where FR and FS are resisting and driving forces, respectively; and dR
and dS are the corresponding moment arms around the dam's toe [129].
Consequently, the failure probability may be computed as the
probability of having Z < 0 . “Structural reliability” analysis deals with
the quantitative assessment of the failure probability, given a model of
the uncertainty in the structural, environmental and load parameters
[76]. The reliability estimated as a measure of the structural safety can
be used in a decision process. For example, a lower level of the relia-
bility can be used as a constraint in optimization problems [113]. In
order to estimate the structural reliability using the probabilistic con- Fig. 1. Skeleton of the literature survey and the relation between the studied concepts.
cepts it is necessary to include the uncertainty [31] and random vari-
ables (RVs) in the problem definition.
1.3.1. Reliability analysis of concrete dams
To the best of the author's knowledge Bury and Kreuzer [15] is the
1.2. Objectives and organization
first work on calculation of the failure probability in concrete dams.
They made rigid body analysis of a gravity dam under sliding failure
Based on an extensive literature review in Section 1.3, there is no
mode. Gumbel distributional model was assumed for both the annual
comprehensive research on structural reliability and estimation of
peak flood and the ground acceleration.
failure probability in concrete dams using the support vector machine
Baylosis and Bennett [8] evaluated the safety and failure probability
technique. This paper aims to adopt SVM to determine the failure
of a high gravity dam under the static and dynamic conditions. Three
probability of the gravity dams under material, modeling and loading
failure modes were considered (i.e., sliding, overturning and over-
uncertainties.
stressing) and reported that probability of sliding during earthquake is
The main innovation of this research can be summarized as follows:
negligible compared to two others.

• Combination of structural reliability and SVM for the first time in These two were followed by de Araújo and Awruch [29] where both
the concrete properties and the seismic excitation were assumed to be
response analysis of concrete dams.
RVs in finite element analyses. Safety factor against sliding, concrete
• Contrasting both the simple analytical technique and complex finite
cracking and crushing were computed using Monte Carlo simulation
element method in SVM based reliability analysis.
• First application of seismic reliability analysis of concrete dams (MCS) and presented as a cumulative probability.
Horyna [49] is the first solid research on the application of struc-
which accounts for ground motion record-to-record variability.
• Simulation of the system reliability under both the hydrological tural reliability in the concrete dam assessment. He focused on post-
crack dynamic analysis of existing gravity dams and evaluated the re-
hazard (analytical model) and seismic hazard (finite element
liability against sliding. Both the analytical and experimental studies
model).
• Proposing a hybrid FEM-SVM based methodology for seismic relia-
were performed. The ground motion type, PGA and water level were
assumed to be three main RVs in the reliability assessment.
bility assessment of dams.
Jeppsson [54] performed a safety assessment of a concrete column
in an existing dam using both the current Swedish deterministic
First, the fundamental theory of structural reliability (Section 2) as
guideline and reliability analysis. Overturning and sliding were con-
well as the concept of support vector machine (Section 3) are reviewed.
sidered as two limit states. Reliability index was computed as a function
Next, two applications are discussed: 1) a simplified gravity dam model
of coefficient of variation (COV) in the RVs (e.g., uplift pressure, ice
based on LEM theory accounting for the material uncertainty and flood
load, and angle of friction).
loading, and 2) a finite element model of the couple dam-foundation-
Kazemi [58] compared the safety levels resulting from conventional
reservoir system with fracture mechanics based interface joint sub-
seismic stability analysis (allowable stress method) of a typical concrete
jected to large number of real ground motions (Section 4). Results are
gravity block subjected to pseudo-static earthquake load and the ulti-
discussed in Section 5 and the summary and future work will end the
mate limit state design principle (reliability method). A dam monolith
paper (Section 6).
with and without post-tensioned anchors was considered. He found that
while some measure of safety is ensured by following the conventional
1.3. Comprehensive literature review
approach, reliability method provides a consistent level of structural
reliability in the stability analysis of dams.
This subsection provides a comprehensive literature review on the
Saouma [105] combined the concept of reliability index (through
subject studied. Considering that the present paper aims to combine
the point estimate method and Taylor's series finite difference estima-
three concepts, i.e., 1) structural reliability, 2) support vector machine
tion) with finite element fracture mechanics to determine the safety
(SVM), and 3) concrete dam engineering, one-to-one relationship be-
index after rehabilitation. Linear elastic fracture mechanics (LEFM) was
tween these three concepts are reviewed separately as it is schemati-
used for evaluation of the original dam while nonlinear fracture me-
cally shown in Fig. 1. Then, the objectives of this paper are explained.
chanics (NLFM) was adapted for the retrofitted one. Reservoir eleva-
The mathematical definition of reliability is different from fragility
tion, fracture toughness, cohesion and friction angle were taken as
(while they can be connected in some aspects). Readers interested in
normally distributed RVs. He mentioned that the comparison between
fragility analysis of concrete dams may refer to Hariri-Ardebili and
two methods is not possible since input data and analysis types are
Saouma [45] where up to 25 research studies carried out between 1998
different.
and 2016 worldwide are summarized. Nearly all of these publications
Carvajal et al. [18], Carvajal et al. [17] and Carvajal et al. [16]
are limited to seismic fragility curves and surfaces where the prob-
performed reliability analysis of a gravity RCC dam by MCS and first
ability of failure (or any intermediate limit state) is computed as a
order reliability method (FORM). The shear parameters are evaluated
function (usually in the form of log-normal CDF) [121] of ground
using an intrinsic curve formula and the variability is evaluated from
motion intensity measure, e.g. peak ground acceleration (PGA) or the
variability of compressive and tensile strength. Statistical analysis of
structure's first-mode spectral acceleration (Sa (T1) ) [73].

277
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
RCC density, analysis of scatter at different spatial scales, data uni-
fication, and a physical formulation of the RCC intrinsic curve are all
considered. In addition, dependency between the hydraulic loading and
shear strength are studied. Sliding and cracking were taken as limit
state functions. Finally, Peyras et al. [91] proposed a combined method
of risk analysis and reliability assessment and event tree method for
safety analysis of concrete dams.
Papadrakakis et al. [89] and Piliounis and Lagaros [92] proposed a
neural networks (NN) based MCS procedure for reliability analysis of
large concrete dams in conjunction with a nonlinear finite element
analysis for the prediction of the bearing capacity of dams. Young's
modulus, the Poisson's ratio, the tensile strength and the specific frac-
ture energy of the concrete were considered as RVs. Probability of
failure and flood fragility curves were computed consequently.
In a comprehensive PhD dissertation, Westberg [128] investigated
the reliability-based safety assessment of concrete dams. The disserta-
tion starts with explanation of risk analysis and its relation with risk
management in dam safety. Theoretical background of the structural
reliability is then explained and the major failure modes are de-
termined. Finally, impact of the uncertainties in the RVs (e.g., strength
of concrete and rock, headwater, uplift and etc) are discussed in detail.
Moreover, Westberg Wilde and Johansson [129] calculated the system
reliability of a spillway structure consisting of two monoliths. LSs were
sliding at the concrete-rock contact, sliding at the rock mass, and ad-
justed overturning; while RVs were defined by stochastic distributions.
The safety index was calculated by the FORM for each failure mode and
monolith, and the system reliability was approximated by direct in-
tegration of the bivariate normal distribution.
Xu et al. [134] proposed an improved reliability model mixed with
finite element analysis for determination of the failure probability in
concrete dams. The results were compared with the conventional
methods like FORM, second order reliability method (SORM), weighted
regression method (WRM) and space reduced weighted regression
method (SRWRM). They claimed that the new algorithm has good
convergence and stability and enhances the calculation efficiency. Xu
et al. [132] developed a method for functional reliability analysis of
concrete gravity dams. This method which is, in fact, the extension of
the previous paper uses a pseudo excitation method to calculate the
probabilistic characteristics of the model. Finally, Xu et al. [133] fur-
ther improved the model to determine the failure probability of the
interlayer sliding belts under seismic loads. This approach is mainly
based on developing a relationship between failure probabilities of
specified interlayer sliding belts and its elements using the Markov
chain theorem.
Serrano-Lombillo et al. [110] proposed a new methodology to cal-
culate incremental risks based on the event tree framework. The main
advantage of this methodology is the ease with which it can be applied
to systems of several dams: with a single risk model that describes the
complete system and with a single calculation the incremental risks of
the system can be obtained, being able to allocate the risk of each dam
and of each failure mode.
Altarejos-Garcia et al. [2] proposed a methodology to improve the
estimation of the conditional probability of responses in gravity dam-
reservoir system using complex behavior models based on numerical
simulation techniques, together with reliability analysis (at different
levels of precision as well as MCS). This method was applied to a
gravity dam with sliding potential failure mode along the rock-concrete
interface.
In a series of publications, Krounis and Johan [63], Krounis et al. [64],
Krounis et al. [65] and Krounis [62] studied the sliding stability and
failure probability of concrete dams with bonded concrete-rock interfaces.
Both the experimental tests as well as numerical simulations were per-
formed in order to determine the concrete strength parameters. The het-
erogeneity of the interface joint was considered with different spatially
correlation lengths. Also, the impact of partial correlation between the
cohesion and friction angle was studied on the reliability index.
278
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
Wang et al. [124] proposed an interesting method for estimation of
the probability of dynamical systems’ failure. They divided the relia-
bility analysis procedure into two phases: 1) failure mode identifica-
tion, and 2) calculation of the failure probability. Uncertainties due to
earthquake variability, material randomness and selection of the con-
stitutive model for nonlinear analysis were considered (damage plas-
ticity model, smeared crack model, and extended finite element
method). Then, progressive failure analysis was performed to calculate
the failure probability at three levels, i.e., the element failure, path
failure and system failure (the similar way was proposed by Hariri-
Ardebili and Saouma [44] for damage index (DI) calculation in micro,
meta and macro levels of gravity dams). Finally, some meta-models
were proposed to generalize the failure probability.
Liu and Li [71] proposed time-dependent reliability analysis of
concrete gravity dams based on gray theory and a damage mechanics
model. This model uses a concrete constitutive model where the da-
mage changes by age of the structure. Finally, Taylor expansion is used
for quadratic nonlinear term over the structural functional equation and
combined with the engineering application principle. The main ad-
vantage of this method is that it does not need iteration and greatly
reduces the computational effort. Also, it has higher accuracy compared
to FORM.
Altarejos-Garcia et al. [1] proposed a strategy to analyze the dam-
foundation interface joint failure mode under a reliability-based ap-
proach which accounts for the epistemic uncertainty (friction angle and
cohesion). For the probabilistic modeling of the interface joint, the
conventional uniform sampling was compared to spatial variability of
strength parameters (using random field theory) and data from dam
monitoring. Finally, relationships between probability of failure and
factors of safety were obtained.
Shi et al. [111] proposed a new method for reliability analysis of
dam system with multiple failure models. This method determines the
failure domain of a complex structure by resorting to lines rather than
random points. It was used to calculate the failure probability and re-
liability sensitivity of a gravity dam and the results were compared with
the standard MCS. Results showed the efficiency of the line sampling
simulation.
All the previous papers use, to some extent, the statistical data (in
terms of the distributional model) to calculate the failure probability
and reliability index. However, in many realistic cases these data points
are not available. To overcome this problem, Su et al. [117] proposed a
non-probabilistic reliability method into service reliability evaluations
for arch dams which only requires the intervals for uncertain para-
meters. They applied this method on an arch dam according to the time-
varying characteristics of the material parameters and load.
While all the previous papers deal with reliability analysis of ex-
isting concrete dams, Chiti et al. [23] used the concept of reliability for
design optimization of gravity dams subjected to earthquake load. In
this method, the optimal shape of the dam is described by a number of
variables which are found by minimizing the total cost for the given
target reliability. Subset simulation with Markov Chain MCS was
adopted to estimate the accuracy of the failure probability with a
minimum number of samples. They claimed that this new method is
more appropriate than the deterministic approach for the optimal shape
design of concrete gravity dams.
1.3.2. SVM-based reliability analysis
The concept of support vector machine was first introduced as op-
timal margin classifiers in the context of statistical learning theory [11];
however, the term “support vector” was used for the first time by Cortes
and Vapnik [24]. Since then, SVM has been used in many real-world
data processing applications. The fundamental idea in SVM is relatively
straightforward: finding the optimal hyperplane (there might be more
than one) that separates clusters in such a way that the samples with
one category of the target variable falls in one side of the plane while
samples with the other category are on the other side. The vectors near
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
the hyperplane are called the “support vectors”. The overview is sche-
matically shown in Fig. 2 and will be explained later in detail (See
Section 3).
SVM concept has been used in many fields [82] as well as in
structural engineering [20]. Of particular interest in the context of this
paper is its application in structural reliability and estimation of the
failure probability in engineering structures.
Rocco and Moreno [100] used SVM concept for fast structural re-
liability; Li et al. [69] introduced the MCS-based and FORM-based SVM
methods in reliability; Guo and Bai [38] used the LS-SVM in reliability
analysis; Guo and Bai [37] combined the regression with LS-SVM
concepts; Dai et al. [26] proposed an adaptive Markov chain simulation
method for structural reliability, while Wang et al. [125] developed an
optimization technique based on Markov chain Monte Carlo method;
Wei et al. [127] applied particle filtering to estimate the SVM-based
model parameters; Basudhar and Missoum [7] applied the probabilistic
SVM concept to calculate the failure probability; Dai et al. [25] pro-
posed a multi-wavelet support vector regression for efficient reliability
assessment; Jiang et al. [55] generated uniform support vectors to filter
the structural failure functions; das Chagas Moura et al. [28] used the
SVM concept for failure prediction of time series data; and finally, Zhu
and Zhao [138] discussed the applicability of MCS-based and FORM-
based LS-SVM in structural reliability.
Beside the theoretical implementation of the SVM-family in the
structural reliability, different researchers showed real-world applica-
tions. Among them, Zhao [137], Li et al. [70] and Kang et al. [56] used
SVM for slope stability and reliability analysis; Ni et al. [83], Cheng and
Hoang [22] and Liu et al. [72] applied SVM to bridge system; Wang
et al. [126] used LS-SVM for settlement prediction of soft clayey
foundation; Zhao et al. [136] applied LS-SVM to evaluate the tunnel
reliability; Khatibinia et al. [59] used wavelet weighted LS-SVM for
reinforced concrete frames with soil-structure interaction; and finally,
Noori et al. [85] used SVM method for structural health monitoring and
reliability evaluation of structural systems.
1.3.3. SVM concept applied to concrete dams
History of the machine learning methods especially SVM-family in
dam engineering problems goes back no more than a decade. However,
the majority of case studies are applied in the past three years. Su et al.
[119] proposed an early-warning model for dam safety. First, a rough
set theory was used to implement the extraction of main influencing
factors from original monitoring data. Next, based on the problems
existing in build early-warming models of dam safety, SVM method was
applied to establish a model.
Dancea et al. [27] proposed an interesting computer vision tech-
nique in order to enhance the visual inspection process of large concrete
dams. This method allows detection and quantification of calcite de-
posits on dam wall surface (which is a clear sign that water infiltrates
within the dam body). The proposed scheme for identification of cal-
cite/non-calcite areas on the color image of dam wall consists of clas-
sifying the pixels into three classes, using a modified fuzzy c-means
279
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
Fig. 2. Fundamental concept of SVM-based classifica-
tion.
algorithm. The weight for the calcite class is determined using support
vector regression.
Gu et al. [36] proposed the use of least squares support vector
machine (LS-SVM) in back analysis of RCC dams and determination of
the complex mechanical properties. First, the initial samples are uni-
formly designed and then a transversely isotropic model is established
to train the samples. Finally, the complex nonlinear relationship be-
tween relative values of hydraulic components of dam displacements
and mechanical parameters is established. Su et al. [118] also applied a
similar idea for gravity dams. They proposed a criterion for optimal
selection of parameters in back analysis. The parameter sensitivity was
taken as a key index of optimal selection. The uniform design method
was combined with an artificial neural network and SVM to build the
mapping relationship between multiple material parameters and dam
structural responses at different positions.
Cheng and Zheng [21] proposed two multivariate dam safety
monitoring models. One of them is based on latent variables that can be
calculated from the separation process of the environmental and noise
effects. The LS-SVM model was adopted to simulate the nonlinear
mapping. The latent variables are predicted, and the prediction interval
is calculated to provide a control range for the future monitoring data.
They used this model to analyze the monitoring data of the horizontal
displacement and hydraulic uplift pressure of a RCC dam.
Ranković et al. [98] developed a support vector regression model
(nonlinear autoregressive model with exogenous inputs) to forecast the
tangential displacement of a concrete dam. A very large data set was
used for training and testing the model. The model parameters, the
kernel function, the regularization parameter and the tube size of
ϵ -insensitive loss function were specified by the trial-and-error method.
Li et al. [68] adopted the strength reserve factor method to simulate
gradual failure of RCC dams. The concept of entropy and catastrophe
theory were used to obtain the ultimate bearing resistance and the
failure criterion. First, the most dangerous sliding plane is found using
the Latin hypercube sampling (LHS) and auxiliary analysis of partial
least squares regression. Next, the LS-SVM is used to determine the
safety coefficient of a RCC dam.
Salazar et al. [103] and Salazar et al. [104] critically reviewed some
of the machine learning based predictive models for dam safety as-
sessment. Models based on random forests, boosted regression trees,
neural networks, support vector machines and multivariate adaptive
regression splines where discussed and fitted to predict several target
variables. The prediction accuracy was compared with the conventional
statistical model.
Su et al. [116] proposed some algorithms to improve the capacity of
SVM-based model monitoring to account for the time dependency of the
material and load. To describe dynamically the time-varying mapping
relationship between dam structural behavior and its cause, a method
was developed to update the proposed model in real time by making the
most use of the new observations.
Finally, Salajegheh and Khosravi [102], Khatibinia and Khosravi
[60] and Mahani et al. [75] adopted the SVM to optimal shape design of
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

concrete gravity and arch dams. In all cases, the coupled dam-reservoir- local linear approximation of the LS function in a transformed space,
foundation system was excited by a ground motion record and the and 2) an iterative search for the “design point”. The general solution to
computed and estimated responses were cross-validated. find the safety index was proposed by Hasofer and Lind [47]. The first
step to define the Hasofer-Lind reliability index, βHL , is to normalize the
2. Classical structural reliability basic RVs, X , by transferring them into a standard normal vector, U :
Xi − μ Xi
The limit state presented in Eq. (3) generally has a deterministic Ui =
σXi (11)
nature. It is possible to re-write the limit state function in the most
general probabilistic form as: where μ Xi and σXi are the mean and standard deviation of the random
Z = G (X ) = R (X ) − S (X ) (5) variable Xi . Note that in this new transferred space μUi = 0 and σUi = 1.
Note that the LS function G (X ) in the original space will take the
where X ⊂ M is a random vector of N basic variables form G (U ) in the standard normal space. The Hasofer-Lind reliability
X = X1 , X2 , …, XN . Again, G (X ) ≤ 0 corresponds to failure, while index is defined as the shortest distance from the origin in the standard
G (X ) > 0 represents the safe region. normal space to the new transferred failure surface G (U ) , (this point is
The failure probability, Pf , can be determined using the time in- called design point, U *):
variant reliability analysis as (the convolution integral):
U * = argmin{ u , G (u ) ≤ 0}
∞
Pf =  [R (X ) < S (X )] = ∫−∞ FR (t ) fS (t ) dt u⊂M (12)
∞
= 1 − ∫−∞ FS (t ) fR (t ) dt (6) and subsequently,
βHL = U * = min ( U ) (13)
where the randomness of R and S is expressed by probability density
functions (PDFs) fR and fS , while FR and FS are the corresponding cu- and the “exact” FORM based failure probability for the linear LS
mulative density functions (CDFs). functions can be computed as:
The probability of failure in terms of limit state function can be
expressed as: P fFORM = Φ (−βHL) (14)

Pf =  [G (X ) ≤ 0] = ∫G≤0 fR (R) fS (S ) dRdS (7)

For the general nonlinear case, Eq. (12) may be reformulated as an
optimization problem with Lagrange multiplier, λ , as:
Also, it is possible to determine the failure probability in terms of
1 2
joint PDF of all contributing variables as: L (u; λ ) = u + λG (u )
2 (15)
Pf = ∫{x : G (x )≤0} fX (x ) dx (8) The solution of this equation requires that L (u; λ ) be minimized
with respect to primal variables u and maximized with respect to λ :
Since the integration domain in this equation is only implicitly
available, the direct estimation of the failure probability is very difficult ⎧ ∂L (U *, λ*) = 0
(and usually impossible). However, it is possible to introduce a binary ⎪ ∂u U * + λ*∇G (U *) = 0
→ ⎧
(safe/fail) classifier as: ⎨ ∂L (U *, λ*) ⎨
⎩G (U *) = 0
⎪ =0
⎩ ∂λ (16)
1 if G (x ) ≤ 0
If (x ) = ⎧
⎨ This nonlinear constrained optimization problem can be solved
⎩ 0 if G (x ) > 0 (9)
using the algorithm introduced first by Rackwitz and Flessler [97]
Consequently, the failure probability can be estimated by calcu- where the iterative linearized solution continues around the current
lating the expectation of the introduced classifier, Pf =  [If (X )] [76]. point until it converges with desired accuracy. Detailed formulation of
Thus, the challenge is how to compute/estimate the value of the Pf . this procedure is beyond the scope of this paper; however, the relia-
There are many detailed books on theory and application of the bility index in the optimal fashion can be formulated as [96]:
structural reliability, among them Ditlevsen and Madsen [32], Melchers
[81], Thoft-Cristensen and Baker [122], Tichy [123] and Lemaire [67]. ∂G (U *)
In this paper some of the widely-used methods are adopted and applied βHL =− ∂u ·U * = α *·U *
to the case studies. They are: first-order reliability method (FORM) ∂G (U *)
(Section 2.1) and second-order reliability method (SORM) (Section ∂u (17)
2.2), for the cases in which the LS function is expressed explicitly with
and the associated failure probability is computed again by Eq. (14). αi2
analytical solution, and crude Monte Carlo simulation (MCS) (Section
(independently for each of the ith random variable) is a metric to
2.3), MCS with importance sampling (IS) (Section 2.4) and MCS-LHS
evaluate the local sensitivities [76].
(Section 2.5) for the cases with an implicit LS function.
The above-mentioned mathematical models only present the failure
probability of a single failure mode (e.g., dam sliding or overturning); 2.2. SORM
However, for the structural systems with multiple potential failure
modes (PFMs), the reliability must be considered as a system of the Since the FORM is linearly approximated the LS function at the
failure modes [84]. The system reliability for the ideal series and par- design point, it may over- or underestimate the reliability index for the
allel systems can be formulated as: functions with significant curvature [128]. In such a case the second
order reliability method may be used, where the LS function is ap-
⎧ P f =  [∪{ Gk (x ) ≤ 0}] =  [min { Gk (x ) ≤ 0}]
s
proximated by a quadratic function at the design point (e.g., by a Tailor
⎨ Pfp =  [∩{ Gk (x ) ≤ 0}] =  [max { Gk (x ) ≤ 0}] (10)
series expansion). The approximation is such that the two functions
⎩
have the same tangent hyperplane and second order derivatives at the
design point [74].
2.1. FORM Different methods with different level of complexity and accuracy
were proposed for SORM; however, Breitung [13] derived a simple but
First order reliability method is based on the combination of 1) a practical formula on the basis of asymptotic analysis as:

280
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

fX (x )
P fIS = ∫{x : G (x )≤0} hX (x )
hX (x ) dx
(23)
With the similar analogy of Eq. (19), the failure probability can be
approximated as:
Nsim
1 fX (xj )
Pf =
IS
Nsim
∑ If (xj )
hX (xj )
j=1 (24)
An appropriate choice of hX (x ) is the key point in successful im-
plementation of IS method. Using the FORM results is one method
Fig. 3. COV-Pf -Nsim relationship in crude MCS. which is described in Melchers [81]. Others proposed some guides and
observations on the application of this method in structural reliability
N −1 [14,80,79,51] and [3].
P fSORM = Φ (−βHL) ∏ (1 + βHL κi)−1/2
i=1 (18)
2.5. MCS-LHS
where κi are the principal curvatures of the paraboloid. This equation is
asymptotically exact (as βHL → ∞ while βHL κi is kept constant) for any Latin hypercube sampling was first introduced by McKay et al. [78]
LS function having a single design point [30]. and further extended by many researchers. Among them Iman and
Conover [52], Olsson and Sandberg [86], Owen [88], Stein [115],
2.3. Crude MCS Huntington and Lyrintzis [50], and Ayyub and Lai [6].
LHS guarantees samples to be drawn over the whole range of the
Monte Carlo simulation is used to direct calculation of the failure distribution and proceeds as follows. Given a system with basic RVs,
probability based on joint PDF of all the random variables. Since it is X = (X1, …, XN ) and corresponding distributions D1, …, DN , first the
based on the theory of large numbers, Nsim → N∞, an unbiased esti- range of each variable is split into Nsim non-overlapping intervals of
mator of Pf is given by: equal marginal probability 1/ Nsim . Then, sampling starts with the
random selection of an interval followed by another random selection
Nsim
1 Nfail of a point inside it, Fig. 4. The procedure is repeated until all intervals
Pf
MCS
=
Nsim
∑ If (xj ) =
Nsim have been accessed, and none of them more than once. This procedure
j=1 (19)
is repeated for each of the N RVs. So far, the maximum number of
where Nfail is the number of failed samples, and the hat is the sign of combinations for an LHS with Nsim divisions and N RVs can be com-
estimation. puted as:
The confidence intervals for Pf
MCS
are: Nsim − 1 N −1
LHS
Ncombo = ⎛∏ (Nsim − i) ⎞ = (Nsim !) N − 1
α α
Pf ∈ ⎡Pf ⎛1 + COV MCSΦ−1 ⎛ ⎞ ⎞, Pf ⎛1 − COV MCSΦ−1 ⎛ ⎞ ⎞ ⎤
MCS MCS MCS
⎝ i=0 ⎠ (25)
⎢
⎣ ⎝ ⎝ 2 ⎠ ⎠ ⎝ ⎝ 2 ⎠⎠⎥
⎦
Olsson et al. [87] explained a simple method to construct a M × N
(20)
matrix of samples. Note that there is a risk that some spurious corre-
where Φ (.) is the standard normal CDF and α ∈ [0, 1] is used to cal- lation will appear among the RVs, where some correlation reduction
culate the bounds with confidence level of 1 − α . Convergence of an methods can be adopted to solve this problem. Similar idea can be used
MCS estimator is usually assessed by the coefficient of variation (COV), to apply any desired level of correlation between the RVs. The calcu-
lation of the failure probability, Pf , follows the identical analogy as
LHS
Fig. 3:
explained in crude MCS, Eq. (19). The user should be aware of two
1 − Pf
MCS
points in using the MCS-LHS method:
COVMCS =
Nsim × Pf
MCS
(21)
• The number of samples should be larger than number of RVs
Finally, the associated generalized MCS based reliability index can (Nsim > N ), if a specific correlation is need to be applied among
be expressed as: them.
• Although the MCS-LHS can be used for ordinary failure probabilities
βMCS = −Φ−1 (Pf ) = Φ−1 (1 − Pf )
MCS MCS
(22) (> 10−1); for “small failure probabilities” (10−2 − 10−5 ) still relatively
large number of simulations are required, Fig. 3. In such problems,
limited number of sampling will not reflect the real failure prob-
2.4. MCS-IS ability, Fig. 4. One way to handle this problem is the subset simu-
lation which is used for rare events [5,4].
The crude MCS is in need of many samples as practically there is no
control on the sampled point. A lot of methods have been proposed to 3. Support vector machine
reduce the number of sampling and consequently, reduce the variance
of the responses. Importance sampling is one of these techniques which Support vector machine is among the best algorithms for solving
is originally proposed by Harbitz [39]. This method benefits both the problems in classification and regression. One important advantage of
advantages of crude MCS and FORM. SVM is that the determination of the model parameters can be cast as a
The fundamental idea in IS is to concentrate the distribution of convex optimization problem. Thus, any local solution for the model
sampling in the most important region [66]. A simple way is to move parameters is also a global optimum [12].
the sampling center from the origin in standard normal space to the Suppose a training data set consisting of n data points x1, …, xn in
design point on the failure function [108], Fig. 4. In this method, a new M , with corresponding class labels y1 , …, yn , where yi ∈ { −1, +1} are
sampling PDF, hX (x ) (known as sampling density function) is defined to given. The goal is to map the training data from the input space into a
obtain the samples in the desired region. Thus, the failure probability in higher dimensional feature space (maybe infinite) so that the mapped
Eq. (8) is reformulated as: data is linearly separable, Fig. 2. Therefore, by definition, there exists at

281
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
Fig. 4. Comparison of three simulation-based sampling methods.
least one separating hyperplane of the following form:
f (x) = wT φ (x) +b (26)
where the operator φ: M → H represents the transformation; w ∈ H
is the coefficient vector that determines the orientation of the hyper-
plane, and b ∈  is the bias term. An example of two hyperplanes are
shown in Fig. 2 for the linearly separable data. Among all potential
hyperplanes, the one with the largest margin is selected, where margin
is defined as the smallest distance between the hyperplane and any of
the data points. After determination of the model parameters w and b, a
new test data point x is classified according to the sign of f (x) .
3.1. Standard SVM
In this section, the mathematical formulation and optimization of
SVM in order to maximize the margin around the separating hyperplane
is explained. Since the mapped data is linearly separable, all the
training data points x1, …, xn satisfy the following constraints:
T
⎧ w φ (x i) + b ≥ +1 for yi =+ 1
⎨ wT φ (x i) + b ≤ −1 for yi = −1 (27)
⎩
and these two constraints can be combined into one set of inequalities:
yi (wT φ (x i) + b) − 1 ≥ 0 for all i = 1, …, n (28)
Under this setup, the points that are closest to the separating hy-
perplane, known as “support vectors” (red circles with cross mark in
Fig. 2), lie on the following hyperplane:
yi (wT φ (x i) + b) = 1 (29)
This hyperplane and the separating hyperplane in Eq. (26) are
parallel (they have the same orientation or coefficient vector w ) and
the distance between them is |1 − 0|/ ∥ w ∥ = 1/ ∥ w ∥, where ∥ w ∥ is
the Euclidean norm of w . Thus, the hyperplane which gives the max-
imum margin is obtained by maximizing ∥ w ∥−1, which is equivalent to
minimizing ∥ w ∥2 , subject to the constraints given in Eq. (28). More-
1
over, ∥ w ∥2 is often replaced with 2 ∥ w ∥2 for convenience in working
with quadratic forms.
In order to solve the above constrained optimization problem the
Lagrange multipliers method is used. It can be shown that the following
term should be maximized with respect to a = [a1, …, an]T ∈ n :
n n n
∼ 1
L (a) = ∑ ai − 2
∑ ∑ ai aj yi yj k (xi, xj)
i=1 i=1 j=1 (30)
n that the constraint set is different.
subject to the constraints: ai ≥ 0, i = 1, …, n and ai yi = 0 .
∑i = 1
Here, the “kernel function” is defined by , which represents the dot
product between two mapped data points φ (x i) and φ (xj) . The above
optimization problem falls under the category of convex quadratic
programming (QP) and it can be solved using standard techniques [12].
Given the optimal solution a* = [a1*, …, an*]T in the training procedure, a
282
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
new data point xtest is classified according to the sign of f (xtest) :
n
f (xtest) = ∑ ai*yi k (xtest, xi) + b*
i=1 (31)
where b* can be found by substituting the optimal value
n
w * = ∑i = 1 ai*yi φ (x i) in Eq. (29) and can be expressed in terms of the
kernel function.
An important property of the dual representation is that the separ-
ating hyperplane is optimized without ever having to explicitly com-
pute the coordinates of data points φ (x i) in feature space. Thus, em-
ploying kernel functions is an efficient approach to perform nonlinear
classification by “implicitly” mapping input data into feature space. The
two commonly used families of kernels are polynomial kernels and
radial basis functions (RBF) [95,94]. The polynomial kernel function is
of the form k (x i, xj) = (xTi xj + 1)d with parameter d ∈  . Under this
mapping, it can be shown that coordinates of φ (x i) contain the co-
ordinates of input data x i ∈ M as well as combinations of these M
coordinates [107]. On the other hand, RBF (also known as Gaussian
kernel) takes the form k (x i, xj) = exp(−γ ∥ x i − xj ∥2 ) in which the
mapped data point φ (x i) has infinite dimensionality [107].
In the derivation of SVM as presented so far, it is assumed that the
mapped data points φ (x1), …, φ (xn) are linearly separable. However, in
practice, it is often not possible to find exact separation of the training
data. For example, noise and outliers might be present in the training
data [48]. To make the algorithm work for these cases, the optimization
problem that finds the maximum margin classifier must be re-
formulated. Consider a set of n variables ξ1, …, ξn , then the constraints in
Eq. (28) are modified:
T
⎧ yi (w φ (x i) + b) ≥ 1 − ξi for all i = 1, …, n
⎨ ξi ≥ 0 for all i = 1, …, n (32)
⎩
1 n
In this case, the goal is to minimize ∥ w ∥2 + 2
λ ∑i = 1 ξi ,
where the
parameter λ > 0 controls the relative weighting between two objec-
tives: 1) making ∥ w ∥2 small to find maximum margin classifier, and 2)
ensuring that most of training data points have margin at least 1. To
solve this problem as explained before, Lagrangian function is con-
structed to find the dual representation. Consequently, the following
term should be maximized with respect to the vector a :
n n n
∼ 1
L (a) = ∑ ai − ∑ ∑ ai aj yi yj k (xi, xj)
i=1
2 i=1 j=1 (33)
n
subject to the constraints: 0 ≤ ai ≤ λ, i = 1, …, n and ∑i = 1 ai yi = 0 .
This optimization problem is identical to the separable case, except
3.2. Multi-class SVM
The support vector machine is designed to learn a maximum margin
classifier on data sets with binary labels (i.e., yi ∈ { −1, +1} ). However,
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

n
in practice, data sets might consist of K > 2 classes or labels. In this 1
∥ w ∥2 + λ ∑ (ξi + ξ ′i ).
case, multiple two-class SVMs should be constructed in order to achieve 2 (41)
i=1
a multi-class SVM classifier. To be specific, for a data set with K > 2
This optimization problem involves maximizing the following term
K
()
classes, a set of 2 = K (K − 1)/2 different two-class SVMs should be
with respect to a = [a1, …, an]T ∈ n and a′ = [a′1, …, a′n ]T ∈ n [10]:
constructed to cover all possible pairs of classes. For example, for K = 3
which corresponds to a data set with 3 labels, yi ∈ {1, 2, 3} , the total n n
∼ 1
number of possible pairs equals to three: (1, 2) , (1, 3) , and (2, 3) . This L (a , a′) = −
2
∑ ∑ (ai − a′i )(aj − a′ j ) k (x i, xj)
i=1 j=1
technique is called “one-versus-one” and more details are referred to n n
[10]. − ϵ ∑ (ai + a′i ) + ∑ yi (ai − a′i )
i=1 i=1 (42)
3.3. Least squares SVM
subject to the constraints: 0 ≤ ai ≤ λ, 0 ≤ a′i ≤ λ forall i = 1, …, n
n
and ∑i = 1 (ai − a′i ) = 0 .
Least squares SVM is a generalization of the standard SVM, where Given the optimal solution of the dual representation, the response
constraints in Eq. (32) are replaced by the equality constraints: value for a new test point x is estimated via the following expression:
yi (wT φ (x i) + b) = 1 − ei for all i = 1, …, n (34) n
f (x) = ∑ (ai − a′i ) k (x, xi) + b.
Note that the parameter e = [e1, …, en]T in the above constraints is ty- i=1 (43)
pically used in least squares data fitting. To find the maximum margin
This equation only requires finding the values of kernel function
classifier, the objective is to minimize:
k (x, x i) for i = 1, …, n where the commonly used kernels are linear,
n
1 1 polynomial, and radial basis kernel functions.
J (w, e) =
2
∥ w ∥2 + γ
2
∑ ei2
i=1 (35)
4. Case study
where the parameter γ > 0 controls the relative weighting between
1 n
2
∥ w ∥2 and the sum squared error, ∑i = 1 ei2 . Here, a = [a1, …, an] ∈ n Two examples are studied in this paper. The first one is a gravity
is the solution to the following set of linear equations [120]: dam analyzed analytically based on LEM approach. In this showcase,
T
the limit state function is presented explicitly and the material prop-
⎡0 −y ⎤ ⎡ b ⎤ = ⎡0⎤ erties as well as static loads are all assumed to be random variables. In
⎢ y Ω + γ −1I ⎥ ⎢ a ⎥ ⎢ 1 ⎥
⎣ ⎦⎣ ⎦ ⎣ ⎦ (36) the second example, the finite element model of a gravity dam-foun-
dation system is analyzed numerically subjected to a large set of real
where y = [y1, …, yn ]T ∈ n , 1 = [1, …, 1]T ∈ n , I is an n × n identity
ground motions. In this showcase, the seismic reliability is evaluated.
matrix, and Ω ∈ n × n is the kernel matrix defined by Ωij = yi yj k (x i, xj) .
Thus, the maximum margin classifier is found by solving the above
linear set of equations instead of quadratic programming. 4.1. Analytical model

This example explains the classic reliability analysis of gravity dams

3.4. SVM regression
based on LEM approach. In this method, the dam is assumed to be a
rigid body and the loads are applied in a static fashion. Thus, only the
So far, the applications of SVM in classification problems are ex-
magnitude and the moment arm of the resultant loads are important.
plained. In this section, SVMs are extended to regression problems.
No internal stress is evaluated in this method. The applied loads are: 1)
Regression is used to predict continuous values compared with classi-
self-weight, 2) static uplift pressure, 3) hydrostatic (based on normal or
fication, which is used to predict classes or labels (discrete values).
flood water levels) pressure, 4) silt pressure, 5) ice pressure, 6) wind
To be formal, a training data set consisting of n data points x1, …, xn ,
load, and 7) surface wave load. Note that calculation of nearly all these
with observed response values yi ∈  are given. Here, the goal is to find
loads is straight forward, expect the uplift pressure which is shown in
a linear model of the form:
detail in Fig. 5. Furthermore, Table 1 shows the parameters considered
f (x) = wT φ (x) + b (37) in this analytical showcase and their distributional model (See
Appendix A for the Monte Carlo sampling based input data selection).
where it is used to predict the response values. Hence, the following
Originally, four LS functions were considered: LS1) sliding at the
term should be minimized:
dam-foundation interface, LS2) overturning of the dam around the toe,
n LS3) sliding along the concrete-concrete interface at the neck, and LS4)
1
∥ w ∥2 + λ ∑ Eϵ (f (x i) − yi ) overturning of the neck around the downstream discontinuity.
2 i=1 (38)
However, to avoid the further complexity, only the results of LS1 will be
where the parameter λ > 0 controls the weighting between ∥ w and ∥2 reported. Note that the explicit LS functions follow from Eq. (4).
the error term, known as ϵ -insensitive error function, which is defined
as: 4.2. Numerical model

0 if |f (x) − y| < ϵ
Eϵ (f (x) − y ) = ⎧ This example, illustrates the earthquake transient analysis of a
⎩|f (x) − y| − ϵ otherwise
⎨ (39) gravity dam. The height of the dam is 121.92 m, thickness at the base
and the crest are 95.81 and 9.75 m, respectively. Upstream and
As shown before, the above optimization problem can be re-
downstream faces slopes are 0.05 and 0.78. Four-node and three-node
formulated by introducing two variables ξi ≥ 0 and ξ ′i ≥ 0 . The mod-
plain strain elements are used for modeling the dam body and the
ified constraints are:
foundation. Nonlinearity originates only from the zero-thickness in-
⎧ yi ≥ f (x i) − ϵ − ξ ′i terface elements for modeling the discrete joints at the dam-foundation
⎨ interface [19]. The concrete smeared cracking is not considered. The
⎩ yi ≤ f (x i) + ϵ + ξi (40)
finite element code “Merlin” is used for dynamic analyses [106].
for all i = 1, …, n and the objective function in Eq. (38) should be Applied loads on the system are: 1) self-weight, 2) hydrostatic
modified as well: pressure, 3) dynamic uplift which is automatically updated with crack

283
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Fig. 5. LEM based model description (loads and geometry).

propagation, and 4) seismic loads. Hydrodynamic pressure is modeled
based on Westergaard [130] added mass approach on both the dam
upstream face and the foundation, Fig. 6 (this is to reduce the com-
putational effort in transient analysis). Detailed fluid-structure inter-
action effects on concrete dams can be found in Hariri-Ardebili et al.
[46]. Viscous boundary condition is modeled on the foundation far-end
edges to absorb the outgoing (S- and P-) waves. Furthermore, the elastic
boundary is modeled on one side of the foundation to prevent the rigid
body motion. Dynamic analysis is “restarted” (only displacement is
reseted to zero without altering the initial state of stress) after the initial
static one to facilitate visualization of the deformed shape [106].
Damping is modeled based on mass- and stiffness-proportional Rayleigh
coefficients assuming a 5% critical damping for the system.
All the material properties, geometry and loading (except ground
motion) are kept constant (no epistemic uncertainty). Material prop-
erties for concrete are: mass density ( ρc ) 2250 kg/m 3 , modulus of
elasticity (Ec ) 22.4 GPa, Poisson's ratio (νc ) 0.20, tensile strength ( f ′t )
2.24 MPa. Foundation material properties are: mass density ( ρf )
2600 kg/m3, modulus of elasticity (Ef ) 24 GPa, Poisson's ratio (νf ) 0.25.
Interface joint properties are: tangential stiffness (kt ) 224 GPa, normal
stiffness (kn ) 224 GPa, cohesion (c) 1.9 MPa, friction angle (ϕf ) 38°,
dilatancy angle (ϕd ) 20°, specific mode I and II fracture energies (GFI ,
GFII ) 252/2520 N/m, relative irreversible deformation (γ ) 0.3, max-
imum displacement for dilatancy (uDmax ) 0.01 m, tensile stress and
opening displacement at break point < s1σ , COD> is
< 0.56 MPa, 1.12E − 4 m>, cohesion and sliding displacement at break
point < s1c , CSD> is < 0.375 MPa, 1.26E − 3 m>.
The system is excited only in the horizontal direction. Thus, ground
motion record-to-record variability is the only source of uncertainty in
the system (aleatory uncertainty). Hariri-Ardebili and Saouma [43]
provided a comprehensive list of possible ground motion intensity
measure parameters for concrete dams where up to 70 parameters were
compared based on efficiency, proficiency, sufficiency, practicality and
hazard compatibility. In this paper only the basic ones are used,
Table 2. The selected list includes nearly all the characteristics of a
single ground motion which may affect the structural responses. Each
parameter is quantified by a mathematical definition. These intensity
measures are calculated for one hundred ground motions and in each

Table 1
Parameters considered for the LEM analytical limit state function example.

Parameter Symbol Unit Distributional model Quantity Bound

Width at the base B1 m – 60 –

Width at the crest B2 m – 5 –
Height of the dam H1 m – 80 –
Height of the neck H2 m – 5 –
Location of drainage Ld m Uniform U(5, 15) –
Pre-existing crack at the base Lcr m Lognormal LN (13, 0.4) (0, 30)
Pre-existing crack at the neck Lcrn m Lognormal LN (0.5, 0.4) (0, 1)
Height of the water Hw m Lognormal LN (67, 18) (30, 100)
Height of the silt layer Hs m Normal N(10, 7) (1, 30)
Height of the surface wave Hsw m Normal N(1, 0.5) (0.1, 2)
Concrete mass density ρc kg/m3 Normal N(2400, 100) (2200, 2600)
Water mass density ρw kg/m3 – 1000 –
Silt mass density ρs kg/m3 Normal N(1850, 50) (1750, 2000)
Rock-concrete cohesion crc MPa Lognormal LN (0.62, 0.17) (0.2, 2)
Rock-concrete friction angle ϕrc deg. Normal N(30, 7) (15, 45)
Concrete-concrete cohesion ccc MPa Lognormal LN (0.71, 0.13) (0.5, 2)
Concrete-concrete friction angle ϕcc deg. Normal N(34, 6) (20, 45)
Drain efficiency effD – Uniform U(0.01, 0.99) –
Silt internal friction angle ϕs deg. Normal N(30, 8) (15, 45)
Wind pressure Pwind Pa Uniform U(100, 150) –
Ice pressure Pice Pa Normal N(300, 150) (100,1500)
Ice thickness tice m Normal N(0.4, 0.3) (0.01, 1)

284
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Fig. 6. FEM based model description (loads and boundary con-

ditions).

Table 2
Summary of the selected ground motion intensity measures for gravity dams.

Description Symbol Unit Mathematical definition Fitted model Quantity

Moment magnitude Mw – – Normal N(6.31, 0.62)

Rapture distance Rrup km – Lognormal LN (38.6, 25.6)
Significant duration tsig s t0.95IA − t0.05IA Lognormal LN (14, 8.8)
Peak ground acceleration PGA g max( u¨ (t ) ) Lognormal LN (0.12, 0.07)
Peak ground velocity PGV m/s max( u˙ (t ) ) Lognormal LN (0.062, 0.036)
Peak ground displacement PGD m max( u (t ) ) Lognormal LN (0.019, 0.016)
Root-mean-square of acceleration aRMS m/s2 Lognormal LN (0.147, 0.086)
1
ttot
∫0ttot (u¨ (t ))2dt
Root-mean-square of velocity vRMS m/s Lognormal LN (0.011, 0.007)
1
ttot
∫0ttot (u˙ (t ))2dt
Root-mean-square of displacement uRMS m Lognormal LN (0.006, 0.007)
1
ttot
∫0ttot (u (t ))2dt
Arias intensity IA m/s π
2g
∫0ttot (u¨ (t ))2dt Lognormal LN (0.114, 0.07)

First-mode spectral acceleration Sa (T1) g Sa (T1, ξ = 5%) Lognormal LN (0.124, 0.087)

First-mode spectral velocity Sv (T1) m/s Sv (T1, ξ = 5%) Lognormal LN (0.114, 0.076)
First-mode spectral displacement Sd (T1) m Sd (T1, ξ = 5%) Lognormal LN (0.009, 0.007)
Combined 1–2) spectral acceleration Sa1 − to − 2 g 2 mieff Lognormal LN (0.223, 0.07)
∑i = 1 (Sa (Ti, ξ ))αi , αi =
∑12 mieff
Combined 1–3) spectral acceleration Sa1 − to − 3 g 3 mieff Lognormal LN (0.325, 0.066)
∑i = 1 (Sa (Ti, ξ ))αi , αi =
∑13 mieff

Table 3 5.1. Full analytical model

Comparison of classical reliability methods; full analytical model.
In this case, all the RVs shown in Table 1 are used and LEM is
Method FORM SORM Crude MCS MCS-LHS MCS-IS
performed with LS1. First, the classical reliability analysis methods are
Failure probability (Pf ) 2.27e-2 1.95e-2 1.77e-2 1.78e-2 1.79e-2 performed. Table 3 shows the failure probability and reliability index
Reliability index (β ) 2.000 2.065 2.104 2.102 2.099 obtained from classical methods. As seen, all the methods provide a
β − βMCS −4.9% −1.9% 0.0 −0.1% −0.2% good estimation of the failure probability (compared to the reference
Err (%) = × 100%
βMCS
method; crude MCS). FORM and SORM overestimate Pf ; however, the
reliability index is only 4.9% and 1.9% less than crude MCS. All three
case an appropriate distributional model is fitted to the raw data his- simulation-based methods lead to nearly identical results; however, the
tograms. number of required simulations, Nsim , for the crude MCS, MCS-LHS and
MCS-IS decreases from 1e6 to 1e5 and 1e4, respectively. Fig. 7 com-
5. Results and discussion pares the convergence of these three methods. In each plot the failure
probability and the confidence intervals are shown versus the number
This section presents the reliability analysis of concrete dams based of simulations. Based on this figure, MCS-LHS and MCS-IS reach enough
on three examples: stability even with 50,000 and 5000 simulations, respectively.
It is also possible to calculate the local sensitivity of the parameters,
• Full analytical model (Section 5.1) Eq. (17). Based on the FORM results, the following RVs contribute
• Reduced analytical model (Section 5.2) 99.99% of the total results: 1) Hw= 56.55%; 2) crc= 14.03%; 3) Lcr =
• Finite element model (Section 5.3) 12.84%; 4) ϕrc= 10.51%; 5) eff D= 5.37%; 6) ρc = 0.51%; 7) Ld = 0.10%;
8) Hs= 0.08%; and other RVs nearly zero.
Next, the application of support vector machine in estimation of

285
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Fig. 7. Estimated failure probability and the confidence intervals for three simulation-based methods; full analytical model.

failure probability is investigated. To train binary SVM classifier, 1) a ̂ (i) = −1}| re-
responses for the test data points Xtest . Finally, |{i: y SVM
set of Ntrn data points in 17 that contains 17 RVs in Table 1, and 2) the presents the total number of failed cases based on the prediction made
corresponding responses (+ 1 for safe and − 1 for fail) are assumed to be by the trained SVM.
given. Here, these Ntrn responses are calculated explicitly using the
Algorithm 1. Estimation of Pf using SVM classifier.
analytical solution. The training data can be viewed as a matrix
Xtrn ∈  Ntrn × 17 , where each row represents a set of material properties
and/or static loads. The responses (safe/fail) for these Ntrn simulations Input: Xtrn ∈  Ntrn × 17 , ytrn ∈  Ntrn , Xtest ∈  Ntest × 17
are also stored in a column vector ytrn ∈  Ntrn . Output: estimated failure probability Pf
SVM

The goal is to learn a classifier from the training data that allows us Training SVM classifier:
to predict responses (safe/fail) for a new set of material properties and/ 1: for i = 1, …, 17 do
or static loads without using the analytical model. Similar to the 2: μi =MEAN(Xtrn (:,i) )
training step, the test data can be stored in a matrix Xtest ∈  Ntest × 17 ,
3: σi=STD(Xtrn (:,i) )
where each row corresponds to the set of 17 RVs. Then, the responses ∼
4: Xtrn (:,i) ← (Xtrn (:,i) − μi )/ σi
are predicted for these Ntest cases using the trained SVM classifier and
stored in a column vector y SVM ̂ ∈  Ntest . Finally, the failure probability 5: end for
6: ∼
can be estimated as the number of failed cases, i.e., number of entries of svm − model←FITCSVM(Xtrn , ytrn )
̂
y SVM that are equal to − 1, normalized by the total number of cases. Estimation via trained SVM:
Furthermore, there is an important technical aspect in training the 7: for i = 1, …, 17 do
SVM classifier. Note that 17 RVs used in this example have different 8: ∼
Xtest (:,i) ← (Xtest (:,i) − μ )/ σi
i
range of values. For example, parameter ρc takes values from
9: end for
(2200, 2600) while tice takes values from (0.01, 1) . Therefore, it is im- ∼
10: ̂ ←PREDICT(svm − model, Xtest )
y SVM
portant to standardize RVs before solving the optimization problem to
learn the SVM classifier [48] (this step is similar to Eq. (11) by trans- 11: Pf
SVM
̂ (i ) = − 1}|/ Ntest
= |{i : y SVM
ferring data to the standard normal space). To standardize RVs or col-
umns of Xtrn in our framework, we should find the sample mean and
standard deviation for each column. The columns are then centered to
So far, the general formulation and implementation of SVM classi-
have mean 0 and scaled to have standard deviation 1. Obviously, the
fiers are discussed to estimate the failure probability. In the following,
columns of test data Xtest should be standardized using the same sample
the various aspects of training SVM classifiers is investigated on the
mean and standard deviation.
accuracy of estimated failure probability. The estimation error is de-
The overall training and estimation procedure is summarized in
fined as the ratio |Pf
SVM
− Pf |/ Pf , where Pf is the reference value for the
Algorithm 1. In order to make the implementation of the proposed
failure probability which is computed earlier in this section (see
method easier for interested readers, MATLAB's built-in functions are
Table 3).
used in the given algorithm. Xtrn (:,i) denotes the i-th column of matrix
First, the effect of number of training points Ntrn is studied for a
Xtrn . The function FITCSVM() is used to solve the optimization problem
certain polynomial kernel function. Fig. 8 shows the mean and standard
discussed in Section 3 and the returned variable “ svm − model ” con-
deviation of the estimation error over 50 runs for various values of Ntrn .
tains parameters such as the coefficient vector and the support vectors.
In each run, a set of Ntrn points in 17 are generated to train the SVM
These parameters are then used in the function PREDICT() to estimate
classifier. The value of Ntrn varies from 1e 4 to 1e5 and from 1e3 to 1e 4 in

Fig. 8. Standard SVM-based Pf estimation using polynomial

kernel function.

286
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
Crude MCS and MCS-IS, respectively. Then, a set of Ntest = 1e6 (Crude
MCS) and Ntest = 1e 4 (MCS-IS) are used to estimate the failure prob-
ability as described in Algorithm 1. As can be seen in Fig. 8, the mean
and standard deviation of Pf estimation error decrease as the number of
training points, Ntrn , increases. This is expected since more training data
often leads to a more accurate classifier. Moreover, the average error is
less than 2% for Ntrn = 40, 000 (Crude MCS) and Ntrn = 4, 000 (MCS-IS).
Thus, it is confirmed that the MCS-IS technique provides better training
data in the sense that the required number of training data points is
decreased by an order of magnitude compared to the Crude MCS. This is
consistent with the results obtained from classical reliability analysis
techniques in which MCS-IS has the advantage over the crude MCS for
small probability problems.
So far, the polynomial kernel function was used to compare dif-
ferent sampling techniques. Next, the impact of kernel function type is
investigated on the training set and estimated failure probability. Thus,
varying number of training data points, Ntrn , are generated based on
MCS-IS technique. Fig. 9(a) compares the mean and standard deviation
of Pf estimation error for polynomial and linear kernel functions. It is
found that the polynomial kernel function leads to more accurate
classifiers for all values of Ntrn . Specifically, for smaller values of Ntrn ,
the advantage of polynomial kernel function over the linear kernel is
more significant. Fig. 9(b) shows the advantageous of polynomial
kernel function compared to RBF kernel where the latter is not a sui-
table kernel to be used for LEM based Pf estimation.
Up to now, all the training was conducted using the standard SVM
and it is found that RBF kernel function leads to larger error compared
to the others. Thus, it is reasonable to test other SVM family methods to
seek possible improvement in the estimated results. Fig. 10 compares
the standard SVM and LS-SVM only for the RBF kernel function. As
seen, LS-SVM decreases the Pf estimation error considerably compare to
the standard version. However, the main disadvantage of LS-SVM is
that it takes longer than the standard one to train the classifier. In fact,
it takes more than a few hours to train the SVM classifier using LS-SVM
for Ntrn = 1e 4 . Thus, we have omitted this value from Fig. 10. In con-
clusion, the standard SVM with the polynomial kernel function has the
best performance in terms of higher accuracy and lower run-time (often
less than a minute).
5.2. Reduced analytical model
Section 5.1 presented the results of the classical reliability method
and the SVM-based Pf estimation for a full analytical model of gravity
dam. However, for most of the existing dams part of the parameters
introduced as random variable are already known (e.g., location of
drainage, mass density, etc). Thus, it is possible to reduce the dimension
of the uncertainty in the problem to only “sensitive unknown” RVs.
This example is similar to the previous one; however, only limited
number of parameters are assumed to be RVs:
• First, three RVs are considered: c , ϕ and H .
rc rc w
• Second, two RVs are considered: c and ϕ . rc rc
287
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
Fig. 10. Standard SVM vs. LS-SVM for full analytical model; MCS-IS, RBF kernel.
The parameters other than those above-mentioned are kept constant
because they are either not important in overall failure probability
(based on previous FORM calculations) or the purpose of the study is to
assess the dam reliability subjected to a specific load combination. The
constant parameters are: Ld = 15 m , Lcr = 10 m , Lcrn = 0.5 m , Hw = 75 m
(whenever it is assumed to be constant), Hs = 15 m , Hsw = 1.0 m ,
ρc = 2400 kg/m3 , ρs = 1850 kg/m3, ccc = 0.7 MPa, ϕcc = 34 deg.,
effD = 0.5, ϕs = 30 deg., Pwind = 150 Pa, Pice = 300 Pa, t ice = 0.5 m .
Furthermore, the impact of correlation between the two RVs (e.g.,
crc and ϕrc ) is investigated on the failure probability. Correlation is ac-
counted using Iman and Conover [52] method. Fig. 11(a) shows the
sampling pattern of the two RVs with different correlation coefficients,
ρ1 − 2 (note that both RVs are truncated, one is normal and the other
lognormal). Fig. 11(b) shows the reliability index as a function of water
level and correlation coefficient. Crude MCS with Nsim = 1e6 simula-
tions is used. As seen, increasing both Hw and ρ1 − 2 reduces the relia-
bility index; however, the variation of β is more intense with respect to
Hw . Moreover, under the lower water levels, the impact of ρ1 − 2 is
higher.
Next, the water level is kept constant (Hw = 75 m ) and the system
reliability is assessed using both the classical methods and SVM. Two
extreme cases are considered: 1) no correlation ( ρ1 − 2 = 0 ), and 2)
strong correlation ( ρ1 − 2 = 0.95). Table 4 compares five classical relia-
bility methods with these extreme boundaries. As seen, in all methods,
accounting for the correlation increases the failure probability in a way
that fully uncorrelated model has about 4–7% failure probability
compared to strongly correlated model. This is also schematically
shown in Fig. 12(a) only for MCS-IS. This plot reveals a nonlinear re-
lation between Pf and ρ1 − 2 .
Finally, a set of Ntrn = 1e 4 training data points are generated based
on MCS-IS technique. Standard SVM optimization in conjunction with
the polynomial kernel function is used to find the separating hyper-
plane as well as the support vectors. Figs. 12(b) and (c) show the op-
timized hyperplanes for the uncorrelated and correlated cases, respec-
tively. Both the hyperplanes have a linear format which is consistent
with the physics of the problem (where both the cohesion and friction
angle have linear relation in the limit state function).
Fig. 9. Impact of kernel function type on the estimated Pf ; MCS-
IS, standard SVM.
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Fig. 11. Impact of correlation on the reliability index; reduced model with 3 RVs.

Table 4 Moreover, using more than 100 real ground motions for a specific
Impact of random variables correlation using different classical reliability methods. site/structure is not common as the quality of selected ground mo-
tions are reduced [53]. For each of those single transient analyses,
Method FORM SORM Crude MCS MCS-LHS MCS-IS
the results can be presented in the form of input parameter (which is
Pf (Uncorrelated RVs) 7.66e-4 3.54e-4 4.31e-4 4.38e-4 4.69e-2 one of the IMs in Table 2) and the output parameter which can be
β (Uncorrelated RVs) 3.169 3.387 3.332 3.328 3.308 classified into two groups:
Pf (Correlated RVs) 1.09e-2 6.81e-3 1.04e-2 1.04e-2 1.04e-2 – Local parameters such as crest displacement, base joint opening,
β (Correlated RVs) 2.298 2.467 2.310 2.310 2.311 and base joint sliding.
PfUnCorr / P Corr 7.0 5.2 4.1 4.2 4.5
f (%) – Global parameters such as damage index (DI). In the present study
the DI is defined as a ratio of the base crack to the total base length
[43]. Thus, due to the finite number of interface elements at the
base the DI will have a discrete nature in the range of [0, 1] (zero
5.3. Finite element model corresponds to no cracking and one for the fully cracked or failed
model).
Considering that the nonlinear transient finite element analysis of
coupled dam-foundation-reservoir is computationally expensive, a hy-
brid method is proposed which is a combination of numerical analyses
• Machine learning techniques: The performed limited finite element
analyses are then used for training an algorithm and the future
and machine learning techniques: predictions. The proposed method works in three levels: first the
non-failure analyses are separated from the failed ones based on
• Numerical analyses: Any modern seismic assessment of concrete binary (Section 5.3.1) and multi-class (Section 5.3.2) classification
dams starts with probabilistic seismic hazard analysis (PSHA) of the methods and next, the non-failure outputs are expanded based on
site and selection of appropriate set of ground motions [9,40,131]. SVM-regression method (Section 5.3.3).
There are several methods that the selected ground motion records
can be applied to the finite element model, e.g. incremental dynamic
analysis, multiple stripe analysis or cloud analysis (this is, in fact, 5.3.1. Detecting failure from non-failure
based on the required outputs for the decision making). In the Considering the available nonlinearity in the finite element model,
present study 100 ground motions are selected and applied as-re- there is a possibility of system failure (or even numerical instability) in
corded to the structural system (cloud analysis technique). Please the model under intense seismic events. In such a case the “local”
note that due to computational demand for each nonlinear transient outputs usually have very large values (if not infinity) compared to the
analysis, we had limitations in generating a comprehensive data set. non-failure models (note that DI is unit for the failed cases). Thus, it is

Fig. 12. Impact of random variables correlation on the failure probability; reduced model with 2 RVs, MCS-IS.

288
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
important to first distinguish between the failure and non-failure si-
mulations.
In this section, support vector machines are employed to learn a
classifier that allows prediction of the output (failure/non-failure) for a
set of ground motion intensity measures. This set contains 15 RVs that
can be viewed as data points in 15 (see Table 2). To measure the ac-
curacy of learned classifier, 100 simulations are divided into two sub-
sets: 1) training set and 2) test set. The training set is used to fit the SVM
classifier and the test set is used for assessment of the generalization
error of the final chosen model. From 83 non-failure cases, 73 data
points are used for training and the remaining 10 data points are used
for testing. Also, 12 out of 17 failure cases are chosen for training and
the remaining 5 data points are used for testing. Thus, the total number
of training data points is Ntrn = 85 and the test set consists of Ntest = 15
data points.
The training and test data can be represented as matrices
Xtrn ∈  Ntrn × 15 and Xtest ∈  Ntest × 15, where each row corresponds to a set
of ground motion intensity measures. The output (failure/non-failure)
for every training data is also stored in a column vector ytrn ∈  Ntrn . The
overall training and prediction procedure is summarized in Algorithm
2. We emphasize that random variables, i.e., columns of training and
test data, must be standardized to have zero mean and unit standard
deviation.
Algorithm 2. Binary and multi-class classification
Input: Xtrn ∈  Ntrn × 15, ytrn ∈  Ntrn , Xtest ∈  Ntest × 15
Output (Binary): estimated output (failure/non-failure) y SVM̂
Output (Multi-class): estimated output (moderate/severe/near-
collapse) y SVM̂
Training SVM classifier:
1: for i = 1, …, 15 do
2: μi =MEAN(Xtrn (:,i) )
3: σi=STD(Xtrn (:,i) )
4: ∼ linear and RBF based SVMs detect 8 correct cases. Thus, the polynomial
Xtrn (:,i) ← (Xtrn (:,i) − μi )/ σi
5: end for
∼ better performance compared to the others. Three pairs of two-class
6: Binary : svm − model←FITCSVM(Xtrn , ytrn ) or Multi-class :
∼
svm − model←FITCECOC(Xtrn , ytrn )
Estimation via trained SVM:
7: for i = 1, …, 15 do
8: ∼ 5.3.3. Regression of non-failure results
Xtest (:,i) ← (Xtest (:,i) − μi )/ σi
9: end for
∼ values of three local output quantities will be investigated (i.e., crest
10: y SVM
̂ ←PREDICT(svm − model, Xtest )
The accuracy of trained SVM classifier is investigated for various
kernel function types. Based on simulations, the trained SVM using
polynomial kernel function classifies 15 test cases successfully. Thus,
the learned SVM model using the polynomial function is able to detect
outputs for all Ntest = 15 test cases. However, trained SVMs using linear
and RBF kernel functions make correct predictions on 11 and 10 test
cases, respectively. This means that the use of polynomial function in
SVM outperforms both linear and RBF kernel functions based on overall
rate of correct classification.
Moreover, a set of experiments are performed on 2 RVs, instead of
15 initial ones, to get some intuition and visualize the decision
boundary. In each case, one of the RVs is set to be Sa (T1) (since first-
mode spectral acceleration is the optimal IM parameter in most en-
gineering structures) and the other RV is selected from {PGA, IA, Sv (T1)}
(other candidates for optimal IM based on Hariri-Ardebili and Saouma
[43]). Next, an SVM classifier in conjunction with the polynomial
kernel function is trained on 100 data points in 2 . Recall that 83 of
these data points belong to “non-failure” class and the rest belongs to
“failure” one. In Fig. 13, these 100 data points as well as decision
289
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
boundary and support vectors obtained from the SVM training are
plotted. Note that the non-linear decision boundary corresponds to a
separating hyperplane in the feature space induced by the nonlinear
polynomial kernel function. It is found that all three cases show rea-
sonable classification of two classes (failure/non-failure). As seen,
classification based on < Sa (T1), Sv (T1) > pair is more meaningful (more
smoothed hyperplane) compared to the other pairs as there is a high
correlation between these two IM parameters. They both depend on the
spectral responses at the fundamental mode. On the other hand, clas-
sification based on < Sa (T1), IA > pair shows highly scattered data points
and a curvy hyperplane as IA depends on acceleration time history and
its duration.
5.3.2. Detecting damage index of non-failure models
For non-failure models, the values of damage index belongs to the
set {0.15, 0.25, 0.35, 0.45, 0.55, 0.65, 0.75} . Again, the discrete nature of
this global output is enforced by the finite number of interface joints
elements at the dam-foundation interface. Having the damage index, it
is possible to classify the dam response based on predefined limit state
values. Hariri-Ardebili [41] proposed a model where DI = 0.3 and 0.6
are the thresholds separating the “moderate”, “severe” and “near-col-
lapse” damage states. These two thresholds are employed in this study
to split the outputs into three classes which correspond to low-to-
moderate, severe and near-collapse damages. Therefore, given a set of
training data points in 15, the goal is to learn a three-class SVM clas-
sifier that makes predictions for new test data points. This multi-class
classification process can be implemented by replacing FITCSVM() with
FITCECOC() in Algorithm 2.
Based on Section 5.3.1, from 83 non-failure simulations, Ntrn = 73
data points in 15 are used as training data and the remaining Ntest = 10
points are used for test data. The impact of kernel function type is in-
vestigated for polynomial, linear, and RBF kernel functions. Based on
simulations, the SVM classifier trained using the polynomial function
makes correct predictions on 9 out of Ntest = 10 cases. However, the
kernel function based SVM performs quite well on this data and has
SVMs are shown in Appendix B (Fig. A.18) for the three IM parameters
discussed in Fig. 13 and three damage states (i.e., moderate, severe, and
near-collapse).
In this section, the performance of SVM regression on estimating
displacement, joint opening, and joint sliding). The goal is to learn a
regression model based on a training set that allows us to predict values
of these three quantities for new test data points. As discussed in
Section 5.3.1, from 83 non-failure simulations, Ntrn = 73 data points are
selected for training data and the remaining Ntest = 10 data points are
used for test data. For each quantity, a SVM regression is trained to
estimate the corresponding values. Let Xtrn ∈  Ntrn × 15 and
Xtest ∈  Ntest × 15 represent the training and test data. Moreover, the va-
lues of three output quantities for training data points are stored in
column vectors y trn (1) (2)
, y trn (3)
, and y trn in  Ntrn . Algorithm 3 summarizes the
overall training and estimation process for the SVM regression tech-
nique. The MATLAB's built-in function FITRSVM() is used to fit a SVM
regression model on the training data and the learned model is used to
estimate values of three output quantities for new test data points.
Algorithm 3. SVM regression for three output responses
Input: Xtrn ∈  Ntrn × 15 , y trn
(1) (2)
, y trn (3)
, y trn ∈  Ntrn , Xtest ∈  Ntest × 15
̂ , y (2)
SVM
Output: estimated output responses y (1) ̂ , y (3)
SVM
̂
SVM
Training SVM regression models:
1: for i = 1, …, 15 do
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Fig. 13. Binary classification of seismic responses with two RVs.

2: μi =MEAN(Xtrn (:,i) )
3: σi=STD(Xtrn (:,i) )
4: ∼
Xtrn (:,i) ← (Xtrn (:,i) − μi )/ σi
5: end for
6: for i = 1, 2, 3 do
7: ∼ (i)
svm − model(i)←FITRSVM(Xtrn , y trn )
8: end for
Estimation via trained model:
9: for i = 1, …, 15 do
10: ∼
Xtest (:,i) ← (Xtest (:,i) − μi )/ σi
Fig. 15. SVM regression for non-failure crest displacement simulations; in [mm].
11: end for
12: for i = 1, 2, 3 do
13: ∼
̂ i) ←PREDICT(svm − model(i), Xtest )
y (SVM
14: end for displacement highly depend on the elastic properties of the dam body
and partially on the base cracking.
Finally, the values of SVM-based responses are plotted versus the
Fig. 14 reports the normalized estimation error for three output FEM-based ones in Fig. 15 for crest displacement. Based on the previous
quantities (where the difference between SVM-based and FEM-based observation, polynomial kernel function is used to train SVM regression
responses is divided by the value of FEM-based response). Specifically, on all 83 non-failure simulations. It is observed that all these points are
the impact of kernel function type is investigated on the accuracy of the close to the so-called “line of equality” showing the accuracy of SVM
estimated quantities. Fig. 14(a) shows that the trained SVM regression regression. There are several metrics to measure the accuracy of the
using polynomial kernel function outperforms both linear and poly- prediction, some are quantified in Table 5. This shows a high accuracy
nomial kernel functions. For crest displacement, the median error for for the estimated data points based on all error metrics.
the polynomial kernel function is smaller than the median error for the
other two kernel functions. Thus, it is recommended to use the poly-
nomial kernel function to estimate values of crest displacement. How- 6. Summary and future research
ever, based on Figs. 14(b) and (c), it is found that the linear kernel
function is the optimal candidate to estimate values of the joint opening This paper presented a comprehensive research on reliability ana-
and sliding. A possible explanation for this observation is that the va- lysis of concrete gravity dams. The analyses were performed in two
lues of joint opening/sliding are highly dependent on the damage index levels: simplified analytical model which accounts for the hydrological
(dam-foundation base cracking) which has discrete value. Therefore, hazard and detailed finite element analysis subjected to seismic hazard.
this dependence can lead to poor performance of nonlinear techniques On the other hand, structural reliability was investigated based on two
since it becomes more challenging to find optimal parameters of poly- different tools: classical reliability analysis techniques (and more spe-
nomial and RBF kernel functions. Note that the values of the crest cifically first- and second-order reliability methods, importance

Fig. 14. SVM regression on the local seismic responses.

290
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Table 5 damage level). Thus, the complete set of 100 analyses required about 2
Quantifying the SVM regression accuracy. months computations on a Dell Workstation (Core i7 CPU, 8 GB RAM).
On the other hand, the runtime of the SVM models is in the order of
Error metric Mathematical definition Quality Quantity
evaluation couple hours.
Since most of the instant results and findings are discussed in the
Mean squared error ∑iN= 1 (yi − F (xi))2 Smaller value 4.732e-7 details within the manuscript, Fig. 16 proposes a step-by-step hybrid
MSE =
N
methodology based on SVM and FEM for seismic reliability analysis of
Root mean squared ∑iN= 1 (yi − F (xi))2
Smaller value 6.879e-4
error RMSE = concrete dams. This technique uses all three main characteristics of the
N
Coefficient of ∑iN= 1 (yi − F (xi))2 Larger value 0.989 SVM (regression, binary- and multi-class prediction) along with prob-
2
R =1−
determination ∑iN= 1 (yi − y )2
abilistic seismic hazard and structural analysis adopted from perfor-
Adjusted R2 2
Radj = R − (1 − R2)
2 p Larger value 0.989 mance based earthquake engineering (PBEE) [93].
N−p−1
Sum of squared error N
SSE = ∑i = 1 (yi − F (xi ))2 Smaller value 3.88e-5
1. PSHA First, probabilistic seismic hazard analysis is performed and
two sets of ground motions are selected based on seismicity of the
p = number of parameters of the model; −=mean.
dam site.
● A small group with up to 100 ground motions and a large set with
up to 1000 ones.
sampling, Monte Carlo simulation and Latin Hypercube sampling ● The smaller group may includes only the real ground motions
techniques) and machine learning techniques (and more specifically which reflect the actual characteristics of earthquake records.
support vector machine (SVM), least-squares SVM, multi-class SVM, ● The statistics of the small set is analyzed and the optimal in-
and support vector regression). tensity measure parameters are selected. These are later used for
Results were presented for different combinations of the material the larger set as well.
uncertainty, loading condition, and seismic input intensity. In the first ● Since the number of recorded ground motions are limited in the
example, different classical analyses techniques were contrasted and PEER [90] website, the large set of ground motions might be
efficiency of them were compared. Application of SVMs with different combination of real ground motions and the generated ones
kernel functions and sampling size were discussed. Impact of correla- [99,135].
tion, ρ1 − 2 , between the random variables on the failure probability was
discussed and showed that Pf increases with increase in ρ1 − 2 . In the 2. FEM An efficient finite element model is developed and the poten-
second example, a FEM-SVM based hybrid methodology was proposed tial failure modes are implemented by introducing appropriate
for reduction of number of nonlinear seismic analyses and reliability nonlinearities (discrete or smeared crack). The small set of selected
(and later fragility) assessment of concrete dams. Different comparative ground motions are applied to the numerical model and the results
analyses were performed and detailed in the paper. are extracted in two levels: local outputs (crest displacement and
As clear, the main advantage of machine learning techniques (SVM joint response) and global damage index.
in this case) instead of crude finite element method based probabilistic 3. SVM Support vector machine family are used to post-process the
simulations is the lower computational cost. Each of the nonlinear results, classification, training and regression in three levels:
transient time history analyses and their post-processing (in the second ● Level 1: Detecting the failed from non-failed simulation based on
example) takes 10–25 h (depending on the ground motion duration and global damage index and binary SVM training.

Fig. 16. Proposed FEM-SVM based hybrid methodology for reliability assessment of large coupled systems.

291
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
● Level 2: Categorizing the non-failure simulations in different pre-
defined damage states employing optimal intensity measure
parameters.
● Level 3: Predicting the contentious local responses for the non-
failure simulations and developing a meta-model for future re-
liability and fragility assessment of dam with any desired set of
input ground motions (the large earthquake set).
It is found that the family of SVM is an useful and effective tool for
classification, response prediction and reliability analysis of the con-
crete gravity dams with reasonable accuracy. Combining SVM with
classical FEM and reliability technique can be used for fast risk as-
sessment of group of dams and prioritizing the rehabilitation of them
considering limited budget.
The following remarks can be considered in future research:
• Despite many advantageous of SVM technique, it is still not a super
precise classification algorithm specially for “rare events”. One
reason can be attributed to the fact that determination of a clear
separation using linear functions in a projected high-dimensional
feature space is difficult to achieve. Applications of other machine
learning techniques such as artificial neural networks can be in-
vestigated to support the findings in this paper or to improve the
classification. Such an examples can be found in Fedele et al. [34],
Appendix A. Histograms for the loads and model parameters
Fig. A.17 shows the histograms for the loads and model parameters sampled based on crude Monte Carlo simulation technique.
Fig. A.17. Histograms for the crude MCS (1e6 uncorrelated sample points).
292
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
Karimi et al. [57] and Mata [77]. Moreover, hybrid techniques
based on kriging, Naive Bayes, K-Nearest Neighbors, K-Means,
Random Forest, etc. and their accuracy and runtime should be in-
teresting in structural and seismic reliability analysis of concrete
dams.
• The application of other hazard models (e.g. hydrological and aging)
on dam response and the effectiveness of the SVM family in data
regression and prediction. The interaction between the aging, hy-
drological and seismic can be accounted for (multi-hazard reliability
analysis models). Such a framework is already proposed by Hariri-
Ardebili [41] for the capacity function of dams and Ghosh and
Padgett [35] for fragility function of bridges.
• The current paper only considers the uncertainty in the material
properties and the applied loads. However, the case study dams are
deterministic. One may considers the sensitivity of the dam classes
[42] in SVM-based data prediction.
Acknowledgment
The first author would like to express his sincere appreciation to his
former advisor (and the current mentor), Professor Victor E. Saouma at
the University of Colorado Boulder for his enthusiastic guidance and
advice throughout this research.
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki Soil Dynamics and Earthquake Engineering 104 (2018) 276–295

Appendix B. Three pairs of two-class SVMs for seismic responses

Fig. A.18 shows a set of two-class data classification based on three seismic intensity measure parameters and three damage states. The major
observations are:

Fig. A.18. Multi-class classification of seismic responses with two RVs; first, second and third rows correspond to <PGA, Sa (T1) >, <IA, Sa (T1) > and <Sv (T1), Sa (T1) >, respectively.

• The predicted hyperplane has different forms; however, it seems that it has more linear or smoothed form for the data points between the
moderate and near-collapse damage states (minimum and maximum modes). This is shown in the second column.
• <S (T ), S (T ) > pair seems to be most appropriate intensity measure parameters for multi-class SVM as dispersion of data is minimum.
a 1 v 1

References Proceedings of the fifth international conference on structural safety and reliability
(ICOSSAR ’89); 1989. pp. 279–86, San Francisco, California, US.
[9] Bernier C, Monteiro R, Paultre P. Using the conditional spectrum method for im-
[1] Altarejos-Garcia L, Escuder-Bueno I, Morales-Torres A. Advances on the failure proved fragility assessment of concrete gravity dams in eastern canada. Earthq
analysis of the dam–foundation interface of concrete dams. Materials Spectra 2016;0(0). [null].
2015;8(12):8255–78. [10] Bishop CM. Pattern recognition. Mach Learn 2006:128.
[2] Altarejos-Garcia L, Escuder-Bueno I, Serrano-Lombillo A, de Membrillera-Ortuno [11] Boser, BE, Guyon, IM, Vapnik, VN. A training algorithm for optimal margin clas-
M. Methodology for estimating the probability of failure by sliding in concrete sifiers. In 5th ACM workshop on 784 computational learning theory (COLT); 1992.
gravity dams in the context of risk analysis. Struct Saf 2012;36–37:1–13. pp. 144–152. ACM.
[3] Au S, Beck JL. A new adaptive importance sampling scheme for reliability calcu- [12] Boyd S, Vandenberghe L. Convex optimization. Cambridge university press; 2004.
lations. Struct Saf 1999;21(2):135–58. [13] Breitung K. Asymptotic approximations for multinormal integrals. J Eng Mech
[4] Au S, Ching J, Beck J. Application of subset simulation methods to reliability 1984;110(3):357–66.
benchmark problems. Struct Saf 2007;29(3):183–93. [14] Bucher CG. Adaptive sampling - an iterative fast monte carlo procedure. Struct Saf
[5] Au S-K, Beck JL. Estimation of small failure probabilities in high dimensions by 1988;5(2):119–26.
subset simulation. Probab Eng Mech 2001;16(4):263–77. [15] Bury K, Kreuzer H. Assessing the failure probability of gravity dams. Int Water
[6] Ayyub, BM, Lai, K-L. Structural reliability assessment using latin hypercube sam- Power Dam Constr 1985;37(11):46–50.
pling, In: Structural Safety and Reliability; 1989. pp. 1177–1184. ASCE. [16] Carvajal C, Peyras L, Bacconnet C. On the loading/shear strength coupling in the
[7] Basudhar A, Missoum S. Reliability assessment using probabilistic support vector probabilistic formulation of the limit-state in shear for gravity dams. Eur J Environ
machines. Int J Reliab Saf 2013;7(2):156–73. Civil Eng 2010;14(3):283–301.
[8] Baylosis, R, Bennett, R. Safety assessment of an existing concrete gravity dam. In [17] Carvajal C, Peyras L, Bacconnet C, Bécue J. Probability modelling of shear strength

293
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
parameters of rcc gravity dams for reliability analysis of structural safety. Eur J
Environ Civil Eng 2009;13:91–119.
[18] Carvajal, C, Peyras, L, Bécue, J-P, Varon, C, Bacconnet, C, Clergue, D, et al.
Towards a probabilistic assessment of structural safety of gravity dams. In
Proceedings of the 14th german dam symposium, 7th european club of interna-
tional commission on large dams (ICOLD), Munich, Germany; 2007.
[19] Cervenka J, Chandra J, Saouma V. Mixed mode fracture of cementitious bimaterial
interfaces; Part II: numerical simulation. Eng Fract Mech 1998;60(1):95–107.
[20] Çevik A, Kurtoğlu AE, Bilgehan M, Gülşan ME, Albegmprli HM. Support vector
machines in structural engineering: a review. J Civil Eng Manag
2015;21(3):261–81.
[21] Cheng L, Zheng D. Two online dam safety monitoring models based on the process
of extracting environmental effect. Adv Eng Softw 2013;57:48–56.
[22] Cheng M-Y, Hoang N-D. Risk score inference for bridge maintenance project using
evolutionary fuzzy least squares support vector machine. J Comput Civil Eng
2012;28(3):04014003.
[23] Chiti H, Khatibinia M, Akbarpour A, Naseri H. Reliability-based design optimiza-
tion of concrete gravity dams using subset simulation. Int J Optim Civil Eng
2016;6(3):329–48.
[24] Cortes C, Vapnik V. Support-vector networks. Mach Learn 1995;20(3):273–97.
[25] Dai H, Zhang B, Wang W. A multiwavelet support vector regression method for
efficient reliability assessment. Reliab Eng Syst Saf 2015;136:132–9.
[26] Dai H, Zhang H, Wang W, Xue G. Structural reliability assessment by local ap-
proximation of limit state functions using adaptive markov chain simulation and
support vector regression. Comput-Aided Civil Infrastruct Eng 2012;27(9):676–86.
[27] Dancea, O, Tsatos, O, Gordan, M, Vlaicu, A. Adaptive fuzzy c-means through
support vector regression for segmentation of calcite deposits on concrete dam
walls. In Automation Quality and Testing Robotics (AQTR), 2010 IEEE
International Conference on, Volume 3; 2010. pp. 1–6. IEEE.
[28] das Chagas Moura M, Zio E, Lins ID, Droguett E. Failure and reliability prediction
by support vector machines regression of time series data. Reliab Eng Syst Saf
2011;96(11):1527–34.
[29] de Araújo J, Awruch A. Probabilistic finite element analysis of concrete gravity
dams. Adv Eng Softw 1998;29:97–104.
[30] Der Kiureghian A. Structural reliability methods for seismic safety assessment: a
review. Eng Struct 1996;18(6):412–24.
[31] Der-Kiureghian A, Ditlevsen O. Aleatory or epistemic? does it matter? Struct Saf
2009;31:105–12.
[32] Ditlevsen O, Madsen HO. Structural reliability methods, Volume 178. New York:
Wiley; 1996.
[33] Elishakoff I. Safety factors and reliability: friends or foes? Springer
Science & Business Media; 2012.
[34] Fedele R, Maier G, Miller B. Identification of elastic stiffness and local stresses in
concrete dams by in situ tests and neural networks. Struct Infrastruct Eng
2005;1(3):165–80.
[35] Ghosh J, Padgett J. Aging considerations in the development of time-dependent
seismic fragility curves. J Struct Eng 2010;136:1497–511.
[36] Gu C, Li B, Xu G, Yu H. Back analysis of mechanical parameters of roller compacted
concrete dam. Sci China Technol Sci 2010;53(3):848–53.
[37] Guo Z, Bai G. Application of least squares support vector machine for regression to
reliability analysis. Chin J Aeronaut 2009;22(2):160–6.
[38] Guo Z-w, Bai G-c. Classification using least squares support vector machine for
reliability analysis. Appl Math Mech 2009;30:853–64.
[39] Harbitz A. An efficient sampling method for probability of failure calculation.
Struct Saf 1986;3(2):109–15.
[40] Hariri-Ardebili M, Saouma V. Collapse fragility curves for concrete dams: com-
prehensive study. ASCE J Struct Eng 2016;142(10):04016075.
[41] Hariri-Ardebili MA. Performance based earthquake engineering for concrete dams
[Ph.D. thesis]. Boulder, CO: University of Colorado; 2015.
[42] Hariri-Ardebili MA. Analytical failure probability model for generic gravity dam
classes. Proc Inst Mech Eng Part O: J Risk Reliab 2017.
[43] Hariri-Ardebili MA, Saouma V. Probabilistic seismic demand model and optimal
intensity measure for concrete dams. Struct Saf 2016;59:67–85.
[44] Hariri-Ardebili MA, Saouma VE. Quantitative failure metric for gravity dams.
Earthq Eng Struct Dyn 2015;44:461–80.
[45] Hariri-Ardebili MA, Saouma VE. Seismic fragility analysis of concrete dams: a
state-of-the-art review. Eng Struct 2016;128:374–99.
[46] Hariri-Ardebili MA, Seyed-Kolbadi SM, Kianoush MR. FEM-based parametric
analysis of a typical gravity dam considering input excitation mechanism. Soil Dyn
Earthq Eng 2016;84:22–43.
[47] Hasofer AM, Lind NC. Exact and invariant second-moment code format(for relia-
bility analysis in multivariate problems). Am Soc Civil Eng, Eng Mech Div, J
1974;100:111–21.
[48] Hastie T, Tibshirani RJ, Friedman JH. The elements of statistical learning: data
mining, inference, and prediction. Springer; 2011.
[49] Horyna T. Reliability analysis of base sliding of concrete gravity dams subjected to
earthquakes (Ph.D. thesis). University of British Columbia; 1999.
[50] Huntington D, Lyrintzis C. Improvements to and limitations of latin hypercube
sampling. Probab Eng Mech 1998;13(4):245–53.
[51] Ibrahim Y. Observations on applications of importance sampling in structural re-
liability analysis. Struct Saf 1991;9(4):269–81.
[52] Iman R, Conover W. A distribution-free approach to inducing rank correlation
among input variables. Commun Stat-Simul Comput 1982;11(3):311–34.
[53] Jalayer F, De Risi R, Manfredi G. Bayesian cloud analysis: efficient structural
fragility assessment using linear regression. Bull Earthq Eng 2015;13(4):1183–203.
[54] Jeppsson J. Reliability-based assessment procedures for existing concrete
294
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
structures (Ph.D. thesis). Lund University; 2003.
[55] Jiang Y, Luo J, Liao G, Zhao Y, Zhang J. An efficient method for generation of
uniform support vector and its application in structural failure function fitting.
Struct Saf 2015;54:1–9.
[56] Kang F, Li J-s, Li J-j. System reliability analysis of slopes using least squares
support vector machines with particle swarm optimization. Neurocomputing 2016.
[57] Karimi I, Khaji N, Ahmadi M, Mirzayee M. System identification of concrete
gravity dams using artificial neural networks based on a hybrid finite element-
boundary element approach. Eng Struct 2010;32(11):3583–91.
[58] Kazemi, MR. Reliability based analysis and design of anchor retrofitted concrete
gravity dams. In Proceedings of the 13th World Conference on Earthquake
Engineering, 2004.
[59] Khatibinia M, Fadaee MJ, Salajegheh J, Salajegheh E. Seismic reliability assess-
ment of rc structures including soil-structure interaction using wavelet weighted
least squares support vector machine. Reliab Eng Syst Saf 2013;110:22–33.
[60] Khatibinia M, Khosravi S. A hybrid approach based on an improved gravitational
search algorithm and orthogonal crossover for optimal shape design of concrete
gravity dams. Appl Soft Comput 2014;16:223–33.
[61] Krounis, A. Uncertainty in sliding stability analyses of existing concrete gravity
dams with bonded concrete-rock interfaces, 2013.
[62] Krounis A. Sliding stability re-assessment of concrete dams with bonded concrete-
rock interfaces (Ph.D. thesis). KTH Royal Institute of Technology; 2016.
[63] Krounis A, Johan F. The influence of correlation between cohesion and friction
angle on the probability of failure for sliding of concrete dams. Risk analysis, dam
safety, dam security and critical infrastructure management. CRC Press; 2012. p.
75–80.
[64] Krounis, A, Johansson, F., et al. A comparison between two techniques for in-
cluding the influence of progressive sliding failure in structural reliability analyses
of concrete dams. In ISRM international symposium-proceedings of the 8th asian
rock mechanics symposium. International Society for Rock Mechanics, 2014.
[65] Krounis A, Johansson F, Larsson S. Effects of spatial variation in cohesion over the
concrete-rock interface on dam sliding stability. J Rock Mech Geotech Eng
2015;7(6):659–67.
[66] Lee Y, Hwang D. A study on the techniques of estimating the probability of failure.
J Chungcheong Math Soc 2008;21(4):573–83.
[67] Lemaire M. Structural reliability. John Wiley & Sons; 2013.
[68] Li B, Zhang Z, Liu Y, Yang S. Evaluation standard for safety coefficient of roller
compacted concrete dam based on finite element method. Math Probl Eng
2014:2014.
[69] Li H-s, Lü Z-z, Yue Z-f. Support vector machine for structural reliability analysis.
Appl Math Mech 2006;27:1295–303.
[70] Li S, Zhao H-B, Ru Z. Slope reliability analysis by updated support vector machine
and monte carlo simulation. Nat Hazards 2013;65(1):707–22.
[71] Liu, F, Li, H. Analysis of time-dependent reliability of concrete gravity dam da-
mage based on gray theory. In Proceedings of the 3rd International Conference on
Management, Education, Information and Control (MEICI 2015), Shenyang, China;
2015.
[72] Liu Y, Lu N, Yin X, Noori M. An adaptive support vector regression method for
structural system reliability assessment and its application to a cable-stayed
bridge. Proc Inst Mech Eng Part O: J Risk Reliab 2016;230(2):204–19.
[73] Lupoi A, Callari C. A probabilistic method for the seismic assessment of existing
concrete gravity dams. Struct Infrastruct Eng 2012;8:985–98.
[74] Madsen HO. First order vs. second order reliability analysis of series structures.
Struct Saf 1985;2(3):207–14.
[75] Mahani AS, Shojaee S, Salajegheh E, Khatibinia M. Hybridizing two-stage meta-
heuristic optimization model with weighted least squares support vector machine
for optimal shape of double-arch dams. Appl Soft Comput 2015;27:205–18.
[76] Marelli, S, Schobi, R, Sudret, B. Uqlab user manual - structural reliability.
Technical report, Chair of Risk, Safety and Uncertainty Quantification, ETH
Zurich. Report UQLab-V0.92-107; 2016.
[77] Mata J. Interpretation of concrete dam behaviour with artificial neural network
and multiple linear regression models. Eng Struct 2011;33(3):903–10.
[78] McKay M, Beckman R, Conover W. A comparison of three methods for selecting
values of input variables in the analysis of output from a computer code.
Technometrics 1979;21(2):239–45.
[79] Melchers R. Radial importance sampling for structural reliability. J Eng Mech
1990;116(1):189–203.
[80] Melchers R. Search-based importance sampling. Struct Saf 1990;9(2):117–28.
[81] Melchers RE. Structural reliability analysis and prediction. John Wiley & Son Ltd;
1999.
[82] Moguerza JM, Muñoz A. Support vector machines with applications. Stat Sci
2006:322–36.
[83] Ni Y, Hua X, Fan K, Ko J. Correlating modal properties with temperature using
long-term monitoring data and support vector machine technique. Eng Struct
2005;27(12):1762–73.
[84] Nikolaidis E, Ghiocel DM, Singhal S. Engineering design reliability handbook. CRC
Press; 2004.
[85] Noori M, Cao Y, Hou Z, Sharma S. Application of support vector machine for re-
liability assessment and structural health monitoring. Int J Eng Uncertain: Hazards
Assess Mitig 2010;2:3–4.
[86] Olsson A, Sandberg G. Latin hypercube sampling for stochastic finite element
analysis. J Eng Mech 2002;128(1):121–5.
[87] Olsson A, Sandberg G, Dahlblom O. On latin hypercube sampling for structural
reliability analysis. Struct Saf 2003;25(1):47–68.
[88] Owen AB. Controlling correlations in latin hypercube samples. J Am Stat Assoc
1994;89(428):1517–22.
M.A. Hariri-Ardebili, F. Pourkamali-Anaraki
[89] Papadrakakis M, Papadopoulos V, Lagaros ND, Oliver J, Huespe AE, Sánchez P.
Vulnerability analysis of large concrete dams using the continuum strong dis-
continuity approach and neural networks. Struct Saf 2008;30(3):217–35.
[90] PEER, Ground motion database. 〈http://ngawest2.berkeley.edu/〉. Last viewed,
2014 November 2014.
[91] Peyras L, Carvajal C, Felix H, Bacconnet C, Royet P, Becue J, et al. Probability-
based assessment of dam safety using combined risk analysis and reliability
methods - application to hazards studies. Eur J Environ Civil Eng
2012;16:795–817.
[92] Piliounis G, Lagaros ND. Reliability analysis of geostructures based on metaheur-
istic optimization. Appl Soft Comput 2014;22:544–65.
[93] Porter, K. An overview of PEER’s performance-based earthquake engineering
methodology. In Proceedings of the 9th international conference on applications of
statistics and probability in civil engineering (ICASP9), San Francisco, CA; 2003.
[94] Pourkamali-Anaraki F, Becker S. A randomized approach to efficient kernel clus-
tering. IEEE global conference on signal and information processing; 2016. pp.
207–11.
[95] Pourkamali-Anaraki F, Hughes S. Kernel compressive sensing. In IEEE interna-
tional conference on image processing; 2013. pp. 494–8.
[96] Puatatsananon W, Saouma V. Reliability analysis in fracture mechanics using the
first-order reliability method and monte carlo simulation. Fatigue Fract Eng Mater
Struct 2006;29(11):959–75.
[97] Rackwitz R, Flessler B. Structural reliability under combined random load se-
quences. Comput Struct 1978;9(5):489–94.
[98] Ranković V, Grujović N, Divac D, Milivojević N. Development of support vector
regression identification model for prediction of dam structural behaviour. Struct
Saf 2014;48:33–9.
[99] Rezaeian S, Der Kiureghian A. Simulation of synthetic ground motions for specified
earthquake and site characteristics. Earthq Eng Struct Dyn 2010;39(10):1155–80.
[100] Rocco CM, Moreno JA. Fast monte carlo reliability evaluation using support vector
machine. Reliab Eng Syst Saf 2002;76(3):237–43.
[101] Ruggeri, G. Sliding safety of existing gravity dams, final report. ICOLD European
Club; 2004.
[102] Salajegheh J, Khosravi S. Optimal shape design of gravity dams based on a hybrid
meta-heruristic method and weighted least squares support vector machine. Int J
Optim Civil Eng 2011;1(4):609–32.
[103] Salazar F, Morán R, Toledo MÁ, Oñate E. Data-based models for the prediction of
dam behaviour: a review and some methodological considerations. Arch Comput
Methods Eng 2015:1–21.
[104] Salazar F, Toledo M, Oñate E, Morán R. An empirical comparison of machine
learning techniques for dam behaviour modelling. Struct Saf 2015;56:9–17.
[105] Saouma V. Reliability based nonlinear fracture mechanics analysis of a concrete
dam; a simplified approach. Dam Eng 2006;16(3):219–41.
[106] Saouma, V, Červenka, J, Reich, R. Merlin finite element user’s manual. 〈http://
civil.colorado.edu/saouma/pdf/Software/users.pdf〉, 2010.
[107] Scholkopf B, Smola AJ. Learning with kernels: support vector machines, regular-
ization, optimization, and beyond. MIT Press; 2001.
[108] Schuëller GI, Stix R. A critical appraisal of methods to determine failure prob-
abilities. Struct Saf 1987;4(4):293–309.
[109] Schultz, MT, Gouldby, BP, Simm, JD, Wibowo, JL. Beyond the factor of safety:
Developing fragility curves to characterize system reliability. Technical report,
DTIC Document; 2010.
[110] Serrano-Lombillo A, Escuder-Bueno I, de Membrillera-Ortuño MG, Altarejos-
Garcia L. Methodology for the calculation of annualized incremental risks in
systems of dams. Risk Anal 2011;31(6):1000–15.
[111] Shi Z, Gu C, Zheng X, Qin D. Multiple failure modes analysis of the dam system by
means of line sampling simulation. Opt-Int J Light Electron Opt
2016;127(11):4710–5.
[112] Silva F, Lambe TW, Marr WA. Probability and risk of slope failure. J Geotech
Geoenviron Eng 2008;134(12):1691–9.
[113] Sørensen JD. Notes in structural reliability theory and risk analysis. Aalborg
295
Soil Dynamics and Earthquake Engineering 104 (2018) 276–295
University; 2004.
[114] Spross, J. A critical review of the observational method; 2014.
[115] Stein M. Large sample properties of simulations using latin hypercube sampling.
Technometrics 1987;29(2):143–51.
[116] Su H, Chen Z, Wen Z. Performance improvement method of support vector ma-
chine-based model monitoring dam safety. Struct Control Health Monit
2016;23(2):252–66.
[117] Su H, Li J, Wen Z, Fu Z. Dynamic non-probabilistic reliability evaluation and
service life prediction for arch dams considering time-varying effects. Appl Math
Model 2016;40(15):6908–23.
[118] Su H, Wen Z, Zhang S, Tian S. Method for choosing the optimal resource in back-
analysis for multiple material parameters of a dam and its foundation. J Comput
Civil Eng 2016;30:4.
[119] Su, H-z, Wen, Z-p, Gu, C-s. An early-warning model of dam safety based on rough
set theory and support vector machine. In Proceedings of the 5th international
conference on machine learning and cybernetics; 2006. pp. 3455–60, Dalian,
China; IEEE.
[120] Suykens J, Vandewalle J. Least squares support vector machine classifiers. Neural
Process Lett 1999;9(3):293–300.
[121] Tekie P, Ellingwood B. Seismic fragility assessment of concrete gravity dams.
Earthq Eng Struct Dyn 2003;32(2221):2240.
[122] Thoft-Cristensen P, Baker MJ. Structural reliability theory and its applications.
Springer Science & Business Media; 2012.
[123] Tichy, M. Applied methods of structural reliability, volume 2. Springer
Science & Business Media; 2012.
[124] Wang C, Zhang S-r, Sun B, Wang G-h. Methodology for estimating probability of
dynamical system's failure for concrete gravity dam. J Cent South Univ
2014;21:775–89.
[125] Wang, J, Katafygiotis, LS, Noori, MN. Transitional markov chain monte carlo si-
mulation for reliability-based optimization. In Safety, Reliability, Risk and Life-
Cycle Performance of Structures and Infrastructures- Proceedings of the 11th
International Conference on Structural Safety and Reliability; 2013a. ICOSSAR
2013.
[126] Wang Y, Zhao X, Wang B. Ls-svm and monte carlo methods based reliability
analysis for settlement of soft clayey foundation. J Rock Mech Geotech Eng
2013;5(4):312–7.
[127] Wei Z, Tao T, ZhuoShu D, Zio E. A dynamic particle filter-support vector regres-
sion method for reliability prediction. Reliab Eng Syst Saf 2013;119:109–16.
[128] Westberg, M. Reliability-based assessment of concrete dam stability; 2010.
[129] Westberg Wilde M, Johansson F. System reliability of concrete dams with respect
to foundation stability: application to a spillway. J Geotech Geoenviron Eng
2013;139(2):308–19.
[130] Westergaard H. Water pressures on dams during earthquakes. Trans Am Soc Civil
Eng 1933;98:418–33.
[131] Wieland M. Safety aspects of sustainable storage dams and earthquake safety of
existing dams. Engineering 2016;2(3):325–31.
[132] Xu Q, Chen J, Li J. A study on the functional reliability of gravity dam. Energy
Power Eng 2012;4(2):59–66.
[133] Xu Q, Chen J-Y, Zhao C, Li J, Yue H. Failure probability analysis of interlayer
sliding belts of gravity dams under seismic load. Open Mech Eng J 2014;8:85–91.
[134] Xu Q, Li J, Chen J. Probability analysis for the damage of gravity dam. Engineering
2011;3(4):312–21.
[135] Yamamoto Y, Baker JW. Stochastic model for earthquake ground motion using
wavelet packets. Bull Seismol Soc Am 2013;103(6):3044–56.
[136] Zhao H, Ru Z, Chang X, Yin S, Li S. Reliability analysis of tunnel using least square
support vector machine. Tunn Undergr Space Technol 2014;41:14–23.
[137] Zhao H-b. Slope reliability analysis using a support vector machine. Comput
Geotech 2008;35(3):459–67.
[138] Zhu C, Zhao H. Least square support vector machine for structural reliability
analysis. Int J Comput Appl Technol 2016;53(1):51–61.