Professional Documents
Culture Documents
Computers and Geotechnics: Fei Kang, Shaoxuan Han, Rodrigo Salgado, Junjie Li
Computers and Geotechnics: Fei Kang, Shaoxuan Han, Rodrigo Salgado, Junjie Li
a r t i c l e i n f o a b s t r a c t
Article history: This paper presents a system probabilistic stability evaluation method for slopes based on Gaussian pro-
Received 7 February 2014 cess regression (GPR) and Latin hypercube sampling. The analysis is composed of three parts. Firstly, Latin
Received in revised form 7 July 2014 hypercube sampling is adopted to generate samples for constructing the response surface. Then, based on
Accepted 17 August 2014
the samples, Gaussian process regression, which is a popular machine learning technique for nonlinear
Available online 6 September 2014
system modeling, is used for establishing the response surface to approximate the limit state function.
Finally, Monte Carlo simulation is performed via the GPR response surface to estimate the system failure
Keywords:
probability of slopes. Five case examples were examined to verify the effectiveness of the proposed meth-
Slope stability
Response surface
odology. Computer simulation results show that the proposed system reliability analysis method can
System reliability analysis accurately give the system failure probability with a relatively small number of deterministic slope sta-
Monte Carlo simulation bility analyses.
Gaussian processes Ó 2014 Elsevier Ltd. All rights reserved.
Computer experiments
http://dx.doi.org/10.1016/j.compgeo.2014.08.010
0266-352X/Ó 2014 Elsevier Ltd. All rights reserved.
14 F. Kang et al. / Computers and Geotechnics 63 (2015) 13–25
method to consider spatial variability of soil properties. Zhang where g(x) is a limit state function formulated by the uncertain
et al. [19] applied MCS to the identified representative slip surfaces variables to describe the limit state in the space of x, fx(x) is the joint
from a large number of potential slip surfaces to obtain the system probability density function (PDF).
failure probability. Li et al. [20] also developed a system reliability The operating state of the slope can be reflected by the limit
approach to evaluate the slope failure probability using represen- state function. If g(x) > 0, the slope will be safe, else if g(x) < 0 the
tative slip surfaces together with MCS. Cho [21] calculated the fail- slope will be a failure, and g(x) = 0 is the boundary between stable
ure probability of the slope based on multi-point first-order and unstable. The limit state function for probabilistic slope stabil-
reliability method, which gave the probability of the union of ity analysis is usually defined as
approximate events. The approaches mentioned above [19–21]
Z ¼ gðxÞ ¼ F s ðxÞ 1:0 ð2Þ
need identify the representative slip surfaces or failure modes from
a large number of potential slip surfaces. A straightforward way to where Fs is the factor of safety obtained by a deterministic slope sta-
get the system reliability is Monte Carlo simulation [22]. The disad- bility analysis method [31].
vantage is that the computational effort can be extremely intensive. From Eq. (1), we can see that the integral is carried out over the
Therefore, importance sampling [17], subset simulation [23], and failure domain. However, direct evaluation of the D-fold integral is
response surface methods [24] are adopted to improve the efficiency virtually impossible [32]. Therefore, approximate techniques, such
of MCS. Response surface methods can predict the factor of safety as first-order reliability methods (FORM), second-order reliability
instead of the limit equilibrium method or strength reduction tech- methods (SORM) and MCS have been developed to evaluate the
nique. However, except the work of Zhang et al. [24], other works did probability of failure.
not verify the effectiveness of MCS combined with response surface
methods on system probabilistic slope stability evaluation. It has 2.2. Reliability index for slope stability evaluation
previously been verified that a traditional polynomial response sur-
face is unable to approximate the limit states accurately for system The reliability index b is developed to evaluate the comparative
reliability analysis problems of soil slopes [24]. reliability of a system when the exact probability distribution func-
Gaussian processes (GPs) are Bayesian state-of-the-art tools for tion is not known. It is defined as [33]
discriminative machine learning, i.e., regression, classification and
lZ E½gðxÞ
dimensionality reduction [25]. Compared to artificial neural net- b¼ ¼ ð3Þ
rZ r½gðxÞ
works (ANN) and support vector machines (SVM), GPs possess sev-
eral attractions. For instance, training GPs is far simpler than ANN, where lZ = E[g(x)] and rZ = r[g(x)] are the mean value and standard
and unlike in SVM, one can adopt very flexible kernels, including deviation of the performance function. It measures how far the
several free parameters, since their optimization can be efficiently mean of the safety margin g(x) is from zero (assumed to be the fail-
done through maximum likelihood estimation [26]. The use of ure point) in units of number of standard deviations. In Eq. (3) the
Gaussian processes for modeling geotechnical engineering prob- limit state function is assumed normally distributed.
lems is a relatively recent development [27,28]. In this paper, a The relationship between failure probability and reliability
response surface method based on Gaussian process regression index can be expressed as
(GPR) [29] with Latin hypercube sampling (LHS) [30] is presented
Pf ¼ UðbÞ ¼ 1 UðbÞ ð4Þ
for system reliability analysis. GPR is adopted to establish a
response surface that represents the relationship between soil where U is the standard normal cumulative distribution function.
properties and the factor of safety. Training and test data for the Eq. (4) shows that there is a unique relationship between the reli-
model are obtained by LHS and uniform design. Response surface ability index (b) and the probability of failure (Pf).
established by GPR and LHS is integrated with MCS to system fail-
ure probability calculation of soil slopes. Five case examples are 2.3. System reliability analysis and Monte Carlo simulation
examined to verify the effectiveness of the proposed methodology.
Results are also compared with several other recently proposed In slope reliability analysis, there may present numerous slip
methods for system reliability analysis of slopes. surfaces or failure modes. The system reliability is that of all poten-
The rest of the paper is organized as follows. In Section 2, sys- tial slip surfaces, and the failure probability of a system will be lar-
tem reliability analysis of slopes is stated. In Section 3, Gaussian ger than that for any single slip surface [10]. Generally, the FORM
process regression is introduced. In Section 4, GPR-based MCS for approach can only obtain the failure probability associate with the
system reliability analysis is described. In Section 5, simulation probabilistic critical surface, which has the minimum reliability
results are presented. Discussions are given in Section 6, and in index among numerous slip surfaces. System failure probability
Section 7, conclusions are provided. calculation is now recognized as a very important problem in slope
reliability analysis [19].
2. System reliability analysis of slopes The system reliability of a slope can be obtained by Monte Carlo
simulation. MCS is known as a simple random sampling method or
2.1. Probabilistic slop stability analysis statistical trial method that makes realizations based on randomly
generated sampling sets for uncertain variables [34]. There are three
As a complementary approach to deterministic slope stability primary steps to perform MCS: (1) select a distribution type for the
analysis that can obtain a factor of safety, probabilistic slope stabil- random variable; (2) generate a sampling set from the distribution;
ity analysis can reflect the effect of uncertain variables. These (3) conduct simulations using the generated sampling set.
uncertain variables can be represented as a vector x = [x1, When MCS is adopted, the probability of failure is given approx-
x2, . . ., xD]. The variables in the vector are corresponding to uncer- imately by
tain soil parameters of a slope, and D is the number of uncertain Nf
variables been considered in the calculation. The failure probability Pf ¼ ð5Þ
Ns
Pf of the slope can be presented as the following integral
Z where Ns is the total number of trials or simulations and Nf is the
Pf ¼ PfgðxÞ 6 0g ¼ f x ðxÞ dx ð1Þ total number of cases where the failure has occurred. The coefficient
gðxÞ60 of variation of the estimator is [19,35]
F. Kang et al. / Computers and Geotechnics 63 (2015) 13–25 15
sffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 Pf The element Kij is the covariance between values of the latent
COVPf ¼ ð6Þ functions f(xi) and f(xj), and it encodes about the prior of our
Ns Pf
knowledge of nonlinear process among latent functions.
Although MCS has many merits, this approach has not received GPR is used to compute the predictive distribution of the func-
an overwhelming acceptance due to the excessive computational tion values f* at test points X ¼ ½x1 ; x2 ; . . . ; xm . A graphical model
effort that is required. It is noted that most civil engineering works representation of a GP is given in Fig. 1. In the figure, fi denotes f(xi).
have target failure probabilities between 103 and 105 [33]. A The set of latent function values fi indexed by the set of indices xi is
large number of realizations are required to estimate failure prob- fully connected. Each of the connections represents the correlation
abilities accurately in this range. The computation time can be pro- between two latent variables, which are defined by a covariance
hibitively high, especially when the structure exhibits nonlinear function [37].
behavior or the numerical model is rather complex. For example, The distribution of y conditioned on the values of f is given by
when g(x) involves a nonlinear finite-element analysis, large num- an isotropic Gaussian
bers of realizations may not be practical. The efficiency of MCS can
pðyjf ; XÞ ¼ Nðf ; r2n IÞ ð12Þ
be improved by importance sampling [17], subset simulation [23]
and response surface methods [24]. where I is the identity matrix.
From the property of the Gaussian distribution, we can get the
3. Gaussian process regression marginal distribution of y as
Z
Consider a data set S of n observations S = {(xi, yi)|i = 1, . . ., n}, pðyjXÞ ¼ pðyjf ; XÞpðf jXÞ df ¼ Nð0; K þ r2n IÞ ð13Þ
where xi is a D-dimensional input vector, and yi is a scalar output
or target. This set of input/output pairs will be referred to as sam- The joint distribution of the observed target values and the
ple points or experimental points. For the sake of convenience, the function values at the test locations under the prior can be written
inputs are aggregated into a matrix X = [x1, x2, . . ., xn]. The outputs as
are likewise aggregated, y = [y1, y2, . . ., yn]. The regression task is, " #!
given a new input x⁄, to obtain the predictive distribution for the y KðX; XÞ þ r2 I KðX; X Þ
N 0; ð14Þ
corresponding observation y⁄ based on S. f KðX ; XÞ KðX ; X Þ
A Gaussian process is a collection of random variables, any finite
number of which have a joint Gaussian distribution [29]. A Gauss- From (9), using the standard rules for conditioning Gaussians,
ian process is the generalization of a Gaussian distribution. While we can get the predictive distribution
the latter is the distribution of a random variable, the Gaussian pðf jX; y; X Þ Nðf ; covðf ÞÞ ð15Þ
process describes a distribution over functions. From function
space view, the Gaussian process f(x) can be determined by the f ¼ KðX ; XÞ½KðX; XÞ þ r2 I1 y
corresponding mean m(x) and covariance functions that are ð16Þ
defined as follows:
1
covðf Þ ¼ KðX ; X Þ KðX ; XÞ½KðX; XÞ þ r2 I KðX; X Þ ð17Þ
mðxÞ ¼ Eðf ðxÞÞ ð7Þ
The output estimates are made not only with an expected pre-
kðx; x0 Þ ¼ Eððf ðxÞ mðxÞÞðf ðx0 Þ mðx0 ÞÞÞ ð8Þ diction (mean) of latent function f⁄, but also with a measure of
uncertainty (variance).
where k(x, x0 ) is the covariance (or kernel) function evaluated at x
and x0 .
A Gaussian process f(x) can be represented as 3.2. Covariance functions
f ðxÞ GPðmðxÞ; kðx; x0 ÞÞ ð9Þ Selection of an appropriate covariance function for a particular
Usually, for notational simplicity we will take the mean func- problem is an important part of GPs, because the choice of covari-
tion to be zero and the offsets and simple trends can be subtracted ance functions determines the properties of sample functions
out before modeling [36]. drawn from the GP prior (e.g., smoothness, length scales, ampli-
tude, etc.). It is a basic similarity assumption that points with
3.1. GPs for regression inputs which are close are likely to have similar target values,
and thus training points that are near to a test point should be
In GPs, we assume that the relation between the input vector informative about the prediction at that point. Under the Gaussian
and the target is given by
yi ¼ f ðxi Þ þ e ð10Þ y1 x1*
process view, it is the covariance function that defines nearness or [40] is adopted to generate testing samples to test the generaliza-
similarity. We are free in our choice of covariance function, so long tion performance of the established GPR model. Compared to a
as the covariance matrices produced are always symmetric and random sampling method, the two methods can select training
positive semidefinite. We adopt the most commonly-used squared data cover the entire design space uniformly, especially when the
exponential covariance function with automatic relevance deter- number of variables is large and the number of training datasets
mination distance measure [29], is relatively small. Meanwhile, they can avoid repeating runs and
make good use of the resources.
1
kðxp ; xq Þ ¼ r2f exp ðxp xq ÞT Mðxp xq Þ
2
! 4.2.1. Latin hypercube sampling
2 1X D
ðxp;i xq;i Þ2 Latin hypercube sampling is a popular choice for performing
¼ rf exp 2
ð18Þ
2 i¼1 li computer experiments. Suppose that V is a hypercube with dimen-
sion D, the algorithm generates N samples in this hypercube V.
where M = diag(l)2, l = [l1, l2, . . ., lD]T. h = (ln l1, . . ., ln lD, ln rf) are the In the simplest version of this design, the continuous range of
hyperparameters. each variable is partitioned into N equally spaced intervals, each
interval for each variable is sampled exactly once, and the univar-
3.3. Training a Gaussian process model iate sample values are randomly matched across all the variables
to form the N sample points by randomly permuting each factor
The parameters of the kernel functions are called the hyperpa- column in the design [30,41]. The samples generated by this sam-
rameters of the Gaussian process. These hyperparameters can be pling method are distributed uniformly in the hypercube space.
learned by maximizing the log-likelihood of the training outputs Alternative approaches have also been suggested for mating
given the inputs. Considering y Nð0; K þ r2n IÞ, the log-marginal factor levels in LHS. Following the advice to avoid replicate design
likelihood of the hyperparameters is expressed as points, one popular option is the class of maximin LHS [41], in
1 1 1 n which Latin hypercube samples are iteratively generated to find
log pðyjX; hÞ ¼ yT ðK þ r2n IÞ y log K þ r2n I log 2p the best one according to the criterion that maximize minimum
2 2 2
ð19Þ distance between points.
Computer simulation experiments are needed to establish the 4.3. MCS using GPR-based response surface for system reliability
Gaussian process response surface model. Random sampling may analysis
be the most intuitive method for generating training and test sam-
ples [39]. Generally speaking, to achieve a valid response surface A polynomial of fixed degree, e.g., a quadratic polynomial, is
model, the data selected for training must be ‘‘representative’’ of unable to approximate the limit states with highly nonlinearity
the overall design space. In this paper, the LHS is adopted to gen- accurately. Therefore, it is preferable to use a more flexible Gauss-
erate samples for training the GPR model and uniform design ian process regression model, which has a form suited to the
F. Kang et al. / Computers and Geotechnics 63 (2015) 13–25 17
Step 3: Build datasets for GPR. Sample sets consist of the sample
x
0.4 surface.
Step 6: Calculate the probability of failure. MCS is adopted to cal-
0.2 culate the system failure probability based on the established
0
0 0.2 0.4 0.6 0.8 1 Start
x
1
(b)
Define the problem
1
0.6
Prepare data sets: Calculate the factor of safety
x2
18
0.75 Clay 1
1
10
5. Numerical examples
10
Clay 2
4
0
Hard layer To illustrate the efficiency of the proposed approach, five
-10 benchmark examples of probabilistic slope stability analysis prob-
0 10 20 30 40 50 60 70 80 90 lems were considered. Before running a GP model, the hyperpa-
Horizontal distance (m) rameters should be initialized [45]. According to experiences rf
was set as ln rf = 0, li was set as ln li = 0. Computer experiments
Fig. 4. Cross-section of example 1.
points were generated fall within the range [lx 3rx, lx + 3rx] in
example 1–4 [32,39], and [lx 4rx, lx + 4rx] in example 5. The
number of samples generated by LHS to establish the GPR response
surface was firstly tested as N = 10D, if the accuracy was not
Table 1 acceptable, then N = 15D would be adopted. The samples were
Statistical properties of soil parameters in example 1.
assessed by Bishop’s simplified method to obtain the values of fac-
Slope layers Unit weight, c (kN/m3) Undrained strength, cu (kPa) tor of safety. The number of test samples generated by uniform
Mean COV Distribution design was set as 20. The accuracy of GPR model for predicting
Clay 1 19.0 120 0.3 Lognormal
the factor of safety is measured by some statistical criteria, such
Clay 2 19.0 160 0.3 Lognormal as the mean absolute error (MAE), root mean square error (RMSE),
correlation coefficient (R).
500 400
450 350
400 300
350
250
300
cu2 (kPa)
cu2 (kPa)
200
250
150
200
100
150
100 50
50 0
0 -50
0 50 100 150 200 250 300 350 400 450 500 -50 0 50 100 150 200 250 300 350 400
cu1 (kPa) cu1 (kPa)
(a) (b)
Fig. 5. Points generated by MCS with 20,000 samplings under different distributions: (a) lognormal; (b) normal.
4 4
Case points Case points
3.5 Fs (calculated) = Fs (predicted) 3.5 F =F
s (calculated) s (predicted)
3 3
Predicted factor of safety
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.5 1 1.5 2 2.5 3 3.5 4
Calculated factor of safety Calculated factor of safety
(a) (b)
Fig. 6. Comparison between values of factor of safety predicted by GPR model and calculated by the limit equilibrium method in example 1: (a) training data; (b) test data.
F. Kang et al. / Computers and Geotechnics 63 (2015) 13–25 19
Table 2 5
Performance of the GPR-based response surface in example 1. 20 2
Embankment 1 1
10
Data sets MAE RMSE R
Elevation (m)
2.5 Layer 1
Training 0.0116 0.0148 0.9998 10
Testing 0.0543 0.0677 0.9973 Foundation Layer 2
10
(Soft soil)
0
Hard Layer
-8
5.1. Example 1 0 10 20 30 40 50 60 70 80
Horizontal distance (m)
The first example is a slope passing through two clay layers,
Fig. 7. Cross-section of example 2.
taken from the work of Ching et al. [17]. The geometry of the slope
is shown in Fig. 4 and the soil parameters are given in Table 1.
Points generated by MCS with 20,000 samplings under different
distributions are shown in Fig. 5. This example was also analyzed Table 4
Statistical properties of soil parameters in example 2.
by Low et al. [12], Zhang et al. [24], Ji and Low [14] and Cho [21].
In this example, Bishop’s simplified method and the ordinary Layer 1 Layer 2
method of slices yield the same results, because the shear c1 (kPa) u1 (°) c2 (kPa) u2 (°)
strengths are characterized by cohesion cu with friction angle
Distribution Normal Normal Normal –
uu = 0. The factor of safety based on mean values of soil properties Mean 10 12 40 0
was found to be 1.993, which is very close with the value 1.992 COV 0.2 0.25 0.2 0
obtained by Cho [21]. The deterministic critical slip surface passes c (kN/m3) 20 18
through Clay 1 and 2, while the probabilistic critical surface
searched by the FORM only passes through Clay 1 [21]. There are
two representative failure modes or slip surfaces as identified by
the value of MCS. As shown in Table 3, the failure probability
Low et al. [12], Zhang et al. [24], and Cho [21]. The first one passes
obtained by the proposed method is close to the value obtained
through Clay 1 and the second one passed through both layers.
by MCS and several other system reliability analysis methods.
When N = 10D, the obtained failure probability Pf is 0.52% using
the proposed methodology with 100,000 samplings, which is quite
far from the value obtained by MCS method. Therefore we increase 5.2. Example 2
in the number of LHS samples to N = 15D = 30. The comparison
between values of factor of safety calculated by Bishop’s simplified This example is taken from Chowdhury and Xu [11], and it con-
method and predicted by GPR model is shown in Fig. 6. The accu- cerns a fill embankment resting on a clay layer. The geometry of
racy of the GPR model is presented in Table 2. From Table 2 and the slope is shown in Fig. 7 and the soil parameters are given in
Fig. 6, it can be seen that the GPR model perform well to predict Table 4. This example was also analysed by Ji and Low [14], Zhang
the factor of safety. et al. [19] and Cho [21]. The factor of safety based on mean values
Results of probabilistic analysis by different methods for this of soil properties was found to be 1.159, which is very close with
example are listed in Table 3. It can be seen that the single-point the value 1.164 obtained by Cho [21]. The deterministic critical slip
FORM analysis yielded a probability of failure much smaller than surface passes through the embankment fill, while the probabilis-
the failure probabilities obtained by system reliability analysis tic critical surface searched by the conventional FORM passed
methods. Therefore single-mode FORM analysis significantly through both embankment and soft clay foundation [21]. There
underestimated the failure probability, and MCS methods provided are two representative failure modes or slip surfaces as identified
unbiased estimates for the failure probability. Meanwhile, classical by Ji and Low [14], Zhang et al. [19], and Cho [21]. The first one
RSM is not able to approximate the nonlinear performance func- passes through the embankment fill and the second one passed
tion accurately, since the obtained failure probability is far from through both embankment and soft clay foundation.
Table 3
Results of probabilistic analysis by different methods for example 1.
2 2
Case points Case points
Fs (calculated) = Fs (predicted) Fs (calculated) = Fs (predicted)
Predicted factor of safety
1 1
0.5 0.5
0 0
0 0.5 1 1.5 2 0 0.5 1 1.5 2
Calculated factor of safety Calculated factor of safety
(a) (b)
Fig. 8. Comparison between values of factor of safety predicted by GPR model and calculated by the limit equilibrium method in example 2: (a) training data; (b) test data.
Table 5 20
Performance of the GPR-based response surface in example 2.
(33.28,15.24) (48.28,15.24)
Elevation (m)
Data sets MAE RMSE R 15
2
9.14
Training 0.0060 0.0083 0.9995
10 1 Layer 1
Testing 0.0308 0.0429 0.9908
(15,6.1) (48.28,6.1)
5 (0,6.1)
6.1
Layer 2
In this example, the number of LHS samples was set to N = 10D. 0
0 5 10 15 20 25 30 35 40 45 50
The comparison between values of factor of safety calculated by
Bishop’s simplified method and predicted by GPR model is shown Horizontal distance (m)
in Fig. 8. The accuracy of the GPR model is shown in Table 5. From
Fig. 9. Cross-section of example 3.
Table 5 and Fig. 8, it can be seen that the GPR model perform well
to predict the factor of safety. Results of probabilistic analysis by
different methods for this example are listed in Table 6. As shown Table 7
in Table 6, the failure probability obtained by the proposed method Statistical properties of soil parameters in example 3.
is close to the value obtained by MCS and several other system reli- Layer 1 Layer 2
ability analysis methods.
c1 (kPa) u1 (°) c2 (kPa) u2 (°)
Distribution Normal Normal Normal –
Mean 38.31 0 23.94 12
5.3. Example 3
COV 0.2 0 0.2 0.1
c (kN/m3) 19.0 19.0
The third example is presented by Hassan and Wolff [4]. The
geometry of the slope is shown in Fig. 9 and the soil parameters
are given in Table 7. This example was also analyzed by Xu and The factor of safety based on mean values of soil properties was
Low [7], Xue and Gavin [5], Cho [21], Chowdhury and Rao [46], found to be 1.652, which is very close with the value 1.651
Kang et al. [6], Luo et al. [9], Zhang et al. [19]. obtained by Xu and Low [7] using Spencer method with circular
Table 6
Results of probabilistic analysis by different methods for example 2.
2.5 2.5
Case points Case points
Fs (calculated) = Fs (predicted) Fs (calculated) = Fs (predicted)
2 2
Predicted factor of safety
1 1
0.5 0.5
0.5 1 1.5 2 2.5 0.5 1 1.5 2 2.5
Calculated factor of safety Calculated factor of safety
(a) (b)
Fig. 10. Comparison between values of factor of safety predicted by GPR model and calculated by the limit equilibrium method in example 3: (a) training data; (b) test data.
Table 8 15 (20.0,13.5)
Performance of the GPR-based response surface in example 3.
Elevation (m)
(0.0,13.5)
Clay 1 (33.5,9.0)
Data sets MAE RMSE R
10 (38.0,7.5) (60.0,7.5)
(0.0,9.0) Clay 2
Training 0.0012 0.0016 1.0000 5
Testing 0.0159 0.0203 0.9989 (0.0,4.5) Clay 3 (60.0,4.5)
0
0 5 10 15 20 25 30 35 40 45 50 55 60
Horizontal distance (m)
slip surface. The position of deterministic critical slip surface and
probabilistic critical slip surface can be seen in Kang et al. [6]. Fig. 11. Cross-section of example 4.
In this example, the number of LHS samples was set to
N = 10D = 30. The comparison between values of factor of safety
calculated by Bishop’s simplified method and predicted by GPR
model is shown in Fig. 10. The accuracy of the GPR model is shown
Table 10
in Table 8. From Table 8 and Fig. 10, it can be seen that the GPR Statistical properties of soil parameters in example 4.
model performs well to predict the factor of safety. Results of prob-
Slope layers Unit weight, c (kN/m3) Undrained strength, cu (kPa)
abilistic analysis from different methods for this example are listed
in Table 9. As shown in Table 9, the failure probability obtained by Mean COV Distribution
the proposed method is close to the value obtained by MCS other Clay 1 18.0 18 0.3 Normal
reliability analysis methods. In this example, the system effect is Clay 2 18.0 20 0.2 Normal
not obvious [19], and the failure probability of the most critical slip Clay 3 18.0 25 0.3 Normal
Table 9
Results of probabilistic analysis by different methods for example 3. The variables are assumed normally distributed.
2.5 2.5
Case points Case points
Fs (calculated) = Fs (predicted) Fs (calculated) = Fs (predicted)
2 2
Predicted factor of safety
1 1
0.5 0.5
0 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
Calculated factor of safety Calculated factor of safety
(a) (b)
Fig. 12. Comparison between values of factor of safety predicted by GPR model and calculated by the limit equilibrium method in example 4: (a) training data; (b) test data.
Table 11 40
Performance of the GPR-based response surface in example 4.
(50,35) (70,35)
Data sets MAE RMSE R Elevation (m) 35
Layer 1
Training 0.0029 0.0041 0.9999
Testing 0.0621 0.0823 0.9812 30 (54,31) (70,31)
(50,29) Layer 2
25 (40,27) (52,24) (70,24)
(20,25) (30,25)
The critical deterministic slip surface passes through Clay 1 and 2, Layer 3
while the critical probabilistic surface searched by the Zhang et al. 20
20 25 30 35 40 45 50 55 60 65 70
[19] passes through all three layers.
Horizontal distance (m)
In this example, the number of LHS samples was set to
N = 10D = 30. The comparison between values of factor of safety Fig. 13. Cross-section of example 5.
calculated by Bishop’s simplified method and predicted by GPR
model is shown in Fig. 12. The accuracy of the GPR model is shown
in Table 11. From Table 11 and Fig. 12, it can be seen that the GPR
Table 13
model performs well to predict the factor of safety. Results of prob- Statistical properties of soil parameters in example 5.
abilistic analysis by different methods for this example are listed in
Slope layers Unit weight Cohesion (kPa) Friction angle (°)
Table 12. As shown in Table 12, the failure probability obtained by
the proposed method is close to the value obtained by MCS several c (kN/m3) Mean COV Mean COV
other system reliability analysis methods. Layer 1 19.5 0 NA 38 NA
Layer 2 19.5 5.3 0.3 23 0.2
Layer 3 19.5 7.2 0.3 20 0.2
5.5. Example 5
Note: NA = not applicable. Normal distribution is assumed for the random variables.
The fifth example is taken from the work of Ji and Low [14], and
it is adapted from a test problem of Association for Computer Aided methodology with 100,000 samplings, which is quite far from the
Design, Australia (ACADS). The geometry of the slope is shown in value obtained by MCS method. Therefore we increase the number
Fig. 13 and the soil parameters are given in Table 13. This example of LHS samples to N = 15D = 60. Meanwhile, we found that the fac-
was also analyzed by Zhang et al. [19]. The factor of safety based on tor of safety is Fs = 0.970 when the parameters of layer 2 are
mean values of soil properties was found to be 1.405, which is very adopted as lx 3rx and layer 3 with relatively large parameter
close with the value 1.406 obtained by Ji and Low [14] using Spen- values. The critical slip surface in this case is shown in Fig. 14.
cer method with circular slip surface. Because Fs = 1.0 is the boundary of safety and failure, it is hard
When N = 10D and LHS sampling range is [lx 3rx, lx + 3rx], for the LHS to generate enough samples to capture the failure
the obtained failure probability Pf is 1.85% using the proposed mode that only pass through layer 1 and layer 2.
Table 12
Results of probabilistic analysis by different methods for example 4.
35 Table 14
Performance of the GPR-based response surface in example 5.
Elevation (m)
20
20 30 40 50 60 70
perform well to predict the factor of safety. Results of probabilistic
Horizontal distance (m)
analysis by different methods for this example are listed in
Fig. 14. The critical deterministic slip surface passes through layer 1 and 2. Table 15. It can be observed that the value of failure probability
obtained by the proposed method is close to the value obtained
by MCS and several other system reliability analysis methods.
To improve he accuracy, the sampling range is modified to
[lx 4rx, lx + 4rx] and the number of LHS samples was set to
N = 15D. The points generated by MCS with 50,000 samplings are 6. Discussion
shown in Fig. 15. It can be seen that the adopted range can cover
almost all the sampling points. Therefore, it is more conservative. Results of the five examples verified the effectiveness of the
The comparison between values of factor of safety calculated by proposed method for system reliability analysis of soil slopes. From
Bishop’s simplified method and predicted by GPR model is shown the examples, we recommend the range for generating samples
in Fig. 16. The accuracy of the GPR model is shown in Table 14. from LHS should be [lx 4rx, lx + 4rx] for normal distributed vari-
From Table 14 and Fig. 16, it can be seen that the GPR model ables and the number of samples generated by the LHS should be
(a) (b)
Fig. 15. Points generated by MCS with 50,000 samplings of different soil layers: (a) layer 2; (b) layer 3.
2 2
Case points Case points
1.8
Fs (calculated) = Fs (predicted) Fs (calculated) = Fs (predicted)
1.6
Predicted factor of safety
1.4
1.5
1.2
0.8
1
0.6
0.4
0.2
0 0.5
0 0.5 1 1.5 2 0.5 1 1.5 2
Calculated factor of safety Calculated factor of safety
(a) (b)
Fig. 16. Comparison between values of factor of safety predicted by GPR model and calculated by the limit equilibrium method in example 5: (a) training data; (b) test data.
24 F. Kang et al. / Computers and Geotechnics 63 (2015) 13–25
Table 15
Results of probabilistic analysis by different methods for example 5. The variables are assumed normally distributed.
Example 2 for system reliability analysis of slopes, the proposed new method
6 Example 3 has demonstrated strong competitive capabilities in terms of con-
Example 4 venience and efficiency.
5 Example 5
Conflict of interest
4
2 Acknowledgements
[20] Li L, Wang Y, Cao Z, Chu X. Risk de-aggregation and system reliability analysis [34] Choi SK, Grandhi RV, Canfield RA. Reliability based structural
of slope stability using representative slip surfaces. Comput Geotech design. London: Springer-Verlag; 2007.
2013;53:95–105. [35] Naess A, Leira BJ, Batsevych O. System reliability analysis by enhanced Monte
[21] Cho SE. First-order reliability analysis of slope considering multiple failure Carlo simulation. Struct Saf 2009;31(5):349–55.
modes. Eng Geol 2013;154:98–105. [36] Snelson EL. Flexible and efficient Gaussian process models for machine
[22] El-Ramly H, Morgenstern NR, Cruden DM. Probabilistic slope stability analysis learning. PhD thesis, Gatsby Computational Neuroscience Unit. University
for practice. Can Geotech J 2002;39(3):665–83. College London; 2007.
[23] Wang Y, Cao ZJ, Au SK. Practical reliability analysis of slope stability by [37] Yuan J, Wang K, Yu T, et al. Reliable multi-objective optimization of high-speed
advanced Monte Carlo simulations in a spreadsheet. Can Geotech J WEDM process based on Gaussian process regression. Int J Mach Tool Manuf
2011;48(1):162–72. 2008;48(1):47–60.
[24] Zhang J, Huang HW, Phoon KK. Application of the Kriging-based response [38] Bucher CG, Bourgund U. A fast and efficient response surface approach for
surface method to the system reliability of soil slopes. J Geotech Geoenviron structural reliability problems. Struct Saf 1990;7(1):57–66.
Eng 2012;139(4):651–5. [39] Goh ATC, Kulhawy FH. Reliability assessment of serviceability performance of
[25] Pérez-Cruz F, Vaerenbergh SV, Murillo-Fuentes JJ, et al. Gaussian processes for braced retaining walls using a neural network approach. Int J Numer Anal
nonlinear signal processing: an overview of recent advances. IEEE Signal Proc Methods Geomech 2005;29(6):627–42.
Mag 2013;30(4):40–50. [40] Fang KT, Lin DKJ, Winker P, Zhang Y. Uniform design: theory and application.
[26] Verrelst J, Alonso L, Camps-Valls G, et al. Retrieval of vegetation biophysical Technometrics 2000;42(3):237–48.
parameters using Gaussian process techniques. IEEE Trans Geosci Remote [41] Levy S, Steinberg DM. Computer experiments: a review. AStA – Adv Stat Anal
2012;50(5):1832–43. 2010;94(4):311–24.
[27] Pal M, Deswal S. Modelling pile capacity using Gaussian process regression. [42] Lü Q, Chan CL, Low BK. Probabilistic evaluation of ground-support interaction
Comput Geotech 2010;37(7):942–7. for deep rock excavation using artificial neural network and uniform design.
[28] Samui P, Jagan J. Determination of effective stress parameter of unsaturated Tunn Undergr Space Technol 2012;32:1–18.
soils: a Gaussian process regression approach. Front Struct Civ Eng [43] Loeppky JL, Sacks J, Welch WJ. Choosing the sample size of a computer
2013;7(2):133–6. experiment: a practical guide. Technometrics 2009;51(4):366–76.
[29] Rasmussen CE, Williams CKI. Gaussian processes for machine [44] Silvestrini RT, Montgomery DC, Jones B. Comparing computer experiments for
learning. Cambridge (MA): The MIT Press; 2006. the Gaussian process model using integrated prediction variance. Qual Eng
[30] Chen VCP, Tsui KL, Barton RR, Meckesheimer M. A review on design, modeling 2013;25(2):164–74.
and applications of computer experiments. IIE Trans 2006;38(4):273–91. [45] Rasmussen CE, Nickisch H. Gaussian processes for machine learning (GPML)
[31] Duncan JM. State of the art: limit equilibrium and finite-element analysis of toolbox. J Mach Learn Res 2010;11:3011–5.
slopes. J Geotech Eng 1996;122(7):577–96. [46] Chowdhury R, Rao BN. Probabilistic stability assessment of slopes using high
[32] Cho SE. Probabilistic stability analyses of slopes using the ANN-based response dimensional model representation. Comput Geotech 2010;37(7):876–84.
surface. Comput Geotech 2009;36(5):787–97. [47] Feng T, Fredlund M. SVSLOPE: verification manual. Saskatoon: SoilVision
[33] Fenton GA, Griffiths DV. Risk assessment in geotechnical engineering. New Systems Ltd.; 2011.
York: John Wiley & Sons; 2008.