Sunteți pe pagina 1din 8

Computers and Geotechnics 37 (2010) 1015–1022

Contents lists available at ScienceDirect

Computers and Geotechnics


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

Efficient Monte Carlo Simulation of parameter sensitivity in probabilistic


slope stability analysis
Yu Wang ⇑, Zijun Cao 1, Siu-Kui Au 2
Department of Building and Construction, City University of Hong Kong, Tat Chee Avenue, Kowloon, Hong Kong

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

Article history: Monte Carlo Simulation (MCS) method has been widely used in probabilistic analysis of slope stability,
Received 21 April 2010 and it provides a robust and simple way to assess failure probability. However, MCS method does not
Received in revised form 12 July 2010 offer insight into the relative contributions of various uncertainties (e.g., inherent spatial variability of
Accepted 23 August 2010
soil properties and subsurface stratigraphy) to the failure probability and suffers from a lack of resolution
Available online 18 September 2010
and efficiency at small probability levels. This paper develop a probabilistic failure analysis approach that
makes use of the failure samples generated in the MCS and analyzes these failure samples to assess the
Keywords:
effects of various uncertainties on slope failure probability. The approach contains two major compo-
Probabilistic failure analysis
Slope stability
nents: hypothesis tests for prioritizing effects of various uncertainties and Bayesian analysis for further
Monte Carlo Simulation quantifying their effects. Equations are derived for the hypothesis tests and Bayesian analysis. The prob-
Subset Simulation abilistic failure analysis requires a large number of failure samples in MCS, and an advanced Monte Carlo
Hypothesis tests Simulation called Subset Simulation is employed to improve efficiency of generating failure samples in
Bayesian analysis MCS. As an illustration, the proposed probabilistic failure analysis approach is applied to study a design
scenario of James Bay Dyke. The hypothesis tests show that the uncertainty of undrained shear strength
of lacustrine clay has the most significant effect on the slope failure probability, while the uncertainty of
the clay crust thickness contributes the least. The effect of the former is then further quantified by a
Bayesian analysis. Both hypothesis test results and Bayesian analysis results are validated against inde-
pendent sensitivity studies. It is shown that probabilistic failure analysis provides results that are equiv-
alent to those from additional sensitivity studies, but it has the advantage of avoiding additional
computational times and efforts for repeated runs of MCS in sensitivity studies.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction [21–25]. Among these methods, MCS method that integrates with
different deterministic analysis methods (e.g., limit equilibrium
Various uncertainties exist in slope engineering, such as inher- methods [10,11,13], finite element methods [12,14,15]) are gaining
ent spatial variability of soil properties, subsurface stratigraphy, popularity due to their robustness and conceptual simplicity. How-
simplifications and approximations adopted in geotechnical mod- ever, as pointed out by Baecher and Christian [26], although MCS
els. Effects of these uncertainties on probability of slope failure method provides a robust and simple way to assess failure proba-
are often significant, and insight on these effects is of great value bility, it does not offer insight into the relative contributions of var-
for understanding failure mechanisms and designing slope reme- ious uncertainties to the failure probability and suffers from a lack
dial measures. Several probabilistic methodologies have been of resolution and efficiency at small probability levels.
developed to incorporate these uncertainties in slope stability This paper develops a probabilistic failure analysis approach
analysis, such as the First Order Second Moment (FOSM) method that makes uses of the failure samples generated in the MCS and
[1–4], First Order Reliability Method (FORM) [5–9], Monte Carlo analyzes these failure samples to assess the effects of various
Simulation (MCS) method [10–20], and other methods based on uncertainties on slope failure probability. An advanced Monte Car-
artificial neural network, support vector machine or random set lo Simulation called Subset Simulation (Subsim) [27,28] is em-
ployed to improve efficiency of generating failure samples in
MCS and resolution at small failure probability levels. The paper
starts with mathematical formulation of the approach, including
⇑ Corresponding author. Tel.: +852 2788 7605; fax: +852 2788 7612.
hypothesis tests for prioritizing effects of various uncertainties
E-mail addresses: yuwang@cityu.edu.hk (Y. Wang), zijuncao3@student.cityu.
edu.hk (Z. Cao), siukuiau@cityu.edu.hk (S.-K. Au).
and Bayesian analysis for further quantifying their effects. Subset
1
Tel.: +852 3442 6492; fax: +852 2788 7612. Simulation and its implementation in a commonly-available
2
Tel.: +852 2194 2769; fax: +852 2788 7612. spreadsheet environment are briefly described. As an illustration,

0266-352X/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2010.08.010
1016 Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022

the proposed approach is applied to study a design scenario of where P(F|h) is the conditional probability density function (PDF) of
James Bay Dyke [2,10,14]. slope failure for a given h value; P(F) is the probability of slope fail-
ure; P(h|F) is the conditional PDF of h given that the slope has failed;
2. Probabilistic failure analysis approach and P(h) is the unconditional PDF of h. As both P(F) and P(h|F) are esti-
mated from failure samples of MCS and P(h) is given before MCS, Eq.
Probabilistic failure analysis is similar to back-analysis [29–31], (3) can be used to estimate P(F|h) using P(h) and P(h|F) obtained from
which is a common analysis procedure in geotechnical engineering analysis of failure samples. Note that P(F|h) is a variation of failure
that intends to find a set of model parameters that would result in probability as a function of h, and it can be considered as results of
the observed performance of geo-structures. Similarly, probabilis- a sensitivity study of h on slope failure probability. In other words,
tic failure analysis aims to identify a group of uncertain parameters the probabilistic failure analysis approach presented in this paper,
that would significantly affect the slope performance (i.e., the which makes use of the failure samples generated in a single run
probability of slope failure). The back-analysis, however, relies on of MCS for assessment of failure probability, provides results that
the observed performance and it is inapplicable when observed are equivalent to those from a sensitivity study, which frequently in-
performance of the geo-structures of interest is unavailable (e.g., cludes many repeated runs of MCS with different given values of h in
during design analysis of a new geo-structure). On the other hand, each run. Additional computational times and efforts for repeated
probabilistic failure analysis makes uses of the failure samples gen- runs of MCS in the sensitivity study can be avoided using the prob-
erated in the MCS, and it is readily applicable in design analysis for abilistic failure analysis approach described herein.
evaluating the effects of various uncertainties. The proposed prob- In addition, Eq. (3) implies that comparison between the condi-
abilistic failure analysis approach contains two major components: tional probability P(h|F) and its unconditional one P(h) provides an
hypothesis tests for prioritizing effects of various uncertainties and indication of the effect of the uncertain parameter h on failure proba-
Bayesian analysis for further quantifying their effects, which are bility. In general, P(F|h) changes as the values of the uncertain param-
described in the following two subsections, respectively. eter h changes. However, when P(h|F) is similar to P(h), P(F|h) remains
more or less constant regardless of the values of h. This implies that
2.1. Hypothesis tests the effect of h on the slope failure probability is minimal. Such impli-
cation can be used to validate the prioritization obtained from
The effects of various uncertainties on the probability of slope hypothesis tests, as shown in the example of James Bay Dyke later.
failure are prioritized by comparing, statistically, failure samples The resolution of P(F) and P(h|F) is pivotal to obtain P(F|h), and it
with their respective nominal (unconditional) samples. When the depends on the number of failure samples generated in MCS. As
uncertainty of an uncertain system parameter has a significant ef- the number of failure samples increases, the resolution improves.
fect on the probability of slope failure, the mean l of failure sam- For a given slope stability problem, the value of P(F) is constant,
ples of the parameter differs significantly from the mean l0 of its although unknown before MCS. In this case, increasing the number
unconditional samples. The statistical difference between l and of failure samples necessitates an increase in the total number of
l0 is evaluated by hypothesis tests. A null hypothesis H0 and alter- samples in MCS. One possible way to improve the resolution is,
native hypothesis HA are defined as [32] therefore, to increase the total number of samples in MCS at the ex-
pense of computational time. Alternatively, advanced MCS meth-
H0 : l ¼ l0 ods can be employed to improve efficiency and resolution at
ð1Þ
HA : l–l0 small failure probability levels. An advanced MCS called Subset
Then, a hypothesis test statistic ZH of the parameter is formulated as Simulation (Subsim) [27,28,35,36] is used in this paper to calculate
the failure probability and generate failure samples efficiently for
l  l0 probabilistic failure analysis.
ZH ¼ pffiffiffi ð2Þ
r= n
where r is standard deviation of the uncertain parameter and n is 3. Subset Simulation
the number of failure samples. Based on Central Limit Theorem,
ZH follows the standard Normal distribution when n is large (e.g., 3.1. Algorithm
n P 30) [32]. When the failure sample mean l deviates statistically
from the unconditional mean l0 of the parameter, the absolute va- Subset Simulation is an adaptive stochastic simulation procedure
lue of ZH is relatively large. As the absolute value of ZH increases, the for efficiently generating failure samples and computing small tail
statistical difference between l and l0 becomes growingly signifi- probability [27,28]. It expresses a small probability event as a se-
cant, and the effect of the uncertain parameter on failure probability quence of intermediate events {F1, F2, . . ., Fm} with larger conditional
also becomes growingly significant. The absolute value of ZH is probability and employs specially designed Markov chains to gener-
therefore formulated in this paper as an index to measure the ef- ate conditional samples of these intermediate events until the final
fects of the uncertain parameters on failure probability and to pri- target failure region is achieved. For the slope stability problem, the
oritize their relative effects on failure probability. Using the factor of safety (FS) is the key parameter, and let Y = 1/FS be the crit-
absolute value of ZH, the uncertain parameters that have significant ical response [20,35]. The probability of Y = 1/FS larger than a given
effects on failure probability are selected, and their effects are fur- value y (i.e., P(Y = 1/FS > y)) is of interest and let 0 < y1 < y2 <
ther quantified using a Bayesian analysis approach described in    < ym1 < ym = y be an increasing sequence of intermediate thresh-
the next subsection. old values. The sequence of intermediate events {F1, F2, . . ., Fm} are
chosen as Fi = {Y > yi, i = 1, 2, . . ., m} for these intermediate threshold
2.2. Bayesian analysis values. By sequentially conditioning on the event {Fi, i = 1, 2, . . ., m},
the failure probability can be written as
The failure samples generated in the MCS are further analyzed
Y
m
by a Bayesian analysis to quantify effects of various uncertainties. PðFÞ ¼ PðF m Þ ¼ PðF 1 Þ PðF i jF i1 Þ ð4Þ
Let h denote an uncertain parameter selected based on hypothesis i¼2
tests. In the context of the Bayesian Theorem [33,34]
where P(F1) is equal to P(Y > y1) and P(Fi|Fi1) is equal to
PðFjhÞ ¼ PðFÞPðhjFÞ=PðhÞ ð3Þ {P(Y > yi|Y > yi1): i = 2, . . ., m}. In implementations, y1, y2, . . ., ym
Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022 1017

are generated adaptively using information from simulated samples from the failure sample histogram in Xi. The number of bins k in the
so that the sample estimate of P(F1) and {P(Fi|Fi1): i = 2, . . ., m} al- failure sample histogram is estimated as [38]
ways corresponds to a common specified value of conditional prob- n
ability p0 (p0 = 0.1 is found to be a good choice) [35,36]. k ¼ 1 þ log2i ð9Þ
The efficient generation of conditional samples is pivotal in the where ni is the number of the failure samples in Xi. Using Eqs. (5)
success of Subset Simulation, and it is made possible through the and (8), P(F) and P(h|F) can be calculated from Subset Simulation,
machinery of Markov Chain Monte Carlo (MCMC). In MCMC, and subsequently, P(F|h) is calculated using Eq. (3).
Metropolis algorithm [37] is used and successive samples are gen-
erated from a specially designed Markov chain whose limiting sta-
tionary distribution tends to the target PDF as the length of Markov 4. Implementation of Subset Simulation in spreadsheet
chain increases. Details of the computational process of Subset environment
Simulation and its application to slope stability problems are re-
ferred to Au and Beck [27,28] and Wang et al. [20], respectively. The Subset Simulation described above has been implemented
The Subset Simulation provides much more failure samples in a commonly-available spreadsheet environment by a package
than Direct MCS under the same total number of samples, espe- of worksheets and functions/Add-In in EXCEL with the aids of Vi-
cially when the failure probability is relatively small. When com- sual Basic for Application (VBA) [35,36]. The package is divided
pared with failure samples generated in direct MCS where each into three parts: deterministic model worksheet, uncertainty mod-
sample carries equal weight in the calculation of P(F) and P(h|F), el worksheet and Subset Simulation Add-In, which are described
the samples generated by Subset Simulation are conditional sam- briefly in the following three subsections, respectively.
ples and carry different weights for different intermediate events
Fm. Thus, when using these conditional failure samples collected 4.1. Deterministic model worksheet
from Subset Simulation to construct the conditional PDF P(h|F) re-
quired in Eq. (3), a weighted summation by Total Probability The- For a slope stability problem, deterministic model analysis is
orem [33,34] is necessary, which is described in the following the process of calculating factor of safety for a given nominal set
subsection. of values of system parameters. The system parameters include
the geometry information of the slope and the slip surface, soil
properties and profile of soil layers. In this paper, limit equilibrium
3.2. Estimation of P(h|F) based on conditional failure samples
methods (e.g., Swedish Circle method, Simplified Bishop method
and Spencer method) [39] are employed to calculate the factor of
Consider a Subset Simulation that performs m + 1 levels of sim-
safety for the critical slip surface. The calculation process of deter-
ulations. The first level of Subset Simulation is direct MCS, and
ministic analysis is implemented in a series of worksheets assisted
samples of the next level are generated conditional on the samples
by some VBA functions/Add-In [35,36]. From an input–output per-
collected from the previous level. The intermediate threshold val-
spective, the deterministic analysis worksheets take a given set of
ues {yi, i = 1, 2, . . ., m} divide the sample space X of an uncertain
values as input, calculate the factor of safety and return the factor
parameter h into m individual sets {Xi, i = 0, 1, 2, . . ., m}. According
of safety as an output.
to the Total Probability Theorem [34], the failure probability can be
written as
4.2. Uncertainty model worksheet
X
m
PðFÞ ¼ PðFjXi ÞPðXi Þ ð5Þ An uncertainty model worksheet is developed to generate ran-
i¼0
dom samples of uncertain system parameters that are treated as
where X0 = {Y 6 y1}; Xi, i = 1, . . ., m  1 is equal to Fi  Fi+1 (i.e., random variables in the analysis. The uncertain worksheet includes
Xi = {yi 6 Y 6 yi+1}); Xm is equal to Fm (i.e., Xm = {Y P ym}); P(F|Xi) detailed information of random variables, such as statistics, distri-
is the conditional failure probability given sampling in Xi; P(Xi) is bution type and correlation information. The generation of random
the probability of the event Xi. P(F|Xi) is estimated as the fraction samples starts with an EXCEL built-in function ‘‘RAND()” for gener-
of the failure samples in Xi. The failure samples are collected from ating uniform random samples, which are then transformed to ran-
samples generated by Subset Simulation and are based on the per- dom samples of the target distribution type (e.g., normal
formance failure criteria (i.e., Y = 1/FS P 1 for a slope stability prob- distribution or lognormal distribution). If the random variables
lem). P(Xi) is calculated as are considered correlated, Cholesky factorization of the correlation
matrix is performed to obtain a lower triangular matrix, which is
PðX0 Þ ¼ 1  p0 used in the transformation to generate correlated random samples.
PðXi Þ ¼ pi0  p0iþ1 ði ¼ 1; . . . ; m  1Þ ð6Þ Details of the random sample generation process are referred to Au
et al. [36]. From the input–output perspective, the uncertainty
PðXm Þ ¼ pm
0
model worksheet takes no input but returns a set of random sam-
P
Note that P(Xi \ Xj) = 0 for i – j and m
i¼0 PðXi Þ ¼ 1. When P(F), ples of the uncertain system parameters as its output.
P(F|Xi) and P(Xi) are obtained, the conditional probability P(Xi|F) When deterministic model worksheet and uncertainty model
is calculated using the Bayesian Theorem worksheet are developed, they are linked together through their
input/output cells to execute probabilistic analysis of slope stabil-
PðXi jFÞ ¼ PðFjXi ÞPðXi Þ=PðFÞ ð7Þ ity. The connection is carried out by simply setting the cell refer-
Then the conditional PDF P(h|F) of an uncertain parameter h is ences for nominal values of uncertain parameters in
given by the Total Probability Theorem as deterministic model worksheet to be the cell references for the
random samples in the uncertainty model worksheet in EXCEL.
X
m
After this task, the values of uncertain system parameters shown
PðhjFÞ ¼ PðhjXi \ FÞPðXi jFÞ ð8Þ
i¼0
in the deterministic model worksheet are equal to that generated
in the uncertainty model worksheet, and so the values of the safety
where P(h|Xi \ F) is the conditional probability of h estimated from factor calculated in the deterministic modeling worksheet are
failure samples that lie in Xi. In this paper, P(h|Xi \ F) is estimated random.
1018 Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022

4.3. Subset Simulation Add-In


56.0m

When the deterministic analysis and uncertainty model work- 45


(x, y)

Elevation (m)
sheets are completed and linked together, Subset Simulation pro- 3
35 1 3
m ankment γFill; φFill
Emb 1 12.0m
cedure is invoked for uncertainty propagation. In this paper, 25
Claay Cr
Cl Cruust Tcr
Subset Simulation is implemented as an Add-In in EXCEL [35,36]. 15 Marine Clay y SuM 8.0m DTill
S
The userform of the Add-In is shown in Fig. 1. The upper four input 5 Lacustrine Clay y uL TL
fields of the userform (i.e., number of Subset Simulation runs, num- Tiill
T Critical slip
p circle
-5
ber of samples per level N, conditional probability from one level to 0 20 40 60 80 100 120 140 160
next p0, the highest Subset Simulation level m) control the number Distance (m)
of samples generated by Subset Simulation. The total number of
Fig. 2. The James Bay Dyke (modified after [10]).
samples per Subset Simulation run is equal to N + mN(1  p0).
The lower four input fields of the userform record the cell refer-
ences of the random variables, their PDF values, and the cell refer- clay with a thickness TL. The undrained shear strength (i.e., SuM and
ences of the system response (e.g., Y = 1/FS in this paper) and other SuL) of the marine clay and the lacustrine clay were measured by
variables V (e.g., the samples generated in this paper) of interest, field vane tests [2,10,40,41]. The lacustrine clay is overlying on a
respectively. stiff till layer, the depth to the top of which is DTill.
After each simulation run, the Add-In provides the complemen- Six uncertain system parameters have been considered in liter-
tary cumulative density function (CDF) of the driving variable ver- ature (e.g., El-Ramly [10], El-Ramly et al. [11] and Xu and Low
sus the threshold level, i.e., estimate for P(Y > y) versus y = 1/FS in [14]), including the friction angle /Fill and unit weight cFill of
this paper, into a new spreadsheet and produces a plot of it [36]. embankment material, the thickness Tcr of clay crust, the un-
Then, based on the output information obtained, the CDF, histo- drained shear strength SuM of the marine clay, the undrained shear
grams or conditional counterparts (e.g., P(h|F)) of uncertain param- strength SuL of the lacustrine clay, and the depth of the till layer
eters of interest are calculated using the procedures and equations DTill. During the probabilistic failure analysis of the dyke, the six
described in the previous section. uncertain parameters are represented by six independent Gaussian
random variables [10], respectively. Statistics [i.e., mean, standard
deviation and coefficient of variation (COV)] of these six random
5. The James Bay Dyke variables are summarized in Table 1. These statistics are used to
generate random samples for each random variable in uncertain
As an illustration, the probabilistic failure analysis approach is model worksheet. Note that thickness of the lacustrine clay TL is
applied to analyze a design scenario of the James Bay Dyke. The an uncertain variable that depends on Tcr and DTill and has a mean
James Bay Dyke is a 50 km-long earth dyke of the James Bay hydro- of about 6.5 m (see Fig. 2). In addition to these uncertain parame-
electric project in Canada. Soil properties and various design sce- ters, other system parameters are taken as deterministic, including
narios of the dyke were studied by Ladd et al. [40], Soulié et al. an undrained shear strength of 41 kPa for the clay crust and unit
[41], Christian et al. [2], El-Ramly [10], El-Ramly et al. [11] and weights of 19 kN/m3, 19 kN/m3 and 20.5 kN/m3 for the clay crust,
Xu and Low [14]. As shown in Fig. 2, the embankment is 12 m high marine clay and lacustrine clay, respectively [10].
with a 56 m-wide berm at mid-height. The slope angle of the Using the soil properties described above, El-Ramly [10] and Xu
embankment is about 18.4° (3H:1V). The embankment is overlying and Low [14] employed direct MCS methods integrating with limit
on a clay crust with a thickness Tcr. The clay crust is underlain by a equilibrium methods and response surface method, respectively, to
layer of 8.0-m thick sensitive marine clay and a layer of lacustrine evaluate failure probability of the dyke. To enable a consistent
comparison with the analyses by El-Ramly [10], the critical slip
surfaces recommended by El-Ramly [10] are adopted in this paper,
which is circular and always tangential to top of the till layer and
pass through the point (x = 4.9 m, y = 36.0 m). The x coordinate of
the center is fixed at 85.9 m. For each set of random samples, the
critical slip surface is specified uniquely by the value of DTill. In this
paper, the safety factor of the critical slip surface is calculated by
Simplified Bishop method, and two Subset Simulation runs are per-
formed. One has the highest simulation level m = 3 and sample
number N = 1000 per each level, as opposed to m = 4 and
N = 10,000 per each level in the other run.

Table 1
Soil properties of the James Bay Dyke (modified after [10]).

Soil layers Uncertain system Mean Standard Coefficient of


parameters* deviation variation (%)
Embankment /Fill (°) 30.0 1.79 6.0
cFill (kN/m3) 20.0 1.10 5.5
Clay crust Tcr (m) 4.0 0.48 12.0
Marine clay SuM (kN/m2) 34.5 3.95 11.5
Lacustrine clay SuL (kN/m2) 31.2 6.31 20.2
Till DTill (m) 18.5 1.00 5.4
*
All parameters are modeled as Gaussian random variables. Thickness of the
Lacustrine clay layer TL is an uncertain variable that depends on Tcr and DTill and has
Fig. 1. The userform of Subset Simulation Add-In. a mean of about 6.5 m.
Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022 1019

5.1. Simulation results Columns 5 or 6). For a total sample number NT = 20,000, direct
MCS leads to only 48 or 44 failure samples. In contrast, Subset Sim-
Fig. 3 shows a typical complementary CDF for Y = 1/FS (i.e., ulations with m = 3 and NT = 3700 (i.e., Run 1 in Column 5) or m = 4
P(Y > y) versus y), from two Subset Simulation runs with a total and NT = 46,000 (i.e., Run 2 in Column 6) result in a failure sample
sample number NT = 1000 + 3  1000  (1–0.1) = 3700 (i.e., Subsim number of 1128 and 20,482, respectively. This comparison clearly
Run 1) and NT = 10,000 + 4  10,000  (1–0.1) = 46,000 (i.e., Subsim shows that Subset Simulations significantly improve the efficiency
Run 2), respectively. For comparison, the result from a direct of generating failure samples, which enables generation of a large
MCS with 20,000 samples is also plotted. Three consistent failure number of failure samples with relative ease and makes probabilis-
probabilities P(Y P 1) = 0.22%, 0.23%, and 0.25% are estimated, tic failure analysis feasible. Table 2 also shows that, as the value of
respectively, from direct MCS and two runs of Subset Simulations. m increases (e.g., from 3 in Column 5 to 4 in Column 6), the effi-
In addition, the two runs of Subset Simulations provide results that ciency increases as well (e.g., the percentage of failure sample in-
are consistent even at low probability levels [e.g., P(Y P y) = 0.01%] creases from 30.5% to 44.5%).
where the CDF curve from direct MCS becomes erratic.
Table 2 compares the simulation results with those reported by
6. Probabilistic failure analysis results
El-Ramly [10] and Xu and Low [14]. El-Ramly [10] performed direct
MCS with the Simplified Bishop method (i.e., the same limit equi-
With the large number of failure samples generated from Sub-
librium method used in this paper) and obtained a failure probabil-
set Simulations, probabilistic failure analysis are performed for
ity Pf = 0.24%. This Pf value is almost identical to the average of the
the James Bay Dyke, including hypothesis tests for identifying
three Pf values obtained in this paper (i.e., 0.22%, 0.23%, and 0.25%
key uncertainties that have significant effects on slope failure
in Columns 4–6 of Table 2). In addition, Xu and Low [14] combined
probability and Bayesian analysis for further quantifying their
MCS with response surface method to estimate the Pf of the James
effects.
Bay Dyke and obtained a Pf value of 0.33%. Although different
deterministic slope stability analysis methods were used, the Pf
6.1. Hypothesis test results
values obtained compare favourably with each other. This implies
that the probabilistic analysis models for the James Bay Dyke pre-
Based on the failure samples generated from Subsim Run 1, the
sented in this paper work properly. However, note that the Pf val-
hypothesis test statistics ZH defined by Eq. (2) are calculated and
ues summarized in Table 2 are obtained under the assumption of
shown in Fig. 4 for all uncertain parameters. The absolute values
circular slip surfaces. In contrast, non-circular failure mechanism
of ZH varies from less than 2 for the thickness of clay crust Tcr to
has been identified in recent finite element analyses incorporating
about 95 for undrained shear strength of the lacustrine clay SuL.
spatially varying soil properties [14,19]. When the searching of
The decreasing order of the ZH absolute values is: SuL, DTill, cFill,
non-circular slip surfaces is included in the MCS, the corresponding
SuM, /Fill, and Tcr. This implies that the uncertainty of SuL has the
Pf values will increase.
most significant effects on the slope failure probability, while the
Table 2 also compares the number of failure samples in direct
uncertainty of Tcr contributes the least to the failure probability.
MCS (e.g., Columns 2 or 4) with that in Subset Simulations (e.g.,
This result is consistent with that reported by El-Ramly [10] who
employed an Excel spreadsheet – based MCS program @RISK [42]
100 and compared the spearman rank correlation coefficients for vari-
ous uncertain parameters to show that SuL and Tcr are the most and
least influential parameter, respectively. The results can be vali-
10 dated by an independent sensitivity studies on SuL and Tcr, which
are described in next subsection.
P(Y>=y) (%)

1 6.2. Validation of hypothesis test results

Sensitivity studies are performed to explore the effect of SuL and


0.1
Tcr uncertainties on slope failure probability. The coefficient of var-
iation (COV) of SuL and Tcr is varied in the sensitivity studies from
Subsim Run 1 3700 Samples
0.15 to 0.50 and from 0.05 to 0.25, respectively. The SuL COV range
0.01 Subsim Run 2 46000 Samples
adopted in this study follows the typical range of COV of undrained
Direct MCS 20000 Samples
shear strength of clay measured by vane shear tests [43], and the
COV of Tcr varies from half to about twice of the value reported
0.001
0.5 0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 by El-Ramly [10]. Other parameters, including the means of SuL
y=1/FS and Tcr, remain unchanged in the sensitivity studies. About 40
additional Subset Simulation runs are performed to validate the
Fig. 3. Complementary CDF plot from simulations. hypothesis test results, and their results are shown in Fig. 5a and

Table 2
Comparison of simulation results.

Simulation results El-Ramly [10] Xu and Low [14] Direct MCS with EXCEL Subsim with EXCEL Run 1 Subsim with EXCEL Run 2
Failure probability Pf (%) 0.24a 0.33b 0.22a 0.23a 0.25a
Number of failure samples, NF 48 N/A 44 1128 20,482
Number of Total Samples, NT 20,000 N/A 20,000 3700 46,000
Percentage of failure samples, NF/NT (%) 0.24 N/A 0.22 30.5 44.5
a
Pf is calculated by MCS methods integrating with the simplified Bishop method.
b
Pf is calculated by MCS methods integrating with the response surface method.
1020 Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022

and the failure probability is insensitive to the uncertainty on Tcr.


γ Fill The results from Subset Simulations are in good agreement with
those from @RISK and FORM. This validates the results from
φ Fill hypothesis tests that the Tcr uncertainty has the least effect on
the failure probability. It is interesting to note that the probability
T cr of slope failure should, theoretically, increase as the COV of Tcr in-
creases from 0.05 to 0.25, as shown by the FORM results [i.e., a
rather slight increase of the Pf values shown by the open circles
S uM
in Fig. 5a when the COV of Tcr increases from 0.05 to 0.25]. The ef-
fect is, however, so minimal that it is dominated by the MCS
S uL
‘‘noise” (i.e., the random fluctuations of the Pf values obtained from
Subset Simulations and @RISK).
D Till On the other hand, Fig. 5b shows that the slope failure probabil-
0 10 20 30 40 50 60 70 80 90 100
ity increases as the COV of SuL increases. The results from Subset
Simulations are in good agreement with those from @RISK and
Absolute Value of Z H
FORM. When the COV of SuL is 0.15, the probability is less than
Fig. 4. Summary of absolute values of ZH. 0.1%. When the COV of SuL increases to 0.50, the slope failure prob-
ability increases by two orders of magnitude (i.e., increases to
about 10%). The failure probability varies significantly with the
b for Tcr and SuL, respectively. In addition, sensitivity studies on change of the SuL COV. This agrees well with the results from
COVs of Tcr and SuL are also carried out using the Excel spreadsheet hypothesis tests that the uncertainty of SuL has significant effect
– based MCS program @RISK [42] and a FORM calculation spread- on the slope failure probability. Such agreement further validates
sheet developed by Low and his coworkers [5–8]. Fig. 5 includes that the hypothesis test procedure proposed in this paper properly
the results from @RISK and FORM for comparison. prioritizes effects of various uncertainties on failure probability.
Fig. 5a shows that, when the COV of Tcr varies from 0.05 to 0.25,
the slope failure probability fluctuates between 0.19% and 0.43%.
For comparison, the baseline failure probability (i.e., about 0.24% 6.3. Bayesian analysis results
corresponding to the values summarized in Table 1) is also in-
cluded in the figure. The failure probabilities from sensitivity study Based on the failure samples generated from Subsim Run 2, a
using Subset Simulations fall around the horizontal line of 0.24%, Bayesian analysis is performed using Eqs. (3), (7), and (8) accord-

100 100
Subsim
Failure Probability, P(F) (% )

FORM
Failure Probability (%)

@ Risk
isk 10
10

1
1

0.24% 0.1

0.1
0.01
Parametric
P(F|SuL) estimated from sensitivity study
COV=12.0% used by El-Ramly [10]
P(F|SuL) estimated from Bayesian analysis
Series4
0.01 0.001
0.05 0.10 0.15 0.20 0.25 12 14 16 18 20 22 24
COV of T cr SuL(kN/m ) 2

(a) Thickness of clay crust, Tcr (a) θ = undrained shear strength of the lacustrine clay, SuL
100 100
Subsim
Failure Probability, P(F) (% )

FORM
Failure Probability (%)

@ Risk
Risk 10
10

1
1

0.24% 0.1

0.1
0.01
Parametric
P(F|DTill) estimated from sensitivity study
COV=20.2% used by El-Ramly [10]
Series4
P(F|DTill) estimated from Bayesian analysis
0.01 0.001
0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 17 18 19 20 21 22 23 24
COV of S uL DTill (m)
(b) Undrained shear strength of the lacustrine clay, SuL (b) θ = depth of the till layer, DTill

Fig. 5. Effects of coefficient of variation (COV) for different system parameters. Fig. 6. Bayesian analysis results for different system parameters h.
Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022 1021

ingly. Fig. 6 shows the Bayesian analysis results by open circles for 6.4. Validation of Bayesian analysis results
SuL and DTill, which have been identified from the Hypothesis tests
(see Section 6.1) as the two most influential uncertain parameters. Note that variations of failure probability as a function of SuL or
Note that the conditional probability [i.e., P(F|SuL) in Fig. 6a and DTill shown in Fig. 6 can also be obtained from sensitivity studies on
P(F|DTill) in Fig. 6b] obtained from Bayesian analysis is a variation SuL or DTill, which frequently include many repeated runs of MCS
of failure probability as a function of SuL or DTill. Fig. 6a shows that, with different given values of SuL or DTill in each run. Therefore,
as SuL increases from 12 kPa to 24 kPa, the slope failure probability to validate the Bayesian analysis results, about 40 additional Sub-
decreases from more than 10% to less than 0.1%. Similarly, Fig. 6b set Simulation runs are performed with different given values of
shows that, as DTill increases from about 18 m to 23 m, the slope SuL or DTill in each run. Fig. 6 also includes results from these addi-
failure probability increases from about 0.1% to about 10%. It is tional sensitivity runs by open triangles. The open triangles follow
obvious that the values of SuL and DTill have significant effects on trends similar to the open circles (i.e., the Bayesian analysis results)
slope failure probability, and such effects can be quantified explic- and plot closely to the open circles as well. This validates that a
itly from the Bayesian analysis of failure samples. Bayesian analysis of failure samples generated in MCS or Subset
Simulations provides results that are equivalent to those from
100
additional sensitivity studies. In addition, the Bayesian analysis
has the advantage of avoiding additional computational times
and efforts for repeated runs of MCS or Subset Simulations in the
10
sensitivity studies.
Probability (%)

As mentioned before, Eq. (3) implies that comparison between


the conditional probability P(h|F) and its unconditional one P(h)
1 provides an indication of the effect of the uncertain parameter h
on failure probability. This offers a means to check the Bayesian
analysis results with those from hypothesis tests. Fig. 7 compares
0.1 the conditional probabilities [i.e., P(Tcr|F), P(SuL|F) and P(DTill|F)]
P(Tcr|F)
Series4 for Tcr, SuL and DTill and their unconditional ones [i.e., P(Tcr), P(SuL)
Series1
P(Tcr) and P(DTill)]. Fig. 7a shows that P(Tcr|F) and P(Tcr) are almost iden-
0.01 tical for different Tcr values, which is consistent with the hypothe-
2 3 4 5 6 sis test results that Tcr has the least effect on failure probability. In
Tcr(m) contrast, the hypothesis tests show that SuL and DTill are the two
(a) θ = thickness of clay crust, Tcr most influential uncertain parameters. Fig. 7b and c show that
P(SuL|F) and P(DTill|F) differ significantly from their respective
100 unconditional one [i.e., P(SuL) and P(DTill)]. The Bayesian analysis re-
sults agree well with the hypothesis test results.

10
Probability (%)

7. Conclusions

1 This paper developed a probabilistic failure analysis approach


that makes use of the failure samples generated in the MCS and
analyzes these failure samples to assess the effects of various
0.1 uncertainties on slope failure probability. The approach contains
Series4
P(S uL|F) two major components: hypothesis tests for prioritizing effects of
P(SuL)
Series1 various uncertainties and Bayesian analysis for further quantifying
0.01 their effects.
12 14 16 18 20 22 24
2
A hypothesis test statistic ZH was formulated to evaluate the
SuL(kN/m ) statistical difference of failure samples with their respective nom-
(b) θ = undrained shear strength of the lacustrine clay, SuL inal (unconditional) samples. As the absolute value of ZH increases,
the statistical difference between failure samples and their respec-
100 tive nominal samples becomes growingly significant, and the effect
of the uncertain parameter on failure probability also becomes
10 growingly significant. Therefore, the absolute value of ZH is used
as an index to measure the effects of the uncertain parameters
Probability (%)

on failure probability and to prioritize their relative effects on fail-


1
ure probability.
A Bayesian analysis approach was developed to further quantify
0.1 effects of the uncertain parameters that have been identified from
the hypothesis tests as influential parameters. Equations were de-
0.01 rived for estimating conditional PDF [i.e., P(F|h)] of slope failure for
Series6
P(DTill|F) a given value of uncertain parameter h. As P(F|h) is a variation of
Series4
P(DTill) failure probability as a function of h, it can be considered as results
0.001 of a sensitivity study of h on slope failure probability. In other
17 18 19 20 21 22 23 24
D Till (m) words, a Bayesian analysis of the failure samples provides results
that are equivalent to those from additional sensitivity studies. In
(c) θ = depth of the till layer, DTill
addition, it has the advantage of avoiding additional computational
Fig. 7. Conditional probability density function (PDF) (P(h|F)) for different system times and efforts for repeated runs of MCS or Subset Simulations in
parameters h. the sensitivity studies. Furthermore, it was shown that comparison
1022 Y. Wang et al. / Computers and Geotechnics 37 (2010) 1015–1022

between the conditional probability P(h|F) and its unconditional [11] El-Ramly H, Morgenstern NR, Cruden DM. Probabilistic slope stability analysis
for practice. Can Geotech J 2002;39(3):665–83.
one P(h) provides an indication of the effect of the uncertain
[12] Griffiths DV, Fenton GA. Probabilistic slope stability analysis by finite
parameter h on failure probability. This offers a means to check elements. J Geotech Geoenviron Eng 2004;130(5):507–18.
the Bayesian analysis results with those from hypothesis tests. [13] El-Ramly H, Morgenstern NR, Cruden DM. Probabilistic assessment of stability
The resolution of P(F) and P(h|F) is pivotal to obtain P(F|h), and it of a cut slope in residual soil. Geotechnique 2005;55(1):77–84.
[14] Xu B, Low BK. Probabilistic stability analyses of embankments based on finite
depends on the number of failure samples generated in MCS. An element method. J Geotech Geoenviron Eng 2006;132(11):1444–54.
advanced Monte Carlo Simulation called Subset Simulation was [15] Griffiths DV, Jinsong Huang, Fenton GA. Influence of spatial variability on slope
employed to improve efficiency of generating failure samples in reliability using 2-D random fields. J Geotech Geoenviron Eng
2009;135(10):1367–78.
MCS and resolution at small failure probability levels. Subset Sim- [16] El-Ramly H, Morgenstern NR, Cruden DM. Lodalen slide: a probabilistic
ulation algorithm and its implementation in a commonly-available assessment. Can Geotech J 2006;43(9):956–68.
spreadsheet environment were briefly described. [17] Cho SE. Effects of spatial variability of soil properties on slope stability. Eng
Geol 2007;92(3–4):97–109.
As an illustration, the proposed probabilistic failure analysis ap- [18] Grilli ST, Taylor ODS, Baxter CDP, Maretzki SA. Probabilistic approach for
proach was applied to study a design scenario of James Bay Dyke. determining submarine landslide tsunami hazard along the upper east coast of
The hypothesis tests show that the uncertainty of SuL has the most the United States. Mar Geol 2009;264(1–2):74–97.
[19] Cho SE. Probabilistic assessment of slope stability that considers the spatial
significant effect on the slope failure probability, while the uncer- variability of soil properties. J Geotech Geoenviron Eng 2010;136(7):975–84.
tainty of Tcr contributes the least to the failure probability. The [20] Wang Y, Cao Z, Au SK. Practical reliability analysis of slope stability by
hypothesis test results are very consistent with results from inde- advanced Monte Carlo simulations in spreadsheet. Can Geotech J, accepted for
publication.
pendent sensitivity studies. Such agreement validates that the
[21] Rubio E, Hall JW, Anderson MG. Uncertainty analysis in a slope hydrology and
hypothesis test procedure proposed in this paper properly priori- stability model using probabilistic and imprecise information. Comput
tizes effects of various uncertainties on failure probability. Geotech 2004;31(7):529–36.
A Bayesian analysis was performed to quantify explicitly the ef- [22] Schweiger HF, Peschl GM. Reliability analysis in geotechnics with the random
set finite element method. Comput Geotech 2005;32(6):422–35.
fects of SuL and DTill, which have been identified from the Hypoth- [23] Zhao HB. Slope reliability analysis using a support vector machine. Comput
esis tests as the two most influential uncertain parameters. It is Geotech 2008;35(3):459–67.
shown that the slope failure probability change significantly as [24] Cho SE. Probabilistic stability analyses of slopes using the ANN-based response
surface. Comput Geotech 2009;36(5):787–97.
the values of SuL and DTill change. The Bayesian analysis results [25] Ching JY, Phoon KK, Hu YG. Efficient evaluation of reliability for slopes with
have also been validated against those from independent sensitiv- circular slip surfaces using importance sampling. J Geotech Geoenviron Eng
ity studies. In addition, a cross-check between the hypothesis test 2009;135(6):768–77.
[26] Baecher GB, Christian JT. Reliability and statistics in geotechnical
results and the Bayesian analysis results shows that they agree engineering. New York: John Wiley; 2003.
well with each other. [27] Au SK, Beck JL. Estimation of small failure probabilities in high dimensions by
It is worthwhile to note that, although the proposed approach Subset Simulation. Probab Eng Mech 2001;16(4):263–77.
[28] Au SK, Beck JL. Subset Simulation and its application to probabilistic seismic
was developed together with a slope stability analysis problem, performance assessment. J Eng Mech 2003;129(8):1–17.
the approach is general and applicable to other types of geotechni- [29] Luckman PG, Der Kiureghian A, Sitar N. Use of stochastic stability analysis for
cal analyses and engineering problems. Bayesian back calculation of pore pressures acting in a cut at failure. In:
Proceeding of 5th international conference on application of statistics and
probability in soil and structure engineering. University of British Columbia,
Acknowledgements Vancouver, Canada; 1987.
[30] Gilbert RB, Wright SG, Liedtke E. Uncertainty in back-analysis of slopes:
The work described in this paper was supported by a grant from Kettleman Hills case history. J Geotech Geoenviron Eng 1998;124(12):
1167–76.
the Research Grants Council of the Hong Kong Special Administra- [31] Zhang J, Tang WH, Zhang LM. Efficient probabilistic back-analysis of slope
tive Region, China [Project No. 9041484 (CityU 110109)] and a stability model parameters. J Geotech Geoenviron Eng 2010;136(1):99–109.
strategic research grant from City University of Hong Kong (Project [32] Walpole RE, Myers RH, Myers SL. Probability and Statistics for Engineers and
Scientists. 6th ed. Upper Saddle River (New Jersey): Prentice Hall; 1998.
No. 7002455). The financial support is gratefully acknowledged. [33] Au SK. Reliability-based design sensitivity by efficient simulation. Comput
Struct 2005;83(14):1048–61.
Reference [34] Ang HS, Tang WH. Probability concepts in engineering: emphasis on
applications to civil and environmental engineering. 2nd ed. New York: John
Wiley and Sons; 2007.
[1] Tang WH, Yucemen MS, Ang HS. Probability based short-term design of slope.
[35] Au SK, Wang Y, Cao Z. Reliability analysis of slope stability by advanced
Can Geotech J 1976;13:201–15.
simulation with spreadsheet. In: Proceedings of The Second International
[2] Christian JT, Ladd CC, Baecher GB. Reliability applied to slope stability analysis.
Symposium on geotechnical safety and risk. Gifu (Japan): CRC Press; 2009. p.
J Geotech Eng 1994;120(12):2180–207.
275–80.
[3] Hassan AM, Wolff TF. Search algorithm for minimum reliability index of earth
[36] Au SK, Cao Z, Wang Y. Implementing advanced Monte Carlo simulation under
slopes. J Geotech Geoenviron Eng 1999;125(4):301–8.
spreadsheet environment. Struct Saf 2010;32:281–92.
[4] Suchomel R, Masin D. Comparison of different probabilistic methods for
[37] Metropolis N, Rosenbluth A, Rosenbluth M, Teller A. Equations of state
predicting stability of a slope in spatially variable. Comput Geotech
calculations by fast computing machines. J Chem Phys 1953;21(6):1087–92.
2010;37(1–2):132–40.
[38] Sturges H. The choice of a class-interval. J Am Stat Assoc 1926;21:65–6.
[5] Low BK, Tang WH. Reliability analysis of reinforced embankments on soft
[39] Duncan JM, Wright SG. Soil strength and slope stability. 1st ed. New York: John
ground. Can Geotech J 1997;34(5):672–85.
Wiley and Sons; 2005.
[6] Low BK, Gilbert RB, Wright SG. Slope reliability analysis using generalized
[40] Ladd CC, Dascal O, Law KT, Lefebrve G, Lessard G, Mesri G, et al. In: Report of
method of slices. J Geotech Geoenviron Eng 1998;124(4):350–62.
the subcommittee on embankment stability – annex II. Committee of
[7] Low BK. Practical probabilistic slope stability analysis. In: Proceeding of 12th
specialists on sensitive clays on the NBR complex. Société d’Energie de la
Panamerican conference on soil mechanics and geotechnical engineering and
BaieJames, Montréal, Que; 1983.
39th US rock mechanics symposium, M.I.T., Cambridge (MA): Verlag Gluckauf
[41] Soulié M, Montes P, Silvestri V. Modelling spatial variability of soil parameters.
Gmbh Essen; 2003. p. 2777–84.
Can Geotech J 1990;27:617–30.
[8] Low BK, Tang WH. Efficient spreadsheet algorithm for first-order reliability
[42] Palisade Corporation. @Risk and the Decision Tools Suite Version 5.5.1. http://
method. J Eng Mech 2007;133(2):1378–87.
www.palisade.com.2010.
[9] Hong HP, Roh G. Reliability evaluation of earth slopes. J Geotech Geoenviron
[43] Phoon KK, Kulhawy FH. Evaluation of geotechnical property variability. Can
Eng 2008;134(12):1700–5.
Geotech J 1999;36(4):625–39.
[10] El-Ramly H. Probabilistic analysis of landslide hazards and risks bridging
theory and practice. PHD Thesis of University of Alberta; 2001.

S-ar putea să vă placă și