Está en la página 1de 8

Computers and Geotechnics 37 (2010) 905–912

Contents lists available at ScienceDirect

Computers and Geotechnics


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

Back analysis of slope failure with Markov chain Monte Carlo simulation
L.L. Zhang a, J. Zhang b,c,d,*, L.M. Zhang d, W.H. Tang d
a
Department of Civil Engineering, Shanghai Jiaotong University, 1954 Huashan Road, Shanghai, China
b
Key Laboratory of Geotechnical and Underground Engineering of Ministry of Education, Tongji University, Shanghai 200092, China
c
Department of Geotechnical Engineering, Tongji University, Shanghai 200092, China
d
Department of Civil and Environmental Engineering, The Hong Kong University of Science and Technology, Clear Water Bay, Hong Kong

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

Article history: Field observed performance of slopes can be used to back calculate input parameters of soil properties
Received 23 February 2010 and evaluate uncertainty of a slope stability analysis model. In this paper, a new probabilistic method
Received in revised form 19 July 2010 is proposed for back analysis of slope failure. The proposed back analysis method is formulated based
Accepted 26 July 2010
on Bayes’ theorem and solved using the Markov chain Monte Carlo simulation method with a Metropo-
Available online 21 August 2010
lis–Hasting algorithm. The method is very flexible as any type of prior distribution can be used. The
method is also computationally efficient when a response surface method is employed to approximate
Keywords:
the slope stability model. An illustrative example of back analysis of a hypothetical slope failure is pre-
Slope stability
Back analysis
sented. Effects of jumping distribution functions and number of samples on the efficiency of Markov
Reliability chains are studied. It is found that the covariance matrix of the jumping function can be set to be one half
Bayesian analysis of the covariance of the prior distribution to achieve a reasonable acceptance rate and that 80,000 sam-
Monte Carlo simulation ples seem to be sufficient to obtain robust posterior statistics for the example. It is also found that the
correlation of cohesion and friction angle of soil does not affect the posterior statistics and the remedia-
tion design of the slope significantly, while the type of the prior distribution seems to have much influ-
ence on the remediation design.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction judgment and field measurements can be incorporated into the


back analysis in a consistent way. Multiple sets of slope stability
A slope failure implies that the factor of safety of the slope at parameters can also be back calculated simultaneously for slopes
the moment of failure is unity. Based on this information, back involving several types of soil.
analysis is often carried out to improve knowledge on slope stabil- Several probabilistic approaches for geotechnical back analysis
ity parameters, such as soil shear strength parameters and pore are available in the literature: least-squares method, maximum
water pressure parameters at the moment of slope failure. Using likelihood method and Bayesian method. For example, Ledesma
back analysis, important factors that may not be well represented et al. [7] performed geotechnical back analyses for a tunnel excava-
in laboratory testing, such as soil heterogeneity and influence of tion and a large underground cavern based on the maximum like-
fissures and structural fabric on soil shear strength, can be incorpo- lihood approach. The method of maximum likelihood is applicable
rated. Methodologies used for back analysis can be classified into when the number of observed data is larger than the number of
two groups, i.e., deterministic methods [1,2] and probabilistic parameters to be updated. Honjo et al. [8] employed an extended
methods [3–6]. In a deterministic method, the slope stability mod- Bayesian method to back analyze an embankment on soft clay,
el is usually believed or assumed accurate, and the purpose of back which is also based on the assumption that the number of ob-
analysis is to find a set of parameters that would result in the slope served data is larger than the number of parameters to be updated.
failure. In a probabilistic back analysis, however, it is recognized Luckman et al. [3] performed Bayesian updating for estimating
that the slope stability model may not be perfectly accurate and pore pressures using the first order reliability formulations. Gilbert
numerous combinations of slope stability parameters may result et al. [5] and Chowdhury et al. [4] used discrete probability distri-
in slope failure. Using a Bayesian probabilistic back analysis ap- butions in Bayesian updating of uncertain parameters. In their
proach, various sources of information including engineering studies, a continuous random variable should be approximated
as a discrete variable with sufficient small intervals if accurate reli-
ability evaluation is required. Zhang et al. [6] proposed two effi-
* Corresponding author at: Key Laboratory of Geotechnical and Underground
cient methods based on a system identification approach derived
Engineering of Ministry of Education, Tongji University, Shanghai 200092, China.
E-mail addresses: lulu_zhang@sjtu.edu.cn (L.L. Zhang), cezhangjie@gmail.com from Bayesian theory for probabilistic back analysis of slope fail-
(J. Zhang), cezhangl@ust.hk (L.M. Zhang), wtang@ust.hk (W.H. Tang). ures. The probabilistic back analysis methods based on the system

0266-352X/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compgeo.2010.07.009
906 L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912

identification approach are most appropriate when the prior distri- back analysis. According to Bayes’ theorem, the posterior probabil-
bution follows a multivariate normal distribution. ity density function of h is
In the probabilistic back analysis of a slope failure, the observed  
gðhÞ þ le  1
data is that the factor of safety at the moment of slope failure is f ðhjFÞ ¼ k/ f ðhÞ ð4Þ
equal to unity. However, there are often several input parameters
re
to be updated. Hence, the number of observed data is often less where k is a normalization constant to make the probability density
than the number of variables to be updated. Also, the prior infor- function valid.
mation could be non-normal. In this paper, a new method for back When conjugate priors can be adopted, considerable mathemat-
analysis of slope failure is proposed based on Bayes’ theorem and ical simplification of Bayesian updating can be achieved. A list of
solved using Markov chain Monte Carlo (MCMC) simulation. The conjugate distributions is summarized in Ang and Tang [9]. How-
advantages of using MCMC simulation are that it does not require ever, the method of conjugate priors is only applicable to specific
a number of data far greater than the number of variables to be up- prior distributions and likelihood functions, but cannot be used
dated and it can handle any type of prior distribution. A possible to obtain the posterior distribution in Eq. (4). In this study, the
limitation of MCMC simulation is that it requires iterative evalua- Markov chain Monte Carlo (MCMC) simulation method is adopted
tion of the slope stability model, which could be computationally to solve the Bayesian problem in Eq. (4). The advantages of MCMC
intensive. To overcome this problem, a response surface is used simulation include: (1) it does not require that the number of ob-
to approximate the slope stability model, and then MCMC simula- served data is larger than the number of variables to be updated;
tion is carried out based on the fitted response surface. By coupling (2) it can consider any type of prior information; and (3) the sam-
the MCMC simulation with the response surface methodology, the ples from MCMC simulation can be directly used for slope remedi-
proposed method becomes computationally efficient. ation analysis, as will be shown in the illustrative example.
To illustrate the proposed method, an example of back analysis
of a hypothetical slope failure is presented. To obtain efficient Mar- 2.2. MCMC simulation and Metropolis algorithm
kov chains for this example problem, the effects of the size of
jumping distribution and the number of samples on efficiency of The basic idea of MCMC simulation is that drawing samples
Markov chains are first studied. Then the effect of prior distribution from an arbitrary distribution and then correcting those samples
on the back calculated statistics of uncertain parameters and the to better approximate and finally converge to the target posterior
remediation design of the slope are presented. Finally, results using distribution. The samples are drawn sequentially with the distribu-
a simplified method of back analysis based on the system identifi- tion of the current value depending on the last value drawn, which
cation method are compared with those obtained using the pro- forms a Markov chain. The Gibbs sampler and the Metropolis algo-
posed method for verification. rithm are two basic ways to build Markov chains [10,11]. In this
paper, the Metropolis algorithm [12] is adopted to conduct MCMC
2. Proposed method for back analysis simulation. The Metropolis algorithm can be written as follows:

2.1. Probabilistic back analysis based on Bayesian framework 1. Let h0 denote the starting point of a Markov chain. h0 can be ran-
domly drawn from a starting distribution or simply chosen
First, a model correction factor e is used to characterize model deterministically.
uncertainty, which is defined as the difference between the actual 2. For i = 1, 2, . . . , n:
performance and the model prediction: (a) Sample a candidate h from a distribution J(h|h(i1)), where
h is dependent on h(i1). This distribution J(|) is called the
y ¼ gðhÞ þ e ð1Þ jumping distribution or transition kernel of the Markov
chain. The jumping distribution is symmetric in the Metrop-
where y = actual factor of safety; g(h) = predicted factor of safety olis algorithm.
using a given slope stability model; and h = vector of uncertain in- (b) Calculate the ratio of the densities
put parameters. Assume that e follows the normal distribution with
a mean of le and a standard deviation of re. If le and re are known qðh jyÞ
and the actual factor of safety of the slope is equal to Y, the likeli- r¼ ð5Þ
qðhði1Þ jyÞ
hood function is the conditional probability density function of h
as follows: where q(h|y) is the unnormalized posterior density function.
  In this paper,
gðhÞ þ le  Y  
Lðhjy ¼ YÞ ¼ / ð2Þ gðhÞ þ le  1
re qðhjyÞ ¼ / f ðhÞ: ð6Þ
re
where / is the probability density function of a standard normal
variable. In theory, a slope failure implies that the factor of safety
(c) Set h(i) = h with a probability of min(r, 1); otherwise set
of the slope at the moment of failure is equal to unity. In practice,
h(i) = h(i1), which means the jump is not accepted and this
there might be uncertainties in defining and identifying slope fail-
counts as an iteration in the algorithm.
ure. In this study, it is assumed that the slope failure is well defined
(d) Stop the iteration if i = n; otherwise i = i + 1 and go to Step (a).
and identified such that the uncertainties associated with slope fail-
ure definition and identification are minimized. In such a case, the
It can be proven that the Markov chain sample h(i) at the ith step
likelihood function, i.e., the chance to observe the slope failure
approaches the target distribution q(h|y) as i ? 1 [12]. Once the
event is therefore:
posterior samples of h are obtained, the updated probability of fail-
 
gðhÞ þ le  1 ure of a slope can be evaluated by counting the number of the pos-
LðhjFÞ ¼ / ð3Þ terior samples resulting in g(h) + e < 1.
re
The aforementioned Step (c) is basically the acceptance rule for
where F denotes the event of observing slope failure. Let f(h) denote the Metropolis algorithm. If the jump increases the posterior den-
a distribution which represents the knowledge about h prior to the sity, the density ratio is larger than 1. Therefore, h is accepted as
L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912 907

h(i) with a probability of 1. If the jump decreases the posterior den- based on laboratory tests [17] are summarized in Table 1. The de-
sity, the density ratio is smaller than 1. A random number in the sign ground water table shown in Fig. 1 was based on the water le-
interval of [0, 1] is generated. If the random number is less than vel observed in exploratory boring. Investigation after the slope
the density ratio r, h is accepted as h(i). Otherwise, h is rejected failure indicated that the slip surface passed through the bottom
and h(i) is set to be h(i1). In other words, the algorithm always ac- of the highly plastic clay layer. Measurements from the piezome-
cepts steps that increase the posterior density but only sometimes ters installed after the slope failure showed that the ground water
accepts downward steps. A good jumping distribution should be level was essentially as assumed in the design. For the conditions
easily sampled and the jumps should not be rejected too fre- shown in Fig. 1 and using the shear strength parameters shown
quently, otherwise the random walk wastes too much time stand- in Table 1, the computed factor of safety is 1.41, but the slope failed
ing still. How to select the appropriate jumping distribution [17]. We assume that the slope failure was caused by changes in
function for efficient probabilistic back analysis of slope failure will the slope profile and that the soil strength parameters did not
be demonstrated in the illustrated example. change.

2.3. Response surface methodology 3.1. Model uncertainty and prior knowledge about uncertain
parameters
MCMC simulation requires repeatedly evaluating the slope sta-
bility model g(h), which could be computationally intensive. As In this example, a probabilistic back analysis is conducted for
shown in Zhang et al. [6], the relationship between the factor of this slope. First, the shear strength parameters of the two soil lay-
safety and model input parameters along a specific slip surface ers, c1, /1, and /2 are considered as uncertain variables, i.e., h = {c1,
are often quite linear and can be approximated with sufficient /1, /2}. The mean values of the three uncertain variables are as-
accuracy by a second order polynomial function. In this study, sumed to be the same as those design values based on laboratory
g(h) is approximated with a second order polynomial function of tests. The variability of these variables is not reported in Duncan
the following form [17]. According to Phoon and Kulhawy [18], the COV of the friction
angle for sandy soils and clayey soils is between 5% and 20%. The
X
r X
r
y ¼ b0 þ bi hi þ brþi h2i ð7Þ COV of soil cohesion is usually greater than 20% [19]. In this illus-
i¼1 i¼1 trative example, the prior COV values of c1, /1, and /2 are assumed
to be 0.20, 0.15, and 0.15, respectively, which are summarized in
where r = the dimension of h; and b0, b1, . . . , b2r = regression coeffi- Table 2. It is assumed that c1, /1, and /2 are statistically indepen-
cients. The regression coefficients can be obtained by equating the dent and follow the multivariate lognormal distribution. A sensi-
factors of safety calculated using Eq. (7) with those calculated using tivity analysis will be carried out to study the effect of prior
the numerical slope stability model. The above methodology is of- distribution on back analysis.
ten called ‘‘response surface method” and has been used in geotech- The method of Morgenstern and Price [20] is used to calculate
nical reliability analyses by several researchers [13–16]. After the the factor of safety of this cut slope. As the emphasis of this paper
second order polynomial function is calibrated, it will be used in is to improve the knowledge on slope stability parameters through
MCMC simulation for evaluating the likelihood function. The accu- back analysis, the probability density function of the model uncer-
racy of approximating the slope stability model will be examined tainty variable e is not updated in the back analysis. According to
and discussed in the following example. Christian et al. [21], the model uncertainty of the simplified meth-
od of Bishop [22] has a mean of 0.05 and a standard deviation of
3. Illustrative example 0.07. Since the results from Bishop’s simplified method are usually
close to those from the Morgenstern and Price method, it is as-
The illustrative example is adapted from Duncan [17], where a sumed that the model uncertainty of the Morgenstern and Price
back analysis of a hypothetical slope failure was carried out in a method can also be modeled as a normally distributed random var-
deterministic way. The slope was excavated by trimming back iable with a mean of 0.05 and a standard deviation of 0.07, i.e.,
the original hill slope to a steeper configuration. The ground inves- le = 0.05 and re = 0.07.
tigation showed the presence of layers of sandy clay, highly plastic It should be noted the back analysis results depend on the accu-
clay, and dense to very dense sand. The soil profile used in design is racy of the identified soil profile including the accurate location of
shown in Fig. 1. Properties of sandy clay and highly plastic clay the slip surface as well as the understanding of the failure mecha-

30
Vertial drain well
Design ground water table
Vetical distance (m)

20
Horizontal drain

10
Slip surface
Sandy clay
Highly plastic clay
0
Dense to very dense sand

-10
0 25 50 75 100
Horizontal distance (m)

Fig. 1. Cross section of a hypothetical slope (modified from Duncan [17]).


908 L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912

Table 1 a multivariate normal distribution is used as the jumping distribu-


Material properties reported in Duncan [17]. tion. Its mean point is the current point in the Markov chain, and
Sandy clay Highly plastic clay the covariance matrix Chj is equal to nCh, where Ch is the covariance
c1 (kPa) /1 (deg) c1 (kN) c2 (kPa) /2 (deg) c2 (kN) matrix of the prior distribution of h and n is a scaling factor. To
study the effect of scaling factor n, we established three Markov
14.4 35.00 20.8 0.00 25.00 20.8
chains with n = 0.01, n = 0.5, and n = 10, respectively. As an exam-
ple, the values of 2000 samples of c1 generated from the three
chains are shown in Fig. 3a–c. It can be seen that n has an impor-
Table 2 tant effect on the behavior of the Markov chain. When n = 0.01 as
Prior statistics of random variables.
shown in Fig. 3a, the Markov chain moves slowly in the posterior
c1 (kPa) /1 (deg) /2 (deg) space, taking a long time for the chain to travel from one side of
Mean 14.4 35.0 25.0 the posterior space to the other side. When n = 10, there are many
COV 0.20 0.15 0.15 horizontal traces in Fig. 3c, indicating the current point in a Markov
Standard deviation 2.88 5.25 3.75 chain is the same as its previous point, i.e., the chain does not
move. This is because when the value of n is too large, the random
samples drawn from the jumping function will frequently fall into
nism. The back analysis results may be meaningless if the analysis regions with a small posterior density, and hence the suggested
is carried out based on an unrealistic profile or based on a wrong samples are frequently rejected. For comparison, when n = 0.5,
failure mechanism. It is assumed in this study that both the soil the Markov chain can move actively in the posterior space as
profile relevant to the back analysis and the location of the slip sur- shown in Fig. 3b.
face are well defined through post-failure investigation. Moreover,
it is assumed that the slope stability model can correctly reflect the
failure mechanism. Back analysis of slope failures when the soil
profile is uncertain is out of the scope of this study. 20

3.2. Construction of response surface


15
c1 (kPa)

To construct a second order polynomial function to approxi-


mate the slope stability model, the factors of safety of the slope
along the failure surface are first estimated at the following seven
10
design points: {lc1, l/1, l/2}, {lc1 ± 3rc1, l/1, l/2}, {lc1, l/1 ± 3r/1,
l/2}, and {lc1, l/1, l/2 ± 3r/2}. The slope stability model can then
be approximated by the following equation:
5
gðhÞ ¼ 5:767  102 þ 8:816  103 c1 þ 5:132  103 /1 0 500 1000 1500 2000
2 Sample number
þ 3:561  10 /2  6:698  106 c21 þ 1:148  10 4
/21
þ 3:793  104 /22 ð8Þ 20

To check the adequacy of the slope stability model, 16 more


points are randomly drawn from the space of h, and the factors
of safety corresponding to these parameters are evaluated with 15
c1 (kPa)

both the Morgenstern–Price method and with the second order


polynomial function, as shown in Fig. 2. The results from the two
approaches are very close. Hence, the following probabilistic back 10
analysis is based on the second order polynomial function.

3.3. Back analysis using MCMC simulation


5
0 500 1000 1500 2000
To employ MCMC simulation for back analysis of slope failure, Sample number
the jumping functions should first be determined. In this example,
20

2.0
15
c1(kPa)

1.5
Fs2

10
1.0

0.5 5
0.5 1.0 1.5 2.0 0 500 1000 1500 2000
Fs1 Sample number
Fig. 2. Comparison of factors of safety calculated using the response surface Eq. (7) Fig. 3. Samples for c1 in a Markov chain when n takes different values: (a) n = 0.01;
(Fs1) with those calculated using the Morgenstern–Price method (Fs2). (b) n = 0.5; and (c) n = 10.
L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912 909

Table 3
Mean values and standard deviations of the posterior statistics estimated from 10 Markov chains with n = 0.01.

l00c1 l00/1 l00/2 r00c1 r00/1 r00/2 q00c1;/1 q00c1;/2 q00/1;/2


5000 Mean 13.05 28.31 15.91 2.34 3.18 1.25 0.06 0.19 0.37
Std 0.86 0.75 0.23 0.58 0.33 0.08 0.25 0.10 0.11
10,000 Mean 12.90 28.65 15.89 2.38 3.47 1.29 0.06 0.18 0.40
Std 0.50 0.03 0.03 0.04 0.02 0.02 0.02 0.01 0.01
20,000 Mean 12.82 28.67 15.87 2.32 3.49 1.31 0.03 0.18 0.42
Std 0.28 0.34 0.09 0.08 0.10 0.05 0.09 0.07 0.04
40,000 Mean 12.99 28.61 15.89 2.42 3.52 1.32 0.08 0.18 0.41
Std 0.15 0.30 0.06 0.09 0.18 0.04 0.06 0.04 0.03
80,000 Mean 12.81 28.70 15.88 2.42 3.51 1.31 0.06 0.18 0.40
Std 0.12 0.20 0.05 0.11 0.20 0.02 0.03 0.03 0.03

Table 4
Mean values and standard deviations of the posterior statistics estimated from 10 Markov chains with n = 0. 5.

l00c1 l00/1 l00/2 r00c1 r00/1 r00/2 q00c1;/1 q00c1;/2 q00/1;/2


5000 Mean 12.76 28.75 15.89 2.29 3.46 1.28 0.05 0.17 0.41
Std 0.10 0.21 0.04 0.09 0.10 0.03 0.05 0.05 0.03
10,000 Mean 12.82 28.65 15.91 2.38 3.48 1.31 0.07 0.18 0.41
Std 0.11 0.13 0.05 0.08 0.07 0.02 0.03 0.03 0.02
20,000 Mean 12.84 28.64 15.90 2.41 3.47 1.30 0.06 0.18 0.39
Std 0.09 0.12 0.03 0.06 0.07 0.02 0.02 0.02 0.01
40,000 Mean 12.84 28.65 15.89 2.38 3.47 1.29 0.06 0.18 0.40
Std 0.07 0.03 0.03 0.04 0.02 0.02 0.02 0.01 0.01
80,000 Mean 12.85 28.64 15.89 2.40 3.49 1.30 0.06 0.18 0.40
Std 0.05 0.03 0.01 0.04 0.04 0.01 0.01 0.01 0.01

Table 5
Mean values and standard deviations of the posterior statistics estimated from 10 Markov chains with n = 10.

l00c1 l00/1 l00/2 r00c1 r00/1 r00/2 q00c1;/1 q00c1;/2 q00/1;/2


5000 Mean 12.87 28.54 16.01 2.38 3.46 1.35 0.11 0.19 0.38
Std 0.33 0.76 0.11 0.29 0.26 0.17 0.11 0.13 0.15
10,000 Mean 12.81 28.45 15.94 2.32 3.39 1.30 0.08 0.19 0.37
Std 0.29 0.37 0.19 0.17 0.24 0.09 0.09 0.07 0.08
20,000 Mean 12.81 28.45 15.94 2.32 3.39 1.30 0.08 0.19 0.37
Std 0.29 0.37 0.19 0.17 0.24 0.09 0.09 0.07 0.08
40,000 Mean 12.90 28.67 15.85 2.41 3.49 1.31 0.10 0.17 0.40
Std 0.15 0.18 0.06 0.11 0.11 0.04 0.08 0.05 0.04
80,000 Mean 12.84 28.73 15.90 2.40 3.52 1.29 0.07 0.19 0.40
Std 0.11 0.15 0.05 0.09 0.08 0.05 0.03 0.03 0.02

To investigate the effect of n on posterior statistics, ten Markov the posterior statistics decrease as expected. It means the esti-
chains with n = 0.01 are first started from different initial points. mates of the posterior statistics become more accurate when the
Based on the samples from each chain, the posterior statistics of number of the samples in a chain increases. Also noted is that
c1, /1, and /2, i.e., the mean values, standard deviations and corre- the variability of the posterior correlation coefficients is larger than
lation coefficients, can be calculated. It is expected that, when the that of the posterior mean values and standard deviations, indicat-
number of samples drawn from each chain is sufficiently large, the ing that the posterior correlation coefficients are more difficult to
inferred posterior statistics from different chains will be similar. As estimate accurately. For comparison, ten Markov chains with
at the beginning stage the Markov chain may not reach the equilib- n = 0.5 and 10 Markov chains with n = 10 are established. The mean
rium or stationary state, it is often suggested that the first few sam- values and standard deviations of the posterior statistics inferred
ples should be discarded and not used for posterior statistics from the Markov chains when n = 0.5 and n = 10 are summarized
inference. However, there is no systematic and universal way to in Tables 4 and 5, respectively. It can be seen that the estimated
calculate the number of samples that should be discarded [11]. It posterior statistics when n takes different values are generally con-
is also suggested that one does not indeed need to discard the ini- sistent, indicating that the Metropolis algorithm is quite robust. It
tial samples in a Markov chain, since when the number of samples can also be observed that, when the numbers of samples are the
drawn from a Markov chain increases, the effects of the first few same, the posterior statistics inferred from the Markov chains with
samples on the posterior inference results gradually diminish n = 0.5 have smaller variability. Thus, to achieve the same accuracy
[23]. In this study, the first 500 samples are discarded in each with a Markov chain with n = 0.5, more samples should be drawn
chain. Table 3 compares the mean values and standard deviations from Markov chains when n = 0.01 or n = 10 are adopted. In other
of the posterior statistics of c1, /1, and /2 inferred from the ten words, the Markov chain with n = 0.5 is more efficient in collecting
Markov chains when the number of samples drawn from each representative samples from the posterior distribution.
chain are 5000, 10,000, 20,000, 40,000, and 80,000, respectively. An appropriate value of n for the jumping distribution is in fact
As the number of samples increases, the standard deviations of closely related to the acceptance rate, which is defined as the
910 L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912

1.0 It can be seen that the acceptance rate decreases with n. Based on
Fig. 4, when the n values adopted are 0.01, 0.5, and 10, the corre-
Acceptance rate

0.8 sponding acceptance rates are about 0.86, 0.30, and 0.02, respec-
tively. The efficiency of the Markov chain with n = 0.5 is higher
0.6
than those based on n = 0.01 and n = 10 as shown in Tables 3–5.
0.4 Therefore, it is consistent with the suggestions made by Gelman
et al. [12]. In practical application of MCMC simulation, one could
0.2 determine the relationship between n and the acceptance rate first,
and select a n that can result in an acceptance rate in the range of
0.0 0.2–0.4 for MCMC simulation.
0.01 0.1 1 10
ξ Table 4 also shows that when the number of samples is 80,000,
the variability of the posterior statistics inferred from different
Fig. 4. Effect of scaling factor of jumping distribution on the acceptance rate of Markov chains is very small. Hence, 80,000 samples are sufficient
MCMC. to obtain a robust estimation of the posterior statistics. The Markov
chains with 80,000 samples when n = 0.5 are shown in Fig. 5. The
proportion of samples accepted in the Markov chain. Gelman et al. histograms of these samples are presented in Fig. 6. The mean val-
[12] recommended that when the acceptance rate is around 20– ues of the three soil parameters all decrease, indicating that the
40%, the Markov chain is most efficient in collecting samples of prior distributions overestimate the soil strength parameters. To
the posterior density function. For the slope problem studied here, confirm this observation, the failure probability of the slope with
the relationship between acceptance rate and n is plotted in Fig. 4. the ground water table at the moment of failure is calculated based

12000
30
Number of samples μc1” = 12.85
σc1” = 2.40
8000
20
c1 (kPa)

4000
10

0
0 0 5 10 15 20 25 30
0 2 4 6 8 c1 (kPa)
Sample number x104
12000
55
μφ1” = 28.64
Number of samples

σφ1” = 3.49
8000
φ 1 (degree)

40

4000
25

0
10 15 20 25 30 35 40 45
0 2 4 6 8 φ1
Sample number φ1 (degree)
x104
12000
25
μφ2” = 15.89
Number of samples

σφ2” = 1.30
8000
φ2 (degree)

20

4000
15

0
10 10 12 14 16 18 20 22
0 2 4 6 8
φ2
φ2 (degree)
Sample number x104

Fig. 5. Plot of samples with MCMC simulation steps (n = 0.5). Fig. 6. Histograms of posterior c1, /1 and /2 (n = 0.5, 8000 samples).
L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912 911

5 3.5. Effect of prior distribution


Prior 1
Prior 2
Reliability index
4 Prior 3 In the previous sections, c1, /1, and /2 are assumed to be statis-
3 tically independent and follow the multivariate lognormal distri-
bution. In the literature [19,25], cohesion and friction angle of
2 soil are sometimes reported to be negatively correlated. In some
cases, the distributions of soil cohesion and friction angle are sug-
1
gested to follow the normal distribution. To examine the effect of
0 prior distribution, the slope is back analyzed with different
12 14 16 18 20 assumptions about prior distribution. Denote the previous case
Ground water level (m) where prior c1, /1, and /2 are assumed to be statistically indepen-
Fig. 7. Reliability indices of the slope at different ground water levels beneath the
dent and follow the multivariate lognormal distribution as Prior 1.
crest of the slope. For the second case (Prior 2), it is assumed that the three variables
follow the multivariate lognormal distribution with the correlation
coefficient between c1 and /1 equal to 0.5. For the third case
(Prior 3), the three variables follow the multivariate normal distri-
Table 6 bution and are statistically independent. The back calculated val-
Comparison of posterior statistics with different prior distributions.
ues of the mean, standard deviation, and correlation coefficients
l00c1 l00/1 l00/2 r00c1 r00/1 r00/2 q00c1;/1 q00c1;/2 q00/1;/2 of the three random variables with different assumptions about
Prior 1 12.85 28.64 15.89 2.40 3.49 1.30 0.06 0.18 0.40 their prior distributions are summarized in Table 6. The back calcu-
Prior 2 14.53 29.50 15.57 2.60 3.56 1.26 0.46 0.03 0.33 lated results change when the prior distribution changes, indicat-
Prior 3 13.31 29.74 14.50 2.84 4.70 1.76 0.04 0.23 0.57 ing that in the probabilistic back analysis the prior knowledge
about the parameters also has an important effect on the back
analysis results. Thus, it is important to obtain high quality prior
information.
on the prior distribution of h. With a Monte Carlo simulation with To examine the effect of prior distribution on the remediation
1  107 samples, the prior probability of failure is found to be design, the reliability indices of the slope at different levels of
7.7  105, indicating the slope failure is very unlikely. In reality, ground water table are also calculated when different prior distri-
however, the slope failed. Thus, the prior distribution of h used butions are used for back analysis. The results are also plotted in
for reliability analysis overestimates the soil strength parameters. Fig. 7. It can be seen that the reliability indices of the slope calcu-
Based on the posterior samples of MCMC simulation under the lated with and without considering the correlation in the prior dis-
conditions described previously, the failure probability of the slope tribution are similar, and hence the correlation coefficient between
at the failure ground water table is 9%, which is far larger than the c1 and /1 does not affect the remediation design significantly. It can
prior probability of failure. Therefore, with the information of slope also be observed that, when the multivariate normal distribution is
failure, the estimated probability of failure of the slope is in- assumed as the prior distribution (Prior 3), the calculated reliability
creased, which is more consistent with the actual situation. indices at different levels of ground water table are much smaller
than those with a prior distribution of multivariate lognormal dis-
tribution (Prior 1). The multivariate normal distribution seems to
be a conservative assumption. The reason for this is that with the
3.4. Application in remediation design
assumption of a multivariate normal distribution, the input param-
eters have more chances to take smaller values. Hence, the mean
As suggested in Duncan [17], one possible way to stabilize the
factor of safety is smaller and the probability of failure is greater.
slope is to lower down the ground water table using vertical drain
wells with horizontal drains drilled from the bottom of the slope to
drain the water from these wells by gravity. Based on the MCMC 3.6. Comparison with the method based on sensitivity analysis
samples of the posterior distributions, the reliability indices of
the slope at different levels beneath the slope crest of the lowered Zhang et al. [6] suggested a simplified method based on sensi-
ground water table are calculated and shown in Fig. 7 (the case of tivity analysis for probabilistic back analysis of slope failure. The
Prior 1). The assumptions and results of cases Prior 2 and Prior 3 simplified method is basically a system identification method
will be presented in the next section. Fell [24] reviewed landslide and can yield analytical solutions for posterior mean and covari-
risk assessment and acceptable risk and concluded that for natural ance matrix of random variables if the prior distribution is multi-
landslides a community may accept an annual specific voluntary variate normal and the prediction model is linear. In this study,
risk of loss of life in the order 103–104. For involuntary risks, par- Prior 3 assumes a multivariate normal prior distribution. As a com-
ticularly man-made risks, a risk of 105–106 may be a limiting va- parison, the results obtained using the simplified method is shown
lue. If a target reliability index of 3.0 is desired, the ground water in Table 7. It shows that the results from the MCMC simulation and
table needs to be lowered to about 16.5 m. As a comparison, Dun- the simplified method are consistent. Using the simplified method
can [17] suggested that the ground water table be lowered to based on sensitivity analysis, it is assumed that the slope stability
15.6 m based on results of deterministic back analysis. model is linear. The differences between the results from MCMC

Table 7
Comparison of prior and posterior statistics calculated using MCMC and a simplified method based on sensitivity analysis.

l00c1 l00/1 l00/2 r00c1 r00/1 r00/2 q00c1;/1 q00c1;/2 q00/1;/2


MCMC (Prior 3) 13.31 29.74 14.50 2.84 4.70 1.76 0.04 0.23 0.57
Simplified method 13.47 30.26 14.99 2.86 5.00 1.67 0.03 0.37 0.64
912 L.L. Zhang et al. / Computers and Geotechnics 37 (2010) 905–912

simulation and the simplified back analysis method based on sen- References
sitivity analysis are mainly due to nonlinearity of the slope stability
model. [1] Wesley LD, Leelaratnam V. Shear strength parameters from back-analysis of
single slips. Geotechnique 2001;51(4):373–4.
[2] Tiwari B, Brandon TL, Marui H, Tuladhar GR. Comparison of residual shear
strengths from back analysis and ring shear tests on undisturbed and remolded
4. Conclusions specimens. J Geotech Geoenviron Eng 2005;131(9):1071–9.
[3] Luckman PG, Der Kiureghian A, Sitar N. Use of stochastic stability analysis for
The research reported and findings from this paper can be sum- Bayesian back calculation of pore pressures acting in a cut at failure. In: Lind N,
editor. Proc 5th int conf on application of statistics and probability in soil and
marized as follows: struct engr, Vancouver. Ontario: University of Waterloo Press; 1987. p. 922–9.
[4] Chowdhury R, Zhang S, Flentje P. Reliability updating and geotechnical back-
(1) A probabilistic method that couples Markov chain Monte analysis. In: Jardine RJ, Potts DM, Higgins KG, editors. Advances in geotechnical
engineering: the Skempton conference. London: Thomas Telford; 2004. p.
Carlo (MCMC) simulation and the response surface method 815–21.
is presented in this paper for back analysis of slope failure. [5] Gilbert RB, Wright SG, Liedtke E. Uncertainty in back analysis of slopes:
The method is ideally suitable for back analysis of slope Kettleman Hills case history. J Geotech Geoenviron Eng 1998;124(12):1167–76.
[6] Zhang J, Tang WH, Zhang LM. Efficient probabilistic back-analysis of slope
failure, because it does not require the number of observa- stability model parameters. J Geotech Geoenviron Eng 2010;136(1):99–109.
tions being larger than the number of parameters to be [7] Ledesma A, Gens A, Alonso EE. Parameter and variance estimation in
updated, and it is applicable to any type of prior geotechnical back analysis using prior information. Int J Numer Anal
Methods Geomech 1996;20:119–41.
information.
[8] Honjo Y, Liu WT, Soumitra G. Inverse analysis of an embankment on soft clay
(2) The acceptance rate of Markov chain decreases with the size by extended Bayesian method. Int J Numer Anal Methods Geomech
of the jumping distribution. For the example investigated in 1994;18:709–34.
this study, when the covariance matrix of the jumping func- [9] Ang AHS, Tang WH. Probability concepts in engineering: emphasis on
applications to civil and environmental engineering. New York: John Wiley &
tion is one half of the covariance of the prior distribution of Sons; 2007.
the random variables, a reasonable acceptance rate of 0.3 [10] Thomas L, John SJH. Bayesian methods: an analysis for statisticians and
can be achieved. interdisciplinary researchers. New York: Cambridge University Press; 1999.
[11] Gill J. Bayesian methods: a social and behavioral sciences
(3) The number of samples that should be drawn from a Markov approach. London: Chapman & Hall; 2002.
chain can be determined via a sensitivity analysis. For the [12] Gelman BA, Carlin BP, Stem HS, Rubin DB. Bayesian data
example studied in the paper, when the number of samples analysis. London: Chapman & Hall; 2004.
[13] Wong FS. Slope reliability and response surface method. J Geotech Eng
is 80,000, the variability of posterior statistics inferred from 1985;111(1):32–53.
different Markov chains starting from various initial points [14] Xu B, Low BK. Probabilistic stability analyses of embankments based on finite-
is very small. Hence, 80,000 samples seem to be sufficient element method. J Geotech Geoenviron Eng 2006;132(11):1444–54.
[15] Babu GLS, Srivastava A. Reliability analysis of allowable pressure on shallow
to obtain robust posterior statistics for this example foundation using response surface method. Comput Geotech
problem. 2007;34(3):187–94.
(4) While the correlation of cohesion and friction angle of soil [16] Mollon G, Dias D, Soubra AH. Probabilistic analysis of circular tunnels in
homogeneous soil using response surface methodology. J Geotech Geoenviron
does not affect the remediation design significantly, the type
Eng 2009;135(9):1314–25.
of the prior distribution seems to have much influence on [17] Duncan JM. The use of back analysis to reduce slope failure risk. Civil
the remediation design. engineering practice. Spring/summer; 1999. p. 75–91.
[18] Phoon KK, Kulhawy FH. Evaluation of geotechnical property variability. Can
Geotech J 1999;36(4):625–39.
[19] Rethati L. Probabilistic solutions in geotechnics. New York: Elsevier Science
Pub. Co.; 1988.
Acknowledgements [20] Morgenstern NR, Price VE. The analysis of the stability of general slip surfaces.
Geotechnique 1965;15(1):79–93.
The work in this paper was substantially supported by grants [21] Christian J, Ladd C, Baecher G. Reliability applied to slope stability analysis. J
Geotech Eng 1994;120(12):2180–207.
from the Natural Science Foundation of China (Project No. [22] Bishop AW. The use of the slip circle in the stability analysis of slopes.
50809038), the Shanghai Educational Development Foundation Geotechnique 1955;5(1):7–17.
(Project No. 2008CG17), Ph.D. Programs Foundation of Ministry [23] Driscoll TA, Maki KL. Searching for rare growth factors using multicanonical
Monte Carlo methods. SIAM Rev 2007;49(4):673–92.
of Education of China (Project No. 200802481128), Shanghai Lead-
[24] Fell R. Landslide risk assessment and acceptable risk. Can Geotech J
ing Academic Discipline Project (Project No. B208), and the Re- 1993;31:261–72.
search Grants Coucil (RGC) of the Hong Kong Special [25] Cherubini C. Reliability evaluation of shallow foundation bearing capacity on
Administration Region, China (Porject No. 622206). c0 , /0 soils. Can Geotech J 2000;37(1):264–9.

También podría gustarte