Professional Documents
Culture Documents
Model Parameters
J. Zhang1; Wilson H. Tang, Hon.M.ASCE2; and L. M. Zhang, M.ASCE3
Abstract: Back-analysis of slope failure is often performed to improve one’s knowledge on parameters of a slope stability analysis
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
model. In a failed slope, the slip surface may pass through several layers of soil. Therefore, several sets of model parameters need to be
back-analyzed. To back-analyze multiple sets of slope stability parameters simultaneously under uncertainty, the back-analysis can be
implemented in a probabilistic way, in which uncertain parameters are modeled as random variables, and their distributions are improved
based on the observed slope failure information. In this paper, two methods are presented for probabilistic back-analysis of slope failure.
For a general slope stability model, its uncertain parameters can be back-analyzed with an optimization procedure that can be imple-
mented in a spreadsheet. When the slope stability model is approximately linear, its parameters can be back-analyzed with sensitivity
analysis instead. A feature of these two methods is that they are easy to apply. Two case studies are used to illustrate the proposed
methods. The case studies show that the degrees of improvement achieved by the back-analysis are different for different parameters, and
that the parameter contributing most to the uncertainty in factor of safety is updated most.
DOI: 10.1061/共ASCE兲GT.1943-5606.0000205
CE Database subject headings: Landslides; Slope stability; Models; Parameters; Failure investigation; Probability; Reliability;
Bayesian analysis.
Author keywords: Landslides; Slope stability; Failure investigation; Probability methods; Reliability; Bayesian analysis.
冉 冊
One common concern in back-analysis is that the prediction −1
from the theoretical slope stability model g共兲 may be subjected H TH
C兩d = + C−1 共6兲
to model error. In such a case, even if the calculated factor of 2
safety is unity, the actual factor of safety of the slope may not be
unity. To quantify the effect of model imperfection, the model
uncertainty can be modeled as a random variable, which is de-
fined as
H= 冏 冏
g共兲
=
共7兲
terizing the modeling uncertainty. For simplicity, assume fol- The back-analysis based on Eqs. 共5兲–共7兲 is called “simplified
lows the normal distribution with a mean of and a standard probabilistic back-analysis” in this paper. Compared with the
deviation of . method by optimization, the simplified method is easier to apply,
With the above assumptions, the improved distribution of as it does not involve any minimization procedure. The limitation
can also be approximated by a multivariate normal distribution of the simplified method is that it is applicable only when g共兲 is
关see Appendix I based on Tarantola 共2005兲兴. Let 兩d and C兩d largely linear. Fortunately, based on the authors’ experience, many
denote the improved mean and covariance matrix of , respec- stability models are in fact rather linear in spite of the fact that
tively. As a multivariate normal distribution can be fully deter- numerical iterations are usually involved in slope stability mod-
mined by its mean and covariance matrix, the task in the els, which was also noticed by Mostyn and Li 共1993兲.
probabilistic back-analysis is then reduced to determining 兩d The following comments are made regarding the two methods
and C兩d. proposed in this paper for back-analysis of slope failure.
For a general slope stability model g共兲, 兩d is a point that 1. In the two methods described above, the emphasis is to im-
maximizes the chance to observe the slope failure event, and it prove the knowledge on slope stability parameters through
denotes the most probable combination of parameters that had led back-analysis. The probability density function of the model
to the slope failure event. 兩d can be obtained by minimizing the uncertainty variable is not updated in the back-analysis.
following misfit function 2S共兲 关see Appendix I based on Taran- 2. The probabilistic back-analysis procedures described above
tola 共2005兲兴: are based on the assumption that the prior distribution of
can be represented or approximated by a multivariate normal
关g共兲 + − 1兴T关g共兲 + − 1兴 distribution. This assumption simplifies the implementation
2S共兲 = + 共 − 兲TC−1共 − 兲 of the back-analysis methods. If this assumption is not satis-
2
fied, the prior distribution of can be transformed into a
共2兲 multivariate normal distribution using techniques like Nataf
The improved covariance matrix of , C兩d, which describes the transformation 共Der Kiureghian and Liu 1986兲, and the trans-
magnitude of uncertainty in each component of as well as the formed parameters can then be back-analyzed using the
dependence relationships among various components of , can be above procedures. The transformation would introduce addi-
determined as follows 关see Appendix I based on Tarantola tional computational work into the back-analysis. In the ex-
共2005兲兴: amples presented in this paper, the prior distributions are
assumed to be multivariate normal.
C兩d = 冉 冊
G TG
2
+ C−1
−1
共3兲
3. The credibility of the back-analysis results depends on the
accuracy of the identified soil stratigraphy 共including the ac-
curate location of the slip surface兲 as well as the understand-
冏 冏
ing of the failure mechanism. The back-analysis results may
g共兲 be meaningless if the back-analysis is carried out based on an
G= 共4兲
=兩d
unrealistic stratigraphy or based on a wrong failure mecha-
nism. It is assumed in this study that both soil stratigraphy
where G = row vector representing the sensitivity of g共兲 with relevant to the back-analysis and the location of the slip sur-
respect to at 兩d. face are well defined through postfailure investigation. More-
The back-analysis based on the above equations is called over, it is assumed that the slope stability model g共兲 can
“probabilistic back-analysis method by optimization” in this correctly reflect the failure mechanism. Back-analysis of
paper. At a first glance, the application of this method seems not slope failures when the stratigraphy is uncertain is out of the
easy, as in Eq. 共2兲 the minimization is coupled with the slope scope of this study.
stability model g共兲. However, as will be shown later in this 4. Eqs. 共2兲–共7兲 show that the back-analyzed distribution of
paper, the minimization can be automated in a commonly avail- depends on and C, i.e., the prior knowledge about , as
able spreadsheet, thus avoiding the possible troublesome pro- well as the slope failure information. Hence, the back-
gramming work. analyzed distribution is a combination of the prior knowl-
The expressions in Eqs. 共2兲–共4兲 are indeed quite general, and edge about and the knowledge learned from the slope
are applicable no matter g共兲 is linear or not. However, when failure event. As will be seen later in this paper, a change in
g共兲 is approximately linear, 兩d and C兩d can be determined prior distribution will cause changes in the back-analyzed
analytically with the following equations without resorting to the distribution of . This implies that, in a probabilistic back-
minimization procedure 关see Appendix II based on Tarantola analysis, information from other sources on the slope stabil-
共2005兲兴: ity parameters is also important.
Elevation (m)
points are more separated. However, among various points, IV by FMSW (2000)
30
the point that gives the minimal value of 2S共兲 is most plau- II
sible because it has the maximum likelihood, and the poste- V
rior mean and covariance matrix should be estimated at this
II V/IV II V
point. II
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
7. It is known that soil property varies spatially. If spatial vari- Horizontal distance (m)
250
冤 冥
problem, the point that minimizes the misfit function is unique,
32 0 0
and this point is then regarded as 兩d. Based on 兩d and C兩d
C = 0 12 0 共8兲 calculated from the optimization method, the improved distribu-
0 0 0.12 tion for each parameter can be obtained. The prior and the im-
proved probability density functions for c, , and ru are compared
The units for c and in the above and C are kPa and degree, in Fig. 4.
respectively. To be clear, the prior distribution used here is called Although in this study the factor of safety is calculated using
Prior 1. the method of Spencer 共1967兲, other slope stability analysis meth-
In the method of Spencer 共1967兲, all the three equilibrium ods can also be used in a similar fashion as long as they can be
conditions 共horizontal force equilibrium, vertical force equilib- implemented in a spreadsheet. In fact, many slope stability mod-
rium, and moment equilibrium兲 are satisfied. Christian et al. els have been implemented in spreadsheets, such as Bishop’s sim-
共1994兲 suggested that the model uncertainty of the simplified plified method 共Bishop 1955; El-Ramly et al. 2002兲, Janbu’s
method of Bishop 共1955兲 has a mean of 0.05 and a standard generalized method 共Janbu 1987; Low and Tang 1997兲, and Chen-
deviation of 0.07. As the results from Bishop’s simplified method Morgenstern’s method 共Chen and Morgenstern 1983, Low et al.
are usually close to rigorous methods that satisfy all three equi- 1998兲. Although in this paper it is shown that the probabilistic
librium conditions, it is assumed that the model uncertainty of the back-analysis method by optimization can be implemented in a
1
Step 1: Build the slope stability model
2 Slice# α b h u c tanφ s A B C D Q Q*cos( α - θ )
3 1 1.428 1.058 3.672 6.106 7.676 0.78 371.5 0.105 0.075 -0.03 9E-04 -48.8 -36.1
4 2 1.025 1.058 5.963 9.915 7.676 0.78 371.5 0.49 0.133 -0.13 0.044 -48.5 -45.7
21 19 0.315 1.058 2.091 3.477 7.676 0.78 371.5 0.885 -0.27 -0.03 0.08 33.05 30.76
22 20 0.098 1.058 0.717 1.193 7.676 0.78 371.5 0.826 -0.43 -0 0.044 42.37 35.18
23 Posterior mean, obtained by ∑Q ∑Q*cos(a- θ )
24 γ H θ,μθ|θ|d minimizing N33 1E-07 1E-07
25 17 20.66 c 7.676
26 φ 37.96 β Fs Fs
27 ru 0.171 0.689 0.956 0.956
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
28
29
Step 2: Calculate the posterior mean Eq. (2)
30
31 μθ Cθ
32 8 9 0 0 με σε Observed data 2S( θ)
33 38 0 1 0 0.05 0.07 1 0.067
34 0.15 0 0 0.01
35
36 Posterior covariance,
Step 3: Calculate the posterior covariance matrix obtained by Eq. (3)
37
Sensitivity vector, Eq. (4)
38 Cθ|d
39 7.435 -0.17 0.103
40 G 0.027 0.027 -1.61 -0.17 0.981 0.011
41 0.103 0.011 0.003
42
43 Notes:
44 1. γ and H denote the unit weight of soil and the slope height, respectively;
45 2. β and F s denote the angle of interslice force and factor of safety of the slope, respectively;
46 3. α , b , h , and u are the base inclination angle, width, height and pore pressure of a slice, respectively;
47 4. s, A, B, C, D are five intermediate variables that facilitate coding in the spreadsheet;
48 5. When ∑Q = 0 and ∑Q*cos(a- θ ) = 0, force and moment equirlibriums are satisfied, respectively.
49
spreadsheet, we do not mean that this method must be imple- constraint of factor of safety equal to unity at the moment of slope
mented in the spreadsheet. Other tools can also be used. failure. More specifically, in the posterior distribution a higher
Fig. 5 shows the relationships between the factor of safety and pore water pressure should be accompanied by higher shear
c, , and ru around . These relationships are rather linear in the strength parameters such that the calculated factor of safety is
likely ranges of these parameters. Therefore, the slope stability close to unity. Since the variance of is very small, the variation
parameters may also be back-analyzed using the simplified in shear strength parameters is dominated by the change in cohe-
method. To implement the simplified method, one can perform a
sion. This makes the correlation coefficient between c and ru
sensitivity analysis at point to get g共兲 and H first, and then
larger than the correlation coefficients among other variables.
substitute them into Eqs. 共5兲 and 共6兲 to calculate 兩d and C兩d,
respectively. As g共兲 and H can be calculated using a conven-
tional slope stability program, the simplified method can be Information Allocation in Back-Analysis
implemented even without resorting to a spreadsheet. To check In the probabilistic back-analysis methods, the distributions of all
the accuracy of the simplified method, the updated probability uncertain parameters are modified by the slope failure event si-
density functions for c, , and ru obtained from this simplified multaneously. A logic question that one may raise is: would the
method are also plotted in Fig. 4. The results obtained from the degrees of modification achieved by the back-analysis be the
simplified method are almost identical to those from the optimi- same for all the parameters? This question is answered by Fig. 4,
zation procedure.
which shows that the changes in the probability distributions for
The probabilistic back-analysis also provides dependence in-
the three parameters are in fact not the same. Among the three
formation among the model parameters. Let c,, c,ru, and ,ru
parameters, while the probability distribution of ru is changed
denote the correlation coefficients between c and , c and ru, and
and ru, respectively. In Prior 1, c, , and ru are assumed statis- significantly, the prior and modified distributions for are almost
tically independent, so the values of c,, c,ru, and ,ru are all identical. The degree of improvement for a parameter seems to be
zero in the prior distribution. After the back-analysis, the values related to its contribution to the uncertainty in factor of safety. To
of c,, c,ru, and ,ru become ⫺0.06, 0.69, and 0.20, respec- see this, a simple uncertainty analysis can be performed to calcu-
tively. The correlations among the variables in the posterior dis- late the variance of the predicted factor of safety based on the
tribution reflect the slope failure event, which imposes a prior knowledge 共Ang and Tang 2007兲
Probability density
Optimization method
Factor of safety
0.2 Simplified method 1.0
0.1 0.5
0.0 0.0
0 5 10 15 20 0 3 6 9 12 15
c (kPa) c (kPa)
1.5
0.8 (b)
(b) Prior
Factor of safety
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
Probability density
1.5
12 (c)
Prior
Factor of safety
(c)
Probability density
3
0.0
0 -0.10 0.00 0.10 0.20 0.30 0.40
-0.2 0.0 0.2 0.4 0.6 0.8 1.0 ru
ru
Fig. 5. 共a兲 Relationship between factor of safety and c 共 = 38°, ru
Fig. 4. Comparison of prior and back-analyzed distributions based = 0.15兲; 共b兲 relationship between factor of safety and 共c
on Prior 1 共Shek Kip Mei slope兲 = 8 kPa, ru = 0.15兲; and 共c兲 relationship between factor of safety and
ru 共c = 8 kPa, = 38°兲 共Shek Kip Mei slope兲
var共Fs兲 ⬇ 冏 冏 g共兲
c
2
=
var共c兲 + 冏 冏
g共兲
2
=
var共兲
Fig. 7. For comparison, the prior distributions of c, , and ru are
冏 冏
also plotted in this figure. Both the optimization method and sim-
g共兲 2
plified method are used to calculate the posterior distributions.
+ var共ru兲 + 2 共9兲
ru = The two methods yield practically the same results, as shown in
Fig. 7. Among the three variables, while c and ru are more up-
in which Fs = factor of safety. Through Eq. 共9兲, one can identify dated, the distribution of is almost not changed. Fig. 8 shows
the contribution of uncertainty from each parameter to the vari- the contribution of uncertainty from each variable to the variance
ance of Fs, and the results are presented in Fig. 6. The degrees of
changes in the probability distributions in Fig. 4 for the three
variables are consistent with their contributions to the uncertainty
in Fs as shown in Fig. 6: while the probability distribution of ru is 80
most updated, the probability distribution of is almost not
Contribution to variance of Fs (%)
changed.
60
To further verify the above conclusion, the slope failure is
back-analyzed with a different prior distribution on ru, i.e., ru is
normally distributed with a mean of 0.20 and a standard deviation 40
of 0.07. The mean value of is also increased to 39°. The prior
mean for in this case is = 兵8 , 39, 0.20其T. The prior covariance
matrix in this case is 20
冤 冥
32 0 0
C = 0 1 2
0 共10兲 0
cc' f'φ ru
ru eε
0 0 0.072
The above prior distribution is called Prior 2. The posterior dis- Fig. 6. Contributions to the variance of factor of safety 共Fs兲 from
tributions of c, , and ru calculated based on Prior 2 are shown in different parameters based on Prior 1 共Shek Kip Mei slope兲
Probability density
Probability density
Optimization method Optimization method
0.2 Simplified method 0.2 Simplified method
μc|d = 8.04 kPa μc|d = 7.71 kPa
σc|d = 2.57 kPa σc|d = 2.80 kPa
0.1 (Optimization method) 0.1 (Optimization method)
0.0 0.0
0 5 10 15 20 0 5 10 15 20
c (kPa) c (kPa)
0.8 0.8
(b) Prior (b) Prior
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
Probability density
Probability density
0.6 μφ|d = 39.00 o Optimization method 0.6 μφ|d = 38.02 o Optimization method
σφ|d = 0.99 o Simplified method σφ|d = 1.00 o Simplified method
0.4 (Optimization method) 0.4 (Optimization method)
0.2 0.2
0.0 0.0
35 37 39 41 43 34 36 38 40 42
ο ο
φ () φ ()
12 12
(c) Prior (c) Prior
Probability density
Probability density
0 0
-0.2 0.0 0.2 0.4 0.6 0.8 1.0 -0.2 0.0 0.2 0.4 0.6 0.8 1.0
ru ru
Fig. 7. Comparison of prior and back-analyzed distributions based Fig. 9. Comparison of prior and back-analyzed distributions based
on Prior 2 共Shek Kip Mei slope兲 共i兩d and i兩d denote back- on Prior 3 共Shek Kip Mei slope兲
analyzed mean and standard deviation of i, respectively兲
冤 冥
32 − 1.5 0
Contribution to variance of Fs (%)
60
C = − 1.5 12 0 共11兲
0 0 0.12
40
The slope failure is then back-analyzed again using Prior 3, and a
comparison between prior and posterior distributions of c, , and
ru are shown in Fig. 9. Both the optimization method and simpli-
20 fied method are used to calculate the posterior distribution, and
the two methods again yield practically identical results, as shown
in Fig. 9. Through the back-analysis, the values of c,, c,ru, and
0 ,ru are changed from ⫺0.50, 0, and 0 to ⫺0.51, 0.62, and
cc' f'φ ru
ru eε ⫺0.12, respectively. The negative correlation coefficient between
and ru is caused by the combined effects of the positive corre-
Fig. 8. Contributions to the variance of factor of safety 共Fs兲 from lation between c and ru and the negative correlation between c
different parameters based on Prior 2 共Shek Kip Mei slope兲 and . For comparison, the posterior values of c,, c,ru, and ,ru
冤 冥
Slip surface
10
Clay 2 10.222 0 0 0
0 5.472 0 0
C = 2 共12兲
Clay 3 0 0 6.77 0
0 0 0 6.002
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
冤 冥
sure has important effects on the updating of shear strength pa-
rameters. When the knowledge on pore water pressure is poor, the 1 − 0.10 − 0.39 − 0.00
uncertainty caused by pore pressure may dominate the uncertainty − 0.10 1 − 0.37 − 0.00
兩d = 共13兲
in Fs. Therefore, the information collected from the back-analysis − 0.39 − 0.37 1 − 0.01
is mainly allocated to reduce the uncertainty in pore water pres- − 0.00 − 0.00 − 0.01 1
sure rather than to reduce the uncertainty in the shear strength
parameters. However, if pore water pressure uncertainty is small It shows that the strength parameters are negatively correlated as
or pore water pressure is not involved in the back-analysis, the expected to satisfy the constraint that the calculated factor of
probability distributions of shear strength parameters can be im- safety based on the combination of posterior parameters should be
proved more significantly. To illustrate this, the failure of Con- close to unity as a reflection of the slope failure event. Among the
gress Street Cut 共Ireland 1954兲 is analyzed as follows. four uncertain model input parameters, the correlation between
Congress Street Cut 共Fig. 10兲 was located mainly in saturated su1 and su3 as well as the correlation between su2 and su3 are
clays and failed in an undrained manner in 1952 during the con- relatively stronger. The marginal distributions of su1, su2, and su3
struction of Congress Street in Chicago 共Ireland 1954兲. The slip are more updated so these variables have more freedom to adjust
surface passed through four different layers of soil, i.e., a sandy their correlation coefficients.
fill overlaying three saturated clay layers. The cohesion of the
sandy fill was zero 共Oka and Wu 1990兲. In the back-analysis, the
factor of safety of this slope is calculated using Bishop’s simpli- Summary and Conclusions
fied method 共1955兲. Let su1, su2, and su3 denote the average und-
rained shear strengths of the three clay layers, and denote the The research reported in this paper and findings from the paper
average friction angle of the sandy fill. The parameters to be are summarized as follows:
back-analyzed in this case are = 兵su1 , su2 , su3 , 其T. 1. Back-analysis of slope failure is often carried out to improve
For this slope, sampling was carried out using 51 mm diameter one’s knowledge on the parameters of a slope stability
Shelby tubes with samples taken at least 1 m apart 共Tang et al. model. To back-analyze multiple sets of slope stability pa-
1976兲. Based on reported triaxial tests, the mean values of su1, su2, rameters simultaneously under uncertainty, the back-analysis
and su3 are 43.14, 32.65, and 37.35 kPa, respectively, and the can be implemented in a probabilistic way, in which uncer-
standard deviations for these variables are 10.22, 5.47, and 6.77 tain parameters are modeled as random variables, and their
kPa, respectively 共Yucemen et al. 1973; Tang et al. 1976兲. Vari- distributions are improved using observed slope failure infor-
ance reduction due to spatial variability was considered in deriv- mation.
0 20
0 20 40 60 80
s u1 (kPa)
0
c1
s u1 sc2u2 sc3
u3 φf εm
0.5
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
(b) Fig. 12. Contributions to the variance of factor of safety 共Fs兲 from
Prior
different parameters 共Congress Street Cut兲
Probability density
0.4
Back analysis
0.3 μsu2|d = 31.26 kPa
0.2 σsu2|d = 5.24 kPa failure caused by rainfall infiltration. The stability model of
this case is largely linear, and the back-analysis results from
0.1 the two methods are indeed similar.
0
4. Case studies in this paper show that the degree of change in
10 30 50 70 the probability distribution of a parameter achieved by the
s u2 (kPa) back-analysis depends on its contribution to the uncertainty
in the factor of safety: the probability distribution of the pa-
rameter that contributes the greatest uncertainty to the factor
0.5 of safety will be changed most. This conclusion is valid if the
(c)
Prior parameters to be updated are not correlated in the prior dis-
Probability density
0.4
Back analysis
tribution. For a slope failure in which the knowledge in pore
0.3
μsu3|d = 32.67 kPa water pressure is poor, the uncertainty in pore water pressure
0.2 σsu3|d = 4.27 kPa may dominate the back-analysis. In such a case, the knowl-
edge on shear strength parameters may not be effectively
0.1 improved by probabilistic back-analysis.
0
10 30 50 70
s u3 (kPa) Acknowledgements
0.4
Back analysis
0.3 μφ|d = 29.95 o
σφ|d = 6.00 o Appendix I. Theory of Probabilistic Back-Analysis
0.2
for a General Slope Stability Model
0.1
0 Basic Theory
0 20 40 60
The general probabilistic back-analysis theory is first introduced
φ (ο)
here. Let g共 , r兲 be a general model to predict the response of a
Fig. 11. Comparison of prior and back-analyzed distributions system, where and r are vectors denoting unknown and known
共Congress Street Cut兲 parameters, respectively. In the following text, r is dropped from
the prediction model for simplicity. Let d be a vector representing
the actual system response. The objective of the probabilistic
2. Two methods are presented for probabilistic back-analysis of back-analysis is to improve the probability distribution of based
slope failure, i.e., a method based on optimization for a gen- on some observed system response dobs through the prediction
eral slope stability model, and a simplified method based on model g共兲.
sensitivity analysis. While the simplified method is easier to Before the probabilistic back-analysis, one may already have
apply in practice, it is only applicable when the slope stabil- some prior knowledge on from various sources. Assume the
ity model is approximately linear. prior knowledge on can be denoted by a multivariate normal
3. For efficient analysis, a spreadsheet template has been devel- distribution
oped to implement the back-analysis method based on opti-
mization. Also, the simplified method involves simply a
sensitivity analysis together with a conventional slope stabil-
冋
1
f共兲 = k1 exp − 共 − 兲TC−1共 − 兲
2
册 共14兲
ity program. The results from the optimization method and where k1 = normalization constant to make the probability density
the simplified method are compared in a case study of slope function valid 关i.e., making the integration of probability density
冋 册
tribution with a mean of m and a covariance matrix of Cm. The
1 magnitude of m measures the degree of bias in the model pre-
f共d兩dobs兲 = k2 exp − 共d − dobs兲TCd−1共d − dobs兲 共15兲
2 diction. For example, m = 0 means the model is unbiased, and in
this case the model uncertainty is described by Eq. 共16兲. In the
where k2 is another normalization constant to make the probabil-
case where m ⫽ 0, although the prediction from g共兲 is biased,
Downloaded from ascelibrary.org by DALHOUSIE UNIVERSITY on 06/15/13. Copyright ASCE. For personal use only; all rights reserved.
再 1
f共d兩兲 = k3 exp − 关d − g共兲兴TCm−1关d − g共兲兴
2
冎 共16兲 Application to Back-Analysis of Slope Failure
In back-analysis of slope failure, the observation data is that the
where k3 is again a normalization constant to make the probability factor of safety at the moment of slope failure is unity, which is a
density function valid. scalar. In such a case, the model uncertainty, the observation data,
With the above assumptions, the improved knowledge of and the observational uncertainty are all scalars. Let d2 , , and
considering both f共兲 and dobs can be described by its posterior 2 denote Cd, m, and Cm in the back-analysis of slope failure,
probability distribution 共Tarantola 2005兲 respectively, where d and are the standard deviations of ob-
servation and model uncertainty variables, respectively. Note that
f共兩dobs兲 = k4 exp兵关g共兲 − dobs兴TCD−1关g共兲 − dobs兴 in back-analysis of slope failure dobs = 1. Assume the observa-
+ 共 − 兲TC−1共 − 兲其 共17兲 tional uncertainty is negligible, d = 0. In such a case, Eqs. 共22兲,
共19兲, and 共20兲 become Eqs. 共2兲–共4兲, respectively.
where k4 = normalization constant to make the probability density
function valid; f共 兩 dobs兲 = posterior distribution of given dobs;
= prior means vector of ; C = prior covariance matrix of
Appendix II. Probabilistic Back-Analysis Theory for
and CD = Cd + Cm.
an Approximately Linear Slope Stability Model
The posterior density function of may not have a closed-
form expression except when g共兲 is linear. Let ⴱ be the point
For a general linear prediction model g共兲 = H, the posterior
where f共 兩 dobs兲 is maximized. If the prediction model g共兲 is
mean and covariance of can be determined using the following
nonlinear but is linearizable in the neighborhood of ⴱ, f共 兩 dobs兲
relationships without resorting to the optimization procedure
can be approximated by a multivariate normal distribution with a
共Tarantola 2005兲
mean of 兩d and a covariance matrix of C兩d, where 兩d and C兩d
are 共Tarantola 2005兲
兩d = + CHT共HCHT + CD兲−1共dobs − H兲 共23兲
ⴱ
兩d = 共18兲
ⴱ can be found by maximizing Eq. 共17兲, or equivalently, mini- 兩d = + CHT共HCHT + CD兲−1共dobs − H − m兲 共25兲
mizing the misfit function 2S共兲, which is defined as follows
共Tarantola 2005兲:
If the slope stability model is largely linear, Eqs. 共25兲 and 共24兲
2S共兲 = 关g共兲 − dobs兴TCD−1关g共兲 − dobs兴 + 共 − 兲TC−1共 − 兲 may be used to determine the improved mean and covariance
matrix of approximately, respectively. As mentioned before, in
共21兲
the back-analysis of slope failure, the observed information is a
The general probabilistic back-analysis method described above scalar. The scalar form of Eqs. 共25兲 and 共24兲 are Eqs. 共5兲 and 共6兲,
is also known as theory of system identification 共e.g., Schweppe respectively. The accuracy of this simplified method has been
1973兲. checked in the case study.
and geotechnical back-analysis.” Proc., Advances in Geotechnical En- Statistics and Probability in Soil and Struct. Engr., University of Brit-
gineering: The Skempton Conf., R. J. Jardine, D. M. Potts, and K. G. ish Columbia, Vancouver, Canada.
Higgins, eds., Thomas Telford, London. Mostyn, G. R., and Li, K. S. 共1993兲. “Probabilistic slope stability
Christian, J., Ladd, C., and Baecher, G. 共1994兲. “Reliability applied to analysis—State-of-play.” Probabilistic methods in geotechnical engi-
slope stability analysis.” J. Geotech. Engrg., 120共12兲, 2180–2207. neering, K. S. Li and S.-C. R. Lo, eds., Balkema, Rotterdam, The
Der Kiureghian, A., and Liu, P. L. 共1986兲. “Structural reliability under Netherlands.
incomplete probability information.” J. Engrg. Mech., 112共1兲, 85–
Oka, Y., and Wu, T. H. 共1990兲. “System reliability of slope stability.” J.
104.
Geotech. Engrg., 116共8兲, 1185–1189.
El-Ramly, H., Morgenstern, N. R., and Cruden, D. M. 共2002兲. “Probabi-
Schweppe, F. C. 共1973兲. Uncertain dynamic systems, Prentice-Hall,
listic slope stability analysis for practice.” Can. Geotech. J., 39共3兲,
665–683. Englewood Cliffs, N.J.
El-Ramly, H., Morgenstern, N. R., and Cruden, D. M. 共2005兲. “Probabi- Spencer, E. 共1967兲. “A method of analysis of embankments assuming
listic assessment of stability of a cut slope in residual soil.” Geotech- parallel inter-slice forces.” Geotechnique, 17共1兲, 11–26.
nique, 55共1兲, 77–84. Tang, W. H., Yucemen, M. S., and Ang, A. H.-S. 共1976兲. “Probability
FMSW. 共2000兲. “Report on the Shek Kip Mei landslide of 25 August based short-term design of soil slopes.” Can. Geotech. J., 13共3兲, 201–
1999.” Rep. Prepared by Fugro Maunsell Scott Wilson Joint Venture 215.
for the Geotechnical Engineering Office, Government of Hong Kong, Tarantola, A. 共2005兲. Inverse problem theory and methods for model pa-
Hong Kong. rameter estimation, 2nd Ed., Elsevier Science, New York.
Geotechnical Engineering Office 共GEO兲. 共2000兲. Guide to rock and soil Tiwari, B., Brandon, T. L., Marui, H., and Tuladhar, G. R. 共2005兲. “Com-
description, Government of Hong Kong, Hong Kong. parison of residual shear strengths from back-analysis and ring shear
Gilbert, R. B., Wright, S. G., and Liedtke, E. 共1998兲. “Uncertainty in tests on undisturbed and remolded specimens.” J. Geotech. Geoenvi-
back-analysis of slopes: Kettleman Hills case history.” J. Geotech. ron. Eng., 131共9兲, 1071–1079.
Geoenviron. Eng., 124共12兲, 1167–1176. Wesley, L. D., and Leelaratnam, V. 共2001兲. “Shear strength parameters
Ireland, H. O. 共1954兲. “Stability analysis of Congress Street open cut in from back-analysis of single slips.” Geotechnique, 51共4兲, 373–374.
Chicago.” Geotechnique, 4共4兲, 163–168. Yucemen, M. S., Tang, W. H., and Ang, A. H.-S. 共1973兲. “A probabilistic
Janbu, N. 共1987兲. “Slope stability computation.” Embankment-dam engi- study of safety and design of earth slopes.” Structural Research Series
neering: Casagrande volume, R. C. Hirschfeld and S. J. Poulos, eds., No. 402, Dept. of Civil Engineering, University of Illinois at Urbana,
Kriege, New York, 47–86. Ill.