Slope Stability Analysis Based On Experimental Design

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Slope Stability Analysis Based on Experimental Design

Srdan Kostic, Ph.D.1; Nebojša Vasovic, Ph.D.2; Duško Sunaric, Ph.D.3

Abstract: In this paper, the authors propose an analytical model for the prediction of the slope safety factor as a function of basic geometrical
parameters (slope height H and slope angle b ) and soil factors (bulk density g , cohesion c, angle of internal friction w , and pore water pressure
coefficient ru). Research was performed by applying the statistical technique of experimental design, for which the input data were provided by
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

stability analyses of different homogeneous slopes with a circular slip surface using the Spencer limit equilibrium method. The proposed model
represents a nonlinear equation of a simpler form and higher prediction accuracy than those of the existing mathematical expressions, with pre-
dominant linear effect of the individual factors and significant influence of the two-factor interactions. Linear terms in a derived equation indi-
cate a positive effect of c or w and a negative effect of H, b , g , or ru on slope stability. Because of two-factor interactions, the effect of c is
highly dependent on H, b , g , and w , whereas the influence of w is predetermined by the values of b and ru. DOI: 10.1061/(ASCE)GM.1943-
5622.0000551. © 2016 American Society of Civil Engineers.

Introduction approach dates back to Taylor (1937), who proposed a simple dia-
gram for the estimation of safety factors using a dimensionless pa-
Estimation of slope stability represents one of the main tasks in rameter Ns, slope angle b , and angle of internal friction w . Many
geotechnical design, primarily because it enables (1) evaluation of succeeding analyses after this initial proposition suggested numer-
geometrical constraints for stable slopes in road construction, (2) ous stability charts, which provided a very reliable first estimation of
calculation of forces needed for the proper design of landslide resto- slope stability (Eid 2014).
ration, and (3) determination of soil strength parameters using a Apart from the aforementioned analytical and numerical meth-
back-analysis approach. In general, three distinct groups of methods ods, which are based on traditional and well-known principles,
are typically used in slope stability analysis: analytical methods, nu- strong development of artificial intelligence techniques has lead to
merical techniques, and stability charts. Traditional limit equilibrium wide implementation of artificial neural networks (Das et al. 2011;
techniques are commonly used among analytical methods, resulting Erzin and Cetin 2013; Liu et al. 2014), support vector machines
in accurate safety factors for engineering purposes (Duncan and (Zhao 2008; Tan et al. 2011; Li et al. 2013), and genetic algorithms
Wright 2005). Reliability of such methods is confirmed through their (Zolfaghari et al. 2005; Sun et al. 2008; Li et al. 2010; Yang et al.
implementation in modern software packages, including Plaxis 2004) in slope stability analyses. All these methods represent reli-
(Brinkgreve 2002), Slide 5.02, and GEO-SLOPE (GEO-SLOPE able and easy-to-use tools for estimation of safety factors.
International 2008; Aryal 2006). Nevertheless, although analytical Moreover, some of these techniques, such as genetic algorithms,
approaches are easy to use and provide an intuitive approach toward yield explicit mathematical expressions for the prediction of a slope
a final solution, numerical methods have been extensively applied safety factor (Yang et al. 2004; Ahangar-Asr et al. 2010;
recently because of the development of fast computational solutions. Manouchehrian et al. 2014; Alavi and Gandomi 2011). These equa-
Numerical methods mainly cover the application of FEM (Zhang tions are typically based on nonlinear dependence of safety factors
1999; Zheng et al. 2005; Huo and Zhai 2012) and techniques based on various geometrical slope properties or physicomechanical pa-
on the upper-bound theorem of plasticity limit analysis (Donald and rameters of soil, and they are very convenient for the fast calculation
Chen 1997; Kim et al. 2002). However, even though these methods of safety factors and the quick estimation of the effect of input pa-
provide an estimation of slope stability, which is reasonable and of- rameters. This type of prediction model is also derived in this paper.
ten more reliable than limit equilibrium analyses, their implementa- The final goal of this study was to develop a prediction model for
tion requires high computational time and powerful software appli- slope safety factors, including the estimation of the effect of different
cations, which are not always possible to achieve, especially when input factors on slope stability. One may wonder about the merit of
there is limited time for a geotechnical report to be prepared or in developing such a model as a function of six basic parameters, despite
field conditions. In such cases, slope stability charts are typically the fact that a similar problem already has been studied (Yang et al.
used as a suitable tool for quick estimation of slope stability. This 2004; Ahangar-Asr et al. 2010; Manouchehrian et al. 2014; Alavi and
Gandomi 2011). There might be two reasons for this. First, to formu-
1
Institute for Development of Water Resources Jaroslav Cerni,  late this model we used the methodology of experimental design
11226 Pinosava, Belgrade, Serbia (corresponding author). E-mail: (Fisher 1935; Deming and Morgan 1996), which was not previously
srdjan.kostic@jcerni.co.rs applied in the area of geotechnical analysis, even though it has already
2
Dept. of Applied Mathematics, Faculty of Mining and Geology, Univ. provided reliable results in pharmaceutical technology (Doornbos
of Belgrade, 11000 Belgrade, Serbia. et al. 1990) and analytical and food chemistry (Cela et al. 2013; Hibbert
3 
Institute for Development of Water Resources Jaroslav Cerni, 11226
2012). Second, the authors wished to estimate the impact of each term
Pinosava, Belgrade, Serbia.
Note. This manuscript was submitted on November 25, 2014;
in a proposed model on slope stability (i.e., to perform sensitivity analy-
approved on June 4, 2015; published online on January 28, 2016. sis), which was not provided by the previous suggested expressions.
Discussion period open until June 28, 2016; separate discussions must be This paper is organized as follows. Applied methods are briefly
submitted for individual papers. This paper is part of the International described in the next section, and the main results of the analyses
Journal of Geomechanics, © ASCE, ISSN 1532-3641. performed are given in Experimental Results. Accuracy of the

© ASCE 04016009-1 Int. J. Geomech.

Int. J. Geomech., 04016009


developed model is verified by stability analyses of random slopes in determination of up to three-factor interaction which is shown to
Verification of the Proposed Model, and comparison of its forecast- have negligible impact on Fs in the present case.
ing precision with the existing equations is given in Comparison with 5. Estimation of the effect of input parameters. Once the statistical
Existing Prediction Models. A brief review of the results is given in value of the created mathematical model is confirmed, the contri-
the final section, together with suggestions for further research. bution of each input parameter to Fs can be defined and mutually
compared by analyzing their standardized effects in a Pareto chart
or by comparing values of the corresponding coefficients.
Applied Methods Moreover, the influence of each of the individual factors, as well as
the impact of two-factor interactions, is further evaluated by ana-
In this paper, a design of experiments (DoE) methodology was lyzing the corresponding one-dimensional and three-dimensional
applied to develop a prediction model for slope safety factors response surfaces. In that way, parameters with the greatest influ-
(Deming and Morgan 1996). Based on the range of input factors, this ence on Fs are identified, giving valuable information about the
method suggests statistically adequate combinations of input factors, factors that are actually controlling the slope stability.
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

which are uniformly distributed across the chosen parameter space. 6. Validation of the proposed model. After development of the
As a next step, the safety factor (Fs) is calculated for each parameter mathematical model, extensive statistical analysis using new
combination using the Spencer limit equilibrium method, as the sim- input data (which were not used for model building) has to be per-
plest limit equilibrium method that satisfies all the equilibrium condi- formed to finally confirm the practical application of the devel-
tions. After performing these analyses and recording the results (Fs), oped model. Statistical analysis of a developed model includes
a mathematical relation between the examined parameters and safety using the ANOVA test, assessing R2, adjusting R2, and predict-
factor can be created using linear regression and least-squares ing R2 values, as well as the p value and standard error (SE).
approximation, and the contributions of each parameter to Fs can be
defined and mutually compared. Once the model was created, it was
Experimental Results
used for navigation of experimental space defined by the range of
input parameters, without having to perform additional stability anal-
The initial phase of the research consisted of defining the input pa-
yses. This way, the search for a desired or optimal response value
rameters that were expected to significantly influence the slope safety
(i.e., a minimum safety factor) was done through mathematical simu-
factor. In this case, six input parameters were chosen: slope height H,
lation or by reading the appropriate graphical response surfaces.
slope angle b , bulk density g , cohesion c, angle of internal friction
On the basis of the described DoE methodology, the performed
w , and pore water pressure coefficient ru, which are routinely exam-
research consisted of the following important stages:
ined in slope stability analyses, because they predetermine the condi-
1. Selection of the potentially important factors, on the basis of
tions for the soil failure to occur (Erzin and Cetin 2013; Bishop and
previous analyses and research experience. One should note
Morgenstern 1960). However, because of limitations of the applied
that, in this stage, a reasonable number of factors need to be
method, the authors could not examine the impact of bedrock depth
selected, because a large number of input parameters would
on slope stability, because experimental design is used only for the
require a greater number of stability analyses.
input factors that are not mutually dependent. In this case, the depth
2. Definition of factor intervals, which were also set according to
research experience, which reflects what is typically expected
in situ.
3. Selection of experimental design and performing the stability analy-
ses. In this stage, adequate experimental design was selected, which
provided a well-planned and statistically reliable scheme of combi-
nations among input parameters that were set in each analysis. In
general, response surface experimental designs are recommended
for building reliable prediction models. Such a design class exam-
ines each input parameter on three levels to establish a quadratic de-
pendence among the investigated slope and soil properties and Fs.
4. Building of a mathematical model by applying the multiple linear
regression and least-squares method of approximation. Response
surface design enabled development of the following model: Fig. 1. Sketch of the one case of analyzed slopes and review of the
examined input parameters
X
N N X
X N
Y ¼ b0 þ b i Xi þ bij Xij (1)
i¼1 i¼1 j¼1
Table 1. Examined Range of Values for the Chosen Input Parameters
where Y is the response; and b 0, b i, b ij, and b ii correspond to Lower interval Central interval Higher interval
coefficients of the intercept, linear, interaction, and quadratic Input level level level
term (i = j), respectively; and X denotes the effect of influential parameters (coded value) (coded value) (coded value)
(input) factors. This type of model enables the independent esti-
H (m) 6 (−1) 8 (0) 10 (þ1)
mation of linear and quadratic dependence of the response on
b (degrees) 25 (−1) 47.5 (0) 70 (þ1)
each factor, including the assessment of two-factor interactions.
g (kN/m3) 16 (−1) 18 (0) 20 (þ1)
The latter interactions can be valuable in real problems, because
c (kPa) 0 (−1) 25 (0) 50 (þ1)
the impact of simultaneous variation of two factors can contrib-
w (degrees) 10 (−1) 30 (0) 50 (þ1)
ute differently to the response value than the influence of individ- ru 0 (−1) 0.25 (0) 0.5 (þ1)
ual factors. One should note that response surface design enables

© ASCE 04016009-2 Int. J. Geomech.

Int. J. Geomech., 04016009


of the bedrock (d) was not independent of the slope height (H), values), the higher level of the interval (denoted as þ1 in coded fac-
because their relation would have to satisfy the following condition: tor values), and the central point of the interval (denoted as 0 in
d ≥ H. Hence, in this study, the authors examined only the homogene- coded factor values). Coding of real factor values was necessary to
ous slopes with the bedrock at great depth with no impact on the slope enable comparison of the impact of each input parameter on Fs in-
stability (Fig. 1). dependently from their units of measurement.
In the next preparatory phase of the research, a range of values After the choice of the most influential input parameters and the
for each input parameter was chosen (Table 1) according to previ- corresponding ranges of their values, an appropriate experimental
ous studies (Erzin and Cetin 2013; Bishop and Morgenstern 1960). design was selected. To obtain as much valuable information about
It can be seen that each of the six factors was examined on three lev- the relations amonginput parameters andFs as possible, response sur-
els: the lower level of the interval (denoted as −1 in coded factor face (Box–Behnken) design was selected. This design belongs to the
response surface class of experimental designs in which each factor is
varied on three levelsand the total number of required experiments for
Table 2. Input Data for Developing Prediction Model of Fs Using
Experimental Design k investigated factors is 2k(k – 1) points plus central points (Doornbos
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

et al. 1990). In this analysis, the experimental plan consisted of 48 sta-


Case No. H (m) b (degrees) g (kN/m3) c (kPa) w (degrees) ru Fs bility analyses, which were conducted by using the limit equilibrium
1 6 25 18 0 30 0.25 0.893
Spencer method in Slide as the simplest method that satisfies all equa-
2 10 25 18 0 30 0.25 0.88
tions of equilibrium (Table 2). Location of critical slip surface was
3 6 70 18 0 30 0.25 0.128
determined by using the Auto Refine Search method, with the limit of
4 10 70 18 0 30 0.25 0.117
5 6 25 18 50 30 0.25 4.683 Table 3. Developed Mathematical Model and Its Statistical Significance
6 10 25 18 50 30 0.25 3.39 Coefficient Coefficient
7 6 70 18 50 30 0.25 3.147 Model for coded for real
8 10 70 18 50 30 0.25 1.852 term factor values factor values p values SE
9 8 25 16 25 10 0.25 1.793
Intercept 1.7462 0.2049 0.0411
10 8 70 16 25 10 0.25 1.132
H −0.3118 −0.0040 <0.0001 0.0336
11 8 25 20 25 10 0.25 1.532
b −0.7120 −0.0375 <0.0001 0.0336
12 8 70 20 25 10 0.25 0.936
g −0.1116 0.0046 0.0020 0.0336
13 8 25 16 25 50 0.25 4.059
c 1.3234 0.1617 <0.0001 0.0336
14 8 70 16 25 50 0.25 1.452
w 0.5624 0.0871 <0.0001 0.0336
15 8 25 20 25 50 0.25 3.713
ru −0.3782 0.4981 <0.0001 0.0336
16 8 70 20 25 50 0.25 1.225
Hc −0.3038 −0.0061 <0.0001 0.0411
17 8 47.5 16 0 30 0 0.573
b c −0.1933 −0.0003 0.0020 0.0581
18 8 47.5 20 0 30 0 0.573
b w −0.4719 −0.0010 <0.0001 0.0411
19 8 47.5 16 50 30 0 3.43
g c −0.1208 −0.0024 0.0450 0.0582
20 8 47.5 20 50 30 0 2.937
cw 0.1519 0.0003 0.0130 0.0582
21 8 47.5 16 0 30 0.5 0.028
w  ru −0.3351 −0.0670 <0.0001 0.0582
22 8 47.5 20 0 30 0.5 0.028
b2 0.2446 0.0005 <0.0001 0.0471
23 8 47.5 16 50 30 0.5 2.656
c2 −0.1184 −0.0002 0.0165 0.0471
24 8 47.5 20 50 30 0.5 2.183
R2 0.9869
25 6 47.5 18 0 10 0.25 0.098
Adjusted R2 0.9819
26 10 47.5 18 0 10 0.25 0.096
Predicted R2 0.9706
27 6 47.5 18 50 10 0.25 3.044
28 10 47.5 18 50 10 0.25 1.894
29 6 47.5 18 0 50 0.25 0.642
30 10 47.5 18 0 50 0.25 0.575
31 6 47.5 18 50 50 0.25 4.196
32 10 47.5 18 50 50 0.25 2.98
33 8 25 18 25 10 0 1.827
34 8 70 18 25 10 0 1.125
35 8 25 18 25 50 0 4.79
36 8 70 18 25 50 0 2.093
37 8 25 18 25 10 0.5 1.458
38 8 70 18 25 10 0.5 0.949
39 8 25 18 25 50 0.5 2.942
40 8 70 18 25 50 0.5 0.715
41 6 47.5 16 25 30 0 2.6
42 10 47.5 16 25 30 0 1.918
43 6 47.5 20 25 30 0 2.258
44 10 47.5 20 25 30 0 1.704
45 6 47.5 16 25 30 0.5 1.91
46 10 47.5 16 25 30 0.5 1.05 Fig. 2. Values of Fs determined by the Spencer limit equilibrium
47 6 47.5 20 25 30 0.5 1.587 method versus the predicted value of Fs, indicating high statistical reli-
48 10 47.5 20 25 30 0.5 1.246 ability of the proposed model (internal validation)

© ASCE 04016009-3 Int. J. Geomech.

Int. J. Geomech., 04016009


a 1-m minimum depth to avoid the occurrence of kinematically determined, including one intercept, six linear term coefficients, 15
impossible slipsurfaces (e.g., attheslopesurface). two-factor interaction coefficients, and six quadratic term coeffi-
cients. Theoretically, only 28 stability analyses are necessary to
determine the required 28 coefficients. However, the applied
Prediction Model DoE methodology suggested 48 analyses, which left sufficient
degrees of freedom for obtaining the model of higher statistical
Once the slope safety factors for different geometrical and soil prop- reliability. In other words, to create a model that is able to simu-
erties of the slope were calculated (Table 2), the mathematical late real slope conditions, the position of factor levels in experi-
Model 1 was built. In the present study, 28 coefficients had to be mental space should be carefully distributed, (i.e., it should fill
the experimental space uniformly and provide as much informa-
tion as possible). This way, the created mathematical model is
not only the numerical presentation of the performed experi-
ments, it is also the description of the entire system.
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

The adequate mathematical model for the prediction of Fs was


created using multiple linear regression and least-squares methods
of approximation. The preliminary statistical analysis enabled the
identification and omission of evidently insignificant model terms
(the p value was significantly higher than 0.05), giving this model
(Model 2) in the following form (Table 3):

Fs ¼ 0:2049  0:004  H  0:0375  b þ 0:0046  g þ 0:1617

 c þ 0:0871  w þ 0:4981  ru  0:0061  H  c

 0:0003  b  c  0:001  b  w  0:0024  g  c

þ 0:0003  c  w  0:067  w  ru þ 0:0005


Fig. 3. Pareto chart of standardized factor effects (SE = 0.165)  b 2  0:0002  c2 (2)

Fig. 4. Effect of individual input parameters on slope safety factor Fs: (a) Fs = f(c); (b) Fs = f( w ); (c) Fs = f(H); (d) Fs = f( b ); (e) Fs = f(ru); (f) Fs =
f( g ); while a single parameter was varied, other factors were held constant at their average values (H = 8 m, b = 47.5°, g = 18 kN/m3, c = 25
kPa, w = 30°, and ru = 0.25)

© ASCE 04016009-4 Int. J. Geomech.

Int. J. Geomech., 04016009


Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 5. Three-dimensional response surface of Fs as a function of two-factor interactions: (a) Fs = f(c,H); (b) Fs = f(c, b ); (c) Fs = f(c, g ); (d) Fs =
f(c, w ); (e) Fs = f(ru, w ); (f) Fs = f( b , w ); while two parameters were being changed, other input factors were held constant at their average values (H =
8 m, b = 47.5°, g = 18 kN/m3, c = 25 kPa, w = 30°, and ru = 0.25); dark contour lines denote the boundary of slope stability

© ASCE 04016009-5 Int. J. Geomech.

Int. J. Geomech., 04016009


The obtained values of the coefficient of determination lower to a higher level is not the same for H = 6 m and H = 10 m
(R2 > 0.98) and adjusted coefficient of determination (adjusted (i.e., the value of Fs increases more slowly with the rise of cohesion
R2 > 0.98) listed in Table 3 indicate that the model describes the ana- for the higher slopes). Similarly, although H is moved from mini-
lyzed system with high statistical reliability. Moreover, the high pre- mum to maximum values, the decrease of Fs is not the same for c =
dicted value of R2 (0.9706), which is a marker of internal validation, 0 kPa and c = 50 kPa. For example, if a constant value of c = 20 kPa
suggests that the predictive power of the proposed model is is assumed, it is clear that the increase of slope height from 6 to
extremely high, indicating its great practical reliability. Values of Fs 10 m will induce an occurrence of slope instability. Such mutual
determined by the Spencer limit equilibrium method were compared effects of c and H on Fs are acceptable under real conditions,
with the predicted value of Fs, indicating a high value of correlation because these parameters affect opposite forces acting on a slope:
coefficient (Fig. 2). the increase of c contributes to resisting forces, whereas the higher
values of H affect more the disturbing forces.
Similar conclusions can be made for the interaction of c and b
Impact of Input Data [Fig. 5(b)], in which the most negative effect on Fs is estimated for
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

the high slope angles and low cohesion, which is expected, because
After a reliable mathematical model had been created and its statis-
lower cohesion reduces the resisting forces and higher slope angle
tical significance had been confirmed, the effect of each input pa-
increases the disturbing forces.
rameter on slope safety factor was estimated. The impact of each
As for the coaction of c and g [Fig. 5(c)], it is clear that g has
term in a proposed Model 2 could be evaluated by analyzing the
almost no impact on the effect of c, which has already been indi-
coefficients for coded factor values (Table 3) or by calculating the
cated by the analysis of a single-factor effect and by the lowest
standardized effect values, which were obtained by dividing each
value of its coefficient effect in the Pareto chart (−2.075).
coefficient by its SE (Fig. 3). The absolute values of coefficient
Conversely, the simultaneous decrease of c and w , as shown in
effects in the Pareto chart denote the impact magnitudes of each
Fig. 5(d), induces the slope instability, which is also expected under
input parameter, and their sign suggests the increasing (þ) or
real conditions, because the decrease of both c and w reduces the
decreasing (−) effect on the safety factor.
resisting forces acting on a slope.
The obtained Pareto chart (Fig. 3) indicates the linear depend-
Regarding the mutual effect of w and ru, as seen in Fig. 5(e), the
ence of the safety factor on all input parameters, whereas the signifi-
increase of w for a constant value of ru increases the slope stability
cance of squared terms for b and c also suggests a quadratic rela-
in different ways. For high values of ru, the increase of Fs with the
tion among Fs and these two factors but with small effects (5.19 and
−2.51, respectively). Regarding the impact of individual factors, it rise of w is rather small, whereas the effect of w is more significant
is clear that soil cohesion c has the strongest influence on Fs for dry slopes (ru = 0).
(39.39), whereas the subsequent important factor (i.e., the slope In contrast to previous cases, an interesting feature appears in
angle b ) shows almost half of the impact of c (−21.19). The effect Fig. 5(f), which presents the coaction of b and w on Fs. It is appa-
of the remaining individual input parameters on Fs decreases in the rent that the effect of w on Fs is significantly different for mild and
following order: w > ru > H > g . Hence, bulk density has the least
influence on the slope stability, which was already stated by Jiang
and Yamagami (2006), who showed that bulk density does not have
significant influence on slope stability, as long as the constant pore
pressure ratio ru is assumed throughout the whole slope.
The influence of individual factors on Fs, while the remaining
input parameters are kept on average level, is shown in Fig. 4. It
can be seen that cohesion c and angle of internal friction w show
positive influence on Fs [Figs. 4(a and b)], induced by the
increase of resisting forces acting on a slope. Conversely, the
remaining factors H, b , g , and ru have a negative effect on Fs.
Indeed, the increase of slope height and angle (H and b ), which
significantly affects the disturbing forces, induces a decrease in
Fs [Figs. 4(c and d)]. High values of pore pressure (ru) reduce
the resisting forces on slope similarly, which results in a lower
safety factor [Fig. 4(e)]. Also, the increase of bulk density
slightly decreases the value of Fs [Fig. 4(f)], which is expected
because the increasing g affects the disturbing forces more than
the resisting ones, especially under the effect of pore water pres-
sure (Zhao et al. 2010).
One has to notice a moderate curvature in Figs. 4(a and d), which
indicates that c and b also have significant quadratic terms (Table
3), confirming their strong impact on slope stability.
Nevertheless, although the analysis of a single parameter effect
on Fs provided reasonable results, the existence of two-factor inter-
actions in the proposed Eq. (2) indicates that the simultaneous varia-
tion of some input parameters will contribute differently to slope Fig. 6. Coaction of slope angle b and angle of internal friction w on
stability than the effect of individual factor variation. With the most slope stability for the worst-case scenario regarding the slope stability:
influential factor, cohesion c, it practically means that its impact on H = 10 m, g = 20 kN/m3, c = 0 kPa, and ru = 0.5; the dark line denotes
Fs will be different for various values of H, b , g , and w (Fig. 5). As the boundary of slope stability
shown in Fig. 5(a), the increase of Fs induced by c changing from a

© ASCE 04016009-6 Int. J. Geomech.

Int. J. Geomech., 04016009


Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 7. Coaction of cohesion c and slope angle b on slope stability for: (a) the most favorite values (H = 6 m, g = 16 kN/m3, w = 50°, and ru = 0); (b)
the least favorite values of other parameters (H = 10 m, g = 20 kN/m3, w = 10°, and ru = 0.5); dark lines denote the boundary of slope stability

steep slopes. In the former case, the rise of w strongly affects Fs by Table 4. Random Properties of Slopes for Verification of the Proposed
increasing its value, which proves that this two-factor interaction has Model 2
the strongest influence on slope stability among all the remaining
Numerical
interactions (Fig. 3). In the latter case, one can observe smaller example No. H (m) b (degrees) g (kN/m3) c (kPa) w (degrees) ru
changes of Fs with the increase of w . Moreover, another important
feature arises in Fig. 5(f). It is apparent that, because of the quadratic 1 6.5 62 16.8 9 15 0.1
dependence of Fs on b , the minimum of the response function in 2 7.2 36 17.2 46 22 0.22
Eq. (2) is captured within the investigated region (Fs = 1.125), mean- 3 8.4 52 17 24 44 0.41
ing that for w = 10 − 15°, a further increase of b from 45 to 70° does 4 9.6 33 19.4 18 35 0.39
not have a significant impact on Fs. Such effects of b and w on Fs are 5 6.8 43 18.3 38 11 0.13
further qualitatively confirmed for the worst-case scenario (H = 10 m, 6 7.3 66 19 44 29 0.26
c = 0 kPa, ru = 0.5, and g = 20 kN/m3), as shown in Fig. 6. Slopes 7 9.1 58 16.2 21 33 0.48
with high values of height, bulk density of soil, and a high level of 8 7 28 19.6 15 12 0.31
underground water, including the lowest value of c, apparently cannot 9 9 61 17.5 4 48 0.2
withstand slope angles higher than b  35° for small values of w . In 10 6.9 44 18.9 12 19 0.11
other words, further increases of b in this case do not have any signifi-
cant impact on Fs, which confirms our previously stated observation.
Even though the two-factor interactions (Fig. 5) give rather clear Verification of the Proposed Model
and intuitive insight in their qualitative effects on slope stability,
one should be aware of the fact that mutual impact of these factors To further confirm the reliability of the proposed Model 2, its
on Fs is shown for average constant values of other parameters. predictive power was tested by analyzing 10 examples of slopes
However, because the effect of c on Fs, for example, will be differ- with random geometrical characteristics and soil properties, dif-
ent for various values of H, b , g , and w , the coaction of all of these ferent from those used for building the mathematical model
factors should be taken into account. Therefore, the mutual effects (Table 4). Exact values of the safety factors for slopes given in
of c and b are shown in Fig. 7 for the most favorite values of other Table 4 were determined by using the Spencer limit equilibrium
input parameters for slope stability (H = 6 m, g = 16 kN/m3, w = method (Table 5).
50°, and ru = 0) and for the least favorite scenario (H = 10 m, g = 20 The obtained results indicate a high level of prediction accuracy,
kN/m3, w = 10°, and ru = 0.5). with the 95% of predicted values of Fs falling within 62 SE (Table 5).
As seen in Fig. 7(a), the effects of c and b on Fs for the best- The values of Fs determined by the Spencer limit equilibrium method
case scenario are qualitatively the same as for the average constant arecomparedwithitspredictedvaluesinFig.8.
values of other input parameters [Fig. 5(b)]. However, in the worst-
case scenario [Fig. 7(b)], the impact of the coaction of these param-
eters on Fs is significantly different. In this case, the stronger effect
of c over b , as shown in Fig. 3, is obvious, because for constant b Comparison with Existing Prediction Models
the safety factor significantly rises with the increase of c.
However, in the opposite case, for constant value of c, safety fac- To justify the need for a new prediction model, its prediction
tor Fs increases slightly with the decrease of slope angle b . accuracy has to be compared with the existing mathematical

© ASCE 04016009-7 Int. J. Geomech.

Int. J. Geomech., 04016009


Table 5. Values of Fs Determined by Spencer Method and Proposed Table 6. Range of Input Parameters for the Prediction Eqs. (3)–(5)
Model
Model by Model by
Low High Model by Ahangar-Asr Manouchehrian
Numerical Fs1 Fs2 Absolute 95% 95% Input Yang et al. et al. et al.
example (Spencer (proposed error: Mean confidence confidence parameter (2004) (2010) (2014)
no. method) model) Fs1  Fs2 SE level level
H (m) 3.66–214 3.66–214 3.6–214
1 0.730 0.637 0.093 0.083 0.47 0.81 b (degrees) 16–53 16–53 16–53
2 3.139 3.254 0.115 0.063 3.13 3.38 g (kN/m3) 12–28.44 12–28.44 12–28.44
3 1.533 1.487 0.046 0.061 1.36 1.61 c (kPa) 0–150.05 0–150.05 0–50
4 1.536 1.603 0.067 0.060 1.48 1.72 w (degrees) 0–45 0–45 0–45
5 2.224 2.160 0.064 0.070 2.02 2.30 ru 0.11–0.5 0–0.5 0–0.5
6 2.076 2.202 0.127 0.067 2.07 2.34
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

7 0.954 0.859 0.095 0.062 0.73 0.98


8 1.181 1.138 0.043 0.079 0.98 1.30 Table 7. Input Parameters of Sample Slopes for Comparison of the
9 0.615 0.556 0.059 0.084 0.39 0.73 Proposed Model 2 with the Existing Prediction Eqs. (3)–(5)
10 1.112 0.992 0.12 0.061 0.87 1.11
Numerical
example no. H (m) b (degrees) g (kN/m3) c (kPa) w (degrees) ru

4 1 6.6 26 16.8 9 15 0.1


2 7.2 36 17.2 46 22 0.22
3 8.4 42 17 24 41 0.41
4 9.6 33 19.4 18 35 0.39
5 6.8 38 18.3 38 11 0.13
3
Predicted values of Fs

6 8.8 45 19.8 44 29 0.26


7 9.1 51 16.2 21 33 0.48
8 7 28 19.6 15 12 0.31
9 9 53 17.5 4 44 0.2
2
10 6.9 32 18.9 12 19 0.11

which were developed by Yang et al. (2004), Ahangar-Asr et al.


1 (2010), and Manouchehrian et al. (2014), respectively, for the same
input parameters as those in the present study but using different
r=0.995 methods. Model 3 was developed using the method of genetic pro-
SE=0.066 gramming, whereas Model 4 was derived using a combination of
0 the least-squares method and a genetic algorithm, the last of which
0.5 1 1.5 2 2.5 3 3.5
was also used for developing Model 5. The range of input parame-
Values of Fs obtained by Spencer's method
ters used for deriving these models is given in Table 6.
As shown in Table 6, the analyzed range of input parameters is
Fig. 8. Comparison of values of Fs determined by the Spencer limit
broader for H, g , and c (except for the model by Manouchehrian
equilibrium method and its values predicted by the proposed Model 2,
indicating reliable predictive power of the developed mathematical
et al. 2014) than that from the model presented here. The range of
expression (external validation)
input values for ru is the same as that in the Model 2, with a slight
difference in the model by Yang et al. (2004), whereas the input
range for w is smaller when compared with that of the present
model. Conversely, the range for b significantly differs from that in
expressions. For this purpose, the following three models were
the present analysis (smaller values of both lower and upper limits).
analyzed (Models 3–5): This difference in the examined parameter range could cause differ-
1 ent predictions of Fs using the developed Model 2 and existing pre-
Fs ¼ ð1:406 þ 0:039c  csc2 w Þ þ 1:25ð1  ru Þtan w cot b diction Eqs. (3)–(5).
H
To examine the drawbacks/benefits of implementing the pres-
1
þ 0:156sec w þ 0:011 csc b (3) ent Model 2 over the existing prediction Eqs. (3)–(5), 10 sample
g cases of slopes, the properties of which fall within the input
ranges of all models, were chosen (Table 7). The obtained
1:49H results are given in Table 8.
Fs ¼   1:8ru2 þ tanð w Þ½2:59  2:18 tanð b Þ þ 0:014  c
g2 As seen in Table 8, proposed prediction Model 2 gives a signifi-
 5:19  105 c2 þ 0:817; (4) cantly higher prediction accuracy when compared with existing
Models 3–5. For simpler visualization, the prediction accuracy of
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi these models is shown in Fig. 9.
pffiffiffi 1:372
Fs ¼ 1:979 þ 3:25  107 expð g Þ þ 0:083 c  h i2
expðtan w Þ
Conclusion

 0:673 lnðtan b Þ  0:19 lnðHÞ  3:515ðru Þ3 (5) In this paper, a prediction model for safety factor (Fs) of homogene-
ous slope was developed by using the method of experimental

© ASCE 04016009-8 Int. J. Geomech.

Int. J. Geomech., 04016009


Table 8. Values of Fs Determined by the Spencer method, Proposed design. It was assumed that six main factors predetermine the stabil-
Model 2, and Existing Eqs. (3)–(5) ity of slope: height H, slope angle b , bulk density g , cohesion c,
Fs1 Fs2 angle of internal friction w , and pore pressure coefficient ru. Results
Numerical (Spencer (proposed Fs3 Fs4 Fs5 of the performed analysis indicate general linear dependence of Fs
example no. method) Model 2) (Model 3) (Model 4) (Model 5) on the aforementioned input parameters, including a nonlinear
quadratic influence of cohesion and slope angle. Moreover, the con-
1 1.313 1.128 (SE = 0.094) 1.788 1.295 1.547
ducted research implies the significant impact of two-factor interac-
2 3.139 3.254 (SE = 0.063) 2.530 1.634 1.734
tions on the slope stability. Influence of a single parameter lies in
3 1.822 1.883 (SE = 0.057) 1.350 1.323 1.570
increase of Fs with the increase of c and w , whereas Fs decreases
4 1.536 1.603 (SE = 0.060) 1.368 1.563 1.651
with the increase of H, b , and ru, for assumed average constant val-
5 2.334 2.275 (SE = 0.072) 6.747 1.386 1.357
ues of other input parameters. The impact of bulk density g on slope
6 2.230 2.331 (SE = 0.067) 1.650 1.405 1.608
stability is not significant, because the value of Fs decreases from
7 1.136 1.045 (SE = 0.063) 0.987 0.555 1.035
1.85 to approximately 1.62 with the increase of g from 16 to 20 kN/
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

8 1.181 1.138 (SE = 0.079) 2.597 1.120 1.361


9 0.799 0.752 (SE = 0.068) 1.136 0.464 1.312
m3. This is further confirmed by the analysis of two-factor inter-
10 1.378 1.290 (SE = 0.066) 1.594 1.350 1.526
actions, with no significant change in Fs for any values of g
with an increase of cohesion c, while other parameter values are
held constant at their average values.
However, because the two-factor interactions significantly affect
the slope stability, the impact of certain single input parameters will
change with different values of other factors. This is confirmed by

3.5 7
Values of Fs predicted by model (2)

Values of Fs predicted by model (3)


3

5
2.5

4
2
3
1.5
2

1
1

0.5 0
0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2 2.5 3 3.5
(a) Values of Fs obtained by Spencer’s method (b) Values of Fs obtained by Spencer’s method

2
Values of Fs predicted by model (4)

Values of Fs predicted by model (5)

1
0.5 1 1.5 2 2.5 3 3.5 0.5 1 1.5 2 2.5 3 3.5
(c) Values of Fs obtained by Spencer’s method (d) Values of Fs obtained by Spencer’s method

Fig. 9. Values of Fs determined by the Spencer method versus the values of Fs predicted by: (a) Model 2 (r = 0.99); (b) Model 3 (r = 0.46); (c)
Model 4 (r = 0.72); (d) Model 5 (r = 0.59); it is clear that Model 2 provides the highest prediction accuracy, with the smallest SE

© ASCE 04016009-9 Int. J. Geomech.

Int. J. Geomech., 04016009


the analysis of every significant two-factor interaction in the pro- Alavi, A. H., and Gandomi, A. H. (2011). “A robust data mining approach
posed Model 2, while other factors are held constant at their average for formulation of geotechnical engineering systems.” Eng. Comput.,
values. It turns out that the minimum value of safety factor is 28(3), 242–274.
obtained for the following two-factor interactions: the highest value Aryal, K. P. (2006). “Slope stability evaluations by limit equilibrium and fi-
nite element methods,” Ph.D. thesis, Norwegian Univ. of Science and
of b and the lowest value of c (for the c– b interaction), the lowest
Technology, Trondheim, Norway.
values of c and w (for the c– w interaction), high values of b Bishop, A. W., and Morgenstern, N. R. (1960). “Stability coefficients for
( b > 45°) and the lowest value of w (for the b – w interaction), and earth slopes.” Geotechnique, 10(4), 129–153.
the lowest value of w and the highest value of ru (for the ru– w inter- Brinkgreve, R. B. J. (2002). Plaxis 8.2 user’s manual, Delft University of
action). Regarding the remaining two significant interactions, those Technology and Plaxis bv, Delft, the Netherlands.
of c–H and c– g , minimum values of Fs are obtained for the lowest Cela, R., Ordoñez, E. Y., Quintana, J. B., and Rodil, R. (2013).
values of c, regardless of the values of H or g . Further analysis of “Chemometric–assisted method development in reversed–phase liquid
slope stability for the best- and worst-case scenarios has confirmed chromatography.” J. Chromatogr. A, 1287, 2–22.
Das, S. K., Biswal, R. K., Sivakugan, N., and Das, B. (2011).
the predominant effect of cohesion on slope stability. Moreover, the
“Classification of slopes and prediction of factor of safety using dif-
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

obtained results indicated that for the worst-case scenario (H = 10 ferential evolution neural networks.” Environ. Earth Sci., 64(1),
m, g = 20 kN/m3, c = 0 kPa, and ru = 0.5), slope cannot maintain 201–210.
high values of slope angle. Deming, S. N., and Morgan, S. L. (1996). Experimental design: a chemo-
The proposed model was further verified by analyzing 10 sam- metric approach, Elsevier, Amsterdam.
ples of slopes with random geometrical and soil properties within Donald, I. B., and Chen, Z. (1997). “Slope stability analysis by the upper
the examined range of input parameters. The obtained results indi- bound approach: Fundamentals and methods.” Can. Geotech. J., 34(6),
cate a high level of statistical reliability, with the predicted values of 853–862.
Fs within the 95% prediction interval and with a high correlation Doornbos, D. A., Smilde, A. K., DBoer, J. H., and Duineveld, C. A. A.
coefficient (r = 0.996). When compared with the existing prediction (1990). “Experimental design, response surface methodology and multi-
criteria decision making in the development of drug dosage forms.”
equations proposed by Yang et al. (2004), Ahangar-Asr et al.
Scientific computing and automation (Europe), E. J. Karjalainen, ed.,
(2010), and Manouchehrian et al. (2014), the proposed model esti- Elsevier, Amsterdam, 85–95.
mates the value of Fs with much higher accuracy. Along with this, Duncan, J. M., and Wright, S. G. (2005). Soil strength and slope stability,
the proposed models provide relatively simple relations among Wiley, New York.
slope properties and safety factors, whereas the existing equations Eid, H. (2014). “Stability charts for uniform slopes in soils with nonlinear
give a highly nonlinear dependence of Fs on input parameters, failure envelopes.” Eng. Geol., 168, 38–45.
including logarithmic, exponential, and cubic terms and trigono- Erzin, Y., and Cetin, T. (2013). “The prediction of the critical factor of
metric relations, whose impact on Fs cannot be easily evaluated. It safety of homogeneous finite slopes using neural networks and multiple
regressions.” Comput. Geosci., 51, 305–313.
has to be emphasized that different predictive power of developed
Fisher, R. A. (1935). The design of experiments, Haffner Press, New York.
model and existing equations could arise from slightly different GEO-SLOPE International. (2008). Stability modeling with SLOPE/W
ranges of the examined input parameters. 2007: An engineering methodology, 3rd ed., Calgary, AB, Canada.
One should note that the method of experimental design is Hibbert, D. B. (2012). “Experimental design in chromatography: A tutorial
applied in slope stability analysis for the first time in this paper. review.” J. Chromatogr. B, 910, 2–13, 22333438.
This method belongs to soft computing techniques, and it can pro- Huo, Y., and Zhai, H. (2012). “The study on slope stability analysis based
vide reliable results with a sufficient minimum of input parameters. on finite element method.” Adv. Mater. Res., 575, 70–74.
In this way, the number of analyses needed for calculation of Fs for Jiang, J.-C., and Yamagami, T. (2006). “Charts for estimating strength
an earth slope is significantly reduced without a loss of statistical parameters from slips in homogeneous slopes.” Comput. Geotech.,
33(6–7), 294–304.
reliability and prediction accuracy.
Kim, J., Salgado, R., and Lee, J. (2002). “Stability analysis of complex soil
The prediction equation, which is proposed by the current analy-
slopes using limit analysis.” J. Geotech. Geoenviron., 10.1061
sis, represents a basic model for slope stability prediction. Further /(ASCE)1090-0241(2002)128:7(546), 546–557.
analyses in this direction should include the effect of the bedrock Li, S., Zhao, H.-B., and Ru, Z. (2013). “Slope reliability analysis by
depth and the impact of external static and dynamic forces. The final updated support vector machine and Monte Carlo simulation.” Nat.
model should also involve the possibility of locating the position of Hazards, 65(1), 707–722.
critical slip surface. In that case, such a model would provide a full Li, Y.-C., Chen, Y.-M., Zhan, T. L. T., Ling, D.-S., and Cleall, P. J. (2010).
picture of slope stability, and it would have a more significant prac- “An efficient approach for locating the critical slip surface in slope sta-
tical application. bility analyses using a real-coded genetic algorithm.” Can. Geotech. J.,
47(7), 806–820.
Liu, Z., Shao, J., Xu, W., Chen, H., and Zhang, Y. (2014). “An extreme
Acknowledgments learning machine approach for slope stability evaluation and predic-
tion.” Nat. Hazards, 73(2), 787–804.
Manouchehrian, A., Gholamnejad, J., and Sharifzadeh, M. (2014).
This research was partly supported by the Ministry of Education,
“Development of a model for analysis of slope stability for circular
Science, and Technological Development of the Republic of mode failure using genetic algorithm.” Environ. Earth Sci., 71(3),
Serbia (Contract No. 176016). Special thanks also go to the 1267–1277.

Institute for the Development of Water Resources (Jaroslav Cerni) Slide 5.02 [Computer software]. Rocscience, Toronto.
for extensive communication during the research. Sun, J., Li, J., and Liu, Q. (2008). “Search for critical slip surface in slope
stability analysis by spline-based GA method.” J. Geotech.
Geoenviron., 10.1061/(ASCE)1090-0241(2008)134:2(252), 252–256.
References Tan, X.-H., Bi, W.-H., Hou, X.-L., and Wang, W. (2011). “Reliability anal-
ysis using radial basis function networks and support vector machines.”
Ahangar-Asr, A., Faramarzi, A., and Javadi, A. A. (2010). “A new Comput. Geotech., 38(2), 178–186.
approach for prediction of the stability of soil and rock slopes.” Eng. Taylor, D. W. (1937). “Stability of earth slopes.” J. Boston Soc. Civil Eng.,
Comput., 27(7), 878–893. 24, 197–246.

© ASCE 04016009-10 Int. J. Geomech.

Int. J. Geomech., 04016009


Yang, C., Tham Feng, X. T., Wang, Y. J., and Lee, P. K. K. (2004). “Two- Zhao, H.-B. (2008). “Slope reliability analysis using a support vector
stepped evolutionary algorithm and its application to stability analysis machine.” Comput. Geotech., 35(3), 459–467.
of slopes.” J. Comput. Civil Eng., 18(2), 145–153. Zhao, L.-H., Li, L., Yang, F., Luo, Q., and Liu, X. (2010). “Upper bound anal-
Zhang, X. (1999). “Slope stability analysis based on the rigid finite element ysis of slope stability with nonlinear failure criterion based on strength
method.” Geotechnique, 49(5), 585–593. reduction technique.” J. Cent. South Univ. Technol., 17(4), 836–844.
Zheng, H., Liu, D. F., and Li, C. G. (2005). “Slope stability analysis based Zolfaghari, A. R., Heath, A. C., and McCombie, P. F. (2005). “Simple
on elasto-plastic finite element method.” Int. J. Numer. Methods Eng., genetic algorithm search for critical non-circular failure surface in slope
64(14), 1871–1888. stability analysis.” Comput. Geotech., 32(3), 139–152.
Downloaded from ascelibrary.org by San Diego State University on 01/29/16. Copyright ASCE. For personal use only; all rights reserved.

© ASCE 04016009-11 Int. J. Geomech.

Int. J. Geomech., 04016009

You might also like