Professional Documents
Culture Documents
,(((
height. As input units authors use the previously obtained TABLE I: INPUT-OUTPUT FACTORS FOR ANN MODEL
results of stability analyzes using Spencer’s method and Case Output factor=f (input factors)
No. Fs=f(H, ȕ, Ȗ, c, ij, ru, d/H)
experimental design [20].
1 1.43 = f(14, 14, 19.5, 9, 13, 0, 2)
The paper is organized as follows. Brief description of 2 3.80 = f(14,14, 19.5, 90, 13, 0, 2)
the applied methods is given in Section 2. Section 3 is 3 2.22 = f(14, 14, 19.5, 9, 22, 0, 2)
devoted to the results of the analysis, including the 4 4.85 = f(14, 14, 19.5, 90, 22, 0, 2)
5 0.86 = f(14, 14, 19.5, 9, 13, 0.5, 2)
definition of the values of input factors, development of 6 3.03 = f(14, 14, 19.5, 90, 13, 0.5, 2)
ANN model, and the internal and external verification of 7 1.25 = f(14, 14, 19.5, 9, 22, 0.5, 2)
the model accuracy. Discussion and conclusion are given 8 3.58 = f(14, 14, 19.5, 90, 22, 0.5, 2)
9 11.91 = f(3, 14, 19.5, 49.5, 17.5, 0, 1)
in Section 5, together with suggestions for further 10 2.78 = f(25, 14, 19.5, 49.5, 17.5, 0, 1)
research. 11 11.02 = f(3, 14, 19.5, 49.5, 17.5, 0.5, 1)
12 2.01 = f(25, 14, 19.5, 49.5, 17.5, 0.5, 1)
II. APPLIED METHODS 13 7.89 = f(3, 14, 19.5, 49.5, 17.5, 0, 3)
In order to develop a reliable model for estimation of 14 2.51 = f(25, 14, 19.5, 49.5, 17.5, 0, 3)
15 6.60 = f(3, 14, 19.5, 49.5, 17.5, 0.5, 3)
earth slope stability, authors engage several methods. 16 1.66 = f(25, 14, 19.5, 49.5, 17.5, 0.5, 3)
Starting ranges of values for the input parameters (slope 17 10.52 = f(14, 3, 19.5, 49.5, 13, 0.25, 1)
height H and slope inclination ȕ, unit mass Ȗ, cohesion c, 18 2.06 = f(14, 25, 19.5, 49.5, 13, 0.25, 1)
19 13.23 = f(14, 3, 19.5, 49.5, 22, 0.25, 1)
angle of internal friction ij, pore pressure coefficient ru 20 2.43 = f(14, 25, 19.5, 49.5, 22, 0.25, 1)
and bedrock depth d) are chosen as representative values 21 5.97 = f(14, 3, 19.5, 49.5, 13, 0.25, 3)
for clay-marl deposits in Belgrade [20]. Within the 22 1.85 = f(14, 25, 19.5, 49.5, 13, 0.25, 3)
23 8.71 = f(14, 3, 19.5, 49.5, 22, 0.25, 3)
defined starting intervals different combinations of input 24 2.27 = f(14, 25, 19.5, 49.5, 22, 0.25, 3)
parameters are established using the method of 25 7.83 = f(3, 3, 19.5, 9, 17.5, 0.25, 2)
experimental design, with the final aim to uniformly 26 5.06 = f(25, 3, 19.5, 9, 17.5, 0.25, 2)
27 1.87 = f(3, 25, 19.5, 9, 17.5, 0.25, 2)
cover the examined parameter space, so that developed 28 0.76 = f(25, 25, 19.5, 9, 17.5, 0.25, 2)
model could provide reliable and statistically significant 29 34.25 = f(3, 3, 19.5, 90, 17.5, 0.25, 2)
responses. For each combination of input parameters, 30 8.38 = f(25, 3, 19.5, 90, 17.5, 0.25, 2)
slope safety factor is calculated using Spencer's method of 31 10.55 = f(3, 25, 19.5, 90, 17.5, 0.25, 2)
32 2.09 = f(25, 25, 19.5, 90, 17.5, 0.25, 2)
slices, which is considered to be the simplest method 33 1.61 = f(14, 14, 16, 9, 17.5, 0.25, 1)
which satisfies all the equlibrium equations of forces and 34 1.43 = f(14, 14, 23, 9, 17.5, 0.25, 1)
moments [21]. These values of safety factors, together 35 6.10 = f(14, 14, 16, 90, 17.5, 0.25, 1)
36 4.66 = f(14, 14, 23, 90, 17.5, 0.25, 1)
with corresponding combinations of parameter values are 37 1.51 = f(14, 14, 16, 9, 17.5, 0.25, 3)
used for creating the adequate estimation model by 38 1.38 = f(14, 14, 23, 9, 17.5, 0.25, 3)
applying mutlilayer feed-forward neural network with 39 4.34 = f(14, 14, 16, 90, 17.5, 0.25, 3)
40 3.47 = f(14, 14, 23, 90, 17.5, 0.25, 3)
Levenberg-Marquardt learning algorithm, and 60% of 41 8.23 = f(3, 14, 16, 49.5, 13, 0.25, 2)
data for training, 15% for validation and 25% for testing. 42 1.93 = f(25, 14, 16, 49.5, 13, 0.25, 2)
Results obtained are then compared with the results of 43 6.08 = f(3, 14, 23, 49.5, 13, 0.25, 2)
44 1.63 = f(25, 14, 23, 49.5, 13, 0.25, 2)
Spencer's method for values of input factors different 45 9.11 = f(3, 14, 16, 49.5, 22, 0.25, 2)
from those used for model derivation, and using the 46 2.62 = f(25, 14, 16, 49.5, 22, 0.25, 2)
existing mathematical relation defined for the similar 47 6.96 = f(3, 14, 23, 49.5, 22, 0.25, 2)
48 2.29 = f(25, 14, 23, 49.5, 22, 0.25, 2)
range of input values as in the present case. 49 10.73 = f(14, 3, 16, 49.5, 17.5, 0, 2)
50 2.64 = f(14, 25, 16, 49.5, 17.5, 0, 2)
III. RESULTS 51 9.34 = f(14, 3, 23, 49.5, 17.5, 0, 2)
52 2.14 = f(14, 25, 23, 49.5, 17.5, 0, 2)
A. Definition of input factor values 53 7.52 = f(14, 3, 16, 49.5, 17.5, 0.5, 2)
For starting intervals of the chosen input parameters, 54 2.02 = f(14, 25, 16, 49.5, 17.5, 0.5, 2)
55 6.20 = f(14, 3, 23, 49.5, 17.5, 0.5, 2)
authors chose the most commonly determined values for 56 1.57 = f(14, 25, 23, 49.5, 17.5, 0.5, 2)
clay-marl deposits in Belgrade [20]: 57 2.74 = f(14, 14, 19.5, 49.5, 17.5, 0.25, 2)
Slope height H (m): 3-25
Slope inclination ȕ (0): 3-25 B. Neural network model
Unit mass Ȗ (kN/m3): 16-23 Following the suggestion of Rumelhart et al. [22],
Cohesion c (kPa): 9-90 Lippmann [23] and Sonmez et al. [24] one hidden layer
Friction angle ij (0): 13-22 was chosen in present study, while authors chose 14
Pore pressure coefficient ru: 0-0.5 hidden nodes as suggested by Kannellopoulas and
Bedrock depth d/H: 1-3 Wilkinson [25], who proposed that number of hidden
Using the predetermined ranges from Table I, 57 nodes should be two time larger than the number of input
factors.
different combinations of input factors were derived using
The initial weights were set in the range [-1.0;1.0]. In
experimental design, so as to uniformly cover the
all the examined cases, the total data set has been divided
examined parameter space. For these combinations, slope as following: 60% for training (34 recordings), 15% for
safety factor is calculated using Spencer’s method of validation (9 recordings) and 25% for testing (14
slices [20]. These data are used for derivation of neural recordings), which corresponds well with the suggestion
network model (Table I).
of Nelson and Illingworth [26] who supported the idea of
20-30% of data for testing.
Considering different units of input factors, their
scaling was necessary, and it was performed in the
following way:
scaled value =
(max.value − unscaled value) (1)
(max.value − min.value)
In that way, numerical values of the analyzed
parameter were normalized in the range of [0,1].
In present paper, in order to create an adequate ANN
model for estimation of slope stability, based on the
calculated data, a three-layer artificial neural network is
chosen using Levenberg-Marquardt learning algorithm.
This training algorithm is commonly considered as the
fastest method for training moderate-sized feed-forward
neural networks [27], and it is the first choice for solving
the problems of supervised learning, which is the case in
present analysis. A sigmoid function was chosen for the
activation function, as the most common transfer function
implemented in the literature [24].
Fig. 2. Comparison of the scaled values of Fs estimated by
The first task in development of a reliable ANN model
ANN model and Spencer’s method of slices using training set
is to exclude the possibility of overfitting, when the ANN (a), validation set (b) and testing set (c), for ANN model with
model tends to memorize instead of learning the data. In 14 hidden nodes.
order to minimize overfitting, any increase in accuracy
over the training data set should yield an increase in Regarding the properties of residuals, obtained results
accuracy over a data set that the network has not trained indicate that residuals for the testing set represent
on it. As it is clear from Figure 1, the training stopped sequence of independent numbers that follow normal
when the validation error increased for six iterations, distribution, indicated by p-value larger than 0.05 for
which occurred at iteration 3. In present case, the result is Anderson-Darling test (Fig. 3).
reasonable because the final mean-square error is small
(MSE=0.0035), the test set error and the validation set
error have approximately the same properties and no
significant overfitting has occurred by iteration 3 (where
the best validation performance occurs).
C. Internal verification
In order to further confirm the reliability of the
proposed model (2), its predictive power is tested by
analyzing 5 random samples of slopes with geometrical
characteristics and soil properties, different from those
used for deriving the model (2) (Table III). Exact values
REFERENCES
[1] A.W. Bishop, “The Use of the Slip Circle in the Stability Analysis
of Slopes,” Geotechnique, Vol. 5, pp 7 – 17, 1955.
[2] N. Janbu, ‘’Slope Stability Computations,’’ in Embankment Dam
Engineering - Casagrande Volume, R.C. Hirschfeld and S.J.
Poulos, Ed.,New York: John Wiley and Sons, 1973, pp 47-86.
[3] E. Spencer, A method of analysis of the stability of embankments
assuming parallel interslice forces. Geotechnique, Vol. 17pp. 11-
26, 1967.
[4] N.R. Morgenstern and V.E. Price, ‘’The analysis of the stability
of general slip surfaces,’’ Geotechnique, Vol. 15, pp. 79-93, 1965.
[5] D.W. Taylor, ‘’Stability of earth slopes,’’ Journal of the Boston
Society of Civil Engineering, pp. 24-31, 1937.
[6] N. Janbu, "Stability Analysis of Slopes with Dimensionless
Parameters," in Soil Mechanics Series, No. 46, Harvard
University, 81 p., 1954.
[7] A.W. Bishop, and N.R. Morgenstern, ‘’Stability coefficients for
earth slopes,’’ Geotechnique, Vol. 10, pp. 129-150, 1960.
[8] E. Hoek and J.W. Bray. Rock slope engineering, third Edition,
Fig. 4. Comparison of the slope safety factor determined B.C., Canada, 1981.
by the Spencer's method (light gray column), proposed model [9] Y. Huo and H. Zhai ‘’The study on slope stability analysis based
(2) (dark gray column) and by the existing equation (3) (black on finite element method,’’ Advanced Materials Research, Vol.
column). Values for input factors are given in Table III. 575, pp. 70-74, 2012.
[10] L.H. Zhao, L. Li, F. Yang, Q. Luo and X. Liu, ‘’Upper bound
analysis of slope stability with nonlinear failure criterion based on
As it can be seen in Fig. 4, proposed prediction model strength reduction technique,’’ Journal of Central South
(2) provides similar prediction accuracy as the existing University of Technology, Vol. 17, pp. 836-844, 2010.
model (3). One could notice that Eq. (3) does not include [11] S.K. Das, R.K. Biswal, N. Sivakugan and B. Das, ‘‘Classification
the impact of bedrock depth on slope stability. Still, it of slopes and prediction of factor of safety using differential
evolution neural networks,’’ Environmental Earth Sciences, Vol.
provides similar prediction accuracy as the derived model 64, pp. 201-210, 2011.
(2), which implies that the effect of bedrock depth on [12] Y. Erzin and T. Cetin, ‘’The prediction of the critical factor of
slope stability, in present case, is minimal. safety of homogeneous finite slopes using neural networks and
multiple regressions,’’ Computers and Geotechnics, Vol. 51, pp.
305-313, 2013.
[13] K. Gelisli, T. Kaya and A.E. Babacan, ‘’Assessing the factor of
IV. CONCLUSION safety using an artificial neural network: case studies on
In present paper, we developed a prediction model for landslides in Giresun Turkey,‘’ Environmental Earth Sciences,
Vol. 73, pp. 8639–8646, 2015.
safety factor (Fs) of homogeneous slope based on
[14] A. Gomes Correia, P. Cortez, J. Tinoco and R. Marques,
artificial neural networks. It was assumed that seven main ‘’Artificial Intelligence Applications in Transportation
factors predetermine the slope stability: height H, slope Geotechnics’’ Geotechnical and Geological Engineering, Vol.
inclination ȕ, unit mass Ȗ, cohesion c, angle of internal 31, pp. 861–879, 2013.
[15] J.Lai, J. Qiu, Z. Feng, J. Chen and H. Fan, ‘’Prediction of Soil
friction ij, pore pressure coefficient ru and bedrock depth
Deformation in Tunnelling Using Artificial Neural Networks,’’
d/H. As input values, authors use the results of stability Computational Intelligence and Neuroscience
analyzes of 57 different slopes, which are previously set Vol. 2016 Article ID 6708183, 16 pages, 2016.
using in situ determined range of values and [16] V. Kohestani and M. Hassanlourad, "Modeling the Mechanical
Behavior of Carbonate Sands Using Artificial Neural Networks
experimental design. In testing phase, proposed model (2)
and Support Vector Machines," International Journal of
shows high correlation coefficient (R=0.95) and small Geomechanics, doi: 10.1061/(ASCE)GM.1943-5622.0000509,
MSE (0.0035) for scaled values. The proposed model is 04015038, 2015.
verified by analyzing five samples of slopes with random [17] J. Tinoco, A. Gomes Correia, P. Cortez, ‘’A novel approach to
predicting Young's modulus of jet grouting laboratory
geometrical and soil properties within the examined range
formulations over time using data mining techniques,’’
of input parameters. The obtained results indicate high Engineering Geology, Vol. 169, pp. 50–60, 2014.
level of estimation accuracy, with the difference from the [18] M. M. Yadollahia, A. Benlia and R. Demirbog, ‘’Prediction of
results of Spencer’s method smaller than 0.32. When compressive strength of geopolymer composites using an
artificial neural network,’’ Materials Research Innovations Vol.
compared to the existing mathematical expression
19, pp. 453-458, 2015.
proposed by Yang et al. [28], derived model provides [19] D. J. Armaghani, R. Shoib, K. Faizi, A. Rashid, ‘’Developing a
similar estimation accuracy. hybrid PSO–ANN model for estimating the ultimate bearing
One could notice that, on the basis of the results of capacity of rock-socketed piles,’’ Neural Computing and
Applications doi: 10.1007/s00521-015-2072-z
conducted research, model based on artificial neural
[20] S. Kostiü, N. Vasoviü and D. Jevremoviü. 2015. Stability of earth
networks is able to provide reliable results reducing the slopes under the effect of main environmental properties of
number of analyzes needed for calculation of Fs, without weathered clay-marl deposits in Belgrade (Serbia). Environmental
the loss of statistical reliability and estimation accuracy. Earth Sciences, Vol. 75, doi: 10.1007/s12665-016-5339-5
[21] J.M. Duncan, S.G. Wright, Soil strength and slope stability. New
It should be emphasized that model proposed by the
York: Wiley.
current analysis, represents a simplified model for [22] D.E. Rumelhart, G.E. Hinton and R.J. Williams, ‘’Learning
estimation of slope safety factor. Future studies in this internal representation by error propagation’’ In Parallel
direction should include the derivation of estimation distribution processing, D.E.,Rumelhart and J.L.McCleland, Eds.
1986, 1318-1362
model for heterogeneous slope conditions, with the
[23] R.P. Lipmann, ‘’An introduction to computing with neural nets,’’
possibility of locating the rupture surface. IEEE ASSP Mag, Vol. 4, pp. 4-22, 1987.
[24] H. Sonmez, C. Gokceoglu, H.A. Nefeslioglu, A. Kayabasi,
‘’Estimation of rock modulus: For intact rocks with an artificial
neural network and for rock masses with a new empirical
equation,’’ International Journal of Rock Mechanics and Mining
Sciences, Vol. 43, pp. 224-235, 2006.
[25] I. Kanellopoulas and G.G. Wilkinson, ‘’Strategies and best
practice for neural network image classification,’’ International
Journal of Remote Sensing, Vol. 18, 711-725, 1997.
[26] M. Nelson, W.T. Illingworth, A practical guide to neural nets,
Reading MA: Addison-Wesley, 1990.
[27] B. Tiryaki, ‘’Application of artificial neural networks for
predicting the cuttability of rocks by drag tools’’ Tunneling and
Underground Space Technology, Vol. 23, pp. 273-280, 2008.
[28] C. Yang, X.T. Tham Feng, Y. J. Wang, and P.K.K. Lee, “Two-
stepped evolutionary algorithm and its application to stability
analysis of slopes.” Journal of Computing in Civil Engineering,
Vol.18, pp. 145–153, 2004.