This is the published version.

Li, A.J., Khoo, S.Y., Wang, Y. and Lyamin, A.V. 2014, Application of neural network to rock slope stability assessments. In Hicks, Michael A., Brinkgreve, Ronald B.J.. and Rohe, Alexander. (eds), Numerical methods in geotechnical engineering, CRC Press, London, England, pp. 473-478. This is the published version. 2014, Taylor & Francis Reproduced by Deakin University with the kind permission of the copyright owner. Available from Deakin Research Online: http://hdl.handle.net/10536/dro/du:30070564

Numerical Methods in Geotechnical Engineering Hicks, Brinkgreve & Rohe (Eds) 2014 Taylor & Francis Group, London, 978-1-138-00146-6 Application of neural network to rock slope stability assessments A.J. Li, S.Y. Khoo & Y. Wang School of Engineering, Deakin University, Geelong, Australia A.V. Lyamin Centre for Geotechnical & Materials Modelling, University of Newcastle, Newcastle, Australia ABSTRACT: It is known that rock masses are inhomogeneous, discontinuous media composed of rock material and naturally occurring discontinuities such as joints, fractures and bedding planes. These features make any analysis very difficult using simple theoretical solutions. Generally speaking, back analysis technique can be used to capture some implicit parameters for geotechnical problems. In order to perform back analyses, the procedure of trial and error is generally required. However, it would be time-consuming. This study aims at applying a neural network to do the back analysis for rock slope failures. The neural network tool will be trained by using the solutions of finite element upper and lower bound limit analysis methods. Therefore, the uncertain parameter can be obtained, particularly for rock mass disturbance. 1 INTRODUCTION Generally speaking, rock masses are inhomogeneous, discontinuous media composed of rock material and naturally occurring discontinuities such as joints, fractures and bedding planes. These features make any analysis very difficult using simple theoretical solutions, like the limit equilibrium method. However, predicting the stability of rock slopes is a classical problem for geotechnical engineers. Currently, the conventional linear Mohr-Coulomb criterion is widely used to assess rock mass strength. It could be due to the Mohr-Coulomb strength parameters, cohesion and friction angle, being the required inputs for most computer programs. In fact, it is known that the failure envelope of rock masses is nonlinear (Hoek 1983, Sheorey 1997, Ramamurthy 1995). It means that the Mohr-Coulomb criterion would not be suitable for representing rock mass strength. Li et al. (2009, 2012) have indicated that using a linear failure criterion will overestimate the factor of safety (F) for rock slopes. It is due to the nonlinearity is more pronounced at the low confining stresses that are operational in slope stability problems (Fu & Liao 2010). This statement agreed with the finding of Li et al. (2012) that the majority stress conditions from the failure surface are located in Region 1, as shown in Figure 1. As discussed by Merifield et al. (2006), the Hoek- Brown failure criterion is one of the few non-linear criteria used by practising engineers to estimate rock mass strength. In this paper, this yield criterion is also adopted as the failure envelope for the rock masses of the slopes. The latest Hoek-Brown failure criterion Figure 1. Hoek-Brown and equivalent Mohr-Coulomb failure criteria. (Hoek et al. 2002) for rock masses is expressed by the following equations: where 473

with the magnitudes of m b, s and a relying on the geological strength index (GSI), which describes the rock mass quality, and σ ci and m i representing the intact uniaxial compressive strength and material constant respectively. The parameter D is a factor that depends on the degree of disturbance. The suggested value of the disturbance factor is D = 0 for undisturbed in situ rock masses and D = 1 for disturbed rock mass properties. The investigation of Li et al. (2011) showed that the differences in the factor of safety evaluations for rock slopes can be significant if rock mass disturbance is considered, particularly for the cases with low GSI. In addition, Hoek et al. (2002) indicated that the disturbance factor (D) should be determined with caution. The importance of estimating D can therefore be seen. As highlighted by Burland (1989), some of the geotechnical parameters used in the analysis may not be accurately measured directly from laboratory tests due to effects of sample disturbance and errors of tests. The back analysis or the observational method, as suggested by Peck (1969), is thus often applied to determine the representative and/or dominant soil parameters based on field observations in practice. This paper will use one of the optimisation techniques, Artificial Neural Network (ANN), either to assess the rock slope stability or to calculate the uncertain parameter in a back analysis. The purpose of this study is to perform rock slope stability analysis more accurately and quickly. 2 METHODOLOGY 2.1 Finite element upper and lower bound limit analysis methods By using finite element upper and lower bound limit analysis methods (Lyamin & Sloan 2002a, b, Krabbenhoft et al. 2005), a non-dimensional stability number, Equation 5, was proposed by Li et al. (2008). Equation 5 is also employed in this study to evaluate rock slope stability. where N r is the stability number, γ is the unit weight of the rock mass, and H is the height. It should be noted that safety factors obtained from Equation 5 are only the same as the conventional definition when F = 1 (Li et al. 2012). The average solutions obtained from the numerical upper and lower bound limit analysis methods (Lyamin & Sloan 2002a, b, Krabbenhoft et al. 2005) are used to train ANN. To perform training for ANN so that it can act as the back analysis tool, slope angle (β), GSI, m i and D are chosen as the training inputs while N r is the desired training output of the ANN. Figure 2. Single-hidden layer neural network. 2.2 Artificial neural network Since ANN has been proven to be a universal approximator, the linear combinations of the nonlinear neurons and weights, after proper training or selections, can approximate any linear or nonlinear functions. It is a nice property that motivates us to choose a single hidden layer feed forward neural network to map β, GSI, m i, and D to N r. The trained ANN will be treated as a continuous differentiable mapping of the inputs to the outputs. A single-hidden layer feed forward neural network is described in Figure 2. In this paper, the extreme learning training machine (ELM) (Huang et al. 2006) is used to train our network. It is worth pointing out that in the area of neural computing, gradient based back propagation (BP) algorithm is commonly used for weight training. The BP algorithm is computed based on the error between the ANN output and the desired output. The output weights are then updated based on this error iteratively until the error converges to zero or any predefined sufficiently small number. This is a very time consuming process and the error convergence rate is slow. Using the ELM training algorithm, the weights of the ANN can be randomly assigned and the ANN can be treated as a linear system. But the most important feature of the ELM algorithm is its batch learning capability that trains the ANN in one operation of the global optimization and therefore the learning speed of the ELM can be significantly faster than the BP and all the BP like algorithms. Because of this remarkable merit, ELM has recently received a great deal of attention in computational intelligence, with application and extension to many other areas. The input data vector x(k) and the output data vector y(k) can be expressed as follows: 474

and the ith output of the neural network, y i (k) can be expressed as: Based on the study of Yu et al. (2004), the terminal SD algorithm can be constructed using a novel error criterion Suppose we have N training input vectors x(1), x(2),, x(n) and N desired output data vectors y d (1), y d (2),, y d (N) for training the ANN in Figure 1. it is easy to get where η 1, η 2 > 0, p, q (p > q) are odd integers. The corresponding SD algorithm is as follows: Based on the universal approximation property of ANN and the least square estimation of the optimal weight for ANN to approximate the desired output, N r, we can write with where α j1 and w T j, j = 1,..., M are optimal constants and x is the optimal constant vector that is unknown with Matrix G is called the hidden layer output matrix. Based on the ELM training algorithm, the input weights and the biases of the hidden layer of the ANN can be randomly assigned. The output weight matrix, G, of the ANN can be computed in a single iteration where and The optimal constants α j1 and w T j for j = 1,..., M can be obtained using the training process of the ANN using the training algorithm, Equation 11. For back analysis applications, given any new desired output, N r, our goal is to adaptively construct the optimal inputs, x = [β GSI m i D ], such that the trained ANN can optimally approximate N r. If the input vector, x, of the ANN, Equation 8, is continuously updated according to Equation 15 to approximate the desired output, N r, then the ANN is stable and the error e = y N r will converge to zero in a finite time based on the Lyapunov function, Equation 18. 2.3 Terminal steepest descent algorithm As mentioned previously, the Back Propagation learning algorithm for training the ANN is very slow. It is partly due to the steepest descent (SD) training algorithm that is used to minimise the approximation error at the output layer of the ANN to adjust the output weights. This study will use a fast steepest descent algorithm named terminal steepest descent algorithm motivated by the finite-time stability theory (Yu et al. 2004, Khoo et al. 2013). Let us first discuss the steepest descent algorithm. SD learning requires that, at any time, the performance of the dynamic system be assessable through a certain error function that measures the discrepancy between the trajectories of the dynamical system and the desired behavior. The least square error criterion is often selected as the error function: Since ξ i = 0, V will converge to zero in a finite time based on the finite-time stability theory (Yin et al. 2011, Khoo et al. 2013). Since V = 0.5e 2 and V = 0, it implies that e = 0. This means y = N r is reached in a finite time. It should be noted that this study is the first paper where the finite-time stability theory is implemented to deal with civil engineering problems. This can make the analysis more time-effective. 475

Table 1. The obtained safety factors from developed tool based on ANN. Figure 3. Training performance of ANN using ELM algorithm with 2000 data set. Figure 4. Performance of the trained ANN with 220validation data set. 3 TRAINING AND VALIDATION In this paper, β, GSI, mi, and D are chosen as the inputs while Nr is the output of the ANN. This means that the ANN has 4 inputs and 1 output. For the training, 2000 training data are selected randomly. Each data is an average stability number generated by the numerical upper and lower bound limit analysis methods. For the construction of a continuous differentiable mapping of the inputs to the outputs, we consider an ANN with 200 hidden nodes. The input weights are generated randomly within [ 1, 1], and the sigmoid function which is continuously differentiable, is used as the nonlinear activation function of the hidden nodes. The output weight matrix α is computed using the ELM algorithm (Eq. 11). Figure 3 are the training results based on 2000 data. The validation data set is displayed in Figure 4. It is noted that, after training with the ELM algorithm, the highly accurate continuous input-output mapping performance has been achieved. 4 CASE STUDIES By using ANN, the developed tool has a more direct function. It can be used to predict the safety factors (F) for rock slopes. An assumed case presented by Li D Nr F 0.0 0.7 1.0 4.22 17.4 54.3 8.25 2.00 0.64 et al. (2008) is employed herein to verify the accuracy of the tool. The slope has the following parameters: the slope angle β = 60, the height of the slope H = 25 m, the intact uniaxial compressive strength σ ci = 20 MPa, geological strength index GSI = 30, intact rock yield parameter mi = 8, and unit weight of rock mass γ = 23 kn/m3. Based on the information above, σ ci /γh = 20000/(23 25) = 34.8. Table 1 shows the Nr obtained from the ANN tool for different values of D. The factor of safety can be calculated as (σ ci /γh )/Nr. The calculated F is also displayed in Table 1 where F = 8.25, 2.00, and 0.64 for D = 0.0, 0.7, and 1.0, respectively. The computing time is very short (few seconds) for each case. It shows the rock slope assessment can be done immediately, and thus the developed tool is useful for practising engineers. In fact, F should decrease with increasing D. Therefore the trend obtained is reasonable. In addition, the presented F in Li et al. (2008) is of 8.7 for D = 0.0 which is quite close to the presented result in this study. It should be noted that F = 8.7 (Li et al. 2008) was obtained based on the observation of stability charts, and thus a certain level of the visual error would exist. Based on the discussion above, the accuracy and use of this developed tool have been verified. Moreover, the developed tool can be used to find an uncertain parameter in back analyses by considering F = 1. An assumed case investigated by Li (2009) and Li et al. (2011) is employed as well. Based on the previous studies (Hoek et al. 2002, Li et al. 2011), D plays an important role in rock slope evaluations, however it is difficult to be estimated. Therefore, D is the parameter which will be identified in this study. The assumed slope has the slope angle β = 60, the height of the slope H = 50 m, the intact uniaxial compressive strength σ ci = 10 MPa, geological strength index GSI = 30, intact rock yield parameter mi = 8, and unit weight of rock mass γ = 23 kn/m3. Hence, σ ci /γh = 10000/(23 50) = 8.7. The only unknown parameter is the rock mass disturbance (D). If this slope is assumed to be failed, the back calculated D is 0.43, as shown in Figure 5. Compared with the results in Li (2009) and Li et al. (2011), where F = 1.9 for D = 0.0 and F = 0.51 for D = 0.7 were presented, D = 0.43 is reasonable. Figure 5 also reveals that the convergence can be done very quickly (less than 5 seconds). It can be seen that D achieves a constant in a short time. Simultaneously, the error converges to zero. An interesting phenomenon is found that the input starting point of D does not influence the final obtained results. 476

slope stability analyses. The training data are obtained based on the finite element upper and lower bound limit analysis solutions. The developed tool can provide either prompt rock slope stability estimations or back calculations. It is very useful for practicing engineers, particularly for decision making. REFERENCES Figure 5. Convergences of D and Error for the back analysis. Figure 6. Convergence of D with negative starting point. Figure 7. Convergence of D with starting point larger than 1. In Figure 6, D is starting from a negative value and then reaches 0.43. Similar trend can be seen in Figure 7 that D finally achieves 0.43, even if the starting point is larger than 1. In fact, the convergence would take longer time, but it is still less than 5 seconds. From the engineering perspective, the correct result is more important. It should be stated that both of these inputs are not reasonable values for D because it must be between 0 and 1. However, this result implies that the developed tool is very useful to help junior engineers to pick up experiences in engineering judgments. Although their inputs are incorrect, this tool can help them to find the right answer in back analyses. 5 CONCLUSIONS This study adopts the techniques, Artificial Neural Network (ANN) and terminal steepest descent algorithm, to develop a fast evaluation tool for rock Burland, J.B. 1989. Ninth Laurits Bjerrum memorial lecture: small is beautiful-the stiffness of soils at small strain. Can. Geotech. J., 26: 499 516. Fu, W. & Liao, Y. 2010 Non-linear shear strength reduction technique in slope stability calculation. Computer and Geotechnics, 37(3): 288 298. Hoek, E. 1983. Strength of jointed rock masses. Geotechnique, 33(3): 187 223. Hoek, E., Carranza-Torres, C. & Corkum, B. 2002. Hoek- Brown Failure criterion-2002 edition. Proceedings of the North American rock mechanics society meeting. Toronto. Huang, G.B., Chen, L. & Siew, C. 2006 Universal approximation using incremental constructive feedforward networks with random hidden nodes. IEEE Transactions on Neural Networks, 17(4): 879 892. Huang, G.B., Zhou, H., Ding, X. & Zhang, R. 2012. Extreme learning machine for regression and multiclass classification. IEEE Transactions on Systems, Man, and Cybernetics, 42(2): 513 529. Khoo, S., Yin, J., Man, Z. & Yu, X. 2013. Finite-time stabilization of stochastic nonlinear systems in strict-feedback form. Automatica, 49(5): 1403 1410. Krabbenhoft, K., Lyamin, A.V., Hjiaj, M. & Sloan, S.W. 2005. A new discontinuous upper bound limit analysis formulation. International Journal for Numerical Methods in Engineering 63: 1069 1088. Li, A.J. 2009. Two and three dimensional stability analyses for soil and rock slopes. PhD, The University of Western Australia Li, A.J., Cassidy, M.J., Wang, Y., Merifield, R.S. & Lyamin, A.V. 2012. Parametric Monte Carlo studies of rock slopes based on the Hoek Brown failure criterion. Computers and Geotechnics, 45: 11 18. Li, A.J., Lyamin, A.V. & Merifield, R.S. 2009. Seismic rock slope stability charts based on limit analysis methods. Computers and Geotechnics, 36(1 2): 135 148. Li, A.J., Merifield, R.S. & Lyamin, A.V. 2008. Stability charts for rock slopes based on the Hoek-Brown failure criterion. International Journal of Rock Mechanics & Mining Sciences 45(5): 689 700. Li, A.J., Merifield, R.S. & Lyamin, A.V. 2011. Effect of rock mass disturbance on the stability of rock slopes using the Hoek-Brown failure criterion. Computers and Geotechnics, 38(4): 546 558. Lyamin, A.V. & Sloan, S.W. 2002a. Lower bound limit analysis using non-linear programming. International Journal for Numerical Methods in Engineering 55: 573 611. Lyamin, A.V. & Sloan, S.W. 2002b. Upper bound limit analysis using linear finite elements and non-linear programming. International Journal for Numerical and Analytical Methods in Geomechanics 26: 181 216. Man, Z., Lee, K., Wang, D., Cao, Z. & Khoo, S. 2012. Robust single-hidden layer feedforward network-based pattern classifier. Neural Networks and Learning Systems, 23(12): 1974 1986. Merifield, R.S., Lyamin, A.V. & Sloan, S.W. 2006. Limit analysis solutions for the bearing capacity of rock masses 477

using the generalised Hoek-Brown criterion. International Journal of Rock Mechanics & Mining Sciences, 43: 920 937 Peck, R.B. 1969. Advantages and limitations of the observational method in applied soil mechanics. Geotechnique, 19(2): 171 187. Ramamurthy, T. 1995. Bearing capacity of rock foundations, In: Yoshinaka & Kikuchi (eds.), Rock foundation, Rotterdam: Balkema. Sheorey, P.R. 1997. Empirical rock failure criteria. Rotterdam, Netherlands: A. A. Balkema. Yin, J., Khoo, S., Man, Z. & and Yu, X. 2011 Finite-time stability and instability of stochastic nonlinear systems. Automatica, 47(12): 2671 2677. Yu, S., Yu, X. & Man, Z. 2004. A fuzzy neural network approximator with fast terminal sliding mode and its applications. Fuzzy Sets and Systems, 148(3): 469 486. 478