Abstract:

When the traditional limit equilibrium method is used for slope reliability analysis，because of the performance function is implicit and the form is complicated，the iterative process of solving the function becomes complicated and the computational efficiency is low.Aiming at the above problems，a CNN-LSTM model method was proposed.The principle of this method is to first extract the data features by using convolutional neural network（CNN），and then predict the slope failure probability by using short and long time memory network（LSTM）.On the basis of fully considering the value range of the CNN-LSTM model’s hyperparameters，the five-factor and four-level orthogonal test table was used to design the hyperparameters.Finally，the convolutional output dimension of the first layer and the second layer of the CNN network architecture in the CNN-LSTM model were determined to be 64 and 8 respectively.Dropout ratio is 0.5，the number of the first layer of the LSTM structure is 5 units and the number of the second layer of hidden layers is 20 units，respectively.The 420 slope sample data collected from central and western regions of China were used to train the model according to the ratio of 7∶3 between the training set and the verification set，and the optimal parameters of the CNN-LSTM model were obtained. Finally，Yanshanji landslide was taken as an example to illustrate the feasibility of the model method.The CNN-LSTM model was compared with Monte Carlo method（MCS），response surface method，single CNN，LSTM model and multiple linear regression model in terms of computational efficiency and failure probability prediction.The results show that：（1）When the MCS sampling times is 10 000，compared with the traditional MCS，although the CNN-LSTM model has a relative error of 4.35% in predicting the slope failure probability，in terms of computational efficiency，the CNN-LSTM model takes 45.28 s and the MCS takes 119 s，so the CNN-LSTM model increases the efficiency nearly 2 times.（2）When the single CNN model and LSTM model both adopt two-layer architecture，although the number of parameters of the CNN-LSTM model is not optimal，it has excellent performance in terms of calculation time and prediction accuracy of failure probability due to the small overfitting risk of the model.Compared with the multiple linear regression model，the relative error of CNN-LSTM prediction is 4.35%，and that of multiple linear regression is 34.78%.Through the above two points，the CNN-LSTM model can well complete the analysis of slope reliability，and avoid solving the implicit performance function，and the work efficiency is high.

Key words: slope, reliability, orthogonal test design, hyperparameter, CNN-LSTM model