Research on Slope Stability Based on Improved PSO-BP Neural Network

Abstract:

Slope stability research is extremely important for the prevention and control of geologic hazards， but since factors affecting slope stability are rather diverse， indefinite， and nonlinear， slope stability analysis is always a hot but difficult problem. Studies have shown that neural network prediction models can be effectively applied in slope stability analysis. However， such models also have disadvantages of low accuracy in prediction， poor robustness， and slow convergence. Thus， an improved slope stability prediction model is proposed based on the PSO-BP model. In this model， input parameters include bulk density， cohesion， internal friction angle， slope angle， slope height and pore pressure ratio， and the output parameter is safety factor. The model borrows the idea of mutation in genetic algorithm to improve the global optimization， and applies the negative gradient descent principle of the energy function to improve the convergence speed. With the data cleaning process， eighty highquality slope data are obtained from over a hundred pieces of raw data. Then fifty data are randomly selected as the test data. Finally， a ten-fold cross-validation method is used to verify the accuracy of the model. Comparisons in multiple dimensions are also made with other models. The results show that： （1） Compared with traditional models， the presented model has significant improvement in aspects of convergence speed， accuracy， and robustness；（2） With small sample data， applying K-fold crossvalidation to the slope stability neural network prediction model can effectively avoid the contingency of the results；（3） The model has a small error of 4.31%， which meets the engineering accuracy requirements， so the model can provide reference for slope stability analysis and disaster prevention in real engineering projects.

Key words:slope stability; back propagation neural network; particle swarm optimization; disaster prevention; K‑fold cross validation