Slope stability analysis based on convolutional neural network and digital twin

In order to reduce damages caused by slope instability and landslide disasters, it is of great significance to find an efficient, accurate and time-saving method for slope stability analyses. This paper proposes a convolutional neural network based on digital twin models to predict the safety factor of a slope and be evaluate its stability state. In order to solve the problem of lack of the CNN training samples, the digital twin method is resorted to generate 4000 slope models from 10 real slopes by fine-tuning the geometric coordinates and material parameters of their soil layers. The finite element computation of the safety factor of these 4000 slope models were realized by using the parametric analysis of ABAQUS platform and 4000 slope datasets were obtained to serve as the CNN training samples. With the geometric coordinates and material parameters of the slopes as the CNN input and the slope safety factor as the CNN output, the slope safety factor can be effectively predicted. The results show that the prediction accuracy for the testing set reaches 96% and the root mean square error is 0.079. Compared with the finite element modeling time, the prediction time is greatly shortened. The evaluation accuracy of stability states for the 10 real slopes has reached 100%, which indicates that the CNN model has good generalization ability and prediction effect and has practical significance in engineering applications.