Slope reliability analysis based on deep learning of digital images of random fields using CNN

Considering the spatial variability of soil strength, a deep learning model for the characteristics of random fields is proposed for reliability analysis of slope stability. The random fields of a soil slope are discretized by the Karhunen-Loeve expansion method, and the discretized results are converted into digital images. Then, a convolutional neural network (CNN) surrogate model is established to approach the implicit relationship between the images and the responses of the performance function. Based on the surrogate model, the probability of failure of the slope is calculated. When training the CNN surrogate model, the Latin-Hypercube sampling technique, Bayesian optimization and 5-fold cross-validation are employed to improve the accuracy. Finally, the effectiveness of the proposed method is demonstrated by two case studies, namely a single-layer saturated clay slope under undrained conditions and a two-layered cohesive soil slope. The results show that in the case of high dimensions and small probability, the proposed CNN deep learning model can approximate the original model accurately, and significantly reduce the computational cost of slope reliability analysis considering the simulation of the random fields.