Data augmentation for CNN-based probabilistic slope stability analysis in spatially variable soils

Abstract

A novel methodology that involves the coupling of Convolutional Neural Networks (CNNs) and a data augmentation technique is proposed for slope reliability calculations. The methodology starts from generating a small set of random field samples, which are then calculated using the shear strength reduction method in the finite-difference scheme to obtain the associated factors of safety. Based on the theoretical relationship between the factor of safety and soil property values derived from the underlying mechanism of the shear strength reduction method, an innovative data augmentation technique is developed to enhance both the quantity and comprehensiveness of the dataset. A CNN model is then trained using the augmented dataset to learn the relationship between the factors of safety and random fields and predict the probability of slope failure. The effectiveness of the methodology is illustrated and validated using a c-φ soil slope and a multi-layered Su slope. The results show that CNNs are effective in interpreting high-dimensional random fields. In addition, through using the data augmentation technique, not only has the predictive capability of the CNN model been effectively boosted, particularly for cases involving low probability levels of failure, but the computational efficiency of the slope reliability analysis has also been improved.