Probabilistic slope stability analysis based on the Hermite-logistic regression approach

Abstract

In slope reliability analysis, conventional surrogate model-based analysis methods, such as response surface method, Kriging method, and neural networks method, often rely on the safety factor of slopes for analysis. However, the calculation of safety factors requires repeated iterations using strength reduction, leading to low efficiency in reliability analysis. Addressing this challenge, this manuscript proposes an improved slope reliability analysis method to improve analysis efficiency. This method, which considers the spatial variability of soil parameters, is based on the principles of binary classification concept. It employs the Karhunen-Loève (K-L) expansion to discretize the soil of the slope and generate a random field. By combining Hermite polynomials with logistic regression approach, a surrogate model is established. Using the intrinsic program in FLAC^3D for convergency determination, the stability classification (stable or unstable) for each slope is carried out without reducing the soil strength parameters (using original soil strength parameters). The classification results serve as response values for the Hermite-logistic regression surrogate model, establishing an implicit relationship between random variables and slope stability. The effectiveness of this Hermite-logistic regression method is verified through examples of undrained saturated clay slopes and c-φ soil slopes. The findings indicate that the Hermite-logistic regression model demonstrates remarkable computational efficiency when compared to conventional random finite element calculations, all while maintaining high computational accuracy. Specifically, the proposed method reduces the computational cost by at least a factor of ten while ensuring the attainment of precise results. In addition, a sensitivity analysis is performed to investigate the influence of slope geometric parameters and spatial variability parameters on slope stability and reliability.