关键词: 贝叶斯更新, 非饱和土, 流固耦合分析, 机器学习, 变形预测模型

Abstract: Displacement prediction has always been an effective means for conducting the prediction and prevention of landslide disasters. Geotechnical parameters are key input information for landslide deformation calculation, and are rarely considered in the current study. How to conduct the probabilistic prediction of the landslide deformation based on the quantitative characterization of the uncertainties in geotechnical parameters with limited monitoring data is still a prominent difficulty. A case of unsaturated soil slope under rainfall infiltration is investigated and the coupled hydro-mechanical analysis is performed. Based on the spare monitored pore water pressure data, the probabilistic back analysis of geotechnical parameters is efficiently accomplished via the DREAM_zs algorithm. A set of random samples are obtained via the Latin hypercube sampling in accordance with the prior distribution of the geotechnical parameters, and they are utilized to calculate the displacements at the slope toe through the numerical software ABAQUS. Then, a coupled numerical-mechanical displacement prediction model is established through the multiple adaptive regression splines (MARS) and LightGBM algorithm. Based on this model, the displacements at the slope toe are predicted with the stationary posterior samples and the statistical analysis is performed accordingly. It is found that the DREAM_zs algorithm can perform the probabilistic back analysis of geotechnical parameters using limited monitored data with high efficiency and fast convergence. In addition, the proposed displacement prediction model breaks through the limitation of displacement prediction with indirect monitored data such as pore water pressure. And the occurrence probability of slope deformation is also obtained. Furthermore, this study provides a novel idea and attempt for slope deformation prediction.

Key words: Bayesian updating, unsaturated soils, coupled hydro-mechanical analysis, machine learning, displacement prediction model