Machine learning-based improved method for estimating long-term failure probability with high efficiency of bank slope

Abstract

Affected by water level and rainfall, the failure probability of bank slope is of great time-dependent characteristics. Combining the random field theory with Monte Carlo simulation strategy to calculate long-term failure probability is a time-consuming when considering the spatial variability of geotechnical materials. Therefore, how to efficiently predict the long-term failure probability remains an urgent problem, which is vital to ensure the safety of bank slopes. This study combines remote sensing technology with machine learning methods to conduct comprehensive analysis on the long-term stability and failure probability of bank slope. Firstly, taking an actual bank slope as an example, the failure mechanism and time-dependent characteristics of it is studied with aid of the remote sensing technology. Subsequently, the stability of the bank slope within a year is quantified by the safety factor and seepage field obtained from the numerical model. The failure probability of the bank slope in the same year is calculated by the random field model, and the influence of uncertain parameters on the failure probability is drawn. Three deep learning models, such as multilayer perceptron, convolutional neural networks, and long short-term memory, are adopted to predict the long-term failure probability in 10 years. The results show that the failure probability increases with any of the uncertain parameters, such as the coefficient of variation, correlation coefficient of shear strength, or scale of fluctuation. The utilization of the novel method proposed by this study can efficiently depict the time-dependent characteristic of the long-term failure probability, and it is applicable in predicting the future failure probability of the bank slope. Among three models, the long short-term memory model shows better performance in predicting the time-dependent failure probability. The input data amount and the ratio of the training and test sets have an insignificant effect on the prediction results.