Application of Machine Learning Methods for Predicting the Factor of Safety in Rock Slopes

Abstract

The mining industry is driven by an increasing worldwide demand for minerals that necessitatesthe continuous improvement of safe and cost-effective excavation operations. Slope stability is a primary concern for every mine to avoid significant safety issues and economic loss. Deterministic and stochastic methods for estimating the Factor of Safety (FOS) have been widely implemented and improved for several decades to overcome the uncertainties and interactions among geotechnical parameters in rock slopes, but these methods are limited in analyzing the nonlinear input/output relationships between a large number of variables and in their application to different geotechnical conditions. Recent progress in machine learning (ML) can serve as an alternative method for forecasting FOS by modeling the relationships between inputs and outputs, providing a more reliable estimation of slope stability. In this research, a literature review of ML applications for slope stability assessment was conducted to gain a comprehensive understanding of the topic. In addition, four ML models—Gaussian Process Regressor (GPR), Support Vector Regressor (SVR), Random Forest (RF), and a hybrid Genetic Algorithm–Multi-Layer Perceptron (GA-MLP)—were implemented and compared on two real-world datasets. The datasets have similar geotechnical parameters from the excavated slopes and use the measured FOS as the output label. The first dataset was collected from a highway excavation site in China, and the second from a mining operation in Peru. The geotechnical parameters included in the datasets are slope height, slope angle, unit weight, cohesion, and friction angle. Each dataset was split into training, validation, and testing sets. The accuracy and consistency of the four ML models on both datasets were subsequently determined. The SVR model showed the best accuracy on the mining dataset, while the GPR model was superior on the highway dataset and demonstrated more consistency across both.