Slope stability analysis using a surrogate model with varying sampling precision: a case study of open-pit mine dump slopes

In open-pit mining operations, ensuring slope stability is critical for maintaining safety and efficiency. However, the limited availability of samples presents significant challenges for comprehensive stability analyses. This study proposes a novel small-sample agent modelling method that combines GoogLeNet with finite difference techniques to evaluate slope stability. The methodology utilises the shear strength discounting approach within a finite difference framework to compute the factor of safety using random field samples. Subsequently, datasets of varying sample sizes (400, 800, and 1200) are employed to train the convolutional neural network (CNN) and the GoogLeNet model, enabling the prediction of slope instability based on factor of safety derived from the finite difference calculations. Validation was conducted using real-world slope stability data from open-pit mine discharge sites. The findings indicates that this machine-learning-based image recognition approach efficiently assesses slope stability, even with limited samples. Notably, GoogLeNet demonstrated superior accuracy and robustness compared to traditional CNN models. Additionally, the analysis revealed that the coefficient of variation significantly influences the stability of earth discharge sites. These findings provide a practical reference for fast slope stability assessments in mining operations.