Comparative study of unbalanced mining disaster risk level prediction based on artificial intelligence algorithms

Abstract Predicting mining disaster risk levels is a critical component of intelligent mining systems. This study utilizes five common mining disaster datasets to predict various risk levels. By analyzing correlation coefficients and feature importance for each dataset, optimal evaluation indicators are identified. The Shapley Additive Explanations model is then applied to enhance interpretability. To address the presence of outliers and imbalanced data categories, the Mahalanobis Distance Discriminant Method and the Synthetic Minority Oversampling Technique algorithm based on Tomek Links are used for data preprocessing. Subsequently, Support Vector Machine, Random Forest, Extreme Gradient Boosting, one-dimensional Convolutional Neural Networks, and multi-Grained Cascade Forest algorithms are applied to the five mining disaster datasets. Comparative analysis reveals that the Deep Forest algorithm demonstrates superior performance and generalization in predicting stability levels of goaf, slope stability, rockburst intensity levels, pillar stability, and Hanging Wall stability, with prediction accuracies of 92.31%, 96.77%, 92.50%, 91.67%, and 95.00%, respectively. This research provides a systematic solution for mining disaster classification prediction, offering technical support and a scientific theoretical basis for intelligent mining development and mining safety operations.