Abstract: Rock pillars are critical support structures that ensure safe mining operations in underground mines. Accurately predicting rock pillar stability is essential for ensuring safety in underground spaces. Consequently, we propose a novel robust random forest algorithm to predict rock pillar stability. First, we established a dataset comprising 317 hard rock pillars. We select six feature variables as input parameters through recursive feature elimination with cross-validation: pillar width, pillar height, the ratio of pillar width to height, uniaxial compressive strength, average pillar stress, and the ratio of average pillar stress to uniaxial compressive strength. Additionally, we employ chained equation multiple imputation and isolated random forest algorithms to address missing values and outliers. Second, to address decision redundancy and performance loss potentially caused by low-quality decision trees in the random forest algorithm, we introduce a decision tree purification mechanism and a precision weighting strategy. This enhances decision-making efficiency and prediction accuracy, leading to the development of a robust random forest model for predicting hard rock pillar stability. Finally, we evaluate the accuracy, advancement, and reliability of the models through performance evaluation, model comparison, model explanation, and engineering validation. The results indicate that the proposed algorithm achieves satisfactory predictive performance without requiring parameter optimization. The model’s predictive accuracy reaches 88.9%, outperforming other machine learning models, thereby providing effective guidance for evaluating the stability of hard rock pillars in underground mines.

Key words: underground mine, hard rock pillar, stability prediction, robust random forest, machine learning