Abstract In deep hard rock mining operations, pillar stability plays a pivotal role in maintaining excavation safety and maximizing resource recovery. This study presents an ensemble machine learning (EnML) framework that integrates multiple predictive models with interpretable decision mechanisms. A dataset comprising 246 pillar samples was utilized, with pillar width (w), width-to-height ratio (w/h), uniaxial compressive strength (UCS) of surrounding rock, and stress-to-strength ratio (Ps/UCS) selected as input features. Through Friedman rank-based statistical analysis, KNN, RF, SVM, and LGBM were identified as base learners, while a multi-layer perceptron (MLP) served as the meta-learner within a stacked ensemble architecture. To enhance model performance, a Modified Equilibrium Optimizer (m-EO) was employed for hyperparameter tuning. Furthermore, Shapley Additive Explanations (SHAP) and Individual Conditional Expectation (ICE) methods were implemented to provide both global and local interpretability of the classification process. The EnML model achieved accuracy scores of 0.946 and 0.929 on the test set and real-world validation cases, respectively, demonstrating improved classification performance compared to existing predictive methods. Interpretability analysis revealed that Ps/UCS was the most influential predictor, followed by UCS, with geometric features such as w and w/h contributing modestly in boundary cases. Overall, the proposed framework offers a generalizable and interpretable approach for intelligent stability evaluation of hard rock pillars in deep mining environments. Highlights <UnorderedList Mark="Bullet"> <ItemContent> A stacked learning framework optimized by the Modified Equilibrium Optimizer is used to predict deep rock pillar stability. </ItemContent> <ItemContent> Fifteen supervised learning models are statistically compared using Friedman and Nemenyi tests to construct the ensemble model. </ItemContent> <ItemContent> Global and local explainability analysis shows that the stress-to-strength ratio is the dominant factor in failure prediction. </ItemContent> <ItemContent> A two-dimensional decision boundary analysis is introduced to visualize classification behavior across stability zones. </ItemContent> </UnorderedList>

Pillar Stability Assessment in Deep Mining Using a Stacked Ensemble Model with Interpretable Features

期刊论文
发表日期：

2026年1月28日

查看全文

Hongning Qi,
Yingui Qiu,
Jian Zhou

摘要

Abstract

In deep hard rock mining operations, pillar stability plays a pivotal role in maintaining excavation safety and maximizing resource recovery. This study presents an ensemble machine learning (EnML) framework that integrates multiple predictive models with interpretable decision mechanisms. A dataset comprising 246 pillar samples was utilized, with pillar width (w), width-to-height ratio (w/h), uniaxial compressive strength (UCS) of surrounding rock, and stress-to-strength ratio (Ps/UCS) selected as input features. Through Friedman rank-based statistical analysis, KNN, RF, SVM, and LGBM were identified as base learners, while a multi-layer perceptron (MLP) served as the meta-learner within a stacked ensemble architecture. To enhance model performance, a Modified Equilibrium Optimizer (m-EO) was employed for hyperparameter tuning. Furthermore, Shapley Additive Explanations (SHAP) and Individual Conditional Expectation (ICE) methods were implemented to provide both global and local interpretability of the classification process. The EnML model achieved accuracy scores of 0.946 and 0.929 on the test set and real-world validation cases, respectively, demonstrating improved classification performance compared to existing predictive methods. Interpretability analysis revealed that Ps/UCS was the most influential predictor, followed by UCS, with geometric features such as w and w/h contributing modestly in boundary cases. Overall, the proposed framework offers a generalizable and interpretable approach for intelligent stability evaluation of hard rock pillars in deep mining environments.

Highlights

A stacked learning framework optimized by the Modified Equilibrium Optimizer is used to predict deep rock pillar stability.

Fifteen supervised learning models are statistically compared using Friedman and Nemenyi tests to construct the ensemble model.

Global and local explainability analysis shows that the stress-to-strength ratio is the dominant factor in failure prediction.

A two-dimensional decision boundary analysis is introduced to visualize classification behavior across stability zones.