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

Area of Science:

• Geosciences and Mining Engineering
• Artificial Intelligence and Machine Learning

Background:

• Accurate prediction of mining disaster risk is essential for intelligent mining systems.
• Existing methods often struggle with data outliers and class imbalance, impacting prediction accuracy.
• Interpretability of predictive models is vital for trust and practical application in mining safety.

Purpose of the Study:

• To develop and evaluate a robust machine learning framework for predicting multiple mining disaster risk levels.
• To identify optimal evaluation metrics and enhance model interpretability using explainable AI techniques.
• To address data preprocessing challenges like outliers and imbalanced datasets in mining disaster prediction.

Main Methods:

• Utilized five common mining disaster datasets.
• Applied correlation coefficients and feature importance for indicator selection.
• Employed Shapley Additive Explanations (SHAP) for model interpretability.
• Implemented Mahalanobis Distance Discriminant Method and Synthetic Minority Oversampling Technique (Tomek Links) for data preprocessing.
• Compared Support Vector Machine, Random Forest, Extreme Gradient Boosting, 1D Convolutional Neural Networks, and Deep Forest algorithms.

Main Results:

• Deep Forest algorithm exhibited superior performance across all five datasets.
• Achieved high prediction accuracies: 92.31% (goaf stability), 96.77% (slope stability), 92.50% (rockburst intensity), 91.67% (pillar stability), and 95.00% (Hanging Wall stability).
• Data preprocessing techniques effectively handled outliers and imbalanced data, improving model robustness.

Conclusions:

• The Deep Forest algorithm offers a powerful and generalizable solution for mining disaster classification.
• The proposed systematic approach provides technical support and a scientific basis for intelligent mining and safety.
• Enhanced interpretability through SHAP values aids in understanding disaster prediction drivers.