Temporal machine learning for predicting accidents and violations in the mining industry

Abstract This study investigates the use of temporal machine learning models to predict future accidents and violations at individual mines using historical data from the Mine Safety and Health Administration (MSHA). The primary datasets utilized were MSHA’s Accident, Injuries, and Violations records. Mine accidents were categorized by accident classification, and violations were categorized by relevant Part Sections. The data was cleaned and organized by mine type and commodity, then divided into subsets for training, validation, and testing. A range of configurations, including model type, cutoff thresholds, learning rate, number of hidden layers, data processing methods, data division strategy, observation window length, training subset percentage, and activation functions, were evaluated using the underground coal dataset. These configurations were compared based on F1-score, precision, and recall, identifying the best-performing algorithms. The top-performing models were then tested on the most frequently cited accident types and violation categories across different mine types and commodities. Results showed that the temporal models outperformed random chance, particularly for high-frequency incident types, indicating that such models can aid in anticipating future safety events. The underground coal dataset yielded the highest F1-scores, suggesting that model performance may vary depending on dataset characteristics and selection criteria. Predictive performance further improved when combining top-performing F1-score algorithms with those optimized for precision or recall, or by adjusting prediction cutoff thresholds.