A comparative analysis of neural network and ensemble learning models for automated root cause classification in fatal mine accident reports

Mining is one of the highest-risk industries due to the hazardous nature of coal extraction, a primary power generation source. This study uses 1305 mining accident reports, recorded from 1995 to 2015 in India, to determine the root causes of accidents and devise effective safety measures. The methodology follows a structured approach combining expert domain knowledge with advanced natural language processing techniques. We employed a novel hybrid vectorisation strategy combining N-grams, TF-IDF and word embeddings to analyse accident reports. We used stratified 10-fold cross-validation to address the imbalanced data distribution. This study compared three machine learning models: CatBoost, XGBoost and neural networks. CatBoost demonstrated superior performance with a 0.91 F1-score and 0.98 PR-AUC, outperforming both XGBoost (0.88, 0.96) and neural networks (0.84, 0.93). The developed system predicts accident root causes with 91% accuracy, providing a robust framework for improved decision-making and enhanced industry safety standards.