Explaining Condition-Based Maintenance for Metro Trains Using Ensemble ML with XAI and Hybrid Deep Learning Models

Metro train systems serve as a backbone for urban transportation but are increasingly vulnerable to operational failures, which can disrupt services and compromise passenger safety. Ensuring their reliability requires robust and accurate methods for failure classification and remaining useful life (RUL) estimation. However, existing approaches often face significant challenges, including imbalanced datasets, high computational costs, reliance on synthetic data, difficulties in processing large time-series datasets, and risks of overfitting.To address these challenges, this research proposes a comprehensive framework of ensemble machine learning methods like Stacking, Random Forest, Bagging, and XGBoost and advanced hybrid deep learning models like TCN-GRU which combines the temporal feature extraction capability of Temporal Convolutional Networks (TCNs) with the sequential modelling efficiency of Gated Recurrent Units (GRUs). Furthermore, resampling techniques like down-sampling for large time series dataset were employed to retain over 90% of the original data variance, significantly reducing computational overhead. Explainable AI (XAI) tools, such as SHAP and LIME, were integrated only into the ensemble methods to enhance interpretability.The proposed framework was evaluated on a real-world metro train dataset, with a focus on the air production unit due to its critical role in train brakes. Results demonstrated that ensemble methods like Stacking performed robustly in failure classification, and the hybrid TCN-GRU model achieved the best results in remaining useful life (RUL) prediction. Also, the results shows that the integration of SHAP offers actionable insights into ensemble models like xgboost, allowing maintenance teams to identify and adjust key variables, thus reducing the chances of future failures.This study underscores the potential of ensemble and hybrid models for advancing predictive maintenance in public transportation. The proposed approach offers significant benefits, including cost reductions, enhanced reliability, optimized performance, and improved safety. Future research will focus on extending these methods to other industrial systems, emphasizing real-time deployment and adaptive learning for broader applicability.