Data-Driven Failure Prediction in Metro Trains: An ML-Based Approach to Predictive Maintenance

Abstract

This study aims to predict metro-train failures using machine learning (ML) models, specifically focusing on XGBoost and Random Forest. Data from the MetroPT dataset were labeled based on time intervals relative to failure: from the start of the dataset to 24 hours before failure, 24 to 2 hours before, 2 hours prior, and the failure event itself. Sensor data, including analog signals (pressure, temperature, current) and digital signals (control, discrete signals), as well as GPS data (latitude, longitude, speed), were utilized to train and evaluate the models. The XGBoost model outperformed Random Forest, achieving high accuracy and precision in failure prediction, particularly within the 12-to-2-hour window before failure. The models' performance was assessed through multi-class ROC curves, reflecting the true and false positive rates across the four time categories. A confusion matrix further illustrated the accuracy by comparing predicted and true labels. Feature importance analysis highlighted key variables like DV-pressure, Reservoirs, oil_compressor, TP3, and TP2, which had high predictive power for failures. This work supports predictive maintenance, enabling timely interventions to extend the remaining useful life (RUL) and conduct root cause analysis (RCA).