Conference Agenda

In rail systems, maintenance activities in one subsystem may influence incidents in other subsystem. However, these relationships are rarely analyzed using integrated operational and maintenance data. Besides, existing predictive maintenance studies mainly focus on individual components or sensor-based monitoring, while event-based operational data and maintenance decision support remain underexplored. Thus, this research proposes a machine-learning (ML) based predictive maintenance framework for rail systems using integrated event-based operational and maintenance data. The framework processes include data preprocessing, descriptive analysis, feature engineering and predictive modelling, as well as model performance evaluation. Operational incident and maintenance records from an urban railway signalling system consisting of 98 train sets over the period of 2024–2025 was used as a case study in this study. Incident records, maintenance data, and operational information are firstly integrated. Since maintenance and incident records are largely text-based and subjective depending on maintainer interpretation, preprocessing techniques were applied to transform unstructured records into structured and interpretable datasets suitable for machine-learning based analysis and modelling. Furthermore, integrated operational and maintenance-related datasets provide opportunities to investigate hidden relationships and operational dependencies between railway subsystems. Descriptive analysis is subsequently performed to identify recurring operational incident patterns and relationships between maintenance activities and incident occurrences, especially wheel re-profiling activities and overspeed incidents. Linear Regression and several machine learning techniques including Support Vector Regression (SVR), Random Forest, and Gradient Boosting, were applied to build rail-incident forecasting models. Model performance evaluation results showed that Gradient Boosting achieved the best predictive performance, with a Mean Absolute Error (MAE) of 8.14 days and an R-squared (R²) value of 0.82. Moreover, feature importance analysis also indicated that wheel position was the most influential feature in the predictive model. It significantly supported that maintenance-related and event-based operational data contain meaningful information in forecasting cross-system incidents in rail systems for predictive maintenance.