Predictive Analytics for Rail Infrastructure Maintenance Using Multi Modal Data

Sammanfattning

Failures in critical railway switches severely impact network operations, necessitating advanced maintenance strategies. This thesis presents a data-driven framework leveraging Point Operating System (POSS) current consumption data from two Stockholm metro lines to automate fault classification and explore early degradation indicators for predictive maintenance. Methodology involved POSS data preprocessing, feature extraction, Dynamic Time Warping (DTW) for anomaly detection, and development of XGBoost fault classifiers trained on labels K-Means clustering output and expert fault catalogues. The classifier achieved 91% balanced accuracy emphasizing power and movement/lock phase features. While long lead-time degradation trends were not consistently found in raw features, increased feature variability often preceded corrective maintenance. The classifier probability trends were generally reactive to developed faults. Challenges with free-text maintenance logs highlighted the need for improved data standardization. Study limitations include the use of POSS data over two years with no additional sources of data about the physical attributes of the switch. This research contributes a validated framework for automated switch diagnostics and provides insights into degradation indicators, supporting a shift towards more data-informed, proactive railway maintenance to enhance network reliability.