Data-Driven Approaches for Track Monitoring Using In-Service Trains

This thesis explores data-driven approaches for monitoring rail-infrastructure from the dynamic response of a train in revenue-service. Presently, track inspection is performed either visually or with dedicated track geometry cars. Collecting and analyzing vibration data from in-service trains can offer more economical and more reliable monitoring. The high frequency with which in-service trains travel each section of track means that faults can be detected sooner than with dedicated inspection vehicles, and the large number of passes over each section of track makes a data-driven approach statistically feasible. Developing such a data-driven approach requires modeling the state of the tracks from the collected data, then detecting track anomalies as the model changes over time. Building consistent models from different passes is challenging due to the variation in the train’s speed from pass to pass, the uncertainty in the train’s position, and changes in the properties of the train itself. We study two ways of modeling the state of the tracks to address these challenges: explicit models where the track profile itself is estimated, and implicit models, where features extracted from the collected data are used to imply information about the tracks. In addition, we explore change detection methods appropriate for both modeling approaches; these would allow for monitoring to occur without human supervision. Finally, for network-level monitoring to be practical, we study how data from multiple sensors and multiple trains could be fused together. Data fusion could enable more accurate representations of the state of the tracks, and more rapid detection of track changes after they occur. The track modeling, change detection and data fusion approaches presented in this thesis are validated with simulations and with data collected from two instrumented trains. This collected data includes more than 500 passes through a 40km rail network over a three year period. We demonstrate that the proposed sensing, signal processing, and data analysis can detect numerous types of track anomalies and could facilitate safer, more efficient rail-infrastructure in the future.