Data-Driven Forecasting of Track Geometry Degradation

This paper presents a methodology to monitor the temporal evolution of the track longitudinal level. It is designed to be integrated into a condition monitoring system installed on in-service vehicles, and it relies on the sequential analysis of degradation time series. To estimate the expected growth of a defect, time-weighted linear regressions with different forgetting factors are computed at each iteration. This way, a probability density function of degradation trajectories is generated. Given predefined threshold amplitudes associated to maintenance interventions, the proposed methodology allows to predict the threshold-crossing times. The application to different defects recorded along the monitored railway line shows that the method captures non-linear degradation trends and progressively refines the forecast of the threshold-crossing time when new data are available. When implemented on inservice vehicles, the proposed strategy can provide useful information to support maintenance decisions.