A Machine Learning Approach for Track Condition Assessment Through Repeated Historical Data Analytics

Condition monitoring of rail infrastructure is an important task to ensure the safety and ride quality. The increasing travel demands of the rail network due to higher miles traveled requires regular monitoring of the infrastructure and efficient processing of the data for timely decision-making. Despite the regular data collection on different parameters such as acceleration and track geometry, the data processing is commonly performed to document the track performance and maintenance without further knowledge discovery to realize all the potential from historical data. Motivated by the wealth of historical track data in practice, this paper investigates the feasibility of using onboard data that is repeatedly collected over a period of time on a segment of track to potentially identify changes to the track. The proposed approach has been envisioned to learn from repeated historical time-series data to identify both the location and timing of unexpected changes to the track system. To account for stochastic nature of the collected data, associated with the temporal mismatch between the time-series across different inspection runs, we propose a framework by adopting the concept of Matrix Profile without relying on time series synchronization. The approach divides the entire data into smaller track segments, performs extensive similarity search of time-series signatures, and associate locations with higher dissimilarity to changes of the track either due to maintenance or a potential defect. To demonstrate the efficacy and potential of the method, evaluation on both synthetic data and the field geometry data from a revenue-service Class I railroad has been conducted. The findings provide promising results in predicting the location of track changes with a reasonably high degree of certainty, with an automated offline analysis.