Rail anomalies detection: A comparative analysis of three self-supervised models on real data

Monitoring the state of railways infrastructure is crucial for travel safety. These inspections are mostly accomplished by means of dedicated and expensive vehicles, which cause significant disruption to the normal operativity of a line. A recent emerging solution is to equip train vehicles with low cost sensors (mostly accelerometers) able to scan tracks, with the goal to detect anomalies as the train passes. This strategy has shown to be a powerful and effective tool to monitor the conditions of the infrastructure. In this work, we investigate the implications of processing time series collected using these accelerometers on train vehicles axes, by leveraging self-supervised anomaly detection techniques. This is motivated by the high economic value of the approach, with respect to other alternatives. Our work has the advantage of being performed on scarcely available real data annotated with track features, namely weldings positions. Specifically, we explore the performance of three different anomaly detection approaches in identifying the position of rail weldings in an accelerometric signal obtained from a real vehicle. By comparing the performance of three models and considering their features and differences, we provide insight about the anomaly detection problem in a real scenario concerning aspects of the problem that influence model selection, also considering the computational effort needed to train each considered model.