Anomaly detection for railway switch monitoring data to enable condition-based maintenance

Abstract

Railway switches are a crucial railway asset since they enable train operators to route trains by changing tracks. A switch failure can compromise the availability of a larger part of the infrastructure in cases when train traffic heavily relies on that switch. Nearly one third of the total costs of railway maintenance is spent for switches and crossings. Thus there is a need for increased switch reliability and costs reduction. Nowadays tens of thousands assets around the world are remotely monitored. Switch monitoring data in combination with weather information can be combined to characterize switch functioning and to determine how it is influenced by the weather. Such a characterization allows the detection of anomalies i.e. emerging and sudden failures under different weather conditions. In this paper results from data-based switch-specific models for anomaly detection, which account for temperature influence, are presented. These results are validated against two annotated data sets based on experts' assessment. It is found that the model capabilities strongly depend from switch to switch. A model trained with features (derived from the monitoring data) that are narrowly distributed within small temperature intervals has a very good performance; otherwise the performance is poor. Additionally the influence of rain and humidity on the switch functioning was explored by including data from the closest weather station. No correlation was found probably due to the fact that the available weather information is only a proxy of the local conditions at the switch, stressing the importance of measuring weather parameters at the asset.