Anomaly Detection and Forecasting Methods Applied to Point Machine Monitoring Data for Prevention of Switch Failures

Abstract

Railway switches are a crucial asset since they enable trains to change tracks without stopping. Switch failures can compromise a larger part of the railway infrastructure, which can have a negative impact on reputation and revenues. Switches are a costly asset due to frequent inspections, maintenance and renewal of components. Therefore knowing current and future asset condi-tion can be helpful in optimizing switch maintenance to prevent complete failure. The goal of the research presented here is to exploit switch condition monitoring and weather data to identify switch failures on an early stage. Approaches for detection of anomalous switch behavior and prediction of failures are developed. To validate the anomaly detection results obtained by applying the Isolation Forest algorithm, two different annotated data sets are considered. It is found that the anomaly detection approach performs well when applied to a switch, which is characterized by narrow feature distributions within temperature bins. Moreover first results from an Autoregressive Integrated Moving Average model for failure evolution prediction are presented.