Learning to Predict Component Failures in Trains

Trains of DB Schenker Rail AG create a continuous logfile of diagnostics data. Within the company, methods to use this data in order to increase train availability and reduce costs are researched. An interesting and promising application is the prediction of train component failure. In this paper, we developed and evaluated a method that utilizes the diagnostic data to predict future component failures. To do so, failure codes were aggregated and a flexible labeling scheme is introduced. In an extensive experiment section, three different failure types are examined, a combination of them is evaluated, and different parametrizations are inspected in more detail. The results indicate that a prediction for all of the different types indeed is possible starting from days up to weeks ahead of the failure. However, the level of data-quality and its quantity still have to be increased considerably to yield high quality models.