Using Event Data to Build Predictive Engine Failure Models

Abstract

Diesel engine failures are one reason for delays and breakdowns on the UK rail network, resulting in significant fines and related financial penalties for a train operating company. Preventing such failures is the ultimate goal, but forecasting or predicting future failures before they occur would be highly desirable. In this study, we take real world Diesel Multiple Unit sensor data, recorded in the form of event data, and repurpose it for the remote condition monitoring of critical diesel engine operations. A methodology based on windowing of data is proposed that demonstrates the effective processing of event data for predictive modelling. This study specifically looks at predicting engine failures, and through this methodology, models trained on the processed data resulted in accuracies of 88%. Explainable AI methods are then utilised to provide feature importance explanations for the model’s performance. This information helps the end user understand specifically which sensor data from the larger dataset is most relevant for predicting engine failures. The work presented is useful to the railway industry, but more specifically to train operator companies who ideally want to foresee failures before they occur to avoid significant financial costs. The methodology proposed is applicable for the predictive maintenance of many systems, not just railway diesel engines.