Integrated Multiple-Defect Detection and Evaluation of Rail Wheel Tread Images using Convolutional Neural Networks

Abstract

The wheel-rail interface is regarded as the most important factor for the dynamic behaviour of a railway vehicle, affecting the safety of the service, the passenger comfort, and the life of the wheelset asset. The degradation of the wheels in contact with the rail is visibly manifest on their treads in the form of defects such as indentations, flats, cavities, etc. To guarantee a reliable rail service and maximise the availability of the rolling-stock assets, these defects need to be constantly and periodically monitored as their severity evolves. This inspection task is usually conducted manually at the fleet level and therefore it takes a lot of human resources. In order to add value to this maintenance activity, this article presents an automatic Deep Learning method to jointly detect and classify wheel tread defects based on smartphone pictures taken by the maintenance team. The architecture of this approach is based on a framework of Convolutional Neural Networks, which is applied to the different tasks of the diagnosis process including the location of the defect area within the image, the prediction of the defect size, and the identification of defect type. With this information determined, the maintenancecriteria rules can ultimately be applied to obtain the actionable results. The presented neural approach has been evaluated with a set of wheel defect pictures collected over the course of nearly two years, concluding that it can reliably automate the condition diagnosis of half the current workload and thus reduce the lead time to take maintenance action, significantly reducing engineering hours for verification and validation. Overall, this creates a platform or significant progress in automated predictive maintenance of rolling stock wheelsets.