Wheelflat detection and severity classification using deep learning techniques

Wheel flats are one of the most common types of defect found in railway systems. Wheel flats can result in decreasing passenger comfort and noise if they are slight, or serious incidents such as derailment if they are severe. With the increasing demand for railway transport, the speed and weight of rolling stock tend to increase, which results in relatively rapid deterioration. The occurrence of wheel flats is also affected by this increasing demand. To perform preventative maintenance for wheel flats, to keep wheelsets in a proper condition and to minimise maintenance costs, the ability to detect and classify wheel flats is required. This study aims to apply deep learning techniques to detect wheel flats and classify wheel flat severity. The deep learning techniques used in the study are a deep neural network (DNN), a convolutional neural network (CNN) and a recurrent neural network (RNN). 1608 samples, simulated using D-Track, a dynamic behaviour simulation software package, are used to develop machine learning models. Three different aspects of the models are evaluated, namely overall accuracy, the ability to detect wheel flats and the ability to classify wheel flat severity. The results from the study show the DNN has the highest overall accuracy of 96%. In addition, the DNN can be used to detect wheel flats with nearly 100% accuracy. The CNN performs better than the RNN in terms of overall accuracy and wheel flat detection. However, the RNN performs better than the CNN in wheel flat severity classification. Overall, the DNN offers the best approach for detecting wheel flats and classifying their severity.