Condition monitoring for fault diagnosis of railway wheels using recurrence plots and convolutional neural networks (RP-CNN) models

Abstract

RPThe wheel condition monitoring when the train in operation is significant task to prevent the occurrence of unexpected event. In this study, the piezoelectric sensors were installed on the railway track to collect the dynamic voltage-and-strain signals when the train wheels pressed them. These one-dimensional time series signals were transformed to the two-dimensional Recurrence Plots (RP) images as an input data sets for two deep learning models, Xception and EfficientNet-B7. The binary classification, Normal or Faulty as the diagnostical output to indicate the health state of the train wheels in that time. Five metrics were selected to evaluate the performance of two models, namely Accuracy, Precision, Recall, Miss Rate, and AUC. The results show that both models perform the high accuracy of 91.1% to the wheel condition classification. Furthermore, EfficientNet-B7 shows better performance in Recall, Miss-rate, and AUC metrics than those of Xception to express the premium ability in defective wheel identification, which is crucial for this application. Therefore, the efficientNet-B7 is selected as a favorable machine learning classifier for the fault diagnosis of rolling stock wheels. It is significant contribution to train wheel condition monitoring and health management since it provides the effective diagnostic information for maintenance decision to decrease the occurrence of unexpected event.