Detection of Isolated Track Defects Using Axle Box Acceleration from Multibody Simulation and Recurrent Neural Network

Abstract

Railway track inspection is typically conducted cyclically, which is slow, budget-intensive, and time-consuming, as it can only be performed during non-service periods when trains are not in operation. Currently, there are emerging concepts and research studies utilizing railway defect detection through acceleration measurements from axle box mounted on service trains, as this approach offers easier implementation, reduced costs, and time efficiency compared to conventional methods. However, the diagnostic process remains challenging due to the high complexity and large volume of data involved. This research presents a method for applying Recurrent Neural Network (RNN) that works with time-series data in diagnosing railway defects by classifying three specific types of isolated defects: squats and corrugation. The training samples for the neural network were derived from axle box acceleration data obtained through multibody simulation analysis that modeled various railway conditions, speeds, and defect characteristics. A total of 360 samples from the multibody model were used for training, validation, and testing. The study results showed that the Recurrent Neural Network (RNN) achieved an average defects classification accuracy of 91.67%. Furthermore, the study findings revealed that the developed RNN model can accurately predict the location and classify defects, with particularly high accuracy in predicting the location and defects caused by long-pitch corrugation. The study concluded that the developed RNN model can efficiently learn and classify defects from axle box acceleration data obtained from multibody simulation. This approach can be applied as a guideline for railway damage diagnosis in future maintenance applications.