A Preliminary Study on 2D Convolutional Neural Network-Based Discontinuous Rail Position Classification for Detection on Rail Breaks Using Distributed Acoustic Sensing Data

Abstract

Rail breaks, which are crucial maintenance issues for the railways, require immediate inspection and maintenance as it can cause severe railway accidents. Though, it is still difficult to promptly detect rail breaks that occur during train operation even with the advances in maintenance and inspection technologies. In this research, as a preliminary study on rail break detection system, a deep learning-based discontinuous rail position classification method, which is using vibration data obtained from distributed acoustic sensing (DAS) system during train operation, is proposed. To analyze the vibration data, a preprocessing algorithm for determining train occupancy is applied first. After that, the data in the space–time domain occupied by the train is converted to the spectrogram which is in the frequency domain by using short-time Fourier transformation (STFT). In the third step, the spectrogram images are applied to the proposed 2D convolutional neural network (2D CNN) model and the network detects discontinuous rail positions along the track, which are geometrically distinct from continuous welded rails, such as rail breaks. In order to evaluate the superiority of the proposed network model, performance comparison tests with other existing models were conducted with data collected from an actual railway line. From the results, the proposed model could achieve 99.17%, 93.33%, 87.5%, 90.32% for accuracy, precision, recall, and F1 score, respectively, and the results show overwhelming detection performance compared to other models.