Railway Joint Detection Using Deep Convolutional Neural Networks

Abstract

Railway maintenance is crucial to the safety and efficiency of railway operation. Condition monitoring of railway infrastructure has become more and more important that railway companies move to take advantage of artificial intelligence (AI) based technologies. Successful deployment of the technology will enable railway companies to conduct proper predictive maintenances before defects and failures take place so as to improve operation safety and efficiency. This paper presents an end-to-end time series classification approach for the detection of rail joints on railway track using acceleration data by training Convolutional Neural Networks. The advantages of this approach are: 1) working with raw data to reduce the heavy preprocessing of data; and 2) being able to detect joints on left or right rail using one model. Two convolutional networks of ResNet and FCN are investigated and compared. The experimental results show both networks obtain good performance.