Railway track circuit fault diagnosis using 1DCNN considering locality dependencies and multiscale feature fusion

Railway track circuit plays a vital part in railway signal system. Accurate detection and identification of faults in track circuits are crucial for the safety of rail transportation. In this paper, a novel framework based on one-dimensional convolutional neural network with multiscale feature fusion is presented for detecting the faults of track circuits via condition monitoring data. Locality dependencies among the monitoring variables are innovatively noticed and explained. Convolution layers in one-dimensional convolutional neural network are used to learn these dependencies. Parallel, serial and dense — three multiscale feature fusion methods are proposed to further enhance the performance of feature learning and diagnosis. Multiscale feature fusion captures more rich diagnosis information at different scales. The experimental results and the comprehensive comparison analysis with respect to the traditional classifiers have demonstrated the superiority of the proposed method. The dense feature fusion produces the most robust and reliable promotion on fault diagnosis performance among the three proposed multiscale feature fusion methods.