A Novel Parallel Multi-Scale Attention Residual Network for the Fault Diagnosis of a Train Transmission System

The data-driven intelligent fault diagnosis method has shown great potential in improving the safety and reliability of train operation. However, the noise interference and multi-scale signal characteristics generated by the train transmission system under non-stationary conditions make it difficult for the network model to effectively learn fault features, resulting in a decrease in the accuracy and robustness of the network. This results in the requirements of train fault diagnosis tasks not being met. Therefore, a novel parallel multi-scale attention residual neural network (PMA-ResNet) for a train transmission system is proposed in this paper. Firstly, multi-scale learning modules (MLMods) with different structures and convolutional kernel sizes are designed by combining a residual neural network (ResNet) and an Inception network, which can automatically learn multi-scale fault information from vibration signals. Secondly, a parallel network structure is constructed to improve the generalization ability of the proposed network model for the entire train transmission system. Finally, by using a self-attention mechanism to assign different weight values to the relative importance of different feature information, the learned fault features are further integrated and enhanced. In the experimental section, a train transmission system fault simulation platform is constructed, and experiments are carried out on train transmission systems with different faults under non-stationary conditions to verify the effectiveness of the proposed network. The experimental results and comparisons with five state-of-the-art methods demonstrate that the proposed PMA-ResNet can diagnose 19 different faults with greater accuracy.