High Impedance Ground Faults Diagnosis in High-Speed Railway Traction Networks Based on ReliefF and Self-attention CNN

High impedance ground faults are widely recognized as one of the most frequent failures in high-speed railway traction networks. Due to the complex multi-conductor structure and significant transient disturbances of the traction networks, it is challenging to identify high impedance ground faults. In this paper, a faults diagnosis method based on feature selection and deep learning is proposed to identify high impedance ground faults in traction networks. The wavelet denoising is used to reduce noise from measured data, and the characteristics of the time domain, frequency domain as well as time-frequency domain applicable to the traction network are extracted. As a result, the latent characteristics of the fault signals are successfully identified and explored. An optimized ReliefF algorithm is applied to realize feature selection in order to enhance the robustness of the diagnosis method. A convolutional neural network-based model is established by integrating the self-attention mechanism (i.e., self-attention CNN) to ensure the high accuracy of the model. Test results have demonstrated the noteworthy superiority of the method in high impedance ground faults tasks, in comparison to several existing technologies.