A Robust Wavelet-integrated Residual Network for Fault Diagnosis of Machines with Adversarial Training

In engineering applications, the performance of intelligent fault diagnosis models can be significantly impacted by noise interference in the signals. This paper aims to address this issue by analyzing the influence of noise interference on diagnosis models, focusing on the network structure and training methods. Based on the analysis findings, a wavelet-integrated residual network (WResNet) is proposed to improve the noise-robustness. WResNet integrates discrete wavelet transformation (DWT) into the residual network architecture to mitigate potential problems related to frequency aliasing caused by traditional down-sampling operations. By incorporating DWT, WResNet could reduce the impact of noise interference. In addition, a gradient-based adversarial training method is adopted for optimizing the loss function of WResNet. By minimizing the maximal risk for label-preserving fluctuations of input signals, adversarial training is able to enhance the stability of WResNet. The effectiveness of WResNet is validated by using the monitoring data from a motor with different signal-noise-ratio. The results show that compared with ResNet and the method that using wavelet transformation as a pre-processing step, WResNet is able to achieve higher diagnosis accuracy while owning better noise-robustness.