An interpretable frequency-enhanced domain adaptive network for cross-domain fault diagnosis of rotating machinery

The performance of domain adaptive fault diagnosis methods can be limited under noisy environments and variable operating conditions, and existing domain adaptive methods lack interpretability. Therefore, to address the above issues, an interpretable frequency-enhanced domain adaptive network (IFEDAN) for cross-domain fault diagnosis of rotating machinery is proposed in this paper. First, the time-domain signals are converted into the frequency domain using the Fast Fourier Transform (FFT) to enhance the representation of frequency-domain fault features. Additionally, the Morlet wavelet is introduced in the initial layer of the model for weight initialization, which enhances the model’s ability to capture fault features. Then, a frequency-enhanced residual block is designed, which not only helps the model to capture more transferable features, but also further enhances the useful features from both local and global perspectives. Finally, Entropy Maximum Mean Difference (EMMD) loss is designed. EMMD uses the entropy value to determine the bandwidth of the Gaussian kernel, which enhances the stability of the training and decision boundaries. Validation is performed on the public dataset, the roller gear (RG) dataset and the Lanzhou University of Technology (LUT) dataset. The results show that IFEDAN has excellent cross-domain diagnostic performance. When performing cross-domain diagnosis between different datasets, the average diagnosis accuracy of IFEDAN reaches 93.81 %, which is higher than the comparison methods.