Deep Transfer Learning With Generalized Distribution Matching Measure for Rotating Machinery Fault Diagnosis

Diagnostic models based on deep transfer learning hold the potential to apply diagnostic knowledge across relevant machinery. However, existing methods suffer from several drawbacks. First, the random initialization of convolutional kernels lacks interpretability and may lead to suboptimal solutions, affecting diagnostic accuracy and training convergence. Second, treating all channels equally limits the ability of the model to capture crucial information. Third, traditional domain alignment methods do not effectively utilize high-order features for matching global or local feature information. To address these issues, a novel domain adaptation method named multiorder statistics matching sparse wavelet convolutional neural network (MSM-SWCNN) is proposed in this article. In MSM-SWCNN, a multichannel feature extraction module is proposed to utilize sparse wavelet convolutional kernels for comprehensive feature extraction from both time and frequency domain. In addition, a novel generalized distribution matching measure is proposed to align features and distribution between two domains. For eight transfer tasks in two datasets of bearings and gears, the diagnostic accuracy can reach more than 96.6%. The results demonstrate that the proposed method outperforms other methods for rotating machinery fault diagnosis.