滚动轴承域适应迁移学习故障识别方法OA

To address challenges in practical engineering where bearing data contains noise and labeled data is scarce,this study proposed a transfer learning method for fault identification using a wavelet packet 1D-CNN and compound domain adaptation loss.The developed wavelet packet branch attention CNN(WPBA-CNN)integrated wavelet packet analysis and attention mechanisms for noise-resistant feature extraction.A branch maximum mean discrepancy(BMMD)loss was designed for multi-scale branches,and combined cross-entropy loss with fast batch nuclear-norm maximization(FBNM)method to form the domain adaptation compound loss(DACL).Experimental results demonstrate that the accuracy of the WPBA-CNN-DACL method increases by 16 percentage points compared to the TICNN method,and the accuracy of the BMMD component increases by 3.3 percentage points compared to the traditional MMD.The average accuracy rate of the 20 migration tasks reaches 98.24%.These experimental results validate the synergistic advantages of our method in noise suppression and cross domain adaptation,and this method can serve as an effective solution for unlabeled bearing fault diagnosis.