A Multi-Source Domain Unsupervised Multi-Scale Convolution Attention Model for Bearing Fault Diagnosis

The varying operating conditions of machine tools in different work tasks make it difficult to obtain complete bearing health status data, and the distribution of bearing health status data varies notably across different operating conditions. Although the existing unlabeled target domain bearing fault diagnosis methods can achieve global feature alignment, the prediction on class boundary data samples is prone to misjudgment, which limits the diagnostic accuracy. To overcome this issue, this paper develops a multi-source domain unsupervised multi-scale convolution attention (M-MSAC) model for bearing fault diagnosis tasks under different conditions. This model consists of a feature extractor with shared parameters and two classifiers. By using a feature extractor based on multi-scale convolution attention mechanism, deep feature representations of the source and target domains are extracted, and the maximum mean discrepancy (MMD) embedded is used to quantify the distributions of the source and target domains to achieve global feature alignment. In addition, by combining two classifiers and their prediction results, the class boundary alignment of different bearing fault categories in the source and target domains is optimized to further improve the accuracy and robustness of diagnosis. The experiment was validated based on the CRWU dataset, which showed that the M-MSAC method achieves superior prediction accuracy and stronger generalization ability in predicting the health status of bearings.