CFNet: Cross-Domain Bearing Fault Diagnosis Under Different Operating Conditions

The methods based on unsupervised domain adaptation are now widely used in the diagnosis of bearing faults. However, the global dependence of features in the feature extraction process is often overlooked. This global dependence is crucial for accurately diagnosing bearing faults, as it can reveal the distribution and variation patterns of fault signals in the overall structure. Therefore, we propose a Centralized Features Network (CFNet) for bearing fault diagnosis. The core of CFNet lies in its Transformer-based feature extractor, which not only captures the local details of fault signals but also preserves the global dependence, thereby achieving comprehensive analysis of fault signals. Furthermore, an Explicit Visual Center Module is proposed to further improve the fusion of long-distance features and local angular regions. Finally, a subdomain adaptation module is also proposed to transform the features of each domain to achieve subclass alignment of the feature space distribution. We have performed experiments on the CWRU and a self-built dataset using the mentioned model. The results demonstrate the effectiveness of our model.