A Domain Generalization Network with Discriminant Domain-Common Features for Fault Diagnosis Under Unseen Working Conditions

Transfer learning effectively addresses the issue of distributional mismatches between training and testing data in cross-domain fault diagnosis. However, traditional domain adaptation methods heavily rely on testing data during training, which limit their applicability in real industrial scenarios, particularly when obtaining fault samples from the target domain is challenging. To address this challenge, this paper proposes a domain generalization network with discriminant domain-common features for fault diagnosis under unseen working conditions. The core idea is to extract fault features from multiple source domains using a convolutional encoder, leveraging the local maximum mean discrepancy loss and orthogonal loss to separately capture domain-common and domain-specific features. Simultaneously, the Hilbert-Schmidt independence criterion is employed to reduce redundancy among these features. Furthermore, a convolutional decoder is introduced to ensure the integrity of information across multiple source domains through feature reconstruction. Experimental results demonstrate that our method excels on the Paderborn University bearing dataset and achieves superior results across a range of generalization tasks.