Domain anchor-guided cluster matching for intelligent fault diagnosis under distribution discrepancy and category shift

In recent years, domain adaptation (DA) diagnosis methods that can transfer fault knowledge have been widely used for fault diagnosis under variable working conditions. However, in universal DA diagnosis scenarios, where the fault mode set of the target domain may consist of only a subset of the source domain fault modes or contain novel “unknown” fault types, the traditional DA method is difficult to achieve satisfactory results. Therefore, we propose a domain anchor-guided cluster matching network (DCN) for universal DA fault diagnosis under distribution discrepancy and category shift. The proposed method utilizes a cluster matching algorithm that combines a prototype matching strategy and a clustering metric to mine the hidden structural information of the target domain and identify “unknown” fault samples. Then, the domain anchor contrastive learning is proposed to mitigate distribution discrepancy by minimizing the distance between class-level and instance-level anchor pairs. Finally, the test samples are assigned the same labels as the nearest target domain prototype. The performance of the proposed method is analyzed using two fault diagnosis cases. Extensive experimental results show that DCN outperforms previous state-of-the-art fault diagnosis methods in diagnosis scenarios with variable working conditions and category shift.