Universal domain adaptation in rotating machinery fault diagnosis: A self-supervised orthogonal clustering approach

The development of fault diagnosis has been significantly advanced by progress in domain adaptation (DA). Universal Domain Adaptation (UniDA) has garnered considerable attention for its ability to eliminate the assumptions about the target labeling space, effectively handling various scenarios including closed set, partial set, open set, and open-partial set. However, existing UniDA methods often rely heavily on supervised learning within the source domain and fail to adequately explore the intrinsic data structure of the target domain. This limitation hinders the model's ability to recognize unknown faults and reduces its domain adaptation performance. To address this issue, we propose a self-supervised orthogonal clustering network (SSOCN) for UniDA. The core idea of SSOCN fully leverages the structure of the target data to learn discriminative features and achieves adaptive clustering of target domain samples. By using source class centers as clustering points, SSOCN facilitates instance-level feature alignment, enabling the model to effectively address arbitrary category gaps. Furthermore, orthogonal regularization and a known-unknown separation strategy are incorporated to ensure feature orthogonality across different classes and to enhance the recognition of unknown samples, respectively. Extensive experiments across all sub-cases of UniDA demonstrate the effectiveness and superiority of the proposed method.