A universal multi-source domain adaptation method with unsupervised clustering for mechanical fault diagnosis under incomplete data

Recently, due to the difficulty of collecting condition data covering all mechanical fault types in industrial scenarios, the fault diagnosis problem under incomplete data is receiving increasing attention where no target prior information can be available. The existing open-set or universal domain adaptation (DA) diagnosis methods typically treat private fault samples in the target as a generalized “unknown” fault class, neglecting their inherent structure. This oversight can lead to confusion in latent feature space representations and difficulties in separating unknown samples. Therefore, a universal DA method with unsupervised clustering is developed to explore the intrinsic structure of the target samples for mechanical fault diagnosis, where multi-source information on different working conditions is considered to transfer complementary knowledge. First, a composite clustering metric combining single-domain and cross-domain evaluation is constructed to recognize shared and unknown health classes on source–target domains. Second, to alleviate the intra-class shift while enlarging the inter-class gap, a class-wise DA algorithm is suggested which operates on the basis of maximum mean discrepancy. Finally, an entropy regularization criterion is utilized to facilitate clustering of different health classes. The efficacy of the presented approach in the fault diagnosis issues when monitoring data is inadequate has been verified through extensive experiments on three rotating machinery datasets.