Complete knowledge weighted adversarial network and class alignment learning guided open set multi-source domain adaptive fault diagnosis

Abstract Open-set domain adaptation in fault diagnosis has gained significant research interest due to its critical implications for practical engineering. However, the current work insufficiently considers inter-domain sample similarity during the transportability evaluation process, resulting in models that are overconfident in unseen samples. In addition, separability of fault features to the target domain is ignored. To address these limitations, this paper proposes a complete knowledge weighted adversarial network and class alignment learning guided open set fault diagnosis model (CKWAN-CAL). First, complete knowledge weights are established to evaluate the transferability the samples through confidence discrepancy weight mechanism and predictive entropy. The CKWAN is then designed to selectively align the feature distributions of the samples between domains, limiting the interference of unseen samples in the transfer process. Meanwhile, an unknown fault identification network is constructed to establish the clear decision boundary between shared and unknown faults. In addition, pseudo-label guided CAL is introduced with the aim of improving the proximity of identical fault features and the separability between fault features. Through comprehensive experimentation on both public and self-made datasets, incorporating systematic comparative analyses and ablation, proposed method demonstrates statistically significant performance advantages over state-of-the-art methods.