Dual-adaptive unknown faults separation network: An open-set domain adaptation method for industrial fault diagnosis

Abstract

Deep domain adaptation methods have gained significant attention in few-shot intelligent industrial fault diagnosis. However, most existing approaches often struggle when confronted with unknown faults in target domains, posing operational risks. To tackle this problem, an open-set domain adaptation method is presented to improve fault diagnosis accuracy and robustness. The proposed Dual-Adaptive Unknown Faults Separation Network (DAUFSN) approach integrates a shared feature extractor with two specialized networks: a closed-set adversarial unknown separation network and an open-set adversarial domain adaptation network. The closed-set network generates similarity scores to separate known and unknown faults, while the open-set network introduces an extra class, especially for unknown faults, leveraging the similarity scores in the loss functions to separate the unknown faults. This allows the system to both accurately classify known faults and effectively identify unknown ones, addressing the critical challenge of unknown fault identification. Experimental validation using datasets from rolling bearing and multiphase flow processes demonstrates the efficacy and superiority of the DAUFSN method.