Multibinary Similarity Filtering-Enabled Multisource Domain Open-Set Fault Diagnosis for Rotating Machinery Under Cross Conditions

Transfer learning have shown great promise for fault diagnosis of rotating machinery. However, in open-set scenarios, where unknown fault classes may exist and the conditions more closely reflect real-world settings. So fault diagnosis becomes significantly more challenging. We propose an end-to-end multi-binary similarity filtering-based multi-source domain transfer learning method (MBSF-MDTL) for cross-condition open-set fault diagnosis. This method aims to overcome two major challenges in current open-set fault diagnosis: poor cross-condition transferability and the lack of effective multi-source evaluation mechanisms. MBSF-MDTL constructs multiple sub-transfer networks, each handling a source-target domain pair. Within each sub-network, adversarial learning is employed for domain alignment, while the multi-binary similarity filtering strategy is used to distinguish known and unknown classes, enabling the training of a dedicated classifier for the target domain. In addition, a fusion decision module (FDM) integrates diagnostic results from multiple sources using various metrics to produce the optimal fault identification outcome. Extensive evaluations, including effectiveness analysis, comparative experiments, comparisons between multi-source and mixed-source domains, and ablation studies, confirm the robustness and effectiveness of the proposed MBSF-MDTL method for open-set transfer learning fault diagnosis. And the average accuracy across 15 transfer tasks is 84.44%, with an average H-score of 80.81%.