Dual-knowledge collaborative mechanism for multi-source-free domain adaptation in rotating machinery fault diagnosis

Abstract Source-free domain adaptation (SFDA) has gained increasing attention in the field of AI-driven intelligent machinery fault diagnosis, as it addresses scenarios where original source data is unavailable due to privacy constraints. Despite its potential, SFDA within artificial intelligence–based diagnostic frameworks still encounters two key challenges: (1) in real-world industrial applications, multiple working conditions often coexist, leading to complex relationships between target and multiple implicit source domains. (2) In SFDA, where pseudo-labeling is essential to exploit unlabeled target data, most methods depend solely on data structure and overlook model features. To address these issues, this paper proposes a dual-knowledge collaborative mechanism for multi-source SFDA in rotating machinery fault diagnosis. The proposed method first employs entropy maximization with dynamic multi-source fusion to integrate transferable patterns across diverse source domains, enhancing cross-adaption. Then, a dual-knowledge collaboration strategy is designed to jointly leverage source-domain model feature knowledge and target-domain structural cues for confidence-guided pseudo-labeling, which improves label reliability. Furthermore, a relation-consistency module is introduced to enforce semantic alignment and alleviate conflicts between domains. Extensive experiments on two benchmark datasets demonstrate that the proposed approach achieves consistently higher diagnostic accuracy, faster adaptation, and improved robustness against distribution shifts compared with existing SFDA methods.