Multi-source domain adaptation poses more complex challenges compared to traditional single source domain adaptation. While constrained target domain labeling and limited information from a single source can be mitigated, the inherent discrepancies among multiple domains exacerbate the difficulty of fault diagnosis under varying operating conditions, particularly in real industrial systems with diverse and intricate environments. To tackle these issues, a novel Multi-source Inter-domain Feature Discrepancy (MIFD) model is proposed in this paper, which differs from existing multi-source adaptation methods by explicitly modeling inter-domain feature discrepancies instead of solely enforcing a unified shared feature space through global or marginal distribution alignment. In the proposed framework, a three-scale alignment mechanism is introduced to jointly align feature representations, class semantics, and domain distributions, thereby constraining domain shifts at multiple semantic levels while preserving domain-pair-specific characteristics. A discrepancy-aware feature matching module is developed to enable the extraction of reliable and transferable features tailored to specific source–target domain pairs. Furthermore, a class-center and domain alignment strategy is designed to constrain conditional distributions and alleviate pseudo-label bias. In addition, a dual-level weighting scheme is proposed, by which domain contributions are adaptively quantified and irrelevant classes are automatically filtered. Experimental results on two benchmark fault diagnosis scenarios under partial label space settings demonstrate that the proposed MIFD model outperforms state-of-the-art multi-source domain adaptation methods by up to 5.13% on the CWRU dataset and achieves an improvement of 2.16% on the TEP dataset, effectively reducing negative transfer and domain conflicts while enhancing diagnostic robustness under label space inconsistency.
扫码关注我们
求助内容:
应助结果提醒方式:
