A knowledge dynamic matching unit-guided multi-source domain adaptation network with attention mechanism for rolling bearing fault diagnosis

Most current research on multi-source domain adaptation in bearing fault diagnosis focuses on training domain-agnostic networks whose parameters are static. However, it is challenging for static networks to address conflicts across multiple domains when there are domain discrepancies not only between source and target domains, but also between different source domains. Thus, this paper develops a knowledge dynamic matching unit-guided multi-source domain adaptation network with attention mechanism (KDMUMDAN) for bearing fault diagnosis, whose model parameters can dynamically adapt to input samples. KDMUMDAN consists of two modules: a feature extractor with the knowledge dynamic matching unit (KDMU) and two classifiers with attention mechanism. The feature extractor with KDMU is capable of dynamically adjusting the model parameters according to the distribution of input samples to obtain better feature representations, which can effectively facilitate the alignment of source and target domain distributions since it only needs to align the target domain with any part of the set of multi-source domains. Moreover, an attention mechanism is embedded into two classifiers to boost the impact of the more relevant source domain, which can leverage fully the knowledge in multi-source domains to promote data distribution alignment. Experimental results verify that KDMUMDAN has superior bearing fault diagnosis ability across multiple domains.