Transferable adaptive channel attention module for unsupervised cross-domain fault diagnosis

Domain adaptation methods are widely applied to unsupervised cross-domain fault diagnosis. However, the existing studies always treat the extracted features equally and thus cannot effectively tackle the negative transfer caused by those non-transferable features. Besides, complex actual diagnosis scenarios impose higher generalization performance requirements on traditional domain adaptation models. Given all this, we develop a transferable adaptive channel attention module to enhance the positive transfer and improve models' performance. As a practical plug-and-play component, it can be universally applicable to any one of most domain adaptation models with different network structures. To actively guide domain adaptation, the transferable adaptive channel attention module continuously recalibrates the feature maps based on their transferability during training to improve shared features' domain-invariance and category-discriminability. Moreover, by establishing adaptive selection of feature group size and third-order statistical moment matching strategies, the effectiveness and broad applicability of the proposed module are further improved. Without bells and whistles, the results of two transfer diagnosis cases demonstrate the advantages of the transferable adaptive channel attention module for improving various domain adaptation models' accuracy and generalization performance.