A Multi-Gradient Hierarchical Domain Adaptation Network for transfer diagnosis of bearing faults

Domain adaptation is a significant approach to improving the generalizability of diagnosis models of bearing faults in actual engineering scenarios. However, the transfer diagnosis under largely differentiated and loosely distributed cross-domain datasets is still a challenging task. In this work, a Multi-Gradient Hierarchical Domain Adaptation Network (MGHDAN) that can simultaneously learn the transferable domain-invariant and class-discriminative knowledge is proposed to improve the transfer diagnosis performance of bearing faults. First, by aggregating the individual local features into the cell and holistic-level features via the vector of locally aggregated descriptors based network (NetVLAD), the MGHDAN simultaneously aligns the local, cell, and ho-listic features at the global and subdomain level. Then, the intra-class compactness and inter-class separability of cross-domain features are promoted by minimizing the subdomain maximum mean discrepancy (SMMD) be-tween aggregation centers of the NetVLAD and subdomain features. Finally, the contributions of different network-layer to the distribution discrepancy of cross-domain features are adaptively considered by the weighted residual network via the trainable weighting factors, and the attention enhancement of domain-invariant fea-tures is realized. The experiment results show that the proposed MGHDAN method outperforms all the state-of-art global-domain, sub-domain adaptation and hierarchical domain adaptation methods with 27.73%, 7.97% and 36.21% overall average improvement percentages, respectively.

Automated summary

In this work, a Multi-Gradient Hierarchical Domain Adaptation Network (MGHDAN) is proposed to improve the transfer diagnosis performance of bearing faults by simultaneously learning transferable domain-invariant and class-discriminative knowledge.