Search

Domain adaptation is one of the mainstream deep transfer learning strategies to deal with unsupervised fault diagnosis issues. Nevertheless, the existing domain adaptation based diagnosis approaches mainly focus on the alignment of source domain and target domain at global level, which may lead to the confusion between subdomains and finally restrict the improvement of target domain accuracy. In this work, the subdomain adaptation capsule network (SACNet) is developed, and the contributions of the proposed method mainly reflected in the following aspects. Firstly, in the proposed architecture, a novel distance metrics named local joint maximum mean discrepancy (LJMMD) is developed and embedded as part of the objective function. Utilizing the LJMMD, the joint distribution of feature and prediction space of target domain can be matched with that of source domain at subdomain level, which can avoid the limitations of global or single marginal distributions match. Secondly, the capsule network with vector-output is substituted for the classical fully connected adaption layer. Instead of matching specific values, the module value of the capsule vector is set as the object of domain adaption. Thus, the discrepancy between different domains can be admitted to a certain extent, which can endow the model with more excellent generalization ability. The experimental results indicate that, compared with other methods, the suggested LJMMD can achieve higher diagnostic accuracy in most transfer tasks. Besides, the diagnostic accuracy of the developed SACNet can be further improved by nearly 2% after adopting the capsule adaptation layer. [ABSTRACT FROM AUTHOR]