A deep feature alignment adaptation network for rolling bearing intelligent fault diagnosis

Fault diagnostic methods based on deep learning achieve impressive progress recently, but most studies assume that signals from the source domain and target domain share a similar probability distribution. However, the domain shift phenomenon is often unavoidable in practical engineering because of changeable conditions, which hinders the performance of some intelligent methods in fault diagnosis. To tackle the above issue, an unsupervised domain adaptation approach called Deep Feature Alignment Adaptation Network (DFAAN) is proposed in this paper to raise the domain adaptability of fault diagnosis. Firstly, the latent distributions of the two domains are aligned indirectly guided by a Gaussian prior to create a common latent space, which can promote the feature alignment across different domains. Secondly, to better narrow the discrepancy of the feature distribution with the Gaussian prior, a novel discriminative reconstruction distance based on the mechanism of the autoencoder is introduced. Thirdly, an entropy minimum technique is incorporated in the objective function to further increase the transferability of the adaptation method. Diagnostic experiments are conducted on two bearing datasets to illustrate the effectiveness of the proposed approach. The results reveal the superiority of the proposed approach over other typical methods and validate the versatility in multiple diagnostic tasks.