Simulation Data-driven Enhanced Unsupervised Domain Adaptation for Bearing Fault Diagnosis

Abstract

The existing unsupervised cross-domain fault diagnosis studies of bearing usually utilize sufficient experimental data collected from test rigs as the source domains, the marginal distribution and conditional distribution alignments between domains are difficult to be considered simultaneously, and all source-domain samples are endowed with the same importance in the process of domain adaptation. Aiming at the above challenges, a new method of simulation data-driven enhanced unsupervised domain adaptation for bearing fault diagnosis is proposed. The bearing fault data with rich fault information and sufficient label data obtained by numerical simulation is used to construct the source domain, thus reducing the dependence on the resources of test rigs. An enhanced loss function embedded with the joint max mean discrepancy is designed to achieve simultaneous alignments of marginal and conditional distributions between different domains in unsupervised scenarios. A weight allocation mechanism for source domain samples is developed to measure the similarity between each individual source domain sample and target domain samples through domain prediction error and to adaptively allocate their weights to suppress negative transfer. Two sets of experimental data collected from test rigs are used as the target domains to validate the effectiveness of the proposed method. The results show that the proposed method can fully adapt the deep feature distributions of simulation domain and experimental domain to improve cross-domain fault diagnosis accuracy in unsupervised scenarios.