Rolling bearing fault diagnosis method based on multi-manifold label propagation
Here,aiming at problems of imbalanced source domain data and domain offset between two domains to cause fault recognition rate low when applying unsupervised domain adaptive algorithm in rolling bearing fault diagnosis field,a rolling bearing fault diagnosis method based on multi-manifold label propagation was proposed to project data of both source and target domains onto a common subspace,reduce intra-domain and inter-domain differences,balance sample distribution,and improve the accuracy of bearing fault diagnosis under variable working conditions.Firstly,a locally balanced mapping method inside a domain was proposed to map source domain and target domain data to subspace of a manifold,and obtain sample data aligned inside domain.The source domain data was then balanced to obtain the balanced source domain data.Then,a cross domain-manifold structure refinement alignment method was proposed to further map data to a dual-shared subspace,and obtain the refined and aligned sample data.Finally,by dynamically weighting the pseudo label domain adaptation propagation method,high-precision pseudo labels were obtained.Fault diagnosis tests were conducted on both CWRU and self-built bearing datasets.Test results showed that the proposed method can not only have better recognition ability for multiple fault types,multiple fault sizes and composite faults,but also exhibit excellent diagnosis effect when label samples are scarce.