Fault diagnosis of rolling bearing based on cross-domain divergence alignment and intra-domain distribution alienation

When the current transfer learning algorithm is applied to the field of bearing fault diagnosis under different working conditions, it only focuses on reducing the cross-domain distance or the distribution difference within the domain, and does not consider the domain tilt. When the fault samples are scarce, the degradation of recognition ability is more obvious. A fault diagnosis method for rolling bearings based on cross-domain divergence alignment and intra-domain distribution alienation (CDDA-IDDA) is proposed. Firstly, aiming at the cross-domain tilt in the domain data space of variable working conditions, the overall divergence matrix of source domain and target domain is constructed, and the cross-domain divergence alignment is carried out. Then, aiming at the overlapping phenomenon of categories in the domain, based on the distribution adaptation weighted conditional distribution, the spatial distribution of different categories in the same domain is further alienated. Finally, the regularization term is introduced under the framework of structural risk minimization. On the basis of fully retaining the internal structure of the data, a multi-classifier with strong transfer ability is obtained by iteration. Experiments show that the proposed method is better than some mainstream transfer learning algorithms in multi-fault, multi-degree recognition and compound fault diagnosis. In addition, the proposed method still has high diagnostic accuracy when there are few labeled training samples. When the ratio of labeled source domain samples to unlabeled target domain is 1:50 (16 labeled data), the average accuracy of the transfer task reaches 97.78 %.