A novel compound fault diagnosis method for rolling bearings based on graph label manifold metric transfer

Metric transfer learning overcomes the problem of no domain distribution alienation in traditional sample transfer, and also improves the classification accuracy. However, some problems exist; for example, the target domain must have labelled samples without considering the manifold connection between data samples, and the algorithm is not robust under sparse samples. A compound fault diagnosis method for rolling bearings based on graph label manifold metric transfer is proposed. First, the pseudo-label of the target domain is acquired through a weighted graph constructed on the source domain with known fault labels. In order to ensure the accuracy of metrics in manifold space, we adopt the method of graph label propagation (LP) to gradually spread the label information. Then, the sample data are mapped to the low-dimensional manifold space by locality preserving projections (LPP), which gives higher weight to the source domain samples close to the target domain to solve the domain offset problem. At the same time, combined with the labelled samples of the target domain in LP, the manifold distance measure is learned to minimize the intra-class distance and maximize the class spacing, eliminate the overlap phenomenon in the domain and realize the distribution alienation between classes in the domain. After each iteration, unsupervised manifold metric transfer learning is realized by increasing the range of LP and gradually reducing the error of graph LP. Experiments show that the new unsupervised transfer learning method has better fault identification ability than the semi-supervised transfer learning method. In the case of sparse training samples, it still maintains a high diagnostic accuracy, and the highest fault recognition accuracy can reach more than 97% with good robustness.