Fine-grained transfer learning based on deep feature decomposition for rotating equipment fault diagnosis

The study of transfer learning in rotating equipment fault diagnosis helps overcome the problem of low sample marker data and accelerates the practical application of diagnostic algorithms. Previously reported methods still require numerous fault data samples; however, it is unrealistic to obtain information about the different health states of rotating equipment under all operating conditions. In this paper, a two-stage, fine-grained, fault diagnosis framework is proposed for implementing fault diagnosis across domains of rotating equipment under the condition of no target domain data. Considering that the target domain is completely unknown, the main idea of this paper is to decompose multiple source domain depth features to identify domain-invariant categorical features common under different source domains and classify unknown target domains. More impressively, the problems of data imbalance and low signal-to-noise ratio can be properly solved in our network. Extensive experiments are conducted in two different case studies of rotating devices to validate the proposed method. The experiments show that the method in this paper achieves significant results on both bearing and gearbox health status classification tasks, outperforming other deep transfer learning methods.