A novel bearing imbalance Fault-diagnosis method based on a Wasserstein conditional generative adversarial network

In practical industrial applications, rolling bearings are in normal operation most of the time, and the inadequacy of the fault data makes the data itself exhibit unbalanced characteristics. The imbalance of data is difficult to meet the training demand of intelligent networks, which in turn causes its low recognition ability. To solve this problem, this paper proposes a new unbalanced fault diagnosis framework, the Wasserstein conditional generation adversarial network, based on hierarchical feature matching. Unlike traditional generative frameworks, which achieve the overall alignment between the generated distribution and the true distribution by adversarial only. The proposed framework unites Wasserstein loss and hierarchical feature matching loss, and constrain the data generation characteristics from the perspective of global and class-specific to improve the validity of the data. Experiments containing real bearing faults demonstrate the generalization performance of the proposed method, and the results show that the proposed method requires only a small number of 10 samples to reach 92% correct diagnosis, which provides a feasible tool for solving the current industrial data imbalance problem.