A new generative adversarial network based imbalanced fault diagnosis method

In the field of mechanical fault diagnosis, most of the collected signals are normal signals, leading to data imbalance and reduction of fault diagnosis performance. To address the issue, a conditional Wasserstein generative adversarial network with gradient penalty (CWGAN-GP) and the gated recurrent unit recurrent neural network (FDGRU) based method is proposed to improve the classification accuracy. Firstly, the CWGAN-GP is used to generate data. Specially, a Pearson correlation coefficient screening criterion (PCCSC) is proposed to ensure the quality of generated samples. Then, the generated data are added to the original data. Finally, FDGRU is applied for fault recognition. Extensive experiments are conducted, such as the influence of the number of health states, and the imbalance ratio, etc., to prove the effectiveness and stability of the proposed approach. Experimental results illustrate that the proposed approach can significantly enhance the classification accuracy of FDGRU in case of imbalanced data.