Imbalanced Fault Diagnosis Using Conditional Wasserstein Generative Adversarial Networks With Switchable Normalization

Mechanical equipment usually runs under normal condition (NC), making it prohibitively challenging to collect sufficient fault samples and the dataset is prone to imbalanced characteristics, which severely limits the performance of intelligent fault diagnosis methods. In view of this, a conditional Wasserstein generative adversarial network with switchable normalization (SN-CWGAN) is proposed. First, self-attention mechanism and dense convolutional network (DenseNet) are integrated into SN-CWGAN to enhance the transmission of key features, so as to obtain more discriminative feature information. Simultaneously, switchable normalization is performed within discriminators to increase the generalization capability of the SN-CWGAN model. Then, a two time-scale update rule (TTUR) is applied to improve the convergence speed and stability of the model during training. Accordingly, the SN-CWGAN model can generate high-quality fault samples to balance the dataset. Finally, the AlexNet classifier is trained on the balanced dataset to realize fault diagnosis. The effectiveness of the proposed method is validated by two case studies. The diagnostic results and comparative experiments indicate that the proposed method achieves significant improvements in diagnostic accuracy and stability.