Regularized Stable Convolutional Generative Adversarial Network for Imbalanced Fault Diagnosis

The deep learning-based fault diagnosis method faces critical challenges due to the scarcity of fault samples. This article proposes a regularized stable convolutional generative adversarial network (RSCGAN) for data augmentation. Then, an intelligent fault diagnosis strategy is proposed by integrating RSCGAN and 2-D convolutional neural network (2DCNN). Two case studies are implemented on the benchmark bearing data and the incipient agglomeration detection in a chemical fluidized bed reactor. Results exhibit that RSCGAN has obvious contributes to improve the stability of the generative adversarial network (GAN)-based generation model and the fault detection performance even under extremely imbalanced sample conditions. The structural and information similarity between the real data and the generated reaches 0.888 and 0.928, respectively, nearly doubled compared with the existing GAN method. The average accuracy and fault detection rate (FDR) are more than 99%, when the imbalanced ratio continues to increase.