A conditional variational autoencoding generative adversarial networks with self-modulation for rolling bearing fault diagnosis

Abstract

Rolling bearing fault diagnosis with imbalanced data is a challenging task. It is a significant means to augment the data into balanced datasets. A novel data augmentation method named CVAEGAN-SM is proposed to address this issue in this paper. Firstly, to alleviate the overfitting of generative models due to data scarcity, the input data is preprocessed with a joint translating and scaling, whose hyperparameters are fed by the self-modulation output parameters. Secondly, concerning the conditional generative adversarial network, self-modulation is embedded into the generator, which allows the generator to update itself simultaneously relying on the feedback of input and discriminator. Thirdly, A novel model is constructed integrating the conditional variational autoencoder and conditional Wasserstein generative adversarial network with self-modulation. Furthermore, multi-class comparative experiments are conducted to demonstrate the effectiveness and performance of CVAEGAN-SM. Experimental results indicate that CVAEGAN-SM can effectively augment the imbalanced dataset and outperforms other well-advanced methods.