Framework for imbalanced fault diagnosis of rolling bearing using autoencoding generative adversarial learning

This paper proposes a data augmentation model SAE-WGAN (stacked autoencoder with Wasserstein generative adversarial network), to reduce data imbalance condition of bearing fault diagnosis. To stabilize Wasserstein generative adversarial network training, both Wasserstein distance and informative noise vectors from stacked autoencoder are utilized to improve the quality of generated samples. For the quantitative evaluation of generated samples, both normalized cross-correlation and Kullback-Leibler divergence metrics are employed. Experimental validation and comparisons with state-of-art methods are presented to verify the effectiveness of the generation model. The results show an improvement of 6.58%, and 10.23% compared to Generative adversarial network, and Variational autoencoder, respectively. Furthermore, one-dimensional convolutional neural network is utilized for fault classification, and its performance is assessed using the receiver operating characteristic curve and area under curve values. The comparison results revealed the superior performance of the proposed generation model for bearing intelligent fault diagnosis under paucity of faulty data.