A Novel Approach for Intelligent Fault Diagnosis in Bearing With Imbalanced Data Based on Cycle-Consistent GAN

The rise of industrial progress has advanced the growth of deep learning (DL)-driven smart fault diagnosis techniques, particularly for condition-based maintenance (CBM). However, the training of these DL methods relies on large dataset, which is unrealistic to collect because fault signal is not practically viable in real case. To address this issue, this article proposes a conditional auxiliary classier cycle-consistent generative adversarial network restrained by Wasserstein distance with gradient penalty (CAC-CycleGAN-WGP). This model can generate superior-quality signals of the minority classes with stability from majority class. In the experimental section, a stacked autoencoder (AE)-based evaluator is proposed to evaluate the quality of these generated sample, and then imbalanced fault diagnosis is conducted at varying balance ratios based on two benchmarked datasets. The outcomes indicate that the proposed approach is adept at generating fault signals, leading to a notable enhancement in fault diagnosis accuracy as the generated samples are added. Additionally, the efficacy of the proposed framework was benchmarked against other commonly employed techniques. Among them, CAC-CycleGAN-WGP stands out with superior performance.