A novel approach for rolling bearing fault diagnosis based on generative adversarial feature under limited data conditions

Limited fault data severely constrains the accuracy and reliability of deep learning-based methods, posing a great challenge to the non-destructive testing of rolling bearings. Although using Generative Adversarial Networks (GANs) to augment limited data has been validated as a beneficial way, fault diagnosis is still driven by large amounts of data in most GAN-based methods. To tackle this challenge, an approach based on the generative adversarial feature is proposed, which improves the diagnosis performance by extracting critical features from limited data. First, a generative adversarial feature extraction network (FE-GAN) is developed by improving the bidirectional generative adversarial network (BiGAN) to capture featuresfrom limited data. Second, a conditional encoding strategy is constructed for the generator and discriminator of BiGAN, empowering FE-GAN to precisely differentiate high-dimensional features of different data categories. Finally, a Laplace wavelet convolution technique is designed for the encoder of BiGAN, granting FE-GAN the capability to extract impact information of each category from limited data. The experimental results on two real cases, CWRU and MFPT, show that the proposed FE-GAN combined with CNN achieves 92.26% and 89.68% recognition accuracies under the extreme condition of only 5 samples per class, outperforming most existing advanced methods, validating its effectiveness.