Few-Shot GAN: Improving the Performance of Intelligent Fault Diagnosis in Severe Data Imbalance

In severe data imbalance scenarios, fault samples are generally scarce, challenging the health management of industrial machinery significantly. Generative adversarial network (GAN), a promising solution to solve the data imbalance problem, suffers from a negative overfitting issue when trained with few samples. To tackle challenges, this article proposes a Few-shot GAN which uses a sample-rich class to provide a sample distribution paradigm for the sample-poor class. More specifically, the GAN is first pre-trained using a sample-rich class. Then, a fine-tuning strategy based on anchor samples is developed, which on the one hand keeps the generated samples close to the real samples and on the other hand preserves the learned complex sample distributions as much as possible. Experiments demonstrate that the overfitting problem of the GAN with few samples trained is well solved and the diversity of the generated samples is improved. In addition, to avoid the offset of features extracted by the fault diagnosis model due to the addition of numerous generated samples in severe data imbalance scenarios, large-margin learning is introduced to constrain the similarities between the features of the generated samples and the real samples. The performance of the fault diagnosis model is significantly improved when numerous generated samples are added, benefiting the predictive maintenance-based decision and avoiding unexpected economic loss.