IEVAEGAN: An Input Enhancement VAEGAN for Rotating Component Fault Diagnosis With Extremely Limited Data

In the field of fault diagnosis for rotating components, obtaining sufficient fault data is highly challenging. Extremely limited fault data pose a significant obstacle in developing accurate fault diagnostic models. To address fault diagnostic scenarios with fewer than 20 samples per category, this article proposes a novel input enhancement variational autoencoding generative adversarial network (IEVAEGAN). By integrating extracted features with random noise, IEVAEGAN ensures both the similarity and diversity of the generated data. Furthermore, the proposed method integrates spectral normalization (SN) and Wasserstein distance to address issues related to mode collapse and gradient vanishing. In addition, it incorporates dropout layers in the network to reduce overfitting risks. Experimental results on an extremely limited dataset demonstrate that the proposed method effectively generates high-quality samples and enhances fault diagnosis performance. Notably, it achieved accuracies of 97.78% and 95.3% on two validation datasets with only four training samples per category, representing a significant improvement over comparative methods. These findings highlight the effectiveness and promise of the proposed method in overcoming the challenges associated with extremely limited data in fault diagnosis.