CE-FFGAN: A feature fusion generative adversarial network with deep embedded category information for limited sample fault diagnosis of rotating machinery under speed variation

In engineering practice, limited samples and variable speed constrain the accuracy of intelligent diagnosis of rotating machinery. Generative adversarial network (GAN) offers a solution by augmenting data, yet challenges like training instability and mode collapse persist. This article proposes a feature fusion GAN with deep embedded category information (CE-FFGAN) for high-quality time-frequency image generation. An improved feature fusion generator (Encoder-Decoder) structure is designed. It employs multiple skip-connections to capture multi-scale features. The category information matrix is directly embedded in each network layer of decoder, bolstering the category constraints on the generator's output. Subsequently, a novel loss function is formulated by incorporating image reconstruction loss to enhances image fidelity, while spectral normalization in the discriminator prevents gradient vanishing. Dilution experiments were conducted on fault datasets with speed variations. The results indicate that CE-FFGAN is superior to comparison methods and significantly improves the diagnostic accuracy under limited samples.