Generative Adversarial Network With Dual Multiscale Feature Fusion for Data Augmentation in Fault Diagnosis

The performance of intelligent fault diagnosis models heavily depends on the amount of monitoring data available. In the situations of monitoring data insufficient for fault diagnosis, generative adversarial networks (GANs) can augment the existing data to supplement data scarcity, which is a promising approach to improve diagnostic accuracy. However, the quality of the generated samples greatly affects the effectiveness of this method. To address this issue, this article proposes a dual multiscale feature fusion (MSFF) GAN to ensure the similarity between generated and real samples and also to improve the diversity of the generated samples. Specifically, a multiscale feature extraction and fusion module is designed to integrate multiscale feature extraction and fusion. A multiscale feature decision fusion module is constructed to avoid the loss of decision-sensitive features in different healthy states. The design of the dual MSFF enhances the learning ability of the generation model and guarantees the similarity between the generated and real samples. A reconstruction network is established to restrain the error of the latent vectors reconstructed by the generated samples, thereby preventing the overfitting of the generated samples and improving their diversity. Experimental results demonstrate that the proposed model has advantages in the similarity, diversity, and effectiveness of the generated samples, significantly improving the performance of intelligent fault diagnosis.