Generative adversarial network for classification of mechanical fault diagnosis model

The scarcity and imbalance of annotated fault data pose significant challenges to the reliability of intelligent industrial diagnostics. To address this issue, we propose an integrated fault diagnosis framework based on multi-domain feature fusion and generative adversarial networks (GANs). Unlike traditional approaches that treat generation and classification as independent stages, our model unifies these two processes. This method achieves diagnosis by transforming raw vibration signals into multi-domain representations (time domain, frequency domain, and time-frequency domain). The core innovation lies in the restructured generator architecture: a Transformer encoder captures global signal correlations, combined with an Efficient Channel Attention (ECA) mechanism for adaptive recalibration of feature weights, ensuring high-fidelity sample synthesis. Additionally, the model employs a dual-function discriminator that distinguishes genuine from synthetic samples while directly performing multi-class fault classification. Extensive experiments on CWRU and JNU benchmark datasets demonstrate that this approach surpasses existing state-of-the-art algorithms, achieving superior performance in Structural Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR), and diagnostic accuracy. This end-to-end solution effectively mitigates data scarcity challenges in industrial settings.