<p>Railway Wagon bearing fault diagnosis often encounters the challenge of data imbalance, with scarce fault data making it difficult to train efficient and reliable diagnostic models. To address this, Generative Adversarial Networks (GANs) have been introduced to generate high-quality synthetic fault data, balancing the dataset. However, traditional GANs face training difficulties, resulting in imbalanced dynamics between the discriminator and generator and lower-quality generated data. This paper proposes improvements to Auxiliary Classifier Generative Adversarial Networks (ACGAN) to overcome these challenges. The enhancements include integrating a multi-level sampling attention module into the generator to enhance its capacity, introducing the Wasserstein distance as the loss function to improve stability and sample quality, employing spectral normalization to control gradient magnitude, and implementing a gradient penalty strategy to further enhance network stability. Comparative analyses with other GAN variants demonstrate the improved model's superior stability and effective adversarial dynamics. Evaluation metrics confirm the enhanced model's significantly better generation quality. Fault diagnosis on augmented balanced datasets reveals that the proposed model achieves the best diagnostic performance among various models. Additionally, the effect of adding false data on diagnostic accuracy was investigated, showing a positive correlation between the number of samples and diagnostic accuracy. This research offers an effective solution for addressing data imbalance in railway Wagon bearing fault diagnosis.</p>