To address the sample imbalance problem in industrial equipment fault diagnosis, an adaptive sampling generative adversarial network ADWGAN-GP (Adaptive Density Wasserstein GAN with Gradient Penalty) model. The model integrates the adaptive sampling module and the improved WGAN-GP structure, enhances the learning ability of features of a few categories through farthest point sampling, local region querying and multi-scale feature extraction, and introduces a density loss function to optimise the distribution uniformity of the generated samples. The experiments are validated with Case Western Reserve University and University of Paderborn bearing datasets under various sample imbalance ratios (10:1, 5:1, 2.5:1). The results show that the model not only effectively balances the number of samples in each category, but also ensures data quality while maintaining sample diversity, resulting in a diagnostic accuracy of 96.21% under the condition of maximum imbalance ratio (10:1). Verified by confusion matrix analysis, the model shows stable classification effects on all fault categories, confirming the significant advantage of the method in balancing the sample distribution.