To solve the problem of sample imbalance in fault diagnosis of industrial equipment, an adaptive density wasserstein GAN with gradient penalty (ADWGAN-GP) model is proposed. The model integrates an adaptive sampling module and an improved WGAN-GP structure, enhances the learning ability of minority class features through farthest point sampling, local region query and multi-scale feature extraction, and introduces a density loss function to optimize the distribution uniformity of generated samples. The bearing data set is used in the experiment, and the verification is carried out under various sample imbalance ratios (10:1, 5:1, 2.5:1). The results show that the model can not only effectively balance the number of samples in each category, but also maintain the diversity of samples while ensuring the quality of data, so that the diagnostic accuracy rate reaches 96.21% under the condition of the maximum imbalance ratio (10:1). Through the analysis of confusion matrix, the model shows a stable classification effect in all fault categories, which confirms the significant advantage of the method in balancing the sample distribution.