Improved Generative Adversarial Network for Bearing Fault Diagnosis with Imbalanced Data

In the actual industrial process, rolling bearings usually operate under normal conditions, making it challenging to collect fault signals from sensors. However, the emergence of a fault can lead to serious accidents and thus fault diagnosis for rolling bearings is of much significance. Moreover, the data imbalance issue arises due to the significantly higher amount of normal data compared to faulty data, resulting in reduced accuracy of fault diagnosis models. To solve this problem, an auxiliary classification generative adversarial network based on spectrum normalization and gradient penalty (ACGAN-SG) is proposed. In this method, spectral normalization ensures that the generator stably generates samples of different label types, effectively avoiding mode collapse. Additionally, the gradient penalty encourages the model to generate fake samples that closely resemble real ones, thereby further enhancing the quality of the generated samples. These generated samples are then incorporated into the original dataset for data augmentation. Moreover, this paper converts the one-dimensional original signal into a two-dimensional gray sample, enabling the model to extract more fault features and thereby further improving the quality of the generated samples and the overall efficiency of fault diagnosis. Furthermore, compared to the original fault classifier and other excellent data generation models, ACGAN- SG demonstrates the highest fault diagnostic accuracy by experiments on the bearing dataset, verifying the feasibility and superiority of the proposed model.