An improved Auxiliary Classifier Generated Adversarial network for Bearing Fault Diagnosis

The generative adversarial network (GAN) and its derivatives has been applied in data generation, especially auxiliary classifier GAN (ACGAN) provides an alternative solution for fault diagnosis under the condition of unbalanced dataset. In this paper, aiming at the problem of training instability and gradients optimization in training ACGAN model, we propose an improved ACGAN to improve data-enhanced capabilities and the classification accuracy of class-imbalanced fault data. In particular, the improved ACGAN produces more convincing generated samples by improving the measurement mode of distribution distances and enhance the speed and stability of training process by splitting the model structure. Finally, the fault dataset for rolling bearings from CWRU are used to confirm the validity of the proposed model. Experimental result shows that the improved ACGAN performs better when comparing to the original ACGAN across several evaluation metrics.