Adaptive variational autoencoding generative adversarial networks for rolling bearing fault diagnosis

The fault diagnosis of rolling bearings with imbalanced data has always been a particularly challenging problem. With data augmentation methods to complement the imbalanced dataset, the effectiveness of diagnosis will be improved significantly. In this paper, adaptive variational autoencoding generative adversarial networks (AVAEGAN) are developed for data augmentation and applied to fault diagnosis. Firstly, a new adaptive network is constructed so that the network adaptively extracts the key features from data to improve the training performance of the network. Secondly, the adaptive loss calculation method is designed to creatively realize the interaction between the loss of the model and the gradient of the function in the network, forming an adaptive balancing mechanism for stable model training. Finally, an adaptive optimal data seeker is proposed so that the model always finds the optimal data in the generated data for augmenting the dataset and enhancing the performance of fault diagnosis. In addition, multi-class comparison experiments are conducted to verify the effectiveness of the method. The results suggest that AVAEGAN outperforms other augmentation methods when used for fault diagnosis.