Mixed-attention variational autoencoding generative adversarial networks for rolling bearing fault diagnosis

The problem of unbalanced samples has always been an obstacle for the application of intelligent methods in the fault diagnosis of rolling bearings. The method of data augmentation provides a new idea for solving this problem from another perspective. More accurate diagnosis can be achieved by expanding the unbalanced data set. A variational autoencoding generative adversarial network integrating the mixed-attention mechanism is proposed in this research. Firstly, a mixed-attention mechanism is integrated in the basic network to construct a novel self-adaptive network, which promotes the model to selectively focus on the channel and spatial information contained in the potential variables. Meanwhile, in view of the characteristics of small and unclear impact intervals of high-speed vibration signals, the attention mechanism is set to multi-head. Secondly, a novel multi-domain loss mechanism is proposed to enhance the model's sensitivity to discriminative fault signatures in both spectral and envelope feature spaces, achieving the extraction of fault characteristics and successfully completing the generation of diverse data. Finally, a data screening method is designed. The optimized data is used for dataset expansion, which improved the diagnostic performance, and the effectiveness of this approach has better performance against other methods validated through two experiments.