Spectral normalization CycleGAN for bearing fault transfer diagnosis

The advantage of deep learning without prior feature extraction makes it increasingly used in intelligent fault diagnosis of industrial equipment, but the low robustness and high data requirements of deep learning methods hinder practical applications. In order to adapt to complex and variable working conditions in industrial sites, an SN-lDCycleGAN network based on spectral normalization (SN) and cycle-consistent adversarial networks (CycleGAN) was proposed for fault data transfer generation and diagnosis under variable working conditions. Firstly, a lDCycleGAN network adapted to vibrational data generation was built to obtain the mapping relationship between the source and target domains. The network was improved using spectral normalization to effectively prevent the training instability situation. Secondly, adaptive target data were obtained by changing the source domain to achieve variable work transfer. Finally, the quality of data generation was quantitatively evaluated using three evaluation metrics as well as classifier accuracy, and validated using simulation and experimental signals. Experimental results show that SN-lDCycleGAN has a certain transfer effect on 1D vibration signals, which can enhance the variable working condition data and improve the accuracy of the classifier. The stability and generation quality are better than lDCycleGAN. © 2023 Chinese Vibration Engineering Society. All rights reserved.