Multi-mode data augmentation and fault diagnosis of gearbox using improved ACGAN

To address the problems that the existing generative adversarial network (GAN) was difficult to efficiently generate multi-mode fault samples, and its training was unstable, an improved auxiliary classification GAN (ACGAN) was proposed for multi-mode data augmentation and intelligent fault diagnosis of the gearbox to ensure the safe operation of the vehicle. The independent classifier was introduced to construct a new ACGAN framework, and the compatibility between the classification accuracy and discriminant accuracy of classic ACGAN was improved. Wasserstein distances were used to define new adversarial loss functions with smooth properties, so as to overcome the disadvantages of the GAN, such as mode collapse and gradient vanishing. In order to improve the stability during the adversarial training process, the spectral normalization was used to replace the weight clipping to constrain the weight parameters of the discriminator. In order to verify the effectiveness and superiority of the improved ACGAN method, experimental analysis was performed on gearbox samples under six health conditions. Analysis results show that the fault samples generated by the improved ACGAN achieve better quality evaluation results at the data level and feature level, among which the evaluation index based on the structural similarity outperforms the comparison method by 0. 249 3 on average, and the evaluation index based on the maximum mean difference outperforms the comparison method by 0. 696 6 on average. The training process for the improved ACGAN is more stable. Its loss function has better convergence, and it has higher efficiency in multi-mode fault diagnosis scenarios with its training time reduced to 20% of the comparison method. In the case of missing fault samples, the generated fault samples of the improved ACGAN can effectively assist the training of intelligent fault diagnosis models based on deep learning, which can improve the diagnosis accuracy from 75.34% to 97.06%. © 2023 Chang'an University. All rights reserved.