Data Augmentation Method for Fault Diagnosis of Mechanical Equipment Based on Improved Wasserstein GAN

Most of the time the mechanical equipment is in normal operation state, which results in high imbalance between fault data and normal data. In addition, traditional signal processing methods rely heavily on expert experience, making it difficult for classification or prediction algorithms to obtain accurate results. In view of the above problem, this paper proposed a method to augment failure data for mechanical equipment diagnosis based on Wasserstein generative adversarial networks with gradient penalty (WGAN-GP). First, the multi-dimensional sensor data are converted into two-dimensional gray images in order to avoid the interference of tedious parameters preset on the model and the dependence on the professional knowledge of signal preprocessing. Based on this foundation, the gray images of the minority sample are used as the input of WGAN-GP to carry out adversarial training until the network reaches the Nash Equilibrium. Then the generated images are added to the original failure samples, reducing the imbalance of the original data samples. Finally, by calculating the structural similarity index between the generated images and the original images, the difficulty of quantitative evaluation of WGAN-GP data generated by itself is solved. Taking the accelerated bearing failure dataset as an example, the classification prediction effects of different classifiers are compared. The results of multiple experiments shown that the proposed method can more effectively improve the prediction accuracy in the case of sparse fault samples.