Abstract:

In the field of data-driven mechanical system fault diagnosing, there have been many advances in feature extraction and diagnosis models. But in the whole set of diagnostic processes, the lack of failure signals has always been an inevitable problem in diagnosis. Therefore, combined with the GAN model which was effective in the field of image processing, the probability distribution between the generated signal and the real signal were taken as the selection basis of the partial hyperparameter, and the normal phase signal of the bearing was taken as the input of the model. The bearing simulation signal was used as the target of generation, and the failure signal was generated in combination with the mechanical characteristics of the normal stage and the fault characteristics of the simulation signal. Finally, the proposed signal was proved closer to the true failure signal than the simulated signal by means of the general distribution, envelope spectrum, kurtosis and full-life fitting curve of the margin feature.

Key words: generative adversarial networks, bearing, fault diagnosis, failure signal, generating signal, generating model