Improved adversarial learning for fault feature generation of wind turbine gearbox

Sufficient samples collected from different conditions can effectively improve fault diagnosis performance for wind turbine gearboxes. However, it is very difficult and time-consuming to obtain enough data under fault conditions rather than the normal condition for real wind turbines. For this reason, an improved adversarial learning is proposed to generate fault features for the fault diagnosis of wind turbine gearbox with unbalanced fault classes. In the present method, wavelet package transform is first performed on raw data for producing feature space as the input of a generative adversarial network (GAN). To improve adversarial learning capability, a Wasserstein distance with gradient punishment is proposed to guide the fault feature generation of the conditional GAN. The addressed approach was validated using fault diagnosis experiments on the gearbox of an industrial wind turbine. In the experiments, the present method has the best performance compared to peer methods, due to contributions of the improved adversarial learning and the feature space generation. The results show that the present method is capable of dealing with the imbalance samples by generating fault features for the wind turbine gearbox.