Data augmentation on fault diagnosis of wind turbine gearboxes with an enhanced flow-based generative model

Abstract

Deep learning-based fault diagnosis has been rapidly applied in recent years. However, datasets under the normal condition can be easily obtained rather than the faulty datasets. Consequently, such an imbalanced state will be caused between normal and faulty conditions that will lead to poor fault diagnosis. To this end, data augmentation based on an enhanced flow-based generative model (EFBG) is proposed to mitigate the data imbalance problem. Firstly, the Wavelet packet transform is performed on the time series to enhance the feature extraction. Secondly, a flow-based generative model named Glow (Generative flow with invertible 1 × 1 convolutions) is used to augment the faulty datasets. Finally, a random forest is adopted to perform the fault classification. The proposed method was validated in the experiment of the wind turbine gearboxes. Results show that this proposal can increase the fault diagnosis accuracy from 83.26 % to 88.01 %, which demonstrates that this proposal can be set as a data augmentation tool for the fault diagnosis of wind turbine gearboxes.