Generative Adversarial Network Based Multi-class Imbalanced Fault Diagnosis of Rolling Bearing

Abstract

Fault diagnosis of rolling bearing plays an important role for the assessment of system reliability. Meanwhile, the number of fault data tend to be much less than the normal data in the real application. This imbalanced problem will greatly reduce the accuracy of most traditional fault diagnosis methods. Especially for the multi-classification problem, some conventional methods can not have good performance on dealing with unbalanced data. In this paper, a method based on generative adversarial network network which generates data for data unbalanced compensation is proposed. This method use designed generator to generate the virtual data which has significant useful features to puzzle the discriminator. Moreover, the virtual data that out-trick the discriminator can be added into the minor dataset. Finally, the classifier based on Convolutional Neurtal Network will dispose the new dataset. In order to verify the effect of this method, experiments based on major methods and proposed method are executed on the CWRU bearing dataset under different loads, which will reduce the correlation of data over time continuity in order to achieve a more realistic fit. Moreover, the proposed method has been compared with several widely applied methods for imbalanced data in fault diagnosis in terms of accuracy. Finally, the comparative results demonstrate that the proposed method has better performance on dealing with the imbalanced problem in fault diagnosis of the rolling bearing than major methods.