Data augmentation strategy for power inverter fault diagnosis based on wasserstein distance and auxiliary classification generative adversarial network

Abstract

With the rapid development of new energy vehicles, the brushless DC motor (BLDCM) drive system's reliability and safety have attracted extensive attention. The three-phase full-bridge inverter (TFI) of the BLDCM drive system has a high fault occurrence rate under actual working conditions. It is difficult to identify the fault directly, which leads to imbalanced fault datasets. In addition, it is challenging to obtain fault samples directly, which increases the difficulty of fault diagnosis. In response to these problems, a data augmentation method based on Wasserstein distance and auxiliary classification generative adversarial network (WAC-GAN) for TFI fault diagnosis has been proposed. First, based on the Auxiliary Classification Generative Adversarial Network (ACGAN), one-dimensional convolutions are constructed to replace two-dimensional convolutions for the characteristics of a three-phase current signal to improve the extraction efficiency of signal features. Then, the Wasserstein distance is introduced to improve the model's objective function. Based on the principle of the mutual game between the generator and discriminator, the generator can mine the sample distribution characteristics from few fault mode samples and generate numerous fault samples of specific categories to accomplish the purpose of data augmentation. The experimental results show that the fault diagnosis accuracy of the WAC-GAN model under different datasets and different fault modes can achieve satisfactory fault recognition performance. Compared with other data augmentation methods, the effectiveness and superiority of the proposed method has been verified.