Imbalanced Samples Fault Detection Using GANCNN for Power Inverters

Abstract

Aiming at the problem of incomplete fault types existing in power switches fault detection for three phase inverters, a novel diagnosis method based on generative adversarial network (GAN) and convolutional neural network (CNN) is proposed. Firstly, the phase current is used as the fault-sensitive signal, and the fast Fourier transform (FFT) is performed to obtain the frequency domain features, and the normalization preprocessing is performed. Then, the GAN model is used for confrontation training to generate virtual samples by few real sample characteristics, in order to get balanced samples with different fault modes. Finally, convolutional neural network model is built to complete the power inverter fault diagnosis. The experimental results show that GAN-CNN can effectively improve the diagnosis accuracy and stability in the case of sample imbalance.