Generative Adversarial Networks With Comprehensive Wavelet Feature for Fault Diagnosis of Analog Circuits

Abstract

Analog circuits are one of the most commonly used components in industrial equipment. Circuit failure may lead to significant causalities and even enormous financial losses. To cope with this problem, a novel deep learning strategy based on the cross-wavelet transform (XWT) and generative adversarial networks (GANs) is proposed for analog-circuit fault diagnosis. Different from the traditional methods that simply apply the amplitude characteristic on the time-frequency plane, XWT is applied to obtain the spectral images of the fault signals that can be used to capture the comprehensive features with the involvement of the amplitude, phase, and coherence characteristics. Furthermore, a data-regularization trick is employed on these spectral images to construct three-channel feature images data. The input data can reflect the complex relationship effectively between the measured signals and the feature space, with abundant information of multiple domains on the time-frequency plane. Finally, the different distributions for various fault conditions can be learned by the proposed GAN. The adversarial training manner of the generative network and the discriminative network of the GAN can not only mine the complex nonlinear relationship between the root cause of the failure and the signal signature but also alleviate the problem of overfitting that is caused by the limited availability of the training samples. Furthermore, we expand the GAN to a supervised classifier with the coordinative usage of two auxiliary classifiers, and two convolution neural networks are adopted to establish the basic structure of the proposed model. The comparisons of classification performance with other methods indicate that the proposed approach has good potential for analog-circuit fault diagnosis.