A Novel Fault Diagnosis Method for Analog Circuits Based on Conditional Variational Neural Networks

Abstract

To enhance the reliability of analog circuits in complex electrical systems, a novel fault diagnosis method based on conditional variational neural networks (CVNN) is presented in this paper. The CVNN model is constructed by adding a sampling layer to the multi-layer perceptron. The latent variable which has the same distribution with the original signal of the analog circuit is obtained in the sampling layer, where the noise is introduced to improve the generalization performance of the model. The output features of the sampling layer are achieved by resampling on the latent variable, and the variational inference is adopted to estimate unknown parameters of the model. To address the overfitting issue of the CVNN model, the Dropout algorithm and the scaled exponential linear unit function are applied to the hidden layers. Furthermore, the features compressed by the second hidden layer are input into the Softmax classifier for training, and then the trained fault diagnosis model is utilized to identify the fault classes of the analog circuit. The method is fully evaluated with the three typical analog circuits, and the experimental results demonstrate that the fault diagnosis method based on CVNN can achieve better diagnosis accuracy and generalization performance than other typical fault diagnosis methods for analog circuits.