Neural Network Method for Fault Diagnosis of Analog Circuit Based on Kurtosis and Skewness

This paper proposes the method of analog circuit fault diagnosis based on high-order cumulants combined with Information Fusion. It is to extract the original voltage and current signals from output terminal of the circuit under test, to determine their kurtosis and skewness as fault eigenvectors, and to import them into improved BP neural network for fault diagnosis. As for construction of fault eigenvectors, high-order cumulants technique, compared to Principal Component Analysis (PCA) which is based on second order statistics, pays more attention to information neglected by PCA. After Information Fusion is employed to integrate voltage with current as fault eigenvectors, it makes eigenvectors show relatively comprehensive fault information. Diagnosis examples further verify that fault eigenvectors gained in this way have higher recognition rate and diagnosis accuracy.