Multi-Class SVMs with Combined Kernel Function and its Applications to Fault Diagnosis of Analog Circuits

Abstract

Fault diagnosis of analog circuits is really important for development and maintenance of safe and reliable electronic circuits and systems. It can be modeled as a pattern recognition problem and addressed by multi-class support vector machines (SVMs). In this paper, one-against-one SVM and directed a cyclic graph SVM are adopted to diagnose the faulty analog circuit. Aiming at the uncertainty of the node arrangement and the error accumulation phenomenon, the improved directed a cyclic graph SVM based on fisher separability measure in high dimensional feature space and margin of SVM is proposed. To further improve the diagnostic accuracy the combined kernel function based on Lévy kernel function and Gaussian kernel function is adopted. Experimental results show the effectiveness of the proposed method.