Feature Vector Selection Method Using Mahalanobis Distance for Diagnostics of Analog Circuits Based on LS-SVM

Abstract

Multi-fault diagnosis for analog circuits based on support vector machine (SVM) usually used a single feature vector to train all binary SVM classifier. In fact, each binary SVM classifier has different classification accuracy for different feature vectors. However, no one has discussed the optimal or near-optimal feature vector selection problem. Based on Mahalanobis distance, a near-optimal feature vector selection method has been proposed for diagnostics of analog circuits using the least squares SVM (LS-SVM). The selection problems of wavelet types, wavelet decomposition level, and normalization methods have been also discussed. Two filters with parametric faults and a nonlinear half-wave rectifier with hard and parametric faults were used as circuits under test (CUTs). The simulation results showed the following: (1) the accuracies using the feature vector with the maximum MD were better than the average accuracies using all the feature vectors, and were better than most accuracies using a single feature vector. But the computation time using the MD method was an order of magnitude larger than that using a single feature vector; (2) Most the diagnostic accuracies using the maximum MD method were near to the optimal accuracies using the exhaustive method while the computation time was reduced about 20–50 % in comparision to the exhaustive method; (3) the Haar wavelet was the best choice among Daubechie’s wavelet family for all CUTs’ diagnosis; (4) only non-normalization, all-normalization, and part-normalization methods are necessary to be considered for feature vector normalization. The proposed method can obtain a near-optimal diagnostic accuracy in a reasonable time, which is beneficial for analog IC or circuits testing and diagnosis.