Analog circuit fault diagnosis approach using optimized SVMs based on MST algorithm

Abstract

The classification accuracy and efficiency of multiclass SVMs are largely dependent on the SVM combination strategy in analog circuits fault diagnosis. An optimized SVM extension strategy is presented in this paper, which uses minimum spanning tree (MST) algorithm to simplify the SVM structure and decrease the classification errors. By taking the separability measure of fault classes as edge weight of undirected graph extracted from feature space, the tree nodes are generated by bottom-top method, which represents sub-class partition with clustering characteristic. Finally, hierarchical multiclass SVMs are constructed according to the structure of MST obtained. The MST-SVM classifier is expected to improve the diagnosis accuracy because the fault classes with larger margin are preferentially separated. The experimental results on a high-pass filter circuit prove that the MST-SVM method outperforms other conventional SVM approaches in veracity and efficiency of fault diagnosis.