Research And Realization Of Fault Diagnosis System Of Circuit Based On Clustering Combining Svm Dynamic Pruned Binary Tree

Circuit testing and fault diagnosis have been a hot research topic since1960s. With the rapid development of electronic technology and great process of integrated electronic circuits, various theories and methods of components-level fault diagnosis have been to adapt to the needs of large-scale net fault diagnosis, so the fault diagnosis method based on sub-network level in large-scale analog circuits is the urgent requirement to circuits fault diagnosis.The existence of components tolerance greatly restricts the further development of circuit, especially the development of large-scale circuit fault diagnosis, and which becomes popular realm and difficult research problem in circuit fault diagnosis.This paper researches the fault diagnosis approaches at sub-network level in large-scale circuits, focuses on the exploration of the large-scale tolerance circuit diagnosis based on network tearing method and tries to boost the method systematically and practically.In view of the problem whether the torn nodes be accessed or not, this paper researches a modified nodal tearing method. The method introduces pseudo-tearing nodes and makes the network tearing more flexible. Combining with the characteristics of intersected-torn search and multi-level torn search, the paper proposes the testing conditions of tolerant sub-networks by applying interval mathematics and improved genetic algorithm, and which provides a fast and effective method for locating faulty sub-networks in large-scale tolerance circuits.Because circuits are usually with tolerances and the voltage and current of different nodes are sensitive to different fault components, the analysis functions of PSpice are used to collect different fault information, provide data for fault diagnosis.The orthodox wavelet transform (WT) only splits low frequency bands, but wavelet packet transform (WPT) can split low frequency bands and high frequency bands synchronously, wavelet packet transform is used for fault feature extraction. Then corresponding relation between energy of each frequency bands and different fault information is studied, energy features are collected as fault samples features. To traits of different faults of circuit. The simulations results show that wavelet packet transform is effective for feature extraction of fault diagnosis.Support vector machine (SVM) is originally designed for two-class classification, but pattern classification question belongs to multi-classification. Aiming at shorts of several common multi-classification methods based on SVM:one-against-all(OAA), one-against-one (OAO) clustering combine binary-tree support vector machines multi-classification methods is proposed in this paper. Simulation results show that proposed multi-classifiers are feasible to solve fault diagnosis problem, and obtain high classification precision and speed.With MATLAB R2008a platform, fault diagnosis system of analog circuits is developed. Several modules of fault diagnosis system are designed and the clustering combine SVM dynamic pruned binary tree algorithms are realized.A lot of simulation experiments is done, the results show that the proposed multi-class classification algorithm is feasible and effective to improve the diagnosis accuracy and speed.