Study On Simulation-before-test Diagnostic Approach Based On Classifier

Analog circuit fault diagnosis is an attractive research topic which has been regarded as a leading trend in circuit theory for years. It requires that analog circuit run more smoothly and reliably to ensure the rapid development and extensive application of the modernized LSI technology. Thus a further developed method in diagnosing analog circuit fault both theoretically and practically is urgently needed. The present diagnostic method is still a little behind industrial application. That's why the dissertation studies in depth the most practical diagnostic method on analog circuit—CB-SBT (classifier based SBT) which aims at improving diagnostic theory more practically. The dissertation starts with the basic structure and theory of CB-SBT, figuring out the common pattern and features with the purpose of seeking for a new breakthrough in enhancing the ratio of correctly diagnosing and a new way in dealing with the existing problems.Take the fault diagnosis of low-pass filter as an example first, the research is based on CB-SBT diagnosis of the following three classifiers—BP network, RBF network and SVM. The experiments set proper structure and parameter of the classifiers. Meanwhile a comparison of performance is made in the field of correct diagnosis ratio and testing time which results in finding the law and features of CB-SBT diagnostic method in choosing and configuring classifiers.A fault feature extraction algorithm (PSO/k-NN) is put forward to deal with CB-SBT diagnostic method on the base of PSO algorithm and k-NN evaluation criteria. A test is made to check the structure and flow of the above mentioned algorithm and compare its performance with existing PCA, LDA and GA/k-NN with the findings of advantages of PSO/k-NN in extracting fault features.Another achievement lies in the ambiguity sets determination beyond the limit of circuit and ambiguity sets type, which is obtained on the basis of research on the problem of ambiguity sets in CB-SBT diagnostic method. The theory and process of the class difference matrix and class overlap degree matrix in algorithm is presented. Meanwhile, the improved stochastic algorithm of raising the time performance is proposed along with analysis of its performance in accuracy and time is being done. Furthermore, Ambiguity sets determination based on outliers testing method is also realized and examined.The research on the overlapping problems in analog circuit is still one of the important parts in this dissertation. In order to resolve it, a reclassification method is proposed. Take CB-SBT diagnostic method as an example. When the diagnosis is made on a simple resistance circuit, the application of the method can shorten the training process of the classifier and improve the performance of its classifying faults. Furthermore, the improved method is followed to promote the efficiency and reduce the running time.CB-SBT diagnostic method can be used to deal with undefined fault classes. But whether the performance is reliable is in doubt. Nevertheless, the liability is assured by the proposed OCB-SBT (one-class classifier based SBT) framework which can also make it adapt to the overlapping problems. A diagnosis sample is made to show the advantages of OCB-SBT framework over conventional MCB-SBT (multi-class classifier based SBT) and compare the performances of different one-class classifiers.