Study On Fault Diagnosis In Analog Circuits Based On Support Vector Machine

Fault diagnosis theory and methods for analog circuits is still an extremelychallenging research topic in the circuit test field around the world.The rapiddevelopment of semiconductor technology has promoted the wide applications ofanalog integrated circuits and analog/digital mixed-signal circuits.In order to shortenthe time-to-market of the electronic product and increase its reliability,a new theory andfault diagnosis method needs to be established to meet the new requirements of analogcircuit test and fault diagnosis.However,due to the inherent characteristics of analogcircuits,such as its nonlinearity and continuous response,tolerances on componentparameters,etc,as well as the diversity and complexity of their faults,it is difficult forthe conventional or traditional fault diagnosis theories and methods to achieve theexpected results in the practical engineering.Hence,it is very important to explore somehighly efficient fault diagnosis theories and methods to meet the development of analogcircuits.Support vector machine (SVM) based on statistical learning theory,which is aresearch focus both at home and abroad,has provided an effective solution for faultdiagnosis in analog circuits.On the basis of the modern test technology,signalprocessing,system identification and testability analysis,etc,the fault diagnosis methodfor analog circuits based on SVM is deeply researched in the dissertation.The mainresearch contents and achievements are summarized as follows:According to the characteristics of analog circuit faults and SVM with simplestructure,the global optimum solution,good generalization to solve pattern recognitionissues with a few samples,nonlinearity and high dimensionality,the fault diagnosismethod for analog circuit based on SVM is presented.At the same time,the faultdiagnosis system of analog circuit based on SVM and a SVM model of fault diagnosisfor the circuits under test (CUT) is established.The factors which influence theperformance of SVM fault diagnosis model,such as kernel function and kernelparameters,punishment parameters and the combination algorithm of multiclass SVM,are researched.The feature extraction method based on efficient points of response curve is studied in time domain and frequency domain.In order to reduce the faults' featuredimensions,the fault feature selection method based on criteria of Max-Relevance andMin-Redundancy (MRMR) and SVM is presented,and the selection mechanism of theoptimal feature is established.As a result,the complexity of the SVM fault diagnosismodel is effectively reduced.Due to analog circuit fault response included the non-stationary or time-varyinginformation,the fault feature extraction method based on wavelet analysis is studied,and the feature measure of CUT is defined as the RMS value of difference betweennormal and faulty features.Meanwhile,the selection principle of optimal motherwavelet based on feature measurement is presented so as to achieve the faults featureextraction in analog circuits based on adaptive wavelet analysis.As a result of the time-variability of analog circuits and the irregularity of actualmeasured signals,the measured signals have fractal properties in certain scale rangesand the fault feature extraction method of analog circuits based on multfractal analysisis presented.In addition,the procedure for estimating multifractal singularity spectrabased on the wavelet's maximum modulus is researched and 6 parameters related tomultifractal singularity spectra are extracted as faults' features.The response obtained from the limited test nodes is not enough to reflect the stateof each component in CUT,and there are some ambiguity groups in CUT.The study ofapproaches to selecting the diagnosable component set is very helpful for diagnosingthe potential fault in CUT.Firstly,the testability analysis based on the circuit topologyis studied,the potential fault components are grouped by the testable and ambiguitygroup which is attained by use of linear correlation of the testability matrix of CUT,andthe selection of diagnosable components set is realized.Secondly,according to therelationship between transfer function Pole-zero and parameters of components,theapproach to selecting the diagnosable component set is presented based on thePole-Zero sensitivity analysis of CUT.The diagnosable component set can beascertained by the relationship between the Pole and Zero ambiguity groups.Finally,bystudying the fuzzy cluster of the response obtained from accessible node,the approachto selecting the diagnosable component set is presented based on fuzzy cluster analysisand fuzzy cluster validity index theory.Based on the theoretical analysis,targeted at the analog/digital signal benchmark circuits and the circuits widely studied in the references,and simulated the faultscenarios as well as the fault-free operation of CUT by OrCAD10.5,analog circuit faultdiagnosis model based on support vector machine is achieved in Matlab7.0 and asimulating experiment is also carried out to diagnose the soft faults,hard faults andmulti-faults in CUT.The experimental results have testified the feasibility of theproposed method and the validity of the conclusion in this dissertation.Compared withthe widely studied analog circuits fault diagnosis method based on neural network,theproposed method in this dissertation possesses the merits of simpler structure and higherdiagnostic accuracy,and it solves the disadvantage of neural network--local optimalsolution,over-fitting and under-fitting,difficulty in choosing model structures,etc,andrealizes the automatic choice of model structure and good generalization ability forsmall samples to account for the trade off between learning ability and generalizationability of learning machine by minimizing structural risk.