Research Of Machine-Learning-Based Fault Diagnosis Methods Of Analog Circuits

Analog circuits fault diagnosis theory and method is currently a hot researchtopic. The rapid development of modern electronic technology and modular design ofcircuit size and structure proposed a more challenging requirement for analog circuitfault diagnosis. Basically, the problem of analog circuits fault diagnosis is a patternrecognition problem. The two major problems in this study are how to extract thesignal characteristics and how to create a diagnostic machine.The appearance and development of wavelet theory and machine learningalgorithms made the study move into a new level. Currently, it’s becoming a verypopular method to use wavelet analysis to pre-process the test signals and usemachine learning algorithm to do diagnosis. This solution provides a new andeffective way to solve the problem of analog circuit fault diagnosis.Neural network (NN) and support vector machine (SVM) both are effective andpopular methods in machine learning algorithm. The optimization principal of neuralnetwork is based on Empirical Risk Minimization (ERM). NN has a good performancein dealing with large sample case, but its disadvantage is easy to fall into localoptimum. The optimization principal of SVM is based on Structural RiskMinimization（SRM）,which made SVM has a strict mathematical theory basis. Also,SVM has global optimality and better generalization than NN. But at present, SVM ismainly a theoretically algorithm focus on small sample situation.Therefore, based on research about genetic algorithm, immune algorithm, antcolony system, wavelet analysis, NN and SVM, difficulties of NN-based analog faultdiagnosis approach and SVM-based analog fault diagnosis approach are discussed inthis dissertation. Next, some new optimization algorithms of NN and new waveletmethod of analog fault diagnosis are presented. At the last, a technical solutions ofrelevant automatic test and diagnosis system (ATDS) based on DSP，NN and expertsystem is introduced.The main contents and achievements of the paper are as follows:1. Neural-network-based fault diagnosis of analog circuits is proposed. Based onresearch and analysis of NN learning principle and application, back propagation (BP)neural network and radial basis function (RBF) neural network are introduced indetail. Described the application of the two networks in fault diagnosis, and compared the advantages and disadvantages of the two. Also, the applicable condition isdiscussed in detail in this paper.2. A new fault features extraction methods is proposed based on research aboutwavelet theory. The outstanding advantage of wavelet is good time-frequencylocalization. Therefore, we can use wavelet as a preprocessor to do de-noising anddecomposition of test node voltages of analog circuits. Then use the components ofwavelet coefficients to extract fault features. In this paper, a new method based on theroot mean square(RMS) value of the components of wavelet coefficients is proposed.After principal component analysis (PCA) and normalization, the RMS value is usedas the inputs of neural network to identify fault classes. The detailed steps are listedthrough diagnosis examples, and correctness of the methods is also verified.3. A new optimized neural network based fault diagnosis method of analogcircuits is proposed. Aimed to resolved the problem of hard to choose parameters forRBF neural network(RBFNN), a new method based on immune algorithm and antcolony system(ACS-IP) is presented. This method based on the research of immunealgorithm, ant colony system and genetic algorithm, which bring the concept ofimmune factor into ant colony. ACS-IP introduced ‘Antibody concentration’ toimprove global optimization and convergence of RBFNN. This method cansignificantly shorten the running time, at the same time, it can enlarge optimizationsearching space. The specific steps and application results of the new method andtraditional method are discussed in a comparing way in detail in this paper. Andsuperiority of ACS-IP is proved by the comparison. It has less computational work,faster convergence speed and higher diagnostic accuracy.4. A new optimization algorithm is proposed to select Support Vector Machine(SVM) parameters for analog circuit fault diagnosis. The theory and algorithm ofSVM, especially the application in fault diagnosis, is researched in this paper. Withsolid theoretical foundation, SVM has better performance than NN. But SVM hasdefects on dealing with mass data and multi-classification problems. Particularly,there is no developed method to choose parameter of SVM. An improved ant colonysystem is proposed to solve this problem. To get better performance with betterparameters, SVM can be designed as classifiers for fault diagnosis with less trainingand shorter running time. The effectiveness and superiority the proposed method isfurther verified by examples.5. Described the technical solution of Automatic Test and Diagnosis System (ATS). With the theoretical guide of neural network method to do analog circuit faultdiagnosis, the experimental device of DSP-based ATDS is built. This paper introducedthe basic design principles, hardware structure and software realization. The designrules and solution of communication between DSP and PC, DSP-based mainboard,excitement module, analog&digital testing module and printed circuit board (PCB)testing module are all presented in this paper.