Applications Of Subsoace Learning Algorithms In Fault Diagnosis Of Analog Circuits

Analog circuit fault diagnosis is a research topic related to national economy and people's livelihood.Data show that more than 60%of integrated circuits include digital and analog mixed-signal circuits,while more than 80%of faults in hybrid electronic systems occur in the analog part.The test cost of analog circuits can be accounting for 95%of the total test cost.However,it is really hard to diagnose and locate faulty components and parameters due to some inherent difficulties in analog circuits.Although fault diagnosis theory for analog circuits has achieved a lot in the past few decades,breakthroughs have not been made.This dissertation focuses on the key technology of analog circuit fault diagnosis,namely feature extraction methods.The research content is to improve the accuracy of fault diagnosis based on subspace learning algorithms and the method to resist noise for these algorithms is also involved.The specific work is as follows:(1)For the kernel entropy component analysis which is a nonlinear subspace learning algorithm,a crucial step is the estimation of Renyi entropy.It is optimized from two perspectives,namely the full bandwidth matrix kernel probability density function and the higher order kernel probability density function.Two more reasonable and accurate estimations of Renyi entropy are given.The dissertation proves the conclusions theoretically.The experiments show that the improved estimation formulas are more accurate in estimating the entropy.In addition,when the kernel entropy component analysis mainly improved by the entropy estimation is used to extract nonlinear features of analog circuit fault signals,the accuracy of diagnosis is increased.(2)For several manifold algorithms,the dissertation introduces a robust distance to measure the distance between two points in manifolds.The Euclidean distance to determinate neighborhoods is replaced by the diffusion distance.The improvement is designed to enhance the anti-noise ability of these algorithms and reduce the computational complexity.Experiments show that the improved manifold algorithms can reduce the computation time and preserve the geometry structures of noisy datasets.(3)The powerful feature extraction ability of the locally linear embedding and the clustering ability of the diffusion mapping algorithm are explored,besides the statistical characteristics of fault data under Monte Carlo simulation are analyzed.Finally the two algorithms are used to extract the geometric features of the fault data.The results show that the method can completely preserve the geometry structures of the original fault data,for the same type of fault data are mapped to one point rather than a fuzzy class.This is a tremendous advance compared to previous fault feature extraction methods.After extracting the fault data by using the proposed method,the faults are completely separable in the visualization space.(4)Three commonly used norm functions in the cost function are discussed and the applications of these criteria are analyzed.It is theoretically proved that the criterion functions based on the information theory learning are not sensitive to non-Gaussian noise,followed by the relationship between these criteria.The relevant formulas are derived in regard to different non-Gaussian noises and the performance of the criteria functions is given.(5)The information theory learning criterion function and the diffusion distance are applied to the local linear embedding algorithm.Experiments show that the improved algorithm can eliminate non-Gaussian noise on artificial data sets.As a result,the original geometry structures of the datasets in low-dimensional space are preserved.For analog circuit fault signals mixed with non-Gaussian noise,the local linear embedding algorithm improved by the information theory and multidimensional scaling analysis can be used for fault diagnosis.It is shown that the algorithm has high-precision diagnostic capability.