Multi-scale Kernel Methods And Its Application In Electronic System Test

The increasing complexity of the structure and the wide application of intelligent technology make the traditional test methods difficult to meet the needs of modern weapon systems on diagnostic accuracy and speed.As an important guarantee condition of maintainability and reliability,testability is significant to improve weapon system effectiveness and reduce the life cycle cost.In order to provied more effective test method and ensure good testability,kernel methods are used and studied in this paper.Kernel methods are excellent pattern recognition methods for limited samples,they have the advantages of solid theoretical,clear reasoning process and suitable to be used in the field of weapon system testability.The support vector machine theory was researched at first,for simplification of the classification process and improvement of the classification accuracy,the multiclassification methods of support vector machine were studied.Then the basic theory of kernel methods was researched.In order to improve the performance of test methods,while reducing algorithm complexity,the performance improvement of kernel functions was considered as a starting point,the selection of kernel parameter and the constructure of kernel functions were studied.At last,the studies of kernel methods were used to improve the performance of support vector machine.The commonly used multi-classification methods of support vector machine have the disadvantages of complex structure,large amount of calculation and unclassifiable regions.Aiming to the disadvantages,an improved support vector machine based on minimum spanning tree is proposed.Compare with the other classification methods,support vector machine based on minimum spanning tree has the advantages of simple structure and small amount of calculation,the clustering ability of minimum spanning tree can reduce classification error effectively.Fisher separability measure is used as the basis for constructing a minimum spanning tree.To prove the application value of weapon system testability improvement,fault diagnosis of mixer circuit was taken as an example.Accumulated error has significant influence on the performance of minimum spanning tree support vector machine based on Fisher separability measure.In order to solve this problem,two important problems of kernel function,kernel parameters selection and new kernel functions construction were studied.To solve the problems of kernel parameter selection such as complex algorithm and large amount of calcucation,two different kernel parameter selection methods,twice searching method of kernel parameter based on uniform design and kernel parameter selection method based on improved fruit fly optimization algorithm,were proposed for different usage requirements in this paper.Aiming at the problem of complex data structure in system test,the construction of single kernel function and multiple kernel functions were studied.A new kernel function,Rayleigh kernel function,was proposed and its validity were proved by closure properties of kernel functions.The commonly used multiple kernel functions are mostly weighted combination by different single kernel functions and have shortcomings like large amount of calculation and complexity of kernel parameter selection.To avoid the problems,Cauchy kernel function and Raylegh kernel function were extended to multi-scale kernel functions,and a classification integration method of support vector machine was proposed.The performance and effectiveness of the above kernel functions and algorithm were proved by the classification experiment on standard data sets.On the basis of the above studies,single kernel function was replaced by multiscale kernel functions to reduce the classification risks of FMST-SVM.To test the rationality of the multi-scale kernel functions usage in minimum spanning tree support vector machine based on Fisher separability measure,the effect of kernel parameters change on the support vector machine classification model was investigation though classification experiment,and then some qualitative conclusions were given.The effectiveness of the method proposed in this paper was proved by the classification experiment.Then,circuit fault diagnosis problems were taken as examples to prove the effectiveness and application value of the classification method in electronic system testability.