Study On Multi-class Pattern Statistical Recognition Model And Application

In realistic life, the problem of pattern recognition, especially the study of multi-classification pattern recognition is very significant. The field of pattern recognition is very comprehensive, for example, disease diagnosis and physical check-up in the research of iatrology, fingerprint recognition and face recognition when solving a criminal case, the problem of fault diagnosis and detect prediction in the power system, communication system and traffic system etc., all that above can be regarded as the study of pattern recognition. At the same time, the problems such as much more feature indexes, multicollinearity of feature, smaller sample size and noisy data often may be encountered when modeling, they will influence the parameter estimation of pattern recognition model badly, expand the model error and the model's robustness will be destroyed.The SVM method based on the statistical learning theory is one classification method that own the good generalization ability. It displays the unique superiority and the good application prospect in solving the small sample, non-linear and in the high dimension pattern recognition question. But the basic support vector machines is proposes based on two class problems, and few research results in multi-class pattern recognition. Therefore, the study on the SVM statistical pattern recognition model and the method based on the high dimension possesses the remarkable academic value and the important practical value.Aimed at the problems that encountered in the pattern recognition, such as the redundancy of the sample and multi-collinearity of feature indexes, as well as the incertitude of the class information etc., the paper carried on the thorough analysis to the pattern recognition methods, the SVM theory and algorithm and the partial least squares methods. It combined the partial least squares and the fuzzy support vector machine and proposed a multi-class pattern recognition model based on PLS and FSVM The model made use of the thought of PLS to accomplish the pretreatment to the sample data. Then it used FSVM to model for the reduced samples. The contents and the results of the paper mainly contained:(1) The paper has carried on the thorough analysis to the domestic and foreign research actuality of the multi-class pattern recognition methods, especially carried on the thorough study to the SVM theory and algorithm. And carried on the thorough analysis to the partial least squares ideas and methods. I realized the new small sample pattern recognition methods, and built the solid foundation for the further research.(2) The paper proposed one new multi-class pattern recognition model: the model based on PLS and FSVM. The model combined the advantage of the PLS and FSVM. It overcame the influence that the redundancy of the sample and multi-collinearity of feature indexes, as well as the incertitude of the class information etc. on the pattern recognition effect.(3) In order to get the better fuzzy membership, the paper made use of the fuzzy C-mean algorithm when we made certain the fuzzy membership in the FSVM.(4) The paper made use of the data sets in the UCI database and the turbogenerator unit fault diagnosis data to carry on analyzing and appraising in the training speed (the amount of the SVs) and the recognition effect (the correct recognition rate) to the model. Through the experimental result, the model that the paper proposed based on PLS and FSVM possessed the higher performance to compare with the traditional SVM multi-classification algorithm.