The Improved Support Vector Machine And It's Application In Soft Sensing Of White Water

White water concentration is a very important parameter in process of paper production. To measure white water concentration accurately and timely can improve productivity and efficiency , and reduce the extent of pollutants discharge. But the traditional methods can not get the measurement of white water concentration online and need a lot of time. As to this problem, a soft sensing based on Least Square Support Machine was proposed in this paper, and was applied in the soft sensing of white water concentration.This paper presented as follows:(1) Above all, the application of soft sensing in paper production ,the basic principle of SVM applied in function regression, and the basic process of Niche Genetic Algorithms were introduced briefly.(2) The niche technique combined self-adaptive crossover probability is adopted to improve the parallel genetic algorithm, which can maintain and enrich the community diversity of population, avoiding the problem of prematurity and low convergence speed. And the parallel algorithm is proposed to enhance the ability of local searching and accelerate positioning the overall optimal solution. At last, the improved Niche Genetic Algorithm(NGA) was available.(3) On one hand , an improved SVM regression method is proposed ,as to the problem that LS-SVM loses the sparseness of solution .According to the learning error of the SVM modeling , most of the samples with small errors in the variable space are eliminated. Accordingly , the sparseness is recovered. On the other hand , The Niche Genetic Algorithm is mainly used to select and optimize the parameters of LS-SVM (Loss function parameter, Punishment factor, Core function and it's parameters) by making full use of the overall searching ability of genetic algorithm, which offer an effective way for SVM parameters automatic selection.(4) According to the technics of white water, the soft sensing modeling was built up, with the appropriate assistant variables and the improved LS-SVM whose parameters are selected by NGA. The study result shows that the proposed method has better estimate accuracy compared to the generic LS-SVM ,in which the parameters are selected artificially; and it can satisfy the needs of industrial production quite well.