Research On The Generalization And Applications Of Suport Vector Machines

Some variables which characterize the quality of the final product are hard to be measured by instruments in a chemical process. By using soft-senor technique, their estimations can be determined by mathematics models.The Support Vector Machine (SVM) based on the statistical learning theory is a widely used for soft-sensor models. It uses Structural Risk Minimization to supply a efficient solution of small samples, nonlinear and high-dimensional problems. Due to the complexity, the single SVM model doesn't meet the requirements of the actual precision, so the research is made for the theory and the application of the SVM.(1) Based on the support vector machine with gridding search method to search the kernel function, the soft-sensor model which is based on the parameters searched globally is established. The result of simulation shows the algorithm can find the most suitable parameters.(2) K-Nearest Neighbor (KNN) is simple, but when the number of the data is larger, it needs many operations. To solve this problem, an improved KNN algorithm is used to classify the data which was processed by Bayes algorithm and is marked the number of the data point in every son-boxing structure. Simple operation and part of the distance operation is combined to find K nearest distance. Then the parameters are found to build the Compositional Support Vector Machine Model. The result of simulation whose data is picked from a chemical plant shows that the approach developed here can effectively reduce the number of the data which need to know the distance.(3) Fuzzy C-means and other clustering algorithm by changing metric methods are normally require the membership normalized. When the sample data has some noise or isolated points, the result of the clustering isn't good. The likely reason is that they get bigger membership. By improving the constraints of membership, the requirement of normalization is loosened to generate a new clustering algorithm. The result of simulation whose data is picked from a chemical plant shows that the improved method brings better result of the model.