Research And Application Of Support Vector Regression

Support Vector Machine (SVM) is a new data modeling method based on Statistical Learning Theory (SLT). It is built on the theory of VC dimension (Vapnik-Chervonenkis Dimension) and the principle of minimum structure risk. It can solve the practical problems on small sample, nonlinear, high dimension, local minimum points and can get better generalization. The application of SVM in the industrial field is relatively small, so it has more chance of application in this field because of the small sample statistics of chemical industry. In this paper, highlights were focused on Support Vector Regression (SVR) and a series of work on how to improve the algorithm performance and generalization was discussed. More important, it will be applied to industrial production of Bisphenol A (BPA) soft sensor process modeling.After analyzing the SVR in detail, regarding to the limitation with single kernel, SVM based on mixtures of kernels is proposed to improve the generalization ability and model's precision. The mixed kernel introduced in this paper is linearly combined by a local kernel and a global kernel and the influence of different kernels can be tuned by parameter. The application in the soft sensor modeling of BPA shows that this method has better generalization ability.Parameters optimization of SVM usually apply to penalty parameter C and kernel parameterσ. However, with the introduction of mixed kernels, SVM has one more adjustable parameterm . m is used to be gotten by experience and it's not the best parameter. In order to improve the model's precision, the parameters of mixed kernels SVM in this paper are selected by Chaotic Particle Swarm Optimization (CPSO) and the k? fold cross-validation error is used as the fitness function to find the optimal parameters. The application in the soft sensor modeling of BPA shows that this method has better generalization ability.The development of SVM algorithm also includes its combination with data preprocessing. It means that the nature of data will be integrated into the SVM algorithm to generate a new algorithm. Fuzzy C-means clustering (FCM) algorithm can bring the problem about interference between border informations, so the accuracy of model can't be improved a lot. Linear Discriminant Analysis (LDA) is an effective method used to expand the limits of the samples which can make Clustering more precise. The combination of FCM and improved LDA used into SVM by multi-models method can add and expand the algorithm. The application in the soft sensor modeling of BPA shows that this method has a certain practicality.