An Application Of Support Vector Machines For Bisphenol A Soft Sensor

The goal of process control in industrial production is obtain qualified products. So quality control has become the core of controlling. In order to meet the requirements for quality control, it is necessary to obtain quality and physical parameters such as composition, concentration and viscosity. The use of soft-sensor technology can estimate variables by building soft-sensor models and using measurable information, which is the research direction with very broad prospects.Support Vector Machine (SVM) is a new data based modeling method based on statistical learning theory (SLT). It is based on the theory of VC dimension and the principle of structural risk minimization, which can well solve the problem with small samples, nonlinearity, high dimension and local minimum points. At the same time it can bring better generalization ability. The application of SVM in the industrial field is relatively limited. Because of the character of small samples in the chemical field, SVM has broad prospects for application. This paper studies the performance and generalization ability for support vector machine regression (SVR) algorithm and applies it to the soft sensor modeling for Bisphenol A production process.First of all, this paper discusses some basic results of machine learning theory, the development of machine learning, problems to be solved by machine learning, some concepts and methods of statistical theory and methods as well as the development and main research tools of support vector machine. It analyzes a variety of support vector machine algorithms and the property of support vector machine regression (SVR). It explores and compares linear SVR algorithms and discusses the parameter selection of SVR based on the Particle Swarm Optimization method.Secondly, this paper studies the performance and generalization ability of SVR algorithm in the soft sensor modeling of Bisphenol A production process. It also introduces the modeling process.It is found that the accuracy of soft-sensor models is highly related with the selection of SVR parameters including nuclear parameter, penalty parameter andεinsensitive factor. But there is no analytical method to guide the SVR parameter selection. This paper uses Particle Swarm Optimization (PSO) algorithm to select SVR parameters. This method takes SVR parameters as the particle swarm and the minimization of the 5-fold cross-validation error as the adaptation goal. Then it uses PSO with strong global search ability to achieve the parameter optimization. Numerical functions and practical application examples show that the method can obviously improve the generalization ability to soft sensor models...