Application Of Data Mining In The Modeling And Control Of Polymerization Process

It is difficult to model polymerization process well based on internal mechanism of the process, due to the complexity of reaction mechanical, infeasibility of parameter measurements and strong non-linearity in the reaction. Artificial neural networks (ANN) needn't depend on the process's mathematical mechanism model exactly and can approximate any function map in any given precision; which shows particular advantage on modeling of complicated system with deficient prior knowledge. Nevertheless, the training time of ANN is usually too long and there also exists local minimum. To get over the demerit of slow training, this thesis presents some new faster training algorithms and introduces global optimization algorithm - genetic algorithm (GA) in order to overcome local minimum problem. A stacked neural networks is applied to construct predictive model of polymerization process. And also some achievements are applied to thermal bulk polymerization between styrene (ST) and maleic anhydride (MAH) and suspend polymerization of chloroethylene. The main research contents and characteristics in the thesis are as follows:(1) Standard back-propagation (BP) study algorithm is a simple and static optimization method. The implementation of standard BP study algorithm updates the network weights and biases in the direction in which the performance function decreases most rapidly - the negative of the gradient. It seems to be "steepest descent" in local region, but is converge slowly in whole region. As standard BP study algorithm is prone to converge slowly, some improved principles about standard BP study algorithm are discussed. The improved optimization methods are analyzed and compared each other. At last, a Levenberg-Marquardt (LM) algorithm for improving the convergence performance of BP neural network training is proposed. The experimental results show that the LM algorithm for BP neural network can achieve significant improvements in convergence performance.(2) Traditional BP neural network training algorithm may lead to entrap into local optimization. In accordance with the global optimizing ability of genetic algorithm (GA), GA is proposed to overcome the problem of local minima of traditional BP training algorithm. In order to improve the convergence performance of GA, an improved GA is developed to overcome "escape phenomena" of global optimal solution by a dynamic adjustment method of variable range. Numerical calculation is shows that the improved GA algorithmis superior to standard GA.(3) Artificial neural network (ANN) is a effective modeling method, especially for chemical processes with complex mechanism. In this paper, the modeling of semi-continuous bulk co-polymerization of styrene and maleic anhydride based on artificial neural networks was studied. The experimental data were used to train Back-Propagation (BP) network in order to predict the conversion of co-polymerization. The training velocity of the improved method (Marquardt algorithm) for BP network is increased more than 10 times. Under different initial conditions, such as residence time 3~7 h, polymerization temperature 110-120 , and maleic anhydride content in feed 7-10 mass%, the satisfied convergence was obtained. In the case of 3 inputs and 1 output (conversion), the predicted maximum relative error is about 10-15% and average relative error is less than 5%. The good agreement of the predicted conversion with the experimental data shows that the network model can be effectively used to model polymerization process.(4)A stacked neural networks - ridge regression (SNNs -- RR) modeling approach aimed at obtaining improvements in neural network model performance is proposed. A single neural network model developed from a limited amount of data usually lacks generalization capability. The stacked generalization for neural network models by integrating multiple neural networks into an architecture known as (SNNs), and model generalization capability can be significantly improved by using SNNs model. Proper selection of the stacki...