Nonlinear Model Research Based On Data-Driven Method

Flow industry system always includes numerous process variables and all kinds of complicated physical and chemical reactions. It's difficult to model this kind of process. Data-driven models are based on process sampling data and have the characteristics of few understanding of process mechanism, high accuracy, nice generality and so on. The data-driven modeling method is widely used in the field of modeling and optimization in flow industry. It's usually difficult to get accurate models because the sampling data always involve high nonlinearity, noises and variables coupling with each other and so on. This thesis emphatically studies the nonlinear data-driven modeling methods and especially puts an effort on the artificial neural networks and principal curves, which are both good at dealing with nonlinear characteristic. In the thesis, we combined the statistic theory with neural network models and successfully introduce principal curve to nonlinear regression at the first time. The main jobs of this thesis are as follows:According to the fact that industrial data involve noises and nonlinear, collinear and dynamic features, a soft-sensing model based on DPCA-RBF network is proposed. First, pretreatment the data with dynamic PCA to get principal component. Then, build the RBF network model between principal component and the important quality variables. This model effectively decreases the actual modeling variables, noise, dynamic features and so on, reduces the number of parameters and improves the modeling accuracy.Emphasis on nonlinear PCA methods based on neural networks and principal curve. Analyze the characteristics of the two different NLPCA methods in principle and compare the two methods with data sets from nonlinear functions and TE process. The results indicate that NLPCA model based on principal curve is much better than the model based on neural networks with higher precision and much stable score. Besides, the NLPCA model based on principal curve is more suitable to theory of nonlinear principal analysis.According to the fact that industrial data has the characteristic of high dimension, nonlinearity and data coupling with each other, put forward a nonlinear regression model based on principal curve. The model draws lessons from the basic idea of PLS model and extracts latent variables by principal curve simultaneously considering the correlation between dependent variables and independent variables. We introduce polynomial functions to fit nonlinear relations in latent space and validate the model with data set from nonlinear functions and VCM distillation unit respectively. The results show that the model can solve the nonlinearity problems with less latent variables and more accurate results than linear PLS model and polynomial PLS model.