Inferential Estimator Algorithms And Application In Chemical Engineering Modeling

Inferential estimator is the key technique of an inferential control system design.The design of inferential estimator model (soft-sensing model) is the key part of a control system, and also is the core of soft-sensing technique. It's different form mathematical model in general meaning, and emphasizes getting the optimal estimation of a primary variable from secondary variables.In this paper, based on the design methods of inferential estimator, it is studied for Partial Least Squares (PLS) and Support Vector Machine (SVM). PLS is a new multivariate statistical analysis method. In the model of multiple dependent variables to multiple independent variables, and there is high correlation of inner every variables, PLS is more reliable. Besides, SVM is a new data modeling method theoretically based on statistic learning theory. Employing the criteria of structural risk minimization, which minimizes the fitting errors of the sample-data and control simultaneously the complexity of the learning function. Thus, the SVM's generalization is better than others.A new approach to improve the generalization ability of partial least squares regression with Shrink-Magnifying thought is presented in this paper. This approach is to implement shrinking or magnifying the input vector, reducing test error, improving the generalization ability of PLS. Some simulation results show this algorithm is effective and improved the generalization ability of PLS.In ordinary multivariate modeling methods, it is thought that each input sample makes the same contribution to the output of model, which the difference of representativeness between the samples is ignored. In this paper, by introducing the concept of weighted sampling and decrement learning into Partial Least Squares (PLS), a new multivariate regression method, Regulable Weighted PLS (RWPLS) is proposed. Different sample in the training set is weighted differently according to its representativeness to improve the prediction ability, and develop a model with better generalization ability. RWPLS is utilized to develop a soft sensor for the crystallization process of Bisphenol-A, and the results show that the prediction by RWPLS is much more precise.The main idea of inferential control is to calculate primary variable for feedback control, or calculate interference for feedforward control. According to this idea, a new modeling method based on output deviation correction to improve the generalization ability is presented in this paper. This approach introduced the model test error to model output correction and improved the accuracy of predication model. At the same time, it integrated with support vector machine. It is also utilized to develop the soft sensor for crystallization process of Bisphenol-A, and the results show that it has a better predicted precision.