Research On Soft-sensing Methods For Estimating Parameters Of The Concentration In Continuous Stirred Tank Reactor

Continuous Stirred Tank Reactor is a production method of polymeric chemical reaction, which is adopted widely in industry now. There are many typical characteristics of this method just like in the complicated manufacturing process, such as nonlinearity, great inertia, time-variant and so on. The concentration of the reactant is a key chemical parameter in this process, but it is very difficult to be measured on-line because of poor economy or technology. In general, we use the manual sampling analyses method to obtain the sampled value of this parameter. This method would cause much delay in result information, which brings great difficulty in the online supervision and quality control in the manufacturing process. Based on the problems discussed above, this paper focuses on Soft Sensing method of the concentration of reactant.The main idea of Soft-Sensing technology is described as following:through building mathematical modeling between immeasurable key process variables and their related measurable process information, the immeasurable process variables could be predicted. Based on the background of the polymeric chemical reaction in Continuous Stirred Tank Reactor, this paper focuses on the Soft Sensing of the concentration of reactant.This paper mainly includes the following aspects.Firstly, a concentration soft-sensing model is developed based on a multi-neural network. At the beginning, this input information is classified to get a series of sub-models. Then these sub-models are inner connected on a multiple model structure. This multiple modeling method could improve the precision and generalization performance of the model to some extent.Secondly, a hybrid modeling method is presented by combining mechanism modeling and intelligence modeling to estimate the parameters of the concentration of reactant in Continuous Stirred Tank Reactor. It builds the hybrid models of the concentration of reactant in Continuous Stirred Tank Reactor using serial and parallel methods. This hybrid modeling method causes the result that this model reduces the dependence degree of the training data and improves the generalization performance. Finally, the problems of online adjustment of Soft Sensing model are discussed.