Soft Sensing of Polyethylene Molecular Weight Distribution Based on Generalized Dynamic Fuzzy Neural Network

Molecular weight distribution (MWD) is an important quality index of polymer. However, it has been a challenging problem for advanced control and optimization of polymerization process to predict MWD, due to the difficulty of real-time detection. In order to solve the problem of real-time and accuracy on the prediction of polyethylene molecular weight distribution, this paper proposes a mixed prediction model on MWD by combining process information and reaction mechanism. First, by the mechanism analysis, MWD is fitted with the number of different active sites. And then, the number of active sites and parameters of the distribution function can be reasonably obtained via the fitting error analysis. Furthermore, the relationship between process parameters and distribution function parameters is described by generalized dynamic fuzzy neural network (GDFNN), in which the K-means clustering algorithm is used to initialize the structure of GDFNN. During the process of training, the historical data and new information will be utilized to determine whether to add new rules for avoiding the frequent generation of redundant rules. Besides, by using hierarchical learning mechanism, both the overall learning rate in the early stage and the local learning rate in the later stage can be effectively improved. Finally, the simulation experiments are made with four operating variables, i.e., rate of ethylene, rate of butane, temperature of reactor, and pressure of reactor, as input variables of the model, from which the effectiveness of the proposed model is verified.