Melt index prediction by neural networks based on independent component analysis and multi-scale analysis

Reliable estimation of melt index (MI) is crucial for the production of polypropylene. Propylene polymerization process is highly nonlinear and characterized by multi-scale nature with lots of variables and information that are highly correlated and derived at different sample rates from different sensors. A novel soft-sensor architecture based on radial basis function networks (RBF) combining independent component analysis (ICA) as well as multi-scale analysis (MSA) is proposed to infer the MI of polypropylene from other process variables. In the proposed model, ICA is carried out to select the most independent process features and to eliminate the correlations of the input variables, MSA is introduced to acquire approximate and detailed scale information of the process and make the model more robust to mismatches, and RBF networks are used to characterize the nonlinearity in every scale. The approach is evaluated and the results are compared with simplified approaches built with the same data set. The research results confirm the validity of the proposed model.