Wave-Net Based On-Line Quality Inference System for Polymerization Processes

Abstract

The paper proposes an on-line inferential scheme to predict a polymer property, the Melt Index, from off-line, past measured values of this property and available on-line measured process variables. In the on-line inferential scheme, the L∞ Wave-Net, which is an artificial neural network with basis functions drawn from the family of wavelets, is utilized by incorporating fundamentals of chemical engineering. The multistep predictor, formulated in a cascaded structure of identical one-step-ahead Wave-Net based predictors, is adaptively learning the functional relationship between the polymer property and process variables. The application of the scheme is illustrated using operating data of an industrial High Density Polyethylene Process. The results of the proposed scheme show that the inferential scheme built upon the Wave-Net can successfully predict the MI and prove its potential as an on-line monitoring scheme to maintain the polymer quality at the highest level.