Experience of neural networks for inferential estimation in industrial process control

This paper is concerned with viscosity control in a polymerisation reactor. Here the measurement from the viscometer is subject to a significant time delay but the torque from a variable speed drive provides an instantaneous, if noisy, indication of the reactor viscosity. The aim of the research was to investigate neural network based inferential estimation, where the network is trained to predict the polymer viscosity from past torque and viscosity data thus removing the delay and providing instantaneous information to the operators. The paper presents results from off-line training of a feedforward network and describes the study on online viscosity estimation using B-Spline networks