Development of a robust hybrid estimator using partial least squares regression and artificial neural networks.

Abstract

Measurement difficulty is one of the process control issues arising from the complexity and the lack of online measurement devices. One of the alternative solutions to deal with the problem is inferential estimation where secondary variables, such as temperature and pressure are used to predict the unmeasured primary variables that are manly product qualities. This paper presents the estimation of product composition for a fatty acid fractionation column using a hybrid technique. The proposed technique combines partial least square regression (PLS) and artificial neural networks (ANN) in an estimation paradigm to provide better estimation properties. The aim is to take advantage of ANN capability to capture the non-linear relationships as well as the statistical strength of PLS method. The results of process estimation using both PLS and hybrid methods are presented. The significant improvement obtained by the hybrid strategy revealed its capability as potentially viable estimator for product properties in chemical industry.