On-line estimation of process variables using SVM-based MPLS in fed-batch fermentations

Abstract

This paper integrates support vector machines (SVM) into multiway partial least squares (MPLS) resulting in a nonlinear MPLS model. Process data from normal historical batches are used to develop the MPLS model, and a series of single-input-single-output SVM networks are adapted to approximate nonlinear inner relationship between input and output variables. This model can provide a prediction of end-of-batch quality measurements and on-line estimation of process variables such as biomass and product concentration, which are difficult to measure on-line. In addition, the application of a time-lagged window technique not only makes the complementarities of unmeasured data of the monitored batch unnecessary, but also significantly reduces the computation and storage requirements in comparison with the traditional MPLS.