Identification on demand using a blockwise recursive partial least-squares technique

Abstract

Regression techniques that are used for online estimation and control generally yield poor models if the data is not rich enough. Further, restrictions on the lower limit of the forgetting factor, used to prevent ill conditioning of the covariance matrix, confine applications of such techniques to slowly changing processes only. In this paper a modified blockwise dynamic recursive PLS technique that is based on selection of rich data has been proposed for online adaptation and control. Because of its ability to accommodate a wider range of forgetting factors, the technique is found to track the dynamics of slow as well as fast changes in the processes. The proposed technique has been evaluated for adaptation and control using representative chemical processes taken from the literature.