Complex batch processes quality prediction using non-Gaussian dissimilarity measure based just-in-time learning model

Abstract

In modern batch processes, soft sensors have been widely used for estimating quality variables. However, they do not show superior prediction performance due to the self-limitations of these methods and the unique characteristics of batch processes such as time-varying, nonlinearity, non-Gaussianity, multi-phases and batch-to-batch variations. To cope with these issues, a novel non-Gaussian dissimilarity measure based just-in-time learning (JITL) soft sensor is developed in this paper. Unlike the traditional JITL model which uses the distance-based dissimilarity measure for local modeling, the proposed method uses non-Gaussian dissimilarity measure to evaluate the statistical dependency of the extracted independent components to construct the local model, which can well capture the non-Gaussian features in the process data. Furthermore, a novel relevant samples search strategy is introduced into the JITL framework for local modeling, which not only searches the relevant samples along the direction of time axis but also along the direction of batch-to-batch. The proposed search strategy can guarantee the current query sample and the local modeling data belong to the same phase duration and have the smallest process trajectory variations. Hence, the proposed soft sensor is suitable for uneven-phase and batch-to-batch variations batch processes. Meanwhile, the proposed method can well cope with the changes in process characteristics as well as nonlinearity. The effectiveness of the proposed method is verified on the fed-batch Penicillin Fermentation process.