Prediction of Batch-End Quality for an Industrial Polymerization Process

Abstract

In this paper, an inferential sensor for the final viscosity of an industrial batch polymerization reaction is developed using multivariate statistical methods. This inferential sensor tackles one of the main problems of chemical batch processes: the lack of reliable online quality estimates. In a data preprocessing step, all batches are brought to equal lengths and significant batch events are aligned via dynamic time warping. Next, the optimal input measurements and optimal model order of the inferential multiway partial least squares (MPLS) model are selected. Finally, a full batch model is trained and successfully validated. Additionally, intermediate models capable of predicting the final product quality after only 50% or 75% batch progress are developed. All models provide accurate estimates of the final polymer viscosity.