A model-based approach to quality monitoring of a polymerization process without online measurement of product specifications

This paper presents a model-based approach for on-line monitoring of difficult-to-measure quality variables with the description of their dynamic behavior. The strategy comprises the use of a Virtual On-line Analyzer (VOA), based on an empirical model, whose development is supported on the one hand by experimental data, and on the other hand by a complex phenomenological model. A systematic approach to the selection and estimation of parameters of the phenomenological model from experimental data is also presented in order to adjust this model to the behavior of the industrial process. Regarding the experimental data, this complete model has additional information (such as the dynamic behavior of the process, embedded in the knowledge used in its formulation) and it is able to provide synthetic data for VOA training which comprises a NARX (Non-linear AutoRegressive with Exogeneous inputs) neural network model. A real case study comprising quality monitoring in a polymerization system is investigated. The available measurements of the main polymer properties are performed at a low frequency and are not able to represent the dynamic behavior and support decision-making at the operational level, especially during online grade transitions (changes in product specifications). The results show the ability of the neural model to predict the quality variables with a high frequency (small sampling period), enabling and supporting decision-making for quality monitoring of the polymer in real time. (C) 2017 Elsevier Ltd. All rights reserved.