Data-driven soft sensing towards quality monitoring of industrial pasteurization processes

Abstract

In the food and beverage industry many foods, beers and soft drinks usually need to get pasteurized, a process that holds a significant role in the quality and taste of the final product but is difficult to monitor due to the process nature. Soft sensing techniques, also called virtual sensing or surrogate sensing, can be leveraged to monitor the product quality, by using information available from other measurements and process parameters to calculate an estimation of the quantity of interest. In this paper, we develop a soft sensing methodology that is based on machine learning algorithms for continuous, end-to-end estimation of the temperature of products during the pasteurization process, with the vision to serve as an intermediate step towards monitoring live the final quality of the pasteurized products. This work studies a real beer pasteurization process in collaboration with Heineken’s plant in Patras, Greece and the results demonstrate notable performance in temperature prediction accuracy, with average root mean square error (RMSE) of 1.85°C in the test sets. Thus, we claim that it is possible to obtain measurements quite similar to the ones by the respective physical sensors with sufficient accuracy, and our methodology can be considered as a virtual low-cost solution for monitoring product quality in legacy pasteurizer operation.