Soft sensors based on interpretable learners for industrial-scale fed-batch fermentation: Learning from simulations

Abstract

Monitoring bioprocesses is a challenging task where most of the variables of interest can only be measured offline. Soft sensors have emerged as a solution to provide online estimations. This work compares interpretable learners such as CART, M5, CUBIST, and Random Forest as soft sensors for industrial-scale fed-batch fermentation of penicillin production. A structured model of industrial-scale penicillin fermentation is implemented to generate the dataset and train the interpretable learners. Variables such as substrate feed rate, agitation, temperature, pH, dissolved oxygen, vessel volume, CO2, and O2 percent in off-gas are considered as independent (predictors). The CUBIST model has achieved the best results with values of 0.908, 9.916, 3.149, and 1.920 for the coefficient of determination, Mean Squared Error, Root Mean Squared Error, and Mean Absolute Error, respectively. These results demonstrate the feasibility of developing soft sensors based on interpretable models to predict penicillin concentration at an industrial scale.