Comparison of different model-based observers for monitoring substrate and product concentrations in a Penicillium chrysogenum Fed-batch process

Abstract

Real-time process monitoring and control strategies are needed to guarantee the required product quality and the improvement of the manufacturing process. Monitoring in the fermentation industry often relies on the analysis of offline samples. This time-consuming measurement technique hinders the application of automated feedback control. Novel developments in state estimation and the availability of chemometrics makes it nowadays possible to use online collected information in order to estimate non-measured system states. In this work, an industrial strain of Penicillium chrysogenum was used as a model organism in a fed-batch fermentation process with penicillin as a main product. Two vibrational spectroscopy methods (nearand mid-infrared) were introduced. Spectral data was used for the construction of black-box PLS models, which were able to predict the concentrations of soluble components (with prediction errors below 30% for most of the processes). However, it was shown, that constructed PLS models lack transferability. Kinetic modeling was also applied. Nevertheless, in order to be able to react to the unforeseen process changes, kinetic models need real-time process information. Therefore, automated feedback regulation was supported by a model-based observer particle filter. Combination of off-gas measurements, nearand mid-infrared based PLS models, as well as the kinetic model via particle filter led to the establishment of a good, real-time process monitoring strategy. Finally, the established system was validated by a real process and proper process control was successfully reached.