Automatic Control of Bioprocesses - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 58

feeding rate. When parameter uncertainties and noise disturbances occur, they
determined that a recursive least-squares formulation was simple to implement and
improved the resulting set-point tracking.
3.9 Control Based on A Heuristic Procedure
Spectroscopic measurements provide a wide range of information based on the
interaction of electromagnetic radiation with matter. At the same time, this abun-
dance of information is the reason why it is often difficult to interpret. Advanced
mathematical tools such as partial least squares or principle component analysis are
employed to overcome this information overload. Hantelmann et al. [ 71 ] present a
new method to monitor and control S. cerevisiae cultivations by two-dimensional
(2D) in situ fluorescence spectroscopy. They introduce a chemometric model that is
derived from multivariate data analysis. The glucose feeding rate is thereby con-
trolled, predicting the metabolic state directly from the fluorescence intensities. The
glucose concentration was held between 0.4 and 0.5 g L -1 over 11 h, completely
avoiding ethanol formation. They point out that the BioView technology that they
used for cultivation control is suitable for industrial environments.
Schenk et al. [ 72 ] present a soft sensor based on mid-infrared spectroscopy.
They introduce a simple and fast method to calibrate the instrument for Pichia
pastoris fermentations. For this purpose, they assume that only the substrate
concentration will change significantly during the cultivation and that the absor-
bance is proportional to the concentration. The control action is performed using a
PI controller. They propose that, in some cases such as Pichia pastoris, a multi-
variate calibration procedure is not necessary and the measurement of one com-
pound of interest is sufficient (methanol in the case of P. pastoris). The calibration
is performed in situ using two points: one spectrum at the beginning without
carbon source, and one with carbon source. They carried out six cultivations in the
range of 0.8-15 gL -1 to demonstrate the performance of the control system. The
standard error of prediction over all cultivations was 0.12 gL -1 . They point out
that long-term baseline instability had an influence on the accuracy, which could
be addressed using linear correction of the signal. Even though the method was
designed for the special case of P. pastoris, the authors mention a possible
application for other microorganisms.
4 Conclusions
In this chapter, different approaches of open-loop and closed-loop control for
bioprocess automation are discussed. As a result of the diversity of bioprocess
requirements, a single control algorithm cannot be applied in all cases; rather,