Automatic Control of Bioprocesses - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 56

Fig. 6 Basic structure of a fuzzy-ANN local linear model

of neurons will describe a trajectory best, but will possibly not decrease the
required computational power.

3.6.2 ANN-Based Estimation for NMPC

Meleiro et al. [ 64 ] presented results of a MPC strategy of a fuel-ethanol fermen-
tation process using simulations. A neural network was applied as the internal
model for the controller. The authors used an optimization algorithm to determine
the neural network structure as well as the shape of their activation functions,
guiding the parsimonious network architecture. The inputs were the feed flow rate,
cell recycle rate, and flash recycle rate; the output were the biomass, substrate, and
product concentrations. The authors presented results demonstrating successful
control of the biomass, substrate, and ethanol concentrations with set-points
varying between 37 and 32 gL -1 , 10 and 3 gL -1 , and 45 and 40 gL -1 , respectively.

3.7 Probing Feeding Controller Strategy

Velut et al. [ 65 ] presented a probing feeding strategy for E. coli fermentations,
operating close to the maximum oxygen transfer rate capacity. The principle of the
probing feeding strategy is to superimpose a short glucose pulse onto the glucose
feeding flow and evaluate the response in the dissolved oxygen signal. If the
dissolved oxygen level decreases, the feed rate is increased, due to the determined
capacity. If no response can be detected, the feeding rate is decreased. The
technique is combined with DO control, which is performed by adjusting the
stirrer speed and the temperature control to decrease the oxygen demand when
the reactor is at its maximum oxygen transfer capacity. They present the perfor-
mance of the combined controller in an E. coli cultivation. Due to the probing