Multiphase Bioreactor Design

measurement device (Figure 3.7a) it results in an identity mapping which will lead to the
recovery of real process output (RPO). They concluded that simple 3-layered feedforward
neural networks may provide very accurate recovery of data distorted by first order
dynamics, including time delay. Closed-loop control could subsequently be performed

Figure 3.7 (a) Neural network

measurement dynamics compensation;

(b) application to closed-loop control.

with the recovered signal (Figure 3.7b), without the difficulties experienced when feeding
the controller with distorted measurements.

Software sensors/Inferential measuring

The design and implementation of software sensors provides a suitable answer to cope
with the lack of instrumental sensors and have been widely reported in the literature.
Software sensors are algorithms for the on-line estimation of those state variables and
parameters, that are not measurable in real time, from more easily accessible related
measurements. One of the most common industrial application is biomass estimation,
using exhaust gas analysis by means of very simple empirical correlations.
The most serious problem regarding industrial application of software sensors is that
they are based on process models that must be accurate and robust. With this respect, and
as usual, kinetics present the most serious difficulty. For these reasons, in recent years
many authors focused their work in developing algorithms for state observation and
parameters estimation, avoiding the knowledge of the underlying kinetic model
essentially by the application of systems theory.
Over the last 20 years, different approaches have been proposed for state and
parameter estimation in fermentation processes. Extended Kalman filters have been
widely reported. This and other stochastic algorithms have been preferred due to the
problems posed by the large quantities of white noise attached to on-line measurements.
However there are today several useful methods available for successful noise filtering.
Many other works focused on the robustness of model-based estimators. Several
applications using the general theoretical framework developed by Bastin and Dochain
(1990) have been reported. Farza and co-authors (Farza et al., (1997a), (1997b), (1998),
(1999)) have proposed observerbased estimators derived from nonlinear systems theory.

Multiphase bioreactor design 80