Multivariate Data Analysis for Advancing the Interpretation of Bioprocess Measurement and Monitoring Data - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 181

Fig. 9 Physiological state (PS)-based inferential bioprocess control scheme indicating a number
of PS-specific models used for estimating the process variable of interest
In an inferential control scheme as proposed in Fig. 9 , a classification method is
required to allow offline and online prediction of the physiological state of the
culture and therefore allow the appropriate estimation model (Mod 1, Mod 2 or
Mod n in Fig. 9 ) to be used to estimate the process variable of interest, y(est), and
thus drive the controller to maintain the desired set-point.
The PNN approach, described in Sect. 3.3 , was used in this particular case study to
predict three PSs, assigned on the basis of the PCA-CVA, using online data typically
measured during cultivation (e.g. dissolved oxygen, pH, and CO 2 and O 2 concen-
tration in the exhaust gas). Whilst the network successfully predicted the PS for most
of the cultivations throughout the duration of the experiment, further improvement
was required in assigning data into a particular PS based on the PyMS data.
4 Regression Methods
Exploratory data analysis, clustering and classification are important during
bioprocess development and monitoring as highlighted in Sect. 3 . For a range of
applications, in particular for software sensors (Chap. 1), it is very important to use
regression methods which allow the prediction of a desired variable from process
data presented to the model. Given the issues with the measurement of important
biological variables during bioprocessing highlighted in Chap. 1, it is not
surprising that MVDA applications in this area are abundant in the literature.
The most pertinent of these are mentioned below.