Multivariate Data Analysis for Advancing the Interpretation of Bioprocess Measurement and Monitoring Data - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 185

process knowledge base, MVDA and first-principles process models and a group
of functional agents. The role of the agent components was to cooperate with each
other in 'describing the whole process behaviour, evaluating process operating
conditions, monitoring of the operating processes, predicting critical process
performance, and providing guidance to decision-making when coping with pro-
cess deviations'. This demonstrated that, for the protein model system used in this
case, the overall agent-based framework provided superior process performance
compared with the traditional approach of optimising individual processing steps.
4.3 Non-Linear Regression Methods
Section 4.2.1 introduced non-linear variants of PLS with ANNs as one of the
means of building non-linear capability into a standard PLS model. ANNs, in
particular feedforward (sometimes referred to as multilayer perceptrons, MLPs, or
back propagation) and RBF networks, have been reported extensively in the
literature in the last two decades as providing an effective means of predicting a
range of important biological variables during bioprocessing. The basic principles
of these techniques are described in Sect. 4.3.1 with appropriate examples of
application.
However, it should be noted that there are also successful bioprocess applica-
tions of alternative non-linear methods reported in the literature. These include
methods such as genetic algorithms [ 10 ] and support vector regression [ 11 ] and
Petri nets [ 8 ] amongst others. These applications cover optimisation of operating
conditions, predictions of various process variables, such as biomass and product/
metabolite concentrations and representation of biological systems.
4.3.1 Neural Networks
Initially developed in the 1940s from research to model the operation of a
biological neuron by McCulloh [ 36 ], their applications increased significantly with
the introduction of the back propagation learning algorithm in the 1980s [ 59 ].
ANNs share many of the characteristics of biological systems such as parallel
computation, robustness, insensitivity to noise, adaptability and good generaliza-
tion properties i.e. the ability to predict data different from that contained in the
training data set [ 4 ]. They also have demonstrated an ability to generate acceptable
models from limited data sets [ 34 ].
A typical neural network consists of a number of neurons arranged in layers as
indicated for example in Fig. 11 . The neurons in the input layer distribute the
process data through the network of weighted connections to the hidden layer(s),
which perform a non-linear transformation of the data before passing the output
through another set of weights to the output neurons.