Automatic Control of Bioprocesses - Measurement, Monitoring, Modelling, and Control of Bioprocesses - page 53

Fig. 5 Basic artificial neural network with four inputs in the input layer, five neurons in the
hidden layer, and two outputs in the output layer
parameters, here called weighting factors, are calculated by an optimization
algorithm. The weighting factors are used to weight each input of a neuron. The
sum of the weighted inputs is used as an argument of the activation function to
calculate the output of the neuron. A vast amount of different training data is
necessary to build the training sets for a certain problem. Greater variety in the
training data leads to better prediction performance in unknown scenarios and
prevents that the training data are only memorized.
Karakuzo et al. [ 55 ] presented an ANN soft sensor with fuzzy controller for fed-
batch fermentations of baker's yeast. The performance of the controller was
compared with that of a controller using a theoretical model-based estimation of
the specific growth rate. As input to the network, the exhaust gas O 2 and CO 2
concentrations, the feeding rate, as well as the temperature and pH were used (five
input neurons). The neural network consists of six neurons in the hidden layer and
one output neuron to estimate the specific growth rate. For globally robust training
data of their ANN, cultivation datasets under a lot of different process conditions
were necessary: The authors generated a training dataset containing 360 patterns
and an evaluation dataset with the same number of patterns from cultivation data.
The results of the model predictive controller they used for comparison gave
satisfactory estimation for the specific growth rate, however only under fixed
inoculum sizes. The potential of their ANN becomes obvious during the change of
inoculum sizes. The ANN continues to generate reliable estimations for the spe-
cific growth rate. They also applied a fuzzy logic controller for air flow and
feeding control based on the ANN soft sensor specific growth rate estimation.
They performed simulation studies with this controller setup, leading to acceptable
results for large-scale applications.
Gadkar et al. [ 56 ] presented an online adaptive neural network as a soft sensor
that estimates the substrate, ethanol, and biomass concentrations based on dis-
solved oxygen measurements (one input neuron, three output neurons) during
S. cerevisiae fermentation. Their neural network had three hidden layers with ten,
eight, and four neurons, respectively. They discuss the performance with and