Figure 4.19. A hypothetical feedforward ANN with one hidden layer for estimating substrate and biomass concentrations in fed-batch penicillin fermentation (OUR: Oxygen uptake rate, S: Substrate conc., X: Biomass conc., wij and weight vectors associated with interconnected layers).
structure that combines parametric models based on fundamental process knowledge and nonparametric models based on process data.
Three-layer feedforward networks with backpropagation learning algorithm are dominantly used in the applications mentioned above. A hypothetic neural network structure is shown in Figure 4.19 for estimating biomass and substrate concentrations in penicillin fermentation utilizing frequently measured variables such as feed rates and off gas concentrations.
There are many educational and commercial software packages available for development and deployment of ANNs. Most of those packages include data preprocessing modules such as Gensym's NeurOn-Line Studio .
ANNs can be seen as autoassociative regression models. They resemble statistical modeling techniques in that sense. But the lack of statistical inference and robustness issues may cause problems. Care should be taken (e.g., data pre-processing, appropriate selection of input-output data set) during deployment. The advantages/disadvantages summarized in the introductory paragraph of Section 4.6 should be taken into consideration prior to deciding if ANNs are appropriate for a specific application.
Was this article helpful?