Related Experiment Videos
Functional nodes in dynamic neural networks for bioprocess modelling.
M Fellner1, A Delgado, T Becker
1Lehrstuhl für Fluidmechanik und Prozessautomation, Technische Universität München, Aussenstelle Freising-Weihenstephan, Weihenstephaner Steig 23, 85350 Freising, Germany. markus.fellner@epost.de
Bioprocess and Biosystems Engineering
|September 25, 2003
Summary
This study introduces a novel hybrid modeling approach for industrial yeast fermentation, integrating prior knowledge into artificial neural networks (ANNs). This method significantly reduces training data needs by 50% while maintaining accuracy and improving robustness.
Area of Science:
- Biotechnology
- Computational Science
- Chemical Engineering
Background:
- Industrial yeast fermentation requires accurate online monitoring of secondary metabolites.
- Traditional data-driven models often lack robustness and require extensive training data.
- Integrating existing system knowledge can enhance model performance.
Purpose of the Study:
- To present a novel method for directly integrating a-priori knowledge into artificial neural networks (ANNs).
- To apply this hybrid approach for online determination of secondary metabolites in industrial yeast fermentation.
- To demonstrate the benefits of hybrid modeling over pure data-based approaches.
Main Methods:
- Integration of existing system knowledge into ANNs using 'functional nodes'.
- Development of a generalized backpropagation algorithm for hybrid models.
- Incorporation of ordinary differential equations (ODEs) describing diacetyl formation/degradation into a functional node.
- Hybrid integration of the functional node within a dynamic feedforward neural network.
Main Results:
- The hybrid modeling approach significantly outperformed pure data-based modeling in robustness and generalization.
- The amount of necessary training data was decreased by 50% while achieving comparable accuracy.
- All incorrect decisions, based on defined cost criteria, were avoided compared to conventional ANNs.
Conclusions:
- Hybrid modeling, combining a-priori knowledge with experimental data, offers superior performance in industrial yeast fermentation monitoring.
- This approach enhances model reliability and reduces data requirements.
- The method provides a more efficient and accurate way to determine secondary metabolites during fermentation.