2024-03-28T09:23:52Zhttps://www.tdx.cat/oai/requestoai:www.tdx.cat:10803/85142017-09-13T07:47:20Zcom_10803_311col_10803_318
nam a 5i 4500
neural networks
fed-batch bioreactors
process optimizations
process control
Optimization and control of feb-batch fermentation processes by using artificial neural systems
[Tarragona] :
Universitat Rovira i Virgili,
2011
Accés lliure
http://hdl.handle.net/10803/8514
cr |||||||||||
AAMMDDs2011 sp ||||fsm||||0|| 0 eng|c
8468832545
Valencia Peroni, Catalina,
autor
Tesi
Doctorat
Universitat Rovira i Virgili. Departament d'Enginyeria Química
2002
Universitat Rovira i Virgili. Departament d'Enginyeria Química
Tesis i dissertacions electròniques
Giralt, Francesc,
supervisor acadèmic
Giralt i Marcé, Jaume,
supervisor acadèmic
TDX
This work focuses on the application of neural networks in the areas of modelling, identification, control and optimization of biothechnology processes, mainly fed-batch bioreactors. The basic ideas and techniques of artificial neural networks are presented with the notation familiar to control engineers. The applications of a variety of neural network architectures in control and control schemes are first surveyed. Some especific fed-batch bioreactor processes are mentioned to illustrate particular control cases to be examimined in detail and solved. Especifically, a non-linear multivariable bioreactor control problem is used as a case study for model based control techniques. An implementation of direct and inverse process control models based on neural networks that considers biological, thermal and pH effects for this multivariable fed-batch bioreactor is performed and tested. Multilayer perceptrons and radial basis functions neural networks are considered to model this type of non-linear multi-input multi-output (MIMO) dynamic process. The direct models are successfully tested under steady state, dynamic process operation and when a acid disturbance in the process causes a plant/model mismatch. The inverse process model is also successfully tested at the set-point input with a random series of perturbations around the plant operation state. The RBF architecture with goal 3.0 is the best architecture for the direct model of this multivariable process while the best inverse model is based on a MLP 19-11-7-1 trained including past information of the steady states of the process. <br/>On the other hand, optimal control techniques that employ neural networks are studied to optimize the production of invertase in a fed-batch bioreactor. The controlled addition of substrates is used in this bioreactor process to increase productivity when end-product inhibition or catabolite repression are present. Cloned invertase production in Saccharomyces cerevisiae yeast is carried out in fed-batch mode of operation because the enzyme expression is repressed at high glucose concentrations. An optimal glucose feed rate profile is needed to achieve the highest fermentation profit. The controller has to find at each time step an optimal control action that increments the fed-batch bioreactor profitability, even when a disturbance or a set-point change arise. This optimal control action increases the productivity and, within the same optimization process, finds the optimal fermentation ending time. This double optimization is a novelty not met by previous optimization schemes published in the literature. A neuro dynamic programming (NDP) approach coupled with MLP neural networks or fuzzy ARTMAP systems is employed to accomplish these optimization objectives. Fuzzy ARTMAP creates multidimensional category maps by incremental supervised learning. The optimization method utilizes suboptimal control policies as a starting guess. The neural networks are used to build a cost surface in the state space visited by the process. Bellman's iteration is used to improve the cost approximation. The cost surface obtained is implemented into a control system. The controller is tested for different fermentation processes started with different initial fermentation volumes. NDP outperforms other optimization methods employed to find an optimal feeding profile. Besides, it can be used to optimize any fermentation process (starting at different initial conditions) because the future costs (profits) are characterized as a function of system states. The optimal control trajectories found by the controller are similar to the best suboptimal policy for each initial volume. MLP-NDP controllers yield the highest profits, but the manipulated variable trajectories are not smooth. Fuzzy ARTMAP-NDP overcomes this limitation. The best fuzzy ARTMAP-NDP based control system is also tested when an abrupt death of yeast cells occurs. In this case, the controller performance is better than the performance of the fermentation using the best suboptimal policy for the given initial volume.<br/>The integration of control science with neural networks in a unified presentation and identification key areas is a path to follow in future research. Artificial neural networks techniques can be succesfully applied to control fed-batch bioreactors.
r
ES-BaCBU
cat
rda
ES-BaCBU
text
txt
rdacontent
informàtic
c
rdamedia
recurs en línia
cr
rdacarrier