Thème :
Neural and Hybrid Neural Modeling of a Yeast Fermentation Process.
Présentation :
The growth rate of the micro-organisms in biological reactors is described by some complicated kinetic expressions, and the modeling of such reactors is offering many challenges to biochemical engineers. In this paper, two modeling
approaches are described and compared.We use the artificial neural network (ANN) and the hybrid artificial neural network (HANN).The models development is carried out through twosteps procedure consisting of estimation step (training) and validation step (prediction). The ANN’s fail the validation, but the HANN’s are good interpolators even for fewer training data.
The weights of the networks are tuned using the Levenberg Marquardt
opitimization procedure. The simulations employ data from a baker’s yeast production in a fedbatch fermenter and the results show that the HANN technique is a powerful tool for bioprocess modeling, especially when only limited theoretical knowledge of this process is available. Copyright c ° 2008 Yang’s
Scientific Research Institute, LLC. All rights reserved.
Thème :
Logique floue
Présentation :
Fuzzy Estimation of a Yeast Fermentation
The dynamics of fermentation processes are very complex and not completely known. Some state variables are nonmeasurable, and the process parameters are strongly time dependent. Recently, there are some control methods like fuzzy learning and neural networks, which are promising in dealing with non-linearity, complexity, and uncertainly of these processes. These methods are suitable for the modelling of these systems, which are difficult to describe mathematically. The fuzzy learning methods are useful for the modelling, they are less demanding on the mathematical model and a priori knowledge about the processes. Different techniques for estimating the state variables (that are non-measurable) in the fermentation process have been investigated. A non-linear auto-regressive with exogenous input (NARX) model was developed using process data from a pilot bioreactor. The fermentation process is splitted into three phases, where each phase was treated separately. Generally, fuzzy models have a capability of dividing an input space into several subspaces (fuzzy clustering), where each subspace is supposed to give a local linear model. In our work, we used global learning where the local models are less interpretable, but the global model accuracy is satisfying, and the fuzzy partition matrix is obtained by applying the Gustafson-Kessel algorithm. The fermentation parameters are estimated for a batch and a fed-batch culture. The number of inputs to our fuzzy model are three for a first simulation. We used four inputs for a second simulation, in order to detect some correlations among inputs. The results show that estimated parameters are close to the measured (or calculated) ones. The
parameters used in the computation are identified using batch experiments.