Neural network modelling and control of bioprocesses

Mei-Jywan Syu, Purdue University


This dissertation aims to develop neural networks for bioprocess systems. It is shown that back-propagation neural networks (BPNNs) with a saturation-type transfer function can be successfully used in simulation/prediction, system identification, and adaptive control. A function different from the commonly used sigmoid transfer function, namely the saturation-type function, has been employed for the first time ever in neural network modeling in this dissertation. It was first developed for the BPNNs, and subsequently applied to a series of multiple component adsorption systems and a fermentation system. This function shows results superior to those obtained with the sigmoid function inasmuch as it leads to precise simulation, fast convergence, and few required data inputs. BPNNs were first applied to a series of multicomponent adsorption systems. This new approach could successfully simulate and predict adsorption isotherms. Comparison of the BPNNs with the Langmuir isotherms for these adsorption systems was made. Interactions among the compounds in these systems were identified. A novel approach was then developed by implementing the BPNN to model the growth patterns of cells from the initial batch conditions. This model demonstrates tremendous capability in extrapolated as well as interpolated prediction of cell growth. Sensitivity analysis of this system with respect to the initial conditions was performed and the optimal levels of these initial factors were obtained. The dynamic identification of the MIMS/FIA/2,3-butanediol batch fermentation was studied by a neural network. From the previous and current concentrations of the products and the oxygen composition in the gas supply, the neural network is able to perform one-step ahead prediction of the oxygen composition. This dynamic neural network has also the capability in few-step ahead prediction under the existence of process fault. From such identification, the control command could be predicted directly from the model without coupling the process model with the controller model. By implementing the above idea, an efficient neural network predictive adaptive controller was developed for a fed-batch fermentation system. The previous history, which is less important for the current control is erased as the batch progresses. With this neural network adaptive control strategy, the specific growth rate was successfully maintained at a predetermined value.




Tsao, Purdue University.

Subject Area

Bioengineering|Chemical engineering|Artificial intelligence

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server