The Application of Neural Sensors to Fermentation Processes

This paper describes the training and implementation of Artificial Neural Network models for the real-time, on-line estimation of key biological variables in a fed-batch yeast fermentation process. The neural networks are generic nonlinear models that are configured and trained to act as software sensors for biomass concentration.In order to successfully customise a neural sensor it is necessary to give considerable attention to the choice of network inputs. The task of choosing these inputs is somewhat eased by knowledge of the yeast’s metabolism supplemented where necessary by a statistical analysis of the available inputs and outputs. In this study, the inputs that were chosen reflected both the yeast’s metabolic activity and its rate of growth.The experimental investigation was performed using a Ferranti Process Control Computer connected to a pilot-scale fermentation suite. The neural sensor module was written such that the customised neural network (sensor), configured and trained offline, could be implemented on-line using the Ferranti PML sequence language. The training and configuration stages of the neural sensor were conducted using software written in-house and designed to be compatible with the Ferranti based neural sensor module.