A NEURAL-NETWORK STUDY ON THE DYNAMIC IDENTIFICATION OF A FERMENTATION SYSTEM

The objective of this paper is to propose neural networks for the stud y of dynamic identification and prediction of a fermentation system wh ich produces mainly 2,3-butanediol (2,3-BDL). The metabolic products o f the fermentation, acetic acid, acetoin, ethanol, and 2,3-BDL were me asured on-line via a mass spectrometer modified by the insertion of a dimethylvinylsilicone membrane probe. The measured data at different s ampling times were included as the input and output nodes, at differen t learning batches, of the network. A fermentation system is usually n onlinear and dynamic in nature. Measured fermentation data obtained fr om the complex metabolic pathways are often difficult to be entirely i ncluded in a static process model, therefore, a dynamic model was sugg ested instead. In this work, neural networks were provided by a dynami c learning and prediction process that moved along the time sequence b atchwise. In other words, a scheme of two-dimensional moving window (n umber of input nodes by the number of training data) was proposed for reading in new data while forgetting part of the old data. Proper size of the network including proper number of input/output nodes were det ermined by trained with the real-time fermentation data. Different num ber of hidden nodes under the consideration of both learning performan ce and computation efficiency were tested. The data size for each lear ning batch was determined. The performance of the learning factors suc h as the learning coefficient eta and the momentum term coefficient al pha were also discussed. The effect of different dynamic learning inte rvals, with different starting points and the same ending point, both on the learning and prediction performance were studied. On the other hand, the effect of different dynamic learning intervals, with the sam e starting point and different ending points, was also investigated. T he size of data sampling interval was also discussed. The performance from four different types of transfer functions, x/(1 + \x\), sgn(x) . x(2)/(1 + x(2)), 2/(1 + e(-x)) - 1, and 1/(1 + e(-x)) was compared. A scaling factor b was added to the transfer function and the effect of this factor on the learning was also evaluated. The prediction result s from the time-delayed neural networks were also studied.