RAPID PROCESS MODELING - MODEL-BUILDING METHODOLOGY COMBINING UNSUPERVISED FUZZY-CLUSTERING AND SUPERVISED NEURAL NETWORKS

A modeling methodology is suggested aimed to deal with new processes, particularly with process uncertainties, non-linearity and knowledge i nsufficiency. The model architecture is a modular neural network The d ata space is partitioned into several overlapping domains and a neural -network in each domain maps input-output relations. Each data feature vector has a membership value related to every 'local' neural-net. Th e total output is a weighted sum of the local networks outputs. The do mains are defined by unsupervised fuzzy clustering procedure. The mode l architecture enables efficient learning and tuning. Defining the opt imal number of domains (i.e. clusters) is of great importance. The eff ectiveness of several cluster validity measures was compared with the generalization capability of the model and the information criteria su ggested by Akaike. The validity measures were tested with data obtaine d from fermentation model simulations and a fermentation of yeast-like fungus Aureobasidium pullulans. (C) 1998 Published by Elsevier Scienc e Ltd. All rights reserved.