NEURAL-NETWORK FRAMES - APPLICATION TO BIOCHEMICAL KINETIC DIAGNOSIS

The diagnosis of chemical kinetics in chemical plants can be viewed as a process of classification. Recorded data can be associated with dif ferent types of kinetic models and the type of kinetics can be classif ied by comparison with previously recorded data. A new frame for a neu ral network (NN) is proposed in order to carry out the classification. The potentialities of this adaptive, hierarchized frame organized as a ''polystochastic'' model have been investigated here. The underlying approach is based on the use of distances between two paths of observ ed kinetic data. A matrix of distances results from a set of possible kinetic models, and algorithms for classifying models are developed us ing this matrix. Another type of distance, an informational type, is p roposed between two matrices of distances so as to compare one classif ication with another or with a reference classification. Training the net by methods based on informational criteria is proposed and tested. By a fast adaptive procedure, the small number of resulting weights a re adjusted to account for reference cases. The utility of the net is illustrated via the kinetic modeling of a fermentation process. A comp arison with another conventional net is also made.