UTILISATION DES RESEAUX DE NEURONES ARTIFICIELS DE TYPE RBF POUR LA MODELISATION DU REGIME NORMAL A POINT DE FONCTIONNEMENT VARIABLE D’UNE INSTALLATION INDUSTRIELLE

Abstract

This work presents the development of a mathematical model based on stochastic artificial neural networks type RBF (Radial Basis Function) for modeling the normal mode at a variable point of functionning of an industrial installation. The studied industrial facility is a distillation column of methylcyclohexane (C6H11-CH3) from toluene-methylcyclohexane mixture (C6H5-CH3 / C6H11-CH3) which was defined in the mass composition by 23% in methylcyclohexane. Neuronal architecture proposed for the modeling of this system consists of an input layer containing seven neurons, a hidden layer containing nine neurons and an output layer having a single neuron. The hidden layer is activated by a Gaussian function whose center is determined by using the K-means algorithm; however the output layer is activated by a linear function. Regarding existing weight between the hidden layer and the output layer, they are determined by the back-propagation algorithm of the error gradient. The RBF neural architecture type so determined was validated on a new database and has achieved better results compared to conventional methods.