Neural network based adaptive control via temporal pattern recognition

Abstract

This paper presents a neural network approach to adaptive control through pattern recognition techniques. Two interconnected backpropagation networks are trained to translate error patterns resulting from sustained set point changes into predictions of mismatch between current internal model parameters, model gain and model time constant, and those which restore desired performance. The network predictions are then used to update a model based PI controller. The strategy is demonstrated on two simulations and a pilot scale process which are undergoing severe changes in model gain and time constant. The strategy compares favorably against a more traditional rule based pattern recognition approach. On presente dans cet article une approche par reseau neuronal du controle adaptatif au moyen de techniques de reconnaissance des modeles. Deux eeseaux de retropropagation interconnectes sont entrainees a traduire les types d'erreurs resultant de changements constants des points de consigne en prediction de discordance entre les parametres de modele internes courants, le gain de modele et la constante de temps de modele, et ceux qui retablissent la performance desiree. On utilise ensuite les predictions du reseau pour mettre a jour un controleur PI base sur les modeles. Cette strategie est demontree a partir de deux simulations et d'un systeme a l'echelle pilote qui subissent des changements severes dans le gain et la constante de temps de modele. Cette strategie se compare favorablement a une approche de reconnaissance des modeles a partir des regles traditionnelles.