A fault tolerant controller based on neural nets

Describes a method of using neural nets of the multi-layer perceptron type in the task of controlling dynamical systems governed by ordinary differential equations which are subjected to a finite set of faults. The set of faults are assumed to be known a priori, although the event itself is considered to occur in a random manner. The proposed controller has the capability to be trained in order to match a reference model, such as a linear second order system. At the occurrence of a fault, the controller is reconfigured by altering the weight values of the synaptic connections according to informations provided by the level of hormones, which is, in the case of animals, a chemical messenger. In this work, the effect of a hormone is simulated by broadcasting a digital signal, indicating the occurrence or not of a fault