Temporal sequence learning with neural networks for process fault detection

This paper deals with the learning and recognition of simple temporal sequences with artificial neural networks. Two solutions to represent time in neural networks are presented: the spatial representation and the dynamic representation. The model proposed, which uses the dynamic representation of time, is based on recurrent connections. By adding to each input neuron a self-recurrent connection it is proved that a network composed of such input neurons and binary output neurons is able to learn perfectly simple temporal sequences. The learning algorithm is based on a gradient descent of an error defined for binary neurons. This neural network is used for process fault detection and particularly in manufacturing systems in which sequence recognition is done both on the nature of the events and on their temporal constraints