Recurrent neuro-fuzzy system for fault detection and isolation in nuclear reactors

This paper presents an application of recurrent neuro-fuzzy systems to fault detection and isolation in nuclear reactors. A general framework is adopted, in which a fuzzification module is linked to an inference module that is actually a neural network adapted to the recognition of the dynamic evolution of process variables and related faults. Process data is fuzzified in order to reason rather on qualitative than on quantitative values. The fuzzified attributes feed the neural network. Two different network topologies are tested over data simulated by a commissioned simulator of a nuclear reactor: a feed-forward topology and a recurrent topology, where the additional network inputs are considered as delayed activation of output units. The later approach shows better generalization performance for the detection and isolation of a number of security related faults. A graphic interface presents a qualitative representation of symptoms and diagnostic results by colored shades that evolve with time allowing a friendly and efficient communication with operators in charge of the process security.