Neural and genetic-based approaches to nuclear transient identification including 'don't know' response

Abstract This paper presents two approaches based on different artificial intelligence techniques (AI), genetic algorithms (GA) and neural networks (NN), for pattern recognition on the set of transient reactor signals. The first approach describes a possibilistic analysis based on the minimum centroids set method optimized by genetic algorithm. The use of a possibilistic classification provides a set of natural and consistent classification rules leading naturally to a good heuristic, based on the cluster size, to handle unlabeled transients by providing a ‘don't know’ response. In situations where the safety is the decisive parameter, this characteristic is fairly desirable in transient classification systems, since wrong or unreliable classifications can be catastrophic. Application of the proposed approach to a nuclear transient identification problem shows good capability of the genetic algorithm in learning optimized possibilistic classification rules for efficient diagnosis including ‘don't know’ response. The second approach uses two multilayer neural networks trained with the backpropagation algorithm. The first NN does a preliminary transient type identification using as inputs a short set (in a moving time window) of recent measurements of each variable. These inputs were presented to the NN, for each variable, as a set of consecutive values in temporal sequence. The second NN is used to validate the preliminary identification (from the first NN) through a self associative neural simulator used to estimate the value of every state variable. In order to validate both methods, a nuclear power plant (NPP) transient identification problem, comprising postulated accidents, is proposed. Noisy data was used in order to evaluate the method robustness. The results obtained reveal the ability of the methods in dealing with both dynamic identification of transients and correct ‘don't know’ response.