Soft computing tools for transient classification

Any action taken on a process, for example in response to an abnormal situation or in reaction to unsafe conditions, relies on the ability to identify the state of operation or the events that are occurring. Although there might be hundreds or even thousands of measurements in a process, there are generally few events occurring. The data from these measurements must then be mapped into appropriate descriptions of the occurring event(s), which in most cases is a difficult task. A systematic study was carried out with the aim of comparing alternative neural network designs and models for performing this mapping task. Four main approaches have been investigated: radial basis function (RBF) neural networks and cascade-RBF neural networks combined with fuzzy clustering, self-organizing map neural networks, and recurrent neural networks. The main evaluation criteria adopted were: identification accuracy, reliability (i.e. correct recognition of an unknown event as such), robustness (to noise and to changing initial conditions), and real time performance. Additionally, in this paper we describe how ensembles of recurrent neural networks can overcome some of the limitations encountered in these early prototypes, and give an example involving the identification of anomalous events in a PWR 900 MW nuclear power plant.