A fundamental requirement on data to be used for the training of an empirical diagnostic model is that the data has to be sufficiently similar and consistent with what will be observed during on-line monitoring. In other words, the training data has to 'cover' the data space within which the monitored process operates. The coverage requirement can quickly become a problem if the process to be monitored has a wide range of operating regimes leading to large variations in the manifestation of the faults of interest in the observable signal transients. In this paper we propose a novel technique, based on neural networks, aimed at reducing the variability of fault manifestations through a process of 'intelligent normalisation' of transients. The paper includes the application of the proposed method to a nuclear power plant transient classification case study.
Keywords: data preprocessing, fault diagnosis, early fault detection, neural networks, transient classification, nuclear power plants, nuclear energy, intelligent normalisation, training, process monitoring
The IJNKM is a peer-reviewed journal dedicated to the publication of research articles, review papers and technical notes in all domains related to improvement of the peaceful uses of nuclear energy and the development of nuclear sciences- and tec » Read more