Mathematical model of automatic recognition of latent pre-emergency conditions of shaft mechanical sealing systems of NPP main circulation pumps

In order to improve the operational safety of nuclear power plants (NPP) and the support systems for operating personnel of nuclear power units and to exclude possible errors due to human factors, as part of the developed automatic systems for operating diagnostics of the current technical conditions of the main equipment (in particular, the main aggregate modes of the main aggregate circulation pumps), an additional subsystem for the trend of changes in this state was created. This subsystem is designed to enable the early detections of latent initial phases of potentially dangerous violations of the normal course of physical processes in the most vulnerable modes of pumping units and the timely prevention of the development of identified deviations and anomalies in emergency situations. As it is known the essence of the problem consists of that up-todate NPP monitoring-and-control systems being the part of NPP computer-aided manufacturing control systems (CAMCS) have in their base a deterministic approach to logistic analysis of equipment operating conditions to prevent controlled by them parameters from the falling outside preliminary sate safe limits. In this work the approach to building SOM-neural networks regarding the tasks of accidental objects recognition is reviewed. Modified algorithm of study in the recognizing SOMneural structure is proposed in condition of absence of a priori information on the power of classes multitude tobe-recognized. In this article the approach to the training and automated adaptation of diagnostic in conditions of a priory uncertainty of many classes to be recognized is proposed. This approach is implemented on the basis of determination of the moment of disorder of random time series using the auto-regressive model. In this work the approach to the training and automated adaptation of diagnostic neuro-networking structure on the basis of Kohonen's topology in conditions of a priors uncertainty of many classes to be recognized is proposed. This approach is implemented on the basis of determination of the moment of disorder of random time series using autoregressive mode. (author)