Neural Network Software for the Automatic Diagnostics of Nuclear Power Plants’ Components

Read full text Summarise with VIRAT
Authors: G. I. Sharaievskyi
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In this work, the state of existing automated systems for the management of technological processes at nuclear power plants was analyzed, and use of neurocomputing approach to recognition of spectral implementations of stochastic diagnostic signals was proved, which defines actual technical condition of certain NPP equipment. As it is known, the essence of the problem lies in the fact that modern NPP monitoring and control systems, which are the part of computer-aided manufacturing control systems (CAMCS), are based on a deterministic approach to the logical analysis of equipment operating conditions to prevent controlled parameters from exceeding predefined safe limits. It is obviously not enough for reliable prevention of damage that may occur in the main equipment of nuclear power units. In fact, some NPP parameters that are not directly controlled by CAMCS, particularly temperature conditions of reactor heat-exchange devices, may significantly exceed permissible values during operation that can lead to severe damage, for example, heat-exchange crisis causing burnout that may spread throughout the entire nuclear core. To prevent such catastrophic events, it would be necessary to measure local temperature of fuel elements (in reactor about 40,000 practically uncontrolled fuel elements are located). This would require hundreds of thousands of thermocouples on their surfaces, which is technologically impossible. Therefore, the problem of diagnosing abnormal heat-exchange regimes can be solved not by direct temperature measurements of the coolant at the nuclear core inlet as it is now, but by recognizing the spectral characteristics of dynamic components of certain reactor parameters. These include, first of all, signals of dynamic coolant pressure and signals from neutron flux detectors located directly within the reactor core. In this work the approach to building SOM-neural networks regarding the tasks of accidental objects recognition is reviewed. Modified learning algorithm for the SOM-neural structure recognition is proposed in condition of no a priori information on the number of classes to be recognized. This approach is implemented based on the determination of disorder moment in random time series using the autoregressive model.

Fields of Study & Subjects

Fault Detection and Control Systems Neural Networks and Applications Machine Fault Diagnosis Techniques

Publication details

Journal / source
Nuclear Power and the Environment
Volume
34
Issue
1
Pages
31–44
Publication date
1 January 2026
Language
EN