Abstract
Air-operated valves (AOVs) are used to control or shut off the flow in the nuclear power plants. In particular, the failure of safety-related AOV could have significant impacts on the safety of the nuclear power plants and therefore, their performances have been tested and evaluated periodically. However, the current method to evaluate the performance needs to be revised to enhance the accuracy and to identify defects of AOV independently of personal skills. This paper introduce the ANN (Artificial Neural Network) model to diagnose the performance and the condition altogether.
Test facilities were designed and configured to measure the signals such as supply pressure, control pressure, actuator pressure, stem displacement and stem thrust. Tests were carried out in various conditions which simulate defects with leak/clogged pipes, the bent stem and so on. First, the physical models of an AOV are developed to describe its behavior and to parameterize the characteristics of each component for evaluating the performance. Secondly, CNN (Convolutional Neural Network) architectures are designed considering the developed physical models to make a lead to the optimal performance of ANN. To train the ANN effectively, the measured signals were divided into several regions, from each of which the features are extracted and the extracted features are combined for classifying the defects. In addition, the model can provide the parameters of maximum available thrust, which is the key factor in periodic verification of AOV with the required accuracy and classify more than 10 different kinds of defects with high accuracy.
A Constraint-Based Random Search Algorithm for Optimizing Neural Network Architectures and Ensemble Construction in Detecting Loss of Coolant Accidents in Nuclear Power Plants