Application of Hierarchical Agglomerative Clustering to Create a Library of Basic Classes of Acoustic Events of the VVER-1000/1200

Abstract

In water-water power reactors (VVERs), loose, weakly fixed, and foreign objects can appear in the main circulation circuit (MCC) and pose a threat to the integrity of the equipment and the safety of the reactor. In order to detect these items early, the reactor is equipped with a loose parts monitoring system (LPMS). In addition to the detection of loose/weakly fixed parts, the LPMS functions include the classification of reported events. The possibility of applying the classification algorithm is based on the fact that the signals from the triggering of standard equipment are characterized by a high degree of repeatability, even in the presence of noise, while a loose object is characterized by a large stochastic component, and its own deterministic class cannot be formed for it. Classification reduces false alarms and allows signals from routine operations to be detected; the signals from the same process should be assigned to the same class. The idea of the article is to train the LPMS on some archive of data characterizing routine reactor operation, to create a library of basic classes, and to set the boundaries of each class in order to, on one hand, take into account the possible variability of signal parameters due to noise. Having defined the basic classes, we can assert that, if a newly received signal falls into one of the classes, it reflects routine reactor operation, while signals that do not fall into any of the classes can be from the appearance of a loose/weakly fixed item. The article analyzes a set of events accumulated in the archive of one of the operating LPMS. Their clustering is performed, as a result of which classes of events corresponding to routine unit operations are singled out. For each class, the center of the class and permissible limits of deviations from the center are calculated. All obtained class centers are benchmarks by which the LPMS in real time either classifies a newly detected event or characterizes it as unclassified.