Optimal grouping of signals for condition monitoring of nuclear components

Condition monitoring is essential to timely detect abnormal operation conditions in critical Component Structure System (CSS). It can be addressed by developing a (typically empirical) model of the CSS behaviour in normal conditions; during operation, the behavior actually observed in the plant is compared with the behavior predicted by the model representing the CCS in normal conditions: a deviation between the measured and predicted values of the CCS signals reveals the presence of an abnormal condition, e.g. caused by components or instrumentation faults. The present work consider a real case study concerning the condition monitoring of a reactor coolant pump of a Pressurized Water Reactor. Since the performance of a single model monitoring all the 94 signals measured by the sensors and related to the component operational conditions has resulted to be not satisfactory, the possibility of splitting the signals into subgroups and then building a specialized model for each subgroup has been investigated. Different methods to subdivide the signals in groups have been investigated in this thesis work. The comparison of the performances of the different signal groupings has been made with respect to metrics measuring the accuracy, i.e. the ability of the model to correctly and accurately reconstruct the signal values when the plant is under normal operation, and the robustness, i.e. the model ability to reconstruct the signal values in case of abnormal operation and consequent anomalous behavior of some monitored signals. The obtained results have shown that the subdivision of the signals in groups formed by highly correlated signals permits to achieve satisfactory accuracy in the signal reconstruction, whereas the use of groups formed by many signals results in very robust signal reconstructions. Genetic algorithms have been successfully used to find an optimal signal grouping which is a compromise solution between accuracy and robustness.