Robust Abnormality Diagnosis Model Considering Nuclear Power Plants Data Trends

Abstract

The development of artificial intelligence applications in the real world has been driven by enhancements in computer performance, the accessibility of big data, and advancements in deep learning algorithms. For instance, artificial intelligence in the autonomous driving holds the potential to enhance driving convenience. However, accidents can occur when artificial intelligence encounters differences between training and testing environments such as handling complex road or extreme weather conditions. Even in the nuclear field, applying abnormality diagnosis models using artificial intelligence to real nuclear power plants may lead performance degradation due to data differences between the simulator and real nuclear power plant variables. Therefore, we aim to develop an abnormality diagnosis model that trains on simulator data and tests on real plant data using neural networks and fuzzy logic. Conventional artificial intelligence models excel at responding to microscopic data changes, but they are vulnerable to macroscopic changes in test cases that deviate from training situations. To overcome these limitations, we use fuzzy logic that considers data trends in plant variables. Simulation plant variables are transformed into input values representing the degree of change, and input values are passed into fuzzy membership functions like decrease, maintenance, and increase. Consequently, the plant variables are converted into different types of variables that represent how much has changed compared to the past by fuzzification. Fuzzy values enhance robustness of an abnormality diagnosis model on real power plant data with data differences, even when trained solely on clear simulation data.