Robust diagnosis of abnormal events using fuzzy-based feature extraction in nuclear power plants

Abstract

Diagnosis of abnormal events in nuclear power plants is important for ensuring operational safety and preventing human errors. Conventional deep learning-based diagnosis models often fail to ensure robust performance due to data discrepancies between simulator training data and actual plant data. To overcome data discrepancies, we enhanced the robustness of an abnormality diagnosis model through two key approaches. First, we developed a fuzzy-based feature extraction method to handle ambiguity and uncertainty between simulator and plant data by focusing on relatively long-term trends within time-series data rather than short-term data values. Second, we implemented a shallow-layer deep learning model to minimize overfitting on simulator data. Although deeper models may improve training accuracy, they do not always perform better in real-world applications. To evaluate our approach, we employed a synthetic plant dataset that reflects the discrepancies between simulator training data and actual plant data. The proposed model was trained solely on simulator data and tested on synthetic plant data. Experimental results show that the proposed model achieves a diagnostic accuracy of 99.7 %, while conventional models decline as data discrepancies increase. These outcomes illustrate the potential of our approach to improve the safety and reliability of nuclear power plant operations.