Fault identification and diagnosis based on KPCA and similarity clustering for nuclear power plants. (January 2021)

• Record Type: Journal Article
• Authors: Wang, Hang; Peng, Min-jun; Yu, Yue; Saeed, Hanan; Hao, Cheng-ming; Liu, Yong-kuo
• Is Part Of: Annals of nuclear energy. Volume 150(2021)
• Publication Date: 2021-01
• Subjects: Anomaly detection -- Fault diagnosis -- Feature extraction -- Kernel principal component analysis -- Clustering; Nuclear energy -- Periodicals; Nuclear engineering -- Periodicals; 621.4805
• DOI: 10.1016/j.anucene.2020.107786
• Languages: English
• ISSNs: 0306-4549

Abstract

Highlights: The KPCA is utilized for distinguishing system-level malfunction from an anomaly in sensors themselves. Feature extraction is done by KPCA and compared with other manifold learning methods. Different sources of data are used to prevent overfitting. Similarity clustering is combined to verify the results of SVM and assess the severity of malfunctions. Abstract: Nuclear power plant is a complex engineering system with strong coupling, nonlinearity and potential radioactive release risk. Different researchers have proposed numerous fault diagnosis techniques but there is a gap between these techniques and their practical manifestation. In this paper, kernel principal component analysis (KPCA) and similarity clustering are primarily presented for fault diagnosis. First, KPCA is utilized for anomaly detection to distinguish actual faults form abnormal sensor readings. After that, support vector machine is carried out for fault diagnosis. Subsequently, KPCA is also used for feature extraction before clustering algorithms for analyzing fault type and degree. As opposed to other 'black box' data-driven methods, this technique allows the results to be illustrated in a visual form which greatly enhances the interpretability of diagnosis results. Finally, the accuracy of the method is verified with a full scope NPP (Type: Pressurized Water Reactor) simulator.