Early fault detection method for nuclear power plants based on sparse denoising autoencoder and kernel principal component analysis

•A data-driven fault detection method is proposed, which can accurately detect early faults in nuclear power plants.•The proposed grouping strategy effectively overcomes the insensitivity of single models to early faults.•The combination of SDAE and KPCA demonstrates good feature extraction and nonl...

Full description

Saved in:
Bibliographic Details
Published in:Annals of nuclear energy Vol. 220; p. 111460
Main Authors: Yin, Wenzhe, Xia, Hong, Huang, Xueying, Shan, Longfei, Ran, Wenhao, Jia, Zhujun
Format: Journal Article
Language:English
Published: Elsevier Ltd 15.09.2025
Subjects:
Early fault
Fault detection
Kernel principal component analysis
Nuclear power plant
Sparse denoising autoencoder
Fault detection
Nuclear power plant
Kernel principal component analysis
Early fault
Sparse denoising autoencoder
ISSN:0306-4549
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:•A data-driven fault detection method is proposed, which can accurately detect early faults in nuclear power plants.•The proposed grouping strategy effectively overcomes the insensitivity of single models to early faults.•The combination of SDAE and KPCA demonstrates good feature extraction and nonlinear data processing capabilities. Effective fault detection technology is of great significance to the safety and economy of nuclear power plants (NPPs). To accurately identify early faults in NPPs, this study proposes a novel fault detection method based on sparse denoising autoencoder (SDAE) and kernel principal component analysis (KPCA). First, the operating data of NPPs is collected by numerous sensors, and the operating parameters are grouped according to physical properties. Then, the corresponding fault detection model is established according to each parameter group, and each detection model consists of the SDAE and KPCA. The case study evaluated four accident scenarios (LOCA, SLBIC, FHAIC, FHAIAB) across two development degrees (0–1 % and 0–0.1 %). The proposed method achieved fault detection rates of 99.07 %, 95.20 %, 99.73 %, and 99.60 % for the 0–1 % degree with zero false alarms. Even for the subtler 0–0.1 % degree, it maintained a 94.84 % average detection rate and no false alarms. Compared to traditional methods, its average fault detection rate was higher than that of PCA and KPCA by 62.9 % and 32.4 % (0–1 % degree), and by 89.5 % and 88 % (0–0.1% degree), demonstrating its potential application value in NPPs.
ISSN:0306-4549
DOI:10.1016/j.anucene.2025.111460