Unsupervised anomaly detection of nuclear power plants under noise background based on convolutional adversarial autoencoder combining self-attention mechanism

•This study involves an anomaly detection problem of NPPs under a noisy background, which is rarely studied in previous research.•A novel unsupervised anomaly detection method (CAAE-SA) is designed for solving UAD-NPPs.•The self-attention mechanism is proposed to improve the robustness of CAAE-SA.•E...

Full description

Saved in:

Bibliographic Details
Published in:	Nuclear engineering and design Vol. 428; p. 113493
Main Authors:	Sun, Xiang, Guo, Shunsheng, Liu, Shiqiao, Guo, Jun, Du, Baigang
Format:	Journal Article
Language:	English
Published:	Elsevier B.V 01.11.2024
Subjects:	Adversarial convolutional autoencoder Noise background Nuclear power plants Self-attention mechanism Unsupervised anomaly detection Noise background Self-attention mechanism Nuclear power plants Adversarial convolutional autoencoder Unsupervised anomaly detection
ISSN:	0029-5493
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	•This study involves an anomaly detection problem of NPPs under a noisy background, which is rarely studied in previous research.•A novel unsupervised anomaly detection method (CAAE-SA) is designed for solving UAD-NPPs.•The self-attention mechanism is proposed to improve the robustness of CAAE-SA.•Experiment results show that CAAE-SA is superior to the state-of-the-art anomaly detection methods, which can effectively reduce instances of false alarms and failed alarms. Anomaly detection of nuclear power plants (NPPs) is critical to maintaining efficient operations and preventing catastrophic failures. Most existing research about anomaly detection of NPPs assumes that the accident data are sufficient and disregards the influence of the noise, which is not consistent with reality. Thus, this paper presents a novel network based on a convolutional adversarial autoencoder combining self-attention mechanism (CAAE-SA) for solving unsupervised anomaly detection of NPPs (UAD-NPPs) under noisy background. An adversarial learning mechanism is designed to overcome the problems of insufficient feature representation capability and low quality of sample generation in UAD-NPPs by learning more robust and discriminative latent feature representations. A specific prior distribution is introduced in the training process to differentiate between the distributions of abnormal and normal samples in the feature space. Meanwhile, the self-attention mechanism is introduced into the CAAE-SA, which captures the global relationships between measurements to improve the robustness of CAAE-SA in noisy environments. Finally, anomaly detection is realized based on the reconstruction error of the real-time operation data. The experimental results based on the Fuqing Unit 2 full-scale simulator show that the proposed model achieves the highest F-score under different noise degrees compared to other unsupervised methods. Additionally, the CAAE-SA can quickly identify the operation state within 0.04 s, meeting real-time demands. This capability offers a foundation for accident classification and subsequent rescue efforts.
ISSN:	0029-5493
DOI:	10.1016/j.nucengdes.2024.113493