摘要
为利用核动力系统产生的大量时序数据,自动、高效检测出系统动态运行过程中存在的各类异常,本研究提出了一种面向核动力系统多元时序数据的无监督异常检测方法NadGAN,以提高微小异常和新异常的识别率,避免其发展为严重事故。利用门控循环单元(GRU)作为生成对抗网络(GAN)中生成器和判别器的基础模型,提取时间依赖关系,通过对抗训练学习正常数据特征分布。在异常检测阶段,采用自适应阈值方法综合考虑判别误差和重构误差进行异常判定,避免手动调参。在6个仿真数据集上与4种基线方法进行对比实验,结果表明NadGAN性能优于其他对比方法,证明了将所提出方法运用于核动力系统异常检测的有效性和可行性。
To automatically and efficiently detect various abnormalities of the nuclear power system by utilizing the abundant time series data from nuclear power system,this study proposes an unsupervised anomaly detection method(NadGAN)based on nuclear power multivariate time series data.It aims to improve the identification accuracy of both minute and novel anomalies,and avoid them from developing into serious accidents.NadGAN applies Gated Recurrent Unit(GRU)as the fundamental model for both the generator and discriminator in a Generative Adversarial Network(GAN)framework,which is able to extract temporal dependencies and learn the feature distribution of normal data through adversarial training.In the anomaly detection stage,an adaptive thresholding approach is adopted to comprehensively consider the discriminative error and reconstruction error for anomaly identification without manual parameter tuning.Comparative experiments with four baseline methods on six simulated datasets demonstrate that NadGAN outperforms other methods,which further validates the effectiveness and feasibility of applying the proposed method for anomaly detection in nuclear power systems.
作者
万静意
艾庆忠
曾辉
欧阳泽宇
唐雷
赵欣
WAN Jingyi;AI Qingzhong;ZENG Hui;OUYANG Zeyu;TANG Lei;ZHAO Xin(National Key Laboratory of Nuclear Reactor Technology,Nuclear Power Institute of China,Chengdu 610213,China;Nuclear Power Institute of China,Chengdu 610213,China)
出处
《智能计算机与应用》
2025年第5期111-116,共6页
Intelligent Computer and Applications
基金
四川省自然科学基金(23NSFSC2873)。
关键词
核动力系统
多元时序数据
异常检测
生成对抗网络
门控循环单元
nuclear power system
multivariate time series data
anomaly detection
Generative Adversarial Network
gated recurrent unit
