深度图网络驱动的核电系统多级异常检测方法  

Deep Graph Network⁃Driven Multi⁃level Anomaly Detection for Nuclear Power System

作　　者：张乐 成玮[1] 张硕 陈雪峰[1] 常丰田 洪郡滢 马颖菲 彭将 ZHANG Le;CHENG Wei;ZHANG Shuo;CHEN Xuefeng;CHANG Fengtian;HONG Junying;MA Yingfei;PENG Jiang(State Key Laboratory for Manufacturing Systems Engineering,Xi'an Jiaotong University Xi'an,710049,China;China Nuclear Power Engineering Co.,Ltd.Beijing,100840,China;Fujian Fuqing Nuclear Power Co.,Ltd.Fuqing,350318,China)

机构地区：[1]西安交通大学机械制造系统工程国家重点实验室,西安710049 [2]中国核电工程股份有限公司,北京100840 [3]福建福清核电有限公司,福清350318

出　　处：《振动.测试与诊断》2025年第1期88-94,202,共8页Journal of Vibration,Measurement & Diagnosis

基　　金：国家重点研发计划资助项目(2019YFB1705403);王宽诚教育基金会资助项目;中核集团领创项目(J201912021)。

摘　　要：针对深度学习方法未明确学习变量间关系结构、系统异常难以准确检测的问题,提出一种深度图网络驱动的核电系统多级异常检测方法。首先,利用无监督图对比学习方法挖掘系统变量时间序列间相关性,构建与核电系统物理结构匹配的可解释性图结构;其次,基于变分图自编码器重构系统图结构,以重构误差来表征系统运行状态,从系统层面防止非线性突发行为带来的安全性问题;然后,通过半监督图卷积节点分类模型识别系统内部各变量运行状态,实现测点级异常检测;最后,以PCTranACP100仿真机2种基准事故工况数据、国内某核电机组循环水系统监测数据来验证提出方法的有效性。结果表明,系统级异常检测准确率达到93%,86%和90%,证明所提出方法能够准确检测出系统异常情况,可降低电厂单一仪表异常触发的非计划停机概率。Deep learning methods do not explicitly learn the structure of relationships between variables making it difficult to accurately detect system anomalies.To address this issue,a deep graph network-driven multi-level anomaly detection method is proposed for nuclear power systems.First,unsupervised graph contrastive learning is introduced to mine the correlation between variables.Based on that interpretable graph structure is con⁃structed,which corresponds to the physical structure.Second,variational graph auto-encoder is integrated to re⁃construct the graph networks.The reconstruction error is applied to characterize system operating conditions,preventing security problems caused by non-linear or bursting behavior.Then,a semi-supervised graph convolu⁃tion node classification model is applied to determine the operational status of each system variable.Finally,two cases studies are used to verify the effectiveness of the proposed method,including simulation data of the PC⁃TranACP100 and monitoring data of the three-circuit circulating water system.The accuracy of system-level anomaly detection based on correlation modelling is 93%,86%and 90%,respectively.The proposed method is effective in improving the accuracy of system anomaly detection,thereby reducing the probability of un⁃planned downtime triggered by anomalies in a single component.

关 键 词：核电系统 无监督深度图学习 可解释性图结构 多级异常检测 变分图自编码器 

分 类 号：TH133.31[机械工程—机械制造及自动化] TH17