多元统计方法在核取证溯源分析中的研究进展

Research on Multivariate Statistical Analysis Method in Nuclear Forensics Traceability Analysis

国家高度重视核安全,习近平总书记多次在全球核安全峰会上强调核材料溯源取证的重要性。核材料溯源取证,是对核材料及放射性材料进行检查分析,以溯源其性质、制造的时间、地点、方式及预期用途。目前关于该领域的前沿研究方向是基于数据库信息,利用多元统计方法溯源未知核材料的来源信息。该文回顾了近年来的核取证研究进展,并指出存在数据依赖性强、模型缺乏定量分析能力和普适性等不足。为克服这些局限,该文介绍了最新的线性拟合方法实现溯源取证分析。基于经济合作与发展组织核能署(OECD/NEA)于2017年发布的SFCOMP-2.0乏燃料数据库,以不同核素浓度作为样品特征,提出了线性关系假设并通过数据库数据和压水堆(PWR)与沸水堆(BWR)模拟结果进行了检验。结果表明,线性关系假设在PWR和BWR中均适用,并可用于预测燃料的初始富集度和燃耗量。此外,应用3种机器学习方法(逻辑回归、支持向量机、多层感知器)实现对PWR和BWR的分类。最后,应用KNN分类、随机森林和多层感知器算法改进了核取证分类模型,提高了模型对各反应堆的分辨能力。

The state attaches great importance to nuclear safety.General Secretary Xi Jinping has re⁃peatedly emphasized the significance of nuclear material traceability and forensics at the Global Nu⁃clear Security Summits.Nuclear material traceability forensics is the inspection and analysis of nucle⁃ar and radioactive materials to trace their nature,time,place,method and intended use.At pres⁃ent,the cutting-edge research direction in this field is to trace the source information of unknown nu⁃clear materials by using multivariate statistical method based on database information.This paper re⁃views the progress of nuclear forensics research in recent years,and points out that there are still shortcomings such as strong data dependence,lack of quantitative analysis ability and universality of models.In order to overcome these limitations,this paper introduces the latest linear fitting method to achieve traceability forensic analysis.Based on the SFCOMP-2.0 spent fuel database published by OECD/NEA in 2017,the linear relationship hypothesis was proposed with different nuclide concen⁃tration ratios as sample characteristics and tested by the database data and the simulation results of pressurized water reactor(PWR)and boiling water reactor(BWR).The results show that the linear re⁃lation hypothesis is suitable for both PWR and BWR,and can be used to predict the initial enrich⁃ment degree and burnup amount of fuel.In addition,three machine learning methods(LGR,SVM and MLP)are used to classify PWR and BWR.The research presented in this paper provides a new approach and technique for nuclear material traceability and forensics,which is of great theoretical and practical significance for improving the monitoring and prevention capabilities of nuclear safety.Finally,KNN classification,random forest and multi-layer perceptron algorithm are used to improve the nuclear forensics classification model and improve the ability to distinguish each reactor.