大数据和领域知识驱动的食品安全风险预警方法与关键技术研究

This project is aimed at some key technical issues in the field of food safety, such as the difficulty in identifying the risk and the poor accuracy in the early warning and so on, as well as the food safety data have some difficult-to-describe characteristics of multiple heterogeneous sources, the hybrid, and the uncertainty, etc. In order to handle these problems, we proposed a new multi-level quantitative, formal and logical cyberspace description method for the food safety information. Firstly, two methods of the food safety information fusion and the risk key feature extraction are studied to build a multi-level, multi-dimensional and multi-index data clustering fusion model for the food safety. Then the overall characteristics of the big data network structure of the food safety, the evolution analysis and the modeling methods of the risk dynamic behavior are analyzed to establish a comprehensive risk evaluation model with different hierarchical levels. Finally, a Bayesian networks based food safety risk warning model is established by combining with the knowledge of the food safety field and analyzing the transition probability between the different risk points and the sensing body. Thus, a new method of the food safety risk warning method integrated the big data with the domain knowledge fusion driver is proposed. As a case study and application, the big data of the dairy products and the infant formula milk powder products are selected. Then a cloud platform-based for the food safety risk warning decision support system is established aiming at demonstration and applications. The universality and effectiveness of the proposed methods can provide decision-making strategies for the risk warning and safety control of the food regulatory authorities, testing units, and enterprise production process.

针对目前食品安全存在风险识别难、预警准确性差，以及食品安全大数据复杂性难处理的关键技术问题，研究食品安全大数据信息融合与风险关键特征提取方法，提出食品安全信息多层次定量化、形式化与逻辑化的复杂网络空间描述架构，构建多层次、多维度、多指标水平的食品安全数据聚类融合模型；研究食品安全大数据网络建模方法，识别网络结构整体特征和风险动力学行为演变规律，建立食品安全不同层次水平的分级风险综合评价模型；结合食品领域相关知识，分析不同风险点和各检测主体之间的转移概率，建立贝叶斯网络的食品安全风险预警模型，提出大数据和领域知识融合驱动的食品安全风险预警新方法；以乳制品和婴幼儿配方奶粉大数据为研究应用对象，建立基于云平台的食品安全风险预警决策支持系统并示范应用，验证所提方法的普适性和有效性，为食品监管部门、检测单位与企业生产的风险预警和安全控制提供决策依据。

经过对课题系统研究主要研究成果如下。首先针对食品安全大数据复杂性难处理的关键技术问题，通过分析句法依赖的必要性和句子中词语对关系提取的贡献，研究食品安全风险关键特征提取方法，提出了食品安全大数据复杂网络实体关系抽取方法。该方法采用自顶向下的组网策略，形成食品安全知识对象，以检测主体作为实体，以检测属性作为关系。然后针对食品安全大数据多源异构多层次结构的特征，构建多层次、多维度、多指标水平的食品安全知识融合的表示学习方法，其将非对称知识嵌入到流形空间中。其次基于食品安全大数据的时空分布，从网络的结构关系层次性和关系性出发，研究网络结构整体特征和风险动力学行为演变规律，提出食品安全大数据网络拓扑结构知识与模型风险的演化行为分析方法。该方法采用基于灰色关联分析的解释结构建模对食品安全的影响因素进行层次预警分析，可以指导相关部门加强监管，督促企业安全生产。然后针对复杂食品安全检测数据预警建模问题，提出食品安全风险评估与预警方法，基于熵权的层次分析可以得到食品安全成分数据的风险值，利用质量控制分析方法计算相应的风险等级，并通过相关质检部门提供风险管理，有效降低国家风险损失。最后基于以上研究成果，建立基于云平台的食品安全风险预警决策支持系统并示范应用，验证所提出方法的普适性和有效性，为食品监管部门、检测单位与企业生产的风险预警和安全控制提供决策依据。目前累计发表SCI/EI学术论文26篇（第一标注11篇，第二标注15篇），申请发明专利3项，获批软件著作权3项。培养博士生3名其中毕业2名、毕业硕士生7名。特邀报告1次、会议分组报告4次、墙报展示4次。