基于深度学习的多源异构质检大数据融合与分析技术研究

It is urgent for administration of quality and inspection to solve the problem how to realizing scientific decision based on big data and improving the capacity of government governance and risk prevention in the field of product quality. In this study, theories and methods of deep learning are applied in the research of quality and inspection data to compensate the shortage of traditional data analysis method..This project takes big data of quality and inspection as the research object and deep learning as the core technology. Reaching on fusion method based on semantic analysis from different dimensions such as entities matching, entities linking and data conflict processing, and using the method of domain ontology, Markov predict and natural language processing. In combination with the deep learning methods of convolutional neural network, recurrent neural network and Q learning, analysis methods such as knowledge acquisition, representation adapter evolution and predict, visualization are researched in the field of cross-organization and multi-agent big data of quality and inspection. Some researching on knowledge services methods of big data are presented, such as multi-text abstract, knowledge navigation and knowledge recommendation based on the methods of domain ontology, feature words extraction and neural network. Finally, inspection and correction for theoretical model is presented through platform building of big data of quality and inspection which oriented to management decisions. This study will support scientific decision based on big data in the field of quality and inspection, enrich the theories of deep learning and provide technical support and reference sample for the relevant demonstration platform.

基于大数据的科学决策，提升政府治理与风险防范能力，已成为质检部门亟待解决的关键问题。为了弥补传统数据分析方法在质检大数据研究方面的不足，本研究将深度学习的相关理论和方法应用其中。 .本项目以质检大数据为研究对象，以深度学习为核心，综合采用领域本体、Markov链预测和自然语言处理等理论方法，从实体匹配、实体链接和数据冲突处理等角度，研究基于语义分析的质检大数据融合方法；结合卷积神经网络、递归神经网络和Q学习等深度学习方法，研究跨组织多主体质检大数据的知识获取、表征适配、演化与预测和可视化等分析技术；基于领域本体、特征词提取和神经网络等，研究多文本摘要、知识导航、知识推荐等质检大数据知识服务。最后，通过面向管理决策的质检大数据平台构建与典型应用研究，对理论模型进行检验和修正。本研究将有效支持大数据环境下质检领域的科学决策，并丰富深度学习理论，为相关示范平台的构建提供技术支撑和参考范例。

项目组针对质检大数据体量巨大、类型繁多、时效性高以及价值高密度低等特性，开展了质检大数据融合方法、质检大数据分析技术、质检大数据知识服务和质检大数据平台构建研究。具体研究内容包括：基于本体对齐的质检大数据实体匹配方法研究、基于增量记录的质检大数据实体链接方法研究、基于Markov逻辑网的质检大数据冲突解决方法、基于复杂网络的质检信息演化与预测研究、基于主题模型的质检大数据可视化技术研究、基于特征词提取的质检大数据知识导航研究、基于个性定制的质检大数据知识推荐研究、质检大数据平台构建研究。.通过本项目的资助，项目组发表学术论文24篇，其中SCI论文11篇，EI论文13篇；申请专利2项，获得软件著作权6项；完成国家标准报批稿4项；获得科研奖励2项。项目负责人许应成曾通过国家“万人计划”青年拔尖人才的函评，进入了会评答辩阶段，项目主要参与人李明、蔡华利和高晓红等先后晋升为教授或研究员。.培养研究生11名，其中博士生4名，硕士生7名。.本项目研究成果将为产品质量安全事件的信息管理、风险监控和预警决策提供理论基础，进而为有效控制产品质量安全风险，减少及预防系统性、区域性和行业性产品质量安全事件发生提供技术支持。