【作者】 陈遥；

【导师】 吕建华；

【作者基本信息】 东南大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 随着数据管理技术的发展与可穿戴智能设备的普及,个人生活数据受到了越来越多的关注。面向个人生活事件的检测一直是复杂事件检测领域的重要研究内容。通过将个人生活数据与事件检测有效地结合起来,可以更好地指导人们改善生活水平,提高生活质量,具有重要的研究价值和经济价值。简单行为事件主要是对个人身体状态的描述,例如坐立,跑步和走路等简单动作,这些行为事件可以显示出个人每天的运动情况,而适当的运动对于个人健康而言极为重要。面对智能设备中集成传感器快速发展的现状,本文提出了一种基于多传感器和多权重的简单行为事件检测方法MW_KNN,挖掘多个传感器数据在用户特定行为下的联系,并充分考虑到不同传感器对于不同简单行为事件的检测效果具有差异性,设置相应的特征权重,从而提高识别效果。提出了基于查询集过滤的优化方案,解决了KNN时间复杂度的问题。相关实验结果表明,利用多种传感器协同工作的简单行为事件检测模型可以有效提高人体简单行为事件的识别率,同时过滤方案可以有效降低行为检测所需时间。在获取个人实时的运动情况后,结合其他生活数据信息,本文提出了生活数据网络(Observations of daily living Network,ODL Network)的概念,并给出具体构造算法。为了构造典型的生活数据事件库,提出了面向生活数据网络的三种切割方法,分别为基于时间特征、基于空间特征和基于聚类的切割,这三种方法能有效地划分生活数据网络。其次,由于图编辑距离作为衡量两个图相似性的常用方法是NP-Hard问题,本文提出了三种面向生活数据网络的相似性度量方法来进行复杂行为事件检测,分别为基于星型结构的相似性度量方法(Star Mapping Based Network Similarity,StarMapping)、基于元路径的相似性度量方法(Metapath Mapping Based Network Similarity,PathMapping)和基于主结构序列的相似性度量方法(Dominant Structure Sequence Mapping Based Network Similarity,DSSeqMapping),这三种方法充分考虑了生活数据网络的语义信息和结构信息。实验结果表明,面对不同划分方式下的生活数据网络,各个算法的复杂行为事件检测效果也存在差异性,PathMapping更适用于时间特征划分,DSSeqMapping更适用于空间特征划分,StarMapping则更适用于聚类划分,另外,PathMapping的时间复杂度较高,DStarMapping和StarMapping两者相当。之后,针对面向生活数据网络的复杂行为事件实时检测的要求,基于上述三种检测算法提出了三种利用上下界进行过滤的策略,实验表明,三种方法都能有效地提高实时检测的效率。最后,本文设计并实现了一个面向生活数据的事件检测原型系统,用于生活数据的处理和存储以及事件检测技术的应用。更多还原

【Abstract】 With the development of data management technology and the popularity of wearable intelligent devices,personal life data has attracted more and more attention.The detection of personal life events has always been an important research content in the field of complex event detection.By effectively combining personal daily living data with event detection,it can better guide people to improve their living standards and quality of life.It has important research value and economic value.Simple behavioral events are mainly descriptions of individual physical state,such as sitting,running and walking.These behavioral events can highlight the individual’s daily exercise situation,and proper exercise is extremely important for personal health.Faced with the rapid development of integrated sensors in intelligent devices,this thesis proposes a simple behavior event detection method based on multi-sensor and multi-weight,MW_KNN,which can mine the relationship between multi-sensor data and user-specific behavior,and fully consider that different sensors have different detection effects for different simple behavior events,and set corresponding feature weights,so as to achieve the goal.Improve the recognition effect.At the same time,aiming at the high time complexity of distance-based computation in KNN,an optimization scheme is proposed to reduce time complexity by filtering query sets.Relevant experimental results show that the recognition rate of human simple behavior events can be effectively improved by using a simple behavior event detection model based on multisensor collaboration,and the optimization scheme can effectively reduce the need time of behavior detection.After obtaining the real-time movement of individuals and combining with other life data information,this thesis puts forward the concept of Observations of daily living network(ODL Network),and gives the concrete construction algorithm.In order to construct a typical life data event database,three cutting methods for life data network are proposed,which are based on time feature,space feature and clustering.These three methods can effectively partition life data network.Secondly,since graph editing distance is NP-Hard problem as a common method to measure the similarity between two graphs,this thesis proposes three similarity measurement methods for life data network to detect complex behavior events,namely Star Mapping Based Network Similarity(StarMapping),Metapath Based Similarity Measurement(PathMapping)and Dominant Structure Sequence Mapping Based Network Similarity(DSSeqMapping)take full account of the semantic and structural information of life data networks.The experimental results show that in the face of different partitioning modes of life data networks,the detection effect of complex behavior events of each algorithm is also different.PathMapping is more suitable for time feature partition,DSSeqMapping is more suitable for space feature partition,StarMapping is more suitable for clustering partition.In addition,PathMapping has higher time,StarMapping and DSSeqMapping is comparable.Then,in order to meet the requirements of real-time detection of complex behavioral events for life data networks,three strategies of using upper and lower bounds to filter are proposed based on the above three detection algorithms.Experiments show that the three methods can effectively improve the efficiency of real-time detection.Finally,this thesis designs and implements an event detection prototype system for life data,which is used to process and store life data and apply event detection technology.更多还原