【作者】 王捷；

【导师】 李龙江；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 随着新一代工业智能和信息系统的发展,数字孪生技术作为一种模拟、预测和优化物理系统和过程的途径,已经成为了未来推动各领域发展的核心驱动力和新兴研究热点。数字孪生是对现实世界物理实体在其全生命周期内的精准映射,具备实时连续性,其构建和更新依赖于数据源的采集和传输两个过程。然而,由于设备资源受限、数据隐私问题、通信传输故障等原因,交互数据面临缺失和高时延的风险,无法满足数字孪生对于精确性和实时性的要求。如果不应用相应的数据补全机制,特征信息不足和错误的结论将会导致无法构建、更新与物理实体精确对应的数字孪生体。本文围绕数字孪生中“通过物理实体完整数据构建并更新数字孪生体”的实际场景,探究在数据同步时需要同时考虑在线和离线场景中不同的补全需求,所做的主要工作和创新如下:针对数字孪生在线数据流场景,设计了一种基于动态贝叶斯网络的在线数据补全方案。首先,针对贝叶斯网络学习复杂耗时的问题,提出了基于模拟退火和贪心算法的混合并行搜索算法完成贝叶斯网络结构的学习,再使用期望最大化(Expectation Maximization,EM)算法完成参数学习并构建动态贝叶斯网络进行补全。最后,在数据补全的基础上通过效能评估生成相应的改进方案并反馈到动态贝叶斯网络中更新模型参数,实时优化模型的补全效果。通过对比实验分析,表明所设计的方案较现有的贝叶斯补全方法在补全精确度上提升了11%,结构学习的效率和拟合程度分别提升了50%和19%,补全的实时性控制在40ms的范围内,保证了信息交互的精确性和实时性。针对数字孪生系统离线数据流场景,设计了一种基于改进时序和形变长短期记忆网络(Time And Mogrifier,TAM)的离线数据补全方案,深度挖掘数字孪生数据流内不同维度属性间的时序和空间特征,利用融合时空特征对缺失的数据流进行补全,保证了数据对于整个系统的精确性和一致性。通过与其他补全算法进行对比实验,本文所设计的补全方案较现有的基于形变长短期记忆网络的补全方法在补全精度上提升了5%,表明所提出方案能够适应数字孪生系统离线数据流的补全场景,并且保证了数据补全的高精度和一致性。最后,对所设计的在线/离线补全方案进行了对比,表明本文所划分的在线/离线场景的合理性。更多还原

【Abstract】 With the development of the next generation of industrial intelligence and information systems,digital twin technology,as a way to simulate,predict,and optimize physical systems and processes,has become a core driving force and an emerging research hotspot for various fields in the future.Digital twins are an accurate mapping of physical objects in the real world throughout their entire lifecycle,with real-time continuity,and their construction and update depend on the two processes of data source collection and transmission.However,due to factors such as device resource limitations,data privacy issues,and communication transmission failures,digital twins face the risks of missing data and high latency,which cannot meet the requirements for precision and real-timeness of digital twins.If no corresponding data completion mechanism is applied,insufficient feature information and incorrect conclusions will result,making it impossible to construct and update digital twins that accurately correspond to physical objects.This article focuses on the actual scene of digital twins,building and updating digital twins through complete data of physical objects,and explores the need for different data completion requirements in different offline and online scenarios during data synchronization.The main work and innovation of this article are as follows:Against the backdrop of the online data flow scene of digital twin,a dynamic Bayesian network-based online data completion scheme was designed.Firstly,in view of the complex and time-consuming problem of learning Bayesian networks,a hybrid parallel search algorithm based on simulated annealing and greedy algorithms was proposed to complete the Bayesian network structure.Then,the EM algorithm was used to complete the parameter learning and construct a dynamic Bayesian network for completion.Finally,based on the data completion,the efficiency evaluation was used to generate corresponding improvement plans and feedback was used to update the model parameters in the dynamic Bayesian network,optimizing the accuracy of the model’s completion effect real-time.Through comparison experiment analysis,it was shown that the designed scheme was more effective than existing Bayesian completion methods in improving the precision of data completion by 11%,and the efficiency and fit of structure learning were improved by about 50% and 19%,respectively.The real-time accuracy of data completion was controlled within a range of 30-40 ms,ensuring the precision and real-trimness of information interaction.For the offline data stream scenario of digital twin system,an offline data completion scheme based on improved temporal sequence and shape-shifting long short-term memory network is designed to deeply explore the temporal signatures and relational features among different dimensional attributes within the digital twin data stream,and use the fused temporal features to complement the missing data streams to ensure the accuracy and consistency of the data for the whole system.Through comparison experiments with other completion algorithms,the completion scheme designed in this thesis improves the completion accuracy by 5% with Mogrifier-LSTM,indicating that the proposed scheme can be adapted to the completion scenario of offline data streams in digital twin systems,and ensures the high accuracy and consistency of data completion.Finally,a comparison of the designed online/offline completion schemes is carried out to show the reasonableness of the online/offline scenarios delineated in this thesis.更多还原