面向工业物联网的数字孪生建模方法研究

【作者】 刘杨；

【导师】 关心；

【作者基本信息】 黑龙江大学 ， 信息统计技术， 2024， 硕士

【摘要】 工业物联网通过物联网设备将不同的工业设备连接,加深了不同设备之间以及设备与工人之间的联系。随着设备的不断增多,确保工业设备的稳定运行变得愈发重要。数字孪生是一种能够连接物理空间和虚拟空间,以实现对物理实体运行进行估计和分析的新兴技术。在工业物联网中建立的数字孪生模型能够反映工业设备的运行。然而,工业设备复杂的运行模式难以被完全掌握,因此基于机理模型的传统方法不适合用于建立数字孪生。神经网络通过提取数据中的知识,为建立通用的数字孪生提供了思路。本文设计了一个用于工业物联网的数字孪生架构,并提出了一种基于神经网络建立数字孪生模型的方法。神经网络的训练需要大量数据,而工业设备复杂的使用环境会导致数据采集与传输面临挑战,因此产生数据不充分问题。于是,本文提出了面向数字孪生的迁移学习方法,确保数据不充分条件下物理实体能被重构。其次,本文将数字孪生建模问题描述为数字孪生模型和物理实体之间差异最小化问题,并设计了一个深度神经网络以最小化数字孪生的建模损失。自相关和偏自相关被引入,用于判断数据中是否存在显著的时间特征。其后,Informer算法被用于提取序列中的时间特征。最后,为了解决建立数字孪生时会面临的数据不充分问题,本文提出了一个基于深度迁移学习的不充分数据数字孪生建模算法。同时,本文分析了源域和目标域之间的相似性,以确保源域和目标域之间存在公共知识以避免负迁移。进一步,本文选择了领域对抗神经网络,将从源域中提取公共知识用于目标域的学习任务。在基于相似性的基础上,本文定义了迁移率的概念以决定迁移多少知识。基于真实数据集的数值结果验证了所提出的数字孪生建模算法、相似性分析方法和所提出的迁移率是有效的。更多还原

【Abstract】 The Industrial Internet of Things connects various industrial equipment through the Internet of Things devices,deepening the connection between equipment and workers.With the increasing number of devices,the effective management of the Industrial Internet of Things becomes more important for the stable operation of industrial equipment.Digital twin is an emerging technology that can connect physical and virtual space to assess and analyze the operation of physical entities.In this paper,a digital twin architecture for the industrial Internet of Things is designed,and a method for building digital twin models based on neural networks is proposed.The training of neural networks requires a lot of data,and the complex environment of industrial equipment will lead to challenges in data acquisition and transmission,resulting in an insufficient data problem.Therefore,this paper proposes a transfer learning method for digital twins to ensure that physical entities can be reconstructed with insufficient data.Second,this paper describes the digital twin modeling problem as a problem of minimizing the difference between the digital twin model and the physical entity and designs a deep neural network to minimize the modeling loss of the digital twin.Autocorrelation and partial autocorrelation are introduced to determine if there are significant temporal features in the data.The Informer algorithm is then used to extract the temporal features in the sequence.Finally,to solve the problem of insufficient data when building digital twins,this paper designs a digital twin modeling with insufficient data algorithm based on deep transfer learning.Meanwhile,this paper analyzes the similarity between the source and target domains to avoid negative transfer.The similarity reflects that there is public knowledge between the source domain and the target domain.Furthermore,this paper selects the DomainAdversarial Neural Networks,which extracts the public knowledge from the source domain using the adversarial method.Based on the similarity,we define the concept of transfer ratio to determine the amount of knowledge to be transferred.The numerical results based on real data sets verify that the proposed digital twin modeling algorithm,similarity analysis method,and proposed mobility are effective.更多还原