基于特征差异增强的工程装备知识跨项目多源域迁移学习研究  

A multi-source transfer learning study based on feature difference enhancement for inter-project knowledge transfer of construction equipment

作　　者：徐进[1,2] 赵慧祺[1,2] 张泽慧 刘盾 XU Jin;ZHAO Huiqi;ZHANG Zehui;LIU Dun(School of Economics and Management,Southwest Jiaotong University,Chengdu 610000,China;Service Science and Innovation Key Laboratory of Sichuan Province,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学经济管理学院,成都610000 [2]西南交通大学服务科学与创新四川省重点实验室,成都610031

出　　处：《系统工程理论与实践》2024年第3期1097-1113,共17页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金(72171197,62276217);四川省自然科学基金面上项目(2023NSFSC0364)。

摘　　要：工程装备智能化是发展智能建造的重要基础,项目知识是工程装备智能化的知识源泉,因此,工程装备知识跨项目的有效共享与利用是实现智能建造的重要环节.为了增强工程装备知识的跨项目利用效率与效果,本文提出了一种基于特征差异增强的多源域迁移学习框架.该框架利用混合深度神经网络提取源项目的通用时空特征表示,基于项目相似度度量筛选可迁移源项目,通过所设计的特征差异增强方法挖掘多源域的域特殊特征表示并进行集成,在避免负迁移的同时实现工程装备知识的跨项目有效转移.本文使用多个隧道工程项目的数据进行了实验,在六个盾构设备姿态预测知识转移任务的两个预测目标上,该框架相较于基线模型的预测准确性平均提升度分别为86.48%、117.01%,并具有良好的稳健性和情景适应性.实验结果表明:本文所设计的新框架可以挖掘多个源域项目的特性知识并整合其共性知识,通过集成多源域迁移学习的知识来提高知识利用率,为大型工程装备知识的跨项目转移提供了有效的方法和工具,有助于提升施工项目的知识管理与智能建造水平.Engineering equipment intelligence is an important basis for the development of intelligent construction,and project knowledge is the knowledge source of engineering equipment intelligence.Therefore,the effective sharing and utilization of engineering equipment knowledge across projects is an important link to realize intelligent construction.In order to enhance the efficiency and effectiveness of cross-project utilization of engineering equipment knowledge,this paper proposes a multi-source domain transfer learning framework based on feature difference enhancement.The framework uses a hybrid deep neural network to extract the common spatiotemporal feature representation of source items,screens transferable source items based on the project similarity measurement,and mines the domain special feature representation of multiple source domains through the designed feature difference enhancement method and integrates it,so as to avoid negative transfer and realize effective cross-project transfer of engineering equipment knowledge.In this paper,the data of several tunnel engineering projects are used to carry out experiments.In the two prediction objectives of six shield equipment attitude prediction knowledge transfer tasks,the average accuracy improvement of the framework compared with the baseline model is 86.48%and 117.01%,respectively,and it has good robustness and situational adaptability.Experiments show that the new framework designed in this paper can mine the characteristic knowledge of multiple source domain projects and integrate their common knowledge,improve the knowledge utilization rate by integrating the knowledge of multi-source domain transfer learning,provide an effective method and tool for the cross-project transfer of large-scale engineering equipment knowledge,and help improve the level of knowledge management and intelligent construction of construction projects.

关 键 词：工程项目 项目知识转移 工程装备知识 多源域迁移学习 深度学习 

分 类 号：C931.6[经济管理—管理学]