基于深度学习的机械领域知识图谱构建及应用  

Construction and application of knowledge graph in mechanical fields based on deep learning

作　　者：崔硕 张春燕[1,2] 贾家乐 张成东 张胜文 陈凯[1] CUI Shuo;ZHANG Chunyan;JIA Jiale;ZHANG Chengdong;ZHANG Shengwen;CHEN Kai(School of Mechanical Engineering,Jiangsu University of Science and Technology,Zhenjiang 212100;Jiangsu Provincial Key Laboratory of Advanced Manufacturing of Machinery and Equipment,Zhenjiang 212100,CHN)

机构地区：[1]江苏科技大学机械工程学院,江苏镇江212100 [2]江苏省船海机械装备先进制造重点实验室,江苏镇江212100

出　　处：《制造技术与机床》2023年第2期83-89,共7页Manufacturing Technology & Machine Tool

基　　金：江苏省研究生实践创新项目(SJCX22_1921);镇江市重点研发计划项目(GY2019003,GY2020007)。

摘　　要：针对企业内以分散形式进行存储的设计信息难以重用的问题,融合设计模型库和文本资源库构建机械设计领域知识图谱:通过计算机视觉等技术从历史模型库及文档中获取所需领域知识,对于模型库数据,采用聚类算法减少图谱节点冗余;提出一种新的MachineALBERT预训练模型对通用ALBERT进行参数设计,将该模型作为文本的语义编码层,以Bi-LSTM作为标签预测层,加入CRF作为整体标签优化层,搭建实体识别模型;对于关系抽取模型,共享字符编码层,加入CNN层对文本关系进行分类,将非结构化信息转化为结构化三元组,经实体对齐后存储至图数据库。根据构建完成的图谱建立计算机辅助设计可视化系统,以提供实体查询与知识问答等多种功能,提升产品设计效率。Aiming at the problem that it is difficult to reuse the design information stored in a decentralized form within the enterprise,the design model library,and the text resource library are integrated to build a knowledge graph in the field of mechanical design.Obtain the required domain knowledge from historical model libraries and documents through computer vision and other technologies:for model library data,clustering algorithms are used to reduce map node redundancy;A new machine ALBERT pre-trained model is proposed to design parameters for universal ALBERT,which is used as the semantic coding layer of text,Bi-LSTM as the label prediction layer,CRF as the overall label optimization layer,and the entity recognition model is built.For the relationship extraction model,the character encoding layer is shared,the CNN layer is added to classify the text relationships,and the unstructured information is converted into a structured triple,which is stored in the graph database after entity alignment.A computer-aided design visualization system is established based on the completed map,which aims to provide various functions such as entity query and knowledge quizzes to improve product design efficiency.

关 键 词：知识图谱 深度学习 可视化系统 计算机辅助设计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]