基于大语言模型的复杂工业系统设计:以供暖系统设计为例OA
LLM-Based Design of Complex Industrial Systems:A Case Study of Heating System Design
针对当前模块化工程系统方案设计与装配流程中存在的人工设计效率低、易出错、工程制图耗时久,以及通用大语言模型在工业领域因领域数据稀缺、专业术语理解偏差而导致应用不准确、可信度不足等问题,本文将大语言模型应用于下游工业生产制造尤其是模块化工程系统方案设计中,挖掘复杂系统模块间的深度耦合关系,实现复杂系统自动化设计与工程制图自动化.本文基于通义千问2.5(Qwen2.5)构建面向复杂模块化工业系统方案设计的大模型框架(Large language model for design of complex industrial system,DCI-LLM),搭建涵盖模块内部信息与全局系统知识的模块对话数据库,并设计"局部 协同"微调策略,使模型充分学习模块工况知识、深度挖掘模块间耦合关系,同时结合后处理程序集成,实现复杂系统工程制图全流程自动化.本文依托国内某企业供热系统知识数据库构建对话数据集开展验证实验.实验结果表明,微调后的DCI-LLM在模块知识问答、系统全局知识问答任务中的准确率分别可达93.4%、89.3%,模型专业认知能力优异;同时经专业工程师综合评测,该模型具备良好的实际工程应用价值.实际测试证实,DCI-LLM框架可有效实现从自然语言需求到完整工程图纸的端到端自动化设计,突破了传统大模型仅聚焦文本交互或中间代码生成的应用局限.本文研究揭示了基于数据特征驱动的大模型微调策略可有效适配模块化工业领域设计任务,所提框架与方法可为大语言模型在工业领域其他场景的落地应用提供可行技术路径与实施策略参考,在各类工业子领域自动化工程设计工作中具备广阔的应用前景与实用价值.
Modular design of complex engineering systems is a universal technology for rapid system design in the modern industrial field.Generating assembly schemes for modular engineering systems in industrial production is the core of achieving automation in system design.Traditional system design methods based on human experience suffer from low efficiency and poor adaptability.To enable the automatic assembly of modules for engineering systems,the performance matching and correlation between modules need to be accurately identified.Currently,general large language models,with their strong language analysis capabilities,have been applied to multiple industries.However,limited by data barriers in specific industrial field,their application in assembly of system modules is rare.Therefore,constructing high-quality domain data,fine-tuning professional large language models,and developing domain frameworks for automatic design of modular system constitute an important research direction to be explored at present.Based on the Qwen2.5 open-source large model,we utilize industrial system knowledge to construct a component library and propose a large language model for design of complex industrial system(DCI-LLM).Through the proposed low-rank adaptation(LoRA)-freeze-based"local-collaborative"fine-tuning method which combines local module and global knowledge of the system,DCI-LLM automatically generates system composition schemes and produces the related engineering drawings given system design requirements.We use the scheme design of heating systems as an example to verify the effectiveness of the proposed framework.Experimental results show that the fine-tuned DCI-LLM model achieves accuracy rates of 93.4%and 89.3%in answering questions about module knowledge and global knowledge,respectively.Moreover,scores from professional engineers indicate that DCI-LLM has practical application potential in scheme design of complex modular systems.Our work demonstrates that LLMs have significant application prospects in the field of automatic scheme design for complex industrial systems.
蔡鑫;吕宏强;许冉;王波;王琦;王鹤云;刘学军
南京航空航天大学人工智能学院,南京 211106,中国南京航空航天大学航空学院,南京 210016,中国南京航空航天大学人工智能学院,南京 211106,中国青岛凯能环保科技有限公司,青岛 266300,中国青岛凯能环保科技有限公司,青岛 266300,中国青岛凯能环保科技有限公司,青岛 266300,中国南京航空航天大学人工智能学院,南京 211106,中国
信息技术与安全科学
大语言模型低秩自适应微调冻结微调复杂系统设计系统工程
large language model(LLM)low-rank adaptation(LoRA)fine-tuningfreeze fine-tuningcomplex system designsystem engineering
《南京航空航天大学学报(英文版)》 2026 (2)
251-274,24
评论