自动化学报2025，Vol.51Issue(11)：2454-2472,19.DOI:10.16383/j.aas.c250247
工业垂域具身智控大模型构建新范式探索
An Exploration of a New Paradigm for Constructing Industrial Domain-specific Embodied Intelligent Control Large Models
摘要
Abstract
The industrial domain-specific adaptation of large models is an inevitable trend in the evolution of gener-al intelligence towards specialized applications.It is also the core engine driving the intelligent transformation of in-dustries.However,the application of large models in the industrial field encounters several challenges,such as diffi-culties in understanding the implications of industrial time-series data,embedding industrial physical and chemical laws,ensuring the reliability of model outputs,and solving complex industrial problems.To overcome these bottle-necks,a paradigm for developing industrial domain-specific embodied intelligent control large models is proposed.This paradigm innovatively introduces a time-series data meta-modeling approach that converts industrial time-series data into code semantics,thereby improving the large model's ability to interpret and reason with time-series data.Additionally,an industrial law knowledge graph is constructed based on meta-models and integrated into the large model generation process,utilizing deterministic scientific principles to mitigate randomness of generation.A dual-track verification platform combining digital twins and physical entities has been established.The platform employs a virtual-physical embodied feedback mechanism and real-time reinforcement learning to optimize the cred-ibility of the model outputs.A hybrid reward function is designed,combining knowledge graph rule-based scoring with expert evaluations from both virtual and physical validations.By integrating adaptive learning with length regularization strategies,the model overcomes the tendency to"prioritize simplicity over complexity"in solving complex industrial problems.Ultimately,this approach forms a four-layer closed-loop architecture that incorporates domain-specific adaptation,embodied control,credible verification,and embodied feedback.When applied to the non-ferrous metallurgy sector,the first embodied intelligent control large model for non-ferrous metallurgy was con-structed,and experimental validation demonstrated the effectiveness of this paradigm.This establishes a bridge for transitioning large models from laboratory settings to industrial applications,connecting technology with practical implementation.关键词
工业大模型/强化学习/有色冶金/具身反馈Key words
Industrial large model/reinforcement learning/non-ferrous metallurgy/embodied feedback引用本文复制引用
陈致蓬,韩杰,阳春华,桂卫华..工业垂域具身智控大模型构建新范式探索[J].自动化学报,2025,51(11):2454-2472,19.基金项目
国家重点研发计划(2024YFC3908002),国家自然科学基金面上项目(62273359,62373377),国家自然科学基金重大项目(62394340),湖南省青年骨干教师培养对象项目(206030802),湖南省教育厅研究生教改项目(2024JGYB021)资助Supported by National Key Research and Development Pro-gram of China(2024YFC3908002),National Natural Science Foundation of China General Program(62273359,62373377),Major Projects of National Natural Science Foundation of China(62394340),Hunan Province Young Backbone Teacher Training Program(206030802),and Hunan Provincial Department of Edu-cation Graduate Education Reform Program(2024JGYB021) （2024YFC3908002）
