• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 1-22.doi: 10.3901/JME.260776

• 特邀专栏:大模型驱动的高端装备智能运维 • 上一篇    

扫码分享

多模态大模型驱动的高端装备智能运维研究综述

李响1, 俞舒鹏1, 雷亚国2, 杨彬1, 陈思成1, 冯珂2, 李宜峰1, 鲁凡1   

  1. 1. 西安交通大学现代设计与转子轴承系统教育部重点实验室 西安 710049;
    2. 西安交通大学精密微纳制造技术全国重点实验室 西安 710049
  • 收稿日期:2026-05-25 修回日期:2026-07-16 发布日期:2026-11-20
  • 作者简介:李响,男,1990年出生,教授。主要研究方向为工业人工智能、机器视觉、具身智能、大模型等。E-mail:lixiang@xjtu.edu.cn;雷亚国(通信作者),男,1979 年出生,教授。主要研究方向为机械装备智能运维、动力学建模、智能故障诊断与预测等。E-mail:yaguolei@mail.xjtu.edu.cn
  • 基金资助:
    国家自然科学基金(52475125,52435003)和新基石科学基金会科学探索奖(XPLORER-2024-1036)资助项目。

Survey on Multimodal Large Model-driven Intelligent Operation and Maintenance of High-end Equipment

LI Xiang1, YU Shupeng1, LEI Yaguo2, YANG Bin1, CHEN Sicheng1, FENG Ke2, LI Yifeng1, LU Fan1   

  1. 1. Key Laboratory of Education Ministry for Modern Design and Rotor-Bearing System, Xi'an Jiaotong University, Xi'an 710049;
    2. State Key Laboratory for Manufacturing System Engineering, Xi'an Jiaotong University, Xi'an 710049
  • Received:2026-05-25 Revised:2026-07-16 Published:2026-11-20

摘要: 高端装备智能运维对保障现代工业生产安全与提升效率至关重要。然而,传统基于机理建模、经典信号处理、机器学习等方法在高端装备复杂多变的服役工况下,普遍存在泛化能力有限、多模态信息融合不充分、人机交互不足等局限性。随着人工智能技术的飞速发展,多模态大模型凭借其强大的通用表征、跨模态对齐、上下文学习等能力,正推动装备运维从“专用模型”向“通用基础模型”的范式变革。梳理了驱动这一变革的四大关键方法,包括自然语言处理、领域泛化的迁移学习、多模态数据协同学习和智能体构建。在此基础上,系统整理了大模型在故障诊断、寿命预测、维护建议生成及决策执行等典型运维任务的研究进展,并阐述了其在数字孪生、具身智能等场景下的应用可行性,展现了其在数据与知识双轮驱动下的强大潜力。最后,深度剖析当前面临的数据瓶颈严重、通用模型底座缺失等挑战,指明未来研究应聚焦于构建具备感知、推理与执行闭环能力的“视觉-语言-动作”运维智能体,以实现从“感知诊断”向“认知执行”以及从“被动诊断”到“主动干预”的智能化范式跃迁。

关键词: 高端装备, 智能运维, 多模态大模型, 运维智能体

Abstract: The intelligent operation and maintenance of high-end equipment plays a critical role in ensuring production safety and efficiency in modern industry. However, traditional methods based on mechanistic modeling, classic signal processing, and machine learning commonly suffer from limitations under the complex and variable service conditions of high-end equipment, including restricted generalization capabilities, insufficient multi-modal information fusion, and inadequate human-machine interaction. With breakthroughs in artificial intelligence, multi-modal large models, owing to their powerful capabilities in general-purpose representation, cross-modal alignment, and in-context learning, are driving a paradigm shift in equipment maintenance from "specialized models" toward "general-purpose foundation models." Four key methods underpinning this transformation, namely natural language processing, domain generalization in transfer learning, multi-modal data collaborative learning, and intelligent agent construction, are systematically reviewed. On this basis, the research advances of large models in typical maintenance scenarios, including fault diagnosis, remaining useful life prediction, maintenance recommendation generation, and decision-making execution, are systematically organized, and their application feasibility in digital twins and embodied intelligence is elucidated, demonstrating their remarkable potential powered by the dual engines of data and knowledge. Finally, core challenges currently faced by the industry, such as severe data bottlenecks and the absence of general-purpose foundation backbones, are deeply analyzed and identified; meanwhile, a future research focus is directed toward developing vision-language-action maintenance agents equipped with closed-loop perception, reasoning, and execution capabilities, thereby enabling an intelligent transition from perception-based diagnosis to cognitive execution and from passive diagnostics to active intervention.

Key words: high-end equipment, intelligent operation and maintenance, multimodal large models, operation and maintenance agents

中图分类号: