• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 23-35.doi: 10.3901/JME.260673

• 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

基于认知循环的大语言模型智能体驱动旋转机械故障诊断框架

李振宁1,2, 蔺瑞管1,2, 彭聪1,2, 李彦夫1,2   

  1. 1. 清华大学工业工程系 北京 100084;
    2. 清华大学质量与可靠性研究院 北京 100084
  • 收稿日期:2026-03-31 修回日期:2026-07-05 发布日期:2026-11-20
  • 作者简介:李振宁,男,1994年出生,博士后研究员。主要研究方向为机器学习与大语言模型及其在故障诊断中的应用。E-mail:lizhenning@mail.tsinghua.edu.cn;李彦夫(通信作者),男,1982年出生,博士,教授,博士研究生导师。主要研究方向为系统可靠性评价、故障预测及维修优化方法。E-mail:liyanfu@tsinghua.edu.cn
  • 基金资助:
    国家自然科学基金重点国际合作资助项目(W2511076)。

Cognitive Loop Based LLM Agent-Driven Framework for Rotating Machinery Fault Diagnosis

LI Zhenning1,2, LIN Ruiguan1,2, PENG Cong1,2, LI Yan-Fu1,2   

  1. 1. Department of Industrial Engineering, Tsinghua University, Beijing 100084;
    2. Institute for Quality and Reliability, Tsinghua University, Beijing 100084
  • Received:2026-03-31 Revised:2026-07-05 Published:2026-11-20

摘要: 旋转机械故障诊断中,深度学习方法面临模型选择繁琐、流程自动化程度有限及闭环反馈机制缺失等困境。大语言模型(Large language model, LLM)凭借其语义理解与推理能力,为构建自动化诊断流程提供了新的可能,然而现有研究或赋予LLM过高自主权而致使行为难以预测,或停留在浅层工具化应用而忽视其语义理解能力。针对上述问题,研究提出一种语义驱动的端到端故障诊断框架。该框架包含三项关键设计:提出OTAR认知循环范式,将智能体(Agent)决策过程结构化为观察、思考、行动、反思四阶段,建立从状态感知到经验反思的完整闭环,使LLM能够动态干预训练过程并持续积累经验;设计双层优化机制,Epoch级Agent负责实时状态评估与参数调整,Run级Agent负责架构搜索与超参数全局优化,两层通过共享记忆库实现知识传递;构建语义决策与符号执行分离的协作架构,通过结构化接口约束LLM行为边界,将其定位为认知中枢,在确保可控性的同时兼顾优化效果与决策过程透明性。实验结果表明,所提方法在有限迭代预算下相较于传统自动化机器学习方法具有更高的搜索效率与更稳定的收敛特性,消融实验验证了各组件的协同作用,泛化性实验表明该框架可适配多种主流LLM后端。

关键词: 旋转机械, 故障诊断, 大语言模型, 智能体, 自动化机器学习

Abstract: In rotating machinery fault diagnosis, deep learning methods face challenges including cumbersome model selection, limited process automation, and the absence of closed-loop feedback mechanisms. Large language models (LLMs), with their semantic understanding and reasoning capabilities, offer new possibilities for building automated diagnostic pipelines. However, existing research either grants LLMs excessive autonomy leading to unpredictable behavior, or remains confined to shallow tool-based applications that overlook their semantic understanding capabilities. To address these issues, the research proposes a semantics-driven end-to-end fault diagnosis framework. The framework incorporates three key designs: the OTAR (Observe, Think, Act, and Reflect) cognitive loop paradigm, establishing a complete closed-loop from state perception to experience reflection, enabling LLMs to dynamically intervene in training and continuously accumulate experience; a dual-layer optimization mechanism, where the epoch-level agent handles real-time state assessment and parameter adjustment, while the run-level agent handles architecture search and global hyperparameter optimization, with both layers achieving knowledge transfer through shared memory; and a collaboration architecture that separates semantic decision-making from symbolic execution, constraining LLM behavioral boundaries through structured interfaces and positioning them as cognitive hubs, thereby balancing optimization effectiveness and decision-process transparency while ensuring controllability. Experimental results demonstrate that the proposed method exhibits higher search efficiency and more stable convergence compared to traditional automated machine learning methods under limited iteration budgets. Ablation experiments validate the synergistic effects of each component, and generalization experiments indicate that the framework can adapt to multiple mainstream LLM backends.

Key words: rotating machinery, fault diagnosis, large language model, agent, automated machine learning

中图分类号: