• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 226-240.doi: 10.3901/JME.260772

• 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

DiagAgent:基于多智能体思维链推理的轴承故障诊断框架

王杰1,2, 段凯文2, 张继2, 陈是扦1, 翟婉明1   

  1. 1. 西南交通大学轨道交通运载系统全国重点实验室 成都 610031;
    2. 西南交通大学计算机与人工智能学院 成都 611756
  • 收稿日期:2025-12-05 修回日期:2026-07-01 发布日期:2026-11-20
  • 作者简介:王杰,男,1994年出生,博士,助理研究员,硕士研究生导师。主要研究方向为大模型与智能运维。E-mail:jackwang@swjtu.edu.cn
  • 基金资助:
    国家自然科学基金(62506311,62506310)、四川省自然科学基金(2026NSFSC1481,2026NSFSC1479)和中国博士后科学基金(2025M771637)资助项目。

DiagAgent: Multi-Agent Framework for Explainable Fault Diagnosis via Chain-of-thought Reasoning

WANG Jie1,2, DUAN Kaiwen2, ZHANG Ji2, CHEN Shiqian1, ZHAI Wanming1   

  1. 1. State Key Laboratory of Rail Transit Vehicle System, Southwest Jiaotong University, Chengdu 610031;
    2. School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756
  • Received:2025-12-05 Revised:2026-07-01 Published:2026-11-20

摘要: 大语言模型的语义建模与链式推理能力,为机械故障诊断由“结果判别”向“依据解释”拓展提供了新途径。现有方法多依赖单模型微调或时频特征直接拼接,难以解耦时域冲击表征与频域调制机理,且缺乏对生成推理过程的物理一致性约束,易导致单视角误判。针对上述问题,提出融合物理先验与多智能体思维链协同的生成式轴承故障诊断框架DiagAgent。首先,构建多视图物理语义重构机制,将时域统计特征和频域包络特征转化为结构化语义描述,实现振动信号数值特征与文本语义空间的对齐;其次,基于参数高效微调构建TimeAgent与FreqAgent异构专家,分别建模时域冲击感知和频域机理匹配逻辑,生成具有显式因果关系的诊断推理链;最后,引入JudgeAgent,从逻辑自洽性和物理一致性两方面校验推理过程,抑制生成幻觉并完成多专家冲突仲裁。实验结果表明,DiagAgent的综合诊断性能优于主流基线模型,在小样本和复杂噪声工况下仍保持稳定;同时可输出包含诊断结果、特征依据与推理逻辑的诊断报告,兼具可解释性与边缘部署潜力,为工业场景下轴承故障智能诊断提供了新方案。

关键词: 故障诊断, 大语言模型, 思维链, 多智能体协同, 可解释性人工智能

Abstract: The semantic modeling and chain-of-thought reasoning capabilities of large language models provide a new pathway for mechanical fault diagnosis, shifting from result-oriented classification to evidence-based explanation. Existing methods commonly rely on single-model fine-tuning or direct concatenation of time-frequency features, making it difficult to decouple time-domain impact signatures from frequency-domain modulation mechanisms. Moreover, the lack of physical consistency constraints on the generated reasoning process may lead to single-perspective misdiagnosis. To address these issues, this paper proposes DiagAgent, a generative bearing fault diagnosis framework that integrates physical priors with multi-agent chain-of-thought collaboration. First, a multi-view physical semantic reconstruction mechanism is developed to convert time-domain statistical features and frequency-domain envelope features into structured semantic descriptions, enabling alignment between numerical vibration features and textual semantic space. Second, a heterogeneous expert architecture consisting of TimeAgent and FreqAgent is constructed through parameter-efficient fine-tuning. The two agents respectively model time-domain impact perception and frequency-domain mechanism matching, generating diagnostic reasoning chains with explicit causal relationships. Finally, a JudgeAgent is introduced to verify the reasoning process in terms of logical self-consistency and physical consistency, thereby suppressing hallucinations and resolving conflicts among multiple expert perspectives. Experimental results demonstrate that DiagAgent achieves superior overall diagnostic performance compared with mainstream baseline models and maintains robustness under few-shot and complex noisy conditions. In addition, it can generate diagnostic reports containing fault categories, feature evidence, and reasoning logic, showing both interpretability and potential for edge deployment. The proposed framework provides a new solution for explainable intelligent bearing fault diagnosis in industrial scenarios.

Key words: fault diagnosis, large language models, chain-of-thought, multi-agent collaboration, explainable AI

中图分类号: