• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 36-50.doi: 10.3901/JME.260779

• 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

工业大模型赋能的旋转机械装备变工况故障诊断方法

王子胜1,2,3, 许朝峻1,3, 靳亚强1,3, 饶猛4   

  1. 1. 山东大学高端工程机械智能制造全国重点实验室 济南 250061;
    2. 中南大学机电工程学院 长沙 410083;
    3. 山东大学齐鲁交通学院 济南 250061;
    4. 青岛明思为科技有限公司 青岛 266041
  • 收稿日期:2025-10-22 修回日期:2026-04-30 发布日期:2026-11-20
  • 作者简介:王子胜,男,1996年出生,博士,特聘副教授。主要研究方向为工业大模型与具身智能体、智能制造与智能运维等。E-mail:wangzs@csu.edu.cn

Fault Diagnosis Methods for Rotating Machinery under Variable Operating Conditions Empowered by Industrial Large-scale Models

WANG Zisheng1,2,3, XU Chaojun1,3, JIN Yaqiang1,3, RAO Meng4   

  1. 1. State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, Shandong University, Jinan 250061;
    2. College of Mechanical and Electrical Engineering, Central South University, Changsha 410083;
    3. School of Qilu Transportation, Shandong University, Jinan 250061;
    4. Qingdao Mingserve Technology Ltd., Qingdao 266041
  • Received:2025-10-22 Revised:2026-04-30 Published:2026-11-20

摘要: 面向变工况下旋转机械装备故障诊断,提出一种基于提示词工程增强机制的工业大模型。该方法从振动信号中提取多域故障特征,并结合转速工况构建结构化文本,将多源数值特征映射到语言空间。在此基础上,设计由角色、任务和输出三层组成的提示体系,引导大模型按“症状识别-机理推理-严重度判定”执行链式推理,嵌入机械先验与诊断约束。针对早期故障样本少、类别不平衡等问题,进一步引入异常样本感知的自适应提示优化机制,通过难度评分与更新策略自动调整提示模板,提高弱故障的识别稳定性。实验与工程验证结果均表明,与传统深度学习方法及固定提示的大模型基线相比,提出的方法在变转速工程数据集上获得更高的故障识别精度、更一致的严重度分级,并提升诊断解释性。

关键词: 故障诊断, 旋转机械, 大语言模型, 提示词工程, 思维链

Abstract: For fault diagnosis of rotating machinery under varying operating conditions, we proposes a prompt engineering enhanced industrial large-scale model. This method extracts multi-domain fault features from vibration signals and constructs structured text by combining rotational speed conditions, mapping multi-source numerical features to the language space. Based on this, a three-layer prompt system consisting of roles, tasks, and outputs is designed to guide the large model to perform chained reasoning according to "symptom identification-mechanism reasoning-severity determination", embedding mechanical priors and diagnostic constraints. To address the problems of limited early fault samples and class imbalance, an adaptive prompt optimization mechanism that perceives abnormal samples is further introduced. The prompt template is automatically adjusted through difficulty scoring and update strategies to improve the stability of weak fault identification. Verification results from experiment and engineering platforms show that, compared with traditional deep learning methods and large model baselines with fixed prompt, the proposed method achieves higher fault identification accuracy, more consistent severity classification, and improved diagnostic interpretability on variable-speed engineering datasets.

Key words: fault diagnosis, rotating machinery, large language model, prompt engineering, chain of thought

中图分类号: