• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 203-214.doi: 10.3901/JME.260770

• 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

基于语义指导和动态扩展的风电齿轮箱轴承复合故障诊断模型框架

李学艺1,2, 冉光圣1, 李奇2, 王天杨2, 褚福磊2   

  1. 1. 东北林业大学机电工程学院 哈尔滨 150040;
    2. 清华大学机械工程系高端装备界面科学与技术全国重点实验室 北京 100084
  • 收稿日期:2025-11-29 修回日期:2026-07-22 发布日期:2026-11-20
  • 作者简介:李学艺,男,1991年出生,博士,副教授。主要研究方向为故障诊断、智能运维等。E-mail:lixueyi@nefu.edu.cn;王天杨(通信作者),男,1985年出生,博士,副研究员。主要研究方向为现代信号处理、动力学建模、人工智能模式识别、大型旋转机械故障诊断等。E-mail:wty19850925@tsinghua.edu.cn
  • 基金资助:
    国家自然科学基金资助项目(52505091)。

Composite Fault Diagnosis Model for Wind Power Gearbox Bearing Based on Semantic Guidance and Dynamic Extension

LI Xueyi1,2, RAN Guangsheng1, LI Qi2, WANG Tianyang2, CHU Fulei2   

  1. 1. College of Mechanical and Electrical Engineering, Northeast Forestry University, Harbin 150040;
    2. National Key Laboratory of Interface Science and Technology for High-end Equipment, Department of Mechanical Engineering, Tsinghua University, Beijing 100084
  • Received:2025-11-29 Revised:2026-07-22 Published:2026-11-20

摘要: 基于深度学习的风机齿轮箱轴承故障诊断方法通常存在提取特征不可控的问题,并且模型在向未知故障进行扩展时需依赖增量学习,这不仅带来了较高的计算成本,还容易引发灾难性遗忘。针对以上问题,本研究提出一种基于语义指导和动态扩展的复合故障诊断模型框架,该框架构建了振动信号与自然语言指令的协同推理机制。具体而言,模型利用适配器将时域振动信号映射至文本嵌入空间,并通过交叉注意力动态建立信号特征与文本描述的语义关联。这种设计使得文本指令能够显式指导模型关注与故障机理相关的关键特征。当新故障出现时,仅需通过少量样本微调适配器即可实现诊断能力的快速提升。通过对比试验,验证了本框架在复合故障上的准确率达到97.1%。试验还验证了低成本分钟级扩展能力,且对原有故障的诊断精度无显著影响。

关键词: 复合故障, 语义引导, 故障诊断, 多模态融合

Abstract: Feature extraction in deep learning-based fault diagnosis methods for wind turbine gearbox bearings is generally uncontrollable. Furthermore, when extending these models to unknown faults, they rely on incremental learning, which not only incurs high computational costs but is also prone to catastrophic forgetting. To address these issues, this study proposes a compound fault diagnosis model framework based on semantic guidance and dynamic expansion. This framework constructs a collaborative inference mechanism between vibration signals and natural language instructions. Specifically, the model utilizes an adapter to map time-domain vibration signals into a text embedding space, and dynamically establishes semantic correlations between signal features and textual descriptions via cross-attention. This design enables text instructions to explicitly guide the model to focus on key features related to fault mechanisms. When new faults emerge, rapid enhancement of diagnostic capabilities can be achieved simply by fine-tuning the adapter with a small number of samples. Comparative experiments verify that the proposed framework achieves an accuracy of 97.1% on compound faults. The experiments also demonstrate a low-cost, minute-level expansion capability without significantly impacting the diagnostic accuracy of the original faults.

Key words: compound faults, semantic guidance, fault diagnosis, multi-modal fusion

中图分类号: