• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 114-124.doi: 10.3901/JME.260774

• 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

大模型驱动的旋转机械多场景通用故障诊断

邵海东1, 王杰1, 张家乐2, 肖一鸣1, 王海宇1   

  1. 1. 湖南大学机械与运载工程学院 长沙 410082;
    2. 中国船舶科学研究中心 无锡 214062
  • 收稿日期:2026-01-20 修回日期:2026-06-02 发布日期:2026-11-20
  • 作者简介:邵海东(通信作者),男,1990年出生,博士,副教授,博士研究生导师。主要研究方向为故障诊断与寿命预测,数据挖掘与信息 融合。E-mail:hdshao@hnu.edu.cn;王杰,男,1999年出生,博士研究生。主要研究方向为机械智能故障诊断与工业大数据分析。E-mail:jiewang1013@hnu.edu.cn
  • 基金资助:
    江苏省基础研究计划(基础研究重点项目)(BK20253037)、国家自然科学基金(52275104)、湖南省科技创新计划(2023RC3097)、湖南大学整车先进设计制造技术全国重点实验室自主研究课题(72675005)和湖南省研究生科研创新(CX20250537)资助项目。

Large Language Model-driven Generalized Fault Diagnosis Method for Rotating Machinery Across Multiple Scenarios

SHAO Haidong1, WANG Jie1, ZHANG Jiale2, XIAO Yiming1, WANG Haiyu1   

  1. 1. College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410082;
    2. China Ship Scientific Research Center, Wuxi 214062
  • Received:2026-01-20 Revised:2026-06-02 Published:2026-11-20

摘要: 近年来,大模型在多个领域取得突破性进展。卓越的泛化与逻辑推理能力为建立协同应对多源数据集、跨工况迁移及数据不平衡等多场景的通用故障诊断框架提供了可能性。然而,由于连续的振动信号与离散的文本之间存在着显著的模态差异,大模型难以直接迁移至时序数据分析任务。因此,如何有效适配并利用大模型处理时序数据,仍是当前故障诊断领域面临的一项关键挑战。为此,提出了大模型驱动的机械设备多场景通用故障诊断方法。首先,设计了信号-文本协同表征模块,通过从原始振动信号中生成文本描述,并将两者联合编码,让故障特征同时具备多样化的数据形态与深层的语义表达。其次,构造了跨模态特征对齐与融合模块,采用从浅层到深层的渐进式对齐策略,实现了文本特征与时序信号特征的深度融合。最后,基于大模型强大的泛化与逻辑推理能力,构建一个统一的新型故障诊断框架。通过在多个数据集,跨工况和数据类别分布不均衡等多场景下进行综合实验验证,结果表明,所提方法不仅能显著提高故障诊断精度,同时也具备良好的模型鲁棒性与跨场景泛化能力。

关键词: 大模型, 跨模态, 多场景, 旋转机械, 故障诊断

Abstract: In recent years, large language models (LLMs) have achieved breakthrough progress across multiple domains. The exceptional capabilities in generalization and reasoning offer the potential to establish a universal fault diagnosis framework capable of addressing different scenarios, including multiple datasets, cross-condition transfer, and data imbalance. However, owing to the significant modal differences between continuous vibrational signals and discrete textual data, LLMs struggle to be directly applied to time-series data analysis tasks. Therefore, how to effectively adapt and leverage LLMs for processing time-series data in fault diagnosis remains a critical challenge. To this end, this paper proposes a LLM-driven generalized fault diagnosis method for mechanical equipment across multiple scenarios. Firstly, a signal-text collaborative representation module is developed to generate textual descriptions from raw vibration signals and jointly encode both modalities, enabling fault features to exhibit both diverse data modalities and rich semantic representations. Second, a cross-modal feature alignment and fusion module is constructed, which adopts a progressive alignment strategy from shallow to deep layers to achieve deep integration of textual and time-series signal features. Finally, leveraging the strong generalization and reasoning capabilities of LLM, a generalized fault diagnosis framework is established. Extensive experimental results across multiple datasets, cross-condition transfer, and data imbalance scenarios verify that the proposed method delivers substantial improvements in fault diagnosis accuracy, along with robust and transferable generalization performance.

Key words: large language model, cross modal, multiple scenarios, mechanical equipment, fault diagnosis

中图分类号: