• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 134-146.doi: 10.3901/JME.260777

• 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

基于多任务特征关联大模型的旋转机械故障预测健康管理研究

李川1, 郭改芳1, 郭圳2, 蒲自强1   

  1. 1. 重庆工商大学装备系统服役健康保障国际联合研究中心 重庆 400067;
    2. 武汉理工大学交通与物流工程学院 武汉 430063
  • 收稿日期:2025-10-12 修回日期:2026-06-18 发布日期:2026-11-20
  • 作者简介:李川(通信作者),男,1975年出生,博士,特聘教授。主要研究方向为机械故障诊断和智能运维。E-mail:chuanli@21cn.com;郭改芳,女,2003年出生。主要研究方向为机械故障诊断和智能运维。E-mail:guogaifang@ctbu.edu.cn;郭圳,男,1995年出生,博士研究生。主要研究方向为机械故障诊断和智能运维。E-mail:gz9522@whut.edu.cn;蒲自强,男,1993年出生,博士,讲师。主要研究方向为机械故障诊断和智能运维。E-mail:puziqiang@ctbu.edu.cn
  • 基金资助:
    国家重点研发计划(2023YFB3406100)、国家自然科学基金(52575103)、重庆市自然科学基金(CSTB2025NSCQ-LZX0128)和中央高校基本科研业务费(3120600100)资助项目。

Multitask Feature-correlated Large Language Model for Prognostics and Health Management of Rotating Machinery

LI Chuan1, GUO Gaifang1, GUO Zhen2, PU Ziqiang1   

  1. 1. Research Center for System Health Maintenance, Chongqing Technology and Business University, Chongqing 400067;
    2. School of Transportation and Logistics Engineering, Wuhan University of Technology, Wuhan 430063
  • Received:2025-10-12 Revised:2026-06-18 Published:2026-11-20

摘要: 旋转机械的故障预测健康管理(Prognostics and health management,PHM)需要综合处理异常检测、故障分类、寿命预测和维护建议等多项任务,但是现有方法主要聚焦单任务PHM建模。为此,提出一种多任务特征关联大型语言模型,解决旋转机械的综合PHM问题。设计新型的故障寿命关联网络,将振动信号关联为兼具瞬时判别能力和寿命退化敏感性的统一深度特征。构建分支式多任务枢纽,同时生成异常检测、故障分类和寿命预测结果。提出信号与文本对齐机制,将数值信号估计映射到语义嵌入空间以实现跨模态融合。采用低秩适配对大模型进行微调,生成自适应维护建议。在旋转机械的5个公开数据集、1个自建试验台和1个标准型号试验台上开展对比实验,结果表明,所提出的大模型在旋转机械综合PHM任务中性能优于现有单任务模型。

关键词: 特征关联, 大模型, 旋转机械, 故障预测健康管理, 多任务学习

Abstract: Prognostics and health management (PHM) of rotating machinery demands integrated handling of multiple tasks such as anomaly detection, fault classification, lifespan prediction, and maintenance advice. However, existing methods predominantly focus on single-task PHM modelling. To address this issue, a multitask feature-correlated large language model (MFC-LLM) is proposed for comprehensive PHM of rotating machinery. A novel fault-to-life correlation network is designed to convert vibration signals into unified deep features with both instantaneous discrimination ability and lifespan degradation sensitivity. A branched multitask hub is constructed to simultaneously generate estimates for anomaly detection, fault classification, and lifespan prediction. A signal-text alignment mechanism is developed to map numerical signal estimates into the semantic embedding space for cross-modal fusion. Low-rank adaptation is employed for fine-tuning of MFC-LLM and generating adaptive maintenance advice. Comparative experiments are conducted on 5 public rotating machinery datasets, a self-built experimental setup, and a standard test rig. Results demonstrate that the proposed MFC-LLM outperforms existing single-task models in comprehensive PHM tasks for rotating machinery.

Key words: feature correlation, large language model, rotating machinery, prognostics and health management, multitask learning

中图分类号: