• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 51-66.doi: 10.3901/JME.260761

• 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

融合工况先验的多模态动态路由大模型RUL预测方法

刘卫朋1, 刘赛1, 胡启迪2, 王佶宣1   

  1. 1. 河北工业大学人工智能与数据科学学院 天津 300401;
    2. 湖北工程职业学院智能制造学院 黄石 435002
  • 收稿日期:2026-01-30 修回日期:2026-06-25 发布日期:2026-11-20
  • 作者简介:刘卫朋,男,1979年出生,博士,研究员,博士研究生导师。主要研究方向为智能机器人技术、工业软件。E-mail:liuweipeng@hebut.edu.cn;刘赛,男,2002年出生,硕士研究生。主要研究方向为智能感知与先进检测。E-mail:nameliusai@163.com;胡启迪,男,1995年出生,博士研究生。主要研究方向为机器人结构设计与控制、工业软件。E-mail:marshellhu@qq.com;王佶宣(通信作者),女,1995年出生,博士,讲师。主要研究方向为类脑智能算法与闭环控制。E-mail:wangjixuan@hebut.edu.cn
  • 基金资助:
    国家重点研发计划资助项目(2024YFB3310900)。

Multi-Modal Dynamic Routing Large Language Model with Integrated Operating Condition Prior for RUL Prediction

LIU Weipeng1, LIU Sai1, HU Qidi2, WANG Jixuan1   

  1. 1. School of Artificial Intelligence and Data Science, Hebei University of Technology, Tianjin 300401;
    2. School of Intelligent Manufacturing, Hubei Engineering Polytechnic, Huangshi 435002
  • Received:2026-01-30 Revised:2026-06-25 Published:2026-11-20

摘要: 工业设备剩余使用寿命(Remaining useful life,RUL)预测对预测性维护至关重要,但数据稀缺、工况多变及模型泛化能力不足等挑战制约了其预测能力。大语言模型(Large language model,LLM)在时序预测中展现出强大的小样本学习能力,然而现有方法通常采用全局一致的微调策略,缺乏对样本特征和模型内部模块的差异化控制,限制了对异质工业数据中多样化退化模式的有效识别和预测。为此,提出一种改进的融合工况先验的多模态动态路由大语言模型(Operating condition prior-guided multi-modal dynamic routing large language model,PMR-LLM)。该方法通过跨模态注意力机制实现时序特征与文本提示的语义对齐,构建融合工况先验的动态路由机制,并在各Transformer层的6个关键模块独立部署低秩适应(Low-rank adaptation,LoRA)专家池,实现工况感知的细粒度参数适配。在CMAPSS和PHM2010数据集上的实验结果表明,在10%小样本场景下,PMR-LLM相比次优基线方法RMSE分别平均降低7%和13%,并在有无工况标签的跨场景泛化中均表现出较强的适应能力。

关键词: 剩余使用寿命预测, 大语言模型, 小样本学习, 动态路由, 工况先验

Abstract: Predicting the remaining useful life (RUL) of industrial equipment is critical for predictive maintenance,yet its predictive capability is hindered by challenges such as data scarcity,variable operating conditions,and limited model generalization. Large language model (LLM) has demonstrated strong few-shot learning capabilities in time-series forecasting; however,existing methods typically employ globally consistent fine-tuning strategies,lacking differentiated control over sample characteristics and internal model modules,which limits the ability to effectively identify and predict diverse degradation patterns in heterogeneous industrial data. To address this,an improved multi-modal dynamic routing large language model guided by operating condition prior (PMR-LLM) is proposed. Semantic alignment between temporal features and textual prompts is achieved through a cross-modal attention mechanism,a dynamic routing mechanism incorporating operating condition prior is constructed,and low-rank adaptation (LoRA) expert pools are independently deployed across six key modules in each Transformer layer,enabling operating-condition-aware fine-grained parameter adaptation. Experimental results on the CMAPSS and PHM2010 datasets show that,under the 10% few-shot setting,PMR-LLM reduces RMSE by an average of 7% and 13% compared to the second-best baseline methods,respectively,and demonstrates strong adaptability in cross-scenario generalization both with and without operating condition labels.

Key words: remaining useful life prediction, large language model, few-shot learning, dynamic routing, operating condition prior

中图分类号: