• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 186-202.doi: 10.3901/JME.260819

• 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

数据不平衡下轴向柱塞泵故障诊断的LLM提示学习方法

袁科研1,2,3, 靳亚强4,5,6,7, 饶猛7, 卢国梁1,2,3, 左明健4,5,6,7   

  1. 1. 山东大学机械工程学院 济南 250061;
    2. 山东大学高效洁净机械制造教育部重点实验室 济南 250061;
    3. 山东大学金属成形高端装备与先进技术全国重点实验室 济南 250061;
    4. 山东大学齐鲁交通学院 济南 250002;
    5. 山东大学高端装备与工程资产智能运维研究中心 济南 250002;
    6. 山东大学高端工程机械智能制造全国重点实验室 济南 250061;
    7. 青岛明思为科技有限公司 青岛 266041
  • 收稿日期:2025-10-31 修回日期:2026-07-05 发布日期:2026-11-20
  • 作者简介:袁科研,男,1997年出生,博士研究生。主要研究方向为旋转机械故障诊断方向。E-mail:sdu_yuankeyan@mail.sdu.edu.cn;左明健(通信作者),男,1962年出生,博士,教授,博士研究生导师。主要研究方向为系统可靠性分析、维护建模与优化、设备健康管理、信号处理、故障诊断、机器学习。E-mail:zuo.mingjian@mingserve.com

LLM Prompt Learning for Fault Diagnosis of Axial Piston Pumps under Data Imbalance

YUAN Keyan1,2,3, JIN Yaqiang4,5,6,7, RAO Meng7, LU Guoliang1,2,3, ZUO Mingjian4,5,6,7   

  1. 1. School of Mechanical Engineering, Shandong University, Jinan 250061;
    2. Key Laboratory of High-efficiency and Clean Mechanical Manufacture, Shandong University, Ministry of Education, Jinan 250061;
    3. National Key Laboratory of High-End Equipment and Advanced Technology for Metal Forming, Shandong University, Jinan 250061;
    4. School of Qilu Transportation, Shandong University, Jinan, 250002;
    5. Center for Intelligent Operation & Health Management of Advanced Equipment & Engineering Assets, Shandong University, Jinan 250002;
    6. State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, Shandong University, Jinan 250061;
    7. Qingdao Mingserve Technology Co., Ltd., Qingdao, 266041
  • Received:2025-10-31 Revised:2026-07-05 Published:2026-11-20

摘要: 轴向柱塞泵的故障数据在实际工业场景中常呈现显著的不均衡性,即大量正常样本与极少量故障样本共存,这使得严重依赖均衡大数据集的传统深度学习方法面临泛化性能下降的挑战。现有方法多通过数据层面的过采样或欠采样来调整分布,未能从根本上提升模型对少数类细微特征的感知能力。近年来,大语言模型(Large language model,LLM)展现出从海量数据中习得通用特征表示的巨大潜力,其强大的上下文学习与特征捕捉能力为样本不平衡故障诊断提供了新思路。为此,提出一种基于大语言模型的不均衡样本故障诊断框架。首先,通过设计的自相关注意力与混合专家分解机制,自适应地解构振动信号,析离出其周期性与趋势性成分;进而,利用一种概率采样稀疏交叉注意力机制,将分解后的特征重新编程为类文本序列,并与特定设计的提示(Prompt)相结合,以引导LLM聚焦于故障的判别性特征;最终,通过故障映射模块实现故障状态的精准分类。在真实数据集上的实验表明,文中方法在少数类故障的诊断精度上表现优异,为不均衡工况下的轴向柱塞泵故障诊断提供了有效方案。

关键词: 轴向柱塞泵, 故障诊断, 数据不平衡, 大语言模型, 提示学习, 小样本学习

Abstract: Fault data for axial piston pumps in real-world industrial settings often exhibits significant class imbalance, characterized by an abundance of normal samples and a scarcity of fault samples. This imbalance poses a substantial challenge to traditional deep learning methods, which rely on balanced, large-scale datasets and consequently suffer from degraded generalization performance. While existing approaches often address this through data-level techniques like oversampling or undersampling, they fail to fundamentally enhance the model's ability to discern subtle features in minority classes. Recent advances in Large Language Models (LLMs) have demonstrated immense potential in learning universal feature representations from massive datasets, and their powerful in-context learning and feature extraction capabilities provide a new perspective for fault diagnosis with imbalanced samples. To this end, we propose a fault diagnosis framework for imbalanced samples based on LLMs. First, an autocorrelation attention mechanism combined with a Mixture-of-Experts (MoE) decomposition module adaptively deconstructs vibration signals to isolate their periodic and trend components. Subsequently, a probabilistic sampling sparse cross-attention mechanism reprograms the decomposed features into text-like sequences. These sequences, integrated with specifically designed prompts, guide the LLM to focus on discriminative fault characteristics. Finally, a fault mapping module achieves precise fault state classification. Experiments on a real-world dataset demonstrate that the proposed method achieves promising diagnostic performance for minority-class faults, providing an effective solution for fault diagnosis of axial piston pumps under imbalanced operating conditions.

Key words: axial piston pump, fault diagnosis, data imbalance, large language model, prompt learning, few-shot learning

中图分类号: