• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2026, Vol. 62 ›› Issue (18): 157-168.doi: 10.3901/JME.260778

• 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

基于多尺度时间序列大模型的航空交点孔精密镗削刀具状态监测方法

杨展鹏1,2, 华丰3, 许云皓4, 赵鹏远1, 王映均2, 宋海涛5, 程骋1,2   

  1. 1. 华中科技大学人工智能与自动化学院 武汉 430074;
    2. 高端航空装备技术创新中心(四川)有限公司 成都 610000;
    3. 华中科技大学机械科学与工程学院 武汉 430074;
    4. 昆山杜克大学数学与应用数学系 昆山 215316;
    5. 上海人工智能研究院 上海 201100
  • 收稿日期:2025-11-02 修回日期:2026-05-31 发布日期:2026-11-20
  • 作者简介:杨展鹏,男,2001年出生。主要研究方向为数据驱动的机理建模。E-mail:zhanpeng_yang@hust.edu.cn;程骋(通信作者),女,1990年出生。主要研究方向为航空装备智能运维。E-mail:c_cheng@hust.edu.cn
  • 基金资助:
    国家自然科学基金(92467301,T2525012,62373160,52205520,62503185)、航空国创科学基金(010601012502)和湖北省自然科学基金(2024AFB564)资助项目。

Multi-scale Large Time-Series Model for Tool Condition Monitoring in Precision Boring of Airplane Intersection Hole

YANG Zhanpeng1,2, HUA Feng3, XU Yunhao4, ZHAO Pengyuan1, WANG Yingjun2, SONG Haitao5, CHENG Cheng1,2   

  1. (1. School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan 430074;
    2. National Center (Sichuan) of Technology Innovation for Advanced Aviation Equipment Corporation, Chengdu 610000;
    3. School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan 430074;
    4. Applied Mathematics and Computational Sciences/Computer Science, Duke Kunshan University, Kunshan 215316;
    5. Shanghai Artificial Intelligence Research Institute, Shanghai 201100
  • Received:2025-11-02 Revised:2026-05-31 Published:2026-11-20

摘要: 交点孔的加工质量直接关系到飞机的结构性能与疲劳寿命。其镗削过程涉及机床-刀具-工件-夹具的多体耦合,系统状态具有强动态时变性,导致不同刀具间的监测信号存在显著的分布差异。传统端到端模型受限于训练数据的特定分布,难以对全新刀具的全寿命周期状态进行准确泛化。为此,提出一种基于多尺度时序大模型的刀具状态监测方法,该方法引入旋转位置编码与多尺度注意力机制,构建了针对高频非平稳切削信号的通用表征学习架构;采用“无监督预训练-小样本微调”范式,利用海量未标注数据学习信号的本质特征,并通过少量标注样本实现下游监测任务的快速适配。在三轴数控机床交点孔镗削平台上进行验证,实验结果表明,该方法在预测精度与鲁棒性上均显著优于传统深度学习模型;对后刀面磨损宽度的预测平均绝对误差和均方根误差均控制在3 mm以内,验证了时序大模型在高端装备智能运维任务中的有效性与应用潜力。

关键词: 刀具状态监测, 镗削加工, 时序大模型

Abstract: The machining quality of intersection holes is directly related to the structural performance and fatigue life of airplane. The boring process involves a multi-body coupling of the machine, tool, workpiece, and fixture, where the system state exhibits strong dynamic time-varying characteristics, resulting in significant distribution shifts in monitoring signals among different tools. Conventional end-to-end models, limited by the specific distribution of training data, struggle to accurately generalize to the full life-cycle state of new tools. To address this, a tool condition monitoring method based on a multi-scale time-series large model is proposed. Rotary Position Embedding and multi-scale attention mechanisms are introduced to construct a general representation learning architecture for high-frequency non-stationary cutting signals. An "unsupervised pre-training and few-shot fine-tuning" paradigm is adopted, utilizing massive unlabeled data to learn intrinsic signal features and achieving rapid adaptation to downstream monitoring tasks via limited labeled samples. Validation is conducted on a three-axis CNC boring platform for intersecting holes. Experimental results indicate that the proposed method significantly outperforms traditional deep learning models in terms of prediction accuracy and robustness. The prediction errors for flank wear width, measured by Mean Absolute Error and Root Mean Square Error, are controlled within 3μm. The effectiveness and application potential of the time-series large model in intelligent operation and maintenance tasks for high-end equipment are thus verified.

Key words: tool condition monitoring, boring process, large time series model

中图分类号: