• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2025, Vol. 62 ›› Issue (6): 247-256.doi: 10.3901/JME.260191

• 仪器科学与技术 • 上一篇    

扫码分享

基于卷积神经网络注意力机制模块的复合材料损伤自动识别研究

贾康康1, 都军民2,3, 陈飞宇2,3, 时建纬1, 侯国义4, 李成1   

  1. 1. 郑州大学机械与动力工程学院 郑州 450001;
    2. 河南省水下智能装备重点实验室 郑州 450015;
    3. 中国船舶集团有限公司第七一三研究所 郑州 450015;
    4. 南京航空航天大学机电学院 南京 210000
  • 收稿日期:2025-05-08 修回日期:2025-12-15 发布日期:2026-05-12
  • 作者简介:贾康康,男,1992年出生。主要研究方向为先进复合材料自动化检测技术。E-mail:202022202013941@gs.zzu.edu.cn
    李成(通信作者),男,1962年出生,教授,博士研究生导师。主要研究方向为复合材料损伤机理,无损检测技术。E-mail:chengli@zzu.edu.cn
  • 基金资助:
    国家自然科学基金(52175153)和河南省水下智能装备重点实验室开放基金(ZT22064U)资助项目。

AM-1DCNN Based Automatic Recognition of Mechanical Damage in Composite Material

JIA Kangkang1, DU Junmin2,3, CHEN Feiyu2,3, SHI Jianwei1, HOU Guoyi4, LI Cheng1   

  1. 1. School of Mechanical and Power Engineering, Zhengzhou University, Zhengzhou 450001;
    2. Henan Key Laboratory of Underwater Intelligent Equipment, Zhengzhou 450015;
    3. 713th Research Institute of China State Shipbuilding Corporation Limited, Zhengzhou 450015;
    4. College of Mechanical&Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210000
  • Received:2025-05-08 Revised:2025-12-15 Published:2026-05-12

摘要: 为解决碳纤维增强复合材料(Carbon fiber reinforced polymer, CFRP)机械损伤自动化识别困难的瓶颈,本研究借助注意力机制模块(Attention module,AM)和一维卷积神经网络(One dimensional convolutional neural network, 1DCNN),基于非线性超声检测技术提出一种CFRP机械损伤的自动化判断方法。本研究依托超声检测平台试验测量及有限元仿真的方法,获取CFRP在不同类型机械损伤情形下的超声信号,建立数据集并训练模型。训练完成后,使用前述模型对含损伤CFRP的非线性超声检测信号进行自动化判断识别,成功判断CFRP未损伤、穿孔损伤、劈裂损伤和低速冲击损伤等不同状态。研究结果表明,本研究所提模型可扩展性强,收敛速度较快,对CFRP常见机械损伤类型的识别准确率达96.79%,较传统CNN模型提高2.07%,并为CFRP复合材料服役环境损伤的高效自动化诊断奠定理论基础。

关键词: 复合材料, 一维卷积神经网络, 注意力机制模块, 损伤识别

Abstract: In order to solve the bottleneck of automatic identification of mechanical damage of carbon fiber reinforced polymer (CFRP), an attention module (AM) and one dimensional convolutional neural network (1DCNN) damage identification method was proposed, based on the nonlinear ultrasonic detection technique. Firstly, one-dimensional ultrasonic vibration signals of different mechanical damage types of composite materials were obtained by ultrasonic testing platform and finite element simulation method. After the model training was completed, the model was used to judge the damage type of the test data set of carbon fiber composites. Finally, try to identify the signals of different damage types of composites, such as undamaged, perforated, split and low-speed impact damage. The results show that the proposed model has strong scalability and fast convergence, and the recognition accuracy of common mechanical damage types of CFRP is 96.79%, which is 2.07% higher than that of the traditional CNN model, laying a foundation for the automatic diagnosis of damage in CFRP composites in service environment in universities.

Key words: composite material, one-dimensional convolution neural network, attention module, damage identification

中图分类号: