• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2025, Vol. 61 ›› Issue (6): 53-65.doi: 10.3901/JME.2025.06.053

• 仪器科学与技术 • 上一篇    

扫码分享

基于改进迁移学习的CFRP缺陷涡流检测与识别

程军1,2,3, 朱煜龙1,4, 汪步云1,2,3, 梁艺1,2,3, 许德章1,2,3, 刘蒙蒙1,4, 刘有余1,4   

  1. 1. 安徽工程大学安徽省机器人产业共性技术研究中心 芜湖 241007;
    2. 安徽工程大学人工智能学院 芜湖 241000;
    3. 芜湖云擎机器人科技有限公司 芜湖 241007;
    4. 安徽工程大学机械工程学院 芜湖 241000
  • 收稿日期:2024-06-01 修回日期:2025-01-20 发布日期:2025-04-14
  • 作者简介:程军(通信作者),男,1986年出生,博士,副教授,硕士研究生导师。主要研究方向为复合材料无损检测、机器视觉与图像处理。E-mail:chengjun@ahpu.edu.cn
  • 基金资助:
    国家自然科学基金(51605229)、教育部重点实验室开放课题(KFKT202209)、安徽省高校自然科学基金(2023AH050928)、安徽省经信委制造业重点领域揭榜挂帅(JB22031)和安徽未来技术研究院企业合作(2023qyhz35)资助项目。

Improved Transfer Learning Method for Eddy Current Testing and Identification of CFRP Defects

CHENG Jun1,2,3, ZHU Yulong1,4, WANG Buyun1,2,3, LIANG Yi1,2,3, XU Dezhang1,2,3, LIU Mengmeng1,4, LIU Youyu1,4   

  1. 1. Anhui Research Center for Generic Technologies in Robot Industry, Anhui Polytechnic University, Wuhu 241007;
    2. School of Artificial Intelligence, Anhui Polytechnic University, Wuhu 241000;
    3. Wuhu Yunqing Robot Technology Co., Ltd., Wuhu 241007;
    4. School of Mechanical Engineering, Anhui Polytechnic University, Wuhu 241000
  • Received:2024-06-01 Revised:2025-01-20 Published:2025-04-14

摘要: 近年来,碳纤维增强树脂基复合材料(Carbon fiber reinforced polymer,CFRP)缺陷的自动化检测与识别已经成为无损检测领域的热点课题之一。然而,CFRP构件缺陷数据的不足会产生过拟合现象,导致缺陷识别精度低等问题。基于此,提出了一种基于改进迁移学习的CFRP材料缺陷涡流检测方法。首先利用涡流C扫描成像和复平面信号特征提取获得CFRP缺陷样本并进行数据增强,解决了样本不足的问题;然后利用MobileNet V2网络与K-means聚类的方法,在热轧带钢表面缺陷数据集中挑选出特征相似源域图片进行预训练,完成目标域相似特征的提取,减小“负迁移”的影响;最后在特征提取网络中融合卷积注意力模块,减少特征图中背景特征信息的影响,再通过模型迁移的方式将源域训练的网络权值迁移到改进后的Faster R-CNN目标检测模型中,建立CFRP缺陷检测模型。通过对比试验,该方法有效解决了CFRP构件缺陷数据较少的问题,并且具有较高的准确率和鲁棒性。构建出的CFRP缺陷检测模型实现了对裂纹、分层、褶皱三类缺陷的高精度识别,平均精度均值达到了94.62%,相较于传统训练方法与原始网络迁移学习检测精度分别提升了29.31%与2.79%,尤其对于褶皱缺陷的识别精度显著提高,取得了较好的效果,满足了对CFRP构件缺陷的检测要求。

关键词: 涡流检测, 碳纤维复合材料, 迁移学习, K-means聚类, 卷积神经网络

Abstract: In recent years, automated detection and identification of defects in carbon fiber reinforced polymer (CFRP) composites has emerged as a hot topic in the field of nondestructive testing. However, insufficient defect data of CFRP components might result in overfitting and low defect recognition accuracy. Therefore, an eddy current testing approach for CFRP defects using improved migration learning is proposed. Firstly, the methods of eddy current C-scan imaging and complex plane signal feature extraction are used to obtain CFRP defect samples and implement data enhancement, which solves the problem of insufficient samples. The MobileNet V2 network with K-means clustering is then used to pick out source domain images with similar features in the hot rolled strip surface defect data set for pre-training, to complete the extraction of similar features in the target domain and reduce the negative migration. Finally, the convolutional attention module is integrated into the feature extraction network to reduce the influence of background feature information in the feature map. The network weights trained on the source domain are transferred to the improved Faster R-CNN object detection model through model transfer, to establish a CFRP defect detection model. Through comparison tests, the method efficiently overcomes the problem of less defect data for CFRP components while maintaining high accuracy and robustness. The constructed CFRP defect detection model achieves high-precision recognition of three types of defects, namely, crack, delamination, and wrinkle. The mean average precision reaches 94.62%, which is 29.31% and 2.79% higher than that of the traditional training method and the original network migration learning detection accuracy, respectively. The recognition accuracy of wrinkle defects in particular is significantly improved, which achieves a better result and meets the requirements for CFRP defect detection.

Key words: eddy current testing, CFRP, transfer learning, K-means clustering, convolutional neural network

中图分类号: