CFRP低速冲击损伤成像精度提升算法研究  

Research on algorithm for improving imaging accuracy of CFRP low speed impact damage

作　　者：吴项南 程小劲 李琪鑫 尚建华[2] WU Xiangnan;CHENG Xiaojin;LI Qixin;SHANG Jianhua(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620,China;School of Information Science and Technology,Donghua University,Shanghai 201620,China)

机构地区：[1]上海工程技术大学机械与汽车工程学院,上海201620 [2]东华大学信息科学与技术学院,上海201620

出　　处：《航空材料学报》2025年第1期80-90,共11页Journal of Aeronautical Materials

基　　金：国家自然科学基金面上项目(52173219)。

摘　　要：碳纤维增强树脂基复合材料(carbon fiber reinforced polymer,CFRP)遭受低速冲击后会发生微小且隐蔽的损伤,损伤的存在会显著降低CFRP材料的承载力和服役寿命。C扫描是超声成像的常用方法,针对CFRP低速冲击内部损伤的C扫描成像精度低这一问题,使用梯度算子对原始图像进行处理,并利用迁移学习的方法在ResNet18与ResNet50上进行损伤类型的分类训练。为了改善分类模型的性能,提出基于卷积神经网络的成像精度提升算法——图像重建模型(image reconstruction model,IRM),并基于结构相似性系数(structural similarity index,SSIM)提出采用性能指标σEOL验证图像性能的提升水平。迭代训练结果表明,当迭代次数达到200次时,不同种类冲击损伤的σEOL均大于1。为了进一步提升成像精度,引入ResNet残差连接思想,并提出了ResIRM网络。与IRM相比,ResIRM对不同类型撞击损伤的检测精度进一步提升,针对全部冲击类型,平均σEOL可提升0.85%;同时,经ResIRM处理的图像分类模型的梯度显著性热力图表明,ResIRM可以对损伤区域的特征起到强化作用。Carbon fiber reinforced polymer(CFRP)composites has small and hidden damage after low-speed impact,and the existence of damage significantly reduces the bearing capacity and service life of CFRP materials.C-scan represents a conventional ultrasonic imaging method.To address the issue of low imaging precision in C-scan detection of internal damage caused by lowvelocity impact in CFRP,gradient operators were employed to process the original images,and transfer learning methodology was utilized to conduct damage classification training on ResNet18 and ResNet50 architectures.To enhance the classification model’s performance,an image reconstruction model(IRM)based on convolutional neural networks was proposed to improve imaging precision.Additionally,a performance metricσEOL,based on the structural similarity index(SSIM),was introduced to validate the level of image quality enhancement.The iterative training results demonstrate that when the iteration count reaches 200,theσEOL of different types of impact damage is greater than 1.To further improve imaging precision,the ResNet residual connection concept is incorporated,leading to the development of the ResIRM network.Compared to IRM,ResIRM exhibits enhanced detection precision for different types of impact damage,with an averageσEOL improvement of 0.85%across all impact types.Furthermore,the gradient saliency heat maps of the classification model processed by ResIRM indicate that ResIRM effectively reinforces the features in damaged regions.

关 键 词：卷积神经网络 无损检测 损伤重建 超声检测 

分 类 号：V258[一般工业技术—材料科学与工程] TP183[航空宇航科学与技术—航空宇航制造工程]