【作者】 许佳豪； 李丹阳； 吴迪龙； 左炎春； 刘伟； 张思楠； 郭立新； 曹育维；

【Author】 Xu Jiahao;LI Danyang;WU Dilong;ZUO Yanchun;LIU Wei;Zhang Sinan;Guo Lixin;Cao Yuwei;Xidian University;

【机构】 西安电子科技大学；

【摘要】 为突破复合材料内部缺陷"可检不可辨"的瓶颈,提出一种基于"测算融合"的电磁—深度学习协同框架,实现缺陷存在性、类别及几何形状的同步高可信识别。以弹跳射线法(SBR)快速生成宽频带、全极化散射场,结合后向投影(BP)算法重构缺陷三维空间分布,构建高置信理论数据集;在微波暗室获取同频段实测散射数据,二者经测算融合生成含形状标签的混合数据集,用以训练轻量化卷积神经网络(CNN)。该网络在12–18 GHz内150 epoch即收敛,缺陷有无及类别识别准确率达100%,混淆矩阵对角化显著;训练与测试曲线均显示120 epoch后准确率趋1、损失趋0,泛化性能优异。该框架以"算–测–学"闭环方式首次实现复合材料内部缺陷的零漏检、精几何识别,为航空、轨道等领域复合结构在线质量监控提供了可行技术路径。更多还原

【Abstract】 To overcome the "detectable but indistinguishable" bottleneck in characterizing internal defects within composite materials,we proposed an electromagnetic–deep-learning collaborative framework based on computational–experimental fusion.This framework enables simultaneous high-confidence identification of defect presence,category,and geometric shape.We rapidly synthesized wideband,fully-polarized scattering fields using the Shooting and Bouncing Rays(SBR) method.We then reconstructed the three-dimensional spatial distribution of defects by using Back-projection(BP) Algorithm,constructing a high-fidelity theoretical dataset.We acquired experimental scattering data within the same 12–18 GHz band in a microwave anechoic chamber.We fused the simulated and experimental datasets to generate a shape-labeled hybrid dataset for training a lightweight Convolutional Neural Network(CNN).The CNN converged within 150 epochs,achieving 100% accuracy for identifying defect presence and category,with the confusion matrix showing significant diagonalization.Both training and testing curves indicate that accuracy approaches 1 and loss approaches 0 after 120 epochs,demonstrating excellent generalization performance.This "compute-measure-learn" closed-loop framework achieves,for the first time,zero missed detection and precise geometric identification of internal defects in composites.It offers a feasible technical pathway for online quality monitoring of composite structures in aerospace and rail transportation applications.更多还原