摘要
层状粘接结构广泛应用于航空航天领域,为了保证层状粘接结构的高可靠性,当前大多利用无损检测与深度学习相结合的方式来进行生产和维护期间的检测。然而,大多数检测模型依赖大量标注样本,当标注样本较少时,模型性能显著下降。此外,不同结构组合的多样性增加了模型训练的难度。针对这些问题,提出了一种基于迁移学习的层状粘接结构界面太赫兹(THz)识别方法,利用传输矩阵法分析层状粘接结构迁移的可行性,使用传输矩阵法构建仿真信号数据集,旨在解决由于小样本问题导致的深度学习模型泛化能力不足的问题,提高模型在少量标注样本和多样化结构上的检测性能;融合Focal Loss和MMD损失函数,改善类别不平衡带来的影响,增强模型的跨域泛化能力;分析不同网络层对识别结果的影响,确定模型的微调策略,缩短训练时间并减少计算资源。实验结果表明,该方法在层状粘接结构峰值界面识别中表现出较高的精度与鲁棒性,界面识别率达到99.85%,飞行时间误差为0.1 ps。
Laser layered adhesive structures are widely used in the aerospace field.In order to ensure the high reliability of layered adhesive structures,most of the current methods use a combination of non-destructive testing and deep learning to perform inspections during production and maintenance.However,most detection models rely on a large number of labeled samples.When the number of labeled samples is small,t he performance of the model is significantly reduced.In addition,the diversity of different structural combinations increases the difficulty of model training.Aiming at these problems,this paper proposes a terahertz(THz)identification method for layered adhesive structure interface based on transfer learning.The transfer matrix method is used to analyze the feasibility of layered adhesive structure transfer,and the transfer matrix method is used to construct the simulation signal data set.The purpose is to solve the problem of insufficient generalization ability of deep learning model caused by small sample problem,and improve the detection performance of the model on a small number of labeled samples and diversified structures.The Focal Loss and MMD loss functions are combined to improve the impact of category imbalance and enhance the cross-domain generalization ability of the model.Analyze the influence of different network layers on the recognition results,determine the fine-tuning strategy of the model,shorten the training time and reduce the computing resources.The experimental results show that the method has high accuracy and robustness in the peak interface recognition of layered adhesive structures.The interface recognition rate reaches 99.85%,and the flight time error is 0.1 ps.
作者
任姣姣
何伟涵
张霁旸
牟达
李丽娟
REN Jiaojiao;HE Weihan;ZHANG Jiyang;MU Da;LI Lijuan(School of Opto-Electronic Engineering,Changchun University of Science and Technology,Changchun 130022;Zhongshan Institute of Changchun University of Science and Technology,Zhongshan 528400)
出处
《长春理工大学学报(自然科学版)》
2025年第2期55-68,共14页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省自然科学基金项目(YDZJ202301ZYTS242)
中山市第九批创新科研团队项目(GXTD2022010)。
关键词
迁移学习
太赫兹
层状结构
界面识别
传输矩阵法
transfer learning
terahertz
layered structure
interface recognition
transfer matrix method
