摘要
基于监测数据的结构损伤检测对桥梁运营安全十分重要,然而实际桥梁监测数据的标签不足,导致结构损伤识别方法精度不足。为提高小样本监测数据下桥梁结构的损伤识别精度,提出一种基于特征可迁移数字孪生的结构损伤识别方法。该方法采用数字孪生技术缩小数值模型与实际结构之间的误差,并通过数值模型扩充损伤状态的样本数量,形成了物理和数据双驱动的桥梁结构损伤识别方法。在无数据标签情况下,基于损伤敏感与域不变特征,采用迁移学习方法对数值模型和真实结构数据进行训练,并生成实际监测数据的标签,克服了传统方法仅缩小误差的缺陷。采用斜拉桥缩尺模型测试数据验证了所提方法的有效性。研究结果表明:通过特征可视化程序观察到了源域和目标域特征在低维流行空间中的逐渐对齐过程,显著减小了源域和目标域之间的差异,并揭示了无监督域适应方法的学习机制,解决了跨域的损伤检测问题;在没有标记训练数据的情况下,高精度地识别结构损伤位置。
Structural damage detection based on monitoring data is crucial for bridge operational safety.However,the lack of labeled data in real bridge monitoring often results in insufficient accuracy of damage detection methods.To improve the accuracy of damage detection in bridge structures with small sample monitoring data,a structural damage identification method based on feature-transferable digital twins was proposed.In the method,the digital twin technology was used to reduce the error between numerical models and real structures and the sample size of damage states was expanded through the numerical models,forming a physics-and data-driven approach for bridge structural damage identification.In the case of absence of labeled data,the transfer learning was applied based on damage-sensitive and domain-invariant features to train numerical model and real structure data,generating labels for actual monitoring data,thus overcoming the limitation of traditional methods that can only reduce errors.The effectiveness of the proposed method was validated through test data from a cable-stayed bridge scale model.The results show that by a feature visualization process,the gradual alignment of source and target domain features can be observed in low-dimensional latent space,significantly reducing the discrepancy between the two domains.Additionally,the learning mechanism of unsupervised domain adaptation is revealed,solving the cross-domain damage detection problem,and accurately identifying structural damage locations without labeled training data.
作者
鲁乃唯
崔健
肖向远
罗媛
LU Naiwei;CUI Jian;XIAO Xiangyuan;LUO Yuan(School of Civil and Environmental Engineering,Changsha University of Science and Technology,Changsha 410114,China;School of Civil and Environmental Engineering,Hunan University of Technology,Zhuzhou 412007,China)
出处
《振动与冲击》
北大核心
2026年第2期66-75,共10页
Journal of Vibration and Shock
基金
国家自然科学基金项目(52178108,52408175,52478128)
湖南省自然科学基金项目(2024JJ5033)
深圳市科技计划项目(CJGJZD20220517141800001)。
关键词
桥梁工程
损伤检测
数字孪生
迁移学习
域自适应
无监督学习
bridge engineering
damage detection
digital twin
transfer learning
domain-invariant feature
unsupervised learning
