基于元学习的结构损伤定位与量化  

Meta-learning for Structural Damage Localization and Quantification

作　　者：王磊 黄杜康 马亚飞 黄可 WANG Lei;HUANG Du-kang;MA Ya-fei;HUANG Ke(School of Civil Engineering,Changsha University of Science&Technology,Changsha 410114,Hunan,China)

机构地区：[1]长沙理工大学土木工程学院,湖南长沙410114

出　　处：《中国公路学报》2024年第11期38-51,共14页China Journal of Highway and Transport

基　　金：国家重点研发计划项目(2021YFB2600900);国家自然科学基金项目(52008037);湖南省自然科学基金项目(2024JJ5026);湖南省科技创新计划项目(2023GK2036);湖南省研究生科研创新项目(CX20220856)。

摘　　要：针对现有基于深度神经网络的结构损伤识别方法需要大量标记数据的不足,提出了一种基于元学习的结构损伤定位和量化方法。首先,基于人工神经网络建立结构损伤定位和量化模型,学习结构模态参数(频率、振型)与刚度参数之间的非线性映射关系;其次,利用模型无关元学习方法训练损伤定位和量化模型,优化人工神经网络的初始权值参数,以提升损伤定位和量化模型在少量数据下的泛化性能;最后,在少量模态数据下基于训练模型进行快速学习,实现结构损伤的定位与量化。该方法利用模型无关元学习训练策略获取先验知识,可在少量训练数据条件下加速新结构损伤定位和量化任务的学习过程。采用3跨桥梁数值算例和Z24桥工程实例验证所提方法的有效性。结果表明:在少量模态数据下,该方法能有效且准确地定位和量化结构损伤;与传统的人工神经网络方法和迁移学习方法相比,该方法具有更快的收敛速度和更高的识别精度。Existing deep-learning-based methods for structural damage identification rely heavily on massive amounts of labeled data.Therefore,a meta-learning-based approach is proposed for structural damage localization and quantification.First,a structural damage localization and quantification model was established using an artificial neural network.This model was used to learn the nonlinear mapping relationship between structural modal data(frequency and mode shape)and substructure stiffness parameters.Second,a model-agnostic meta-learning strategy was used to train the damage localization and quantification model.The generalizability of the damage localization and quantification models can be improved by optimizing the initial weight parameters of the artificial neural network(ANN).The proposed method utilizes a model-agnostic meta-learning training strategy to acquire prior knowledge,thereby accelerating the learning process for new structural damage localization and quantification tasks with limited training data.The method was verified on a numerical three-span bridge and benchmark project of the Z24 bridge.The results demonstrate that the proposed approach provides efficient and accurate localization and quantification of potential structural damage using limited data.Compared with conventional ANN and transfer learning methods,the method exhibited faster convergence and higher identification accuracy.

关 键 词：桥梁工程 损伤定位和量化 元学习 模态数据 健康监测 人工神经网络 

分 类 号：U448[建筑科学—桥梁与隧道工程]