基于迁移卷积神经网络的桥梁结构损伤识别方法  

Structural damage identification for bridges based on transfer learning and 1D convolutional neural network

作　　者：罗旭欣 陈龙 梁韬 黄天立[1] LUO Xuxin;CHEN Long;LIANG Tao;HUANG Tianli(School of Civil Engineering,Central South University,Changsha 410075,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075

出　　处：《铁道科学与工程学报》2024年第9期3888-3899,共12页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(52078486)。

摘　　要：针对实际桥梁结构损伤模式识别时有限元模型与实际结构存在差异的情况,为了提高有限元数值模拟数据集训练的深度神经网络识别实际桥梁结构损伤模式的准确率,提出一种结合迁移学习(Transfer Learning,TL)和一维卷积神经网络(One Dimensional Convolutional Neural Network,1D-CNN)的结构损伤识别方法。首先,基于结构有限元数值模拟数据训练1D-CNN模型,选择损伤识别效果较好、性能优良的模型作为源模型;然后,将源模型中的网络结构和超参数迁移到实际结构实测数据集(目标域)网络模型的对应位置并冻结,得到预训练模型;最后,使用实测数据微调预训练模型得到目标模型。为验证该方法的有效性,通过3层钢框架结构实验室试验和日本某简支钢桁梁桥的现场试验,对比源模型(模型Ⅰ)、仅采用实测数据训练得到的CNN模型(模型Ⅱ)和采用迁移学习得到的CNN目标模型(模型Ⅲ)等3种神经网络模型的结构损伤模式识别准确率。研究结果表明:3层钢框架结构实验室试验中,3种CNN模型的最高损伤模式识别准确率分别为63.44%,98.44%,99.06%;日本某简支钢桁梁桥的现场试验中,3种CNN模型的最高损伤模式识别准确率分别为59.50%,97.00%,99.50%。针对不同结构,目标模型(模型Ⅲ)的损伤模式识别准确率均最高,收敛速度最快,优于其他2种CNN模型。基于迁移卷积神经网络的桥梁结构损伤识别方法具有较好的实际结构损伤识别能力,为解决数据有限情况下的结构损伤识别问题提供了一种有效的解决途径。Aiming at the discrepancy between the finite element model and the actual structures during the damage pattern identification of the actual bridges,to improve the accuracy of the deep neural network trained on the finite element numerical simulation dataset in identifying the damage patterns of the actual bridges,a method combining Transfer Learning(TL)with 1D Convolutional Neural Network(CNN)for structural damage identification of bridge structures was proposed in this paper.Firstly,the 1D-CNN model was trained based on the structural finite element numerical simulation data,and the model with better damage identification results was selected as the source model.Then,the network structure and weight parameters of the source model were transferred,and the pre-trained model was obtained from the actual structure data.Finally,the target model was obtained by fine-tuning the pre-training model.Through the laboratory test of a three-layers steel frame structure and the field test of a simple steel truss girder bridge in Japan,the structural damage pattern identification accuracy of three neural network models was compared,namely,the source model(model Ⅰ),the CNN model(model Ⅱ)trained only with measured data,and the target model(model Ⅲ)trained by transfer learning.The results confirm that the maximum damage pattern identification accuracy of the three CNN models reaches 63.44%,98.44% and 99.06% respectively in the laboratory test.In the field test,the maximum damage pattern identification accuracy of the three CNN models is 59.50%,97.00%,99.50%,respectively.For different structures,the target model(model Ⅲ)has the highest damage pattern identification accuracy and the fastest convergence rate,which is better than the other two CNN models.The transfer convolutional neural network method has good damage identification ability for the actual bridges.The results can provide an effective way to resolve the problem of structural damage identification in the case of limited data.

关 键 词：桥梁结构 损伤模式识别 一维卷积神经网络 迁移学习 源域 目标域 

分 类 号：U446[建筑科学—桥梁与隧道工程]