基于卷积神经网络和多标签分类的复杂结构损伤诊断

Damage diagnosis of complex structure based on convolution neural network and multi-label classification

在线阅读下载PDF

导出

摘要为研究复杂空间框架节点损伤识别问题,利用多标签分类的优势,构建了多标签单输出和多标签多输出两种卷积神经网络模型,用于框架结构节点损伤位置的判断和损伤程度诊断。针对复杂结构损伤位置判断时工况多、识别准确率不高等问题,提出了一种能对结构进行分层(或分区)处理并同时完成损伤诊断的多标签多输出卷积神经网络模型。分别构建了适用于多标签分类的浅层、深层和深层残差多输出卷积神经网络模型,并对其泛化性能进行了研究。结果表明:提出的模型具有较高的损伤诊断准确率和一定的抗噪能力,特别是经过分层(分区)处理后的多标签多输出网络模型更具高效性,有更快的收敛速度和更高的诊断准确率;利用多标签多输出残差卷积神经网络模型可以从训练工况中提取到足够多的损伤信息,在面对未经过学习的工况时也能较准确判断各节点的损伤等级。 In order to study the damage diagnosis of complex spatial frame joints as the research object,two convolutional neural network models of multi-label single-output and multi-label multi-output were constructed by using the advantages of multi-label classification,which were used to judge the damage location and damage degree of frame structure joints.Aiming at the problems of multiple location conditions and low recognition accuracy of damage joints in complex structures,a multi-label multi-output convolutional neural network model was proposed,which could process the structure hierarchically(or partitioned)and complete the damage diagnosis at the same time.The shallow,deep and deep residual multi-output convolution neural network models for multi-label classification were constructed respectively,and their generalization performance was studied.The results show that the proposed model has high damage diagnosis accuracy and certain anti-noise performance.In particular,the multi-label multi-output network model after hierarchical(or partition)processing is more efficient,with faster convergence speed and higher diagnostic accuracy.Using the multi-label multi-output residual convolution neural network model,enough damage information can be extracted from the dynamic response,and the damage level of each joint can be accurately determined in the face of unlearned conditions.

作者李书进杨繁繁张远进 LI Shujin;YANG Fanfan;ZHANG Yuanjin(School of Civil Engineering and Architecture,Wuhan University of Technology,Wuhan 430070,Hubei,China;School of Safety Science and Emergency Management,Wuhan University of Technology,Wuhan 430070,Hubei,China)

机构地区武汉理工大学土木工程与建筑学院武汉理工大学安全科学与应急管理学院

出处《建筑科学与工程学报》北大核心 2025年第1期101-111,共11页 Journal of Architecture and Civil Engineering

基金国家自然科学基金项目(52378313)。

关键词损伤诊断卷积神经网络多标签分类框架结构深度学习 damage diagnosis convolution neural network multi-label classification frame structure deep learning

分类号 TU312.3 [建筑科学—结构工程]

作者简介李书进(1967-),男,工学博士,教授,博士生导师,E-mail:sjli@whut.edu.cn。