基于多标签卷积神经网络的结构损伤识别  

Structural damage identification based on multi-label convolutionneural network

作　　者：秦世强[1] 苏晟 杨睿 QIN Shiqiang;SU Sheng;YANG Rui(School of Civil Engineering and Architecture,Wuhan University of Technology,Wuhan 430070,Hubei,China)

机构地区：[1]武汉理工大学土木工程与建筑学院,湖北武汉430070

出　　处：《建筑科学与工程学报》2024年第3期108-119,共12页Journal of Architecture and Civil Engineering

基　　金：国家自然科学基金项目(51608408)。

摘　　要：准确识别结构多位置损伤一直是结构损伤识别的难题。为提升结构多位置损伤识别的准确率,提出一种基于卷积神经网络(CNN)的多标签分类(MLC)方法(CNN-MLC)进行结构损伤识别。该方法将结构多个位置损伤识别转换为多标签分类问题,每个损伤位置均用一个对应的标签表示;利用CNN强大的特征提取能力,深入挖掘不同损伤工况之间公共损伤位置的相关性,实现结构多位置损伤识别。通过四层框架结构和一座铁路连续梁桥多位置损伤识别验证了CNN-MLC方法的识别准确率,并将其识别结果与基于CNN的多类别分类(MCC)方法(CNN-MCC)和基于示例差异化算法(InsDif)的多标签分类方法(InsDif-MLC)进行了对比。结果表明:框架结构在两位置和三位置损伤工况下,CNN-MLC方法比CNN-MCC方法的识别准确率分别提升2.50%和9.64%,比InsDif-MLC方法识别准确率提升17.50%和29.28%;对于铁路连续梁桥的两位置损伤和三位置损伤,CNN-MLC方法比CNN-MCC方法识别准确率提升1.63%和6.85%,比InsDif-MLC方法识别准确率提升4.18%和18.49%;随着损伤位置数量的增加,CNN-MLC方法的识别准确率显著提升。Accurate identification of structural multi-site damage has always been a difficult problem in structural damage identification.In order to improve the accuracy of structural multi-site damage identification,a multi-label classification method based on convolution neural network(CNN-MLC)was proposed for structural damage identification.In this method,the multi-site damage identification of the structure was transformed into a multi-label classification problem,and each site damage is represented by a separate label.Using the strong feature extraction ability of CNN,the correlation of common damage site between different damage conditions was deeply mined,and the multi-site damage identification was realized.The CNN-MLC method was verified by multi-site damage identification of a four-story frame structure and a railway continuous beam bridge,and the identification results were compared with those of CNN-MCC and InsDif-MLC.The results show that under two-sites and three-sites damage conditions,the recognition accuracy of CNN-MLC is 2.50%and 9.64%higher than that of CNN-MCC,and 17.50%and 29.28%higher than that of InsDif-MLC.For the two-sites damage and three-sites damage of railway continuous beam bridges,the recognition accuracy of CNN-MLC is 1.63%and 6.85%higher than that of CNNMCC,and 4.18%and 18.49%higher than that of InsDif-MLC.With the increase of the number of damage sites,the recognition accuracy of CNN-MLC is significantly improved.

关 键 词：结构损伤识别 卷积神经网络 多位置损伤 多类别分类 多标签分类 

分 类 号：TU317[建筑科学—结构工程]