基于深度子领域适应卷积神经网络的结构损伤识别  

Structural damage recognition based on deep subdomain adaptation CNN

作　　者：张健飞[1] 曹雨 ZHANG Jianfei;CAO Yu(College of Mechanics and Engineering Science,Hohai University,Nanjing 211100,China)

机构地区：[1]河海大学力学与工程科学学院,南京211100

出　　处：《振动与冲击》2025年第3期251-260,共10页Journal of Vibration and Shock

基　　金：国家自然科学基金项目(12072105)。

摘　　要：针对卷积神经网络(convolutional neural networks,CNN)结构损伤识别模型泛化能力差的问题,提出了一种基于深度子领域适应卷积神经网络(deep subdomain adaptation convolutional neural networks,DSACNN)的结构损伤识别方法。以实际结构为目标域,以有限元模型为源域,根据损伤类别将源域和目标域划分成一系列子领域。在CNN中嵌入子领域适应模块,构建DSACNN模型,通过最小化源域上的损伤分类误差和领域之间的局部最大均值差异,对齐两个领域对应子领域的特征、建立特征与损伤类别之间的映射,从而将源域上的损伤识别能力迁移到目标域之上。模型的训练无需已知目标域样本的损伤标签,采用预训练全局领域适应提高其伪标签的准确率。试验结果表明:与全局领域适应模型相比,基于预训练全局领域适应的DSACNN模型在模拟目标域上准确率最大提高幅度达到21.8%,在实测目标域上提高了9.6%,具有更强的泛化能力。Here,aiming at the problem of poor generalization ability of convolutional neural network(CNN)structural damage recognition model,a structural damage recognition method based on deep subdomain adaptation convolutional neural network(DSACNN)was proposed.An actual structure was taken as target domain and a finite element model was taken as source domain,source and target domains were divided into a series of subdomain according to damage categories.Subdomain adaptation modules were embedded in CNN to construct DSACNN model,by minimizing damage classification error on source domain and local maximum mean discrepancy between two domains,features of corresponding subdomain on two domains were aligned and mappings between features and damage categories were established,thus the damage recognition ability on source domain was transferred to target domain.Training of the model didn’t need know damage labels of target domain samples,and pre-trained global domain adaptation was used to improve the accuracy of pseudo labels.The experimental results showed that compared with the global domain adaptation model,DSACNN model based on pre-trained global domain adaptation has a maximum accuracy improvement of 21.8%on the simulated target domain and 9.6%on the measured target domain,so DSACNN model has stronger generalization ability.

关 键 词：结构损伤识别 子领域适应 局部最大均值差异 卷积神经网络(CNN) 

分 类 号：TU317[建筑科学—结构工程]