不平衡样本下基于CVAE和CNN的结构损伤识别方法  

Method of Structural Damage Identification Based on CVAE and CNN Under Imbalanced Samples

作　　者：蔡东成 张健飞[2] CAI Dongcheng;ZHANG Jianfei(Faculty of Construction Engineering,Sichuan Sanhe College of Professionals,Luzhou 646200,China;College of Mechanics and Materials,Hohai University,Nanjing 211100,China)

机构地区：[1]四川三河职业学院建筑工程学院,四川泸州646200 [2]河海大学力学与材料学院,江苏南京211100

出　　处：《土木工程与管理学报》2023年第2期108-116,129,共10页Journal of Civil Engineering and Management

基　　金：国家自然科学基金(12072105)。

摘　　要：为提高在不平衡样本下结构损伤识别的准确性,提出了一种基于条件变分自编码器(CVAE)数据增强和卷积神经网络(CNN)的结构损伤识别方法。首先,将损伤类别作为约束,构建起基于振动加速度数据的CVAE模型;然后生成损伤加速度数据对初始不平衡数据进行扩充;最后使用CNN对扩充数据集进行特征提取和损伤分类识别。通过对悬臂梁振动台实验与钢框架有限元模拟振动实验两类数据集设置不同不平衡比率,进行了CVAE数据增强的效果对比验证。结果表明:CVAE数据增强有助于CNN损伤识别模型对数据特征的提取,能够提高CNN模型的收敛速度,防止模型过拟合;相对于未经数据增强的数据集,所提方法提高了在极不平衡数据下的损伤分类识别准确率,在两类实验数据集上分别提高了15.10%和15.80%。In order to improve the accuracy of structural damage identification under imbalanced samples,a structural damage identification method is proposed based on conditional variational auto-encoder(CVAE)data augmentation and convolutional neural network(CNN).Firstly,taking the damage category label as constraint,based on vibration acceleration data a CVAE model is constructed;Then the acceleration data with damage label is generated to augment the initial imbalanced data set;Finally,CNN is used for feature extraction and damage classification and recognition of the augmented data set.The effect of CVAE data augmentation is compared and verified by setting different imbalanced ratios on the two data sets of cantilever vibration experiment and finite element vibration simulation of steel frame.The results show that CVAE data augmented model is helpful for CNN damage identification model to extract data features,improves the convergence speed of CNN model,and prevent model overfitting.Compared with the data sets without data augmentation,the proposed method improves the accuracy of damage classification and identification under extremely imbalanced data,up to 15.10%and 15.80%respectively on the two types of experimental data sets.

关 键 词：损伤识别 不平衡样本 数据增强 条件变分自编码器 卷积神经网络 

分 类 号：TU312[建筑科学—结构工程]