强噪声小样本条件下基于图卷积神经网络的结构损伤识别
Structural damage identification based on graph convolutional neural network under strong noise and small sample conditions
李行 1骆勇鹏 2郭旭 1廖飞宇 3鲁四平4
作者信息
- 1. 福建农林大学 交通与土木工程学院,福建 福州 350108
- 2. 福建农林大学 交通与土木工程学院,福建 福州 350108;福建省结构工程与防灾重点实验室(华侨大学),福建 厦门 361021;数字福建智能交通技术物联网实验室,福建 福州 350108
- 3. 福建农林大学 交通与土木工程学院,福建 福州 350108;数字福建智能交通技术物联网实验室,福建 福州 350108
- 4. 中南大学 土木工程学院,湖南 长沙 410075
- 折叠
摘要
基于深度学习的结构损伤识别主要通过捕捉数据特征及内部规律来实现.训练样本不足、噪声干扰均可能导致有效特征及内部规律的挖掘失败.从数据中尽可能挖掘更多的信息用于识别损伤尤为重要.文中提出了基于图卷积神经网络的结构损伤识别方法.首先,为了能够提取更多特征,即同时考虑不同位置传感器之间的相关性和各个传感器数据的自身特性,通过图构造方法将一维振动数据转换为图数据.然后采用图卷积神经网络提取图样本的数据特征并实现快速分类,完成损伤识别的目的.采用卡塔尔大学看台结构模型来验证所提方法的可行性及可靠性,并探讨噪声程度、样本个数、构图方式及相关图卷积网络参数对识别结果的影响.结果表明:与一维卷积神经网络相比,图卷积神经网络模型在强噪声、小样本的情况下具有较高的损伤识别精度.构图方式及图池化方法对识别结果有一定的影响,Path构图方式与Topk池化的识别结果较为稳定且高于其他组合形式.
Abstract
Structural damage identification based on deep learning are mainly realized by capturing the characteristics and internal rules of data.Insufficient training samples and noise interference may lead to the failure of mining effective features and internal laws.It is particularly important to mine information as much as possible from the data for damage identification.To solve these problems,a structural damage identification method based on graph convolutional network(GCN)is proposed.In order to extract more features,considering the correlation between different position sensors and the characteristics of each sensor data,one-dimensional vibration data was converted into graph data by the graph construction method.Subsequently,GCN was used to extract the data features of the graph samples and achieve rapid classification to achieve the purpose of damage identification.The feasibility and reliability of the proposed method were verified by the Qatar University grandstand simulator structure,and the effects of noise level,number of samples,the method of graph construction and convolutional network parameters on the recognition results were discussed.The results show that,compared with 1 dimensional convolutional neural network,the GCN model has higher damage identification accuracy in the case of strong noise and small samples.The method of graph construction and pooling have certain influence on the identification results.The identification results of Path graph and Topk pooling are stable and higher than those of other combination forms.
关键词
结构健康监测/损伤识别/振动响应/深度学习/图卷积神经网络Key words
structural health monitoring/damage identification/vibration response/deep learning/graph convolutional neural network引用本文复制引用
基金项目
国家自然科学基金(51808122)
福建省自然科学基金面上项目(2020J01580)
福建省结构工程与防灾重点实验室开放基金(华侨大学)(SEDPFJ-2018-01)
出版年
2024
NETL
NSTL
万方数据