摘要
为了利用结构振动响应的时间多尺度特征来提升卷积神经网络识别结构损伤的能力,给出了两种用于结构损伤识别的多尺度卷积神经网络,即多尺度输入和多尺度卷积核卷积神经网络。对于多尺度输入卷积神经网络,将通过下采样和滑动平均获取的具有不同时间尺度特征的振动信号输入固定尺寸卷积核的分支卷积神经网络;对于多尺度卷积核卷积神经网络,则将相同的振动信号输入具有不同尺寸卷积核的分支卷积神经网络。然后将各个分支卷积神经网络的输出组合成多尺度特征输入全连接层进行损伤模式的识别。数值试验和振动台试验的结果表明:相比于单一尺度卷积神经网络,多尺度卷积神经网络具有更高的损伤识别精度和抗噪性;对于损伤特征相近的损伤模式具有更好的辨别能力。
To improve the performance of the convolutional neural networks(CNN)for damage identification,two types of multi-scale CNNs,multi-scale input CNN and CNN with multiple kernel sizes,are presented for structural damage identification by utilizing the multi-scale temporal features in structural vibrating response.For multi-scale input CNN,vibrating signals with multi-scale features obtained by down sampling and moving averaging the original signal are fed into branch CNNs with fixed kernel sizes.For CNN with multiple kernel sizes,the same signals are inputted into branch CNNs with kernels of different sizes.The outputs of different branch CNNs are concatenated into a vector with multi-sacle features and inputted into a full connected networks to recognize damage patterns.Numerical and shaking table tests show that compared with single-scale CNN,multi-scale CNNs have higher damage detecting accuracy,stronger anti-noise capacity and stronger discrimination capacity between similar damage patterns.
作者
张健飞
蔡东成
ZHANG Jianfei;CAI Dongcheng(College of Mechanics and Materials,Hohai University,Nanjing 210098,China)
出处
《地震工程与工程振动》
CSCD
北大核心
2022年第1期132-142,共11页
Earthquake Engineering and Engineering Dynamics
基金
国家重点研发计划(2018YFC0406703)。
关键词
深度学习
卷积神经网络
多尺度
结构损伤识别
deep learning
convolutional neural networks
multiple scale
structural damage identification
