摘要
提出一种经验模态分解(empirical mode decomposition, EMD)和密集连接卷积神经网络(densely connected convolutional network, DenseNet)相结合的结构损伤识别网络模型(E-DenseNet)。对采集的加速度信号进行EMD得到多个本征模态函数(intrinsic mode function, IMF)分量,接着剔除皮尔逊相关系数绝对值较小的弱相关IMF分量。根据输入数据的组织方式,设定3种E-DenseNet模型:E-DenseNet1利用强相关IMF分量重构信号建立一维单通道输入数据;E-DenseNet2将各强相关IMF分量分别视作一个通道来建立一维多通道输入数据;E-DenseNet3利用所有强相关IMF分量组成二维矩阵来建立二维单通道输入数据。某简支梁算例分析表明:E-DenseNet1计算速度快但识别精度低,E-DenseNet2计算速度快且识别精度高,E-DenseNet3识别精度高但计算速度慢;与一维多通道残差卷积神经网络(residual network, ResNet)及标准卷积神经网络(convolutional neural network, CNN)相比,E-DenseNet2的识别精度明显更优;E-DenseNet2因而具有兼顾计算效率和识别精度的优点。E-DenseNet2可视化分析表明了其识别过程,对于相同工况下的不同样本,输出层越深其输出特征越相似,直至全连接层给出极大相似输出特征。
A structure damage identification network model(E-DenseNet)that combines empirical mode decomposition(EMD)and densely connected convolutional network(DenseNet)is proposed.The collected acceleration signals undergo EMD to obtain multiple intrinsic mode function(IMF)components,and then the weakly correlated IMF components with small absolute values of Pearson correlation coefficients are removed.According to the organization of the input data,three types of E-DenseNet models are set.E-DenseNet1 reconstructs the signal using strongly correlated IMF components to establish one-dimensional single-channel input data.EDenseNet2 treats each strongly correlated IMF component as a channel to establish one-dimensional multi-channel input data.E-DenseNet3 uses all strongly correlated IMF components to form a two-dimensional matrix to establish two-dimensional single-channel input data.The numerical analysis of a simply supported beam shows that:EDenseNet1 runs quickly with poor damage detection accuracy.E-DenseNet2 is computationally efficient with high damage detection accuracy.E-DenseNet3 provides good damage detection results but is time-consuming.Compared with one-dimensional multi-channel residual convolutional neural network(ResNet)and standard convolutional neural network(CNN),E-DenseNet2 performs much better in damage detection accuracy.It is thus concluded that E-DenseNet2 ensures both the computational efficiency and the damage detection accuracy.The visualization analysis of E-DenseNet2 exhibits its damage detection process that for different samples of the same damage scenario,a deeper layer outputs more similar features until the fully connected layer provides the most similar output features.
作者
吁强
蔡晓丽
李翠
朱学坤
伍晓顺
朱驰
YU Qiang;CAI Xiaoli;LI Cui;ZHU Xuekun;WU Xiaoshun;ZHU Chi(School of Civil and Surveying and Mapping Engineering(Nanchang),Jiangxi University of Science and Technology,Nanchang 330013,China;College of Information Engineering,Jiangxi V&T College of Communication,Nanchang 330013,China)
出处
《地震工程与工程振动》
CSCD
北大核心
2024年第3期61-72,共12页
Earthquake Engineering and Engineering Dynamics
基金
国家自然科学基金项目(51868026)
江西省自然科学基金项目(20202BAB204028)
江西省研究生创新专项资金项目(YC2022-S695)。
关键词
损伤识别
神经网络
动力测试
灵敏度分析
经验模态分解
damage detection
neural network
dynamic test
sensitivity analysis
empirical mode decomposition