摘要
利用振动响应和深度学习进行结构损伤识别时,会遇到需要较多测点数据、损伤识别准确率不高以及网络容易发生过拟合等问题。为此,提出了一种基于卷积神经网络-门控循环单元(convolutional neural networks-gated recurrent unit,CNN-GRU)神经网络并联网络的结构损伤识别新方法。首先,对响应信号进行广义S变换(generalized S-transform,GST)得到其时频图像。然后,分别利用CNN和GRU从时频图像和响应信号中提取时频域特征和时序特征,并将时频域特征和时序特征拼接后输入全连接层和Softmax分类器中进行结构损伤识别。位移激励下的海上风电支撑结构模型试验数据验证结果表明,该方法仅需要一个测点的响应信号,与其他同类方法相比具有更高的识别准确率和效率。
When using vibration response and deep learning for structural damage identification,the problems such as requiring more data of measuring points,low accuracy of damage identification,and over-fitting of the network will be encountered.Therefore,a novel structural damage identification method based on convolutional neural networks-gated recurrent unit(CNN-GRU)parallel neural network were presented in this paper.Firstly,the generalized S-transform(GST)was performed on the measured response signal to obtain the GST time-frequency diagram.Then,CNN and GRU were used to extract time-frequency features and temporal features from time-frequency diagrams and response signals,respectively.The time-frequency features and temporal features were spliced and input into the fully connected layer and the Softmax classifier for structural damage identification.The verification results of the model test data of offshore wind power supporting structure under displacement excitation show that the proposed method only needs the response signal of one measuring point and has higher identification accuracy and efficiency than other similar methods.
作者
李行健
刁延松
吕建达
侯敬儒
LI Xingjian;DIAO Yansong;L Jianda;HOU Jingru(School of Civil Engineering,Qingdao University of Technology,Qingdao 266033,China)
出处
《振动与冲击》
EI
CSCD
北大核心
2024年第20期229-237,共9页
Journal of Vibration and Shock
基金
山东省自然科学基金(ZR2021ME239,ZR2021ME033)。
