摘要
目前在利用CNN网络提取特征的结构损伤识别研究中,仅仅利用1D-CNN和2D-CNN提取的特征进行损伤识别存在准确率低、识别效率不高等问题。提出了一种基于广义S变换和并联神经网络的结构损伤识别方法。为了丰富输入信号的特征维度,利用广义S变换将滤波后的信号转化成时频图,并同时将一维加速度响应信号和二维时频图分别输入1D-CNN和2D-CNN中进行时域和时频域特征提取,并在汇聚层进行特征拼接,然后通过FC层和Softmax层对损伤识别结果进行分类。利用IASC-ASCE SHM Benchmark结构第二阶段试验数据对所提出的并联网络模型进行验证,结果表明,所提出的网络模型与其他同类方法相比具有更高的识别精度和识别效率。
Among the existing researches on structural damage identification that use CNN network to extract features,problems such as low accuracy and low recognition efficiency can be found when only 1D-CNN and 2D-CNN are used to extract features for damage identification.Therefore,this paper proposes a structural damage identification method based on generalized S-transform and parallel neural network.In order to enrich the feature dimensions of the input signal,the filtered signal is converted into a time-frequency diagram by using the generalized S-transform.At the same time,the one-dimensional acceleration response signal and the two-dimensional time-frequency diagram are input into 1D-CNN and 2D-CNN respectively for time-frequency and time-frequency feature extraction,and the characteristics are spliced in the convergence layer.Then,the damage identification results are classified through FC layer and Softmax layer.The proposed parallel network model is verified by the second-stage test data of IASC-ASCE SHM Benchmark structure.The results show that the proposed network model has higher identification accuracy and efficiency than other similar methods.
作者
李行健
吕建达
赵凌云
刁延松
LI Xingjian;L Jianda;ZHAO Lingyun;DIAO Yansong(School of Civil Engineering,Qingdao University of Technology,Qingdao 266525,China)
出处
《青岛理工大学学报》
CAS
2024年第1期26-35,共10页
Journal of Qingdao University of Technology
基金
山东省自然科学基金资助项目(ZR2021ME239)。
关键词
损伤识别
广义S变换
卷积神经网络
时频分析
特征融合
damage identification
generalized S-transform
convolution neural network
time-frequency analysis
feature fusion
