摘要
为了有效提取铁路隧道衬砌结构损伤特征信息,提高损伤识别准确率,提出了一种基于平行卷积神经网络的隧道衬砌结构损伤多域监督学习识别方法。针对隧道衬砌结构振动信号数据结构特征,建立了一种可以融合一维时序信号和二维图像特征提取多域特征信息的平行卷积神经网络(P-CNN)模型。将该方法应用于隧道结构缩尺模型试验,选取1D-CNN、2D-CNN模型进行对比分析。结果表明:P-CNN模型具有最低的训练损失且准确率最高,收敛时迭代次数更少,并且收敛性能始终优于1D-CNN和2D-CNN模型;无论是无损伤还是不同损伤状态的情况,P-CNN的损伤识别准确率为86.25%~91.07%,均高于其他二者,验证了P-CNN模型在异源信号方面具有较好的泛化能力。
To effectively extract feature information of damage in railway tunnel lining structures and improve the accuracy of damage identification,a multi-domain supervised learning identification method for tunnel lining structure damage based on parallel convolutional neural networks(P-CNN)was proposed.Considering the structural characteristics of vibration signal data of tunnel lining structures,a parallel convolutional neural network model capable of integrating one-dimensional time-series signals and two-dimensional image features to extract multi-domain features was proposed.The method was applied to a scaled model test of tunnel structures,and 1D-CNNand2D-CNNmodels were selected for comparative analysis.The results show that,the P-CNN model has the lowest training loss and the highest accuracy,with fewer iterations to convergence,and consistently outperforms the convergence performanceof 1D-CNNand 2D-CNN models.Regardless of whether there is no damage or different damage states,the damage identification accuracy of P-CNN is 86.25%to 91.07%,which is higher than the other two,verifying that the P-CNN model has good generalization ability in heterogeneous signals.
作者
罗忆
万子豪
龚航里
张贤齐
张金瑞
罗涵
LUO Yi;WAN Zi-hao;GONG Hang-li;ZHANG Xian-qi;ZHANG Jin-rui;LUO Han(School of Civil Engineering and Architecture,Wuhan University of Technology,Wuhan 430070,China;Sanya Science and Education Innovation Park,Wuhan University of Technology,Sanya 572025,China)
出处
《武汉理工大学学报》
2025年第3期31-37,共7页
Journal of Wuhan University of Technology
基金
国家自然科学基金面上项目(51979208)
中国博士后科学基金面上资助项目(2024M762516)
关键词
损伤识别
振动信号
平行卷积神经网络
隧道衬砌结构
缩尺模型
damageidentification
vibration signal
parallel convolutional neural network
tunnel lining structure
scaled model
