融合密集卷积网络和注意力机制的拱桥损伤识别  

Damage identification of arch bridges based on dense convolutional networks and attention mechanisms

作　　者：辛景舟 刘倩茹 唐启智 李杰[3] 张洪[1] 周建庭[1,2] XIN Jingzhou;LIU Qianru;TANG Qizhi;LI Jie;ZHANG Hong;ZHOU Jianting(State Key Laboratory of Mountain Bridge and Tunnel Engineering,Chongqing Jiaotong University,Chongqing 400074,China;School of Civil Engineering,Chongqing Jiaotong University,Chongqing 400074,China;Chongqing Construction Science Research Institute Co.,Ltd.,Chongqing 400042,China)

机构地区：[1]重庆交通大学省部共建山区桥梁及隧道工程国家重点实验室,重庆400074 [2]重庆交通大学土木工程学院,重庆400074 [3]重庆市建筑科学研究院有限公司,重庆400042

出　　处：《振动与冲击》2024年第14期18-28,36,共12页Journal of Vibration and Shock

基　　金：国家自然科学基金(52278292);重庆市杰出青年科学基金(CSTB2023NSCQ-JQX0029);重庆市研究生科研创新项目(CYS240455)。

摘　　要：针对传统深度学习方法缺乏对网络特征的差异化利用且损伤识别精度易受环境因素影响的问题,提出了一种融合密集卷积网络(DenseNet121)和混合注意力机制(convolutional block attention module,CBAM)的拱桥损伤识别方法。首先,获取拱桥加速度响应数据,利用连续小波变换将其转换成时频图,形成拱桥损伤识别数据集;其次,将CBAM嵌入DenseNet121模型,加强断层特征的传播和特征的差异化利用,经训练得到拱桥损伤识别模型;然后,基于测试集评估损伤识别模型的性能,引入t分布随机邻域嵌入非线性降维技术对特征进行可视化分析;最后,通过数值案例验证了该方法的可行性和鲁棒性,并应用于劲性骨架拱肋的损伤识别。结果表明:所提方法可增强有用信息的权重,实现网络特征的差异化利用;与传统方法相比,该方法在单损伤和多损伤识别中准确率分别达到了91.67%和92.78%,准确率更高,且具有较强的鲁棒性和实用价值。Traditional deep learning methods lack the differential utilization of network features and the damage identification accuracy is easily affected by the environmental factors.To this end,an arch bridge damage identification method was proposed based on the dense convolutional network DenseNet121 and the convolutional block attention module(CBAM).Firstly,the arch bridge acceleration data were obtained and converted into time-frequency maps using the continuous wavelet transform to construct the damage identification dataset.Secondly,the CBAM was embedded into the DenseNet121 model to enhance the propagation of the fault features and the differential utilization of the features,and the arch bridge damage identification model was obtained after the training.Then,the performance of the damage identification model was evaluated based on the test set and the t-distributed stochastic neighbor embedding nonlinear dimensionality reduction technique was introduced to visualize the features.Finally,the feasibility and robustness of the method were verified by numerical cases,and the proposed method was further applied to the damage identification of the stiff skeleton arch ribs.The results indicate that the proposed method can increase the weight of useful information and realize the differential utilization of network features.Compared with the traditional method,the method achieves the accuracy of 91.67%and 92.78%respectively in single damage identification and multi-damage identification.It is of higher identification accuracy,strong robustness,and practical value.

关 键 词：桥梁健康监测 拱桥 损伤识别 DenseNet121 注意力机制 特征可视化 

分 类 号：U446[建筑科学—桥梁与隧道工程]