PDF全文阅读

吴立朋1 ,2 ,  闫新宇1 ,2 ,  金哲诚1 ,2

 

( 1 .  石家庄铁道大学 土木工程学院 ，石家庄 050043 ;

2.  道路与铁道工程安全保障省部共建教育部重点实验室（石家庄铁道大学），石家庄 050043)

摘   要 ：为了及时有效地发现桥梁病害 ，克服目前人工目测方法固有的费时 、费力且易受主观因素影响的缺陷 ，提出一种基于卷积神经网络的轻量化桥梁病害智能识别方法。在保持原 MobileNetV2 网络模型宽度不变的情况下 ，通过减少网络中瓶颈模块的数量 ，并在其中嵌入 Squeeze and Excitation 通道注意力机制 ，使新网络能够更有效地学习和利用特征通道之间的关系 ， 以提升网络识别性能。结果表明： 经过缩减瓶颈模块和嵌入通道注意力机制两次改进的网络模型 ，实现了 98 . 78% 的识别准确率 ，较改进前提高了 2. 85% ,  计算量也较改进前减少了 23 . 02 M ；将改进后的网络模型引入 Matlab 的 GUI 界面 ，将代码可视化为可以独立使用的桥梁病害识别系统，为风化 、钢筋锈蚀 、开裂和孔洞四种桥梁病害的高效识别提供便利。

关键词：桥梁 ；病害识别 ；MobileNetV2 ；注意力机制 ；GUI

中图分类号：U446           

文献标志码：A           

文章编号： 1005- 8249   (2025)  04- 0143- 07

DOI:10. 19860/j.cnki.issn1005 - 8249.2025 .04.026

 

WU Lipeng1 , 2  ,  YAN Xinyu1 , 2 ,  JIN Zhecheng1 , 2

( 1 . School of Civil Engineering , Shijiazhuang Tiedao University , Shijiazhuang 050043 , China;

2. Key Laboratory of Roads and Railway Engineering Safety Control  (Shijiazhuang Tiedao

University) ,  Ministry of Education , Shijiazhuang 050043 , China)

Abstract :  In order to timely and effectively detect bridge diseases , overcome the inherent defects  of the current manual visual inspection method in disease investigation , which is time - consuming , laborious , easy  to  cause  potential  safety  hazards , and easily affected by subjective factors , so as to provide preconditions for targeted bridge management , an intelligent method which is suitable for mobile disease image recognition is proposed. While keeping the width of the original MobileNetV2  network model unchanged , by  reducing  the number  of  bottleneck modules and embedding  the  Squeeze  and  Excitation  channel  attention mechanism in the bottleneck module , the correlation between channels is established , so that the network can learn and utilize the relationship between feature channels more effectively. The  results  show that the accuracy of image recognition is improved ,after two improvements in the  network  model  of reducing  bottleneck  module  and  embedding  channel  attention  mechanism , the recognition accuracy of 98. 78%  is  achieved , which is 2. 85%  higher than before the improvement and the  computational cost is also reduced  by  23 . 02 M ; The  improved  network model  is  introduced  into the  GUI  interface  of  Matlab , and  the code is visualized as a bridge disease identification system that can be used independently ,  which provides convenience for the efficient identification of four bridge diseases :  efflorescence , steel corrosion , cracking and holes .

Key words :  bridge;  disease identification;  MobileNetV2 ;  attention mechanism;  GUI

基金项目： 国家自然科学基金  (51808357) ；中央引导地方科技发展资金项目  (236Z3810G) 。

作者简介：吴立朋  (1981 —) ,  男 ，博士 ，副教授 ，研究方向： 结构工程。

通信作者： 闫新宇  (1998—) ,   女 ，硕士研究生 ，研究方向： 工程结构管理。

收稿日期：2024 - 01 - 19