摘要
为解决火灾后钢筋混凝土(RC)梁表观损伤识别主观性强、效率低以及损伤等级与残余性能关联机制不明确的问题,提出一种基于深度学习的火灾后RC梁损伤智能评估方法。构建并行MobileNetv3损伤分类网络,实现对颜色变化、裂缝形态、混凝土爆裂与钢筋露筋等多特征的高精度识别,融合Swin-Transformer模块,建立损伤等级识别模型;开发了移动终端智能评估系统,集成损伤分类、损伤等级识别、YOLOv5s-D与MB-SPPF-UNet等四种网络模型,实现火灾后RC梁损伤区域定位与损伤等级快速预测。结果表明:裂缝、颜色、爆裂和露筋等四种损伤类型的分类准确率分别达85%、98%、98%和93%;整体损伤等级分类准确率为91%;类激活热力函数图谱表明模型权重集中于实际损伤区域,验证了分类依据的合理性;网络预测损伤等级与基于残余承载力折减推定的损伤等级吻合度为84%。
To address the issues of strong subjectivity,low efficiency,and the unclear relationship between damage levels and residual performance in the visual assessment of fire-damaged reinforced concrete(RC)beams,this study proposed an intelligent multi-modal damage evaluation method based on deep learning.A parallel MobileNetv3-based network was constructed to achieve high-precision recognition of multiple features,including color change,crack pattern,concrete spalling,and exposed reinforcement.By integrating a Swin-Transformer module,a damage level identification model was established.A mobile intelligent evaluation system was developed,incorporating four network models-damage classification,damage level recognition,YOLOv5s-D,and MB-SPPF-UNet-enabling rapid localization of damaged areas and prediction of damage levels in post-fire RC beams.The results demonstrate that the classification accuracy for the four damage types-cracks,color change,spalling,and exposed reinforcement—reaches 85%,98%,98%,and 93%,respectively.The overall accuracy for damage level classification is 91%.Class activation mapping(CAM)indicates that the model's attention is focused on the actual damaged regions,validating the rationality of the classification basis.Moreover,the damage levels predicted by the network show an 84%agreement with those inferred based on residual load-bearing capacity reduction.
作者
刘才玮
田立斌
王鹏霏
苗吉军
LIU Caiwei;TIAN Libin;WANG Pengfei;MIAO Jijun(School of Civil Engineering,Qingdao University of Science and Technology,Qingdao 266520,China;School of Mechanics and Civil Engineering,China University of Mining and Technology,Xuzhou 221000,China)
出处
《建筑结构学报》
北大核心
2025年第11期105-114,共10页
Journal of Building Structures
基金
国家自然科学基金项目(52178487)
山东省自然科学基金项目(ZR2021ME228)。
关键词
RC梁
火灾损伤
深度学习
损伤分类
损伤评估
RC beam
fire damaged
deep learning
damage classification
damage detection
