【作者】 张涛；

【导师】 张强强； 徐阳；

【作者基本信息】 兰州大学 ， 土木工程·结构工程， 2022， 硕士

【摘要】 随着我国城镇化水平不断提高,城市规模和体量不断加大,其应对自然灾害风险挑战和应急能力愈发重要。城市建筑群结构地震破坏定位与分类作为重要的震害评估环节,对政府应急救援决策和震后建筑加固维修具有重要的参考意义。建筑结构地震破坏由于其特殊的形态学特征、复杂的区域性与强背景噪声干扰等因素,使得传统人工检测存在时效性欠缺和主观依赖性强的局限,而传统图像处理如阈值分割、边缘检测等方法存在精度低、速度慢、计算量大等困难。基于计算机视觉的智能感知与定位识别方法,因高效、快速、精度高和复杂场景适用强等优势,可望为结构破坏地震破坏目标检测和分类评估提供参考。本论文基于YOLOv4深度学习模型提出了建筑物结构构件破坏目标检测方法,实现了多类型破坏目标高准确率定位识别,进而发展了多类型结构构件分类的U-Net语义分割模型,验证了构件破坏类型与所对应构件融合评估震害等级的有效性。主要研究内容如下:(1)建立了国内外典型地震结构构件破坏数据集。采用2016年厄瓜多尔与2008年北川震后建筑室内破坏图片作为数据来源,建立了开裂、剥落、掉块、钢筋暴露与钢筋屈曲等五类破坏标签。根据裂缝特殊的几何形态学特征,先行开展针对裂缝破坏的单分类目标检测训练;进一步开展了剥落、掉块、钢筋暴露与钢筋屈曲四类区域性破坏的多分类目标检测研究。(2)发展了复杂交叉裂缝破坏定位与分类方法。在原YOLOv4神经网络的基础上,通过数据增强丰富数据集,并进行了超参数优化。进一步引入Focal Loss损失函数,改进损失函数中的分类损失,引入超参数α与γ,解决了难识别样本识别率低、误检、漏检等问题。在改进原网络结构的基础上,优化裂缝识别效果评分机制,实现了良好的、符合工程实际的裂缝破坏定位识别效果。(3)针对多分类区域性破坏定位与分类问题,通过加权数据增强手段解决了数据不平衡、个别破坏类别占比较小等问题。引入通道注意力机制与空间注意力机制模块,提高破坏区域识别效果。优化网络上采样模块,在上采样模块中采用Mish激活函数进一步提高多分类破坏目标定位准确率。(4)采用U-Net深度学习模型对对室内建筑构件进行分割识别。数据集采用2008年北川与2021年泸县震后建筑室内图片,建筑构件分为墙、柱、梁、板四类。通过改进优化算法、改进学习率下降方法与超参数训练提升构件语义分割精度。通过破坏区域、破坏类别与构件位置信息的融合,得到建筑构件破坏程度信息,最终实现构件地震破坏目标高精度定位和分类评估。更多还原

【Abstract】 With the significant improvement of urbanization and the continuous increase of urban scale and volume,its ability to deal with natural disaster risk challenges and emergency response is becoming increasingly important.As a crucial link of earthquake disaster assessment,earthquake damage identification of urban building complex structure has great reference value for government emergency rescue decision-making and post-earthquake building reinforcement and maintenance.Due to its special morphological characteristics,complex regionality and strong background noise interference,the traditional manual detection has some shortcomings,such as lack of timeliness and strong subjective dependence of professionals.However,traditional image processing methods such as threshold segmentation and edge detection have some challenges,such as low precision,slow speed and large amount of calculation.In contrast,computer vision perception and recognition technology provides a new way for structural damage detection,location and evaluation because of its fast and high precision.In this paper,the damage detection and location method of building structural components based on YOLOv4 deep learning model was proposed,which realized the high-accuracy identification of multiple damage targets of structural components,and then developed the u-net semantic segmentation model for multi type component classification,which verified the effectiveness of the fusion evaluation of component damage and corresponding components.The contents of this study are as follows:(1)The damage data set of typical seismic structural members at home and abroad was established.Using the internal damage images of buildings after the Ecuador earthquake in 2016 and the Beichuan earthquake in 2008 as the data source,five types of damage labels such as cracking,spalling,failure,reinforcement exposure and reinforcement buckling were established.Due to the special geometric and morphological characteristics of cracks,the single classification recognition training for crack damage was carried out first.Then,multi classification target detection training was carried out for four types of regional damage: spalling,failure,reinforcement exposure and reinforcement buckling.(2)The damage identification and location method of complex cross cracks was developed,and the data set was enriched and improved by means of data enhancement.Based on the original YOLOv4 neural network,the hyperparameters optimization was carried out.The Focal Loss function was introduced to modify the classification loss in the original three types of losses: classification loss,positioning loss and confidence loss.The hyperparameters α and γ was introduced into classification loss for solving the problems of low recognition rate,false detection,missing detection and so on.In addition,it can improve the proportion weight of difficult to recognize samples in the loss function,and optimize the weight proportion of background and recognition target area in the loss function.Based on the improvement of the original network structure,the scoring mechanism of the evaluation indicators of crack identification was optimized to achieve a good and practical crack damage location and identification effect.(3)Aiming at the location and identification of multi-categories regional damage,the weighted data enhancement method was used to solve the problems of data imbalance and small proportion of individual damage categories.By introducing the modules of channel attention mechanism and spatial attention mechanism,the effect of damage area recognition was improved.In addition,the upsampling module was optimized,the mish activation function was used in the upsampling module to improve the multi classification damage identification and evaluation indicators.(4)U-Net semantic segmentation Neural Network was used to classify and identify interior building components.The data set adopts the building interior damage images of the Beichuan earthquake in 2008 and the Luxian earthquake in 2021.The building components were divided into four categories: wall,column,beam and slab.Through the improvement of optimization algorithm strategy,the experiment of learning rate reduction method and hyperparameters optimization,the effect of component semantic segmentation and recognition was improved.The weighted combination of damage area,damage category and component location information can obtain the damage degree information of building components,and finally realize a high-precision and high-speed damage degree evaluation method of building components.更多还原