基于深度学习方法的堆石混凝土缺陷识别与分析

【作者】 张宇帆；

【导师】 杨韬；

【作者基本信息】 贵州大学 ， 土木水利， 2023， 硕士

【摘要】 受施工环境、浇筑方式以及堆石体构造等因素的影响,在堆石混凝土浇筑过程中未能排出的气泡发生挤压、变形,进而形成形态、大小各不相同的内部缺陷,最终影响堆石混凝土材料的力学性能,带来无法预知的风险。因此,开展堆石混凝土缺陷的分类、高精度识别、结构表征和力学性能研究,有助于探索施工环境下堆石混凝土自身特性,为实现堆石混凝土高精度数值模拟以及工程实践提供参考。结合工程现场浇筑大体积堆石混凝土试件,以堆石混凝土内部细观缺陷为研究对象,运用深度学习领域最新进展与数值模拟相结合的手段,选用U-net网络模型构建堆石混凝土缺陷分割程序,获取缺陷信息,探究位置特征、形态特征和大小特征,确立缺陷分布模型、分布参数及表征方式,揭示缺陷模型破坏形态、损伤和破裂演化过程。1)使用特制100mm卡纸在工程现场堆石混凝土大体积切割试件中提取缺陷,根据缺陷形成原理、缺陷与胶结面接触程度将缺陷分为四类:自密实混凝土内部孔隙、自密实混凝土内部填充未密实、胶结面孔隙、胶结面填充未密实;采用深度学习网络构建缺陷分类器,网络结构包括主干特征提取网络、加强特征提取网络和预测网络;通过不同光照、角度和缺陷形态的图像增强方法扩充样本,利用反向传播和特定评价指标对模型进行参数修正,最终模型预测精度可达99%。2)通过自主研发堆石混凝土缺陷识别程序,获取各类缺陷轮廓、位置信息,表征堆石混凝土内部缺陷的物理特征。形态上,基本符合椭圆拟合,其圆形度随单个缺陷面积的增大而减小;位置上,胶结面缺陷分布较多且面积较大,其余区域则大致相同;分布上,缺陷大小、个数、孔隙率、拟合椭圆的长轴、短轴均符合对数正态分布。3)根据缺陷分类器构建包含堆石、自密实混凝土、缺陷三种材料的二维数值模型,对不同种类缺陷造成影响展开研究。其中填充未密实缺陷容易引发裂缝,成为破裂过程的起始区域;当孔隙缺陷连成一条直线或接近直线时,容易导致横向或纵向贯通,过早失去承载能力;胶结面缺陷所在位置容易形成薄弱区域,从而改变裂缝的延伸方向。更多还原

【Abstract】 Due to the influence of factors such as construction environment,pouring method,and stone pile structure,bubbles that are not discharged during the stone pile concrete pouring process are squeezed and deformed,forming internal defects of different sizes and shapes,which ultimately affect the mechanical properties of stone pile concrete materials and bring unpredictable risks.Conducting research on defects in stone pile concrete,classifying,accurately identifying,characterizing the structure,and studying the mechanical properties of defects can help explore the characteristics of stone pile concrete under construction environments and provide reference for achieving digital twins,more realistic numerical simulation research,and engineering practice.Based on the construction site poured large volume stone pile concrete specimens,this paper takes internal defects in stone pile concrete as the research object,and uses the latest advances in image recognition and numerical simulation techniques to construct a U-net deep learning model to segment stone pile concrete defects pixel by pixel.By using this program to obtain defect information,the position,shape,and size characteristics are explored,and the defect distribution model,distribution parameters,and characterization methods are established to reveal the failure mode,damage mechanism,and fracture evolution process of the defect model.1)a specially made 10 cm cardboard is used to extract defects from large volume specimens at the engineering scale.According to the principle of defect formation and the degree of contact between the defect and the bond surface,the defects are classified into four categories: internal pores of self-compacting concrete,internal filling of self-compacting concrete,bond surface pores,and bond surface filling.A U-net deep learning network is constructed to build a classifier,which includes a main feature extraction network,feature enhancement network,and prediction network.By using image enhancement based on different lighting,angles,and defect shapes to expand samples,the model is parameterized and modified through backpropagation and specific evaluation indicators.The prediction accuracy of the model can reach 97%.2)based on the independently developed stone pile concrete defect recognition program,various types of defect contours and location information containing physical features are obtained.The shape basically conforms to the elliptical fitting,and the roundness decreases as the defect area increases.On the position,there are more defects on the bond surface with larger areas,while the rest of the region is roughly the same.The sizes,numbers,porosity,and major and minor axes of the fitted ellipse of the defects all conform to the logarithmic normal distribution.3)the impacts caused by various types of defects are different.Among them,filling unconsolidated defects can easily lead to the generation of cracks and become the beginning of the fracture process.When pore defects are connected in a straight line or deviation is small,it is easy to form overall penetration,which leads to premature loss of bearing capacity.The distribution position of the bond surface defect is prone to form a weak surface,which guides the direction of crack propagation.更多还原