Computer Vision-based Seismic Damage Detection And Performance Evaluation Of Reinforced Concrete (RC) Components

Post-earthquake seismic damage detection and residual seismic capacity assessment of building structures are the basis for decision-making of continuous-use,repair or demolition of damaged buildings.The current approach relies on the field surveys conducted by structural engineering professionals,where 1)visible seismic damage is inspected for structural components;2)damage levels of damaged components are determined based on the visual appearances of seismic damage;3)the residual seismic capacity of the building is evaluated by integrating the damage status of individual components.The assessment is therefore subjective,time-consuming and labor-intensive.Through the interdisciplinary innovation of earthquake engineering and artificial intelligence,in this study,the detection of visible seismic damage and evaluation of mechanical property degradation were comprehensively studied for reinforced concrete(RC)structural components using computer vision techniques.The major achievements are as follows.(1)A database was constructed for the semantic segmentation of typical seismic damage of RC components,including cracking,concrete spalling,reinforcement exposure,concrete crushing,as well as reinforcement buckling and fracture.The samplelevel and pixel-level data balancing and background recognition improvement techniques were proposed and developed,and the pixel-level multicategory detection of visible seismic damage of RC components was therefore achieved based on the deep convolutional networks.For the detected crack damage,an effective post-processing technique was developed,which refined the boundaries of the detected cracks and thus improved the accuracy of the subsequent crack width characterization.(2)Based on the binary image of the detected crack field,through the integration and development of image processing techniques including morphological operations,skeletonization and connected-component labeling algorithms,a novel approach was proposed for crack field analysis and characterization,which is efficient and robust for complex situations such as multi-crack intersection and is capable of separating and labeling individual cracks,and calculating the crack geometric properties including crack length,width and angle.(3)By linking the test photos of RC columns with the loading points on the hysteretic curves,a database of the correlated seismic damage photos and mechanical properties was established for RC components controlled by flexural behavior.Two light-weight modules,i.e.,the Patch-Pooling layer and Dilated-and-Separable Convolution block,were designed,and the architecture of multi-task learning was introduced to assemble a novel deep convolutional network for the evaluation of mechanical property degradation of RC components based on the detected visible damage,which enabled effective integration of visual features and spatial topologies of seismic damage,and thus estimated the stiffness and strength reduction factors with favorable accuracy.(4)Focusing on the shake-table test of a large-scale three-story RC structure,the proposed approach was applied to the analysis of the seismic damage photos of RC components including beams and columns.,where 1)the development of multicategory visible seismic damage after various seismic excitations was detected and characterized;2)based on the detected visible damage,the mechanical property degradation was evaluated for the damaged RC components.Through the comparison with the analysis based on the seismic response data,the effectiveness of the proposed approach were verified,and its application potential in the field of seismic damage detection and residual seismic capacity assessment of building structures was demonstrated.