Study On Multi-scale Seismic Damage Identification And Assessment Of Buildings Enhanced By Geometric Constrained Deep Learning

Our country has a vast territory and an extremely complex geological structure,resulting in frequent natural disasters such as earthquakes that pose a serious threat to the service life of urban buildings and people’s life and property safety.Post-earthquake damage assessment is crucial for post-earthquake emergency deployment,recovery and reconstruction.The structural damage monitoring and evaluation method based on structural dynamic response faces challenges such as high cost and difficulty in maintaining a dense sensor array and locating the spatial location of structural local damage accurately.Traditional earthquake damage investigation methods highly rely on the professional experience and subjective judgment of inspectors,making it difficult to meet the needs of rapid and accurate post-earthquake assessment due to limitations such as timeliness,difference in assessment results and limited accessible areas.Due to the diversity and complexity of damage modes,geometric forms and spatial scales of earthquake disasters in urban buildings,conventional identification and assessment methods are not universally applicable in real post-earthquake scenarios.Compared to time-history signals such as dynamic acceleration,image/video data provides a new perceptual information source for accurately assessing the damage degree of urban buildings after an earthquake.In order to realize comprehensive and rapid identification and assessment of earthquake damage of urban buildings,a geometric constrained deep learning framework based on computer vision is proposed in this dissertation.In addition,a multi-scale identification and assessment study of earthquake damage state associated with "building group – building unit – structural component" is systematically carried out.The main research contents of this dissertation are as follows:(1)A theoretical model for multi-scale seismic damage identification of buildings is proposed,using a geometric constraint deep learning approach.The model integrates geometric consistency loss(GC loss)and cross entropy loss(CE loss)to establish a geometric consistency loss function(GCE loss)suitable for semantic segmentation in multi-scale seismic damage identification and evaluation of buildings.The GC loss promotes contour closure,edge smoothing and internal connectivity of the multi-scale seismic damage identification targets of complex buildings,by measuring the consistency of the target boundary and regional geometric features,including the length and curvature of the partition line,and the consistency loss of the partition area.A multiscale seismic damage assessment framework for buildings is established,utilizing multi-scale and multi-source feature image data from satellite remote sensing,UAV low altitude remote sensing,and close-up imaging to achieve pixel-level segmentation and excellent identification and evaluation of large-scale wide-area architectural buildings,mesoscale regional building groups,and small-scale individual building components.Finally,the evaluation results from various scales are integrated to classify the seismic damage state of buildings.(2)A method for the fine recognition and rapid collapse state assessment of densely distributed small target buildings after an earthquake is proposed based on satellite remote sensing images.This is achieved through the development of a semantic segmentation network that enables the recognition and evaluation of buildings after an earthquake.The study focuses on the weight coefficient of GCE loss and how it affects the model segmentation performance.Additionally,the geometric feature optimization performance and multi-level feature extraction capability of GCE loss in the training process are analyzed.The results show that the geometric feature constraint characteristic of GCE loss enables the model to achieve stable extraction of image features and accurate and fast convergence optimization in the training process.As a result,the semantic segmentation accuracy of the model for the dense distribution of small targets is significantly improved,leading to high-precision recognition of postearthquake buildings with complex and fuzzy boundaries.The proposed method achieves a verification m Io U of 86.98% for building recognition and collapse state assessment in actual seismic remote sensing images.(3)To address the challenges posed by UAV low-altitude remote sensing images,such as large-scale changes,significant differences in morphological distribution,serious target occlusion,and complex weather environment interference,this study proposes a method for building recognition and damage state assessment in complex scenes after an earthquake.The proposed method utilizes Quake City Net-M-N(QCNet),an encoder and decoder network architecture with adaptive network parameters,to enhance multi-type weather interference images in complex scenes.Moreover,the model performance is further optimized by integrating GCE loss.The results demonstrate that the optimized QCNet62 network model can achieve a verification m Io U of 87.65% for damaged,severely damaged,and other buildings.Additionally,the study integrates the UAV attitude information and internal and external parameters of the airborne camera and employs a pixel – image – camera-earth multi-coordinate system transformation and geographic/inverse geographic location coding and interpretation algorithm to accurately locate the building in the post-earthquake area.(4)In light of the significant intra-class difference and limited inter-class difference of multi-type and multi-damage state building components,a multi-task fusion method for post-earthquake building component identification and damage state assessment is proposed.The method involves a fusion network that integrates semantic segmentation for component recognition and image classification for component damage state assessment.GCE loss and a pre-training weight freezing strategy were used to optimize and update the parameters of the multi-task fusion deep network model.The results demonstrate that the proposed method achieves a validation accuracy of94.74% for component semantic segmentation and 97.14% for damage state classification.Additionally,it validates the semantic segmentation performance of the proposed geometric constraint deep learning method for multi-class components in post-earthquake buildings.(5)A visual software system for multi-scale seismic damage identification and evaluation of buildings is developed by integrating the semantic segmentation model algorithm.This system allows for post-earthquake building area recognition and collapse state assessment,as well as damage grade assessment of building groups(destroyed/severely damaged/others)and pixel-level recognition and classification of multi-type and multi-damage state building components.The system is based on largescale remote sensing images,mesoscale UAV low altitude remote sensing images,and small-scale component images.The proposed geometric constraint deep learning method for multi-scale earthquake damage identification of buildings is verified for effectiveness and accuracy using the earthquake site of the old county of Beichuan as the research object.