Research Of Identification Method For Bridge Local Damage Based On Deep Learning And Optimal Inversion Analysis

Bridge construction is an important engineering activity related to the economy and people’s livelihoods.Absent modern and intelligent testing methods,many bridges are worn down for years without repair and are still operating in poor condition.Against the backdrop of “new infrastructure” and “building China as a country with a strong transportation network,” innovative studies on the identification of bridge damage that integrate disciplines such as civil engineering,informatics,and mechanics promote technical innovations in bridge-related calculation theories,maintenance modes,and safety assessment,improve the efficiency of bridge-testing equipment,and reduce the work intensity of such equipment,so as to prolong the service life of bridges,reduce operating costs,and fundamentally improve their economic and social benefits.Based on the theories of deep learning,computer vision,and inverse analysis,a fast bridge damage-detection platform is developed to solve the problems encountered in bridge testing.Image-classification,target-detection,and semantic-segmentation algorithms are proposed to identify damage on bridge decks,bridge structure defects,and bridge cracks,respectively.Along with non-contact image measurement,extended finite element and particle swarm optimization algorithms are applied to determine the depths of cracks in reinforced concrete structures.The highlights and research findings of this study are as follows.(1)A fast bridge damage-detection platform based on computer vision and a multi-platform vehicle-mounted fast bridge damage-detection system are designed,including new detection devices and modules for image acquisition,GIS tracking,long-distance image transmission,and wireless power control.A stability-enhancement module offsets natural jitter during image acquisition to ensure posture balance and smooth transition.A dynamic model of the telescopic mechanism and vertical arm are established to determine the structural strength,analyze the dynamic characteristics of bridge structures in operation,and obtain the load and dynamic characteristics of the arm stand.Civilian vehicles are equipped with a CCD image-acquisition module to quickly inspect damage on bridge decks in real time.(2)Based on images of bridge damage collected with different backgrounds,a database of typical bridge damage is created,including images of 14 types of damage on bridge decks,4 types of bridge structure defects,and 6 types of bridge cracks.A fine-grained classification algorithm based on a double-flow network(TBC-Net)is proposed to solve the problem that damage on the bridge deck cannot be detected and identified quickly.The results show that the trained model has an accuracy of 92.3%in classifying the 14 kinds of damage on the bridge deck.To overcome difficulties identifying and locating bridge structure defects under complex light and with shadow backgrounds,a deep learning target detection algorithm(TBO-Net)is proposed to classify,identify,and locate bridge structure defects.The network structure and damage function are optimized based on the single target-detection algorithm,and local damage on bridges of different structures is classified and located using a rectangular recognition framework with a confidence coefficient.The average accuracy reaches 90.43% in the classification and location of cracks,spalling,exposed reinforcing bars,and free lime.In view of difficulties extracting multiple types of cracks in edge detection,a semantic segmentation algorithm based on weak supervised learning(TBS-Net)is proposed to conduct pixel-level detection and quantitative identification of bridge cracks with different backgrounds,such as bridge structural cracks(including underwater),cracks on concrete bridge decks,and cracks on asphalt concrete bridge decks,with accuracy reaching 94.2%.(3)Based on the above findings,as well as digital image correlation(DIC)non-contact image measurement technology,a parametric inversion method is proposed to identify internal cracks of reinforced concrete bridges.Combining a dynamic extended finite element algorithm and particle swarm optimization algorithm,we establish a model to invert the depth of cracks in reinforced concrete structures and deduce the finite element variational process in discontinuous structures.Through loading and unloading experiments on reinforced concrete beams,the rationality and effectiveness of the proposed model are verified by means of joint numerical simulation.By comparing the strengths and weaknesses of static and dynamic inversion optimization,it is found that the measurement response data used for inverse analysis should be obtained by a dynamic scheme as much as possible,and the established inverse analysis model can accurately calculate the depths of cracks in reinforced concrete beam structures.This study develops novel platforms,intelligent algorithms,and damage identification methods.These methods can be used for comprehensive identification of local damage in bridges from the exterior to the interior.