Research On Non-contact Bridge Structure Identification Based On Deep Learning And Bayesian Theory

Since the new century,China ’s economic construction and infrastructure construction are flourishing,and transportation demand has also increased greatly,which poses challenges for the safe operation of various transportation infrastructures.Therefore,it is of great significance to ensure the overall safety of the bridge and prevent accidents by implementing the necessary monitoring work for the existing bridge structure through advanced scientific means,and then quickly and accurately identify the safety performance and damage state of the existing bridge.As the performance index of the bridge structure,the bridge influence line is widely used in the safety state identification of the bridge structure.Large bridges generally acquire bridge structure state information by installing large high-precision contact sensors.Due to economic effects and other factors,a large number of small and medium-sized bridges lack the support of professional high-precision contact detection system.Because the commonly used bridge influence line identification method usually has the shortcomings of closed traffic,damaged pavement,high cost,etc.,therefore,for the structural detection of small and medium-sized bridges,non-contact bridge influence line identification becomes an option.According to the requirements of non-contact bridge influence line recognition,this paper proposes a bridge influence line recognition method based on deep learning and Bayesian theory,and applies it to real bridges.In this paper,theoretical,numerical and experimental studies are carried out :(1)On theoretical research,this paper compares and summarizes the existing classification standards of road vehicles,and puts forward the classification standards of road vehicles suitable for the needs of this paper according to the research requirements of the subsequent influence line identification.This paper discusses the advantages of Bayesian theory in the field of parameter identification,and combines the influence line inversion algorithm based on bridge response data with Markov chain Monte Carlo method.An improved idea of bridge influence line identification method based on Bayesian theory is proposed,which extends the algorithm from the theoretical application scenario of single vehicle single speed condition to the practical application scenario of multi-vehicle multi-speed mixed condition.(2)On the research of numerical simulation,this paper counts a large number of vehicle factory information data,summarizes the axle distance,axle load and axle load distribution coefficient interval of each vehicle in the new classification standard,establishes the mapping relationship between road vehicle models and vehicle basic information,and applies weight coefficients to different values of vehicle weight interval.The YOLOv4 neural network is improved,and the attention mechanism is added to train the recognition model of road vehicle models.The vehicle-bridge coupling model is established by ANSYS software and MATLAB program,and the bridge dynamic response value under the condition of multi-vehicle and multi-speed vehicle is calculated.The feasibility and accuracy of the improved Bayesian theory bridge influence line recognition model are verified.(3)In experimental research,the traffic flow video is obtained at the actual road traffic flow,and the accuracy of YOLOv4 vehicle recognition model is verified by video images.In the laboratory test,the single-span simply-supported T-beam bridge,the two-axis vehicle scale model and the three-axis vehicle scale model are used to conduct the non-contact identification verification test of the bridge influence line.The dynamic response data of the bridge are obtained through camera.By adding mass blocks and changing the vehicle scale model to change the vehicle axial distance and axle load to simulate the mixed multiple conditions,the posterior distribution and the maximum posterior estimation of the bridge displacement influence line are obtained,indicating that the identification error is within an acceptable range.In the field test,the dynamic displacement response data of the bridge is obtained by a fixed camera,and the vehicle type recognition and vehicle axle load information of the bridge vehicle traffic video captured by the UAV or camera are obtained by the YOLOv4 vehicle recognition model.The posterior distribution of the displacement influence line of the bridge is obtained by the proposed Bayesian theory bridge influence line recognition algorithm,and the maximum posterior distribution estimation of the influence line is used as the final recognition result.