Research On Traffic Load Monitoring For Highway Bridge Based On Machine Learning

Faced with the increasing traffic load,the safety risk of the extremely large number of small and medium span Bridges is difficult to ignore.However,there is a lack of a practical and fast monitoring method for bridge load in the road network.Starting from the machine learning method,this paper uses different neural network models to develop three sets of bridge traffic load identification methods with high efficiency and good precision for highway Bridges under different scenarios.The main research contents and achievements include:(1)Aiming at the requirements of traffic load identification in highway Bridges,this paper proposes a set of highway vehicle load identification method based on video and ETC data,and adopts the integration of YOLOv5 algorithm,Deep Sort algorithm,KNN clustering algorithm,LPRNet algorithm and other algorithms.In the real bridge application,after considering the continuous recognition of video frames,the accuracy of the model for multi-target detection can reach 100%.The coordinate transformation relationship between pixel coordinates and world coordinates is established automatically by several representative points in the video image.The maximum relative errors of the method for longitudinal and transverse positions are5.1% and 4.5%,respectively.Through LPRNet,the character recognition of the detected vehicle license plate picture is carried out,and then the vehicle weight information is obtained through the fuzzy matching between the Levenshtein distance and the license plate characters in ETC database.The matching rate is more than 95%.KNN algorithm is used to calculate the axle weight of the vehicle,and the relative prediction error is mostly below 10% under the 95%guarantee rate.The above program is used to reproduce the temporal and spatial distribution of bridge traffic flow.The bridge load response after the simulation calculation of vehicle load sequence obtained by video recognition can be well matched with the monitoring bridge load response measured by SHM system,which provides valuable guidance for the assessment and management of bridge health condition.(2)Aiming at the demand of traffic load identification for ordinary highway Bridges lacking ETC,this paper proposes a set of BWIM method framework based on GAN.A variant network structure Gan-Autoencoder derived from GAN and a network structure based on generative antagonism mechanism are used to realize the dynamic recognition of bridge influence lines and vehicle axle loads respectively.The recognition effect of the proposed network architecture on influence lines and vehicle axle loads is verified by vehicle-bridge coupling dynamic numerical analysis algorithm.Compared with the traditional Moses algorithm and GAN structure based on physical constraints,the axle load recognition network based on adversitygeneration mechanism has a higher recognition accuracy than the traditional Moses algorithm and GAN based on physical constraints in terms of total vehicle weight and axle load.Under B roughness,the proposed method can be used for different vehicle weights and speeds.For vehicles with different wheelbases,the highest relative error of axle weight identification is13.04%,the highest relative error of total weight identification is 1.41%,and the recognition error of total weight is basically within 1%.(3)Aiming at multiple groups of vehicle-induced bridge responses with unclear vehicle weights,this paper proposes a bridge dynamic weighing method based on transfer learning.The method adopts DANN to realize axle load identification of vehicles from different bridge response data,and verifies the identification effect of the proposed network on vehicle axle load by using the improved MSCA program.Compared with the neural network after training transfer under different data amounts,the axle load and total weight of different bridge strain data are predicted and compared.The identification accuracy of DANN model for large vehicles is equivalent to the transfer learning network model after the pre-training model is trained with 30%-40%training data of the source domain data set.