Vehicle Dynamic Load Recognition For Small And Medium Bridges Based On Deep Learning And Multi-source Information Fusion

With the rapid development of bridge construction in China,the importance of bridges in the national economy continues to climb,and as the most significant dynamic load on bridge structures,vehicle loads are the main factor causing the reliability of bridge structures to decrease and fatigue damage to occur.Therefore,accurate identification and tracking of vehicle moving loads on bridges is of great importance for bridge health warning.In this paper,we propose the YOLOv5 Deep SORT-BP(YD-BP method for short)moving load identification method based on deep learning and multi-source information fusion by combining the advantages of bridge dynamic weighing that simultaneously utilizes the vehicle axle position information and bridge dynamic response information,and the moving load identification method to identify the dynamic time-range forces of the bridge.The method firstly uses the YOLOv5 Deep SORT multi-objective tracking algorithm to locate the real-time spatial position of the axle directly with the wheel as the detection and tracking target;secondly uses the localized axle position information to fuse with the dynamic response information of the bridge;finally,the fused information of the two parts is used to identify the vehicle moving load on the bridge with the powerful nonlinear high-fitting mapping capability of the BP neural network.Finally,the fused information is used to identify the vehicle dynamic loads on the bridge with the help of BP neural network.The numerical simulation of vehicle-bridge coupling is carried out for a 30 m span simplysupported T-girder bridge.Firstly,the effectiveness of this paper’s axle positioning method is analyzed by comparing the direct use of dynamic response to identify the axle position.Secondly,the effects of different speeds,vehicle weights,unevenness and noise levels on this method and the traditional method are analyzed under single-vehicle conditions.Finally,the effectiveness of this paper’s YD-BP method is studied under multi-vehicle conditions.The following conclusions are drawn:(1)Using YOLOv5 Deep SORT multi-target tracking algorithm can effectively solve the problems of difficult detection and localization such as axle occlusion,small axle targets and driving out of the detection area.At the same time,this method is more effective in locating the axle than the direct use of bridge dynamic response to locate the axle position without noise interference,which can effectively provide accurate position information for moving load identification,and also provides a new method for axle detection by bridge dynamic weighing method.(2)Under the single-vehicle condition,the noise level has the greatest influence on the recognition result,followed by the vehicle weight,and the speed and road grade have the least influence on the recognition effect when the magnitude of the moving load is recognized by YD-BP and the traditional method in this paper.However,in general,compared with the traditional dynamic load recognition method,the YD-BP method in this paper can suppress the influence of these factors,and the recognition accuracy is higher and the stability is better.Especially when the vehicle weight is small,this paper method compared with the traditional method average relative error from 47.96% reduced to 24.17%,pass rate from 8.82% to 58.82%,can be better applied to the actual engineering bridge vehicles are mostly small vehicle weight traffic status.(3)Under multi-vehicle condition,when using this paper YD-BP method to identify multiple load magnitudes of transverse multi-vehicle,it can effectively separate the dynamic response caused by single axle,the average relative error of identification is 7.93%,and the qualification rate is 70.22%,which has good identification effect.Simulating the real situation by adding noise to the dynamic response data,when the noise interference seriously reaches 10%,the average relative error of recognition is 24.85%,which still has certain recognition effect,but in the actual engineering,it is recommended to filter the dynamic response first to reduce the influence of noise on the recognition accuracy.