Study On Bridge Moving Load Identification Based On Deep Transfer Learning

Bridge is the throat of highway and the key node of transportation interconnection,which plays an important role in the national social and economic development.The most important external load on the bridge during service is the moving load caused by the vehicle.Under the action of long-term moving loads(especially overweight vehicles),the bridge structure is extremely easy to cause damage accumulation,damage aggravation,and even catastrophic accidents in extreme cases.Therefore,it is of great scientific research significance and engineering application value to quickly and accurately identify the moving load on the bridge and timely monitor and warn the over-limit vehicles,so as to improve the management and maintenance level of the bridge and ensure the operation safety of the bridge:Focusing on the rapid and accurate identification requirements of moving loads,this paper uses theoretical analysis,algorithm development,numerical simulation,real bridge verification and other research methods,combined with deep convolutional neural network and transfer learning strategy,to carry out research on bridge moving load identification method based on deep transfer learning.The main research contents and conclusions of this paper are as follows.(1)A bridge moving load identification method based on deep convolutional neural network and transfer learning is proposed.Firstly,the problem of bridge moving load identification is transformed into the classification task of moving load parameters.Secondly,the dynamic response data of the bridge are transformed into time-frequency images as input samples.Then,based on the transfer learning theory,the shallow feature extraction ability of the pre-trained deep convolutional neural network model is transferred to the moving load identification task.Based on the learning of the deep features of the input samples,the identification of the moving load parameters of the bridge is realized.Finally,the input sensitivity analysis is carried out based on the optimal model.The results show that the recognition rates of VGG16,Alex Net,Res Net and Mobile Net V2 models on the test set are all above 94 % in the task of vehicle speed recognition and vehicle weight recognition.Compared with VGG16,Alex Net and Res Net,the training time of lightweight convolutional neural network Mobile Net V2 is reduced by more than 50 %,the vehicle weight recognition speed is increased by42.2 %,and the vehicle speed recognition speed is increased by 50 %.In the vehicle speed identification,the recognition effect of acceleration sample input is better than that of speed and displacement input.In vehicle weight identification,the recognition effect of displacement sample input is better than that of acceleration and velocity input.(2)The influence factors of bridge moving load identification method based on deep convolutional neural network and transfer learning are analyzed.Firstly,a parameterized numerical model is established,and different parameter information is changed by parameterized numerical simulation.Then,the dynamic response data of the bridge under different parameter conditions are obtained and converted into time-frequency image samples.Finally,the deep convolutional neural network is used to learn the features of time-frequency image samples through transfer learning theory,and the moving load parameters are identified.The results show that the higher the roughness level of the road surface,the rougher the road surface,and the lower the recognition accuracy of the model for vehicle speed and vehicle weight information;with the decrease of the number of measuring points,the recognition accuracy of vehicle speed and vehicle weight information is basically unchanged.As the bridge span increases,the signal-to-noise ratio in the sample data increases,and the number of sample classifications increases,the accuracy of model recognition decreases.(3)Real bridge verification of bridge moving load identification method based on deep convolutional neural network and transfer learning.First,let the loading vehicle travel through the test bridge deck at different speeds,and collect the acceleration time history data of the bridge through the sensor;then,the collected acceleration time-history data are transformed into time-frequency images by continuous wavelet transform to construct a sample library.Finally,combined with the transfer learning theory,the lightweight convolutional neural network Mobile Net V2 is used to learn the features of time-frequency image samples,and the trained model is used to classify the test set samples to identify the vehicle speed information.The results show that the accuracy of bridge moving load identification method based on Mobile Net V2 and transfer learning is 88.33 % and 85.00 % respectively in the load test of two real bridges,which can accurately identify the moving load information.