Prediction Of Bridge Dynamic Response Under Earthquake And Vehicle Load Based On Deep Learning

Seismic action and vehicle load are the mainly dynamic load sources of bridge.Therefore,the study of bridge dynamic response analysis under seismic action and vehicle loads has important theoretical and engineering application significance.Conventionally,the establishment of a numerical model is necessary for structural dynamic response analysis.Surrogate model method generates the surrogate model by structural monitoring data.And then it is used to predict the structural response as well as evaluate its safety,which makes up for the shortcomings of the conventional methods.In this dissertation,deep learning based surrogate models are established to predict the dynamic response of bridge structures under earthquake and vehicle loads.Besides,deep learning based safety assessment of bridge is also studied.The research contents can be summarized as follows.(1)A modified Prandtl-Ishlinskii hysteresis model for structural nonlinearity description is established and introduced into the deep feedforward neural network.The modified Prandtl-Ishlinskii deep neural network model is proposed for seismic response prediction of bridge.The prediction ability of the proposed model is valid by numerical simulation and shaking table test data.For the numerical simulation data,the prediction results of the deep learning model are close to the real value in both frequency domain and time domain and all the maximum peak errors are less than 10%.In shaking table test data validation,a matching peak response estimation result can be obtained by deep learning model.(2)Based on Runge-Kutta numerical integration method,Runge-Kutta deep recurrent neural network model is proposed to predict the dynamic response of bridge under seismic loads.The reliability of the model is validated by numerical tests of linear system and nonlinear system.And the proposed model is used to predict the seismic response of a finite element bridge model.The ability of the proposed model in monitoring data processing is studied.(3)The ability of the deep feedforward neural network model(DFNN),deep convolutional neural network model(CNN)and the deep long and short term memory(LSTM)neural network model to predict the dynamic response of bridges under vehicle loads is studied comparatively.Three representative vehicle loading models are established to generate model training data.In general,DFNN,LSTM network model and CNN have considerable accuracy in bridge response time history prediction.When road roughness is ignored,CNN has the shortest training time and the computational efficiency is better than the other two networks.DFNN is the most effective due to the significant reduction of model parameters when road roughness is considered.(4)A method for bridge random vibration analysis based on Bayesian deep learning is proposed.The effectiveness of the Bayesian deep learning model is validated by a 3D stochastic vehicle-bridge coupling model.The results show that the model can effectively obtain the mean and standard deviation time histories of bridge displacement and acceleration response,as well as the time-varying power spectral density and Fourier spectrum,the max error is less than 8%.The prediction results of Bayesian deep learning model agree well with the analysis results of vehicle-bridge coupling dynamic system.(5)For the safety assessment of bridge structures,a reliability analysis method of vehicle-bridge coupling system based on deep learning surrogate model is proposed.Four safety indexes related to bridge dynamic response and wheel-rail force were predicted by deep learning based surrogate model.Monte-Carlo simulation and first order second moment method based on automatic differentiation were used to analyze the reliability and sensitivity of the structure.