Research On Nonlinear Stochastic Seismic Analysis Method Of Girder Bridge Based On Deep Learning

The existing seismic response analysis methods have some defects when dealing with nonlinear stochastic seismic response.Some cannot deal with random seismic effects,some cannot solve nonlinear problems,and some take too long to compute.Therefore,a nonlinear stochastic seismic response prediction method is proposed,and the spatial expression form of geometric information of girder bridge,artificial seismic wave synthesis,pile-soil interaction and other aspects are studied in depth.Based on BIM software Revit,an intelligent analysis module for seismic response of continuous rigid frame bridge is developed.Although the continuous rigid frame bridge is taken as the specific research object,the research results can be extended to all girder Bridges.The specific research content of the paper is as follows:(1)The proposed and comparative verification of three-dimensional space tensor.In order to fully express the spatial geometry information of girder and bridge,a new 3D representation of bridge structure,called 3D Space Tensor,is proposed after analyzing the spatial shape characteristics of girder and bridge structure.Based on this expression form,the 3D Convolution Frequency Prediction Network is constructed and trained to predict the first five frequencies of continuous rigid frame Bridges.The feasibility and effectiveness of the three-dimensional space tensor are verified by comparing the 3D convolutin network with the traditional shallow neural network prediction results and the bridge load test results.(2)The establishment of the beam bridge mode prediction model.Deep learning requires a large number of samples for network training,but it takes a lot of time to compute a single nonlinear dynamic response.In order to solve this problem,transfer learning technology is used to obtain the mode shape prediction network of continuous rigid frame bridge,and then the network is transferred to the earthquake response prediction problem.By establishing the mode tensor and introducing autoencoder technology,the problem of mode shape prediction of continuous rigid frame bridge is transformed into mode shape code prediction,and the dimension of neural network output is reduced.Taking the 3D Space Tensor as the input and the mode shape tensor encoding as the output,the 3D Convolution Mode Shape Code prediction neural network is constructed and trained.It is combined with the mode shape tensor decoder to predict the mode shape of continuous rigid frame bridge.(3)Artificial seismic wave synthesis based on Conditional Generative Adversarial Networks.Aiming at the problem that wavelet analysis can not adaptively decompose seismic signals,the Artificial Seismic Wave Variational Mode Decomposition Generation Algorithm is proposed.The Variational Mode Decomposition method was used to replace the wavelet decomposition,and the NSGA-II multi-objective optimization algorithm was used to optimize the component coefficients of the decomposed eigenmode function.The objective function was to minimize the error of the response spectrum and the adjustment of the component.The natural seismic wave was adjusted to make it conform to the standard design response spectrum.The existing methods of artificial seismic wave synthiesis can’t get rid of the dependence on the natur seismic wave data,so they can’t generate wave date in quantity and automatically.In view of this problem,a Conditional Generative Adversarial Networks,named Intelligent Generation Model of Artificial Seismic Waves,was established.The seismic wave data are used as model’s input,te corresponding response spectrum and the peak acceleration are used as labes of the input.Using the natural seismic waves adjusted by the Artificial Seismic Wave Variational Mode Decomposition Generation Algorithm as the dataset,the Intelligent Generation Model of Artificial Seismic Waves is trianed to achieve that the goal that giving a seismic wave’s label,the model can randomly generate a large number of seismic wave time-history curve meeting the requirements of the specification.(4)Simplified analysis method of pile-soil interaction based on Boulanger model.Aiming at the problems of complex modeling process and low reuse rate of modeling code in Open Sees,the Open Sees for Bridge library is developed based on Open Sees Py library.Aiming at the problems of complex and heavy computation of the existing pile-soil interaction model,a large number of Open Sees finite element models were established based on the Boulanger pile-soil interaction model by using the Open Sees for Bridge library.The influence parameters of the pile-soil interaction were analyzed,and the mathematical law between the pile length and the seismic response was obtained.Baed on the mathematical law,a simlified apprch to solve pilesoil interaction is proposed.(5)Nonlinear stochastic seismic response analysis method of continuous rigid frame bridge based on Deep Learning and its visualization.Based on Transfer Learning,the nonlinear stochastic seismic response prediction model of continuous rigid frame bridge is constructed and trained.The model use the 3D Convolutin Mode Shape prediction network as pre-trained model.The 3D Space Tensor and semismic wave data were taken as the model’s input,and the seismic response at its key position is used as the output.Based on the Monte Carlo principle,the prediction model is combined with the artificial seismic wave intelligent generation model,and a deep learning-based nonlinear stochastic seismic response analysis method for continuous rigid frame Bridges is constructed.Using BIM software Revit as the visualization development platform,the process of the nonlinear stochastic seismic response analysis method is visualized,and the intelligent analysis module of continuous rigid frame bridge is developed.