Research On Rapid Assessment Method For Seismic Damage Of Bridges Based On Deep Learning

Generally,the seismic damage assessment of traditional bridge structures is carried out on-site by senior research scholars or engineering designers.However,this method is limited to the professionalism of the assessors and the disaster site environment,and the assessment time is long and the efficiency is poor.Therefore,the development of a fast,accurate and safe method for seismic damage assessment of bridge structures has important theoretical significance and practical value,and can provide scientific support and theoretical basis for the smooth implementation of disaster relief work.In this paper,a fast damage assessment method for bridge structures based on wavelet time-frequency analysis and convolution neural network is proposed.This method takes into account both time-domain information of ground motion and non-stationary frequency-domain characteristics,and establishes the mapping relationship between ground motion signal and seismic damage of bridge structures.The main research contents are as follows:(1)Wavelet transform is applied to the time-history data of ground motion acceleration to obtain the time-frequency characteristics of ground motion and draw the time-frequency map of wavelet,which is then aggregated into a convolution neural network data set.(2)A three-span continuous girder bridge finite element model is established by Open Sees platform,the bridge structure is analyzed by non-linear dynamic time-history,Park-Ang two-parameter damage model is used to calculate Park damage index,and Green,Yellow and Red tags are used to classify the damage index according to the damage objectives of seismic performance.(3)The widely used convolution neural network Alex Net,VGG16 Net,Res Net34,Res Net101 and lightweight convolution neural networks Mobile Net V2 and Mobile Net V3 are used to train the time-frequency image data set of wavelet.The convolution neural network suitable for seismic damage assessment of bridge structure is explored by combining the research object and objective of this paper with the method of migration learning.The research results indicate that this method has great potential for application in earthquake damage assessment of bridge structures,with the Res Net101 network model having the best recognition performance and a validation set recognition accuracy of83.25%.Compared to Res Net101,the lightweight convolutional neural network Mobile Net V2 reduces model parameters by 20 times,while the recognition accuracy only decreases by 11%.