Research On Displacement Design Spectrum Based On Convolutional Neural Network

As the development of society and the improvement of economic level,people’s attention to the influence of earthquakes has increased from life safety to comprehensive economic losses.Performance-based seismic design theory breaks through the limitation of"guaranteeing life safety"as the main fortification goal in traditional seismic design,which aims to effectively control casualties and economic losses,and ensure the performance of structures.At the same time,the displacement-based seismic design method is considered to be one of the most effective ways to realize the performance-based seismic design theory.For example,the acceleration design spectrum plays an important role in the force-based seismic design,and the displacement design spectrum is a significant basis for determining the seismic action in the displacement-based seismic design method.At present,the displacement design spectrum has become one of the hot research issues in the field of earthquake engineering and civil engineering.Scholars domestic and abroad have carried out plenty of research work on the displacement design spectrum and proposed various effective displacement design spectrum calibration methods.However,there are often significant differences between the seismic action calculated by these displacement design spectrum methods and the one in the current standards,which seriously hinders its application in engineering practice.To solve this problem,this paper selects the ground motion data matching the Chinese standard acceleration design spectrum based on the convolutional neural network,and it obtains the displacement design spectrum matching the Chinese standard acceleration design spectrum based on these data.The research work in this paper is of great significance to promote the application of displacement-based seismic design methods in engineering practice.The research content of this article mainly includes the following aspects:(1)In order to reflect the diversity of ground motions,this paper selects 11462horizontal ground motion components from 5,731 stations of 200 earthquakes with global magnitude greater than 5 and epicentral distance less than 200 km.This paper divides the Chinese standard acceleration design spectrum into 10 categories according to the difference of characteristic periodT_g.With representative training data is the key to using convolutional neural network to identify ground motion data matching the target spectrum.This paper choose wavelet analysis to calibrate the frequency components of actual ground motions so as to obtain training data that matches the target spectrum.(2)In order to obtain the optimal neural network structure form,this paper systematically compared 11 different forms of network structure forms in the application module of the keras deep learning framework.All training data are being input in the form of images,also,the difference between the logarithmic and natural coordinates of the coordinate axis,and the picture format with or without the target spectrum as a background reference are discussed.The pros and cons of the structure mainly refer to the accuracy and loss rate of the training set and test set.The higher the accuracy rate and the lower the loss rate,the better the network structure.(3)This paper classified 11462 horizontal ground motion components by using the trained convolutional neural network structure,and determined the optimal quantity of ground motion record in each group through combining the cumulative square error between the actual ground motion average acceleration response spectrum and the target design spectrum.Based on the principle of calibrating displacement design spectrum in Eurocode,a form of displacement design spectrum suitable for Chinese standards was proposed,and the recommended values of displacement design spectrum parameters are given based on selected data.