Predicting Amplification Factors Of Surface Ground Motion In Lower Hutt Basin Using Convolutional Neural Network

The analysis of local site seismic amplification characteristics is one of the important tasks of seismic safety evaluation.Seismic waves often exhibit anomalous amplification when they pass through a basin with soft soil layers and complex structure,causing the great disasters.Therefore,it is very necessary to study the seismic amplification effect of complex sites.Owing to the limitation of strong earthquake observation conditions,numerical simulation method was used to predict the amplification characteristics of ground motion in the past.However,due to the high computational cost and complex implementation of numerical simulations,it is difficult to consider all factors of seismic wave propagation process in numerical method.There exists the unsatisfactory accuracy in the prediction of seismic ground motion amplification.With the development of observation data,in this thesis,a novel prediction method for the amplification characteristics of the Lower Hutt Valley was proposed using the state-of-the-art convolutional neural network(CNN)combined with real-time seismic signals.It includes exploring the feasibility of CNN prediction method,comparing and analyzing the difference of prediction results with other models,discussing the rationality of numerical model based on real seismic records.The specific work is summarized as follows:(1)A Python program is developed for batch screening and downloading seismic records,extracting and calculating parameters with an automatic process.A large number of training parameters were provided for the establishment of a CNN model.(2)Using the generalized inversion method,by comparing three different spectral ratio calculation formulas,the characteristics of the spectral ratio of each station and the influence of different frequencies on each station are analyzed and studied.It is summarized the influence of site factors on seismic amplification effect in the sites.(3)The program of ground motion amplification feature prediction is realized by using MATLAB platform.Based on real seismic records,four CNN prediction models are established.The results are discussed,including the comparative analysis of the CNN and traditional BPNN,the prediction results of CNN-FSPA and CNN-PSPA models,and the comparative analysis of unrecorded site prediction,so as to verify the stability and accuracy of a CNN in predicting the surface ground motion amplification feature of the basin.(4)The CNN model is improved,and the results of the numerical model of the New Zealand site are compared with the four comparison models.Based on the results of the three-dimensional finite element model,it contains several comparisons,including the comparison of the seismic amplification results under the real earthquake,the comparison of CNN-FSPA-P8 and CNN-FSPA-P4 models with different training parameters analyzing the difference between numerical simulation and CNN-FSPA-P4 model.Finally,to analyze the feasibility of the four prediction models,it is compared that the prediction results without record points based on numerical simulation and the results of CNN models.(5)The work of this thesis is summarized and the shortcomings of this thesis and the future prospects are discussed.