Strong Ground Motion Recording Features Analysis Based On Deep Learning

At this stage,the accurate prediction of earthquakes,especially the short-term and imminent prediction,is still difficult to achieve as a recognized worldwide scientific problem.Earthquakes are irregular.The direct disasters and secondary disasters will bring great casualties and losses.Therefore,in order to prevent earthquakes from causing greater damage and losses,the aseismic analysis,aseismic design,and early warning emergency measures are particularly important.Deep learning has made great progress in the past thirty years and has been widely used in many fields,especially in the field of seismology,such as seismic lithology prediction,seismic event detection and location,seismic phase detection and picking,phase correlation.Because of its ability to build more complex models of things or abstract concepts,deep learning provides a new perspective for the study of earthquake engineering,and also provides a new way for more effective earthquake disaster prevention.In view of this,this thesis establishes deep learning models for key issues such as ground motion input adjustment methods and earthquake early warning.We take,from 2003 to 2019,involving 1698 stations,119,241 waveform recordings of 634 earthquakes download in K-net and Ki K-net as a database,and conducts in-depth research on several issues such as data pre-processing,activation function,optimization function,model framework,parameter tuning.The main research work completed and the innovations results obtained are as follows:1.A seismic magnitude classification model based on convolutional neural networks is proposed,which can provide a reference for selecting appropriate ground motion input adjustment methods.Based on the database,pre-processing steps such as baseline correction and normalization are performed on each acceleration recording.The sampling start and end points are determined by Arias intensity for segmented sampling,and the magnitude 5.5 is used as the classification boundary to form a sample set.After optimization function selection,learning rate,batch size,and other hyper-parameter tuning,the average accuracy of model classification reached 93.3%.Based on this,the model can judge well whether the sample adjusted from the small earthquake has the characteristics of large earthquakes if used as ground motion input.Then,referring to the artificial synthetic earthquake method,a method of adjusting the frequency domain ground motion input is proposed.After model testing,98% of small earthquake records adjusted by this method is recognized as a large earthquake.From this,it is judged that the model classification is based on the relevant characteristics of the seismic recording frequency domain,which can provide a certain theoretical basis for the model.2.Arias intensity trend prediction model based on the fully connected neural network is established,and this model can improve the timeliness of earthquake early warning to a certain extent.Based on the acceleration recordings in the database,the data is sampled in segments after acceleration scaling and baseline correction.According to the importance of the data,the data is sampled at 1Hz,10 Hz,100Hz,and the duration is15 s,10s,and 5s,which does not affect the model training effect and reduces the cost of training time.We use the increase or decrease of the Arias intensity in the future 5s as a label to form a sample set,and build a fully connected neural network model to predict the trend of Arias intensity.After training and tuning,the model finally reached an average prediction accuracy of 76.5%.The model can be combined with an early warning system based on local site conditions,etc.,to improve the effectiveness and advancement of the early warning system.3.The magnitude prediction model at the initial rupture stage is constructed,and explore the possibility of determining the magnitude of the earthquake magnitude during the initial rupture stage.Taking the 0.2% Arias intensity time point as the starting point,taking the 0.2s acceleration record with the 100 Hz sampling frequency as the model input,and taking the magnitude 5.5 as the label boundary as the sample set,a fully connected neural network is constructed to predict the magnitude during the initial rupture stage.After training and tuning,the correct rate of the model reached 81.2%.After that,six samples with different magnitudes but high correlation coefficient(about 0.8)were taken from the sample set,and the magnitude of the samples can still be accurately predicted by the model.Therefore,in the initial rupture stage of an earthquake,even if the acceleration recordings of a large earthquake and a small earthquake have a high correlation coefficient,there are still certain differences in features that are difficult to extract manually at this stage.The research contribution and application value of this thesis are mainly: the use of deep learning,a relatively new method,provides new ideas for the selection of ground motion input adjustment methods,and proposes a model that can predict the trend of Arias intensity and improve the timeliness of early warning.To a certain extent,this thesis advances the research on the possibility of predicting the magnitude of the earthquake at the initial rupture stage.