Research On Magnitude Estimation Method Based On Convolutional Neural Network

Earthquake early warning refers to the technology to quickly estimate relevant seismic parameters after an earthquake occurs,relying on dense earthquake monitoring networks and other infrastructures,predict the magnitude and possible impact of the ground motion that has not yet reached,and issue an alarm before a certain area of ground motion causes damage or impact.Its technical system has high requirements for timeliness,but the available seismic monitoring information is very limited,which is mainly reflected in:(1)there are fewer stations receiving seismic observation data near the local seismic source;(2)the amount of seismic waveform data received by the observation station receives is limited;(3)Observation data needs to be automatically processed in real time.As a manifestation of the magnitude of an earthquake,the estimation of magnitude is an important intermediate link in the earthquake early warning system.In order to improve the accuracy and reliability of the earthquake early warning system using limited waveform information for rapid and continuous magnitude estimation,this paper uses the information of the data within a few seconds after the P wave is triggered,and uses the neural network algorithm in the field of artificial intelligence to construct the magnitude estimation model to grt a more reliable estimation method of the earthquake early warning magnitude.The main work is as follows:(1)The 50,424 sets of three-component records of the KiK-net strong motion network in Japan from 1997 to 2020 were sorted out,and the data was established according to the 7:3 ratio into the training data set and test data set of the neural network by performing P-wave phase identification,baseline correction,filtering,and integration on the data.Magnitude regression fitting was performed on 10 kinds of seismic wave characteristic parameters such as amplitude parameter,period parameter and intensity parameter of the waveform data,and the relationship between the parameters and the magnitude estimation was discussed.(2)Constructed a CNN-M model for magnitude estimation of convolutional neural network based on multiple parameters of ground motion.A total of 10 seismic wave characteristic parameters and epicenter distances in each time window 1-10s after P wave triggering are used as the input of the deep neural network for training and testing,and the CNN-M continuous magnitude estimation model is established.The results show that compared with the ANN-M model which constructed in paper and the magnitude estimation "Pd method" the standard deviation of the CNN-M model is significantly reduced.The standard deviation of the CNN-M model is 0.400 in 1 s time window,which is gradually reduced over time,and it is reduced to 0.251 in the 10s time window;The CNN-M model also showed good performance on issues that are of concern to earthquake early warning such as undere stimation of large earthquakes and overestimation of small earthquakes.Through analysis of earthquake cases,it performed well in two large earthquakes with magnitudes higher than 6.5.(3)The frequency spectrum information is obtained by fast Fourier transform processing the ground motion records in the original time domain in each time window of 1～10s,and the convolutional neural network is used to construct the CNNFFT-M model,and the frequency spectrum information obtained after processing is used Magnitude estimation,analysis with CNN-M model and "Pd method" and test analysis of two earthquake examples show that this model performs well in earthquake test which below M6,but it is still needs improvement in the seismic field which larger than M6.