Study Of Earthquake Early Warning System Based On Deep Machine Learning

Earthquake early warning refers to the rapid and accurate reporting of earthquake information before the destructive energy of the earthquake reaches the warning area,among them,the main scientific problems are the judgment of earthquake events(the detection of P-wave arrival),the location of the earthquake source and the estimation of the magnitude of the earthquake.For earthquake early warning systems,rapid and accurate P-wave pickup on arrival is essential.Traditional P-wave arrival picking algorithms include long-short-time averaging method STA/LTA,Akaike information criterion method AIC,etc.,which have the following problems:missing P-wave arrivals,mistakenly picking up background noise as P-wave,and low pickup accuracy.In order to solve these problems,this research proposes a deep machine learning-based P-wave arrival pickup method(Ppick),which is trained using Japanese underground seismic data..Ppick has a low false pickup rate,only 21 of the6578 noise samples are falsely picked up as P waves.The missed pick-up rate of Ppick decreases with the increase of the length of the P wave in the window:when the P wave duration in the window is between0-1s,the missed pick-up rate is the highest(92,1.4%);when the P wave duration in the window is between2-3 s,the missed pick-up rate is the lowest,and only one of the 6578 P-wave samples is missed.The test results show that Ppick’s P-wave arrival detection in earthquake early warning has broad application prospects.Rapid and accurate estimation of the epicenter location is also crucial for earthquake early warning.Traditional seismic source location algorithms are usually based on linear,nonlinear and probabilistic methods.They have the problems of large computational loss or insufficient location accuracy.There are also deep machine learning methods that directly use seismic waves to realize source location,but these algorithms use a long time.In this study,a fully connected neural network-based source localization algorithm DNNLOC is proposed.It takes the latitude and longitude of the first triggered stations and the distance difference between two stations and the source as input features,and outputs the latitude,longitude and depth of the source.The positioning accuracy of DNNLOC increases with the increase of the number of stations,and the standard deviation of the epicentral distance error and the standard deviation of the focal depth error decrease with the increase of the number of stations,the results of the test set show that DNNLOC also has broad application prospects in earthquake source location in earthquake early warning.Quickly and accurately estimating magnitudes is a critical step in earthquake early warning systems.Similarly,this study constructed a fast and accurate magnitude estimation algorithm CNN-EQMG based on the convolutional neural network CNN,which achieved far better accuracy than the traditional magnitude estimation algorithm P_dmethod.Compared with the traditional method,which realizes the magnitude estimation through the empirical relationship between the characteristic parameters of the seismic wave and the magnitude in the first few seconds,CNN-EQMG directly extracts the features related to the magnitude from the seismic wave and gets rid of the dependence on empirical experience and knowledge.CNN-EQMG uses a large number of surface strong earthquake records from Japan and Chile for training(98,257 records),validation(31,429 records)and testing(40,638 records),and uses strong earthquake records from the United States and New Zealand for generalization performance testing(583 records).From the results,the accuracy of CNN-EQMG increases as the input duration increases,and is better than the accuracy of the P_dmethod:when the initial seismic wave duration is 3 s in the range of magnitude 4 to6.4,the CNN-EQMG method’s accuracy of estimating the magnitude is 1.5 times that of the P_dmethod.In the range of magnitude 6.5 to 9,the CNN-EQMG method is 1.2 times more accurate than the P_dmethod.When the initial seismic wave increases from 3 s to 10 s,the CNN-EQMG method can continuously improve the accuracy of estimating the magnitude with the increase of the seismic wave duration,and is always higher than the P_dmethod,especially for earthquakes of magnitude 4-6.4,the CNN-EQMG method can estimate the magnitude of the earthquake at the first 3 s with the accuracy 1.2 times of that by the P_dmethod when the wave duration is 10 s.The CNN-EQMG method automatically learns more relevant features from the first-arrival seismic waves,which greatly improves the accuracy and timeliness of magnitude estimation,and can provide faster and more accurate magnitude estimation for earthquake early warning systems.