Magnitude Estimation For Earthquake Early Warning Based On Artificial Intelligence

Magnitude estimation is an important work in earthquake early warning(EEW),the judgment of earthquake influence area and the release of EEW information all depend on the rapid magnitude estimation.Meanwhile,magnitude is also one of the important parameters for EEW system to release early warning information.Due to the requirement of EEW timeliness,the determination of magnitude can not wait until the end of the earthquake.It is necessary to use the initial information of P-wave to predict magnitude immediately,and to update the magnitude prediction in real-time with the increase of data,in order to obtain more accurate and reasonable results.Because the traditional method only uses single characteristic parameters of the initial waveform to establish a linear relationship to predict the final magnitude with the large discreteness,especially the problem of small earthquake overestimation and large earthquake underestimation needs to be solved urgently.In order to estimate the EEW magnitude more accurately and in view of the problems existing in the above EEW magnitude estimation,this paper uses the Japanese K-net strong motion data triggered by P-wave without requirement of the SNR and epicenter distance for strong motion data,at 1s?10s time window after the P-wave arrival,taking 1s as the time interval.Based on multiple characteristic parameters as the input of artificial intelligence method,two artificial intelligence magnitude estimation models are established,and the evaluation method of magnitude estimation accuracy is proposed.The main content of this paper is as follows:(1)Researching on magnitude estimation based on SVM.Using the Japanese K-net strong motion data after P-wave arrival,and based on the classical SVM method in the field of artificial intelligence machine learning,12 characteristic parameters including amplitude parameters,period parameters,energy parameters and derivative parameters are selected as the input of SVM,and SVM-M model is constructed.Meanwhile,these parameters as the input of the model also make the SVM-M model more interpretable.The results show that at the 3s time window after P-wave arrival,compared with the traditional magnitude estimation?_c method and the P_d method,the error and discreteness of the SVM-M model magnitude estimation are obviously improved,the magnitude estimation of the SVM-M model are not affected by the SNR and epicenter distance,and the overestimation of small earthquakes is obviously improved.Based on the magnitude estimation of the SVM-M model,the real-time magnitude estimation analysis of other 49 events in different magnitude segments are carried out.The results indicate that for the events with M_JMA3?M_JMA6.4,a more reliable magnitude estimation can be given at the first few seconds at the station where the first P-wave arrival.For the events with M_JMA3?M_JMA8,the magnitude estimation error is large at the 1s time window of the station where the first P-wave arrival,but with the increase of the time window,the error decreases gradually,and at the 4s time window of the station where the first P-wave arrival,a more reliable magnitude estimation can also be obtained.(2)Researching on magnitude estimation based on deep convolutional neural network.Also using the Japanese K-net strong motion data in the time window after P-wave arrival,based on the deep convolutional neural network,using the same 12characteristic parameters as the input of the deep convolutional neural network,DCNN-M is constructed.The results show that at the 3s time window after the P-wave arrival,the standard deviation of the magnitude estimation error of the DCNN-M model is smaller than that of the?_c method and the P_d method.Meanwhile,the magnitude estimation of the DCNN-M model is independent on the SNR and epicenter distance,and the problem of overestimation of small earthquakes has been obviously improved.The same 49 earthquake events,for events with M_JMA3?M_JMA6.4,a more reliable magnitude estimation can be obtained at the first few seconds of the station where the first P-wave arrival;for the events with M_JMA6.4?M_JMA8,with the increase of the time window of the station where the first P-wave arrival,the problem of magnitude underestimation has been obviously improved.(3)Evaluation of the accuracy of EEW magnitude estimation.For the magnitude estimation of a single station,according to the magnitude estimation of the SVM-M model and the DCNN-M model,the definition error is acceptable within 0.6 magnitude units.And for the real-time magnitude estimation of multiple stations,the area surrounded by the estimated magnitude of multiple stations and the catalog magnitude is defined as the error area.Based on the definition of the accuracy of magnitude estimation for single and multiple stations,the magnitude estimation of SVM-M model and DCNN-M model are analyzed.The results show that in the case of single station or multiple stations,the accuracy of magnitude estimation of SVM-M model and DCNN-M model increases with the increase of time window;meanwhile,in the real-time estimation of EEW magnitude of multiple stations,for M_JMA3.2,M_JMA4.4 and M_JMA5.4 events,a more reliable magnitude estimation can be obtained at the first few seconds of the station where first P-wave arrival.