Machine-learning-based Engineering Emergency Response Parameters Estimation For Earthquake Early Warning

Earthquake early warning（EEW）is one of the effective ways to reduce natural disasters and casualties during earthquakes and take emergency measures for critical infrastructures to avoid secondary disasters.The core algorithm of EEW is to use the initial information of seismic waves to quickly estimate the EEW parameters.However,due to the lack of observation data,the accuracy of parameter estimation is low.Especially for earthquake early warning of critical infrastructures,the sparse station layout makes it impossible to use the average of multiple stations to reduce the discreteness of parameter estimation.The accuracy of earthquake early warning parameters based on a single station needs to be improved urgently.Machine learning has emerged in recent years.It extracts effective features from the data stream and deduces the characteristics of the new mechanism on this basis to improve the prediction ability of the system,which has been initially applied in seismology.Based on classical machine learning methods such as generative adversarial network（GAN）,random forest（RF）,and support vector machine（SVM）,this study uses K-NET and Ki K-net strong ground motion data from Japan to establish the training,testing,and verification data set with the single station,and extract effective features that are highly correlated with predicted target parameters to build a machine learning algorithm for seismic event identification,magnitude estimation,and emergency response parameter prediction for engineering earthquake early warning.On this basis,this study:（1）explores the algorithm of seismic event identification based on GAN,especially the interference elimination method of non-seismic events;（2）explores the earthquake magnitude estimation method based on GAN,and analyze its performance of large earthquake underestimation,small earthquake overestimation and magnitude classification;（3）explores the onsite prediction algorithm of earthquake emergency response parameters for critical infrastructures（high-speed railway,nuclear power,gas）based on SVM,and analyze its prediction performance according to confusion matrix.The main results of the research are as follows:（1）Based on the wave characteristics extracted by an unsupervised learning algorithm,the study provided a seismic event identification algorithm using GAN.The algorithm uses the generator to simulate false seismic waveforms and the discriminator is used to judge the true and false waveforms to improve the classification performance of the neural network,transforming the complex seismic identification problem into a simple binary-classification problem of earthquakes and microtremors.The results show that the GAN can distinguish more than 99%of seismic waves and microtremor noise,which is a significant improvement compared with the traditional STA/LTA method,which has an identification rate of 85.61%.At the same time,the classification performance of the combined model of GAN and SVM,and the combined model of GAN and RF are also explored.The results show that more than 99%of earthquakes and microtremors can be distinguished.Especially,all the square wave and sine and cosine signals can be eliminated.Therefore,linking SVM or RF after the neural network model can effectively eliminate non-earthquake events,which can provide technical support for future earthquake early warning applications.（2）Considering the difference in wave characteristics between large earthquakes and small earthquakes,a magnitude classification model based on GAN,the combined model of GAN and SVM are proposed.This study analyzed the influence of different training time windows,and the"traffic light"system is used to purify the results of the demarcation point,so as to maximize the performance of magnitude classification.The results show that the combined model can improve the accuracy of small and medium earthquakes by 22.75%and that of large earthquakes by 6.42%compared with the GAN with a higher confidence,and it is not affected by the time window.At the same time,based on three original waveform records,a magnitude estimation algorithm based on GAN is constructed,and the probability of the absolute value of the model error within the range of 0.88 magnitude is 95.52%with the input of displacement records.Compared with traditionalτ_c and _dP methods,the GAN model based on displacement records improves the error standard deviation by 1.6553 and 0.3878,respectively.And under the same magnitude range,the model can increase the error percentage by 57.98%and 23.32%,respectively,which verifies the accuracy of the GAN model in magnitude estimation,and effectively alleviates the overestimation of small earthquakes and the underestimation of large earthquakes.（3）With the advantage of machine learning being superior to traditional single parameter in the prediction of ground motion parameters,the study selected the characteristic parameters through theoretical analysis,and trained SVM to predict the emergency response parameters of critical infrastructures(the high-speed railway PGA_0.05-5Hz,the gas SI value and the nuclear power CAV),and then provide the optimal feature parameter combination with the smallest error standard deviation under different input feature parameters.The results show that for the three target parameters,the6-parameter combination Pa&Pv&Pd&CAV&Ia&IV2 can obtain the smallest error standard deviation.At the same time,based on the thresholds of the three target parameters,the study calculated the confusion matrix under each threshold,and determined the performance of each target parameter under the first threshold.For the gas SI,the accuracy under the first threshold can reach 100%.With the increase of the threshold,the accuracy of each target parameter continues to increase,which verifies the superiority and availability of the emergency response parameter prediction model combined with machine learning.