Research On Machine Learning On-Site Threshold Warning Method Based On Strong Motion Data In China

Earthquake early warning is to rapidly provide corresponding warning information to the target area before destructive seismic waves arrive after an earthquake,in order to seize time for implementing relevant emergency measures and achieve the purpose of earthquake prevention and disaster reduction.However,estimating the size and intensity of potential damage zone caused by earthquakes in the early stages is the foundation and key for rapid and accurate release of earthquake warning.The traditional estimation method of potential damage area of on-site earthquake mainly determines the τ_c value when magnitude M=6 and the P_d value of PGV when intensityⅠ≥Ⅶ through the statistical empirical relationship between the characteristic period parameters τ_c and amplitude parameters P_d of P-band and the seismic magnitude M and the seismic peak PGV,and a "four level discriminative model" is established with these two parameters as the threshold to judge the potential seismic damage at the ground of the station.However,this method still has some obvious shortcomings.For example,high requirements for signal-to-noise ratio and significant errors in the results of on-site ground motion and magnitude prediction.Therefore,the judgment method still needs further improvement.In order to solve the above problems,this paper selects all the strong motion data recorded by the China Earthquake Networks Center from 2007 to 2020 as the basis,replace two traditional single parameter prediction equations with two multi-parameter machine learning prediction models related to magnitude M and peak ground motion parameter PGV.According to its test results and combined with the threshold value of China’s instrumental intensity calculation standard,this paper establishes a "new four level discriminative model" to estimate the potential damage zone of the earthquake.The main research work of this paper is as follows:(1)According to the research demand,all seismic events with magnitude M>3 recorded by the China Earthquake Networks Center between 2007 and 2020 were selected.After P-wave picking and integrity screening of seismic records,8169 sets of available seismic records were obtained.Then,the target parameters and the corresponding 13 characteristic parameters under the 3s time window were calculated after integral calculation and ground motion synthesis of these seismic records.Based on the parameter inputs of the two machine learning prediction models used in this paper,the corresponding feature parameters are corrected and normalized for distance effects.(2)Input relevant feature parameters into two machine learning prediction models according to specific sequence combinations to obtain corresponding prediction results,and compare them with traditional the comparison of single parameter prediction methods,finally,error and residual analysis was conducted on the prediction results of machine learning,indicating that the machine learning method used in this paper has higher stability and superiority in earthquake magnitude estimation and PGV prediction.(3)According to the test results of the two prediction models,referring to the determination method of parameter threshold in the traditional "four level discriminative model" and combining with the Chinese instrumental intensity calculation standard,it is finally determined that in this paper,with M=6 and PGV=11.5cm/s as the thresholds,a "new four levels" discriminative model for potential damage areas in situ is established,and a confusion moment is established according to the relationship between the threshold,measured values and predicted values.Based on this,all possible scenarios for early warning results(such as successful alarm SAP,successful non alarm SAN,false alarm FA,and missed alarm MA)were defined.Combined with the common evaluation indexes of classification models in machine learning,the performance of the two machine learning prediction models and the rationality of threshold parameter setting were evaluated.Finally,it was concluded that the two prediction models had high accuracy and the threshold setting was also reasonable.(4)In order to further verify whether the prediction results based on the machine learning method and the discriminant result of potential damage area of site given by"New Four Leveling model" can meet the actual needs,four earthquake events of magnitude 6 or above were applied for offline verification.The final results show that three earthquake events can quickly and accurately determine the earthquake type and damage situation after the earthquake,and provide corresponding effective warning time.At the same time,comparing the warning results of all stations with the actual alarm results,it was found that although there were still cases of false alarms or omissions in the warning results of some stations,the discrimination results using predicted values and true values were the same when compared with the actual postearthquake intensity survey map.The vast majority of stations with warning results of level 3 were distributed within the range of Ⅰ≥Ⅶ.This further shows that the threshold setting of M=6 and PGV=11.5cm/s in this paper is reasonable.Based on machine learning,it is feasible to use the "New Four Leveling discriminative model" to estimate the potential damage area of earthquakes,and its early warning results are basically in line with the actual situation.