| Destructive earthquakes may bring a series of serious consequences such as casualties and damage to facilities.Instrumental seismic intensity indicates the degree of damage to the ground and engineering buildings.Reliable and fast access to instrumental seismic intensity can provide a scientific basis for emergency response,disaster diagnosis,and emergency response.Therefore,real-time prediction of instrumental seismic intensity can help personnel and facilities to take countermeasures in disasters and provide reliable information for post-disaster rescue.Currently,most of the methods for real-time instrumental seismic intensity prediction are based on P-wave extraction of a single feature parameter to estimate the peak parameters of the full-time range,but it is difficult to characterize all the information of ground motion with a single feature.This is mainly due to the high dispersion of the statistical relationship between the single ground motion feature parameter and the maximum target parameter,and the variability of the prediction results for different parameter types.Therefore,in this paper,a real-time seismic intensity prediction framework is proposed and based on machine learning algorithms to predict seismic intensity in real time.Taking the Japanese region as an example,a specific application study is conducted around the proposed framework,and the main work is as follows:1.A reliable real-time seismic intensity prediction framework is proposed.The framework mainly consists of four steps:constructing a ground motion database,building and optimizing a prediction model,validating and interpreting the model,and using the best model prediction,which ensures the accuracy and reliability of continuous seismic intensity prediction.Meanwhile,61,726 sets of ground motion records(each set contains 3-components)recorded by the Kyoshin Network(K-NET)strong ground motion network in Japan from 2010 to 2022 were collected and selected as data.First,these strong-motion records were pre-processed and screened such as zero-line correction,and then the seismic phases were automatically picked up by the STA/LTA method and the AIC method and manually checked.Finally,the parameters of the triggered ground motion of the stations were extracted,and the final seismic intensity of each station was calculated to construct a ground motion database.2.A real-time seismic intensity prediction model based on multi-parameter driven is proposed.In this paper,multiple parameters are used to construct the prediction model,which more comprehensively characterizes the information contained in the P-wave,and an ensemble learning algorithm is used to solve the problem of insufficient nonlinear fitting capability in traditional methods.Subsequently,the Pearson coefficient is used to define the correlation between the feature parameters,and the 24 feature parameters are optimized into 14 feature parameters according to their correlation degree.The purpose is to solve the problem that the input parameters are blind,and the model that eliminates redundant and uncorrelated parameters reduces the complexity and improves the prediction efficiency.Also,the xgBoost model with optimal performance is interpreted using Partial dependence plot(PDP),Individual conditional expectation(ICE)and Shapley additive explanations(SHAP)methods for the model.to improve the credibility of the machine learning black box model and to identify the parameters that have a significant impact on the model,as well as to determine the impact of parameter changes on the prediction results for the selection of feature parameters.3.A continuous seismic intensity prediction model based on time-series data-driven model is proposed.In this paper,the prediction model is constructed based on the Long Short-Term Memory(LSTM)algorithm,which does not require pre-processing of ground motion data and directly uses the original records as input and makes predictions at each time step.The model is then trained,tested,and compared with traditional Pd prediction methods,and finally the March 2022 MJMA 7.3 earthquake event is selected as a case study to validate the model. |
