| 基于卷积神经网络的地震预警震级持续估算方法研究 |
| |
| 引用本文： | 江炳根,马强,陶冬旺.基于卷积神经网络的地震预警震级持续估算方法研究[J].世界地震工程,2022,38(1):213-228. |
| |
| 作者姓名： | 江炳根 马强 陶冬旺 |
| |
| 作者单位： | 1 中国地震局工程力学研究所, 黑龙江 哈尔滨 150080;2 中国地震局地震工程与工程振动重点实验室, 黑龙江 哈尔滨 150080 |
| |
| 基金项目： | 国家重点研发计划课题(2017YFC1500802);中国地震局工程力学研究所基本科研业务费专项资金(2016A03,2014B08);国家自然科学基金高铁联合基金资助项目(U1534202)。 |
| |
| 摘 要： | 为提高地震预警震级快速持续估算结果的准确性，本文构建了基于多种地震动特征参数的卷积神经网络震级估算CNN-M模型。该模型基于日本KiK-net强震动观测记录，利用其P波触发后3~10s时间窗内的幅值参数、周期参数、烈度参数、信噪比参数共11种地震波特征参数以及震中距参数作为输入。本文所建立的CNN-M模型随着地震发生后时间窗的推移可持续进行震级估算。通过研究震中距效应和场地效应对模型的影响，其结果表明:在使用井下数据和震中距参数时，模型表现最佳，其震级估算结果标准差逐渐降低，3 s时间窗标准差为0.336，10 s时间窗时降低至0.251，较之震级估算"Pd方法"，估算结果准确性有了较大提高。在2014年11月22日长野Mj6.7级地震震例分析中，各时间窗内都有多个台站可准确估算出实际震级。在2021年3月20日宫城Mj6.9级地震的近场台站持续震级估算测试和实时震级估算测试中，CNN-M模型展现出了较高的准确性和稳定性，震级均值误差较小。以上研究表明:本文构建的CNN-M模型具有稳定可靠的预警震级持续估算能力，可为"国家地震烈度速报与预警工程"项目建设提供震级估算方法参考。
|
| 关 键 词： | 地震预警 P波 卷积神经网络 震级估算 长野地震 宫城地震 |
Continuous estimation of earthquake early warning magnitude based on Convolutional Neural Network |
| |
| JIANG Binggen,MA Qiang,TAO Dongwang.Continuous estimation of earthquake early warning magnitude based on Convolutional Neural Network[J].World Information On Earthquake Engineering,2022,38(1):213-228. |
| |
| Authors: | JIANG Binggen MA Qiang TAO Dongwang |
| |
| Affiliation: | 1 Institute of Engineering Mechanics, China Earthquake Administration, Harbin 150080, China;2 Key Laboratory of Earthquake Engineering and Engineering Vibration of China Earthquake Administration, Harbin 150080, China |
| |
| Abstract: | In order to improve the accuracy of the earthquake early warning magnitude estimation，this paper constructs a convolutional neural network which named CNN-M model based on multi-parameter of seismic wave for magnitude estimation.This model is based on Japanese KiK-net strong motion observation records，using the amplitude parameters，period parameters，intensity parameters，and signal-to-noise ratio parameters within the time window of 3～10 s after the P wave is triggered，11 kinds of seismic wave characteristic parameters，as well as epicenter distance parameter are used as input. The CNN-M model established in this paper can continuously estimate the magnitude with the time window elapses after the earthquake occurs. By studying the influence of the epicentral distance effect and the site effect on the model，the results show that when using borehole data and epicentral distance parameters，the model performed the best. The standard deviation of the magnitude estimation results gradually decreased，the standard deviation of the 3 s time window was 0.336，and it was reduced to 0.251 at the 10s time window. Compared with the "Pd method" of magnitude estimation，the accuracy of the estimation results has been greatly improved. In the analysis of the Nagano Mj6.7 earthquake on November 22，2014，there are multiple stations in each time window that can accurately estimate the actual magnitude. In the continuous magnitude estimation test and real-time magnitude estimation test of the near-field station the Miyagi Mj6.9 earthquake on March 20，2021， the CNN-M model showed high accuracy and stability，and the magnitude error was relatively high small. The above research shows that the CNN-M model constructed in this paper has a stable and reliable ability to continuously estimate the early warning magnitude，which can provide a reference for the construction of the "National Earthquake Intensity Quick Report and Early Warning Project" project. |
| |
| Keywords: | earthquake early warning P wave Convolutional Neural Networks magnitude estimation Nagano-ken earthquake Miyagi earthquake |
| 本文献已被 维普 万方数据 等数据库收录！ |
| 点击此处可从《世界地震工程》浏览原始摘要信息 |
|
点击此处可从《世界地震工程》下载全文 |
