利用深度学习与物理特征预测仪器地震烈度  

Predicting instrumental seismic intensity using deep learning and physical features

作　　者：郑周 林彬华 金星[1,2,3] 于伟恒 李军[2] 韦永祥[2] 王士成[2] 李水龙 周施文 ZHENG Zhou;LIN BinHua;JIN Xing;YU WeiHeng;LI Jun;WEI YongXiang;WANG ShiCheng;LI ShuiLong;ZHOU ShiWen(Institute of Engineering Mechanics,China Earthquake Administration,Harbin 150088,China;Fujian Earthquake Agency,Fuzhou 350003,China;Key Laboratory of Earthquake Engineering and Engineering Vibration of China Earthquake Administration,Harbin 150088,China)

机构地区：[1]中国地震局工程力学研究所,哈尔滨150088 [2]福建省地震局,福州350003 [3]中国地震局地震工程与工程振动重点实验室,哈尔滨150088

出　　处：《地球物理学报》2024年第7期2712-2728,共17页Chinese Journal of Geophysics

基　　金：国家自然科学基金(42104062);中国地震局地震科技星火计划(XH23024A)联合资助。

摘　　要：快速准确地预测记录台站的仪器地震烈度,并在地震预警过程中向潜在用户提供及时和准确的警报至关重要.本文提出了一种结合深度学习和物理参数特征的算法,用于更可靠地预测仪器地震烈度.收集并处理了2001—2021年日本K-NET和KiK-net强震仪台网记录的3386次地震事件作为研究样本.对这些地震事件进行截取、基线校正、质量筛选等预处理,共得到25714条三通道地震波形.使用18000条地震记录(90%训练,10%验证)构建了一个窗长为3秒的仪器地震烈度预测模型(CNN-PP),并对7714条地震记录进行了测试.结果表明,CNN-PP模型在预测仪器地震烈度方面优于传统的单一特征参数方法.此外,离线震例测试结果显示,CNN-PP模型的报警成功率达到95.03%,没有出现误报情况,为解决地震预警仪器地震烈度的测定提供了一种潜在方法.Rapidly and accurately predicting instrumental seismic intensity at recording sites is essential for providing timely and reliable alerts to potential users during earthquake early warning(EEW).This paper proposes an algorithm that combines deep learning with physics features to predict instrumental seismic intensity more reliably.A total of 3386 seismic events recorded by the K-NET and KiK-net strong-motion seismograph networks in Japan from 2001 to 2021 were collected and processed as research samples.25714 three-channel seismic waveform was obtained after pre-processing of these seismic events,which included interception,baseline correction and quality screening.Using 18000 seismic records(90%for training and 10%for validation),a 3 s window length instrumental seismic intensity prediction model(CNN-PP)was constructed,and 7714 seismic records were tested.The results show that the CNN-PP model outperforms the traditional single feature-parameter model in predicting instrumental seismic intensity.In addition,the offline earthquake case test results show that the CNN-PP model's successful alarm rate reaches 95.03%without false alarms,which provides a potential method for solving the problem of EEW instrumental seismic intensity determination.

关 键 词：地震预警 仪器地震烈度 深度学习 卷积神经网络 物理参数特征 

分 类 号：P315[天文地球—地震学]