«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年08期

页码:

166-174

栏目:

新型计算应用系统

出版日期:

2024-08-10

文章信息/Info

Title:

Ensemble Learning Method for Earthquake Prediction Combining Electromagnetic and Acoustic Features

文章编号:

1673-629X(2024)08-0166-09

作者:

刘英杰; 黄嘉琦; 姜玉凤; 邵宇琪; 杨文韬; 于紫凝; 郑海永

中国海洋大学 电子工程学院,山东 青岛 266404

Author(s):

LIU Ying-jie; HUANG Jia-qi; JIANG Yu-feng; SHAO Yu-qi; YANG Wen-tao; YU Zi-ning; ZHENG Hai-yong

School of Electronic Engineering,Ocean University of China,Qingdao 266404,China

关键词:

地震预测; 机器学习; 集成学习; 特征融合; 数据驱动; 临震特性; 地震三要素

Keywords:

earthquake prediction; machine learning; ensemble learning; feature fusion; data driven; impending seismic characteristics; three elements of earthquake

分类号:

TP391

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0145

摘要:

地震是极具破坏性与不确定性的自然现象,在人们毫无察觉的情况下地震发生在人口稠密区时,将严重危害人们生命财产安全。 人们不断努力了解地震的物理特征和物理危害与环境之间的相互作用,以便在地震发生前发出适当的警报。 可靠的地震预测应包含对地震信号的分析,但是这些信号在地震发生前不明显;因此使用数据驱动机器学习的方法来分析这些信号与地震的联系并预测地震。 通过建立观测台网连续监测与地震发生相关的各种物理量或化学量,据此获取的地震前兆信息是地震预测的研究基础。 地震发生前,地球物理场发生显著变化,伴随电磁和地声等多种前兆信号,其中电磁和地声信号具有临震特性,是开展地震临震观测预测研究的重要数据来源;因此对地下的电磁扰动和地声信号进行实时监测,获取长期观测数据用于数据驱动机器学习方法预测地震。 该文基于 AETA 数据的临震模型预报,针对多分量地震监测预测系统(Acoustic and Electromagnetic Testing All in one system,AETA)在川滇地区记录的电磁和地声数据,提取时域和频域特征,采用基于随机森林算法、轻量级梯度提升决策树和极度随机树的集成学习方法共同预测该区域的发震情况,选取发震概率最大的子区域中心位置作为震中预测结果,进一步训练 LightGBM 回归模型以预测此子区域的震级,按周对地震三要素进行预测。 实验结果表明,该方法在川滇地区地震风险预测上,准确率可达 0. 64,震级预测的平均误差为 0. 38,最小误差为 0. 00,具有良好的预测效果。

Abstract:

Earthquakes are highly destructive and unpredictable natural phenomena. When they strike densely populated areas without warning,they pose a severe threat to human life and property. Efforts are ongoing to understand the physical characteristics of earthquakes and the interplay between physical hazards and the environment,to issue timely warnings before earthquakes occur. Reliable earthquake prediction should involve the analysis of seismic signals,but these signals are often not apparent before an earthquake;hence,data-driven machine learning methods are employed to analyze the relationship between these signals and earthquakes and predict earthquakes. A network of observation stations is established to continuously monitor various physical and chemical quantities associated with earthquake occurrence,and the precursor information obtained forms the basis of earthquake prediction. Before an earthquake,the geophysical field undergoes significant changes,accompanied by a variety of precursor signals such as electromagnetic and geo-acoustic signals. These e-lectromagnetic and geo - acoustic signals, indicative of an imminent earthquake, are crucial data sources for the study of imminent earthquake observation and prediction. Therefore,real-time monitoring of underground electromagnetic disturbances and ground acoustic signals is conducted to obtain long-term observation data for data-driven machine-learning methods to predict earthquakes. We analyze the Electromagnetic and Acoustic data recorded by the Acoustic and Electromagnetic Testing All in one system (AETA) in Sichuan and Yunnan. Based on AETA data,we extract the characteristics of the time and frequency domains. An ensemble learning method based on the random forest algorithm,lightweight gradient boosting decision tree,and extremely random tree is used to predict earthquakes in this region. The center location of the subregion with the highest probability of earthquake occurrence is selected as the predicted epicenter,and the LightGBM regression model is further trained to predict the magnitude of this subregion,with the three elements of earthquakes being predicted weekly. The experimental results show that the accuracy of the proposed method can reach 0. 64,the average error of earthquake magnitude prediction is 0. 38,and the minimum error is 0.00,demonstrating the effectiveness of the proposed prediction meth-od.

常用功能

更新日期/Last Update: 2024-08-10