基于随机子空间和Ada Boost集成学习的地震事件性质辨识研究  

Identification and classification of earthquake types based on Cov-SSI and Ada Boost ensemble learning algorithm

作　　者：吴涛[1,2,3] 庞聪 江勇[1,2,3] 丁炜 廖成旺 WU Tao;PANG Cong;JIANG Yong;DING Wei;LIAO ChengWang(Institute of Seismology,CEA,Wuhan 430071,China;Key Laboratory of Earthquake Geodesy,CEA,Wuhan 430071,China;Seismological Bureau of Hubei Province,Wuhan 430071,China)

机构地区：[1]中国地震局地震研究所,武汉430071 [2]中国地震局地震大地测量重点实验室,武汉430071 [3]湖北省地震局,武汉430071

出　　处：《地球物理学进展》2022年第3期981-988,共8页Progress in Geophysics

基　　金：国家自然科学基金(41174053);湖北省自然科学基金(ZRMS2020000813);中国地震局地震研究所和应急管理部国家自然灾害防治研究院基本科研业务费专项资助项目(IS201856290,IS2018126178,IS201726156);中国综合地球物理场仪器研发专项(Y201707)联合资助。

摘　　要：在常规强震动监测工作中,常记录到人工爆破、塌陷等非天然地震动事件信号,以及电磁环境干扰、设备故障等引起的异常信号,这对地震预警预报工作中的地震精确识别极为不利.设计一种新型天然地震事件特征提取及性质辨识方法,克服以往全人工或半人工识别地震的不足之处,引入协方差驱动的随机子空间(Cov-SSI)方法,经过Hankel矩阵构造、奇异值分解、系统定阶、特征值分解及系统状态矩阵方程求解等步骤,从目标记录中提取出系统阶次、系统协方差矩阵奇异值累加和、有效奇异值个数、模态频率及系统特征根实部等特征参数,应用决策树为弱学习器的AdaBoost集成学习算法对目标事件进行性质辨识.实验数据选取九寨沟M7.0级天然地震事件强震动记录与干扰数据,结果表明:整体辨识准确率达到90%以上;在同等条件下,辨识准确率、召回率、F-Measure等性能指标皆优于SVM、KNN、DAC等传统机器学习方法;该法在天然地震事件性质准确辨识领域,有一定的参考价值.In conventional strong vibration monitoring work,signals of non-natural ground shaking events such as artificial blasting and collapse are often recorded,as well as anomalous signals caused by electromagnetic environment disturbances and equipment failures,which are extremely detrimental to the accurate identification of earthquakes in earthquake early warning and forecasting work.A new natural seismic event feature extraction and property identification method is designed to overcome the shortcomings of previous fully or semi-artificially identified earthquakes by introducing the covariance-driven stochastic subspace(Cov-SSI)method,which extracts the system order,system state matrix equation solution from the target records through the steps of Hankel matrix construction,singular value decomposition,system order fixation,eigenvalue decomposition and system After the steps of Hankel matrix construction,singular value decomposition,system ranking,eigenvalue decomposition and system state matrix equation solution,we extracted the characteristic parameters such as system order,system covariance matrix singular value accumulation sum,number of effective singular values,modal frequency and system eigenroot real part from the target records,and identified the properties of the target events based on the AdaBoost integrated learning algorithm which applied decision tree as a weak learner.The results show that the overall recognition accuracy of the method is over 90%,and the performance indexes such as recognition accuracy,recall rate and F-Measure are better than those of SVM,KNN,DAC and other traditional machine learning methods under the same conditions.It has some application value in the field of natural seismic event identification.

关 键 词：地震事件分类 随机子空间 集成学习 模态参数 决策树 

分 类 号：P315[天文地球—地震学]