地球物理学进展 ›› 2018, Vol. 33 ›› Issue (5): 1844-1853.doi: 10.6038/pg2018BB0607

• 固体地球物理及空间物理学(大气、行星、地球动力学、重磁电及地震学、地热学) • 上一篇    下一篇

基于自编码网络的浅源和深源诱发型微震识别

杨德贺(),王秀英(),申旭辉,陈佳维,卫清   

  1. 中国地震局地壳应力研究所 (地壳动力学重点实验室),北京 100085
  • 收稿日期:2018-01-04 修回日期:2018-06-18 出版日期:2018-10-20 发布日期:2019-01-11
  • 通讯作者: 王秀英 E-mail:ydhmmm@163.com;xiuyw@sohu.com
  • 作者简介:杨德贺,男,1985年生,博士,助理研究员,从事地球物理观测数据挖掘等方面研究.
  • 基金资助:
    国家自然科学基金(41504071);中国地震局地壳应力研究所基本科研业务专项(ZDJ2015-07)

Discrimination of shallow and deep induced-microearthquakes based on autoencoder network

YANG De-he(),WANG Xiu-ying(),SHEN Xu-hui,CHEN Jia-wei,WEI Qing   

  1. Key Laboratory of Crustal Dynamics, Institute of Crustal Dynamics, China Earthquake Administration, Beijing 100085, China
  • Received:2018-01-04 Revised:2018-06-18 Online:2018-10-20 Published:2019-01-11
  • Contact: Xiu-ying WANG E-mail:ydhmmm@163.com;xiuyw@sohu.com

摘要:

地震波形传播的复杂多变性,导致传统互相关分析方法难以识别诱发型微震事件的深度类型.本文基于微震波形的时域、频域及时频域特征,利用自编码网络 (SAE)构造具有可鉴别性的特征空间,提升对深源和浅源诱发型微震事件的分类精度.首先,针对440个诱发型微震事件,构建了大小为40的特征空间;其次,利用遗传算法 (GA)和关联规则特征选择方法 (CFS)对特征空间进行初步筛选,得到特征重要性程度较强的谱矩心和线性度,通过分类验证了谱矩心与震源深度有强相关性;然后,将筛选结果输入到自编码网络,采用基于无监督学习的方法获得新的特征空间;最后,利用逻辑回归 (LR)对新特征空间进行交叉验证分类.与利用初步筛选的特征结果进行分类相比,利用4层的自编码网络模型对40特征进行交叉验证分类,所得正确率(Accuracy)和接收者操作特征曲线(ROC)曲线下方的面积(AUC)分别从84.5%提高到90.91%及84.31%提高到87.14%,结果表明自编码网络提高了分类模型对低能量诱发型微震事件的识别精度.

关键词: 微震深度识别, 自编码网络, 特征空间, 谱矩心

Abstract:

Due to the variability and complexity of propagation for seismic wave, it is difficult to identify the depth of induced micro-earthquakes using the traditional method based on cross-correlation. In this paper, we utilize Autoencoder Network to construct the feature space that has distinctive competence, based on time, frequency and time-frequency domain from micro-seismic waveform, which improves the accuracy of classification of micro-earthquakes. For a total of 440 events, we extract 40 features, filter them using Genetic Algorithm (GA) and Correlation-based Feature Selection. Spectral centroid and the degree of rectilinearity are important for classification. We find that there is a strong correlation between spectral centroid and depth of the source. Then, when those features are input into Autoencoder Network, we can acquire new feature space through unsupervised learning. Using Autoencoder Network with 4 layers based on cross validation, compared to use Logistic Regression (LR) in cross validation, the classification accuracy and area under the curve of ROC (AUC) based the new feature space are improved from 84.5% to 90.91%, 84.31% to 87.14% respectively, compared with the classification based on 40 features. Autoencoder Network improves the recognition accuracy of low energy induced micro-seismic.

Key words: Key Laboratory of Crustal Dynamics, Institute of Crustal Dynamics, China Earthquake Administration, Beijing 100085, China

中图分类号: