地球物理学进展 ›› 2018, Vol. 33 ›› Issue (4): 1331-1338.doi: 10.6038/pg2018BB0326

• 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

地震和爆破事件源波形信号的卷积神经网络分类研究

陈润航,黄汉明(),柴慧敏   

  1. 广西师范大学计算机科学与信息工程学院,桂林 541004
  • 收稿日期:2017-10-10 修回日期:2018-07-11 出版日期:2018-08-20 发布日期:2018-09-13
  • 通讯作者: 黄汉明 E-mail:huanghm@gxnu.edu.cn
  • 作者简介:陈润航,男,1990年生,硕士研究生,主要研究方向为信号处理与模式识别.(E-mail: crunhang@163.com)
  • 基金资助:
    国家自然科学基金资助((41264001))

Study on the discrimination of seismic waveform signals between earthquake and explosion events by convolutional neural network

Run-hang CHEN,Han-ming HUANG(),Hui-min CHAI   

  1. College of Computer Science and Information Engineering, Guangxi Normal University, Guilin 541004, China
  • Received:2017-10-10 Revised:2018-07-11 Online:2018-08-20 Published:2018-09-13
  • Contact: Han-ming HUANG E-mail:huanghm@gxnu.edu.cn

摘要:

本文首先从震源波形中提取梅尔频率倒谱系数(MFCC)图,然后采用卷积神经网络(CNN)进行地震波形信号的震源类型—天然地震和爆破事件—分类识别.事件为首都圈及其附近的72个天然地震和101个人工爆破事件,用于提取梅尔频率倒谱系数图的波形信号为各观测台站波形3分量中的垂直分量波形.在各个事件的所有观测台站的垂直分量波形中,通过滑动窗口按同一准则去除被噪声淹没的部分台站波形,只选择留下未被噪声淹没的台站波形.每一个事件有107个观测台站,故有107份垂直分量波形,而不同事件被留下未被噪声淹没的波形则有几份至几十份不等.然后提取被留下未被噪声淹没的波形的梅尔频率倒谱系数图,以梅尔频率倒谱系数图作为CNN的输入,CNN的输出则为波形的震源类型(天然地震事件或爆破事件).若以单份波形为识别单元,采用五折交叉验证法进行测试,得到的平均准确率为95.78%.使用训练集中单份波形为识别单元,提取梅尔频率倒谱系数图,采用CNN训练出了天然地震事件与爆破事件波形分类器,一个事件在测试集中的多份波形信号通常不会都被正确识别,很可能有些波形被识别为天然地震事件,另一些波形被识别为爆破事件;这时,若识别单元改为事件,一个事件各台站的有效垂直分量波形中,超过一半的波形被识别为某一事件类型,则这个事件被归类为该事件类型,得到的正确识别率为97.1%.实验结果表明:卷积神经网络在天然地震事件与爆破事件的识别方面表现出色.这说明MFCC与卷积神经网络可以用于识别天然地震和爆破事件,尤其是深度学习更值得在地震信号处理方面做进一步的研究.

关键词: 天然地震事件, 爆破事件, 震源波形, 震源类型识别, 梅尔频率倒谱系数, 卷积神经网络

Abstract:

This paper firstly extracts Mel Frequency Cepstrum Coefficient (MFCC) map from seismic source waveform, and then applies the Convolutional Neural Network (CNN) to discriminate seismic waveform signals between earthquake and explosion events. The events are 72 earthquake and 101 man-made explosion events in the Beijing region and nearby, and the waveform signal used to extract MFCC map is the vertical component of the 3 components of the waveform in each observation station. For any selected waveform (vertical component of 3 components) of all observation stations of anyone event, whether the waveform being messed with noises is judged by the same criterion, only the waveform that is not messed with noise is selected. Otherwise the waveform is discarded. Although there are 107 observation stations for each event, certainly there are 107 vertical components, after noise-messing waveforms being discarded, there are maybe only several to dozens of vertical components remaining for one event. Then, the MFCC map of the waveform that is left without noise or with small noise are extracted, and the MFCC map are used as the input of the CNN, and the output of the CNN is the seismic source type of the waveform (earthquake or explosion). If a single waveform is taken as a recognition unit, and 5-fold cross validation test was adopted, the average recognition rate is 95.78%. Using the single waveform in the training set as the recognition unit, extracting MFCC map, adopting CNN training strategy, a classifier is formulated for discriminating earthquake and explosion events. In testing, multi waveform signals from the same event are not usually classified consistently for the same event, it is likely that some are classified as earthquake events, some are classified as explosion events. If the event is regarded as the recognition unit, for the same event, more than half of the waveforms are identified as a type of event, this event is classified as this type of event, and then the correct recognition rate is 97.1%. The experimental results show that the convolutional neural network exhibits excellent performance in earthquake and explosion event recognition. This shows that MFCC and convolutional neural networks can be used to identify earthquake and explosion events, and deep learning theory may also be worth further study in seismic signal processing.

Key words: earthquake event, explosion event, seismic source waveform, seismic source type recognition, Mel Frequency Cepstrum Coefficient (MFCC), Convolutional Neural Network (CNN)

中图分类号: