甘露^1,2,,
吴庆举^1,,,
黄清华²,
唐荣江¹
1. 中国地震局地球物理研究所, 北京 100081
2. 北京大学地球与空间科学学院, 北京 100081

基金项目: 宁夏回族自治区重点研发计划重大(重点)项目(2018BFG02011), 国家自然科学基金项目(41674094)资助


详细信息

作者简介: 甘露, 女, 1992年生, 博士研究生, 主要从事深部地球物理成像研究.E-mail: ganlu@pku.edu.cn

通讯作者: 吴庆举, 男, 理学博士, 研究员, 博士生导师.现主要从事壳幔结构探测与研究.E-mail: wuqj@cea-igp.ac.cn

中图分类号: P315

收稿日期:2020-04-17
修回日期:2021-04-29
上线日期:2021-07-10

Quick selection of receiver function based on convolutional neural network

GAN Lu^1,2,,
WU QingJu^1,,,
HUANG QingHua²,
TANG RongJiang¹
1. Institute of Geophysics, China Earthquake Administration, Beijing 100081, China
2. School of Earth and Space Sciences, Peking University, Beijing 100871, China


More Information

Corresponding author: WU QingJu,E-mail:wuqj@cea-igp.ac.cn

MSC: P315

--> Received Date: 17 April 2020
Revised Date: 29 April 2021
Available Online: 10 July 2021


摘要
摘要:地震三分量波形数据中提取的接收函数受震源复杂性及随机噪声等因素的影响，往往出现一些波形异常现象，需要在资料解释前予以剔除.当接收函数数量较多时，人为挑选质量合格的接收函数将耗费大量的时间.为了高效的挑选高质量的接收函数，本文利用深度学习卷积神经网络（Convolutional Neural Network，简称CNN）方法来对接收函数的质量进行判断.我们使用华北地区和阿巴嘎地区地震观测台站接收的9833个不同的地震事件建立训练集，并用1521个新的地震事件作为检测数据集，得到的训练集的准确率和召回率均达到99%以上，测试集的准确率和召回率分别达到95.3%和92.4%.我们还使用了训练集和测试集以外的数据进行验证，并对比了不同CNN评估结果所对应的波形图，实验证明评估结果与实际的接收函数波形对应良好.此外，对于某些台站的接收函数，可能存在如下问题：波形虽然具有很好的一致性，但由于不符合常规意义下质量好的标准导致CNN无法识别.为解决该问题，本文首先对一定方位角和震中距范围内的接收函数相互求取二范数，再对二范数较低的结果进行统计，并与CNN的挑选结果进行对比，挑选出合格的数据.
关键词: 接收函数/
卷积神经网络/
快速挑选

Abstract:The receiver function extracted from seismic three-component waveform data is often affected by source complexity and random noise, and some waveform anomalies often appear, which need to be eliminated before data interpretation. When there are many receiver functions, it will take a lot of time to select the normal receiver function artificially. In order to select high-quality receiver function efficiently, this paper used convolutional neural network (CNN) to judge the quality of receiver function. We used 9833 different earthquake events received in North China and Abaga area for training, and 1521 earthquake events were used for validation. The accuracy and recall rate of the training datasets are above 99%, and the accuracy and the recall rate of the test datasets are 95.3% and 92.4%, respectively. We also used the CNN to evaluate other data outside the training and the test datasets to get different types of waveforms, which proves that the evaluation results correspond well with the expected receiver function waveform. In addition, however, CNN cannot efficiently evaluate some receiver function which have good consistency but do not meet conventional standard of good quality. In order to solve this problem, this paper calculates the 2-norm of the receiver function within a certain azimuth and epicentral distance, and makes statistics on the results with lower 2-norm, and compares the results with the CNN output to select final qualified data.
Key words:Receiver function/
Convolution neural network/
Quick selection



PDF全文下载地址:

http://www.geophy.cn/data/article/export-pdf?id=7911b19e-790a-4ec8-8dcb-5c3ddb0c4bd8