Acta Scientiarum Naturalium Universitatis Pekinensis ›› 2019, Vol. 55 ›› Issue (3): 451-460.DOI: 10.13209/j.0479-8023.2018.036

Previous Articles     Next Articles

Using Artificial Intelligence to Pick P-Wave First-Arrival of the Microseisms: Taking the Aftershock Sequence of Wenchuan Earthquake as an Example

CAI Zhenyu, GE Zengxi   

  1. School of Earth and Space Sciences, Peking University, Beijing 100871
  • Received:2018-05-18 Revised:2018-06-24 Online:2019-05-20 Published:2019-05-20
  • Contact: GE Zengxi,E-mail:zge(at)


蔡振宇, 盖增喜   

  1. 北京大学地球与空间科学学院, 北京 100871
  • 通讯作者: 盖增喜,E-mail:zge(at)
  • 基金资助:


In order to accurately and quickly pick up P-wave first-arrival of a large number of seismic events, deep learning method is introduced into the micro seismic P-wave first-arrival picking problem. The structure of convolution neural network is adjusted to apply to the characteristics of the seismic waveform data and first-arrival picking problem. The algorithm takes a 10s-window three-component seismic waveform data as input instead of scanning the continuous waveform. So the running time is far less than traditional methods such as STA/LTA and template matching. The algorithm is applied to aftershocks of 2008 Wenchuan earthquake in July and August, using 7467 manual picked first-arrival data as training dataset. Among the 1867 testing data, 98.9% of the P arrival times picked using this algorithm have an error less than 0.5 s compare to the results picked manually. This method can still maintain good pick-up capability under the condition of low signal-to-noise ratio.

Key words: artificial intelligence, machine learning, deep learning, wavelet transform, first-arrival picking


为了准确而迅速地拾取大量地震事件的P波初至, 将深度学习方法引入微地震P波初至到时拾取研究中, 对卷积神经网络的结构进行改造, 以便适应地震波形数据的特点 P波初至拾取的要求。该算法只需要输入10 s窗口的三分量地震波形数据, 就可以自动地判定P波初至时刻, 无需扫描连续波形, 运算时间远远小于长短窗、模板匹配等传统方法。使用该算法训练汶川地震主震后2008年7—8月7467条人工拾取的余震P波初至到时, 将得到的模型对测试集中 1867条数据的计算结果与人工拾取结果对比, 误差小于0.5 s者占比达到98.9%。在低信噪比条件下, 该方法仍能保持较好的拾取能力。

关键词: 人工智能, 机器学习, 深度学习, 小波变换, 初至拾取