Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (6): 113-124.DOI: 10.3778/j.issn.1002-8331.2109-0428

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Seismic P-Wave First-Arrival Picking Model Based on Spatiotemporal Attention Mechanism

LI Yu, HAN Xiaohong, ZHANG Ling, ZHANG Haixuan, LI Gang   

  1. 1.College of Data Science, Taiyuan University of Technology, Taiyuan 030000, China
    2.College of Software, Taiyuan University of Technology, Taiyuan 030000, China
    3.College of Information and Computer, Taiyuan University of Technology, Taiyuan 030000, China
  • Online:2023-03-15 Published:2023-03-15



  1. 1.太原理工大学 大数据学院,太原 030000
    2.太原理工大学 软件学院,太原 030000
    3.太原理工大学 信息与计算机学院,太原 030000

Abstract: Aiming at the problems of low accuracy and poor robustness of the existing earthquake first-arrival picking algorithm, a seismic P-wave arrival picking network based on deep learning is designed. This network is encoder-decoder structure, which can identify seismic signal sequence point by point. The encoder uses multi-scale feature extractor for feature extraction and fusion of input data to improve feature utilization ratio. The multi-scale residual structure is used to deeply mine the hidden feature information in the data to improve the nonlinear fitting ability of the model. Then, the spatiotemporal attention mechanism is added to the decoder to improve the network’s perception of the first-arrival features. Finally, a deep coding feature fusion module is proposed to effectively avoid the pollution of feature sequence while ensuring the integrity of features. The experimental results show that under the three error thresholds of 0.1 s, 0.2 s and 0.3 s, the picking hit rate of the proposed network are 75.04%, 94.6% and 97.37%, respectively, the mean absolute error and mean square error are 0.092 s and 0.036. Compared with the existing traditional and deep learning first-arrival picking methods, it has higher P-wave first-arrival picking accuracy.

Key words: phase arrive picking, deep learning, sequence processing, spatiotemporal attention, feature fusion

摘要: 针对现有地震到时拾取算法精度较低、鲁棒性较差等问题,设计了一种基于深度学习的地震P波到时拾取网络,该网络为编解码结构,可实现地震波形序列的逐点预测。网络编码器对输入数据进行多尺度特征提取与融合,提高特征利用率;利用多尺度残差结构深度挖掘数据中隐藏特征信息,提升模型非线性拟合能力;在解码网络中加入时空注意力机制,提高网络对到时特征的感知能力;提出深层编码特征融合模块,在保证特征完整性的同时有效避免融合特征过程中出现的特征序列污染问题。实验结果表明,提出的网络在0.1?s、0.2?s、0.3?s三个误差阈值下,拾取命中率分别为75.04%、94.6%、97.37%,平均绝对误差和均方误差为0.092?s、0.036,相比现有传统方法与深度学习到时拾取方法,具有更高的P波到时拾取精度。

关键词: 震相到时拾取, 深度学习, 序列处理, 时空注意力, 特征融合