一种基于PPNet的地震直达P波到时拾取方法

A picking method of direct seismic P-wave arrival time based on PPNet

摘要:

针对现有地震直达P波到时拾取网络精度低、误差大等问题,结合UNet++编码、解码器,融入特征过滤器设计一种具有地震震相特征分析与融合能力的轻量级P波到时拾取网络PPNet,实现对地震P波的高精度、低误差拾取。首先,该网络在编码器模块采用大卷积核、低通道数的卷积层,对输入的地震信号进行深度特征提取;其次,在解码器模块的特征还原过程中加入特征融合机制,补全特征信息,避免序列特征污染问题;最后,仅对编码器后三个下采样模块添加特征过滤器,深入挖掘特征序列,通过细化P波到时特征,提升到时拾取精度。实验结果表明,提出的网络在0.1 s、0.2 s、0.3 s误差阈值下P波拾取率分别为80.73%、94.01%、97.81%,平均绝对误差0.078 s,均方误差0.021,与现有P波拾取传统方法和深度学习算法相比性能更优。

Abstract:

To overcome the low accuracies and high error rates of existing P-wave arrival time picking networks, we developed a P-wave pyramid network (PPNet). This lightweight network was designed for high-precision and low-error picking of seismic P-waves by combining the UNet++ codec with a feature filter. Remarkably, the resulting PPNet could analyze and combine seismic phase characteristics. First, a convolution layer with a large convolution kernel and low channel number was incorporated into the encoder module for deep feature extraction from input seismic signals. Subsequently, a feature fusion mechanism was introduced into the feature restoration process of the decoder module to complement the feature information and prevent sequence feature contamination. Finally, feature filters were selectively applied to the final three downsampling modules of the encoder for deep feature sequence exploration, consequently refining the P-wave arrival features and improving pickup accuracy. Experimental results revealed that the P-wave pickup rates of the proposed network reached 80.73%, 94.01%, and 97.81% under error thresholds of 0.1, 0.2, and 0.3 s, respectively, with an average absolute error of 0.078 s and a mean square error of 0.021. Thus, the proposed network outperformed traditional P-wave pickup methods and deep-learning algorithms.