基于多尺度注意残差网络的地震波形分类研究

Seismic waveform classification based on a multiscale attention residual network

摘要:

选用2010年2月—2016年12月发生在北京顺义及河北三河等首都圈邻近区域的117个地震事件(包括54个天然地震事件和63个非天然地震事件——爆炸事件)作为研究对象,利用文章所提出的多尺度注意残差网络对其中的天然地震事件和爆炸事件波形进行二分类。首先,对原始地震波形进行简单预处理并截取成相同长度的地震时序数据,直接将其作为网络模型的输入;其次,选用含有残差模块的深度神经网络作为基础网络,利用深度神经网络对特征的自动提取能力,省略了传统波形分类需要提前提取时域波形的特征作为分类算法输入的步骤;然后,融合通道注意力机制(ECA)并对其进行改进,将空间维度的信息融入通道信息,优化了网络对关键信息的关注,更好地聚焦重要特征;最后,使用空间金字塔池化代替最大池化进行多尺度特征融合,得到更多的特征信息,构成多尺度注意残差网络。实验结果表明,最高分类准确率为97.11%,平均分类准确率为96.53%,证明了多尺度注意残差网络在地震波形分类任务中的有效性,为震源类型识别工作提供了一种新的方法。

Abstract:

A total of 117 seismic events (54 natural earthquakes and 63 explosions) that occurred in the Capital Circle Region (Shunyi, Beijing, Sanhe, and Hebei) from February 2010 to December 2016 were selected in this paper. The multiscale attention residual network was proposed and used to classify the waveforms of earthquakes and explosions. The original seismic waveform was simply preprocessed and intercepted into seismic time series data with the same length, which was directly used as the input of the network model. Then, the deep neural network with the residual module was selected as the basic network. The step of advanced extraction of time-domain waveform features as the input of classification algorithm in traditional waveform classification can be omitted by using the automatic feature extraction ability of the deep neural network. Next, the efficient channel attention mechanism was integrated and improved, after which information from the spatial dimension was integrated into the channel information, thus optimizing the network's attention to key information and resulting in better concentration on essential features. Finally, the multiscale feature fusion was performed using spatial pyramid pooling instead of maximum pooling to obtain more feature information. Ultimately, a multiscale attention residual network was formed. Experimental results show that the highest classification accuracy of the multiscale attention residual network is 97.11%, and the average classification accuracy is 96.53%. The results demonstrate the effectiveness of this approach in seismic waveform classification and provide a new optional approach for seismic-source type identification.