基于深层卷积神经网络的震级快速估算方法

当前地震预警中的震级估算方法是通过初至几秒地震波的特征参数与震级的经验关系来实现的, 这些特征参数依赖于人的经验和主观判断, 没有充分利用初至地震波中与震级相关的信息, 制约了震级估算效果.对此, 本文利用深层卷积神经网络(Deep Convolutional Neural Networks, CNN)直接从初至地震波中自动提取特征, 实现端到端的震级快速估算.CNN方法以单台站的初至竖向地震波作为主输入, 震中距、震源深度以及V_s30作为辅助输入, 震级作为输出.利用日本和智利的大量地表强震记录对CNN方法进行训练(98257条记录)、验证(31429条记录)和测试(40638条记录), 利用美国和新西兰的强震记录进行泛化性能测试(583条记录), 并与应用最为广泛的峰值位移Pd方法进行对比.结果表明, 当初至地震波时长为3s时, 在4~6.4级范围内, CNN方法估算震级的准确率是Pd方法的1.5倍, 在6.5~9级范围, CNN方法估算震级的准确率是Pd方法的1.2倍; 当初至地震波从3s增加到10s时, CNN方法能够随着地震波时长的增加不断提高估算震级的准确率, 并且始终高于Pd方法, 特别是对于4~6.4级地震, CNN方法在初至3s地震波时估算震级的准确率是Pd方法在初至10s地震波时的1.2倍; 随着地震波时长的增加, CNN方法对于震级饱和问题的改善效果优于Pd方法; CNN方法具有较好的泛化能力, 在训练数据集之外的区域, 比Pd方法估算震级更准确.相比于人为定义的特征参数, CNN方法从初至地震波中自动学习到了与震级更为相关的特征, 这些特征极大地改善了震级估算的准确性和时效性, 可以为地震预警系统提供更快速更准确的震级估算.

A fast magnitude estimation method based on deep convolutional neural networks

The magnitude estimation method in earthquake early warning is traditionally realized by the empirical relationship between the characteristic parameters based on the initial few seconds of arriving seismic waves and the magnitude. These characteristic parameters depend on human experience and subjective judgment, and do not make full use of the information related to magnitude in the initial seismic waves, which restricts the effect of magnitude estimation. To address this issue, this paper uses a deep convolutional neural network (CNN) to directly extract features from initial seismic waves to achieve end-to-end magnitude estimation. CNN uses the vertical component of initial seismic records from a single station as the main input, epicentral distance, focal depth, and Vs30 as auxiliary inputs, and magnitude as its output. The proposed CNN was trained (98, 257 records), verified (31, 429 records) and tested (40, 638 records) by using a large number of surface accelerograms from Japan and Chile, and the generalization performance of the proposed CNN was tested by accelerograms (583 records) from the United States and New Zealand, and the results were compared with the most widely used peak displacement Pd method. The comparison shows that when the duration of initial P wave is 3 s, the accuracy of magnitude estimation by the CNN is 1.5 times that of the Pd method for magnitude range of 4~6.4, and 1.2 times that of Pd method in the magnitude range of 6.5~9. When the duration of initial seismic wave increases from 3 s to 10 s, the CNN continuously improves the accuracy of magnitude estimation with the increasing duration of initial seismic waves, and it is always better than the Pd method. It should be noted that for the magnitude range of 4~6.4, the accuracy of magnitude estimation by the CNN using the initial 3 s seismic wave is 1.2 times that of the Pd method using the initial 10 s seismic wave. With the increase of seismic wave duration, CNN performs better than Pd in addressing the magnitude saturation issue. Compared with manually defined characteristic parameters, CNN automatically learns features that are closely related to magnitude from the initial seismic waves, and these features greatly improve the accuracy and timeliness for the magnitude estimation. The proposed approach can provide faster and more accurate magnitude estimation for earthquake early warning systems.