基于深度学习卷积神经网络的地震波形自动分类与识别

发展高效、高精度、普适性强的自动波形拾取算法在地震大数据时代背景下显得越来越重要.波形自动拾取算法的主要挑战来自如何适应不同区域的不同类型地震事件的分类与筛选.本文针对地震事件-噪音分类这一问题, 使用13839个汶川地震余震事件建立数据集, 应用深度学习卷积神经网络(CNN)方法进行训练, 并用8900个新的汶川余震事件作为检测数据集, 其训练和检测准确率均达到95%以上.在对连续波形的检测中, CNN方法在精度和召回率上优于STA/LTA和Fbpicker传统方法, 并能找出大量人工挑选极易遗漏的微震事件.最后, 我们应用训练好的最优模型对选自全国台网的441个台站8天的连续波形数据进行了识别、到时挑取及与参考地震目录关联, CNN检出7016段波形, 用自动挑选算法拾取到1380对P, S到时, 并与540个地震目录事件成功关联, 对1级以上事件总体识别准确率为54%, 二级以上为80%, 证明了CNN模型具有泛化能力, 初步展示了CNN在发展兼具效率、精度、普适性算法, 实时地震监测等应用上具有巨大潜力.

Waveform classification and seismic recognition by convolution neural network

The development of efficient, high-precision, and universal automatic waveform pick-up algorithm is more and more important in the background of earthquake big data. The main challenge comes from how to adapt to the classification of different types of seismic events in different regions. In this paper, according to the seismic event-noise classification problem, a convolutional neural network method was used to train the dataset based on 13839 Wenchuan earthquake aftershocks and 8900 new Wenchuan aftershock events were used as the test data set. The training and detection accuracy rates were both over 95%. In the detection of continuous waveforms, the CNN method is superior to the traditional methods of STA/LTA and Fbpicker in precision and recall rate, and can find a large number of manually selected microseismic events that are easily missed. Finally, we use the trained optimal model to identify 8-day continuous waveform data from 441 stations nationwide. CNN detects 7016 waveforms, then we pick up 1380 pairs of P and S arrival times using an automatic picking algorithm, finally the pick-ups were successfully associated with 540 earthquake catalog events. The overall recognition accuracy of events above magnitude 1 was 54% and 80% above magnitude 2, while in some areas such as Sichuan and Xinjiang the detection rate is higher. It is shown that CNN neural network has broad application prospects in the real-time earthquake detection and location.