基于深度卷积神经网络的地震震相拾取方法研究

地震震相拾取是地震数据自动处理的首要环节,包括了信号检测、到时估计和震相识别等过程,震相拾取的准确性直接影响到后续事件关联处理的性能,影响观测报告的质量.为了提高震相拾取的准确性,进而提高观测报告质量,本文采用深度卷积神经网络方法来解决震相拾取问题,构建了多任务卷积神经网络模型,设计了分类和回归的联合损失函数,定义了基于加权的分类损失函数,以三分量地震台站的波形数据作为输入,同时实现对震相的检测识别和到时的精确估计.利用美国南加州地震台网的200万条震相和噪声数据对模型进行训练、验证和测试,对于测试集中直达波P、S震相识别的查全率达到98%以上,到时估计的标准偏差分别为0.067s,0.082s.利用迁移学习和数据增强,将模型用于对我国东北地区台网的6个台站13000条数据的训练、验证和测试中,对该数据集P、S震相查全率分别达到91.21%、85.65%.基于迁移训练后的模型,设计了用于连续数据的震相拾取方法,利用连续的地震数据对该算法进行了实际应用测试,并与国家数据中心和中国地震局的观测报告进行比对,该方法的震相检测识别率平均可达84.5%,验证了该方法在实际应用中的有效性.本文所提出的方法展示了深度神经网络在地震震相拾取中的优异性能,为地震震相和事件的检测识别提供了新的思路.     

基于深度学习的地震事件分类识别 ——以山东地震台网记录为例

为实现天然地震与爆破、塌陷事件类型的快速高效识别,文中应用深度学习技术中的卷积神经网络模型,设计了基于单个事件单个台站波形记录的深度学习训练模块和基于单个事件多个台站波形记录的实时测试模块。以每个事件P波到时最早的5个台站记录到的原始三分向波形为输入,分别采用目前主流的Alex Net、VGG16、VGG19、Goog Le Net 4种卷积神经网络结构进行学习训练,结果显示各类卷积神经网络结构对训练集与测试集的识别准确率均达93%以上,且各个网络在训练过程中的训练集与测试集的准确率及代价函数的走势曲线基本一致。其中,Alex Net网络结构的识别准确率最高,测试集为98.51%,且未发生过拟合现象; VGG16、VGG19网络结构的准确率次之;Goog Le Net网络结构的识别准确率相对较低。为检验深度学习卷积神经网络在数字地震台网实时运行过程中的事件判别效能,选取训练好的Alex Net卷积神经网络开展基于单个事件多个台站波形记录的事件类型判定检验。最终结果显示,在山东台网实时触发的110个M≥0.7事件中,共有89个事件的类型被准确识别,准确率约为80.9%。具体到各个类型事件中,天然地震的准确率约为74.6%;爆破的准确率约为90.9%;塌陷事件的准确率为100%。若删除其中由于波形失真而造成的类型识别错误事件,则天然地震的识别准确率将提高至91.4%,而所有事件的整体识别准确率也将由80.9%提高至91.7%,与目前地震台网日常工作中人工判定的识别准确率基本相当。这表明,深度学习技术可以快速高效地实现天然地震与爆破、塌陷的事件类型识别。     

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.     

基于U形卷积神经网络的震相识别与到时拾取方法研究

精确获取震相到时是地震定位和地震走时成像等研究的重要基础.近年来,随着地震台站的不断加密,地震台网监测到的地震数量成倍增长,发展快速、准确、适用性强的震相到时自动拾取算法是地震行业的迫切需求.本文在前人工作基础上,发展了Pg、Sg震相自动识别与到时拾取的U网络算法（Unet_cea）,使用汶川余震和首都圈地震台网记录的89344个不同震级、不同信噪比的样本进行训练和测试.研究表明,U网络能够较好地识别Pg、Sg震相类型和拾取到时,Pg、Sg震相的正确识别率分别为81%和79.1%,与人工标注到时的均方根误差分别为0.41 s和0.54 s.U网络在命中率、均方根误差等性能指标上均明显优于STA/LTA和峰度分析自动拾取方法.研究获得的最优模型可以为区域地震台网的自动处理提供辅助.     

基于深度学习的地震检测模型在区域台网的泛化性研究

将识别地震的深度学习算法PhaseNet应用于四川台网和首都圈台网, 对该模型的泛化能力进行了测试和评估。 首先利用2010年1月至2018年10月首都圈台网199个地震台站记录的29328个事件(M_L0~M_L4)所对应的126761段事件波形, 以及 2019年4—9月四川及邻省部分台网227个地震台站记录的16595个事件(M_L0~M_L6.0)所对应的120233段事件波形分别建立了SC和CA测试数据集, 并用预训练好的PhaseNet模型进行P、 S震相自动识别和到时拾取, 并将拾取结果与人工拾取结果在不同误差阈值下进行对比。 测试结果表明, PhaseNet在两个数据集上具有良好的震相检测能力(误差阈值为0.5 s), 其P、 S震相检测的F₁值都超过0.75, 具有比较稳定的准确拾取P波到时能力(误差阈值0.1 s), 其检测F₁值均超过0.6, 而S波到时拾取的F₁值分别为0.33(SC)和0.53(CA)。 进一步分析了测试结果与震中距、 震级、 信噪比、 台站所处地域之间的关系, 为下一步继续训练更优化的模型指明了方向。 研究结果表明, PhaseNet算法在区域台网地震自动检测和到时拾取方面有很大的应用潜力和提升空间, 可以为区域台网的自动编目工作提供辅助。     

广东地区重复地震识别及其在台网定位评价中的应用

利用广东数字地震台网2010年1月至2013年10月的地震震相到时和波形资料,首先对地震目录进行完整性分析,选出在MC震级0级以上的3 969个地震进行波形互相关分析。设定在至少三个台站记录的垂直分量波形相关系数大于0.8的两个事件为重复地震对,共识别出广东地区的重复地震1 612个,占总数的41%。根据前人"重复地震震中位置间的差异约为四分之一优势波长"的研究成果,将筛选出的重复地震对用于定量判断地震目录中的震相拾取误差和评估台网定位精度,结果显示:广东地震台网的震相拾取误差约80%在0.3 s内,约70%在0.2 s内,40%多在0.1 s内;内陆定位误差较小,这与该地台站密集、方位分布较好有关,而沿海定位误差相对较大。     

地震检测与震相自动拾取研究

针对微震事件易受噪声干扰等特点,本文将STA/LTA方法和基于方差的AIC方法（var-AIC）相结合,在震相到时初步拾取的基础上,使用台站的德洛内（Delaunay）三角剖分及台站间最大走时差约束来减少噪声干扰的影响. 利用到时进行地震定位之后,根据台站预测到时,在设定的时间窗内对地震震相进行更精细的分析. 特别是针对微震事件信噪比低的特点,设计了基于偏振分析的拾取函数,根据窗内STA/LTA方法和var-AIC方法的拾取结果自动选择合适的值作为震相到时. 最后,对西昌流动地震台阵2013年304个单事件波形数据的分析处理和检验结果表明,本文方法较传统方法具有更高的地震事件检测能力和更高的震相拾取精度.      

基于深度神经网络的地磁观测数据高压直流输电干扰事件识别

随着越来越多高压直流输电线路的投入运行,地磁观测数据质量受到了严重影响.现有以人工或半人工方法识别高压直流输电干扰事件的工作量也随着受干扰范围的不断扩大和地磁观测仪器的增多而成倍增加.为了高效、准确地识别地磁观测数据中的高压直流干扰事件,本文基于卷积神经网络和长短期记忆神经网络,提出了一种高压直流输电干扰事件自动识别深度学习模型.利用2012年1月1日至2014年12月31日地磁台站原始观测数据,结合专家标注的持续时间在2 h内的高压直流输电干扰事件记录,制作高压直流输电干扰样本34360条,正常样本34360条.模型在训练集上的准确率达到了94.12%,验证集上的准确率达到了92.94%,测试集上的准确率达到了92.86%.初步研究表明深度学习方法在识别地磁观测数据中的高压直流输电干扰事件中具有较高的准确率,为下一步自动识别地磁观测数据中的车辆干扰、基建工程干扰、轻轨干扰等其他干扰事件提供了一种新的思路.     

地震和爆破事件源波形信号的卷积神经网络分类研究

本文首先从震源波形中提取梅尔频率倒谱系数(MFCC)图,然后采用卷积神经网络(CNN)进行地震波形信号的震源类型—天然地震和爆破事件—分类识别.事件为首都圈及其附近的72个天然地震和101个人工爆破事件,用于提取梅尔频率倒谱系数图的波形信号为各观测台站波形3分量中的垂直分量波形.在各个事件的所有观测台站的垂直分量波形中,通过滑动窗口按同一准则去除被噪声淹没的部分台站波形,只选择留下未被噪声淹没的台站波形.每一个事件有107个观测台站,故有107份垂直分量波形,而不同事件被留下未被噪声淹没的波形则有几份至几十份不等.然后提取被留下未被噪声淹没的波形的梅尔频率倒谱系数图,以梅尔频率倒谱系数图作为CNN的输入,CNN的输出则为波形的震源类型(天然地震事件或爆破事件).若以单份波形为识别单元,采用五折交叉验证法进行测试,得到的平均准确率为95.78%.使用训练集中单份波形为识别单元,提取梅尔频率倒谱系数图,采用CNN训练出了天然地震事件与爆破事件波形分类器,一个事件在测试集中的多份波形信号通常不会都被正确识别,很可能有些波形被识别为天然地震事件,另一些波形被识别为爆破事件;这时,若识别单元改为事件,一个事件各台站的有效垂直分量波形中,超过一半的波形被识别为某一事件类型,则这个事件被归类为该事件类型,得到的正确识别率为97.1%.实验结果表明:卷积神经网络在天然地震事件与爆破事件的识别方面表现出色.这说明MFCC与卷积神经网络可以用于识别天然地震和爆破事件,尤其是深度学习更值得在地震信号处理方面做进一步的研究.     

基于机器学习算法的银川台阵资料自动化震相分析和定位

近年来快速发展的机器学习算法显著提高了震相拾取的精度和效率。 采用卷积神经网络和递归神经网络的震相识别方法对银川台阵2019年6~7月的连续波形数据进行事件检测和P、 S震相拾取, 并通过快速震相关联和事件定位得到了银川地区较全的地震目录。 结果表明, 当震相数小于10时, 虽然可以检测出较多事件, 但分布呈弥散状, 与区域地震活动特征不符。 进一步对震相数≥10的事件进行了人工复核。 总体而言, 随着震相数量的增加, 事件的误检率逐步降低。 震相数16是该地区自动检测和定位结果准确性的拐点。 当震相数≥20时, 全部召回了地震目录中的13个地震事件, 二者平均定位差异4.27 km。 经过人工复核, 检测到的真实地震事件为区域内地震目录中事件数量的9倍。 本文使用的基于机器学习和快速震相关联和定位方法的流程可在确保准确率的基础上降低人工检测的难度, 提高地震检测的效率。     

Seismic phase picking in China Seismic Array using a deep convolutional neuron network*

Seismic phase picking is the preliminary work of earthquake location and body-wave travel time tomography. Manual picking is considered as the most accurate way to access the arrival times but time consuming. Many automatic picking methods were proposed in the past decades, but their precisions are not as high as human experts especially for events with low ratio of signal to noise and later arrivals. As the increasing deployment of large seismic array, the existing methods can not meet the requirements of quick and accurate phase picking. In this study, we applied a phase picking algorithm developed on the base of deep convolutional neuron network (PickNet) to pick seismic phase arrivals in ChinArray-Phase III. The comparison of picking error of PickNet and the traditional method shows that PickNet is capable of picking more precise phases and can be applied in a large dense array. The raw picked travel-time data shows a large variation deviated from the traveltime curves. The absolute location residual is a key criteria for travel-time data selection. Besides, we proposed a flowchart to determine the accurate location of the single-station earthquake via dense seismic array and phase arrival picked by PickNet. This research expands the phase arrival dataset and improves the location accuracy of single-station earthquake.     

Active Source Seismic Identification and Automatic Picking of the P-wave First Arrival Using a Convolutional Neural Network

In seismic data processing, picking of the P-wave first arrivals takes up plenty of time and labor, and its accuracy plays a key role in imaging seismic structures. Based on the convolution neural network (CNN), we propose a new method to pick up the P-wave first arrivals automatically. Emitted from MINI28 vibroseis in the Jingdezhen seismic experiment, the vertical component of seismic waveforms recorded by EPS 32-bit portable seismometers are used for manually picking up the first arrivals (a total of 7242). Based on these arrivals, we establish the training and testing sets, including 25,290 event samples and 710,616 noise samples (length of each sample:2s). After 3,000 steps of training, we obtain a convergent CNN model, which can automatically classify seismic events and noise samples with high accuracy (> 99%). With the trained CNN model, we scan continuous seismic records and take the maximum output (probability of a seismic event) as the P-wave first arrival time. Compared with STA/LTA (short time average/long time average), our method shows higher precision and stronger anti-noise ability, especially with the low SNR seismic data. This CNN method is of great significance for promoting the intellectualization of seismic data processing, improving the resolution of seismic imaging, and promoting the joint inversion of active and passive sources.     

Automatic onset phase picking for portable seismic array observation

Automatic phase picking is a critical procedure for seismic data processing, especially for a huge amount of seismic data recorded by a large-scale portable seismic array. In this study is presented a new method used for automatic accurate onset phase picking based on the proporty of dense seismic array observations. In our method, the Akaike＇s information criterion （AIC） for the single channel observation and the least-squares cross-correlation for the multi-channel observation are combined together. The tests by the seismic array observation data after triggering with the short-term average/long-term average （STA/LTA） technique show that the phase picking error is less than 0.3 s for local events by using the single channel AIC algorithm. In terms of multi-channel least-squares cross-correlation technique, the clear teleseismic P onset can be detected reliably. Even for the teleseismic records with high noise level, our algorithm is also able to effectually avoid manual misdetections.     

Improving automatic location procedure by waveform similarity analysis: An application in the South Western Alps (Italy)

The accuracy of automatic procedures for locating earthquakes is influenced by several factors such as errors in picking seismic phases, network geometry, modeling errors and velocity model uncertainties. The main purpose of this work is to improve the performances of the automatic procedure employed for the “quasi-real-time” location of seismic events in North Western Italy by developing a procedure based on a waveform similarity analysis and by using only one seismic station.To detect “earthquake families” a cross-correlation technique was applied to a data set of seismic waveforms recorded in the period 1985-2002, in a small test area (1600 km²) located in the South Western Alps (Italy). Normalized cross-correlation matrices were calculated using about 2700 seismic events, selected on the basis of the signal to noise ratio, manually picked and located by using the Hypoellipse code. The waveform similarity analysis, based on the bridging technique, allowed grouping about 65% of the selected events into 80 earthquake families (multiplets) located inside the area considered. For each earthquake family a master event is selected, manually re-picked and re-located by using Hypoellipse code. Having chosen a reference station (STV) on the basis of the completeness of the available data set, an automatic procedure has been developed with the aim of cross-correlating new seismic recordings (automatically picked) to the waveforms of the events belonging to the detected families. If the new event is proved to belong to a family (on the basis of the cross-correlation values), its hypocenter co-ordinates are defined by the location of the master event of the associated family. The performance of the proposed procedure is tested and demonstrated using a data set of 104 selected earthquakes recorded in the period January 2003-June 2004 and located in the test area. The automatic procedure is able to locate, associating events with the multiplets detected by the waveform similarity analysis, about 50% of the test events, almost independently of the accuracy of the automatic phase picker and without the biasing of the network geometry and of the velocity model uncertainties.     

CRPN: A cascaded classification and regression DNN framework for seismic phase picking*

Current deep neural networks (DNN) used for seismic phase picking are becoming more complex, which consumes much computing time without significant accuracy improvement. In this study, we introduce a cascaded classification and regression framework for seismic phase picking, named as the classification and regression phase net (CRPN), which contains two convolutional neural network (CNN) models with different complexity to meet the requirements of accuracy and efficiency. The first stage of the CRPN are shallow CNNs used for rapid detection of seismic phase and picking P and S arrival times for earthquakes with magnitude larger than 2.0, respectively. The second stage of CRPN is used for high precision classification and regression. The regression is designed to reduce the time difference between the probability maximum and the real arrival time. After being trained using 500,000 P and S phases, the CRPN can process 400 hours’ seismic data per second, whose sampling rate is 1 Hz and 25 Hz for the two stages, respectively, on a Nvidia K2200 GPU, and pick 93% P and 89% S phases with the error being reduced by 0.1s after regression correction.     

重庆地区重复地震识别及在数字地震台网定位评价中的应用

利用重庆数字地震台网2010年1月至2017年12月的地震波形资料和观测报告,选出5个研究区1 251个M_L≥1.5地震进行波形互相关计算,识别出358对同时被2个地震台站记录且各台波形互相关系数(cc)不小于0.8的重复地震对,涉及342个地震事件,约占地震总数的27%。将筛选出的重复地震对用于定量判断地震目录中震相拾取误差及评估台网定位精度,结果显示:重庆数字地震台网的垂直定位误差约为3 km,水平定位误差约为5 km,Pg、Sg震相拾取误差分别为0.5 s和0.7 s;其中武隆区地震定位和震相拾取精度最高,綦江区最低。