白鹤滩水库库区基于深度学习的震相拾取与地震定位

近年来深度学习技术广泛应用于震相拾取与地震定位研究,采用深度神经网络搭建的EQTransformer模型对白鹤滩水库库区34个数字地震台站2016—2018年记录的连续数据进行P、 S波震相拾取,并通过REAL进行震相关联和初步定位,然后使用VELEST和hypoDD地震定位算法优化地震位置。研究表明,基于深度学习的震相拾取,与白鹤滩水库地区传统的人工处理方法相比显示出更高的效率,EQTransformer模型可拾取与人工拾取相当的P、 S波震相到时,其时间差的均值分别为0.03 s和0.07 s,符合正态分布。REAL初步定位后的地震个数(13815个)接近常规目录(7862个)的2倍,最终通过hypoDD获得了7108个高精度定位地震。估算的震级比常规目录中的震级平均低0.27,震级差值集中在0.7以内,最小完备震级由常规目录的M_L1.4更改为M_L0.6+0.27,填补了部分常规目录的震级空白,丰富了研究区域内的中小型地震。     

机器学习在地震检测与震相识别的应用综述

在地震学研究中地震检测与震相识别是最基础的环节,其拾取速度和精度直接影响其在地震精确定位以及地震层析成像中的应用效率和精度。近年来,机器学习在地震学领域中引起广泛关注。机器学习可以改进传统地震检测和震相识别方法,使它们能达到更加准确,识别率更高的效果。把机器学习方法按照监督学习和无监督学习分类介绍,并对机器学习方法流程进行总结,并对目前在地震检测与震相识别方面应用较为广泛的机器学习方法(卷积神经网络、指纹和相似性阈值、广义相位检测、PhaseNet、模糊聚类)进行综述。结果表明:机器学习在地震事件检测和震相识别将会是主要的手段。数据驱动的机器学习在地震学中的应用和物理模型的联合运用将是未来的发展趋势。     

基于机器学习和台阵相关性的水力压裂微地震事件自动识别及到时拾取

利用密集台阵对水力压裂微地震进行监测将有助于优化储层压裂、揭示断层活化.为满足密集台阵海量采集数据的处理需求,本文建立了一种综合运用多种机器学习方法和台阵相关性的、无需人工干预的自动处理流程,从而能够快速得到高质量的密集台阵震相到时目录.该综合策略包括：(1)利用迁移学习在连续波形中快速检测地震事件;(2)利用U型神经网络PhaseNet自动拾取P波、S波震相;(3)利用三重线性剔除法,结合密集台阵到时相关性剔除异常到时数据和地震事件;(4)利用K-means和SVM两类机器学习算法,进一步区分发震时刻接近的多个地震事件,减小事件漏拾率.通过将该流程应用于四川盆地长宁—昭通页岩气开发区微地震监测数据,并将自动处理结果与人工拾取结果进行比对发现,二者在震级测定、定位以及走时成像结果等方面具有很好的一致性,表明本文处理流程结果精度可达到手动处理精度.本文结果为密集台阵地震监测数据的高效、高精度处理提供了新思路.     

基于样本增强的卷积神经网络震相拾取方法

为了快速、高效地从地震数据中识别地震事件和拾取震相,本文利用基于样本增强的卷积神经网络自动震相拾取方法,将西藏林芝地区L0230台站3个月数据作为训练集,该区内另外6个台站连续1个月的波形数据作为测试集,采用高斯噪声、随机噪声拼接、随机挑选噪声、随机截取地震事件等4种样本增强的方法扩增训练集,以提高自动震相拾取技术的准确率。结果显示:样本增强前模型在测试集上的地震事件识别准确率为80%,样本增强后提升至97%,表明样本增强有效地提高了模型的泛化性能和抗干扰能力;在0.5 s误差范围内,震相自动拾取准确率高于81%,在1.0 s误差范围内,准确率高于95%;利用基于样本增强的卷积神经网络震相拾取方法能够检测出人工拾取震相中误标和漏检的震相。      

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。     

上海及邻区微震检测与定位

本文利用基于图像处理器加速的模板匹配定位法(Graphics Processing Unit-based Match&Locate, GPU-M&L)和双差定位法(HypoDD),对上海及邻区13个台站记录的2011年至2020年共10年的连续地震数据资料进行分析.首先从中国地震台网中心提供的146个地震事件目录中挑选了136个地震事件作为模板事件,使用模板匹配定位技术对上海及邻区10年的连续资料进行遗漏地震事件的扫描和检测,共识别出824个地震事件,约为台网中心提供地震目录事件数量的5.5倍.然后对识别出的地震事件通过深度去噪方法(DeepDenoiser)将信号与噪声分离,并对去噪后地震波形的频率和振幅特性分析来进一步确认识别出的地震事件.同时利用基于机器学习的震相拾取技术(PhaseNet),对去噪后的333个地震事件进行了震相拾取.检测后的地震目录完备震级由台网目录的M_c1.0降为M_c0.8.最后利用双差定位法对479个地震事件进行精定位,精定位的结果显示,上海地区整体地震活动性较弱,地震的空间分布相对较为分散,定位后...     

基于喜马拉雅台阵的内蒙古地震监测能力薄弱区拾震能力分析

基于中国地震科学台阵资料及内蒙古、甘肃地震台网资料,采用结合台阵策略的震相拾取深度学习方法APP,开展内蒙古地震监测能力薄弱区——阿拉善右旗拾震能力研究。研究结果显示,APP方法检测到了人工目录中97.8%的地震,地震拾取总数为人工目录地震数的22倍。经tomoDD方法定位后,地震深度分布较符合内蒙古西部的地质构造特征。对震源深度与断裂位置间相关性的初步分析显示,深度随纬度变化中有5条深度“集中条带”与研究区7条断裂的位置相对应,深度随经度变化中有4条深度“集中条带”与研究区7条断裂的位置相对应。分析认为,APP拾取方法在实际地震资料应用中展示出较强的泛化能力,可为增强固定地震台网对于监测能力薄弱地区微震的识别能力,以及优化地震台网布局、提高监测能力薄弱地区的地震监测水平等提供参考。     

基于深度学习到时拾取自动构建长宁地震前震目录

将深度学习到时拾取、震相关联技术与传统定位方法联系起来,构建一套连续波形自动化处理与地震目录自动构建流程,对于高效充分利用地震资料,提升微震检测能力具有十分重要的意义.我们应用最新发展的迁移学习震相识别技术、震相自动关联技术,对长宁M S6.0地震震中附近21个台站震前半个月(6月1日—6月17日)的连续记录波形进行P、S震相识别、震相自动关联和初步定位,并应用传统绝对定位和相对定位技术得到了长宁地震震前微震活动的绝对和相对定位目录.其中绝对定位目录能在较小的误差范围匹配85%的人工处理目录,其发震时刻平均误差为0.36±0.07 s,震级平均误差为0.15±0.024级,水平定位平均误差为1.45±0.028 km,其识别的1.0级以下微震数目是人工的8倍以上,将长宁地震震前微震目录的检测下限提升至M L-1左右,证明了基于深度学习到时识取和REAL(Rapid Earthquake Association and Location,快速震相关联和定位技术)震相自动关联来构建微震目录具有较好的实用性.我们的自动地震目录揭示了长宁M S6.0主震所发生的区域震前异常频繁的微震活动,以及与区域内盐矿注水井的关联性,更好地描绘了这些微震活动的时空演化特征,其空间活动性分布特征与长宁M S6.0余震序列的分布一致.     

东祁连山北缘断裂带基于深度学习的密集台阵地震事件快速检测与定位研究

为监测东祁连山北缘断裂带附近的地震活动性,布设包含240台短周期地震仪的面状密集台阵,进行约30 d的连续观测。首先使用基于深度学习的多台站地震事件检测算法（CNNDetector）进行地震事件检测,然后使用震相拾取网络（PhaseNet）对地震事件进行P波和S波到时拾取,其次使用震相关联算法（REAL）进行震相关联及初定位,最后使用双差定位（hypoDD）进行地震重定位,最终的精定位地震目录中共有517个地震。在密集台阵观测期间,中国地震台网正式地震目录中共有39个位于台阵内的地震事件,相比而言,密集台阵检测到大量小于0级的地震。因此通过布设密集台阵,可提高活动断裂微地震活动性的监测能力。与历史地震空间分布相比,密集台阵地震精定位分布具有较好的一致性,表现出更明显的线性分布特征。基于地震分布,发现研究区域存在与地表断层迹线走向不同的隐伏活跃断裂。     

基于深度学习震相拾取和密集台阵数据构建青海玛多Ms7.4地震震源区高分辨率地震目录

高分辨率地震目录有助于描绘断层的精细结构和认识发震断裂的构造形态及发震机制.基于玛多地震科考布设的短周期台阵数据,本文利用深度学习自动拾取P/S波震相、震相关联、绝对定位、相对定位等定位流程,构建了玛多Ms7.4地震后第14天至第43天的高分辨率地震目录,揭示了玛多震源区主震西侧以及主震向东20 km区域范围内的地震序...     

Earthquake detection in the Jiangsu region,China using graphics-processing-unit-based Match & Locate and rapid earthquake association and location*

Earthquake detection and location are essential in earthquake studies, which generally consists of two main classes: waveform-based and pick-based methods. To evaluate the ability of two different methods, a graphics-processing-unit-based Match & Locate (GPU-M&L) method and a rapid earthquake association and location (REAL) method are applied to continuous seismic data recorded by 24 digital seismic stations from Jiangsu Seismic Network during 2013 for comparison. GPU-M&L is one of waveform-based methods by waveform cross-correlations while REAL is one of pick-based method to associate arrivals of different seismic phases and locate events through counting the number of P and S picks and travel time residuals. Twenty-six templates are selected from the Jiangsu Seismic Network local catalog by using the GPU-M&L. The number of newly detected and located events is about 2.8 times more than those listed in the local catalog. We both utilize a deep-neural-network-based arrival-time picking method called PhaseNet and a short-term/long-term average (STA/LTA) trigger algorithm for seismic phase detection and picking by applying the REAL. We then refine seismic locations using a least-squares location method (VELEST) and a high-precision relative location method (hypoDD). By applying STA/LTA and PhaseNet, 1006 and 1893 events are associated and located, respectively. The newly detected events are mainly clustered and show steeply dipping fault planes. By analyzing the performance of these methods based on long-term continuous seismic data, the detected catalogs by the GPU-M&L and REAL show that the magnitudes of completeness are 1.4 and 0.8, respectively, which are smaller than 2.6 given by the local catalog. Although REAL provides improvement compared with GPU-M&L, REAL is highly dependent on phase detection and picking which is strongly affected by signal-noise ratio (SNR). Stations at southeast of the study region with low SNR may lead to few detections in the same area.     

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.     

Application of Machine Learning Methods in Arrival Time Picking of P Waves from Reservoir Earthquakes

Reservoir earthquake characteristics such as small magnitude and large quantity may result in low monitoring efficiency when using traditional methods. However, methods based on deep learning can discriminate the seismic phases of small earthquakes in a reservoir and ensure rapid processing of arrival time picking. The present study establishes a deep learning network model combining a convolutional neural network (CNN) and recurrent neural network (RNN). The neural network training uses the waveforms of 60 000 small earthquakes within a magnitude range of 0.8-1.2 recorded by 73 stations near the Dagangshan Reservoir in Sichuan Province as well as the data of the manually picked P-wave arrival time. The neural network automatically picks the P-wave arrival time, providing a strong constraint for small earthquake positioning. The model is shown to achieve an accuracy rate of 90.7% in picking P waves of microseisms in the reservoir area, with a recall rate reaching 92.6% and an error rate lower than 2%. The results indicate that the relevant network structure has high accuracy for picking the P-wave arrival times of small earthquakes, thus providing new technical measures for subsequent microseismic monitoring in the reservoir area.     

Generalization of PhaseNet in Shandong and its application to the Changqing M4.1 earthquake sequence

Waveforms of seismic events, extracted from January 2019 to December 2021 were used to construct a test dataset to investigate the generalizability of PhaseNet in the Shandong region. The results show that errors in the picking of seismic phases (P- and S-waves) had a broadly normal distribution, mainly concentrated in the ranges of −0.4–0.3 s and −0.4–0.8 s, respectively. These results were compared with those published in the original PhaseNet article and were found to be approximately 0.2–0.4 s larger. PhaseNet had a strong generalizability for P- and S-wave picking for epicentral distances of less than 120 km and 110 km, respectively. However, the phase recall rate decreased rapidly when these distances were exceeded. Furthermore, the generalizability of PhaseNet was essentially unaffected by magnitude. The M4.1 earthquake sequence in Changqing, Shandong province, China, that occurred on February 18, 2020, was adopted as a case study. PhaseNet detected more than twice the number of earthquakes in the manually obtained catalog. This further verified that PhaseNet has strong generalizability in the Shandong region, and a high-precision earthquake catalog was constructed. According to these precise positioning results, two earthquake sequences occurred in the study area, and the southern cluster may have been triggered by the northern cluster. The focal mechanism solution, regional stress field, and the location results of the northern earthquake sequence indicated that the seismic force of the earthquake was consistent with the regional stress field.     

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.     

Comparison of two earthquake early warning location methods

According to earthquake catalog records of Fujian Seismic Network, the T _now method and the four-station continuous location method put forward by Jin Xing are inspected by using P-wave arrival information of the first four stations in each earthquake. It shows that the four-station continuous location method can locate more seismic events than the T _now method. By analyzing the results, it is concluded that the reason for this is that the T _now method makes use of information from stations without being triggered, while some stations failed to be reflected in earthquake catalog because of discontinuous records or unclear records of seismic phases. For seismic events whose location results can be given, there is no obvious difference in location results of the two methods and positioning deviation of most seismic events is also not significant. For earthquakes outside the network, the positioning deviation may amplify as the epicentral distance enlarges, which may relate to the situation that the seismic stations are centered on one side of epicenter and the opening angle between seismic stations used for location and epicenter is small.     

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.