【作者】 徐震；

【导师】 王涛；

【作者基本信息】 南京大学 ， 固体地球物理学， 2020， 硕士

【摘要】 在走时成像过程中,震相到时拾取占用了大量的人力与机时,其准确性也是地震波速结构成像的关键所在。过去的几十年中地震学家发展了各种各样的震相到时自动拾取方法,这些方法在一定程度上减少了地震数据处理的工作量,然而传统自动拾取方法都各有其局限性。近些年来,地震数据量呈指数级增长,而走时成像精细度要求越来越高,因此需要发展新的震相自动拾取技术。近些年人工智能尤其是深度学习技术在数据挖掘,图像和语音识别,目标检测等多领域都取得了引人注目的成果。深度学习在提取数据特征方面有着得天独厚的优势,地震学家也开始将其引入并应用到地震学的各项研究。基于深度学习,本文分别从波形分类和图像分割两方面提出两种主动源P波初至的自动拾取方法。首先,基于噪声和地震信号在波形上的差异,建立并训练卷积神经网络(Convolutional Neural Network,CNN)进行分类,进而根据输出为地震事件的概率进行P波初至自动拾取。利用江西景德镇实验中由可控震源车产生的地震波信号被短周期地震仪记录到的垂向道分量数据,人工拾取了7242条P波初动到时,通过数据增强生成25290条地震样本和710616条噪声样本(时窗长度均为2s)。利用这些样本,训练得到一个对地震和噪声进行自动分类的卷积神经网络。本研究将训练所得的卷积神经网络扫描连续地震记录,输出不同时间窗波形为地震信号的概率,并将概率最大处对应的时刻作为P波初动到时。测试结果显示,CNN方法对地震和噪声的检测正确率均达到了99%以上,且具有P波初至高精度的自动拾取能力(平均拾取误差:<0.10 s)。同时,与传统的短长时窗比方法(STA/LTA)相比,在对信噪比较低的记录CNN能达到更好的自动拾取效果。其次,将单炮记录作为整体图像,构建U-Net进行图像边界提取,从而实现P波初至到时的自动拾取。利用中石油东方地球物理公司提供的1450炮共炮点道集数据,使用预处理过的地震道集和初至标记图像训练U-Net。利用U-Net在图像分割方面的优势将地震道集初至前和初至后的两部分分割开,通过提取边界的方式进行初至到时拾取。本文测试了U-Net对共炮点道集的拾取误差,结果显示误差小于0.02s的样本约占总测试样本的63%。基于本文研究结果,在今后拾取主动源P波初至时,可以根据不同观测系统借鉴这两种方法,从而更加准确高效地进行P波初至到时的自动拾取。更多还原

【Abstract】 In seismic data processing,the picking of arrivals takes plenty of time and labor,and its accuracy plays a key role in imaging seismic structures.In the recent decades,seismologists have developed a variety of methods to pick up arrival times and achieved some improvement on seismic data processing.However,the traditional automatic picking methods have limitations.With the exponential increase of seismic data,new methods with high resolution for automatically picking arrivals of seismic phases are urgently needed.Recently,artificial intelligence,especially deep learning technology,has made remarkable achievements in data mining,image recognition,speech recognition,target detection,and other fields.Because deep learning is good at extracting features from data,it has been widly applied in the seismological studies.In this study,we propose two methods for the active-source P wave first-arrival picking based on deep learning by waveform classification and image segmentation.First,we construct and train the Convolutionl Neural Networks(CNN)based on the difference between noise and seismic signal.Emitted from MINI28 vibroseis in the Jingdezhen seismic experiment,the vertical components of seismic waveforms recorded by EPS 32-bit portable seismometers are used for manually first-arrivals picking(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.Second,taking common-shot record as a whole image,we construct the U-Net to extract the image boundary and pick P-wave first arrivals.We train the U-Net with first arrivals from 1450 common-shot gather records provided by Bureau of Geophysical Prospecting INC.With the advantage of the U-Net in image segmentation,the segments before and after the first arrivals of common-shot gather records are separated,and the first arrivals are picked up by extracting the boundary.Our results present a small error less than 0.02 s for 63% of test samples.According to the characteristics of seismic observation system,we can adopt the CNN or U-Net for automatically picking-up the P-wave first arrivals from active sources.更多还原