复数域匹配追踪近地表Q值估计及深度学习建模  

Complex domain-matching pursuit for near-surface Q-estimate and deep learning modeling

作　　者：李拥军[1] 宋炜[2] 唐传章[3] 史应龙 王泽丹 陈树光 刘静[4] 王标 LI Yongjun;SONG Wei;TANG Chuanzhang;SHI Yinglong;WANG Zedan;CHEN Shuguang;LIU Jing;WANG Biao(Bayan Exploration and Development Branch,North China Oilfield,PetroChina,Bayannur 750306,China;School of Geophysics,China University of Petroleum(Beijing),Beijing 102249,China;Exploration Department Division of Huabei Oilfield Company,PetroChina,Renqiu 062550,China;Exploration and Development Research Institute of Huabei Oilfield Company,PetroChina,Renqiu 062550,China)

机构地区：[1]中国石油天然气股份有限公司华北油田巴彦勘探开发分公司,内蒙古巴彦淖尔750306 [2]中国石油大学(北京)地球物理学院,北京102249 [3]中国石油天然气股份有限公司华北油田分公司勘探事业部,河北任丘062550 [4]中国石油天然气股份有限公司华北油田分公司勘探开发研究院,河北任丘062550

出　　处：《石油物探》2021年第1期123-135,共13页Geophysical Prospecting For Petroleum

基　　金：中国石油天然气股份有限公司科技攻关课题“河套盆地新区新领域勘探潜力与高效勘探关键技术研究”(2019DG0815);国家科技重大专项“致密气有效储层预测技术”(2016ZX05047002)共同资助。

摘　　要：常规基于微测井资料的品质因子(Q值)求取方法,通常需要拾取初至时间并按时窗求取初至波频谱,因而工作量较大,并且受选取的窗函数参数大小的影响。为此,提出了一种基于非零相位雷克子波的复数域快速匹配追踪分解并结合对数谱比法估算微测井数据近地表Q值的方法。根据微测井数据初至波能量最强的特点,利用非零相位雷克子波匹配追踪提取第一个匹配原子,其频谱表达了初至波的能量,中心时间则包含了初至旅行时信息,从而实现了用于谱比法Q值估算的初至波频谱和旅行时参数的自动提取。在谱比计算中引入整形正则化算子提高算法的稳定性,并采用优化反演算法求出稳定的谱比值,以保证对数谱比法近地表Q值估计的精度。将工区内不同测点求取的品质因子函数作为已知样本标签,通过深度学习训练形成近地表Q值的多元非线性回归算子,建立三维近地表品质因子模型。模型数据和实际数据的处理结果表明,该方法自动、高效、稳定,且抗噪能力强,将获得的近地表Q值模型用于Q值补偿,可有效提高地震资料的分辨率。The conventional method for obtaining quality factors based on uphole data usually requires to pick up the first-arrival time and calculate the first-arrival wave spectrum in a given time window.This method is time-consuming,and its outcome depends on the selection of the window function.In this study,a method for estimating the near-surface quality factor using uphole data is proposed,which combines the log-spectral ratio method with the fast matching pursuit decomposition method in a complex field based on a non-zero phase Ricker wavelet.As the first-arrival wave energy of uphole data is the strongest,the first matching atom can be automatically extracted through non-zero phase Ricker wavelet matching pursuit.The spectrum of the atom not only expresses the energy of the first-arrival wave,but also permits to identify the first-arrival travel time.The near-surface quality factor can be estimated from the spectrum and travel time of the matching atom at different depths using the log-spectral ratio method.However,the spectral ratio method for the Q-value is not stable when the denominator attains a small value;therefore,a shape-regularization operator is introduced in the spectral ratio calculation,and a stable spectral ratio is obtained through an optimized inversion algorithm.Taking the estimated quality factor function of different positions in the survey area as the known sample label,a multiple nonlinear regression operator of the near-surface quality factor is assembled through deep-learning training,and a three-dimensional near-surface quality factor model is established using the regression operator.By processing both synthetic and actual data,it is demonstrated that the method is automatic,efficient,stable,and has a strong noise-reduction ability.The Q-value model of the near surface used for Q-value compensation can effectively improve the resolution of seismic data.

关 键 词：非零相位雷克子波 匹配追踪 对数谱比 整形正则化 深度学习 微测井 

分 类 号：P631[天文地球—地质矿产勘探]