波形反演应用初探

发布时间：2018-06-27 21:49

本文选题：波形反演 + 时间域　；参考：《中国石油大学(华东)》2015年硕士论文

【摘要】：随着人类对地下未知环境的探寻,以及经济发展对资源需求的增加,如何更精确和更高效的描述地下构造进而找寻资源变得尤为重要。较为精确的速度又是进行地下构造成像的必备条件,因此业界对高品质速度模型的探索一直关注。全波形反演方法(FWI)却在精细反演地下速度等参数方面具有潜力。其利用振幅相位信息,结合地震波运动学与动力学特性,通过波动方程拟合,不断匹配更新参数,使模型数据与实际数据残差变得最小,进而完成反演过程,求得所需参数变量。论文就全波形反演的理论进行研究。对基于波动方程的时间域以及频率域的正演反演的基本公式进行详细推导,给出全波形反演的基本框架流程。在此基础上,就梯度、步长,反演策略等几个关键问题进行了着重介绍。论文实现了时间域以及频率域的波形反演方法,并分别建立了不同的并行反演流程。时间域采用基于MPI的炮并行反演策略,而频率域采用MPI+MUMPS并行反演方法。不同点是,频率域将波场分为几个并行区域,其正演是对所有炮的不同波场区域同时并行求解,而不是时间域的炮之间并行。这样既降低了内存消耗,又提高了效率。全波形反演根据模型域与数据域间的相互反馈特征完成反演,而数据域与模型域之间也并非一一对应。这就造成了其多解性的特点,目标函数存在着极多的极小值,波形反演是强非线性的反演,因此论文研究了多尺度化的反演方法。为了进一步解决波形反演应用中面临的有效收敛问题,本文针对梯度预处理优化算法展开研究,实现了共轭梯度法、有限内存的BFGS反演优化算法、近似hessian对角元素法。FWI理论上往往能取得不俗的效果,作为一种精细的反演方法,有着无尽的潜能,受着人们的关注,但却在工业生产应用中见不到它的影子。面对实际资料,波形反演显示出了短板的一面。基于这一点,论文对全波形反演在实际中的应用进行了简单的初步尝试,并对其反演应用中所出现的问题进行分析,说明了为什么波形反演方法尚未在工业生产中应用的一些原因。就其对初始模型依赖的问题进行了复频域的尺度化反演策略的改进,并通过模型验证,证明了方法的可行性。为提高处理效果,采用尺度化方法对实际资料进行处理,将早至波与反射波数据依次反演出大尺度和小尺度信息。
[Abstract]:With the exploration of underground unknown environment and the increasing demand for resources by economic development, how to describe the underground structure more accurately and efficiently becomes more important. More accurate velocity is also a necessary condition for underground structure imaging, so the exploration of high quality velocity model has been paid attention to. Full wave inversion (FWI) has potential in fine inversion of underground velocity and other parameters. By using amplitude and phase information, combined with the kinematics and dynamics of seismic wave, and by fitting the wave equation, it continuously matches the updated parameters, which minimizes the residual error between the model data and the actual data, and then completes the inversion process. The required parameter variables are obtained. The theory of full waveform inversion is studied in this paper. The basic formulas of forward inversion in time domain and frequency domain based on wave equation are deduced in detail, and the basic frame flow of full waveform inversion is given. On this basis, several key problems, such as gradient, step size and inversion strategy, are emphatically introduced. In this paper, the waveform inversion methods in time domain and frequency domain are implemented, and different parallel inversion processes are established. MPI parallel inversion strategy is used in time domain and MPI MUMPS parallel inversion method is used in frequency domain. The difference is that the wave field is divided into several parallel regions in frequency domain, and the forward modeling is to solve the different wave field of all guns simultaneously, rather than parallel between the guns in time domain. This not only reduces memory consumption, but also improves efficiency. The full waveform inversion is based on the mutual feedback characteristics between the model domain and the data domain, and the data domain and the model domain are not one-to-one correspondence. This leads to the characteristics of multi-solution, the objective function has many minimum values, and the waveform inversion is a strong nonlinear inversion, so the multi-scale inversion method is studied in this paper. In order to solve the problem of effective convergence in waveform inversion application, this paper studies the gradient preprocessing optimization algorithm and implements the conjugate gradient method and the BFGS inversion optimization algorithm with limited memory. Approximate hessian diagonal element method. FWI can get good results in theory. As a fine inversion method, it has endless potential and has been concerned by people, but it can not be seen in the industrial production application. In the face of the actual data, the waveform inversion shows the side of the short plate. Based on this point, this paper makes a simple preliminary attempt on the application of full waveform inversion in practice, and analyzes the problems in the application of full waveform inversion. Some reasons why waveform inversion method has not been applied in industrial production are explained. The scaling inversion strategy in complex frequency domain is improved for its dependence on the initial model, and the feasibility of the method is proved by model verification. In order to improve the processing effect, the scale method is used to process the actual data, and the large scale and small scale information are inversed between the early arrival wave and the reflected wave data.
【学位授予单位】：中国石油大学(华东)
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：P631.4

【参考文献】