Research On Seismic Anisotropy Inversion Method Based On Dictionary Learning And Sparse Representation

Seismic anisotropy inversion is a technique that uses actual seismic data from ground observations to estimate anisotropic parameter values of underground media,and then performs reservoir prediction.At present,the method of inverting anisotropic parameters through longitudinal wave data is more widely used.It includes longitudinal wave amplitude anisotropic inversion and velocity anisotropic inversion,however,velocity anisotropy inversion can only reflect the trend of formation anisotropy parameters.Therefore,this paper uses the amplitude anisotropy inversion method to study.The inversion method based on amplitude anisotropy still has many difficult problems in practical applications,which are mainly manifested in the following aspects:(1)The inversion problem itself is an ill-posed problem.Amplitude-based anisotropic inversion requires simultaneous inversion of elastic parameters and anisotropic parameters.The ill-posed characteristics such as multi-solution and stability are more serious.And the commonly used Tikhonov regularity is difficult to adapt to the heterogeneous and non-stationary characteristics of fractured reservoirs reflected by anisotropic parameters,resulting in inconsistent prediction results and actual reservoir conditions.(2)The influence of anisotropic parameters on amplitude is relatively small compared to elastic parameters.In the inversion iteration process,how to improve the sensitivity needs further study.(3)There is a certain correlation between various parameters.Existing inversion methods do not take into account the correlation between similar parameters and the spatial variation of parameter correlation characteristics.(4)Few anisotropic parameter logging data in actual inversion will lead to major problems in initial model construction of the work area,then affect the inversion results.In view of the above problems,this thesis introduces dictionary learning and sparse representation methods to carry out research on the anisotropic inversion method with longitudinal wave amplitude.The main research work of the paper includes:(1)Anisotropic parameter inversion method based on dictionary learning and sparse representation is proposed according to Azimuth AVO prestack inversion framework of simultaneous inversion of elastic and anisotropic parameters.This method learns the dictionary of elastic parameters and anisotropic parameters separately from the actual logging data,and constructs the anisotropic and elastic parameter prior information through dictionary sparse representation for prestack inversion.The algorithm realizes the acquisition of prior information of adaptive anisotropy parameters based on datadriven logging data and as a constraint for inversion,which can avoid the problem that traditional prior information needs to pre-set mathematical assumptions,and is more suitable for prediction of non-stationary and strong anisotropic reservoirs such as fractured reservoirs.In the process of inversion,for the problem that the amplitude is not sensitive to anisotropic parameters,this article introduces the linear regression relationship between the anisotropic parameters and the longitudinal wave velocity as a constraint,which effectively improves the disturbance of anisotropic parameters during the inversion process,and makes the results are more consistent with the actual logging data.(2)In order to solve the influence of the correlation between anisotropy parameters and elastic parameters on inversion,this paper proposes a seismic anisotropy inversion method based on joint dictionary.The algorithm can learn the joint dictionary of elastic parameters and anisotropic parameters simultaneously,and then use to constrain the inversion of each parameter.The joint dictionary can extract the multi-parameter association relationship at the same spatial structure,and realize the synchronous characterization of the multi-parameter spatial structure and association relationship.Model tests and actual data test show that the inversion results of this algorithm are more accurate and reliable.(3)In order to solve the problem of lack of anisotropic logging data,the paper attempts to use online dictionary learning method for semi-supervised learning inversion.First,based on the correlation criterion,the reliable results of the aforementioned joint dictionary inversion are selected as new samples to join the training data,and the common features of the new samples and logging data are obtained through online dictionary learning,and the dictionary is gradually expanded.Finally,the expanded online learning dictionary is used for sparse representation to achieve anisotropic parameter inversion and improve the whole area data inversion result.This method provides a set of feasible solutions to the inversion of reservoir parameters where logging samples are scarce,and the actual data verifies the feasibility of the program.Through the research of the above-mentioned methods,this paper hopes to provide effective solutions for the improvement of instability,incompatibility and multi-solution of seismic anisotropic inversion,so as to promote the development of reservoir prediction.