Prestack Seismic Reflection Pattern Research Using Tensor Dictionary

Seismic reflection pattern research mainly uses the distinction of different seismic reflection signals to make the reflector get the correct pattern assignment,so as to obtain the target horizon's information about the lithology,fluid and stratum deposition.Traditional seismic reflection pattern research is based on post-stack seismic data,but the superimposition process loses a lot of important details in the formation.Therefore,this thesis uses pre-stack seismic data conduct seismic reflection pattern research.In the actual oil and gas exploration process,it is difficult to obtain logging information and label samples.The label data obtained by conventional methods is far from being able to meet the needs of intelligent algorithms which about complex geological structures' exploration and development.Most of the existing methods were converted the high-dimensional pre-stack seismic data into one-dimensional vectors for processing,which loses the characteristics of the amplitude change with offset,azimuth and spatial changes in the pre-stack seismic data.Furthermore,in pre-stack seismic data,it ignores the influence of noise on reflection mode analysis results.Therefore,this article will focus on pre-stack seismic label data construction,tensor dictionary learning,transfer learning,etc.The analyses of the main research work and innovativeness are as follows:(1)In order to solve the problem of less seismic data label samples,this thesis designs a method for constructing pre-stack seismic label data.The distribution law of logging data is extracted by the joint probability density function and the gradient of adjacent points on the upper and lower strata.Then,the relevant information of the target horizon's stratigraphic lithology can be obtained,and with the combination of the designed stratigraphic framework model,the pre-stack seismic label data can be generated.The method enriches the diversity of label samples and provides a data basis for the subsequent intelligent recognition of reflection patterns.(2)In order to solve the problem of high noise and high data dimension of pre-stack seismic signals,this thesis proposes and implements a tensor discriminant dictionary learning algorithm.This method uses Tucker decomposition to obtain the core sparse tensor of multi-dimensional pre-stack data to extract features for classification,and introduces Pearson correlation coefficient to measure the correlation degree of different types of sparse tensors,and combines the reconstruction error with the sparse coefficient error make judgments.This method had considered the data dimension and noise immunity at the same time,and reduces the error of pattern assignment.(3)The direct introduction of pre-stack seismic label data into the actual work area has the problem of domain difference.At the same time,in order to take into account the advantages of tensor dictionary learning,this thesis combines tensor dictionary learning and domain adaptation to train a set of alignment matrix sets,so that the constructed pre-stack seismic label data is used as the source domain,and the pre-stack data in the target work area is used as the source domain.The target domain aligns the core sparse tensors from the two domains into the invariant core tensor quantum space,thereby effectively reducing the difference between the source domain and the target domain.This thesis compares the proposed method with the traditional method through artificial synthetic data and actual work area data,which proves the effectiveness of the proposed method and provides technical support for the study of refined seismic reflection models.