Semi-Supervised Based Pre-stack Seismic Signal Reflection Pattern Analysis

Seismic reflection pattern analysis mainly analyzes the difference between the signals of different reflection element in the stratum,and then divides different reflection elements into corresponding reflection patterns,so as to obtain the geological structure of the target interval,and provide guidance for subsequent oil and gas development.The existing seismic reflection pattern analysis is mostly based on post-stack data.Compared with the pre-stack data,the post-stack data reduces the amount of data under the premise of improving the signal-to-noise ratio,but it also obscures many tiny information in the stratum,which cannot meet the needs of refined interpretation.This thesis starts from the high-dimensional pre-stack seismic data,analyzing and extracting the texture attributes to preserve the tiny information in the stratum.In the actual oil and gas exploration process,the labeled information obtained through logging and drilling is extremely rare,which cannot support effective supervised learning.Second,the traditional analysis results can only be qualitative analysis,and cannot be quantitatively evaluated.Quantitative analysis is an urgent problem to be solved at present.In addition,the pre-stack data is more affected by noise than the post-stack data.In order to alleviate these problems,this thesis conducts research on semi-supervised learning,quantitative evaluation,and spatial information,and proposes corresponding solutions.The specific work and innovation of this article are summarized as follows:1.In view of the fact that there is too rare prior knowledge in seismic exploration and it is impossible to carry out supervised learning,this thesis proposes a semi-supervised learning method with global optimization,which introduces label information in the training process of self-organizing map neural network.These labeled information assist unlabeled data to enhance the training effect of the learner.This method also optimizes the mapping relationship between network neurons and seismic reflection patterns,solves the limitations of self-organizing map neural networks,and reduces the confusion of pattern assignment.2.The traditional seismic reflection pattern analysis methods cannot quantitatively evaluate the results,while quantitative analysis is an important basis for the reliability of the results.Based on this situation,this thesis uses the probabilistic neural network to calculate the membership degree of seismic reflection elements to different reflection patterns.Based on this result,three measurement for evaluating the results are defined,which provide different perspectives for evaluating the results.3.In order to solve the influence of noise on pre-stack seismic data,this thesis introduces a prior spatial information to restrict the pattern assignment process of reflectors by analyzing the continuity of strata.By analyzing the spatial transmission of prior information,this method corrects the data-driven results in the form of probability,and reduces the influence of noise on pre-stack data,so that the final pattern analysis results are more consistent with the actual formation conditions.In this thesis,the proposed methods and traditional methods are applied to synthetic data and actual work area data,and the advantages of the proposed methods are proved by comparison.