Research On Prestack Seismic Waveform Classification Method Based On Semi-Supervision

It is important to determine the underground reservoir by seismic waveform classification.The existing seismic waveform classification methods mainly aims at the post-stack seismic waveform.The post-stack seismic waveform is the lateral summation of the pre-stack waveform,which leads to much seismic information lost.Therefore,this thesis mainly studies how to use pattern recognition technology to classify pre-stack seismic waveforms.The pre-stack seismic waveform data is too large to be classified directly using existing methods,otherwise there will be dimensional disaster.In actual exploration,a small amount of logging information is often used.A small amount of logging information is used as the tag to carry out semi-supervision and reduction of seismic data and remove redundant information.However,the logging label information is too little to train the supervised classifier.In view of the above problems,this thesis studies from pre-stack seismic waveform semi-supervised dimensionality reduction,neighborhood information,similarity measurement and semi-supervised learning,the specific work is as follows:1.Proposing the pre-stack seismic waveform classification method based on semi-supervised dimensionality reduction.1)In order to solve the problem of high dimension,this thesis introduction of a semi-supervised dimensionality reduction algorithm.Using the known logging data in dimension reduction at the same time be able to feature extraction,expand the same kind of similarity and increase the differences of different categories,which means the output can be a good expression of the original signal.The algorithm can effectively reduce the dimension of data so that the subsequent classification can be calculated effectively.Moreover,the tag information is used to train a similarity measurement matrix,which is more consistent with the existing logging data.2)Considering the continuity of the strata and the adjacent reflection waveform has strong correlation and the actual seismic data is often noisy,this thesis propose a k-means algorithm based on neighborhood information.The neighborhood information by the window way to select the optimal neighborhood to join the process of clustering.The algorithm is more consistent with the structural features of seismic data,which canreduce the influence of noise in seismic data and make the classification results better.2.Propose a SFCM-S algorithm deal with pre-stack seismic waveform based on ensemble learning.1)It is difficult to describe the actual geological structure with a clear classification resul.This paper introduces the FCM algorithm of spatial constraint considering the continuity of strata.At the same time,the semi-supervised FCM-S algorithm is used to optimize the clustering process.SFCM-S algorithm can make good use of spatial constraints and logging label information constraints to enrich the seismic information.2)A new ensemble learning algorithm,E-SFCM-S,was proposed to solve the problem of feature selection.The SFCM-S algorithm is used as a sub-learning device to further improve the accuracy of classification results by selecting the different feature sets of the reduced dimension as the input of the sub-learning device.