Study On Seismic Facies Recognition Methods Based On Feature Learning

Machine learning is an important tool for the intelligent interpretation of seismic data and well logging data in reservoir prediction.During its implementation,feature learning or feature representation is an important part of the inference and prediction for targets based on a learning model.According to the learning mode,feature learning can be divided into two categories: supervised learning and unsupervised learning.In the supervised learning model,seismic facies identification and lithology identification are classification tasks,and there are common problems such as overfitting,imbalance problem,or inconsistent distribution of data domains.In the unsupervised learning mode,cluster analysis is the main research at present.And the similarity measurement of seismic data,the representation of structural features and the visualization method are the keys that affect the results of cluster analysis.In addition,from the data point of view,no matter which learning mode is used,both the noise in the data and multiple solutions will seriously affect the effect of the learning method.To solve above problems,the following research work has been carried out in this paper:(1)In the task of seismic facies classification,a classification method based on convolutional neural network and spatial classification probability framework is proposed in this paper.Based on the spatial classification probability framework,this method uses a one-dimensional convolutional network to classify seismic attributes,and combines the posterior probability and spatial distribution probability information of the seismic attribute classifier to optimize the classification results.In the practical application in a carbonate rock area,compared with the traditional classification algorithm,this method reasonably enhances the continuity of seismic facies prediction results in the region with cluttered waveforms,reduces the risk of overfitting,and improves the classification precision of the seismic facies.(2)In the task of waveform clustering analysis,a joint clustering seismic facies analysis method is proposed in this paper,which takes the wavelet transform coefficients as the input and embeds Self-Organizing Maps(SOM)in deep sparse autocoding.This method is suitable for the situation where the clustering results are unstable due to factors such as horizon picking error,strong heterogeneity and noise interference.In the test with model data,for the samples with horizon error and velocity disturbance,the proposed method effectively improves the performance of clustering.And in the practical application of multi-level channel facies identification,by combining with the color space mapping technology,this method achieves better performance than the two-stage clustering method.(3)For the situation where the thickness of formation changes,the SOM waveform clustering analysis method is improved in this paper.Considering that the traditional SOM algorithm based on Euclidean distance is difficult to apply to variable-length waveforms,this paper uses DTW distance as the measure function of SOM,modifies the update process of its parameters,and improves the computational efficiency of the algorithm by introducing the strategy of stratified sampling and hierarchical mapping.This method has been effectively applied in the seismic facies analysis of the marine carbonate reservoirs.(4)Similar to the consideration of spatial continuity in seismic facies analysis,in the lithology classification task,the influence of the spatial lithology probability distribution on the classification results also needs to be considered.For the task of lithological sequences classification,a bidirectional recurrent neural network connected with conditional random fields is proposed to improve the prediction ability of lithology sequences.The algorithm can simultaneously obtain the classification information of logging data sequences and lithology label sequences in the vertical space,and obtain more spatial information of the classification target.However,the risk of overfitting increases due to the increased spatial constraints on the sequence.In this paper,the sequence classification algorithm is further improved by using the classification probability of the support vector machine with minimum structured risk.In the case application of sand and mudstone classification task,this method effectively reduces the risk of overfitting,while retaining its advantage of obtaining spatial information.(5)An adversarial verification method based on a deep learning network is proposed in this paper to evaluate the similarity of structural features of seismic data.This method trains a domain discriminant network and uses its separability to measure the similarity of the data.Through the experiments,we compares and analyzes the similarity evaluation method based on adversarial verification and the method of calculating multi-kernel maximum mean difference and KL(Kullback-Leibler)divergence.In the aspect of similarity evaluation of the seismic data reflection structures,the index calculated by the adversarial verification method is close to the multi-kernel maximum mean difference.In the practical applications,this index can not only provide a basis for selecting training data,but also indicate the prediction risk of a single test sample.