Research And Application For Seismic Facies Recognition Based On Semi-Supervised Learning

The success of deep learning models is inseparable from a large amount of labeled data.However,the cost of acquiring labels in many application scenarios is too high,the development of supervised learning is limited.Semi supervised learning has become a development trend of deep learning.In the field of petroleum exploration,the amount of the unlabeled data is far greater than that of labeled data.Unlabeled data can provide additional geological information,semi supervised learning can improve the accuracy of seismic facies recognition.In view of this problem,this thesis proposes three semi supervised deep learning models for lithology recognition of seismic data.The main contributions include:(1)This thesis proposes a semi supervised deep learning model based on SILPU(SLP-AMCNN).The model uses the Mixup method for linear enhancement of the data,and introduces jump connection mode and characteristic channel attention to build convolution network.Two algorithms are proposed: improved label propagation algorithm based on uncertainty(ILPU)and semi supervised label propagation algorithm based on uncertainty(SILPU).ILPU calculates the similarity matrix of samples through cosine similarity and KNN algorithm,and introduces the calculation of pseudo label uncertainty to improve the label propagation algorithm.SILPU uses the depth network to extract the characteristics of the original data,and realizes the depth semi supervised algorithm.It is verified that the recognition accuracy of the algorithm is better than other comparison algorithms on three general data sets.It is verified that the recognition accuracy of SILPU algorithm is better than other comparison algorithms on three general data sets.The model carries out lithology recognition experiments on two seismic data sets.Compared with other models,SLP-AMCNN model can achieve better results in accuracy and lithology profile.(2)This thesis proposes an integrated semi supervised deep learning model based on SILPU(E-SLP-AMCNN)and a semi supervised deep learning ensemble model based on SILPU(S-SLP-AMCNN).Using bagging idea to predict pseudo labels,the ESLP-AMCNN is obtained.Using KL divergence to calculate the predicted distribution distance between base learners in the integrated model,the S-SLP-AMCNN is obtained.The two models carry out lithology recognition experiments on two seismic data sets.Compared with other models,the three models can achieve better results in accuracy and lithology profile.(3)Combining seismic facies recognition,this thesis proposes the postoptimization processing algorithm(POP).Combining spatial information of the strata,POP algorithm uses a two-dimensional sliding window to correct the prediction results,which enhances the model’s generalization ability for noisy data,reduces wrong classification results of samples and effectively improves the recognition effect.