Lithology Prediction Of Well Logging Based On Generative Adversarial Network And Transfer Learning

Machine learning is widely used in well logging,and lithology prediction has become an important research directio.In the thesis 5 machine learning algorithms are used to predict lithology,and the problems of small data and unbalanced lithologic label are exposed.Adversative neural network can generate a large number of high-quality data to expand the data volume.In this thesis,Info GAN is used to generate any number of specified types of logs according to lithology,layer and well location,and the generated logs conform to the distribution of real logs.However,directly adding generated data and training with real data is not enough to solve the data problems,transfer learning is used to extract general features of generate logging data and real logging data,using a large amount of generated well logging data to pretraining the optimized machine learning model,then according to the relatively small number of real logging data to fine tune the pretraining model,the problems of small samples and unbalanced labels are further overcome.The final results show that with Info GAN and transfer learning,the lithology prediction accuracy is improved from 83% to 87% for small samples problem.And the prediction effect of lithologic category with less proportion is greatly improved for the problem of label imbalance,the F1-Score of argillaceous siltstone,which accounts for7%,is increased from 0.47 to 0.71.The F1-Score of pebbly fine sandstone,which accounts for 10%,is increased from 0.73 to 0.81.It can be seen from the results that the method in this thesis can generate a large number of reliable labeled logging data under the condition of small set of samples and unbalanced lithologic label categories based on generative adversarial neural network,so after the precess of transfer learning,the effect of lithologic prediction can be effectively improved.