Lithologic Identification Model Based On Machine Learning

Carbonate oil and gas reservoirs are another important oil-producing oil and gas reservoir after structural oil and gas reservoirs with good physical properties.The lithologic identification of carbonate reservoirs is the key and difficult point of reservoir geological interpretation.Current hot research directions in the field of artificial intelligence——Machine Learning,provides new technical means for solving the problem of lithologic identification.This paper compares and summarizes machine learning lithology recognition models from different angles,and improves active learning,which is applied to the selection of training samples for lithology recognition.The paper systematically describes and analyzes the supervised learning and unsupervised learning models in machine learning,and compares and analyzes the factors that affect the lithology recognition effect of the supervised learning model: These include:(1)The application of linear normalization and non-linear normalization in logging data are compared,The changes in the recognition effect of different machine learning models caused by the different standardization of logging data are summarized.(2)The correlation between different attributes of lithology are studied.Different isometric reduction methods are used to process logging data and applied to different machine learning models to compare the recognition results.(3)The difference in lithology recognition effect caused by the parameter settings of different machine learning models is analyzed,And the lithology recognition machine learning model is summarized.The acquisition of lithology labels requires drilling to take cores,which is costly.Therefore,reducing the amount of lithology labels is a key research direction.The paper uses active learning combined with Euclidean metric to screen unlabeled samples: Active learning screens out the highvalue samples in the unlabeled sample set to form the sample set to be labeled,and further use the Euclidean metric to judge the similarity of the samples in the sample set to be labeled,and improve the diversity of labeled samples.The results show that under the same labeled sample size,the machine learning model trained on the training set combined with the active learning screening of the Euclidean metric has better lithology recognition accuracy than random selection and traditional active learning.