Abstract
As the accumulation of logging data and the enhancement of computational power, machine learning technology has been progressively applied to logging interpretation field such as lithology identification. However, in traditional data-driven lithology identification model, the implied variation information of logging curve and coupling relationships among features are not fully mined. Additionally, feature extraction cannot filter out information redundancy and noise. We propose logging data representation enhancement approach for lithology identification based on feature decomposition, selection and transformation, converting the raw logging curves into an improved high dimensional representation with more effective information and less noise. Local mean decomposition is used to extract the variation information of logging curves from multiple depth scales and add them to the features of adjacent samples. Considering the different contribution of features to lithology identification, an optimized feature selection method based on Shapley additive explanation is designed to reduce redundant and noisy information in logging data. To mine the complementary information among sequence features, a representation learning model integrating feature transformation and lithology classification is developed by multi-grained scanning and cascading extreme learning machine. The effectiveness and generalization of the proposed approach are verified on the baseline and shale oil field datasets. The results show that the proposed approach can make the logging data acquire more valid information through representation enhancement, which helps to achieve high-accuracy lithology identification.
NSTL