Data mining of well logs using decision tree based model
Thanyaboon Sudhasirikul, Waruntorn Kanitpanyacharoen · 2018
Well-logging is a geophysical survey which provides insights into subsurface geology of an interested borehole. However, the interpretation of well logging data is a time-consuming process and requires an interpreter’s experience. Quantitative approaches are attempted to improve time efficiency. This study uses machine learning model which is one of applied statistics to classify well log lithology and focuses on creating new features or columns to improve model performance which is not widely studied while many studies have been focused on choosing the best model. Data are from the National Petroleum Reserves in Alaska and consist of 11 wells which are 200,000 data in total. Ensemble tree model which shows outstanding performance in previous studies is used to created basic model to classify 3 rock types; mudstone, sandstone, and limestone. The performance of basic model reaches 57.6% of average F1 score and is further improved by incorporating four engineered features. The first feature is known as upsampling and downsampling which is used to manage imbalanced dataset. The second feature involves a calculation of M and N indexes from density, neutron, and sonic logs. To effectively compare and scale data from different wells, the third feature is created through standardization, normalization, and ranking. The fourth feature is developed to reduce data sensitivity and manage outliers by incorporating trimming and winsorizing methods. Results from a combination these features show that upsampling can not improve the model while the effect from downsampling is inconclusive. M and N indexes can slightly improve the model by 1%. The best model combination involves normalization and trimming, which improves the average F1 score by 1%. Hyperparameters of the best model combination such as gamma, max_depth, learning_rate, and n_estimators are tuned to develop the final model, which reaches 60.5% of average F1 score. Further improvement of the classification model can be done by incorporating relative position within a lithologic formation and marine/non-marine indicator.