Lithology Identification Combining LightGBM-Recursive Feature Elimination and Residual CNN

A logging lithology identification model combining LightGBM-recursive feature elimination (RFE) and one-dimensional residual convolutional neural network (CNN) is constructed. Firstly, the outliers are identified by the box diagram and missing values are filled by Lagrange median, and then the Min-Max normalization is performed to improve data quality. Secondly, the contribution rate of logging parameters is analyzed using LightGBM. Furthermore, the optimal logging parameter subset is obtained by combining random forest-RFE and cross-validation method, effectively solve the over-fitting in feature selection. Finally, considering the complex nonlinearity and spatial correlation between logging data, the optimized subset of the simplest logging parameters is used as input to train the one-dimensional CNN, and the residual module is added to further improve the model accuracy and generalization ability. The experimental results on 10000 logging data with four lithology in the actual work area show that, compared with the other five methods, the feature selection method based on LightGBM and random forest-RFE has the highest lithology identification accuracy and data dimension reduction rate, and the lowest memory and CPU consumption. The accuracy and F1 value of the model on the test set are 92% and 83%, respectively.