Lithology Identification Oriented On Low Quality Well Logging Data Unbalancing Problem

Lithology identification based on conventional well-logging data is of great importance for geologic features characterization and reservoir quality evaluation in the exploration and production development of petroleum reservoirs.However,there are some limitations in the traditional lithology identification process:(1)It doesn't consider the effect of the data imbalance problem on the minority lithology identification.(2)Traditional machine learning lithology identification model is inefficient in hyper-parameters optimization stage.To solve above problems,this thesis takes an oil field in the sea area of China as the research object,takes data processing as starting point,uses machine learning technology as research method,carries out research on lithology identification systematically through well logging data preprocessing,data balance and data classification aspects.It aims to solve the well logging data imbalance and model hyper-parameters optimization inefficiency problems.This greatly improves the model identification performance and provides guidance for improving geologic features characterization and reservoir quality evaluation.Firstly,to reduce the error of the measurement process,the correlation function method is used for depth correction of the original logging data,the nonlinear wavelet transform threshold value method is used for data denoising,and the minimum maximum normalization method is used for data normalization.It provides real data support for subsequent data balancing process.Further,the synthetic minority over-sampling technique(SMOTE)is used to balance well logging data to solve data imbalance problem.New samples are synthesized by analyzing the minority class samples and then added to the data set,which effectively overcomes the shortcomings of the generalization lacking and low data utilization in the traditional data balance algorithms.It provides a basis for lithology identification model to distinguish minority lithology accurately.On that basis,to solve the hyper-parameters optimization inefficiency of traditional lithology identification models,this thesis researches and proposes a lithology identification method based on particle swarm optimization(PSO)and gradient boosting decision tree(GBDT)algorithms.By using the heuristic information of the optimization problem to guide the hyper-parameters search,the search domain is significantly reduced,and the optimization efficiency is greatly improved.The efficiency of the proposed method is verified by experimental results.After the correlation function method,the nonlinear wavelet transform threshold value method and the minimum maximum normalization method were used for the well logging data preprocessing,the whole lithology identification accuracy increased from 84.06% to 95.32%,which reduced the measurement errors of well logging data effectively.After the SMOTE method was used to balance well logging data,the identification accuracy on the tuff increased from 84.68% to 96.77%,which improved the minority lithology identification accuracy effectively.On this basis,the classification model hyper-parameters optimization efficiency could be improved greatly by combining PSO and GBDT algorithms.The whole lithology identification performance could also be improved and high accuracy,95.94%,could be achieved.