Research On Lithology Identification Method Based On CNN-LSTM Network And Ensemble Learning

Lithologic identification is an important step in oil and gas exploration.In the process of logging,there is often a problem of lack of continuity in the logging characteristic data due to machine failure or manual operation error.In addition,due to the complex and changeable geological environment,the lithology data often show non-equilibrium characteristics,which may lead to the unsatisfactory effect of lithology identification.At present,to solve the problem of continuous loss of logging characteristics,the commonly used completion methods include constant interpolation and regression interpolation,etc.,but these methods do not fully consider the change of logging curve with depth.To solve the problem of lithology data imbalance,the commonly used equalization methods include sampling method,etc.How to determine the sampling proportion is the key to obtain highquality equalization data.Therefore,it is the key to improve the accuracy of lithology identification to solve the problem of continuous missing of logging curve features and unbalanced lithology data.Firstly,to solve the problem of continuous missing of logging curve features,a completion algorithm based on convolutional cyclic neural network was proposed.The algorithm first uses convolutional neural network to extract the important features of logging curve,and then trains the extracted features with long and short time memory cyclic neural network.This process takes into account the spatial relationship of different logs,increases the dimension of combination characteristics,and takes full advantage of the sequence characteristics of logs that vary with depth.Secondly,a piecewise random sampling method is proposed to solve the problem of lithology data imbalance.The algorithm first divides multiple intervals equally between the maximum and minimum number of lithology categories,and randomly determines a sampling point in each interval to reduce human interference.Then,the number of each sampling point and all lithologic categories are compared in a circular manner to determine whether each lithologic data is over-sampled or under-sampled,so that multiple balanced data sets can be constructed.After several balanced data sets are trained by the model,the best balanced data set is selected.Finally,the missing curve feature completion algorithm based on convolutional cyclic neural network and piecewise random sampling method are combined to design the lithology identification model based on Bayesian optimization of lightweight gradient elevator.The accuracy of lithology identification is improved by solving the problem of continuous missing of logging curve features and unbalance of lithology data.