Application Study Of Heterogeneous Machine Learning Integration In Logging Curve Recovery And Lithology Recognition

Reservoir description is a technique involving the quantitative distribution of reservoir characteristics,which can predict the reserve growth of mature and marginal fields at low cost and risk while ensuring the real geological model and petrophysical parameters.Lithology identification,as an important branch of reservoir description,is in the stage of rapid development and technological reform with the maturity of logging technology in recent years.The logging curve has the advantages of high vertical resolution,good continuity and convenient data collection.Therefore,lithology identification based on seismic logging data is an important research topic at present,which is conducive to guiding the remaining oil mining.According to the actual production demand of oil field,this paper studies lithology identification from three aspects: outliers removal,log curve recovery and log integrated network lithology identification.(1)Wrong measurement data caused by imperfect equipment and record.A method for eliminating outliers based on dichotomous space pauta criterion is proposed.This method solves the problem that there are a large number of extreme outliers in the logging data,and USES the traditional pauta criterion method,which causes the mean and variance to be greatly affected by outliers in the statistical process,and eliminates the case of high error rate.In this method,the distribution parameters of the original data are estimated more reliably and accurately by the binary space representation of the original data.(2)For the distortion or missing of logging data in some well sections due to man-made or instrument failure.This paper presents a recovery model of missing log data by integrating deep feature learning network and crossover network.The model is composed of two networks,one of the crossover network is composed of multiple layers,it explicitly in automatic way cross application feature information,each layer based on the existing high-order interaction,effective learning characteristics,highly nonlinear interaction relationship between recycling residual thought deep tectonic network at the same time,reducing model complexity.In the other network,feature selection is carried out through the constructed tree model,and the sparse vectors obtained from the tree model are transformed into dense vectors through the embedded layer as the input of the neural network.The nonlinear expression ability of the model is improved through the combination of network,the interaction relationship between features is captured effectively,and the irrationality of artificial feature selection is reduced.Finally,the fusion network model is used to recover the missing log data with high precision.(3)In order to improve the accuracy of lithology identification,this paper proposes a lithology identification method based on integrated network structure.Firstly,the statistical features were extracted from the logging curve,and then the corresponding low-frequency subband was calculated by using the discrete cosine transform(DCT)to obtain the low-frequency extraction features.In order to reduce the impact of noise signal on lithology identification,difference calculation was carried out on the features to construct the characteristic attributes of difference value.In addition,each attribute was clustered according to the unsupervised clustering algorithm of k-means to construct the bucket splitting feature.In this paper,a semi-supervised neural network is used to correct noise labels.Then build integration network structure of lithologic identification methods,the method adopts the Stacking of integration framework model of the multivariate statistical analysis,within DNN,Ada Boost and SVM as base in the integrated network structure learning,GBDT as secondary learning of study,using five cross validation way to training the model,finally get more lithology classification results.