亚洲免费av电影一区二区三区,日韩爱爱视频,51精品视频一区二区三区,91视频爱爱,日韩欧美在线播放视频,中文字幕少妇AV,亚洲电影中文字幕,久久久久亚洲av成人网址,久久综合视频网站,国产在线不卡免费播放

        ?

        Identification of low-resistivity-low-contrast pay zones in the feature space with a multi-layer perceptron based on conventional well log data

        2022-06-02 05:00:00LunGaoRanHongXieLiZhiXiaoShuaiWangChenYuXu
        Petroleum Science 2022年2期

        Lun Gao,Ran-Hong Xie,Li-Zhi Xiao,Shuai Wang,Chen-Yu Xu

        State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum (Beijing),Beijing,102249,China

        Keywords:Low-resistivity-low-contrast (LRLC) pay zones Conventional well logging Machine learning DBSCAN algorithm Multi-layer perceptron

        ABSTRACT In the early exploration of many oilfields,low-resistivity-low-contrast (LRLC) pay zones are easily overlooked due to the resistivity similarity to the water zones.Existing identification methods are model-driven and cannot yield satisfactory results when the causes of LRLC pay zones are complicated.In this study,after analyzing a large number of core samples,main causes of LRLC pay zones in the study area are discerned,which include complex distribution of formation water salinity,high irreducible water saturation due to micropores,and high shale volume.Moreover,different oil testing layers may have different causes of LRLC pay zones.As a result,in addition to the well log data of oil testing layers,well log data of adjacent shale layers are also added to the original dataset as reference data.The densitybased spatial clustering algorithm with noise (DBSCAN) is used to cluster the original dataset into 49 clusters.A new dataset is ultimately projected into a feature space with 49 dimensions.The new dataset and oil testing results are respectively treated as input and output to train the multi-layer perceptron(MLP).A total of 3192 samples are used for stratified 8-fold cross-validation,and the accuracy of the MLP is found to be 85.53%.

        1.Introduction

        A low-resistivity-low-contrast (LRLC) pay zone refers to a pay zone with a resistivity Rtclose to that of its adjacent water zone R0,i.e.,the formation resistivity index I (I=Rt/R0,Archie,1942) is low,and,usually,I <2.An LRLC pay zone is better considered as a relative concept than an absolute concept,as pay zones in different oilfields can have very different resistivity ranges.For example,the resistivities of LRLC pay zones and water zones in the Attaka Oilfield in Indonesia range from 2 to 5 Ohmm and 1 to 5 Ohmm,respectively,but the resistivities of LRLC pay zones and water zones in the Lakhmani Oilfield in India range from 10 to 30 Ohmm and 15 to 25 Ohmm,respectively(Worthington,1997).LRLC pay zones are widely distributed all over the world,including in Louisiana and Oklahoma in the United States,Hainan and Tarim in China,the Gulf of Mexico,the Middle East,etc.However,they are also easily overlooked by conventional log analysis due to the low contrast in resistivity.

        In recent years,some novel well logging methods have been demonstrated to perform well in the identification of LRLC pay zones,such as NMR well logging (Guru et al.,2008;Rima et al.,2012;Belevich and Bal,2018;Heidary et al.,2019) and pulse neutron well logging (Simpson and Menke,2010).However,the data acquired from these novel well logging methods are usually absent in many oilfields,and conventional well log data are the only complete data that can be used.

        Many researchers have sought model-driven solutions to the identification of LRLC pay zones for decades.One kind of method is to calculate the water saturation of a formation to determine whether or not it is hydrocarbon-bearing.The most widely used model for the calculation of water saturation from conventional well log data is the Archie's equation(Archie,1942).However,if an LRLC pay zone has similar resistivity to its adjacent water zone,the water saturation calculated by the Archie's equation will be overestimated.Hill and Milburn (1956) found that the relationship between rock resistivity and formation water resistivity is nonlinear due to the effect of clay.Based on this relationship,Waxman and Smits (1968) developed a model that explains the electrical conductivity of shaly sand.Givens (1987) proposed a conductive rock matrix model that separates the rock into two parallel components,i.e.,a conductive pore network containing fluids that are free to move,and the remainder of the rock,which may be conductive due to conductive minerals and immobile conductive water.With the deepening of the study on LRLC pay zones,recent methods tend to identify them based on the causes of the LRLC phenomenon.Worthington(1997)proposed a systematic workflow to interpret LRLC pay zones based on different causes,and different models were deployed for water saturation calculation.To identify LRLC pay zones caused by fresh drilling mud invasion,Li et al.(2010) established a three-step well logging identification method that contains conventional well logging cross-plot,drilling mud invasion factor analysis and well comparison;Pratama et al.(2017) proposed an integrated workflow that combines petrographical analysis,rock typing,and petrophysical analysis to calculate the shale volume,water saturation,and effective porosity,and eventually identifies the LRLC pay zones.Mashaba and Altermann (2015) considered clay-bound water and silt-bound water,which were the main causes of low resistivity in their study area,and improved the Archie's equation to calculate the water saturation.However,most of these model-driven methods are only locally applicable,and cannot yield satisfactory results when the causes of LRLC pay zones are complicated and the types of hydrocarbon layers (e.g.,gas layer,gas/oil layer,oil layer,etc.) are various.

        With the increasing scale and sophistication of data,the development of machine learning algorithms has shed light on complicated geophysical problems from a data-driven perspective(Bergen et al.,2019).Unsupervised algorithms like clustering and graphical models are usually used for well log data or seismic data preprocessing (Kang et al.,2019),reservoir multiscale modeling(Esmaeilzadeh et al.,2020),lithology identification (Feng et al.,2018;Li et al.,2021),and production schedule optimization for naturally fractured reservoirs (Liu and Forouzanfar,2017).Regarding supervised algorithms,artificial neural networks(ANNs)are a popular tool for solving both regression and classification problems.ANNs can be categorized into different types based on their structures.Convolutional neural networks (CNNs) have a unique convolutional layer and a pooling layer,which are suitable for processing image data.CNNs are therefore used for seismic horizon tracking (Yang and Sun,2020),seismic impedance inversion(Das et al.,2019),low-frequency noise suppression(Zhao et al.,2020),and seismic fault detection (Cunha et al.,2020;Yang et al.,2021).Recurrent neural networks (RNNs),including the long short-term memory(LSTM)network,have an advantage in dealing with sequential data,and are used to generate NMR T2distributions(Li and Misra,2019),optimize seismic waveform inversions (Sun et al.,2020),pore fluid identification (Zhou et al.,2021) and obtain pore size distributions (Li et al.,2020).The multi-layer perceptron (MLP) is widely applied together with some heuristic algorithms to bridge scales for seismic phase picking (Chai et al.,2020),seismic facies analysis (Bagheri and Riahi,2015),and shear wave travel time estimation (Anemangely et al.,2017).

        In the present work,the powerful data processing and information extraction abilities of machine learning algorithms are employed in the construction of an implicit relationship between conventional well log data and LRLC pay zones.The density-based spatial clustering algorithm with noise(DBSCAN)is first applied to cluster the input dataset.By calculating the Euclidean distances from the data to the cluster centroids,the dataset can be projected into a feature space with a higher dimension.Eventually,the optimally structured MLP is deployed to train the data for the identification and classification of LRLC pay zones.

        2.Methodology

        2.1.DBSCAN algorithm

        DBSCAN is a clustering algorithm based on the density distribution of a dataset,and it can exclude noise points in the dataset.Before the introduction of the DBSCAN algorithm,the following definitions are provided.

        Given a dataset D={x1,x2,…,xn},Nε(xi),the ε neighborhood of point xi,is defined as the collection of points in dataset D whose distance to point xiis less than ε,i.e.,Nε(xi)=Point xiis called the core point when its ε neighborhood contains no fewer than n points,i.e.,|Nε(xi)|≥n.Point xjis directly density-reachable to point xi(denoted as xj?xi) when point xjis in the ε neighborhood of point xi.Point xjis densityreachable to point xi(denoted as xj→xi) when xj?xp?…?xq?xi.Based on these definitions,the DBSCAN algorithm is described as follows:

        1) Determine the parameters ε and n;

        2) Find all the core points in dataset D and collect them into another dataset Ω;

        3) Randomly choose one core point xiin Ω,find its densityreachable points in D,and gather xiand its density-reachable points in dataset Cias one single cluster;

        4) Update Ω by excluding the core points contained in cluster Ci,i.e.,Ω=Ω-Ci;

        5) Go to step 3 if Ω is not empty.

        2.2.Multi-layer perceptron

        The basic structure of an MLP includes an input layer,several hidden layers,and an output layer.The more hidden layers an MLP has,the more powerful its learning ability will be,and the more risk of overfitting it will have.Fig.1 presents the structure of an MLP with two hidden layers.Given a labeled dataset A={(x1,y1),(x2,y2),…,(xn,yn)},its training steps in this MLP are described as follows.

        1) Input layer →hidden layer 1 →hidden layer 2

        where viis the output of hidden layer 1 and also the input of hidden layer 2,and f is the activation function of the neurons,common choices of which are the rectified linear unit,sigmoid,and tanh functions.

        Fig.1.Structure of an MLP with two hidden layers.

        2) Hidden layer 1→hidden layer 2→output layer

        The MSE function is then minimized to find the optimal weight and bias parameters w1,w2,b1,and b2of the MLP.

        For a classification problem,the cross-entropy(CE)between the predicted outputand the real output yican be calculated as follows:

        The CE function is then minimized to find the optimal weight and bias parameters w1,w2,b1,and b2of the MLP.

        To avoid the overfitting problem in the training steps,dropout regularization is applied to randomly neglect some neurons in the hidden layers by setting the weights of these neurons to 0.

        3.Case application

        Fig.2.The resistivity histogram of water layers and oil layers from 15 wells.

        The study area is located in W Oilfield,China.It includes two sets of Paleogene hydrocarbon-bearing strata.The formation lithology is dominated by pebbled sandstones with low porosity,low permeability,and a complex pore structure.Fig.2 presents the resistivity histogram of the water layers and oil layers from 15 wells in the study area.It is evident that the resistivities of the water layers are in the range of 6 Ohmm-40 Ohmm,and the resistivities of the oil layers are in the range of 12 Ohmm -36 Ohmm.Fig.3 exhibits the well logs of a water layer and an oil layer from well#2.The 1st track contains spontaneous potential log (SP),gamma ray log (GR) and caliper log (CAL).The 2nd track contains deep investigation induction log (ILD) and laterolog 8 log (LL8).The 3rd track contains density log (DEN),acoustic log (AC) and compensated neutron porosity log (CNL).The 4th track is the oil testing results.The contrast between the resistivity logs of these layers is low.Overall,it can be concluded that the reservoirs in the study area are typical LRLC pay zones.

        The workflow of this case study is shown in Fig.4.In section A,the causes of the LRLC pay zones in the study area are analyzed,which offers instructions for the data preprocessing described in section B.In section C,by deploying the DBSCAN algorithm,the dataset is projected into the feature space.In section D,the new dataset and the oil testing results are respectively treated as the input and output to train an MLP.

        3.1.Cause analysis of LRLC pay zones

        The water salinity in the study area has a very complicated distribution,and ranges from 421 mg/L -89997.5 mg/L.Fig.5 shows the water salinity histogram of the study area,from which we can see the water salinity distributes widely and unevenly.A water layer whose formation water has a very low water salinity will have a relatively high resistivity,whereas a hydrocarbon layer whose formation water has a high water salinity will have a relatively low resistivity.Fig.6 exhibits an example to show the influence of the water salinity on the resistivities of difference layers.In Fig.6,the layers of well#6 are tested as gas/oil layer and the water salinity in this depth interval is 12000 mg/L.The layers of well#13 are tested as water layer and the water salinity in this depth interval is 3381 mg/L.However,the resistivity ranges of well#6 and#13 are similar.Therefore,we can see the influence of the complex distribution of formation water salinity on the formation resistivity.

        From the cast thin sections in Fig.7,we can see there are many dissolution pores in the core samples.The pore size distributions show that there are two kinds of pore sizes.The first peak is around 0.01 μm,and the second peak is around 1 μm.Core porosity of the study area is in the range of 4.54%-12.7%and the porosity of most core samples is around 8%.Core permeability of the study area is in the range of 0.20 mD-2.49 mD and the permeability of most core samples is around 0.5 mD.Fig.8 is the porosity and permeability histograms that exhibits the specific porosity and permeability distribution of the core samples.Due to the small pore size,low porosity and low permeability,the irreducible water saturation can be very high.A high irreducible water saturation can result in the high conductivity of the pay zone.Based on the clay mineral X-ray diffraction analysis presented in Fig.9,the clay content is around 20%,and the mixed-layer illite/smectite,which has a very high cation exchange capacity,composes about 50% of all the clay mineral content.

        Above all,it can be concluded that the causes of the LRLC pay zones in this area are the complex distribution of formation water salinity,high irreducible water saturation due to micropores,and high shale volume.Due to the complicated causes,common methods cannot obtain satisfactory classification results.Fig.10 shows the cross-plots of ILD with other well logs.We can see that it is difficult to divide the oil layers and water layers in these crossplots.Therefore,it is necessary to apply the machine learning methods to establish an explicit and robust model for LRLC pay zone identification and classification.

        Fig.3.The well logs of a water layer and an oil layer from well #2.

        Fig.4.The workflow of this case study.

        Fig.5.The formation water salinity histogram.

        3.2.Data preprocessing

        Considering the complicated water salinity distribution and pore structure,different oil testing layers may have different LRLC causes.Therefore,it is reasonable to find the shale layers adjacent to the oil testing layers based on the gamma ray(GR)well logs and use them as reference data.As a result,14 dimensions are included in the dataset,which includes GRot,GRsh,SPot,SPsh,ACot,ACsh,CNLot,CNLsh,DENot,DENsh,ILDot,ILDsh,LL8ot,and LL8sh.Footnote‘ot’ stands for ‘oil testing layer’,and ‘sh’ stands for ‘a(chǎn)djacent shale layer’.

        Another problem is well logging data size mismatch between the oil testing layers and its adjacent shale layers.For example,an oil testing layer is 3 m thick (30 samples),whereas its adjacent shale layer is 2 m thick(20 samples).Due to the size mismatch,they cannot be combined as different dimensions of a dataset.To solve this problem,the well log value histogram of the shale layer is created and fitted with a Gaussian distribution.The mean value of the Gaussian distribution is then taken as the representative value and duplicated to make it have the same size as the oil testing layer.Eventually,the well log dataset can be obtained,as partly shown in Table 1.

        Fig.6.Comparison of the gas/oil layers of well #6 and the water layers of well #13.

        Fig.7.Cast thin sections of two core samples with their corresponding MICP curves and pore size distributions.

        Fig.8.The porosity and permeability histograms of the core samples in the study area.

        Fig.9.Clay mineral X-ray diffraction analysis of 22 core samples in well #7.

        The dataset is normalized base on Equation (6):

        where x is the sample to be normalized,and x′is the normalized sample.

        There are 10 types of layers in the study area based on the oil testing results,namely dry layers(DL),water layers(WL),oily water layers(OWL),low-production gas layers(LPGL),low-production oil layers (LPOL),low-production gas/oil layers (LPGOL),water/gas/oil layers(WGOL),water/oil layers(WOL),gas/oil layers(GOL),and oil layers(OL).With 3192 samples in the dataset,the oil testing results can be expressed as a matrix with a size of 3192×10.In the matrix,each type of layer has a unique corresponding vector.The vector of a dry layer,for example,is (1,0,0,0,0,0,0,0,0,0),which means that the probability of a dry layer is 1 and the probability of other types is 0.Table 2 reports the number of samples,vectors,and descriptions of each type of layer.

        Table 1Part of the well log dataset of an oil testing layer (only the AC,CNL,DEN logs are displayed).

        Overall,the normalized well log dataset (denoted as dataset P)with a size of 3192×14 can be obtained,as can the matrix of the oil testing results with a size of 3192 × 10.

        3.3.Application of the DBSCAN algorithm

        To apply the DBSCAN algorithm,the neighborhood parameters ε and n must first be determined.Generally,n=d+1,where d is the dimension of samples in dataset P;therefore,in this work,n=15.Moreover,ε can be determined by the silhouette coefficient:

        where Siis the silhouette coeffciient of xi,aiis the mean Euclidean distance from xito the other samples in its cluster,and biis the shortest Euclidean distance from xito the centroids of other clusters.After the silhouette coeffciients of all the samples in dataset P are calculated,a mean silhouette coefficientcan be obtained,which can be used to evaluate the clustering result.A higherindicates a better clustering result.Fig.11 presents thecurve with different values of ε in the range of 0.1-0.5.The optimal value of ε=0.19 can be obtained whenreaches its maximum.

        By applying ε=0.19 and n=15 to the DBSCAN algorithm,dataset P can be clustered into 49 clusters.The centroid cj,j=1,2,...,49,of each cluster is the arithmetic mean of the data samples in the cluster.

        Dataset P is then projected into the feature space of the clusters by calculating the Euclidean distance from each sample in dataset P to the centroids.The sample in the feature space can be expressed as

        From Equation(8),it can be seen that the samples in the feature space have 49 dimensions,whereas the samples in dataset P have 14 dimensions.It is known that a dataset projected into a higher dimension is more easily classified than the dataset with a lower dimension.Therefore,the dataset in the feature space is used as the input of the MLP in the next section.

        3.4.Application of the MLP

        Fig.10.Cross-plots of ILD with other well logs.

        The performances of MLPs with different structures are investigated to find the optimal structure,as shown in Table 3.Stratified 8-fold cross-validation is applied to split the input and output data into 8 folds.7 folds are treated as the training dataset,and the final fold is treated as the test dataset.After 8 loops,every fold has been treated as the test dataset,and the test accuracy of the whole dataset can be obtained.To be specific,the weight and bias parameters of the MLP in each loop are trained from the beginning to make sure they are not affected by the other loops.The CE function is used as the loss function,and the steepest gradient descent method is applied to minimize it.The second column in Table 3 reports the structures of the MLPs.As an example,“49-128-64-32-10” means that this MLP has 5 layers,and the numbers of neurons in each layer are 49,128,64,32,and 10,respectively.The complexity can be calculated based on the structure via Equation(10):

        where n is the number of layers in the MLP,liis the number of neurons in the ith layer,and rdris the dropout rate in each layer.Because the complexity is equal to the number of coefficients to be trained in an MLP,the higher the complexity,the longer the run time required by the MLP will be.For structures No.1-No.9,each MLP is run five times,and the mean accuracy is calculated.From Table 3,it can be seen that the accuracy increases with the complexity.Structure No.9 has the highest mean accuracy of 85.97%.However,structure No.4 has a mean accuracy of 85.53%,which is very close to that of structure No.9.Moreover,structure No.4 has lower complexity than structure No.9.Therefore,for a slight sacrifice of accuracy in exchange for greater efficiency,structure No.4 is assumed to have the best performance.

        Table 2Number of samples,vectors,and descriptions of each type of layers.

        To further analyze the prediction result,a confusion matrix is obtained to show the detailed numbers of actual and predicted types of samples.The precision and recall of each type of sample can be calculated from the confusion matrix,and are respectively defined as Equations (11) and (12):

        Fig.11.Silhouette coefficient with different values of ε.

        where TP is the number of true positive samples,F(xiàn)P is the number of false positive samples,and FN is the number of false negative samples.Table 4 presents the confusion matrix of the 2nd MLP from structure No.4.Taking the water layer in Table 4 as an example,the TP of the water layer is 761 (the 4th row of the 3rd column in Table 4),the FP of the water layer is 59(the sum of the 3rd column in Table 4 excluding its 4th row),and the FN of the water layer is 230(the sum of the 4th row in Table 4 excluding its 3rd column).

        From Table 4,it can be seen that the recalls of the water layer and water/oil layer are relatively low,because many water layer samples are predicted as low-production gas layer,low-production oil layer,and water/oil layer samples.Consequently,the precisions of these three types of samples are affected.

        Table 3Performances of MLPs with different structures.

        Table 4Confusion matrix of the 2nd MLP from structure No.4.

        Table 5Performances of MLP with different structures in Case 1.

        Table 6Performances of MLP with different structures in Case 2.

        To intuitively show the effectiveness of this method,we exhibit the well logs,oil testing results and predicted results of well#2,as shown in Fig.12.In Fig.12,the 1st track contains SP,GR and CAL.The 2nd track contains ILD and LL8.The 3rd track contains DEN,AC and CNL.The 4th track is the oil testing results.The 5th track is the predicted results.Layers where the predicted results are inconsistent with the oil testing results are marked by symbol ‘*’.We can see the oil testing results of most layers have a very good agreement with its corresponding predicted results.Thus,the effectiveness of the proposed method is well demonstrated.

        4.Discussions

        4.1.The necessity of including the well log data of adjacent shale layers and projecting the dataset into the feature space

        Two cases are designed for comparison with the proposed method:

        Case 1:Only the well log data of the oil testing layers are included in the dataset;

        Case 2:The well log data of the oil testing layers and adjacent shale layers are included in the dataset.

        The difference between Case 1 and Case 2 is that the well log data of adjacent layers are included in Case 2.The difference between Case 2 and the proposed method is that the dataset is not projected into the feature space in Case 2.

        Fig.12.The well logs,oil testing results and predicted results of well #2.

        Tables 5 and 6 exhibit the performances of MLPs with different structures for Case 1 and Case 2.The highest mean accuracy of Case 1 achieved by the MLP with structure No.5 is 73.10%,and the highest mean accuracy of Case 2 achieved by the MLP with structure No.5 is 81.42% from structure No.5,which is 8% higher than that for Case 1.Also,the mean accuracy of the proposed method is 85.53%,which is 4%higher than that of Case 2.Overall,the necessity of including the well log data of the adjacent shale layers and projecting the dataset into the feature space is revealed.

        4.2.Results comparison with support vector classification

        Support vector classification (SVC) is a kind of support vector machine (SVM),but SVC is used to solve classification problems.SVC can be classified into two categories,linear SVC and kernel SVC.The common choices of kernel functions are polynomial function,radial basis function (RBF) and sigmoid function.We apply these methods above to compare their predicted results with the results of the proposed method,as shown in Table 7,where the regularization parameter used in the SVC methods is 0.1.From Table 7,we can see that the accuracy of these method is smaller than the accuracy of the proposed method,which is 85.53%.

        Table 7Predicted results of SVC.

        4.3.Analysis of the reason for the poor prediction results of water layers and water/oil layers

        The recalls of the water layers and water/oil layers are around 70%,which are relatively low compared to the recalls of other types of layers.Specifically,there are 991 samples of water layers,95 of which are predicted as water/oil layers;moreover,there are 361samples of water/oil layers,46 of which are predicted as water layers,and 55 of which are predicted as oil layers.

        The possible reason for this phenomenon is the scale difference between the well log data and oil testing results.The depth interval of well log data in the study area is 0.1 m,whereas the oil testing results are an overall evaluation of a whole layer,which is usually several meters thick.For instance,the oil testing results in well#2 indicate that a water/oil layer has an oil production of 0.5 m3per day,a water production of 3.07 m3per day,and is 3.2 m thick,which means that this layer contains 32 pieces of well log data.Therefore,the well log data below the water/oil interface are very likely to be predicted as water layers,while the well log data above the water/oil interface are very likely to be predicted as oil layers.

        4.4.Problems to be optimized

        The determination of neighborhood parameters is very essential for DBSCAN algorithm.This paper offers a general but not optimal procedure to calculate these parameters.There are also infinite choices for the number of layers and neurons in an MLP.Perfect balance between accuracy and efficiency is still needed to be found.

        Another problem is the category imbalance in the training dataset.The water layer samples and gas/oil layer samples take up half of the whole sample numbers,which may affect the predicted results.Oversampling the minority categories or undersampling the majority categories is a feasible way to adjust the imbalance.

        5.Conclusions

        A large amount of core analysis shows that the causes for those LRLC pay zones include complex distribution of formation water salinity,high irreducible water saturation due to micropores,and high shale volume.A machine learning method for LRLC pay zones identification and classification is proposed.After the result analysis and discussion,several conclusions are drawn:The proposed machine learning methods can effectively learn the relation between conventional well logging data and oil testing results.This method is applicable when the causes are complicated;Considering the complexity and accuracy,the optimal structure of the MLP is ‘49-512-256-128-64-10’ with a mean accuracy of 85.53%;By comparing the results of Case 1 method to the results of Case 2 method,we can see there is a great accuracy improvement after adding the data of adjacent shale layers into the dataset;By comparing the results of proposed method to the results of Case 2 method,we can see the necessity of projecting the dataset into feature space.

        Acknowledgements

        The work was funded by the Strategic Cooperation Technology Projects of CNPC and CUPB (ZLZX2020-03).

        Nomenclature

        AC Acoustic logs

        ANN Artificial neural network

        CE Cross entropy

        CNL Compensated neutron porosity logs

        CNN Convolutional neural network

        DBSCAN Density-based spatial clustering algorithm with noise

        DEN Density logs

        DL Dry layer

        FN False negative

        FP False positive

        GOL Gas/oil layer

        GR Gamma ray logs

        ILD Deep investigation induction logs

        LL8 Laterolog 8 logs

        LPGL Low-production gas layer

        LPGOL Low-production gas/oil layer

        LPOL Low-production oil layer

        LRLC Low-resistivity-low-contrast

        LSTM Long short-term memory

        MICP Mercury injection capillary pressure

        MLP Multi-layer perceptron

        MSE Mean squared error

        OL Oil layer

        OWL Oily water layer

        RBF Radial basis function

        RNN Recurrent neural network

        SP Spontaneous potential logs

        SVC Support vector classification

        SVM Support vector machine

        TP True positive

        WL Water layer

        WGOL Water/gas/oil layer

        WOL water/oil layer

        校园春色日韩高清一区二区| 四虎精品视频| 国产精品白浆视频一区| 自拍偷拍另类三级三色四色| 国产成人自拍视频播放| 国产电影一区二区三区| 又黄又爽的成人免费视频| 亚洲va欧美va人人爽夜夜嗨| 国产91在线精品观看| 日韩乱码人妻无码系列中文字幕| 一品二品三品中文字幕| 久久综合色鬼| 亚洲产在线精品亚洲第一页| 久久狼精品一区二区三区| 亚洲熟妇无码一区二区三区导航| 尤物yw无码网站进入| 好看午夜一鲁一鲁一鲁| 丝袜美腿国产一区二区| 日本高清视频永久网站www| 亚洲精品成人网久久久久久| 国产日韩AV无码免费一区二区| 国产精品一区二区三区av在线| 免费无码精品黄av电影| 免费做爰猛烈吃奶摸视频在线观看| 狠狠综合亚洲综合亚色| 国产精品毛片av毛片一区二区| 牛牛在线视频| 四月婷婷丁香七月色综合高清国产裸聊在线 | 国产成人精品一区二三区孕妇| 亚洲精品无码久久久久av老牛| 91av小视频| 国产一区二区三区乱码在线| 中文字幕一区二区中文| 熟妇丰满多毛的大隂户| 国产成人亚洲精品电影| 男人一插就想射的原因| 免费观看交性大片| 日韩AV不卡六区七区| 日本久久精品在线播放| 国产在线一区二区三区四区 | 久久久久久无码AV成人影院|