Tight sandstone macroscopic bedding seam logging identification method based on machine learning
By using machine learning-based decision trees, support vector machines, and principal component analysis, combined with core and imaging logging data, the problem of identification error of high-angle fractures and near-horizontal bedding fractures in tight sandstone reservoirs was solved, achieving highly accurate bedding fracture identification with a single sand body and single well accuracy rate of over 80%.
Patent Information
- Application Number
- CN202310588584.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-23
- Publication Date
- 2024-11-26
AI Technical Summary
Existing fracture logging identification methods have significant identification errors in tight sandstone reservoirs, making it difficult to accurately distinguish between high-angle fractures and near-horizontal bedding fractures. Furthermore, existing methods mainly rely on complex nonlinear relationships and weak logging responses, which are easily masked by argillaceous interlayers.
A machine learning-based approach, combining decision tree, support vector machine, and principal component analysis, is employed. By establishing a sample training dataset and utilizing core, imaging logging, and conventional logging data, the nonlinear response of fractures at different angles is automatically learned and identified. The identification results are then corrected using imaging logging images.
It achieves efficient identification of macroscopic bedding fractures in dense sandstone, with a single sand body and single well matching rate of over 80%, improving the accuracy and reliability of identification.
Smart Images

Figure CN119027799A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of petroleum geological analysis, and in particular to a method for well logging identification of macroscopic bedding fractures in dense sandstone based on machine learning. Background Art
[0002] Bedding fractures are commonly found in sedimentary reservoirs, and their significance for oil and gas exploration and development is becoming increasingly prominent. Therefore, effectively identifying the development degree of bedding fractures is of great significance for oil and gas exploration and development. Although many important theoretical results have been achieved in the research on fracture logging identification, most of them are aimed at structural fractures, and there are few logging identification methods specifically for bedding fractures.
[0003] The existing fracture logging identification methods mainly include: comprehensive index method, fractal theory, wavelet analysis method, main controlling geological factor analysis method, etc. Since the logging response caused by fractures is mostly a complex nonlinear relationship, and the opening of bedding fractures in the ground is usually very small, the logging response caused is very weak, which is easily covered by the logging response caused by mud interlayers; and the main controlling geological factor analysis method mainly fits the mathematical relationship between the degree of development of bedding fractures and the main controlling geological factors, and then fits the relationship between the main controlling geological factors and sensitive logging curves. It is an indirect identification method of bedding fractures. Therefore, these existing methods have large identification errors and cannot accurately distinguish high-angle fractures from near-horizontal bedding fractures. At present, there is still a lack of methods to accurately identify bedding fractures in tight sandstone reservoirs. Summary of the invention
[0004] In order to solve the problems existing in the prior art, the present invention provides a method for well logging identification of macroscopic bedding fractures in dense sandstone based on machine learning, which is used to solve the problem that there is a large identification error in the prior art and that high-angle fractures and near-horizontal bedding fractures cannot be accurately distinguished. The method of the present invention has the advantages of the ability to capture complex nonlinear relationships, the ability to automatically learn models, and the ability to mine hidden patterns in data. The method mainly solves the problem of nonlinear well logging response caused by fractures of different angles through decision tree algorithm, support vector machine algorithm and principal component analysis method, and can automatically classify fractures of different angles according to the logging curve value.
[0005] To achieve the above object, the present invention provides the following technical solutions:
[0006] A method for identifying macroscopic bedding fractures in tight sandstone by logging based on machine learning comprises the following steps:
[0007] Step 1, establishing a sample training data set based on core logging, imaging logging and conventional logging data;
[0008] Step 2: Standardize the sample training data set, then combine the logging curves in pairs to establish a cross plot, analyze the logging response trends of the three types of data sets on the same cross plot, and finally select the fracture sensitivity logging curve according to the response trend;
[0009] Step 3, processing the sample training data set through the decision tree algorithm to identify macroscopic bedding fractures;
[0010] The sample training data set is processed by the support vector machine algorithm to identify fractures and non-fracture segments, and to identify macro-bedding fracture segments and inclined fracture segments;
[0011] The sample training data set was processed by principal component analysis to identify macroscopic bedding fractures;
[0012] Step 4: The overlapping area of bedding fractures identified by the decision tree algorithm, support vector machine algorithm and principal component analysis method is used as the final bedding fracture identification area.
[0013] Preferably, in step 1, the development sections of fractures at different angles are determined on the drilling core; then the imaging logging images of the corresponding depths are calibrated, and the imaging logging image features of fractures at different angles are analyzed and summarized; the natural fractures appearing in the imaging logging images are used to calibrate the conventional logging curves of the corresponding depths and read the corresponding data values to establish a sample training data set.
[0014] Preferably, in step 1, outlier detection is performed on the sample training data set;
[0015] The specific process includes the following: first, the mudstone interlayers are eliminated according to the empirical value of the mud content in the study area, and then the lithology, porosity, resistivity and conductivity logging curve values are read in the bedding fractures, inclined fractures and non-fracture development sections in the tight sandstone, and the labels of "bedding fractures", "inclined fractures" and "non-fractures" are attached. Then, the box plot of the anomaly detection module is used to detect outliers, detect abnormal logging values in the data set, and eliminate abnormal logging data.
[0016] Preferably, the standardization process is to normalize the well logging data in the sample training data set and place them in one dimension.
[0017] Preferably, in step 3, the sample training data set is processed by a decision tree algorithm, and the training data set is randomly divided into a modeling data set and a verification data set. The decision tree algorithm will automatically divide the numerical ranges of different conventional logging curves into three types of labels, and give a tree-like discrimination flowchart to identify macroscopic bedding fractures.
[0018] Preferably, in step 3, the sample training data set is processed by the support vector machine algorithm, the inclined fractures and macroscopic bedding fractures are uniformly labeled with "fractures", and the training data set is randomly divided into a modeling data set and a verification data set. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and provide a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function, so as to identify fractures and non-fracture sections.
[0019] Preferably, in step 3, the sample training data set is processed by a support vector machine algorithm to establish a training data set containing only macroscopic bedding fractures and inclined fractures; then the training data set is randomly divided into a modeling data set and a verification data set, and the support vector machine algorithm automatically fits a suitable dividing line for the two types of labels and gives a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function, so as to identify macroscopic bedding fracture sections and inclined fracture sections.
[0020] Preferably, in step 3, the sample training data set is processed by principal component analysis, and the training data set is randomly divided into a modeling data set and a validation data set. The principal component analysis algorithm will obtain three principal components PC 1 、PC 2 、PC 3 , PC 1 、PC 2 、PC 3 The intersection charts were established by combining two of them, and the logging response trends of the three types of data were analyzed on the principal component intersection charts. Different data boundaries were fitted as discriminant formulas to identify macroscopic bedding fractures.
[0021] Preferably, the formula of the principal component analysis algorithm is
[0022] PC i =a 1i ×Z x1 +a 2i ×Z x2 +.....+a pi ×Z xp
[0023] where x = {x 1 ,x 2 …, x p}={AC,DEN…,CNL} is a data set consisting of multiple logarithmic curves, a pi is the eigenvector corresponding to the eigenvalue of the covariance matrix Σ of x, Z xp is the normalized value of the original logarithmic curve.
[0024] Preferably, in step 4, the recognition result is corrected by imaging logging images. If the error exceeds a preset range, the parameters of the decision tree algorithm, support vector machine algorithm and principal component analysis method and the ratio of the modeling data set to the verification data set are continuously adjusted and optimized.
[0025] Compared with the prior art, the present invention has the following beneficial technical effects:
[0026] The present invention provides a method for well logging identification of macroscopic bedding fractures in dense sandstone based on machine learning. The nonlinear well logging response problem caused by fractures at different angles is solved by decision tree algorithm, support vector machine algorithm and principal component analysis method, and fractures at different angles can be automatically classified according to the well logging curve value. The method has the advantages of the ability to capture complex nonlinear relationships, the ability to automatically learn models, and the ability to mine hidden patterns in data. The bedding fractures identified by this method are mainly macroscopic bedding fractures that can be observed in imaging well logging images. Therefore, its accuracy can be verified by imaging well logging. Verification in the Sanbianbei area shows that the consistency rate of single sand bodies and single wells is higher than 80%. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A schematic diagram of a method for well logging identification of macroscopic bedding fractures in dense sandstone based on machine learning according to the present invention;
[0028] Figure 2 It is the imaging logging bedding fracture characteristics and the corresponding conventional logging curve abnormal reflection map;
[0029] Figure 3 This is an outlier detection diagram using a box plot (quartiles);
[0030] Figure 4 Cross-plot for well logging data;
[0031] Figure 5 It is a schematic diagram of the decision tree algorithm;
[0032] Figure 6 This is a partial process display of the decision tree algorithm;
[0033] Figure 7 Processing for support vector machine algorithm;
[0034] Figure 8 This is the result of support vector machine algorithm processing;
[0035] Fig. 9 This is the intersection plate of PC1, PC2 and PC2, PC3 of the extended group in the Sanbianbei area;
[0036] Fig.10 Machine learning identification for conventional well logging and imaging logging verification results. DETAILED DESCRIPTION
[0037] The present invention is further described in detail below in conjunction with specific embodiments, which are intended to explain the present invention rather than to limit it.
[0038] The present invention provides a method for well logging identification of macroscopic bedding fractures in dense sandstone based on machine learning, comprising the following steps:
[0039] (1) Based on core, imaging and conventional logging, a sample training data set is established. The fracture development sections at different angles are determined on the cores taken from the wells. Then, the imaging logging images at the corresponding depths are calibrated, and the imaging logging image features of fractures at different angles are analyzed and summarized. In order to exclude the possible influence of core-induced fractures on the results, the natural fractures appearing in the imaging logging images are used to calibrate the conventional logging curves at the corresponding depths and read the corresponding data values to establish a sample training data set.
[0040] (2) Detection of outliers in the sample training data set. First, mudstone interlayers were removed based on the empirical value of the shale content (SH) in the study area. Then, the lithology, porosity, resistivity, and conductivity logging curve values were read in the bedding fractures, inclined fractures, and non-fracture development sections of the tight sandstone, and the labels of "bedding fractures", "inclined fractures", and "non-fractures" were attached. Then, the box plot (quartiles) of the anomaly detection module in Python, Matlab, and other software was used to detect outliers, detect abnormal logging values in the data set, and eliminate abnormal logging data.
[0041] (3) The logging data are standardized, and then the logging curves are combined in pairs to establish an intersection chart. The logging response trends of the three types of data sets are analyzed on the same intersection chart. Finally, the fracture-sensitive logging curves are selected according to the response trends, such as AC, DEN, SP, CILD, CLL8, RLLS, and RLLD.
[0042] (4) Decision tree algorithm to identify macroscopic bedding fractures. The decision tree algorithm in software such as MATLAB and Python was used to process the training data set. The training data set was randomly divided into a modeling data set and a validation data set. The decision tree algorithm automatically divided the numerical ranges of different conventional logging curves into three types of labels and gave a tree-like discrimination flow chart. Due to differences in rock components, the limits of conventional logging values for discriminating bedding fractures in different study areas were different.
[0043] (5) Support vector machine algorithm to identify fractures and non-fracture segments. The support vector machine algorithm in software such as MATLAB and Python is used to process the training data set. The inclined fractures and macroscopic bedding fractures are uniformly labeled as "fractures", and the training data set is randomly divided into a modeling data set and a validation data set. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and provide a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function.
[0044] (6) Support vector machine algorithm identifies macroscopic bedding fracture segments and inclined fracture segments. First, a training data set containing only macroscopic bedding fractures and inclined fractures is established; then the training data set is randomly divided into a modeling data set and a validation data set. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and give a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function.
[0045] (7) Identify macroscopic bedding fractures using principal component analysis. Use the principal component analysis algorithm in software such as MATLAB and Python to process the training data set. The training data set is randomly divided into a modeling data set and a validation data set. The principal component analysis algorithm will obtain three principal components, PC 1 、PC 2 、PC 3 , and the corresponding calculation formula:
[0046] ①PC i =a 1i ×Z x1 +a 2i ×Z x2 +.....+a pi ×Z xp
[0047] In the formula, x={x 1 ,x 2 …, x p}={AC,DEN…,CNL} is a data set consisting of multiple logarithmic curves, a pi is the eigenvector corresponding to the eigenvalue of the covariance matrix Σ of x, Z xp is the normalized value of the original logarithmic curve. 1 、PC 2 、PC 3 The intersection chart is established by combining two of them, and the logging response trends of the three types of data are analyzed on the principal component intersection chart, and different data boundaries are fitted as the discriminant formula. The more obvious the conventional logging response of the fracture is, the greater the absolute value of the coefficient in the principal component PCi.
[0048] (8) Verification and application of macroscopic bedding fracture logging identification method. The overlapping area of bedding fractures identified by decision tree algorithm, support vector machine algorithm and principal component analysis method is used as the final bedding fracture identification area, and the identification results are corrected by imaging logging images. If there is a large identification error, the parameters of decision tree algorithm, support vector machine algorithm and principal component analysis method, as well as the ratio of modeling data set to verification data set are continuously adjusted and optimized.
[0049] Example
[0050] like Figure 1 As shown, the present invention provides a method for well logging identification of macroscopic bedding fractures in dense sandstone based on machine learning, which specifically includes the following steps:
[0051] Step 1: Establish a sample training data set based on core, imaging logging and conventional logging. Determine the fracture development sections at different angles on the cores; then calibrate the imaging logging images at the corresponding depths, and analyze and summarize the imaging logging image features of fractures at different angles ( Figure 2 ); In order to eliminate the possible influence of core-induced fractures on the results, the natural fractures appearing in the imaging logging images are used to calibrate the conventional logging curves at the corresponding depths and read the corresponding data values to establish a sample training data set. The logging data refers to the specific AC, DEN, SP, etc. The training data set refers to the logging values of AC, DEN, and SP at a certain depth, such as the values corresponding to 1500m and 1600m.
[0052] Step 2: Detect outliers in the sample training data set. Remove mudstone interlayers based on the empirical value of mud content (SH) in the study area (the SH limit of sandstone and mudstone in the Yanchang Formation of the Ordos Basin is generally 40), and then read the lithology, porosity, resistivity and conductivity logging curve values in the bedding fractures, inclined fractures and non-fracture development sections of the tight sandstone, and attach "bedding fractures", "inclined fractures" and "non-fractures" labels. There are at least 50 samples for each type of label, otherwise the conformity rate of the recognition results will be reduced; then use the box plot (quartile) of the anomaly detection module in Python, matlab and other software to detect outliers, detect abnormal logging values in the data set, and remove abnormal logging data ( Figure 3 ).
[0053] Step 3: standardize the logging data, then combine the logging curves in pairs to establish a cross plot, analyze the logging response trends of the three types of data sets on the same cross plot, and finally select the fracture sensitive logging curves according to the response trend, such as AC, DEN, SP, CILD, CLL8, RLLS, RLLD, etc. Figure 4As shown. The standardization process is to put the logging values of AC, DEN, etc. into one dimension, which is similar to normalization. Because the AC value may be 180 and the DEN value may be 2.5, the AC and DEN logging values are changed to about -1 to 1 through standardization. Selecting fracture sensitive logging curves is to select logging curves that will cause the logging curve value to increase or decrease significantly due to the presence of fractures.
[0054] Step 4, the decision tree algorithm identifies macroscopic bedding fractures. The decision tree algorithm in software such as MATLAB and Python is used to process the training data set. The rule of thumb in the machine learning monograph "Predictive Analysis and Data Mining-RapidMiner Implementation" by American Vijaykotu Bala Deshpande is to use the "stratified sampling" module to use 70% to 90% of the original data as the modeling data set, and the rest as the test verification data set, that is, the ratio of the modeling data set to the verification data set is 7:3 to 9:1. Since the response effects of different logging parameters are not completely synchronized, in order to achieve better results, after testing, the training data set containing about 150 samples is randomly divided into a modeling data set and a verification data set in a ratio of 7:3, which is the best effect. According to this number of samples and ratio, the decision tree algorithm will automatically divide the numerical ranges of different conventional logging curves into three types of labels, and give a tree-like discrimination flowchart ( Figure 5 and Figure 6 ). Due to the differences in rock composition, the limits of conventional logging values for identifying bedding fractures vary in different study areas. The advantage of the decision tree algorithm is that it is very intuitive and there is no need to standardize the logging curves.
[0055] Step 5, support vector machine algorithm identifies fractures and non-fracture segments. The support vector machine algorithm in software such as matlab and Python is used to process the training data set. According to the data set ratio test results mentioned in (4), the inclined fractures and macro-bedding fractures are uniformly labeled "fractures". The training data set containing 150 samples is randomly divided into a modeling data set and a verification data set in a ratio of 7:3. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and give a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function. After calculation, the extended group Y in the Sanbianbei area is Y = 0.191×CAL-0.409×DT+0.228×GR+0.123×RFOC-0.016×RILD+0.031×RILM+1.318×SP-0.637( Figure 7 and Figure 8 ).
[0056] Step 6: Use the support vector machine algorithm to identify macroscopic bedding fractures and inclined fractures. First, establish a training data set that contains only macroscopic bedding fractures and inclined fractures. Based on the data set ratio test results mentioned in (4), randomly divide the training data set containing about 100 samples into a modeling data set and a validation data set in a ratio of 7:3. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and provide a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function.
[0057] Step 7: Identify macroscopic bedding fractures using principal component analysis. The principal component analysis algorithm in software such as MATLAB and Python is used to process the training data set. According to the data set ratio test results mentioned in (4), the training data set containing about 150 samples is randomly divided into a modeling data set and a validation data set in a ratio of 7:3. The principal component analysis algorithm will obtain three principal components: 1 、PC 2 、PC 3 , and the corresponding calculation formula:
[0058] PC i =a 1i ×Z x1 +a 2i ×Z x2 +.....+a pi ×Z xp
[0059] In the formula, x={x 1 ,x 2 …, x p}={AC,DEN…,CNL} is a data set consisting of multiple logarithmic curves, a pi is the eigenvector corresponding to the eigenvalue of the covariance matrix Σ of x, Z xp is the normalized value of the original logarithmic curve. The mathematical formulas for PCi in the tight sandstone of the Yanchang Formation in the Sanbianbei area of the Ordos Basin are shown in Figures 2 to 4. When applied to other areas, the selection and coefficients of conventional logging curves may change:
[0060] PC 1 =0.074×SP+0.068×GR-0.548×DEN+0.563×CNL+0.609×AC-0.025×(CILD
[0061] -CLL8)
[0062] PC 2 =0.638×SP+0.376×GR+0.293×DEN+0.199×CNL-0.064×AC-0.568×(CILD
[0063] -CLL8)
[0064] PC 3 =-0.051×SP+0.842×GR+0.143×DEN-0.040×CNL+0.101×AC+0.535×(CIL
[0065] D-CLL8)
[0066] PC 1 、PC 2 、PC 3 Combine two by two to create an intersection map ( Fig. 9 ), the logging response trends of the three types of data are analyzed on the principal component intersection chart, and different data boundaries are fitted as the discriminant formula. The more obvious the conventional logging response of the fracture is, the greater the absolute value of the coefficient in the principal component PCi.
[0067] Step 8, verification and application of macroscopic bedding fracture logging identification method. The overlapping area of bedding fractures identified by decision tree algorithm, support vector machine algorithm and principal component analysis method is used as the final bedding fracture identification area, and the identification result is corrected by imaging logging image. If there is a large identification error, the parameters of decision tree algorithm, support vector machine algorithm and principal component analysis method and the ratio of modeling data set to verification data set are continuously adjusted and optimized, such as Fig.10 shown.
[0068] In view of the current situation that there are few logging identification methods for bedding fractures and the errors are large, the present invention proposes a logging identification method for bedding fractures based on machine learning. The method has the advantages of being able to capture complex nonlinear relationships, automatically learn models, and mine hidden patterns in data. The method mainly solves the nonlinear logging response problem caused by fractures of different angles through decision tree algorithm, support vector machine algorithm and principal component analysis method, and can automatically classify fractures of different angles according to the logging curve values.
[0069] The bedding fractures identified by this method are mainly macroscopic bedding fractures that can be observed in imaging logging images, so its accuracy can be verified by imaging logging. Verification in the Sanbianbei area shows that the coincidence rate between single sand bodies and single wells is higher than 80%.
[0070] Single sand body coincidence rate of bedding fracture identification results
[0071]
[0072] Single well coincidence rate of bedding fracture identification results
[0073]
[0074] At present, similar technologies for bedding fracture identification mainly focus on three aspects: conventional logging identification, seismic prediction of development density and bedding fracture effectiveness evaluation. Similar inventions have not been reported and are not comparable.
Claims
1. A method for identifying macroscopic bedding fractures in tight sandstone by logging based on machine learning, characterized in that: The following steps are included: Step 1, establishing a sample training data set based on core logging, imaging logging and conventional logging data; Step 2: Standardize the sample training data set, then combine the logging curves in pairs to establish a cross plot, analyze the logging response trends of the three types of data sets on the same cross plot, and finally select the fracture sensitivity logging curve according to the response trend; Step 3, processing the sample training data set through the decision tree algorithm to identify macroscopic bedding fractures; The sample training data set is processed by the support vector machine algorithm to identify fractures and non-fracture segments, and to identify macro-bedding fracture segments and inclined fracture segments; The sample training data set was processed by principal component analysis to identify macroscopic bedding fractures; Step 4: The overlapping area of bedding fractures identified by the decision tree algorithm, support vector machine algorithm and principal component analysis method is used as the final bedding fracture identification area.
2. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 1, the development sections of fractures at different angles are determined on the drilling core; then the imaging logging images of the corresponding depths are calibrated, and the imaging logging image features of fractures at different angles are analyzed and summarized; the natural fractures appearing in the imaging logging images are used to calibrate the conventional logging curves of the corresponding depths and read the corresponding data values to establish a sample training data set.
3. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 1, outlier detection is performed on the sample training data set; The specific process includes the following: first, mudstone interlayers are removed according to the empirical value of mud content in the study area; then, the lithology, porosity, resistivity and conductivity logging curve values are read in the bedding fractures, inclined fractures and non-fracture development sections in the tight sandstone, and the labels of "bedding fractures", "inclined fractures" and "non-fractures" are attached; then, the box plot of the anomaly detection module is used to detect outliers, detect abnormal logging values in the data set, and remove abnormal logging data.
4. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: The standardization process is to normalize the logging data in the sample training data set and put it in one dimension.
5. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 3, the sample training data set is processed by the decision tree algorithm, and the training data set is randomly divided into a modeling data set and a validation data set. The decision tree algorithm will automatically divide the numerical ranges of different conventional logging curves into three types of labels, and give a tree-like discrimination flowchart to identify macroscopic bedding fractures.
6. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 3, the sample training data set is processed by the support vector machine algorithm, and the inclined fractures and macro-bedding fractures are uniformly labeled as "fractures". The training data set is randomly divided into a modeling data set and a verification data set. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and give a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function, so as to identify fractures and non-fracture sections.
7. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 3, the sample training data set is processed by the support vector machine algorithm to establish a training data set containing only macroscopic bedding fractures and inclined fractures; then the training data set is randomly divided into a modeling data set and a verification data set. The support vector machine algorithm will automatically fit a suitable dividing line for the two types of labels and give a calculation formula for the boundary line. The selection and coefficient of the conventional logging curve in the calculation formula are related to the number of samples and the selected calculation function, so as to identify macroscopic bedding fracture sections and inclined fracture sections.
8. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 3, the sample training data set is processed by principal component analysis, and the training data set is randomly divided into a modeling data set and a validation data set. The principal component analysis algorithm will obtain three principal components PC1, PC2, and PC3. PC1, PC2, and PC3 are combined in pairs to establish an intersection chart. The logging response trends of the three types of data are analyzed on the principal component intersection chart, and different data boundaries are fitted as discriminant formulas to identify macroscopic bedding fractures.
9. The method for identifying macroscopic bedding fractures in dense sandstone by well logging based on machine learning according to claim 1, characterized in that: The formula for the principal component analysis algorithm is PC i =a 1i ×Z x1 +a 2i ×Z x2 +.....+a pi ×Z xp where x={x1,x2…,x p }={AC,DEN…,CNL} is a data set consisting of multiple logarithmic curves, a pi is the eigenvector corresponding to the eigenvalue of the covariance matrix Σ of x, Z xp is the normalized value of the original logarithmic curve.
10. The method for identifying macroscopic bedding fractures in tight sandstone by well logging based on machine learning according to claim 1, characterized in that: In step 4, the recognition results are corrected by imaging logging images. If the error exceeds the preset range, the parameters of the decision tree algorithm, support vector machine algorithm and principal component analysis method, as well as the ratio of the modeling data set to the verification data set, are continuously adjusted and optimized.
