Dense glutenite micro-crack automatic identification method based on electric imaging logging

Abstract

The invention provides an electric imaging logging-based dense glutenite micro-crack automatic identification method, which comprises the steps of 1, preprocessing electric imaging logging original data, 2, carrying out interpolation restoration on blank stripes between polar plates in an electric imaging image, 3, carrying out sliding processing on a whole-well-periphery dynamic imaging image, and 4, carrying out sliding processing on the whole-well-periphery dynamic imaging image. The method comprises the following steps of 1, setting a feature extraction algorithm and corresponding different radiuses, 2, marking a training set, 3, setting a classifier, 4, training the classifier, 8, calculating parameters of the microcracks, and 9, generating a result map. The method provides powerful technical support for reservoir evaluation, exploration and development of compact glutenite.

Description

Technical Field

[0001] This invention relates to the fields of oil and gas well logging evaluation, image information processing, mathematical geology, computer vision and machine learning, and in particular to an automatic identification method for microfractures in tight sandstone and conglomerate based on electrical imaging logging. Background Technology

[0002] As older oilfields in eastern China enter a period of high water cut, conventional oil and gas discoveries are becoming increasingly difficult. Tight oil and gas (including tight sandstone and conglomerate) has become the most important replacement area and production growth point, holding strategic significance for ensuring energy security. Tight sandstone and conglomerate reservoirs are widely distributed in several major oil and gas basins, possessing enormous resource potential and gradually becoming one of the main forces for increasing oilfield reserves and production. However, due to their complex and heterogeneous characteristics in terms of sedimentary genesis, rock composition, and pore structure, the evaluation and prediction of oil and gas "sweet spots" are extremely difficult, posing a challenge to exploration and development. Natural microfractures are commonly developed in tight sandstone and conglomerate, which not only increase the effective reservoir space but also greatly improve the reservoir's permeability and reduce the difficulty of fracturing development. Therefore, accurate and efficient identification and evaluation of microfractures has become a critical technical bottleneck that urgently needs to be addressed in the exploration and development of tight sandstone and conglomerate reservoirs.

[0003] According to the current oil and gas industry standard "SY / T 5386-2010 Detailed Rules for Calculating Proven Reservoirs in Fractured Oil (Gas) Reservoirs", for non-carbonate rocks, microfractures are defined as fractures with a width between 0.001 and 0.01 mm, while fractures with a width less than 0.001 mm are called hyperfractures. Due to their extremely small width, microfractures and hyperfractures are difficult to observe with the naked eye, even in firsthand drilling core samples. Thin core sections are typically prepared, and further analysis using methods such as cast thin sections, fluorescence thin sections, scanning electron microscopy, and CT is used for further determination. However, the number of core samples is usually limited, resulting in insufficient coverage of the reservoir section. Furthermore, the narrow scale of the samples leads to a lack of representativeness and very weak statistical significance, and may even cause the missed fracture development zones. In addition, the process of preparing and analyzing rock samples is costly, time-consuming, and prone to subjective errors (such as inadvertently creating artificial cracks and misidentifying them as natural fractures). Compared with core testing, evaluating fractures using well logging techniques offers advantages such as speed, continuity, in-situ accuracy, and high efficiency. It can cover the entire reservoir section, revealing the vertical distribution of fractures under real formation conditions, making large-scale, regional fracture evaluation possible. Among these, electrical imaging logging offers the highest resolution and coverage, generating intuitive images of the formation within the wellbore, which is beneficial for fracture identification. Furthermore, fracture parameter calculation formulas established using mature physical models and numerical simulation techniques can quantitatively evaluate fractures with widths down to the micrometer level.

[0004] However, current fracture identification in electrical imaging logging still primarily relies on human-computer interaction, requiring interpreters to manually identify and pick up fracture trajectories based on their personal experience and understanding. This approach is highly subjective, making it difficult to guarantee the consistency of interpretation standards and the reliability of interpretation results. Moreover, for fracture-developed layers, manual picking is extremely time-consuming and labor-intensive, failing to meet the needs of efficient exploration. Therefore, previous researchers have conducted extensive work to improve the accuracy and efficiency of fracture identification in electrical imaging logging by using traditional feature extraction algorithms or establishing corresponding machine learning models. However, traditional feature extraction techniques rely on the design of complex features, making it difficult to guarantee the universality and robustness of their algorithms. Meanwhile, pure machine learning algorithms require manual labeling of image samples beforehand to form a rich and accurate training dataset, which increases the difficulty of modeling and introduces uncertainty. Furthermore, and more importantly, these studies mainly focus on macroscopic fractures, while explorations specifically targeting microfractures have been almost entirely unreported to date.

[0005] Accurate fracture parameter calculation is crucial for the exploration and development of tight sandstone and conglomerate. It not only quantifies fracture development characteristics through core parameters such as fracture density (FD), fracture length (FL), fracture width (FW), and fracture porosity (FP)—for example, identifying high-quality "sweet spots" with a fracture density (FD) > 5 fractures / m and a fracture porosity (FP) > 0.1%, and distinguishing between effective and ineffective fractures with permeability based on fracture width (FW), thus avoiding wasted development resources—but also provides a basis for optimizing fracturing schemes. For instance, it determines the number of fracturing stages based on fracture length (FL) and fracture density (FD), and selects fracturing fluid viscosity based on the actual fracture width (FW), reducing single-well fracturing costs. Furthermore, continuous parameter curves can guide well network deployment optimization, ultimately contributing to the efficient development of tight oil and gas—a strategic energy replacement resource—and providing technical support for increasing oilfield reserves and production.

[0006] However, in current production practice, after identifying fractures based on electrical imaging logging images, it is necessary to first introduce conventional resistivity logging data, manually establish the calibration relationship between electrical imaging logging data and conventional logging resistivity, then convert the electrical imaging logging data into resistivity values, and finally combine these with mud resistivity measurements and parameters related to the logging instrument structure, substituting them into empirical formulas to solve for the corresponding fracture parameters. These processes require manual intervention from engineers and involve numerous additional parameters, resulting in not only many uncertainties but also low efficiency in processing and interpretation.

[0007] Furthermore, compared to macroscopic fractures, microfractures in dense sandstone and conglomerate are smaller in scale and have weaker signals. In electrical imaging logging images, they appear as "narrow, dark lines" with little grayscale difference from the image background, making them easily masked by background noise. Simultaneously, microfractures are often discontinuously distributed with no clear trajectory, making them difficult to capture visually or using traditional feature extraction algorithms. If manual picking is used, the densely distributed, irregular, and discontinuous fracture trajectories in the microfracture development section will create an extremely heavy processing workload. If only simple machine learning algorithms are used, the difficulty in sample labeling makes it hard to guarantee the quality of the training set, resulting in insufficient generalization ability of the model. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an automatic identification method for microfractures in tight sandstone and conglomerate based on electro-imaging logging. This invention, based on the processing of electro-imaging logging data, integrates feature extraction and machine learning algorithms to establish an automatic identification and parameter calculation method for microfractures, providing strong technical support for reservoir evaluation and exploration and development of tight sandstone and conglomerate. This invention does not distinguish between microfractures and ultramicrofractures, but collectively refers to fractures with a calculated width of less than 0.01 mm (i.e., 10 μm) as microfractures.

[0009] The present invention adopts the following technical solution: An automatic identification method for microfractures in tight sandstone and conglomerate based on electrical imaging logging includes: Step 1. Preprocess the raw electrical imaging logging data, complete the correction of depth, amplitude and azimuth, perform dynamic enhancement processing to generate dynamic electrical imaging images, and use gradient color scales for color matching to form pseudo-color images.

[0010] Step 2. Interpolate and repair the blank stripes between the electrodes in the electrical imaging image to generate a dynamic imaging map of the entire well circumference.

[0011] Step 3. Perform sliding processing on the whole-well dynamic imaging map of the tight sandstone and conglomerate reservoir section according to a certain window length (the length of the current processing depth segment) and step size (the distance between adjacent processing windows), and cut it into multiple sample images of fixed size.

[0012] Step 4. Training Set Labeling: Using an image labeling tool, carefully trace the micro-crack outlines of some sample images and delineate some sample images (including gravel, macro-cracks, and sandy skeletons, etc.) to form closed polygonal regions. Label the pixels within these regions uniformly as either "crack" or "background." Merge and save all labeled images as the training set for machine learning.

[0013] Step 5. Set the feature extraction algorithm and corresponding radii to extract crack information in the image in multiple dimensions and at multiple scales.

[0014] In step 5, various features characterize the image from different dimensions and are key objects of image analysis in the field of computer vision. This invention uses a Gaussian blur algorithm as a denoiser to eliminate background noise in the image, and uses a Sobel filter, Gaussian difference, and Hessian matrix eigenvalues ​​as edge detectors to identify the boundaries of cracks in the image. At the same time, six different radii σ are used to extract features at different scales in the image.

[0015] Different radii σ represent the neighborhood range involved in feature operations on each pixel. For example, when σ=2, Gaussian blur and other operations will be performed within a square region centered on the target pixel with a side length of 2 pixels, and the result will be assigned to the current pixel. The six different radii are σ=0, 1, 2, 4, 8, and 16; σ=0 indicates that the feature operation is performed only on the target pixel itself.

[0016] 1) Gaussian blur algorithm: The image is convolved with Gaussian kernels of different radii σ to generate a blurred version of the original image (the larger the σ, the more blurred the image).

[0017] 2) Sobel filter: Based on the Gaussian blur algorithm, it calculates the gradient of the intensity (values ​​of the R, G, and B channels) of each pixel; 3) Difference of Gaussians: Subtraction is performed between every two Gaussian blurred images; 4) Hessian Matrix Eigenvalues: Based on Gaussian blur, a Hessian matrix is ​​generated at each pixel, and then its corresponding eigenvalues ​​are calculated. Assume the second derivatives of the target pixel in its neighborhood of σ in the x, xy, yx, and yy directions are respectively... Then its Hessian matrix is Then, the following eight eigenvalues ​​are calculated:

[0018] The modulus of the eigenvalue vector of a matrix: Sum of the elements on the main diagonal of the matrix: The value of the matrix determinant: First eigenvalue: Second eigenvalue: Normalized feature spacing squared (used to quantify the degree of difference between two feature values ​​at the calibrated scale): Normalized eigenvalue difference intensity (the final intensity of the original eigenvalue differences after scaling): The direction of the eigenvector: Under the above settings, each image sample generates 70 corresponding feature maps: including 1 original image, 5 Gaussian blur maps (without using σ=0), 10 Gaussian difference maps (the pairwise subtraction of the 5 Gaussian blur maps), 6 Sobel maps (corresponding to 6 different σ values), and 48 Hesse feature maps (corresponding to 8 features and 6 different σ values).

[0019] Step 6. Classifier Setup: This invention employs the Fast Random Forest algorithm for ensemble learning, initialized with 200 trees. Each node randomly selects two features, and the depth of each tree is unlimited. The number of concurrent threads is set to 16. Each tree in the random forest is constructed by randomly sampling samples or features from an independent random distribution vector. When the random forest model makes predictions, each tree makes an independent decision, outputting a predicted class. The final result is determined by a majority vote of all trees.

[0020] Step 7. Combining the above design of feature extraction, manual labeling and machine learning, the manually labeled training set images are used as input to the random forest model classifier to train the classifier. The trained model is saved and applied to unlabeled images to achieve automatic identification of microcracks and form the corresponding crack binarized images.

[0021] Step 8. Calculate the parameters of the microcracks.

[0022] 1) For a single microcrack in the image, its trajectory is fitted with a sine function. The number of crack pixels along the trajectory direction is the apparent length of the microcrack, denoted as . The average number of crack pixels in the direction perpendicular to the trajectory is the micro-crack apparent width, set as... The total number of pixels in a single crack is the apparent area of ​​the microcrack, set as [value missing]. According to the definition, To simplify the calculation, the apparent length of the microcrack is first determined. and the apparent area of ​​microcracks Then solve for the apparent width of the microcrack. ; 2) Unit conversion for single crack parameters: Crack apparent length, crack apparent width, and crack apparent area are converted to meters (m), mm, and meters, respectively. 2 Unit, i.e. in, The distance between any two adjacent pixels in an electrical imaging image is defined according to the principle of electrical imaging logging instruments. ; 3) Since the fracture width presented in electrical imaging logging images is generally larger than the actual fracture width under formation conditions, specialized software is used to calculate the actual width of each fracture. Then, the apparent fracture width calculated by the above method is fitted with the corresponding calculation result from the specialized software using a polynomial: In the formula, The apparent width of the crack, i.e., the width of the single crack mentioned above. , These are the corresponding calculation results from professional software. The fitting coefficients are denoted as .

[0023] A unified fitting formula for the target block is established by selecting typical microfracture intervals, and then applied to the conversion of microfracture width in each well section.

[0024] 4) Calculate the output statistical curve The parameters mentioned above for each image sample are statistically analyzed according to a certain window length and step size, and the following curve is calculated and output: Fracture density (FD): The total number of fractures per unit well section, expressed in units of 1 / m, i.e.: To count the total number of cracks inside the window, The length of the statistical window, in meters; Fracture length (FL): The sum of fracture lengths per unit area of ​​wellbore, measured in meters (m / m). 2 Or 1 / m; that is: in, Let be the length of the i-th crack, in meters. The width of the developed image of the electro-imaging logging is the perimeter of the wellbore, expressed in meters (m). This represents the total area of ​​the well wall within the statistical window.

[0025] Crack Width (FW): The average width of all cracks within the statistical window, in mm. in, The width of the i-th crack is in mm; Crack area ratio (FP): The percentage of the total area of ​​a crack in an image, i.e.: in, Let m be the apparent area of ​​the i-th crack. 2 As a unit; 9) Generation of the result map: By stitching together the various binarized image samples, a fracture image of the entire well section is formed. This image is then combined with the fracture parameter curves and the dynamic electrical imaging map to create a microfracture processing result map of the tight sandstone and conglomerate section. Based on this, a segmented evaluation of the degree of microfracture development can be given.

[0026] The beneficial effects of this invention are: (1) First-ever automated identification of micro-fractures in tight sandstone and conglomerate: Existing technologies mostly focus on macro-fractures, and there are almost no publicly reported technologies for the automated identification of micro-fractures (including ultra-micro-fractures) with a width of less than 0.01 mm (10 μm). This invention, through the fusion design of "multi-dimensional feature extraction + machine learning", develops a technical solution specifically for the subtle features of micro-fractures (such as weak gray-scale differences and discontinuous trajectories), and successfully achieves automated identification of micro-fractures, filling the technical gap in intelligent evaluation of micro-fractures in tight sandstone and conglomerate, and providing key technical support for the prediction of reservoir "sweet spots".

[0027] (2). Solving the inherent defects of traditional feature extraction and machine learning: On the one hand, avoiding the problem of traditional feature extraction technology "relying on complex feature design and poor universality" - by combining multiple algorithms such as Gaussian blur, Sobel filtering, and Hessian matrix eigenvalues, and using 6 different radii (σ=0,1,2,4,8,16), multi-scale and full-dimensional crack information capture is achieved, and the robustness of the algorithm is significantly improved; on the other hand, reducing the "sample labeling dependence" of pure machine learning - by manually labeling typical samples (such as 49 image samples) to construct a training set, and combining the ensemble learning advantages of the fast random forest algorithm (200 trees, 16 threads concurrent), the workload of sample labeling is reduced while improving the model's generalization ability and avoiding the uncertainty of a single algorithm.

[0028] (3) Replacing manual interaction and reducing time and labor costs: Traditional electrical imaging logging fracture identification relies on manual picking by interpreters, which is highly subjective and inefficient (fracture development sections need to be identified pixel by pixel). This invention achieves full-process automated design of "data preprocessing - automatic sample cutting - model prediction - parameter calculation", which can complete the identification of micro fractures in the entire well section without manual intervention. Taking the X well implementation case as an example, the parameter setting of 0.5m window length and 0.125m step length can realize continuous parameter curve calculation, which is dozens of times more efficient than manual processing, meeting the time requirements of large-scale and regional exploration of tight oil and gas.

[0029] (4) Continuous coverage of the entire well section avoids the limitations of core analysis: Core sampling suffers from problems such as "low coverage, weak representativeness, and high cost" (e.g., thin sections of core can only reflect local point characteristics and easily miss fracture development zones). Based on the advantage of the continuity of the entire well section of electrical imaging logging data, this invention reveals the vertical distribution law of microfractures through electrode blank strip interpolation repair (generating a dynamic map of the entire well perimeter) and sliding cutting of samples (achieving coverage without omissions), which has stronger statistical significance; at the same time, it avoids the misjudgment of artificial cracks in rock sample preparation, and the reliability of the results is significantly higher than that of traditional core analysis.

[0030] (5) Multi-dimensional parameter calculation to quantify microcrack development characteristics: This invention not only achieves the "presence or absence" identification of microcracks, but also outputs four core evaluation indicators through a scientific parameter calculation model, forming a complete quantitative evaluation system: 1) Fracture density (FD): The frequency of fracture development is reflected by statistically counting the total number of fractures per unit well section (1 / m); 2) Fracture length (FL): Based on the sum of fracture lengths per unit wellbore area (m / m²), it characterizes the fracture propagation capacity; 3) Fracture width (FW): Corrected to the true formation width by polynomial fitting of apparent width (such as the fitting formula of well X), ensuring accurate classification of microfractures (<0.01mm); 4) Fracture porosity (FP): The percentage of fracture area directly reflects the degree of improvement in the effective reservoir space and permeability.

[0031] (6) High consistency of results, reducing subjective error: Traditional manual interpretation is affected by differences in experience, and it is difficult to unify the interpretation standards for the same well section. This invention ensures the consistency of results when used by different well sections and different interpreters by fixing the feature extraction algorithm and random forest model parameters (such as randomly selecting 2 features for each node and unlimited tree depth); at the same time, the visualization of the binarized image (fractures are red and the background is green) and the result image (overlaying dynamic electro-imaging image and parameter curve) can intuitively verify the matching relationship between microfracture development and parameters (such as high parameters in the developed section and low parameters in the underdeveloped section), further improving the evaluation accuracy.

[0032] (7) Evaluation of “sweet spots” in tight sandstone and conglomerate reservoirs: Tight oil and gas are the core replacement resources after the high water cut period in old oilfields in my country, and microfractures are the key factors to improve reservoir permeability and reduce the difficulty of fracturing development. This invention can provide a basis for screening “sweet spot” sections (well-developed microfractures and strong permeability) by accurately identifying microfractures and quantifying their parameters, thus helping oilfields increase reserves and production.

[0033] (8) The technology is highly universal and can be extended to multiple oil and gas basins: This invention takes Well X (FMI instrument, 8.5in wellbore) as an example, but its technical solution does not depend on specific instruments or wellbore size. By adjusting the preprocessing parameters of the electro-imaging data (such as depth / amplitude / azimuth correction algorithm), sample cutting window length / step size, and fitting coefficient (to adapt to the formation characteristics of different blocks), it can be extended to the evaluation of tight sandstone and conglomerate reservoirs in multiple major oil and gas basins in my country, providing technical support for national energy security. Attached Figure Description

[0034] Figure 1 For the restoration and segmentation of dynamic images from electro-imaging.

[0035] Figure 2 For feature extraction of electro-imaging image samples, (a) is the original image, (b) Gaussian blur (σ=16); (c) Difference of Gaussians (σ=2,16); (d) Sobel filter (σ=4); (e) Modulus of the eigenvalue vector of the Hessian matrix (σ=8); (f) Value of the determinant of the Hessian matrix (σ=16); (g) First eigenvalue of the Hessian matrix (σ=4); (h) Sum of the main diagonal elements of the Hessian matrix (σ=8).

[0036] Figure 3 The image shows the labels for the training set. The areas marked with green polygons represent the background image, while those marked with small red rectangles and filled with blue represent micro-cracks.

[0037] Figure 4 This is a random forest classifier model. The micro-cracks / background in the image represent the judgment and classification of each pixel.

[0038] Figure 5 This is a flowchart illustrating the overall process of automatic identification and calculation of microcracks.

[0039] Figure 6 The results show the microcrack prediction results for a single sample image. (a) is the dynamic image of the entire wellbore by electrical imaging and the superimposed microcrack identification results (red part); (b) is the binary image of the microcrack, with a green background and red microcracks.

[0040] Figure 7 This is a graph showing the fitted relationship between the apparent width and actual width of the microcrack.

[0041] Figure 8 This is a diagram showing the results of the treatment of microfractures in the dense sandstone and conglomerate of Well X. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0043] like Figure 5 As shown, an automatic identification method for microfractures in tight sandstone and conglomerate based on electrical imaging logging includes: Step 1. Preprocess the raw electrical imaging logging data, complete the correction of depth, amplitude and azimuth, perform dynamic enhancement processing to generate dynamic electrical imaging images, and use gradient color scales for color matching to form pseudo-color images.

[0044] Step 2. Interpolate and repair the blank stripes between electrodes in the electrical imaging image to generate a dynamic imaging map of the entire well circumference (e.g., Figure 1 (As shown).

[0045] Step 3. The full-well-circumferential dynamic imaging map of the tight sandstone and conglomerate reservoir section is processed by sliding according to a certain window length (the length of the current processing depth segment) and step size (the distance between adjacent processing windows), cutting it into multiple sample images of fixed size. (e.g.) Figure 1 (As shown).

[0046] Step 4. Training Set Labeling (Manual Labeling): Using an image labeling tool, carefully trace the micro-crack outlines of some sample images and delineate some sample images (including gravel, macro-cracks, and sandy skeletons, etc.) to form closed polygonal regions. Label the pixels within these regions uniformly as either "crack" or "background" (e.g., ...). Figure 3 (As shown). All labeled images are merged and saved as a training set for machine learning.

[0047] Step 5. Set the feature extraction algorithm and corresponding radii to extract crack information in the image in multiple dimensions and at multiple scales.

[0048] In step 5, various features characterize the image from different dimensions and are key objects of image analysis in the field of computer vision. This invention uses a Gaussian blur algorithm as a denoiser to eliminate background noise in the image, and uses a Sobel filter, Gaussian difference, and Hessian matrix eigenvalues ​​as edge detectors to identify the boundaries of cracks in the image. At the same time, six different radii σ are used to extract features at different scales in the image.

[0049] Different radii σ represent the neighborhood range involved in feature operations on each pixel. For example, when σ=2, Gaussian blur and other operations will be performed within a square region centered on the target pixel with a side length of 2 pixels, and the result will be assigned to the current pixel. The six different radii are σ=0, 1, 2, 4, 8, and 16; σ=0 indicates that the feature operation is performed only on the target pixel itself.

[0050] 1) Gaussian blur algorithm: The image is convolved with Gaussian kernels of different radii σ to generate a blurred version of the original image (the larger the σ, the more blurred the image).

[0051] 2) Sobel filter: Based on the Gaussian blur algorithm, it calculates the gradient of the intensity (values ​​of the R, G, and B channels) of each pixel; 3) Difference of Gaussians: Subtraction is performed between every two Gaussian blurred images; 4) Hessian Matrix Eigenvalues: Based on Gaussian blur, a Hessian matrix is ​​generated at each pixel, and then its corresponding eigenvalues ​​are calculated. Assume the second derivatives of the target pixel in its neighborhood of σ in the x, xy, yx, and yy directions are respectively... Then its Hessian matrix is Then, the following eight eigenvalues ​​are calculated:

[0052] The modulus of the eigenvalue vector of a matrix: Sum of the elements on the main diagonal of the matrix: The value of the matrix determinant: First eigenvalue: Second eigenvalue: Normalized feature spacing squared (used to quantify the degree of difference between two feature values ​​at the calibrated scale): Normalized eigenvalue difference intensity (the final intensity of the original eigenvalue differences after scaling): The direction of the eigenvector: Under the above settings, each image sample generates 70 corresponding feature maps (e.g., Figure 2 As shown): including 1 original image, 5 Gaussian blur maps (without using σ=0), 10 Gaussian difference maps (the pairwise subtraction of the 5 Gaussian blur maps), 6 Sobel maps (corresponding to 6 different σ values), and 48 Hesse feature maps (corresponding to 8 features and 6 different σ values).

[0054] Step 6. Classifier Setup (Machine Learning): This invention employs the Fast Random Forest algorithm for ensemble learning (e.g., Figure 4 As shown in the diagram, the system is initialized with 200 trees. Each node randomly selects two features, and the depth of each tree is unlimited. The number of concurrent threads is set to 16. Each tree in the random forest is constructed by randomly sampling samples or features from an independent random distribution vector. When the random forest model makes a prediction, each tree makes an "independent decision" and outputs a predicted category. The final result is determined by the "majority election" of all trees.

[0055] Step 7. Combining the above feature extraction (70 feature images), manual labeling, and machine learning design, the manually labeled training set images are used as input to a random forest classifier to train the classifier. This allows the classifier to automatically learn the distribution of "background" pixels and "crack" pixels in the feature space of each sample image (i.e., the corresponding 70 feature images mentioned above), thereby forming a corresponding mapping criterion (not an explicit mathematical expression, but rather a parameter matrix corresponding to the model's weights and biases). The trained model is saved and applied to unlabeled images to achieve automatic identification of micro-cracks, forming the corresponding crack binarized images.

[0056] Step 8. Calculate the parameters of the microcracks.

[0057] 1) For a single microcrack in the image, its trajectory is fitted with a sine function. The number of crack pixels along the trajectory direction is the apparent length of the microcrack, denoted as . The average number of crack pixels in the direction perpendicular to the trajectory is the micro-crack apparent width, set as... The total number of pixels in a single crack is the apparent area of ​​the microcrack, set as [value missing]. According to the definition, To simplify the calculation, the apparent length of the microcrack is first determined. and the apparent area of ​​microcracks Then solve for the apparent width of the microcrack. ; 2) Unit conversion for single crack parameters: Crack apparent length, crack apparent width, and crack apparent area are converted to meters (m), mm, and meters, respectively. 2 Unit, i.e. in, The distance between any two adjacent pixels in an electrical imaging image is defined according to the principle of electrical imaging logging instruments. ; 3) Since the fracture width presented in electrical imaging logging images is generally larger than the actual fracture width under formation conditions, specialized software is used to calculate the actual width of each fracture. Then, the apparent fracture width calculated by the above method is fitted with the corresponding calculation result from the specialized software using a polynomial: In the formula, The apparent width of the crack, i.e., the width of the single crack mentioned above. , These are the corresponding calculation results from professional software. The fitting coefficients are denoted as .

[0058] A unified fitting formula for the target block is established by selecting typical microfracture intervals, and then applied to the conversion of microfracture width in each well section.

[0059] 4) Calculate the output statistical curve The parameters mentioned above for each image sample are statistically analyzed according to a certain window length and step size, and the following curve is calculated and output: Fracture density (FD): The total number of fractures per unit well section, expressed in units of 1 / m, i.e.: To count the total number of cracks inside the window, The length of the statistical window, in meters; Fracture length (FL): The sum of fracture lengths per unit area of ​​wellbore, measured in meters (m / m). 2 Or 1 / m; that is: in, Let be the length of the i-th crack, in meters. The width of the electrical imaging logging image is the perimeter of the wellbore, expressed in meters (m). This represents the total area of ​​the well wall within the statistical window.

[0060] Crack Width (FW): The average width of all cracks within the statistical window, in mm. in, The width of the i-th crack is in mm; Crack area ratio (FP): The percentage of the total area of ​​a crack in an image, i.e.: in, Let m be the apparent area of ​​the i-th crack. 2 As a unit; 9) Generation of the result map: By stitching together the various binarized image samples, a fracture image of the entire well section is formed. This image is then combined with the fracture parameter curves and the dynamic electrical imaging map to create a microfracture processing result map of the tight sandstone and conglomerate section. Based on this, a segmented evaluation of the degree of microfracture development can be given.

[0061] Example The following uses electrical imaging logging data from the tight sandstone section of Well X (using an FMI instrument, with a wellbore diameter of 8.5 inches) as an example to illustrate the specific implementation method: 1. Preprocess raw electrical imaging logging data to generate dynamic images. Use the Heat color scale (transitional colors: white-yellow-brown-black) to create pseudo-color images. Use specialized software to repair blank bands between electrodes, generating a dynamic imaging map of the entire well circumference.

[0062] 2. The dynamic imaging image of the entire well is cut with a window length of 0.5m and a step size of 0.5m, and then cut into image samples of equal size with the beginning and end connected, which are used to generate a binarized image of the entire well; the dynamic imaging image of the entire well is cut with a window length of 0.5m and a step size of 0.125m, which is used to calculate the continuous fracture parameter curve.

[0063] 3. Select 49 typical image samples, manually label the microcracks and background, and merge and save them as a training set.

[0064] 4. Based on the set Gaussian blur, Sobel filter, Gaussian difference and Hessian matrix eigenvalues ​​and six different radii, extract 70 feature maps from the training set samples, train the set fast random forest model, and save the trained model.

[0065] 5. Use the trained model to predict all image samples and generate corresponding binarized images of microcracks. The results are shown in the appendix. Figure 6 As shown.

[0066] 6. The binarized images of micro-fractures segmented by a 0.5m window length and a 0.5m step length are stitched together end to end to form the whole-well fracture image.

[0067] 7. The binarized images of micro-fractures segmented with a window length of 0.5m and a step size of 0.5m are stitched together end to end to form a full-well fracture image.

[0068] 8. Establish the microcrack width conversion formula for the target block, such as... Figure 7 As shown, however, the fracture width of this well section is converted to the actual value.

[0069] 9. Calculate and output statistical curves with a window length of 0.5m and a step size of 0.125m, that is, statistically analyze and output one data point for each image sample. Among them, the perimeter of the 8.5-inch wellbore (i.e., the true width of the electrical imaging image) is 678mm.

[0070] 10. By comprehensively plotting dynamic electrical imaging maps, fracture images of the entire well section, and fracture parameter curves, a microfracture treatment result map of the tight sandstone and conglomerate section is formed. For example... Figure 8 As shown, the calculated crack widths are all below 0.01 mm, belonging to the micro-crack level. Furthermore, in sections with well-developed micro-cracks, the values ​​of various crack parameters are higher, while in sections with underdeveloped micro-cracks, the values ​​of various crack parameters are relatively lower. Therefore, the processing results based on this invention can effectively and continuously evaluate the relative development degree of micro-cracks in tight sandstone and conglomerate.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic identification of microfractures in tight sandstone and conglomerate based on electrical imaging logging, characterized in that, include: Step 1. Correct depth, amplitude and orientation, perform dynamic enhancement, generate dynamic electro-imaging image, use gradient color marks for color matching, and form pseudo-color image; Step 2. Interpolate and repair the blank stripes between the electrodes in the dynamic electro-imaging image from Step 1 to generate a dynamic imaging image of the entire well circumference; Step 3. The full-well-circumference dynamic imaging image from Step 2 is processed by sliding according to a certain window length and step size, and it is cut into multiple image samples of fixed size; Step 4. Set the feature extraction algorithm and corresponding different radii to extract crack information from the image samples in Step 3 in multiple dimensions and at multiple scales. Step 5. Use an image labeling tool to trace the microcrack outlines of some image samples from Step 4, delineate some background images, and form closed polygonal regions in sequence. Label the pixels inside these regions as either cracks or background. Merge and save all labeled image samples for use as a training set for machine learning. Step 7. Combining the feature extraction in Step 4, the manual labeling in Step 5, and the random forest classifier, the manually labeled training set images are used as input to the random forest model classifier to train the random forest classifier. The trained model is saved and applied to unlabeled image samples to achieve automatic identification of microcracks and form the corresponding crack binarized image. Step 8. Calculate the parameters of the microcracks; Step 9. Stitch together the binarized image samples from Step 7 to form a fracture image of the entire well section. Combine this image with the fracture parameter curve from Step 8 and the dynamic electrical imaging image from Step 1 to draw the microfracture processing result map of the tight sandstone and conglomerate section.

2. The method according to claim 1, characterized in that, The window length refers to the length of the current processing depth segment, and the step size refers to the distance between adjacent processing windows.

3. The method according to claim 1, characterized in that, In step 4, the feature extraction algorithm includes: Gaussian blur algorithm as a noise reducer, and Sobel filter, difference of Gaussians and eigenvalues ​​of Hessian matrix as edge detectors; The different radii include: using six different radii σ to extract features at different scales from image samples.

4. The method according to claim 1, characterized in that, The background image includes gravel, macroscopic cracks, and a sandy skeleton.

5. The method according to claim 1, characterized in that, Step 8 includes: For a single microcrack in the image sample, fitting its trajectory with a sine function, where the number of crack pixels along the trajectory direction is the apparent length of the microcrack, denoted as . The average number of crack pixels along the vertical direction of the trajectory is the micro-crack apparent width, set as... The total number of pixels in a single crack is the apparent area of ​​the microcrack, set as [value missing]. According to the definition, To simplify the calculation, the apparent length of the microcrack is first determined. and the apparent area of ​​microcracks Then solve for the apparent width of the microcrack. .

6. The method according to claim 5, characterized in that, Step 8 also includes unit conversion of single fracture parameters: fracture apparent length, fracture apparent width and fracture apparent area are converted to m, mm and m2, respectively 2 units, i.e. in, The distance between any two adjacent pixels in an electrical imaging image is determined by the principle of electrical imaging logging instruments. ; Because the fracture widths shown in electrical imaging logging images are larger than the actual fracture widths under formation conditions, specialized software is used to calculate the actual width of each fracture. Then, the calculated apparent fracture widths are fitted using a polynomial method with the corresponding calculation results from the specialized software. In the formula, The apparent width of the crack, i.e., the width of a single crack. , These are the corresponding calculation results from professional software. These are the fitting coefficients; A unified fitting formula for the target block is established by selecting typical microfracture intervals, and then applied to the conversion of microfracture width in each well section.

7. The method according to claim 4, characterized in that, Step 8 also includes: calculating and outputting statistical curves: statistically analyzing the above parameters for each image sample according to a certain window length and step size, and calculating and outputting the following curves: Fracture density FD: The total number of fractures per unit well section, expressed in units of 1 / m, i.e.: To count the total number of cracks inside the window, The length of the statistical window, in meters; Fracture length FL: The sum of fracture lengths per unit area of ​​wellbore, expressed in meters (m / m). 2 Or 1 / m; that is: in, Let be the length of the i-th crack, in meters. The width of the electrical imaging logging image is the perimeter of the wellbore, expressed in meters (m). This represents the total area of ​​the well wall within the statistical window; Crack width FW: The average width of all cracks within the statistical window, in mm, i.e.: in, The width of the i-th crack is in mm; Crack area ratio (FP): The percentage of the total area of ​​a crack in an image, i.e.: in, Let be the apparent area of ​​the i-th crack, in meters. 2 .

Images