A method for accurate identification and 3D network reconstruction of surrounding rock fractures integrating deep learning

By introducing the DE-YOLOv11 model with enhanced detail convolution and composite loss function, combined with spatial uncertainty quantification and adaptive sampling strategies, the problem of unstable fracture identification in borehole imaging technology is solved, achieving stability in fracture identification and completeness in parameter extraction, thus improving the reliability of rock mass fracture analysis.

CN123118914APending Publication Date: 2026-11-17XIJING UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611373640.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-07
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing borehole imaging technology suffers from numerous interference factors, difficulty in identifying minute fractures, and a lack of adaptive processing strategies when fracture morphology deviates from the sinusoidal curve, resulting in unstable fracture segmentation results and incomplete extraction of geometric parameters in rock fracture identification.

Method used

We adopt a deep learning approach, introduce the DE-YOLOv11 segmentation model with detail enhancement convolution and boundary constraint composite loss, combine spatial uncertainty quantization and area hierarchical adaptive sampling strategy to enhance high-frequency edge response and match differentiated sampling density, and construct a three-dimensional crack network spatial structure.

Benefits of technology

This improves the stability and computational efficiency of crack identification, enhances the integrity and reliability of crack geometric parameters, and provides a reliable foundation for subsequent engineering analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118914A_ABST
    Figure CN123118914A_ABST
Patent Text Reader

Abstract

This invention discloses a method for accurate identification and 3D network reconstruction of surrounding rock fractures using deep learning, relating to the fields of rock engineering and geological exploration. The invention includes preprocessing borehole imaging unfolded images, inputting them into a deep learning network with detail enhancement convolutional modules for fracture segmentation; calculating joint roughness coefficients using an adaptive sampling strategy on the segmentation results; constructing a connectivity probability model for fracture pairs to evaluate 3D connectivity; and determining supplementary borehole locations through multi-objective optimization. This invention, by integrating deep learning segmentation, adaptive roughness calculation, and multi-factor connectivity analysis, can improve the reliability of fracture identification and 3D network reconstruction in complex borehole images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of rock engineering and geological exploration technology, and in particular relates to a method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures by integrating deep learning. Background Technology

[0002] Borehole imaging technology is a geophysical exploration method that uses optical imaging principles to continuously scan the borehole wall. This technology involves lowering a high-definition camera into the borehole to perform a 360° omnidirectional scan of the borehole wall, acquiring a two-dimensional unfolded image that reflects the internal structural characteristics of the rock mass. In the unfolded borehole image, when a fracture in the rock mass intersects the borehole axis, it theoretically exhibits a sinusoidal curve shape. This geometric feature provides important mathematical basis for the spatial orientation analysis of fractures. Borehole imaging technology has been widely used in mining engineering, tunnel engineering, and slope engineering to detect and assess the development of fractures in surrounding rock.

[0003] Fracture identification is a crucial step in borehole imaging data analysis. In engineering practice, technicians typically employ edge detection-based image processing methods to extract fracture information. Traditional image processing techniques such as the Sobel operator and the Canny edge detection algorithm have been incorporated into some commercial borehole imaging analysis software as tools to assist manual interpretation. In recent years, with the rapid development of deep learning technology, target detection and segmentation models such as the YOLO series and U-Net have been gradually introduced into rock fracture identification tasks. Through training on a large number of labeled samples, end-to-end pixel-level prediction from borehole images to fracture regions has been achieved, yielding certain results in some research projects and engineering pilots.

[0004] However, the aforementioned existing technologies still have significant limitations in practical engineering applications. On the one hand, borehole imaging unfolded images commonly contain various interfering factors such as natural rock textures, drill bit scratches, residual mud on the borehole wall, and lens adhesion stains. These interferences are highly similar to real fractures in grayscale distribution and local morphology, making traditional edge detection operators based on manually set thresholds prone to misjudgment. Furthermore, deep learning models lack a mechanism for preserving high-frequency detail features during multi-layer downsampling, making it difficult to effectively respond to minute fractures with widths of only a few pixels in deep feature maps. The final fracture segmentation results frequently exhibit disconnections, boundary shifts, or missing regions. On the other hand, when the fracture morphology deviates from the standard sine curve, existing methods lack effective adaptive processing strategies, further affecting the completeness of fracture geometric parameter extraction and the reliability of subsequent analysis. Therefore, the following solutions are proposed to address these issues. Summary of the Invention

[0005] The purpose of this invention is to provide a method for accurate identification and 3D network reconstruction of rock fissures by integrating deep learning. By introducing detail-enhanced convolution and boundary constraint composite loss, the high-frequency edge response in the segmentation network is strengthened. Combined with spatial uncertainty quantization, the problem of easy disconnection and blurred boundaries of fine fissures in borehole images under strong interference background is alleviated. At the same time, through the area-layer adaptive sampling strategy, the sampling density is matched according to the geometric complexity of the fissure, which controls the computational cost while retaining key abrupt features, and improves the problem of key feature omission or computational redundancy caused by fixed sampling in the calculation of joint roughness coefficient.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0007] This invention is a method for accurate identification of surrounding rock fractures and 3D network reconstruction that integrates deep learning, comprising the following steps:

[0008] Image preprocessing is performed on the acquired borehole imaging unfolded image to obtain an enhanced image;

[0009] The enhanced image is input into a pre-trained DE-YOLOv11 deep learning segmentation model for pixel-level binarization segmentation. The backbone network of the DE-YOLOv11 deep learning segmentation model uses a detail enhancement convolution module to replace part of the standard convolution. The detail enhancement convolution module sets up parallel standard convolution paths and high-frequency response paths, and outputs a binarized gap mask of the same size as the original image.

[0010] The binary crack mask is used for crack classification and parameter extraction: for cracks that conform to the sinusoidal morphology criteria, a sinusoidal model is used for parameter fitting and inversion to output amplitude, period, phase and centerline position parameters; for cracks that do not conform to the sinusoidal morphology criteria, the roughness calculation process is initiated.

[0011] For each crack in the binary crack mask, the crack area is calculated and layered according to the area size. Based on the layering results, the corresponding sampling strategy is automatically matched to extract the discrete coordinate sequence of the crack contour. The root mean square slope is calculated based on the discrete coordinate sequence, and then the joint roughness coefficient value is obtained by mapping through empirical formula.

[0012] Based on the geometric feature parameters of multiple fractures and the joint roughness coefficient, a four-component coupled connectivity probability model between fractures is constructed, and the connectivity probability of fracture pairs is output. Based on this, a three-dimensional fracture network spatial structure is constructed.

[0013] Based on the spatial distribution of existing boreholes, a three-dimensional spatial uncertainty field is constructed, high uncertainty clustering regions are identified, a multi-objective borehole layout optimization model is established to solve for the optimal location of supplementary boreholes, and the three-dimensional coordinates and priority of supplementary boreholes are output.

[0014] Furthermore, the image preprocessing includes grayscale and normalization processing, contrast-limited adaptive histogram equalization enhancement, multi-scale Gaussian filtering for noise reduction, and blind image inpainting based on a conditional diffusion model.

[0015] Furthermore, the DE-YOLOv11 network also includes an input terminal, a feature fusion network, and a segmentation head; the feature fusion network adopts an FPN+PAN multi-scale feature fusion structure.

[0016] Furthermore, the deep learning segmentation model is trained using a composite loss function, which consists of a weighted combination of segmentation loss, boundary loss, detail preservation loss, and regularization term. The segmentation loss is a combination of weighted binary cross-entropy loss and Dice loss. The boundary loss is the L1 norm of the difference between the Sobel gradients of the predicted mask and the real mask. The detail preservation loss is the squared L2 norm of the difference between the high-frequency components of the predicted probability map and the real mask after high-pass transformation.

[0017] Furthermore, the output of the detail enhancement convolution module is: a weighted sum of the standard convolution operation result and the convolution operation result of the input feature map multiplied by the learnable scaling parameter and the high-pass operator.

[0018] Furthermore, the sinusoidal morphology discrimination is as follows: the extracted crack contour line is fitted with a sinusoidal model, the goodness of fit is calculated, and when the goodness of fit is greater than the first preset threshold and the amplitude is greater than the second preset pixel threshold, it is judged as a sinusoidal crack; otherwise, it is judged as a complex morphology crack.

[0019] Furthermore, the step of stratifying and matching corresponding sampling strategies based on area size specifically involves: when the crack area is less than a first area preset threshold, an equal-interval sampling strategy is matched; when the crack area is between the first area preset threshold and a second area preset threshold, a joint sampling strategy combining curvature adaptive sampling and slope adaptive sampling is adopted; when the crack area is greater than the second area preset threshold, an inflection point adaptive sampling strategy is matched, wherein the inflection point adaptive sampling is based on the second derivative of the contour line to detect inflection points and extreme points for sampling.

[0020] Furthermore, the four-component coupled connectivity probability model includes a distance factor, a morphological similarity factor, a joint roughness coefficient compatibility factor, and a spatial uncertainty factor. The distance factor is calculated using a dual exponential decay model based on the spatial distance and depth difference between fracture pairs. The morphological similarity factor is calculated based on the weighted sum of the area similarity, orientation similarity, and dip angle similarity of fracture pairs. The joint roughness coefficient compatibility factor is calculated based on the similarity of the joint roughness coefficient values ​​of fracture pairs and the degree of proximity to the optimal connectivity value. The spatial uncertainty factor is calculated based on the comprehensive uncertainty of the spatial location of each fracture pair.

[0021] Furthermore, the multi-objective borehole layout optimization model includes objectives of maximizing uncertainty reduction, maximizing connectivity enhancement, and maximizing information coverage. An improved non-dominated sorting genetic algorithm is used to solve the Pareto optimal solution set, and then the TOPSIS approximation ideal solution sorting method is used to prioritize the Pareto optimal solution set. The spatial coordinates corresponding to the top-ranked solutions are selected as the optimal supplementary borehole positions.

[0022] Furthermore, it also includes a crack network visualization output step: cracks identified as complex morphological cracks are characterized by piecewise linear interpolation; different cracks in the three-dimensional crack network spatial structure are marked with different colors according to the joint roughness coefficient value, and the connecting lines between cracks are marked with different line widths or colors according to the connectivity probability value; high uncertainty regions are marked with marker symbols.

[0023] The present invention has the following beneficial effects:

[0024] This invention, by introducing detail-enhanced convolution and composite loss functions, can extract weak crack features and maintain the stability of segmentation results under complex background conditions. By matching different sampling strategies with area layering, it can adapt to the roughness calculation needs of cracks of different sizes, improving computational efficiency and result consistency. By integrating multiple factors such as spatial distance, morphological similarity, surface roughness compatibility, and spatial uncertainty, a connectivity probability evaluation framework is established, which can more comprehensively reflect the correlation between cracks. Through uncertainty analysis and multi-objective optimization, it can provide a reference for the selection of supplementary borehole locations, forming a complete process from image processing to evaluation decision-making, which is conducive to improving the overall reliability of crack identification and network reconstruction, and provides basic support for subsequent engineering analysis.

[0025] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a comparison image showing the effect of DiffBIR noise reduction before and after processing according to the present invention;

[0028] Figure 2 This is a schematic diagram of the overall structure of the DE-YOLOv11 network of the present invention;

[0029] Figure 3This is a comparison image of the binary representation of the crack identification method in this invention.

[0030] Figure 4 This is a comparison diagram of the recognition between the DE-YOLOv11 of this invention and the original YOLOv11;

[0031] Figure 5 This is a 1-1 crack identification confidence heatmap generated by DE-YOLO in this invention;

[0032] Figure 6 This is a heatmap of 1-2 crack identification confidence generated by DE-YOLO in this invention;

[0033] Figure 7 This is a heatmap of 1-3 crack identification confidence generated by DE-YOLO in this invention;

[0034] Figure 8 This is a diagram showing the sinusoidal fitting results of some cracks in this invention;

[0035] Figure 9 This is a comparison image of the original image 3-1 and the crack contour recognition of the present invention;

[0036] Figure 10 This is a comparison image of the original image 3-2 and the crack contour recognition of the present invention;

[0037] Figure 11 This is a comparison image of the original image 3-3 and the crack contour recognition of the present invention;

[0038] Figure 12 The crack conforming to the sinusoidal fitting condition of 3-1 in this invention;

[0039] Figure 13 The crack in this invention conforms to the sinusoidal fitting condition 3-2;

[0040] Figure 14 The crack in this invention conforms to the sinusoidal fitting condition 3-3;

[0041] Figure 15 This is a diagram illustrating the multi-dimensional impact mechanism of the discrete coordinate point selection method on JRC calculation in the context of the present invention.

[0042] Figure 16 This is a diagram illustrating the multi-dimensional impact mechanism of the discrete coordinate point selection method on JRC calculation in the context of in-depth analysis of this invention.

[0043] Figure 17 This invention provides a deep analysis of the multi-dimensional impact mechanism of the discrete coordinate point selection method in image 3-3 on JRC calculation.

[0044] Figure 18 This is a statistical diagram showing the hierarchical distribution of connectivity probability in this invention.

[0045] Figure 19 This is a scatter plot showing the relationship between JRC values ​​and connectivity in this invention.

[0046] Figure 20 This is a three-dimensional connectivity analysis diagram of the multi-hole fracture network of the present invention;

[0047] Figure 21 This is a pie chart showing the hierarchical distribution of the spatial uncertainty region in this invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention is a method for accurate identification of surrounding rock fractures and 3D network reconstruction that integrates deep learning, including:

[0051] Phase 1: Image Preprocessing

[0052] Step 1: Grayscale Conversion and Normalization

[0053] The acquired color borehole imaging unfolded image is converted to grayscale, compressing the three-channel color image into a single-channel grayscale image, preserving image brightness features while reducing subsequent computational complexity. The grayscale conversion uses a weighted average method. In the formula, , , The input image is located at pixel coordinates. The red, green, and blue channel values ​​are then calculated. Grayscale normalization is then performed, linearly mapping the pixel values ​​to... The range eliminates the brightness inconsistency caused by differences in lighting conditions and imaging equipment in different boreholes, allowing the model to focus on the relative grayscale difference between the fracture and the surrounding rock rather than the absolute brightness value.

[0054] Step 2: Adaptive Contrast Enhancement

[0055] Contrast-limited adaptive histogram equalization (CLAHE) is used to enhance the local contrast of the image. Specifically, the image is divided into several non-overlapping sub-blocks (preferably divided into...). (Each sub-block), calculate the cumulative distribution function (CDF) of each sub-block, and for those exceeding the preset contrast limit parameter. The histogram portion is cropped and redistributed, and finally, bilinear interpolation is used to eliminate the boundary effect between sub-blocks to obtain an enhanced image. The preferred range is 2.5 to 3.5.

[0056] Its mathematical expression is:

[0057]

[0058] In the formula, This indicates that the grayscale image output in step 1 is located at pixel coordinates. grayscale value at that location It is a local cumulative distribution function. Contrast limiting parameter (preferred value) ), It is the inverse function of the cumulative distribution function.

[0059] Step 3: Multi-scale Gaussian filtering for noise reduction

[0060] A multi-scale Gaussian filter is employed to suppress sensor noise and optical noise in the image. A Gaussian scale space is constructed, the first... The layer filtering result is:

[0061]

[0062]

[0063] In the formula, This represents the convolution operation. The standard deviation is Two-dimensional Gaussian kernel, , The number of scale samples (preferred) ), The initial scale parameter (preferred range is 1.0~2.0, optimal value) ), . This is the enhanced image output from step 2. The multi-scale filtering results are then weighted averaged across all scales. .

[0064] Step 4: Blind Image Inpainting Based on Diffusion Model

[0065] The processed image is then input into the DiffBIR blind image inpainting network, which is based on a conditional diffusion model. This network effectively removes mud streaks and lens white spots characteristic of borehole imaging while preserving high-frequency details at fracture edges to the greatest extent possible. DiffBIR employs a conditional diffusion model architecture, learning the prior distribution of the image to intelligently distinguish noise from the real signal and output a standard enhanced image. Please refer to [link to relevant documentation]. Figure 1The image shown is a comparison of the effects of DiffBIR noise reduction before and after processing according to the present invention. (A) is the original image with obvious noise interference and blurred edges before noise reduction, and (B) is the enhanced image after DiffBIR processing, in which noise is effectively removed, structural outlines are clear, and detailed textures are accurately restored.

[0066] Phase Two: DE-YOLOv11 Pixel-Level Binary Segmentation

[0067] Step 5: Construct the DE-YOLOv11 segmentation network

[0068] Construct a deep learning segmentation model that includes an input network, a backbone network, a feature fusion network, and a segmentation head. This model processes the input image... ( and Output pixel-level crack probability map (the height and width of the image are given respectively, in pixels). The input layer performs standardization and data augmentation on the preprocessed image, including various strategies such as random scaling, color dithering, and Mosaic, to enhance the model's adaptability to cracks of different scales. The backbone network uses improved convolutional and residual modules for multi-level feature extraction, balancing deep semantic features with computational efficiency. The feature fusion network employs an FPN+PAN multi-scale structure, fusing feature maps from different scales in the backbone network to enhance the detection capability of small cracks. The segmentation head performs pixel-level mask prediction on the feature maps. (See also...) Figure 2 The diagram shows the overall structure of the DE-YOLOv11 network proposed in this invention. It illustrates the structure of the innovative segmentation algorithm DE-YOLOv11, built upon YOLOv11. The main feature of DE-YOLOv11 is the introduction of Enhanced Detail Convolution (DEConv) into the feature extraction process. Traditional convolutions tend to lose edge and detail features of fractures in complex backgrounds, while DEConv effectively enhances the expressive power of shallow features by introducing a detail-preserving mechanism into the convolution operation, thus maintaining the continuity and integrity of fracture edges even under interference from rock textures and drilling marks.

[0069] Step 6: Introduce the detail enhancement convolution module

[0070] In the backbone network, augmented detail convolutions (DEConv) are used to replace some standard convolutions. This module introduces high-frequency response paths in parallel on top of standard convolutions, mathematically as follows:

[0071]

[0072] In the formula, For the input feature map, For standard convolution kernels, For convolution operations, For high-pass operators (specifically, the Laplacian high-pass kernel) (convolution operation with input feature map) , Laplacian high-pass kernel matrix:

[0073]

[0074] In the formula, The learnable scaling parameter (initialized to 0.1) is used to balance the contributions of standard convolutions and detail pathways. This structure allows the network to capture low-frequency semantic information while enhancing high-frequency gap edge responses.

[0075] Step 7: Define the composite loss function

[0076] The training phase employs a composite loss function combining multiple constraints:

[0077]

[0078] In the formula, , , The weighting coefficients for each loss term are as follows, and the preferred values ​​are respectively: , , .

[0079] Segmentation loss A combination of weighted binary cross-entropy and Dice loss is used to address the severe class imbalance problem where the proportion of gap pixels is low (typically less than 10%). The weighted binary cross-entropy is:

[0080]

[0081] In the formula, For pixel coordinates, The crack probability map predicted by the model in pixels The value at that location, To accurately label the mask at the pixel level The value at the location (1 represents the crack, 0 represents the background). Pixel weighting function:

[0082]

[0083] In the formula, This is a map showing the true mask boundaries extracted by morphological gradients or the Sobel operator. , and These are morphological dilation and erosion operations, respectively. The boundary weighting strength parameter (preferred range is 10~30, optimal value) ).

[0084] Dice's loss is:

[0085]

[0086] In the formula, It is a minimal constant (take) ), to prevent division by zero errors.

[0087] The combination of segmentation loss is:

[0088]

[0089] In the formula, and For the weighting coefficients, the preferred values ​​are respectively: , .

[0090] Boundary loss To ensure consistency between the predicted and true masks at their boundaries, Sobel gradient differential is employed. Norm:

[0091]

[0092] In the formula, This indicates that the Sobel gradient operation is performed on the image. for Norm (the sum of the absolute values ​​of all elements).

[0093] Loss of detail Constrain the response of the high-frequency pathways of DEConv to ensure that the output probability map approximates the true mask on the high-frequency components (crack edges):

[0094]

[0095] In the formula, This represents a high-pass transform (using the Laplacian operator). for The square of the norm (the sum of the squares of all elements). Here and All space dimensions are The resolution was kept consistent before the calculation.

[0096] Regularization term Standard weight decay (L2 regularization) is used to constrain network parameters and prevent overfitting.

[0097]

[0098] In the formula, For all trainable parameters of the model, for The square of the norm.

[0099] Step 8: Model Training and Binarization Output

[0100] The network was trained using the AdamW optimizer, with an initial learning rate set to... A cosine annealing learning rate scheduling strategy is adopted (the learning rate is set according to the cosine function from...). decay to The weight decay coefficient is set to The batch size is set to 8-16 (adjusted according to GPU memory capacity; use a larger value when memory is sufficient). After training, the model enhances the input images. Output pixel-level crack probability map ,in This is the end-to-end mapping function defined for the trained DE-YOLOv11 model, from the input image to the probability map. A threshold is used. Perform binarization segmentation to obtain a binary mask of the same size as the original image:

[0101]

[0102] In the formula, This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. The segmentation threshold (preferred value) ). Represents the crack pixels (shown in black in the output binarized image). This represents background pixels (displayed in white in the output binarized image). See also... Figure 3 The image shown is a comparison of the binarized data for fracture identification according to the present invention. The image displays a side-by-side comparison of the original borehole image, the fracture probability heatmap, and the final binarized mask. Please refer to [link / reference]. Figure 4 The image shown is a comparison of the recognition capabilities of the DE-YOLOv11 of this invention and the original YOLOv11. Please refer to [link / reference]. Figures 5 to 7 As shown, these are heatmaps of the confidence level of fracture identification in three typical borehole images generated by DE-YOLOv11. The fracture area shows a high confidence response of red and yellow (0.6~1.0), while the non-fracture area shows a low response of dark blue (close to 0).

[0103] Phase 3: Fitting and Inversion of Sinusoidal Fracturing Parameters

[0104] Step 9: Contour extraction and sine discrimination

[0105] For the binary mask of the output of stage two The OpenCV function `cv2.findContours` was used to extract the contours of the crack region, and the contour with the largest area was selected as the main boundary. The contour points were then... Sort the coordinates (ensure) (Coordinates are monotonically increasing), making the coordinate sequence ( , (Total number of contour points) conforms to the geometric orientation of the crack. The extracted contour lines are fitted using a sinusoidal model, with the model form as follows:

[0106]

[0107] In the formula, The amplitude represents the maximum opening width of the crack; The period is theoretically equal to the borehole circumference; The phase determines the starting position of the crack in the circumferential direction; The centerline location indicates the center depth of the crack in the axial direction. (See also...) Figure 8 The figure shown is a sine curve fitting result of some cracks in this invention. The blue scatter points in the figure represent the original crack contour data, and the red dashed line represents the fitted sine curve. The parameter values ​​and goodness of fit are marked. .

[0108] After initial fitting using the nonlinear least squares method, the goodness of fit is calculated:

[0109]

[0110] In the formula, For the first The actual axial coordinates of each contour point For the model in The predicted value at that location, This is the mean of the axial coordinates of all contour points. If... And amplitude If the crack pattern is accurate (e.g., a single pixel), it is classified as a sinusoidal crack and proceeds to step 10 for parameter refinement; otherwise, it is classified as a complex crack and proceeds to stage four for processing. (See also...) Figures 9 to 14 As shown, this invention has detected fractures that meet the sinusoidal fitting conditions in three different borehole images. The green markings indicate the fracture regions that are determined to be sinusoidal.

[0111] Step 10: Initial parameter estimation

[0112] The initial period of the crack was estimated using Fast Fourier Transform (FFT). This was applied to the centerline data. (Obtained through morphological skeleton extraction) Perform FFT analysis:

[0113]

[0114] In the formula, This is a circumferential data sequence of the crack centerline (length equal to the number of pixels in the image width direction). For frequency variables, For Fast Fourier Transform operations, For modulo operation, The width of the image (by multiplying the image width in pixels by the pixel resolution). get).

[0115] Initial value of amplitude and the initial value of the center line Estimation based on extreme values ​​of contour data:

[0116]

[0117]

[0118] In the formula, This is the sequence of axial coordinates for the crack profile.

[0119] initial phase value Determined by minimizing the fitting error:

[0120]

[0121] In the formula, For the first on the outline The coordinates of each sampling point This indicates that the expression within the parentheses should be minimized. The value of is solved by one-dimensional search (such as the golden section method).

[0122] Step 11: Optimization of Regularization Objective Function

[0123] A physically constrained regularized objective function is constructed, and the optimal parameters are iteratively solved using the Levenberg-Marquardt algorithm (a nonlinear least squares optimization algorithm combining the Gauss-Newton method and gradient descent).

[0124]

[0125] In the formula, The parameter vector to be optimized ( (represents matrix transpose); This represents the total number of contour points; For the model in The predicted value at that location; For the first Weight of each point:

[0126]

[0127] In the formula, For the first The distance from each data point to the borehole centerline Distance attenuation parameter (preferred value) mm), It is a natural exponential function; Regularization coefficient (preferred value) ); For regularization terms:

[0128]

[0129] In the formula, The theoretical perimeter of the borehole ( (Drilling diameter) Periodic penalty coefficient (preferred value) ), The amplitude overshoot penalty coefficient (preferred value) ), For indicator functions, and These are the lower and upper limits of the physical constraints on the amplitude (set according to the vertical dimension of the image, usually taken as...). Pixels (pixels), period The constraint range is .

[0130] The optimal parameters are output after optimization. , , , And calculate the goodness of fit. (Formula same as step 9) and period deviation :

[0131]

[0132] As a quality assessment indicator. and It was determined to be a valid fit.

[0133] Phase 4: Adaptive JRC Calculation Based on Area Stratification

[0134] Step 12: Calculation of fracture area and determination of stratification

[0135] For each crack output in stage two, the binary mask Count the total number of pixels with cracks. Based on image resolution (Pixels per inch, unit: dpi) Converts pixel area to physical area. :

[0136]

[0137] In the formula, This is the conversion factor between inches and millimeters. Based on physical area. Determine the size by layering:

[0138] when When the area is mm², it is determined to be a small area crack;

[0139] when When the area is mm², it is determined to be a medium-area crack;

[0140] when When the area is mm², it is determined to be a large-area crack.

[0141] Step 13: Hierarchical matching sampling strategy

[0142] Based on the stratification result in step 12, the corresponding sampling strategy is automatically matched:

[0143] (1) Small area cracks, equidistant sampling

[0144] Distribute evenly between the start and end points of the outline. One sampling point, The preferred range is 40~60 (optimal value) The x-coordinate of the sampling point is calculated as follows:

[0145]

[0146] corresponding ordinate Obtained through linear interpolation:

[0147]

[0148] In the formula, and These are the minimum and maximum values ​​of the x-coordinate of the contour line, respectively; This represents the total number of original contour points; For linear interpolation functions (based on known discrete points) At the specified x-coordinate Linear interpolation is performed at the location to obtain the corresponding ordinate value.

[0149] (2) Joint sampling of medium-area cracks, curvature adaptive and slope adaptive

[0150] First, calculate the curvature of the contour line at each point:

[0151]

[0152] In the formula, and The outlines are respectively at First derivative (dimensionless) and second derivative (unit: mm) at point ), from discrete points using the finite difference method Approximate calculation.

[0153] Simultaneously calculate the slope change between adjacent points:

[0154]

[0155] In the formula, For the first Local slope of the segment (dimensionless). It is the arctangent function.

[0156] The curvature adaptive sampling criterion is: in the curvature Greater than the preset threshold Dense sampling was conducted in the area. Take the 75th percentile of all curvature values ​​(i.e. , (A function for calculating the 75th percentile). The slope adaptive sampling criterion is: as the slope changes... Greater than the preset threshold Regional encrypted sampling, Take the 75th percentile of all slope changes. The sampling points of the two strategies are combined into a union, and an upper limit is set on the sampling density (the maximum number of sampling points does not exceed 50% of the original contour points) to ensure computational efficiency.

[0157] (3) Adaptive sampling of large-area cracks and inflection points

[0158] Inflection points and extreme points of contour lines detected using second derivatives:

[0159] Inflection point determination: and The signs are opposite on both sides of this point;

[0160] Extreme point determination: .

[0161] The comprehensive sampling strategy is as follows:

[0162]

[0163] In the formula, For the set of inflection points, For the set of extreme points, These are additional uniformly distributed points (in addition to the key points mentioned above) The sampling points are evenly supplemented in the correct direction, so that the total number of sampling points reaches 20% to 30% of the original contour points. (See also...) Figures 15 to 17 As shown, the figures represent the multi-dimensional influence mechanism of the discrete coordinate point selection method on JRC calculation in three typical borehole images, and show the comprehensive comparison results of five sampling methods in terms of JRC distribution comparison, sampling point density influence analysis, and JRC value probability distribution.

[0164] Step 14: Numerical Calculation of JRC Value

[0165] For the discrete coordinate sequence obtained in step 13 ( Given the total number of sampling points, outliers are first filtered using the interquartile range (IQR) method. Specifically: Calculate... median of sequence Lower quartiles and upper quartiles , Eliminate those that meet the requirements Outliers. Simultaneously set a minimum distance threshold. ( (Unit: mm) to prevent division by zero errors.

[0166] Then the root mean square slope was calculated. Using discrete integral form:

[0167]

[0168] In the formula, This represents the total number of sampling points after outlier filtering. For the outline at Total projected length in the direction; For adjacent sampling points in The difference in direction; For adjacent sampling points in The difference in direction.

[0169] Finally, the JRC value is calculated using the Patton empirical formula, and physical constraints are applied.

[0170]

[0171] In the formula, and These are functions that take the maximum and minimum values, respectively, to ensure that the JRC value is constrained within a physically reasonable range. Inside.

[0172] Phase 5: Four-component connectivity probability modeling and supplementary borehole layout optimization

[0173] Step 15: Crack Pair Feature Extraction and Distance Factor Calculation

[0174] For all fractures obtained from steps two through four, extract their spatial center coordinates. (Unit: mm, based on the location of the fracture centerline in the borehole imaging image) (Obtained by superimposing with borehole spatial coordinates) fracture area (Calculation method is the same as step 12), Direction angle (Calculated using the principal axis direction of the contour line), Inclination angle and JRC value. For any crack pair Calculate the Euclidean distance in three-dimensional space:

[0175]

[0176] The distance factor employs a double exponential decay model:

[0177]

[0178] In the formula, The first The coordinates of the crack in three spatial dimensions; Spatial attenuation factor (preferred value) mm), representing the rate at which the horizontal connectivity probability decreases with increasing distance; Depth attenuation factor (preferred value) mm), representing the rate at which the connectivity probability in the depth direction decreases as the difference increases; It is a natural exponential function.

[0179] Step 16: Calculation of morphological similarity factor

[0180] Taking into account the similarity in crack size, orientation, and dip angle:

[0181]

[0182] The component functions are defined as follows:

[0183]

[0184]

[0185]

[0186] In the formula, For the first The area of ​​the crack; For the first The strike angle of the crack, and its range. ; For the first The dip angle of the crack, and its range. ; , , The weighting coefficients (determined through the analytic hierarchy process, with preferred values ​​of 0.25, 0.35, and 0.40) satisfy the following conditions: ; and These are functions for finding the minimum and maximum values, respectively.

[0187] Step 17: JRC Compatibility Factor Calculation

[0188] Based on tribological theory, the JRC compatibility factor is composed of a similarity function and an optimal connectivity function:

[0189]

[0190] In the formula, the similarity function Reflecting the consistency of surface roughness between the two cracks:

[0191]

[0192] and The first Article and Section JRC value of the fracture (dimensionless). This represents the upper limit of the JRC value range.

[0193] Optimal connectivity function This reflects that the average JRC value of the two fractures is close to the optimal connectivity value. The degree of roughness, in physical terms, means that surfaces with moderate roughness have good potential for fluid permeation and expanded connectivity.

[0194]

[0195] In the formula, The optimal value for connectivity extension (preferred range is 7~9, optimal value is) ), This refers to the range of JRC values.

[0196] Step 18: Calculation of Spatial Uncertainty Factor and Mapping of Crack Pairs

[0197] For any point in space ( (Using three-dimensional spatial coordinates, unit: mm), a comprehensive uncertainty assessment model is established:

[0198]

[0199] In the formula, , , The weighting coefficients (preferred values ​​are 0.5, 0.3, and 0.2) satisfy the following conditions: .

[0200] Distance uncertainty components Characterizing spatial points The degree of information loss due to distance from existing boreholes:

[0201]

[0202]

[0203] In the formula, For spatial points The minimum distance to the horizontal projection positions of all existing boreholes. For the first The horizontal coordinates of each borehole Maximum influence distance (preferred value) mm).

[0204] Uncertainty component of fracture density Characterizing spatial points The degree of insufficient understanding of fracture distribution due to distance from known fractures:

[0205]

[0206] In the formula, For spatial points The shortest spatial distance to the nearest crack. For the first The spatial center coordinates of the crack, Standard distance for crack influence (preferred value) mm), This is a function that takes the minimum value.

[0207] Deep uncertainty components Determined based on the distribution density of existing boreholes in different depth ranges: The depth range is divided into multiple intervals, and the number of existing boreholes in each interval is counted. The lower the borehole density in a depth interval, the higher the uncertainty. Specifically, according to... Calculation, where For depth Drilling density at the location, This represents the maximum borehole density value.

[0208] For crack pairs Uncertainty factors Defined as the uncertainty of the spatial location of each of the two fractures. and Weighted average:

[0209]

[0210] In the formula, and The first Article and Section The spatial center coordinates of the crack.

[0211] Step 19: Calculation of four-component coupled connectivity probability and construction of three-dimensional network

[0212] Combining the factors from steps 15 to 18, the crack pair The connectivity probability is:

[0213]

[0214]

[0215] In the formula, , , , The weight coefficients for each factor are determined using a multi-objective optimization method, with optimal ranges of 0.30~0.40, 0.20~0.30, 0.20~0.30, and 0.10~0.20, and their optimal values ​​are respectively... , , , ),satisfy The connectivity probabilities of all gap pairs constitute a connectivity probability matrix. ,in Let be the total number of fractures. A depth-first search (DFS) algorithm is used to identify connected components: starting from any unvisited fracture, proceed along the connected components... The edges are traversed to all reachable fractures, marked as a connected component, and this process is repeated until all fractures are visited. A 3D fracture network spatial structure graph is constructed. (See also...) Figure 18 The figure shown is a statistical chart of the hierarchical distribution of connectivity probability according to the present invention. The horizontal axis represents the connectivity probability interval, and the vertical axis represents the proportion of fracture pairs. The figure shows that most fracture pairs are clustered in the medium-high probability interval of 0.4 to 0.8. Please refer to [link / reference]. Figure 19 The image shows a scatter plot illustrating the relationship between the JRC value and connectivity in this invention. Each point represents a fracture, and the red dashed line represents the trend line of the quadratic polynomial fitting. Connectivity reaches its optimal level when the JRC value is around 10. Please refer to [link / reference]. Figure 20 The figure shows a three-dimensional connectivity analysis diagram of the multi-hole fracture network of the present invention. The thick black line in the figure represents 6 boreholes, and the dots of different colors represent fractures with different JRC values. The line width and color of the connecting lines reflect the connectivity probability strength, and the purple × mark indicates a region with high uncertainty.

[0216] Step 20: Identification of high uncertainty regions

[0217] The study area is discretized into a three-dimensional mesh (mesh size is optimized). ), calculate the comprehensive uncertainty of each grid center point The uncertainty threshold is determined using the quantile method:

[0218]

[0219] In the formula, The function to calculate the 75th percentile, This represents the total number of grid points. The uncertainty is greater than... Grid points are marked as high-uncertainty points, and the DBSCAN clustering algorithm is used to identify continuous high-uncertainty regions. Algorithm parameter settings: neighborhood radius. Preferably 300 mm, minimum number of dots The preferred value is 5. Output the center coordinates and spatial extent of each high-uncertainty cluster region. Please refer to [link / reference]. Figure 21 The image shows a pie chart illustrating the hierarchical distribution of spatial uncertainty regions in this invention. The chart displays the proportion of each uncertainty level, with high uncertainty regions (…). This accounts for 23.7% in total.

[0220] Step 21: Multi-target borehole layout optimization

[0221] Construct a borehole layout optimization model containing three objective functions. For the set of candidate supplementary borehole locations... , Preset the number of additional boreholes:

[0222] Objective 1: Maximize the reduction of uncertainty.

[0223]

[0224] in, For in position After additional drilling, the first The reduction in uncertainty at each grid point (unit: dimensionless) is calculated as the difference in uncertainty value at that point before and after the supplementary drilling. For the first The volume weight of each grid point (proportional to the volume of the grid cell).

[0225] Objective 2: Maximize connectivity enhancement

[0226]

[0227] in, For the first The connectivity potential of the fracture:

[0228]

[0229] In the formula, The total number of cracks. For the first The candidate borehole location to the first The shortest spatial distance between the centers of the cracks. For the first The spatial center coordinates of the crack, To find the maximum value function, For the first Importance index of a crack (take the area of ​​the crack) The ratio of the maximum area of ​​all cracks (dimensionless) For the first JRC value of the crack.

[0230] Objective 3: Maximize information coverage:

[0231]

[0232] In the formula, The total number of sub-regions to be divided for the study area (preferably the space is divided into) (a number of equal-volume sub-regions) For the first The weights of each sub-region (taking the mean of the uncertainty of that sub-region). To cover the indicators, , For the first The center coordinates of each sub-region For the coverage radius (preferred) mm).

[0233] The constraints include:

[0234] Spatial boundary constraints: , (Unit: mm, determined based on the drilling layout range at the engineering site);

[0235] Minimum distance constraint between boreholes: , Preferably 600 mm, The distance is Euclidean.

[0236] Construction accessibility and cost constraints (determining feasible drilling areas based on specific engineering conditions).

[0237] Step 22: Improve the NSGA-II multi-objective optimization solution

[0238] An improved non-dominated sorting genetic algorithm (NSGA-II) is used to solve the multi-objective optimization problem in step 21. The algorithm improvements include: (1) adaptive parameter adjustment: crossover probability and mutation probability Dynamically changes with the number of evolutionary generations. , , For the current algebra, (2) Elite retention strategy: The top 10% of the best individuals in each generation are retained and directly enter the next generation; (3) Constraint handling mechanism: The penalty function method is used to handle constraint violations; (4) Local search operator: Neighborhood fine-tuning is performed on the Pareto front solution. Algorithm parameter settings: The population size is preferably 200, the maximum number of generations is preferably 100, and the initial value of the crossover probability is... Initial value of mutation probability The algorithm outputs a Pareto optimal solution set (a set of mutually exclusive, equivalent optimal solutions).

[0239] Step 23: TOPSIS Prioritization

[0240] The Pareto solution set obtained in step 22 is sorted using an improved TOPSIS method. A decision matrix is ​​then constructed. , The number of solutions in the Pareto solution set. For the first The solution is at the th solution. Function values ​​on each objective ( Corresponding to , , The decision matrix is ​​then vectorized:

[0241]

[0242] Constructing a weighted standardization matrix ,in For the target weight, ( (Represents matrix transpose). Determine the positive ideal solution. and negative ideal solution Calculate the Euclidean distance between each solution and the positive and negative ideal solutions:

[0243]

[0244] Final calculation of relative closeness:

[0245]

[0246] By proximity Sort the solutions from largest to smallest, select the spatial coordinates corresponding to the top 3 solutions as the optimal supplementary borehole locations, and then sort them by priority ( The values, from highest to lowest, correspond to priorities 1, 2, and 3, and the output construction sequence is given.

[0247] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0248] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for accurate identification of surrounding rock fractures and 3D network reconstruction integrating deep learning, characterized in that, The reconstruction method includes the following steps: Image preprocessing is performed on the acquired borehole imaging unfolded image to obtain an enhanced image; The enhanced image is input into a pre-trained DE-YOLOv11 deep learning segmentation model for pixel-level binarization segmentation. The backbone network of the DE-YOLOv11 deep learning segmentation model uses a detail enhancement convolution module to replace part of the standard convolution. The detail enhancement convolution module sets up parallel standard convolution paths and high-frequency response paths, and outputs a binarized gap mask of the same size as the original image. The binary crack mask is used for crack classification and parameter extraction: for cracks that conform to the sinusoidal morphology criteria, a sinusoidal model is used for parameter fitting and inversion to output amplitude, period, phase and centerline position parameters; for cracks that do not conform to the sinusoidal morphology criteria, the roughness calculation process is initiated. For each crack in the binary crack mask, the crack area is calculated and layered according to the area size. Based on the layering results, the corresponding sampling strategy is automatically matched to extract the discrete coordinate sequence of the crack contour. The root mean square slope is calculated based on the discrete coordinate sequence, and then the joint roughness coefficient value is obtained by mapping through empirical formula. Based on the geometric feature parameters of multiple fractures and the joint roughness coefficient, a four-component coupled connectivity probability model between fractures is constructed, and the connectivity probability of fracture pairs is output. Based on this, a three-dimensional fracture network spatial structure is constructed. Based on the spatial distribution of existing boreholes, a three-dimensional spatial uncertainty field is constructed, high uncertainty clustering regions are identified, a multi-objective borehole layout optimization model is established to solve for the optimal location of supplementary boreholes, and the three-dimensional coordinates and priority of supplementary boreholes are output.

2. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning as described in claim 1, characterized in that, The image preprocessing includes grayscale and normalization processing, contrast-limited adaptive histogram equalization enhancement, multi-scale Gaussian filtering for noise reduction, and blind image inpainting based on a conditional diffusion model.

3. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The DE-YOLOv11 network also includes an input terminal, a feature fusion network, and a segmentation head; the feature fusion network adopts an FPN+PAN multi-scale feature fusion structure.

4. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The deep learning segmentation model is trained using a composite loss function, which consists of a weighted combination of segmentation loss, boundary loss, detail preservation loss, and regularization term. The segmentation loss is a combination of weighted binary cross-entropy loss and Dice loss. The boundary loss is the L1 norm of the difference between the Sobel gradients of the predicted mask and the real mask. The detail preservation loss is the squared L2 norm of the difference between the high-frequency components of the predicted probability map and the real mask after high-pass transformation.

5. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The output of the detail enhancement convolution module is a weighted sum of the standard convolution result and the result of the high-pass operator convolution of the input feature map multiplied by the learnable scaling parameter.

6. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The sinusoidal morphology is determined as follows: the extracted crack contour is fitted with a sinusoidal model, the goodness of fit is calculated, and when the goodness of fit is greater than the first preset threshold and the amplitude is greater than the second preset pixel threshold, it is determined to be a sinusoidal crack; otherwise, it is determined to be a complex morphology crack.

7. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The specific steps of stratifying and matching sampling strategies based on area size are as follows: when the crack area is less than a first area preset threshold, an equal-interval sampling strategy is matched; when the crack area is between the first area preset threshold and the second area preset threshold, a joint sampling strategy combining curvature adaptive sampling strategy and slope adaptive sampling strategy is adopted; when the crack area is greater than the second area preset threshold, an inflection point adaptive sampling strategy is matched, wherein the inflection point adaptive sampling is based on the second derivative of the contour line to detect inflection points and extreme points for sampling.

8. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The four-component coupled connectivity probability model includes a distance factor, a morphological similarity factor, a joint roughness coefficient compatibility factor, and a spatial uncertainty factor. The distance factor is calculated using a dual exponential decay model based on the spatial distance and depth difference between fracture pairs. The morphological similarity factor is calculated based on the weighted sum of the area similarity, strike similarity, and dip angle similarity of fracture pairs. The joint roughness coefficient compatibility factor is calculated based on the similarity of the joint roughness coefficient values ​​of fracture pairs and the degree of proximity to the optimal connectivity value. The spatial uncertainty factor is calculated based on the comprehensive uncertainty of the spatial location of each fracture pair.

9. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, The multi-objective borehole layout optimization model includes objectives of maximizing uncertainty reduction, maximizing connectivity enhancement, and maximizing information coverage. An improved non-dominated sorting genetic algorithm is used to solve the Pareto optimal solution set, and then the TOPSIS approximation ideal solution sorting method is used to prioritize the Pareto optimal solution set. The spatial coordinates corresponding to the top-ranked solutions are selected as the optimal supplementary borehole positions.

10. The method for accurate identification and three-dimensional network reconstruction of surrounding rock fractures integrating deep learning according to claim 1, characterized in that, It also includes a crack network visualization output step: cracks identified as complex morphological cracks are characterized by piecewise linear interpolation; different cracks in the three-dimensional crack network spatial structure are marked with different colors according to the joint roughness coefficient value, and the connecting lines between cracks are marked with different line widths or colors according to the connectivity probability value; high uncertainty regions are marked with marker symbols.