A multi-step iterative deep learning prediction method for fracture network of concealed rock mass

Abstract

The application discloses a kind of fracture network multi-step iteration deep learning prediction methods for concealed rock mass, it is related to geological fracture structure intelligent identification and prediction technical field, including, extraction exposed rock area fracture network image, analysis exposed rock area in exposed rock mass fracture network image geometric characteristic parameter;Based on geometric characteristic parameter, exposed rock mass fracture network image is divided into different continuous sequence images, according to continuous sequence image construction training set and verification set;Deep learning neural network model is constructed, and training set and verification set are used to execute training process;Different image block size of continuous sequence image and adjacent sequence overlap rate are set as multiple candidate combinations, generate corresponding training set and verification set for each candidate combination, train and verify deep learning neural network model.Deep learning neural network model is constructed to predict the fracture network in concealed area, realize the continuous extrapolation of the fracture structure in masked area, and reduce the cumulative effect of prediction error.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification and prediction technology of geological fracture structures, and in particular to a multi-step iterative deep learning prediction method for fracture networks in concealed rock masses. Background Technology

[0002] The fracture network developed in natural rock masses determines the spatial distribution and evolution characteristics of groundwater, oil and gas, geothermal fluids, and stress-strain fields, and is a key control factor in engineering construction and resource development. Traditional fracture characterization relies on geological exploration and limited observation profiles. However, fracture structures observed from wells, tunnels, and outcrops are often discrete, local, and easily obscured, exhibiting significant spatial incompleteness and discontinuity. Current research struggles to reconstruct continuous and reliable fracture network structures in invisible areas using this data in two-dimensional and three-dimensional dimensions. In recent years, a large number of studies have introduced deep learning into fracture image interpretation, such as using neural networks like U-Net, SegNet, YOLACT++, and SegFormer to perform semantic segmentation or instance segmentation on CT, borehole core, and outcrop images, significantly improving the accuracy and automation of fracture extraction.

[0003] Existing methods have some shortcomings. In terms of identification and prediction, they usually only depict the in-situ fractures based on observational data, lacking consideration of the continuous spatial distribution and topological connectivity of fractures. Especially in actual mining areas, when some areas are obscured or not observed, the current in-situ fracture identification model cannot reliably extrapolate to the obscured areas. In terms of fracture extrapolation, statistical consistency is usually the goal, resulting in a large degree of randomness in the generated results. It is difficult to predict the specific extension path and fine geometric structure of actual fractures, and the ability to depict flow channels and potential damage surfaces is limited. There is no specific design for the spatial translational extrapolation of fractures, and it is impossible to reconstruct the hidden fracture structure in space. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a multi-step iterative deep learning prediction method for fracture networks in concealed rock masses to solve the problems of unreliable extrapolation of occluded areas and difficulty in reconstructing concealed fractures.

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

[0007] This invention provides a multi-step iterative deep learning prediction method for fracture networks in concealed rock masses. The method includes: extracting fracture network images from exposed rock areas and analyzing the geometric feature parameters of the fracture network images within the exposed rock areas; segmenting the exposed rock mass fracture network images into different continuous sequence images based on the geometric feature parameters, and constructing training and validation sets based on the continuous sequence images; constructing a deep learning neural network model and performing a training process using the training and validation sets; setting different image patch sizes and overlap rates of adjacent sequences as multiple candidate combinations, generating corresponding training and validation sets for each candidate combination, training and validating the deep learning neural network model, calculating the cross-union ratio (CUI) between the predicted fracture network and the actual fracture network, and selecting the candidate combination with the largest CUI as the preferred parameter combination; applying the preferred parameter combination to the training of the actual fracture network and the iterative inference prediction process of the deep learning neural network model to generate prediction results for the concealed rock mass fracture network.

[0008] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the specific steps for extracting fracture network images of exposed rock areas and analyzing the geometric feature parameters of the fracture network images of exposed rock masses within the exposed rock areas are as follows.

[0009] The field outcrop photos are used as fracture network images of exposed rock areas. The exposed rock area fracture network images are then converted to grayscale and noise is filtered out to generate preprocessed fracture network images of exposed rock areas.

[0010] Threshold segmentation, edge detection, and deep learning methods were used to extract the fracture network image of exposed rock mass in the preprocessed exposed rock area;

[0011] Based on skeleton extraction and connected component labeling methods, the geometric morphology of each fracture is extracted to form a fracture parameter sample set.

[0012] Statistical analysis is performed on the sample set of fracture parameters to generate geometric characteristic parameters.

[0013] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the geometric feature parameters include fracture length, aperture, orientation, and spacing.

[0014] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the step of segmenting the exposed rock mass fracture network image into different continuous image sequences based on geometric feature parameters includes the following specific steps.

[0015] The height and width of the image patch are defined based on the crack length and spacing;

[0016] Set the horizontal and vertical overlap rates of the image blocks and label them as horizontal overlap rate and vertical overlap rate, respectively;

[0017] Calculate the horizontal sliding step and the vertical sliding step based on the height, width, horizontal overlap rate and vertical overlap rate of the image block, respectively.

[0018] Within the network image of exposed rock fractures, starting from a random location, image blocks are moved from top to bottom at vertical sliding step intervals and from left to right at horizontal sliding step intervals. One image block is captured at each location, and all image blocks are sorted according to their horizontal position to form a continuous sequence of images.

[0019] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the specific steps for constructing training and validation sets based on continuous sequence images are as follows:

[0020] Select image patches at multiple consecutive locations from a continuous sequence of images to generate a crack image sequence;

[0021] The crack image sequence was divided into a training set and a validation set.

[0022] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the specific steps for constructing the deep learning neural network model are as follows:

[0023] Construct an image-to-image end-to-end deep learning neural network model;

[0024] The deep learning neural network model takes image patches as input and outputs a predicted fracture network image with the same size as the input.

[0025] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the specific steps of executing the training process using training and validation sets are as follows:

[0026] The first crack image patch in the same continuous image sequence in the training set As the starting input for neural network prediction, the second crack image patch in the same image sequence is used. As the supervision target of the first step, the prediction output of the first step is obtained. ,calculate With the second crack image patch Single-step loss between ;

[0027] Will The third crack image patch serves as the input for the second step of the neural network's prediction. As the second step of supervision, the second step of prediction output is obtained. ,calculate With the third crack image patch Single-step loss between ;

[0028] The single-step loss is calculated using a combination of Dice loss and cross-entropy loss.

[0029] The single-step losses of each step are weighted and summed according to preset weights to generate the total loss;

[0030] Each time the parameters are updated, the parameters of the deep learning neural network model are backpropagated and iteratively updated based on the total loss.

[0031] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the training process further includes:

[0032] Inference is performed on the validation set after each round of training. The intersection-union ratio, precision, and recall of the predicted output and the corresponding real crack image patch are calculated for each step. The arithmetic mean of the intersection-union ratio of all iteration steps is calculated and used as the comprehensive evaluation index of the training at the current step.

[0033] During training, a learning rate cosine annealing adjustment strategy and an early stopping strategy are used to terminate the training of the deep learning neural network model.

[0034] The optimal network weights are selected based on the maximum value of the average crossover-union ratio across all iterations on the validation set, and these optimal network weights are used as the final model weights for the deep learning neural network model.

[0035] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the specific steps of calculating the cross-union ratio (CUI) between the predicted fracture network and the actual fracture network, and selecting the candidate combination with the largest CUI as the preferred parameter combination, are as follows:

[0036] Multiple image patch sizes, multiple horizontal overlap rates, multiple vertical overlap rates, and model architectures are selected and combined. Training sets and validation sets are constructed according to each combination, and the training process is repeated to generate multiple sets of trained deep learning neural network models.

[0037] For each group of trained deep learning neural network models, perform multi-step iterative predictions on the same crack network image, calculate the cross-union ratio, Dice coefficient, precision, recall and pixel accuracy between the predicted crack network image and the real crack image patch, and record the corresponding training time and inference time.

[0038] The evaluation metrics for all parameter combinations are compared and analyzed. With the goal of maximizing the average intersection-union ratio of all iteration steps, the optimal image patch size, the optimal horizontal overlap rate, the optimal vertical overlap rate, and the optimal model architecture are selected.

[0039] The preferred combination of parameters is the preferred image block size, preferred horizontal overlap rate, preferred vertical overlap rate, and preferred model architecture.

[0040] As a preferred embodiment of the multi-step iterative deep learning prediction method for fracture networks in concealed rock masses described in this invention, the specific steps for applying the preferred parameter combination to the iterative inference prediction process of the real fracture network training and the deep learning neural network model to generate the prediction result of the concealed rock mass fracture network are as follows:

[0041] A sliding window is constructed on the exposed rock mass fracture network image according to the preferred image block size and preferred overlap rate. The exposed rock mass fracture network image is used as the initial input area, and the concealed rock mass area is used as the target area for iterative inference.

[0042] Multiple initial image blocks are extracted at the boundary between the exposed rock mass fracture network image and the concealed rock mass area along the main fracture distribution direction and the vertical direction. Each initial image block is used as the first block of the sequence. The trained deep learning neural network model is used to iteratively generate the predicted fracture network image in the horizontal direction and the fracture distribution direction until the concealed rock mass area is covered.

[0043] Multiple predicted fracture network images obtained from different starting positions and different vertical zones are spatially overlapped and fused. The predicted probability values ​​of the overlapping areas are weighted and averaged to obtain the prediction results of the concealed rock mass fracture network.

[0044] The beneficial effects of this invention are as follows: By establishing an image patch sequence dataset reflecting the actual geometric and statistical characteristics of fractures, a deep learning neural network model is constructed to predict fracture networks in hidden areas. Using image patch size, overlap rate of adjacent training image sequences, and network structure parameters as optimization parameters, the changes in the predicted images under different parameter combinations are analyzed. The optimal parameter combination is determined with the crossover ratio of all iteration steps as the objective, so as to achieve high-precision continuous extrapolation of fracture structures in hidden areas, reduce the cumulative effect of multi-step prediction errors, better maintain the geometric continuity and topological connectivity of fractures in space, and provide more reliable fracture network input for the simulation of fluid flows such as groundwater, oil and gas, and geothermal energy, as well as rock mass stability analysis. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0046] Figure 1 This is a flowchart of a multi-step iterative deep learning prediction method for fracture networks in concealed rock masses.

[0047] Figure 2 The image shows the fracture network used.

[0048] Figure 3 This is a flowchart for deep learning autoregressive training.

[0049] Figure 4 Scatter plots of intersection-union distributions for different parameter combinations and model selections.

[0050] Figure 5 A predicted fracture network diagram for applying the optimal parameter combination and model.

[0051] Figure 6 The curve shows the crossover ratio (CR) as a function of the iterative prediction distance. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-6 This is one embodiment of the present invention, which provides a multi-step iterative deep learning prediction method for fracture networks in concealed rock masses, comprising the following steps:

[0056] S1. Extract the fracture network image of the exposed rock area and analyze the geometric feature parameters of the fracture network image of the exposed rock mass within the exposed rock area.

[0057] It should be noted that the field outcrop photos are used as the fracture network image of the exposed rock area; the fracture network image is preprocessed with grayscale conversion, noise reduction, and contrast enhancement, and the fracture network image of the rock mass surface is extracted using threshold segmentation, edge detection, or deep learning segmentation methods; the geometric shape of the fracture is extracted line by line based on skeleton extraction and connected component labeling, and the length of each fracture is calculated. Opening degree , towards and spacing The parameters are used to form a sample set of crack parameters. Statistical analysis is performed on the sample set of crack parameters to obtain the cumulative frequency distribution of crack parameters, which is used to guide the selection of subsequent image patch size and sliding direction settings, so that the image patch can cover several complete cracks in space without excessively smoothing local details.

[0058] The expression for calculating the crack length is:

[0059] ;

[0060] in, Indicates the first The length of the crack; This represents different fracture numbers, with values ​​ranging from 1 to... ; Indicates the total number of cracks; Indicates the first The skeleton pixel index of the crack ranges from 1 to ; Indicates the first The number of fracture skeleton points in a single fracture; Indicates the first The first crack The pixel coordinates of the skeleton points.

[0061] The expression for calculating the opening degree is:

[0062] ;

[0063] in, Indicates the first The aperture of the crack; and These represent the pixel coordinates of the left and right crack boundaries along the crack normal direction at the crack skeleton point; subscripts... and subscript These represent right and left, referring to the right side and left side, respectively.

[0064] The expression for calculating the direction is,

[0065] ;

[0066] in, Indicates the first The direction of the crack; Represents the first in the fractured skeleton The coordinates of the starting point of the crack; Represents the first in the fractured skeleton The coordinates of the starting point of the crack.

[0067] The expression for calculating the spacing is,

[0068] ;

[0069] ;

[0070] in, Indicates number With number The spacing between cracks; Indicates the selected scanning direction Above, the projected coordinates of the intersection of the crack and the scan line in the scanning direction.

[0071] S2. Based on geometric feature parameters, the exposed rock mass fracture network image is segmented into different continuous sequence images, and a training set and a validation set are constructed based on the continuous sequence images.

[0072] It should be noted that the height H and width W of the image block are determined according to the crack length and spacing, so that each image block contains several complete cracks and the surrounding background.

[0073] Set the horizontal overlap rate of the image blocks and vertical overlap rate And calculate the horizontal sliding step size. and vertical sliding step size In the fracture network image of exposed rock mass, with To move image blocks from top to bottom at intervals, To move image patches from left to right at intervals, an image patch is cropped at each position, and adjacent image patches are arranged into a continuous sequence according to their horizontal position, such that there is at least one horizontal distance between adjacent image patches. The overlap width; for each group of consecutive sequences, select a subsequence of length m, and denot the crack blocks within it sequentially as... , … ,in The starting image patch of the sequence, The target image patches that are sequentially adjacent along the main direction of the crack spread form a training sample sequence for multi-step prediction.

[0074] The training sample sequences derived from the fracture network images of exposed rock mass are divided into training set and validation set according to spatial location or image number, ensuring that the image blocks in the validation set do not overlap or duplicate with the training set to the greatest extent possible.

[0075] The expressions for calculating the horizontal and vertical sliding step sizes are as follows:

[0076] ;

[0077] ;

[0078] in, Indicates the horizontal sliding step size; Indicates the width of the image block; Indicates the horizontal overlap rate; Indicates the vertical sliding step size; Indicates the height of the image block; This indicates the vertical overlap rate.

[0079] S3. Construct a deep learning neural network model and execute the training process using the training and validation sets.

[0080] It should be noted that the image segmentation neural network constructed with a convolutional encoder-decoder structure is used as a deep learning neural network model, denoted as . The deep learning neural network model takes a single-channel crack image patch as input and outputs a predicted crack network image patch of the same size as the input; for each training sample sequence, with As the first step of network input, the first step of prediction satisfies...

[0081] ;

[0082] in, The first step is to predict the fracture network image; This represents a deep learning neural network model; This indicates the starting image block of the sequence.

[0083] The reference standard for the model used in the first step of prediction is: In the second step, it will be activated by the function Processing As network input, the expression is:

[0084] ;

[0085] in, This represents the predicted fracture network image from the second step. This represents the activation function.

[0086] The model reference standard for the second step of prediction is: And so on, extending to This method achieves autoregressive multi-step iterative prediction of the same network parameters across multiple iteration steps.

[0087] To avoid gradients accumulating infinitely over multiple steps and causing training instability, in the second step and thereafter, and Gradient truncation is performed when the gradient is used as the input for the next step.

[0088] For each iteration step Define single-step loss The weighted combination of Dice loss and binary cross-entropy loss is expressed as follows:

[0089] ;

[0090] in, Indicates iteration step The single-step loss function value; This represents the weighting coefficient, with a value ranging from 0 to 1; Indicates the first Predicting the fracture network image step by step; Indicates the first A true image of the crack; Represents the Dice loss function; This represents the binary cross-entropy loss function.

[0091] The expression for Dice loss is,

[0092] ;

[0093] in, express and The Dice loss function; These represent the predicted gap network image and the actual gap image, respectively. and They represent and In the The value at each pixel; Indicates the total number of iterations; To prevent small positive numbers with a denominator of zero; Indicates the pixel index.

[0094] Binary cross-entropy loss The expression is,

[0095] ;

[0096] in, express and The binary cross-entropy loss function.

[0097] The expression for the total loss is,

[0098] ;

[0099] in, This represents the total loss across all iterations in a single training iteration. This represents the loss weight for each iteration step.

[0100] During training, minimize For network parameters Backpropagation and updates are performed, while cosine annealing and early stopping strategies are used to control the training process. Training is terminated when the average intersection-union ratio (IU) of all iterations on the validation set no longer improves or the validation loss no longer decreases over several consecutive rounds. The results at this point are then recorded. As the final model parameters.

[0101] During the training and validation phases, to evaluate the segmentation performance of the deep learning neural network model at each iteration step, a set of predicted fracture network images is calculated. Collection of real crack images Crossover ratio between and Dice coefficient The expression is,

[0102] ;

[0103] ;

[0104] in, and Describe the first The cross-union ratio and Dice coefficient predicted by step-by-step iteration.

[0105] The arithmetic mean of the intersection-union ratios of each iteration step is used as one of the main indicators for parameter combination and optimization, and the expression is:

[0106] ;

[0107] in, This represents the arithmetic mean of the intersection-union ratios of all iteration steps.

[0108] S4. Set different image patch sizes and adjacent sequence overlap rates of continuous sequence images as multiple candidate parameter combinations. Generate corresponding training and validation sets for each candidate parameter combination, train and validate deep learning neural network models, calculate the crossover ratio between the predicted fracture network and the real fracture network, and use the candidate combination with the largest crossover ratio as the preferred parameter combination.

[0109] It should be noted that, under different image patch sizes, overlap rates of adjacent image sequences, and network structure parameters, a multi-step training process was repeatedly executed to obtain multiple sets of trained deep learning neural network models. Multi-step iterative prediction was performed on the same fracture network image, calculating the intersection-union ratio, Dice coefficient, precision, recall, and pixel accuracy between the predicted fracture network and the real fracture network in the masked region, and recording the corresponding training and inference times. A comparative analysis of the evaluation metrics for all parameter combinations was conducted to... With the goal of maximizing, we select the optimal combination of image patch size, horizontal overlap rate, vertical overlap rate, and model architecture, and set these as standard parameters in engineering applications.

[0110] S5. The optimal parameter combination is applied to the iterative inference prediction process of the real fracture network training and deep learning neural network model to generate the prediction result of the hidden rock mass fracture network.

[0111] It should be noted that a sliding window is constructed on the exposed rock mass fracture network image according to the optimized image patch size and overlap rate. The exposed rock mass fracture network image is used as the initial input region, and the concealed rock mass region is used as the target region for iterative inference. Multiple initial image patches are extracted at the boundary between the exposed rock mass fracture network image and the concealed rock mass region along the main fracture propagation direction and the vertical direction. Each initial image patch is used as the first patch of the sequence. The trained deep learning neural network model iteratively generates a predicted fracture network image in the horizontal and fracture propagation directions until the concealed rock mass region is covered. Multiple predicted fracture network images obtained from different starting positions and different vertical zones are spatially overlapped and fused. The predicted probability values ​​of the overlapping areas are weighted and averaged to obtain the predicted result of the concealed rock mass fracture network.

[0112] It should also be noted that, to better demonstrate the process of this invention, a simulated fracture network is now generated based on the self-similarity of rock fracture growth and fracture merging characteristics, such as... Figure 2 As shown.

[0113] To simulate the growth and distribution of fracture networks in real rock masses, the simulated fracture network has two main orientations: 60° and 120°. The minimum fracture length in the 60° orientation is 150 pixels, the maximum fracture length is 300 pixels, the minimum fracture spacing is 45 pixels, and the maximum spacing is 85 pixels. The minimum fracture length in the 120° orientation is 120 pixels, the maximum fracture length is 250 pixels, the minimum fracture spacing is 30 pixels, and the maximum spacing is 60 pixels.

[0114] Will Figure 2The simulated fracture network shown is divided into two parts horizontally. The left part represents the fracture network of exposed rock mass, and the right part represents the fracture network of concealed rock mass. A suitable fracture network image sequence is selected from the left image for multi-step autoregressive iterative training. The trained deep learning neural network model is used to predict from the left image to the right, and the prediction results are compared with the existing right region.

[0115] according to Figure 2 The interrelationships and combination characteristics of the fracture network were investigated. The candidate values ​​for the iterative image size [H,W] parameter were [256,256], [384,384], and [512,512] to consider the local and combined growth patterns of the fractures. To cover typical cases of fracture image sequences ranging from non-overlapping to highly overlapping, and to examine the impact of overlap on the model's predictive performance, the horizontal overlap rate of the image patches was considered. and vertical overlap rate The parameter values ​​are candidate values ​​of 0, 0.2, 0.5, and 0.8; the number of iterative image patches n in a continuous sequence is 3; and the loss weight for different iteration steps is assigned a value of 1. To fully cover the exposed rock mass fracture network region and meet the model convergence stability requirements, the training set contains 1000 training images, and the validation set contains 600 validation images.

[0116] During the parameter optimization training process, the maximum number of iterations for training the deep learning neural network model is 200 to ensure that the deep learning neural network model is fully trained under all parameter combinations. The optimizer uses AdamW, which is widely used in vision tasks, to accelerate the training process, and the initial learning rate is set to... Furthermore, a cosine annealing strategy is used to decay the learning rate in order to accelerate the convergence of the deep learning neural network model and prevent overfitting.

[0117] It should be noted that the parameter settings above, like the training hyperparameters, are only a representative set of settings given in this study. In practical applications, the number of iteration steps, learning rate, and its decay strategy can be re-searched and optimized according to the data scale, network structure, and computing resources.

[0118] In summary, this invention achieves high-precision continuous extrapolation of the fracture structure in concealed areas by: establishing an image patch sequence dataset reflecting the actual geometric and statistical characteristics of fractures; constructing a deep learning neural network model for predicting fracture networks in hidden areas; using image patch size, overlap rate of adjacent training image sequences, and network structure parameters as optimization parameters; analyzing the changes in the predicted images under different parameter combinations; and determining the optimal parameter combination with the cross-union ratio of all iteration steps as the objective. This reduces the cumulative effect of multi-step prediction errors, better maintains the geometric continuity and topological connectivity of fractures in space, and provides a more reliable fracture network input for simulating the flow of fluids such as groundwater, oil and gas, and geothermal energy, as well as rock mass stability analysis.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-step iterative deep learning prediction method for fracture networks in concealed rock masses, characterized in that: include, Extract fracture network images from exposed rock areas and analyze the geometric feature parameters of fracture network images of exposed rock masses within exposed rock areas; Based on geometric feature parameters, the exposed rock mass fracture network image is segmented into different continuous image sequences. The specific steps are as follows. The height and width of the image patch are defined based on the crack length and spacing; Set the horizontal and vertical overlap rates of the image blocks and label them as horizontal overlap rate and vertical overlap rate, respectively; Calculate the horizontal sliding step and the vertical sliding step based on the height, width, horizontal overlap rate and vertical overlap rate of the image block, respectively. Within the network image of exposed rock fractures, starting from a random location, image blocks are moved from top to bottom at vertical sliding step intervals and from left to right at horizontal sliding step intervals. One image block is cropped at each location, and all image blocks are sorted according to their horizontal position to form a continuous sequence of images. Training and validation sets are constructed based on the continuous sequence of images. Build a deep learning neural network model and execute the training process using the training and validation sets; Different image patch sizes and adjacent sequence overlap rates of consecutive image sequences are set as multiple candidate combinations. For each candidate combination, a corresponding training set and validation set are generated. A deep learning neural network model is trained and validated. The crossover ratio between the predicted fracture network and the real fracture network is calculated. The candidate combination with the largest crossover ratio is used as the preferred parameter combination. The optimal parameter combination is applied to the iterative inference prediction process of real fracture network training and deep learning neural network model to generate the prediction results of hidden rock mass fracture network.

2. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 1, characterized in that: The specific steps for extracting the fracture network image of the exposed rock area and analyzing the geometric feature parameters of the fracture network image of the exposed rock mass within the exposed rock area are as follows. The field outcrop photos are used as fracture network images of exposed rock areas. The exposed rock area fracture network images are then converted to grayscale and noise is filtered out to generate preprocessed fracture network images of exposed rock areas. Threshold segmentation, edge detection, and deep learning methods were used to extract the fracture network image of exposed rock mass in the preprocessed exposed rock area; Based on skeleton extraction and connected component labeling methods, the geometric morphology of each fracture is extracted to form a fracture parameter sample set. Statistical analysis is performed on the sample set of fracture parameters to generate geometric characteristic parameters.

3. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 2, characterized in that: The geometric feature parameters include crack length, opening, orientation, and spacing.

4. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 3, characterized in that: The specific steps for constructing the training and validation sets based on continuous image sequences are as follows. Select image patches at multiple consecutive locations from a continuous sequence of images to generate a crack image sequence; The crack image sequence was divided into a training set and a validation set.

5. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 4, characterized in that: The specific steps for constructing the deep learning neural network model are as follows. Construct an image-to-image end-to-end deep learning neural network model; The deep learning neural network model takes image patches as input and outputs a predicted fracture network image with the same size as the input.

6. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 5, characterized in that: The training process using the training and validation sets involves the following specific steps. The first crack image patch in the same continuous image sequence in the training set As the starting input for neural network prediction, the second crack image patch in the same image sequence is used. As the supervision target of the first step, the prediction output of the first step is obtained. ,calculate With the second crack image patch Single-step loss between ; Will The third crack image patch serves as the input for the second step of the neural network's prediction. As the second step of supervision, the second step of prediction output is obtained. ,calculate With the third crack image patch Single-step loss between ; The single-step loss is calculated using a combination of Dice loss and cross-entropy loss. The single-step losses of each step are weighted and summed according to preset weights to generate the total loss; Each time the parameters are updated, the parameters of the deep learning neural network model are backpropagated and iteratively updated based on the total loss.

7. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 6, characterized in that: The training process also includes, Inference is performed on the validation set after each round of training. The intersection-union ratio, precision, and recall of the predicted output and the corresponding real crack image patch are calculated for each step. The arithmetic mean of the intersection-union ratio of all iteration steps is calculated and used as the comprehensive evaluation index of the training at the current step. During training, a learning rate cosine annealing adjustment strategy and an early stopping strategy are used to terminate the training of the deep learning neural network model. The optimal network weights are selected based on the maximum value of the average crossover-union ratio across all iterations on the validation set, and these optimal network weights are used as the final model weights for the deep learning neural network model.

8. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 7, characterized in that: The calculation of the cross-union ratio (CUI) between the predicted fracture network and the actual fracture network, and the selection of the candidate combination with the largest CUI as the preferred parameter combination, involves the following specific steps. Multiple image patch sizes, horizontal overlap rates, vertical overlap rates, and model architectures are selected and combined. Training and validation sets are constructed for each combination, and the training process is repeated to generate multiple sets of trained deep learning neural network models. For each group of trained deep learning neural network models, perform multi-step iterative predictions on the same crack network image, calculate the cross-union ratio, Dice coefficient, precision, recall and pixel accuracy between the predicted crack network image and the real crack image patch, and record the corresponding training time and inference time. The evaluation metrics for all parameter combinations are compared and analyzed. With the goal of maximizing the average intersection-union ratio of all iteration steps, the optimal image patch size, the optimal horizontal overlap rate, the optimal vertical overlap rate, and the optimal model architecture are selected. The preferred combination of parameters is the preferred image block size, preferred horizontal overlap rate, preferred vertical overlap rate, and preferred model architecture.

9. The multi-step iterative deep learning prediction method for fracture networks in concealed rock masses as described in claim 8, characterized in that: The process of applying the optimized parameter combination to the iterative inference and prediction of the real fracture network and the deep learning neural network model to generate the prediction result of the concealed rock mass fracture network is as follows: A sliding window is constructed on the exposed rock mass fracture network image according to the preferred image block size and preferred overlap rate. The exposed rock mass fracture network image is used as the initial input area, and the concealed rock mass area is used as the target area for iterative inference. Multiple initial image blocks are extracted at the boundary between the exposed rock mass fracture network image and the concealed rock mass area along the main fracture distribution direction and the vertical direction. Each initial image block is used as the first block of the sequence. The trained deep learning neural network model is used to iteratively generate the predicted fracture network image in the horizontal direction and the fracture distribution direction until the concealed rock mass area is covered. Multiple predicted fracture network images obtained from different starting positions and different vertical zones are spatially overlapped and fused. The predicted probability values ​​of the overlapping areas are weighted and averaged to obtain the prediction results of the concealed rock mass fracture network.

Images