A top coal fracture fractal analysis method based on CT imaging

Abstract

This invention discloses a fractal analysis method for top coal seam fractures based on CT imaging, belonging to the field of fracture fractal detection. The method includes the following steps: based on CT scan data of top coal samples, a three-dimensional digital model containing fracture information is obtained through three-dimensional reconstruction; the three-dimensional digital model is processed using image processing methods to extract the fracture network and calculate topological parameters; the fracture network is processed using fractal analysis methods to calculate the global fractal dimension and local dimensional statistical features; the topological parameters, global fractal dimension, and local dimensional statistical features are fused to form a multi-scale feature vector; based on the multi-scale feature vector, a machine learning model is used to obtain the top coal fracture classification result. This invention overcomes the limitations of traditional methods that rely on a single indicator and are highly subjective, achieving objective and accurate identification of the development level of top coal fractures.

Description

Technical Field

[0001] This invention belongs to the field of fracture fractal detection technology, and particularly relates to a fractal analysis method for top coal fractures based on CT imaging. Background Technology

[0002] The development of top coal fractures directly affects coal mining efficiency, mine safety, and resource recovery rate. Accurate identification and classification of top coal fractures play an irreplaceable role in optimizing mining schemes, preventing roof collapse accidents, and improving resource recovery rates. However, traditional top coal fracture analysis methods have many limitations and cannot meet the high requirements of modern coal mining for the accuracy and efficiency of fracture analysis.

[0003] Traditional research on top coal seam fractures primarily relies on field observations, borehole sampling, and some simple physical experiments. Field observations often only yield limited, localized fracture information, failing to comprehensively reflect the complex fracture network structure within the top coal seam. While borehole sampling can obtain core samples at a certain depth, the sampling process may damage the fracture structure, and the limited sample size cannot guarantee representativeness. Furthermore, the data obtained by these traditional methods are mostly two-dimensional, unable to accurately reflect the distribution and connectivity characteristics of top coal seam fractures in three-dimensional space, posing significant challenges for quantitative analysis and classification of fractures.

[0004] In fracture analysis, traditional methods primarily focus on describing the macroscopic morphological characteristics of fractures, such as their length, width, and dip angle, lacking in-depth analysis of the overall structural features of the fracture network. Quantitative analysis of fractures typically employs simple statistical methods, such as fracture density calculation, which fail to adequately consider the interrelationships between fractures and the complexity of the fracture network. Regarding fracture classification, traditional methods are mostly based on empirical judgment, lacking scientific classification standards and effective algorithms, making it difficult to guarantee the accuracy and reliability of the classification results.

[0005] Therefore, this invention proposes a fractal analysis method for top coal fractures based on CT imaging. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a fractal analysis method for top coal fractures based on CT imaging, thereby resolving the issues present in the prior art.

[0007] To achieve the above objectives, this invention provides a fractal analysis method for top coal fractures based on CT imaging, comprising:

[0008] Based on CT scan data of top coal samples, a three-dimensional digital model containing fracture information is obtained through three-dimensional reconstruction.

[0009] The three-dimensional digital model is processed using image processing methods to extract the fracture network and calculate the topological parameters;

[0010] The fracture network was processed using fractal analysis to calculate the global fractal dimension and local dimension statistical characteristics.

[0011] The topological parameters, global fractal dimension, and local dimension statistical features are fused to form a multi-scale feature vector;

[0012] Based on the multi-scale feature vectors, the classification results of top coal fractures are obtained through machine learning model processing.

[0013] Optionally, the process of obtaining a three-dimensional digital model containing fracture information includes:

[0014] Based on CT scan data of top coal samples, a two-dimensional tomographic image sequence was obtained using a high-resolution X-ray scanning system;

[0015] The two-dimensional tomographic image sequence is subjected to adaptive filtering and noise reduction processing to obtain the noise-reduced image;

[0016] The denoised image is processed using a multi-threshold segmentation algorithm to obtain the segmented image;

[0017] The segmented image is reconstructed using a region growing algorithm to generate a preliminary 3D model.

[0018] The preliminary three-dimensional model is subjected to anisotropic diffusion smoothing to obtain an optimized three-dimensional digital model.

[0019] Optionally, the process of processing the three-dimensional digital model using image processing methods to extract the fracture network and calculate the topological parameters includes:

[0020] The three-dimensional digital model is processed using a deep learning-based three-dimensional segmentation network to obtain a binarized crack segmentation result;

[0021] The segmentation results are subjected to skeleton extraction and midline transformation to obtain a simplified fracture network skeleton.

[0022] Based on graph theory, the skeleton of the fracture network is abstracted into a topological graph structure, where vertices represent fracture intersections or endpoints, and edges represent fracture branches.

[0023] Based on the aforementioned topological graph structure, topological parameters including fracture rate, fracture density, average aperture, orientation anisotropy index, and connectivity coefficient are calculated.

[0024] Optionally, during the process of processing the three-dimensional digital model using a deep learning-based three-dimensional segmentation network, a slit segmentation neural network based on the 3D U-Net architecture is constructed to perform slit segmentation on the three-dimensional digital model; the slit segmentation neural network is an encoder-decoder structure, the encoder includes several downsampling modules, and the decoder includes several upsampling modules.

[0025] The process of segmenting the three-dimensional digital model using a 3D U-Net-based slit segmentation neural network includes:

[0026] The three-dimensional digital model containing the crack information is input into the downsampling module of the encoder to obtain a low-resolution high-dimensional feature map; wherein, several downsampling modules are connected in series, and the three-dimensional digital model undergoes two 3×3×3 convolution operations in sequence after passing through each downsampling module. After each convolution, the ReLU activation function is used for nonlinear transformation, and then a 2×2×2 max pooling operation is performed to obtain a feature map with half the size.

[0027] The low-resolution high-dimensional feature map is input into the upsampling module of the decoder to obtain the decoder output feature map; wherein, several upsampling modules are connected in series, and the low-resolution high-dimensional feature map is transposed and convolved sequentially after passing through each upsampling module to enlarge the size of the low-resolution high-dimensional feature map by one time to obtain an enlarged feature map, and the enlarged feature map is concatenated with the feature map passed through the skip connection in the encoder path to obtain a concatenated feature map;

[0028] The decoder output feature map is processed using a 1×1×1 convolutional layer to obtain a probability map with the same size as the input 3D digital model.

[0029] Optionally, the process of obtaining the statistical characteristics of the global fractal dimension and the local dimension includes:

[0030] The global fractal dimension of the fracture network was calculated using the three-dimensional box counting method.

[0031] The three-dimensional model of the fracture network is traversed using the sliding window method, and the local box dimension is calculated at each window position to form a local dimension distribution matrix;

[0032] Statistical features are calculated based on the local dimension distribution matrix;

[0033] The global fractal dimension and the statistical feature quantity are combined to form a fractal feature vector.

[0034] Optionally, the statistical features include: mean, standard deviation, skewness, and peak value.

[0035] Optionally, the process of fusing the topological parameters and the fractal feature vectors to form a multi-scale feature vector includes:

[0036] The topological parameter feature set and the fractal feature set are respectively standardized and preprocessed to obtain the preprocessed topological parameter feature set and the preprocessed fractal feature set;

[0037] The preprocessed topological parameter feature set and the preprocessed fractal feature set are concatenated to form an initial feature vector;

[0038] Based on the initial feature vector, an attention mechanism is used to calculate the weight coefficient of each feature and generate a weighted feature vector;

[0039] Construct a feature interaction network and extract feature interaction terms to obtain interaction features;

[0040] The weighted feature vector is concatenated with the interaction feature to form an enhanced feature vector;

[0041] The enhanced feature vector is then subjected to dimensionality reduction processing to generate the final multi-scale fused feature vector.

[0042] Optionally, the process of obtaining the top coal fracture classification result by processing the multi-scale feature vector through a machine learning model includes:

[0043] The multi-scale feature vectors are input into a pre-trained hybrid classification model for processing. The hybrid classification model includes a parallel gradient boosting decision tree module and a deep neural network module.

[0044] The outputs of the gradient boosting decision tree module and the deep neural network module are weighted and fused to generate the final probability distribution output.

[0045] The development level classification result of top coal fractures is determined based on the probability distribution output.

[0046] The present invention also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described thereon.

[0047] The present invention also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0048] Compared with the prior art, the present invention has the following advantages and technical effects:

[0049] This invention achieves a comprehensive quantitative characterization of top coal seam fracture structure from macroscopic statistics to microscopic heterogeneity by integrating multi-scale fractal features and topological parameters, significantly improving the accuracy and depth of fracture assessment. A machine learning model is used to intelligently classify the integrated features, overcoming the limitations of traditional methods that rely on single indicators and are highly subjective, thus enabling objective and accurate identification of top coal seam fracture development levels. This method provides a reliable quantitative basis for predicting top coal seam caving, and has significant application value for improving the efficiency of fully mechanized longwall mining and coal recovery rates. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0052] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0054] like Figure 1 As shown, this embodiment provides a fractal analysis method for top coal fractures based on CT imaging, including the following steps:

[0055] S1. Based on the CT scan data of the top coal sample, a three-dimensional digital model containing fracture information is obtained through three-dimensional reconstruction. This process includes: acquiring a two-dimensional tomographic image sequence using a high-resolution X-ray scanning system based on the CT scan data of the top coal sample; performing adaptive filtering and noise reduction on the two-dimensional tomographic image sequence to obtain a denoised image; processing the denoised image using a multi-threshold segmentation algorithm to obtain a segmented image; performing three-dimensional reconstruction on the segmented image based on a region growing algorithm to generate a preliminary three-dimensional model; and performing anisotropic diffusion smoothing on the preliminary three-dimensional model to obtain an optimized three-dimensional digital model.

[0056] S2. Image processing methods are used to process the 3D digital model, extract the fracture network, and calculate the topological parameters. This process includes: processing the 3D digital model using a deep learning-based 3D segmentation network to obtain a binarized fracture segmentation result; extracting the skeleton and performing a median transformation on the segmentation result to obtain a simplified fracture network skeleton; abstracting the fracture network skeleton into a topological graph structure based on graph theory, where vertices represent fracture intersections or endpoints, and edges represent fracture branches; and calculating the topological parameters, including fracture rate, fracture density, average aperture, orientation anisotropy index, and connectivity coefficient, based on the topological graph structure.

[0057] Furthermore, as a specific implementation of this embodiment, an image processing method is used to process the three-dimensional digital model. By combining deep learning and mathematical morphology algorithms, the gap network is accurately extracted and its multi-dimensional topological parameters are quantitatively calculated. Specifically, this includes constructing a gap segmentation network to achieve accurate voxel-level extraction, and abstracting the gap network into the relationship between nodes and edges based on a graph theory model to calculate the topological parameters.

[0058] S2.1 Precise Crack Segmentation Based on Deep Learning

[0059] A 3D U-Net-based neural network for slit segmentation was constructed. This network employs an encoder-decoder structure. The encoder consists of four downsampling modules, each containing two 3×3×3 convolutional layers and a ReLU activation function, followed by a 2×2×2 max-pooling layer for downsampling. The decoder uses transposed convolutions for upsampling and fuses feature maps of the corresponding scale from the encoder via skip connections. The network ultimately outputs the predicted probability of each voxel belonging to a slit through a 1×1×1 convolutional layer and a sigmoid activation function. The network training uses a combined loss function: L = α·L dice + β·L focal L dice The Dice loss function is used to handle the foreground-background imbalance problem, with a weighting coefficient α of 0.6; L focal The Focal loss function is used to focus on hard-to-classify samples, with a weight coefficient β of 0.4 and a modulation factor γ of 2.0.

[0060] Furthermore, in the process of processing the 3D digital model using a deep learning-based 3D segmentation network, a slit segmentation neural network based on the 3D U-Net architecture is constructed to perform slit segmentation on the 3D digital model; the slit segmentation neural network is an encoder-decoder structure, with the encoder including several downsampling modules and the decoder including several upsampling modules.

[0061] The process of using a 3D U-Net-based gap segmentation neural network to segment gaps in a 3D digital model includes: inputting the 3D digital model containing gap information into the downsampling module of the encoder to obtain a low-resolution high-dimensional feature map; wherein several downsampling modules are connected in series, and the 3D digital model undergoes two 3×3×3 convolution operations sequentially after each downsampling module, and after each convolution, a ReLU activation function is used for nonlinear transformation, followed by a 2×2×2 max pooling operation to obtain a feature map with halved size; inputting the low-resolution high-dimensional feature map into the upsampling module of the decoder to obtain the decoder output feature map; wherein several upsampling modules are connected in series, and the low-resolution high-dimensional feature map undergoes a transposed convolution operation sequentially after each upsampling module, enlarging the size of the low-resolution high-dimensional feature map by half to obtain an enlarged feature map, and concatenating the enlarged feature map with the feature map passed through skip connections in the encoder path to obtain a concatenated feature map; and processing the decoder output feature map based on a 1×1×1 convolutional layer to obtain a probability map with the same size as the input 3D digital model.

[0062] S2.2 Extraction and simplification of the fracture network skeleton.

[0063] The binary segmentation results output by the neural network are used to extract the skeleton using a three-dimensional thinning algorithm. Boundary voxels that meet specific conditions are removed iteratively, while the center line of the crack with a width of one pixel is retained. Then, the axial transformation algorithm is applied to calculate the shortest distance from each point in the skeleton to the crack boundary, and the crack opening information w(x,y,z) = 2·d(x,y,z) is obtained, where d(x,y,z) represents the shortest distance from the skeleton point (x,y,z) to the boundary.

[0064] S2.3 Topological modeling based on graph theory.

[0065] The extracted skeleton is modeled as an undirected graph G = (V, E, A), where the vertex set V represents the fracture intersections or endpoints, the edge set E represents the fracture branches, and the attribute set A contains the geometric features of each edge: length l i Average opening w i and direction vector θ i Length l i The average aperture w is calculated by accumulating the Euclidean distances of the skeleton point sequence. i The arithmetic mean of the opening values ​​of all skeleton points on this branch; the direction vector θ i The spatial orientation of the skeleton points is obtained by fitting the data through principal component analysis.

[0066] S2.4 Calculation of topology parameters.

[0067] Based on the topology graph G, calculate the following parameter: fissure ratio φ = V pores / V total× 100%, where V pores V is the total volume of the fracture. total The total volume of the sample; the fracture density ρ = N / V total Where N is the total number of fracture branches; average aperture μ w = (Σw i ) / N; Orientation anisotropy index η = λ max / λ min , where λ max and λ min These are the maximum and minimum eigenvalues ​​of the direction vector covariance matrix, respectively; the connectivity coefficient λ = (S max / S total ) × C global Where S_max is the number of edges in the maximum connected component, S total Let C be the total number of sides. global This is the global clustering coefficient, reflecting the degree of clustering of network connections.

[0068] S3. Fractal analysis is used to process the fracture network, and the global fractal dimension and local dimension statistical characteristics are calculated. The process of obtaining the global fractal dimension and local dimension statistical characteristics includes: calculating the global fractal dimension of the fracture network using the three-dimensional box counting method; traversing the three-dimensional model of the fracture network using the sliding window method, calculating the local box dimension at each window position to form a local dimension distribution matrix; calculating statistical characteristic quantities based on the local dimension distribution matrix; and combining the global fractal dimension and statistical characteristic quantities to form a fractal feature vector.

[0069] Furthermore, as a specific implementation of this embodiment, a fractal analysis method is used to process the fracture network. By introducing multifractal theory and adaptive sliding window technology, the global fractal dimension characterizing the overall complexity of the fracture system and the local dimensionality statistical features reflecting its spatial heterogeneity are calculated. Specifically, this includes calculating the global dimension based on an improved three-dimensional box counting method, calculating the local dimensionality distribution by constructing a scale-adaptive sliding window matrix, extracting its multifractal spectrum feature parameters, and finally fusing them to form a high-dimensional fractal feature vector.

[0070] S3.1 Improved three-dimensional box counting method for global dimension calculation.

[0071] The global fractal dimension D is calculated using a weighted box counting method that considers the anisotropy of the fracture direction. global With side length l i A cubic mesh covers the binary fracture model, where l i =2 k (k=1,2,…,n); for each scale l i Count the number of boxes N(l) containing cracks. iAnd calculate the entropy value of the crack direction for each box:

[0072] ;

[0073] Where, p j This represents the proportion of the fracture voxels within the box in the j-th directional interval, where M is the total number of directional intervals. The weighted box number N... w (l i )=∑(W i ⋅N i ), through linear fitting: The slope of D is obtained global .

[0074] S3.2 Scale-adaptive sliding window local dimension calculation.

[0075] A scale-adaptive sliding window algorithm is constructed. First, the optimal window size is calculated based on the fracture density distribution.

[0076] ;

[0077] Where ρ(x,y,z) is the local fracture density. A variable-size sliding window (size range L) is used. opt (±ΔL) Traverse the 3D model and calculate the local box dimension D at each window position. local (x,y,z) generates a three-dimensional local dimension distribution matrix D. map The missing areas are filled using Kriging interpolation.

[0078] S3.3 Multifractal Spectrum Feature Extraction.

[0079] Based on the local dimension distribution matrix D map Calculate the multifractal spectrum. First, calculate the partition function: .in, is the weighting factor, and q is the order (a sequence of integers from -10 to 10). Through linear fitting: Obtain the quality index Finally, the multifractal spectrum is obtained through the Legendre transform.

[0080] S3.4 Construction of fractal eigenvectors.

[0081] The global fractal dimension D global With multifractal spectrum characteristic parameters Δα, B, f max and local dimensional statistical features (mean μ) d Standard deviation σ d skewness S d kurtosis K d The combination of these elements forms a nine-dimensional fractal eigenvector: .

[0082] S4. The topological parameters, global fractal dimension, and local dimensionality statistical features are fused to form a multi-scale feature vector. The process of forming the multi-scale feature vector includes: standardizing and preprocessing the topological parameter feature set and the fractal feature set to obtain preprocessed topological parameter feature set and preprocessed fractal feature set, respectively; concatenating the preprocessed topological parameter feature set and preprocessed fractal feature set to form an initial feature vector; based on the initial feature vector, using an attention mechanism to calculate the weight coefficient of each feature and generate a weighted feature vector; constructing a feature interaction network and extracting feature interaction terms to obtain interaction features; concatenating the weighted feature vector with the interaction features to form an enhanced feature vector; and performing dimensionality reduction on the enhanced feature vector to generate the final multi-scale fused feature vector.

[0083] Furthermore, as a specific implementation of this embodiment, topological parameters, global fractal dimension, and local dimensionality statistical features are fused to form a multi-scale feature vector. By constructing a multimodal feature fusion network based on a gating mechanism and structured interaction modeling, adaptive weighting and deep interaction of features with different physical meanings and scales are achieved. Specifically, this includes using a modality-specific standardization method, feature weighting based on a gating attention mechanism, extraction of higher-order interaction terms, and manifold learning dimensionality reduction, ultimately generating a low-dimensional fusion feature vector with stronger representational capabilities.

[0084] S4.1 Modal Adaptive Standardization. Modal-specific standardization is performed on both the topological parameter feature set and the fractal feature set. The topological parameter feature set includes five parameters: fracture rate, fracture density, average aperture, orientation anisotropy index, and connectivity coefficient. The fractal feature set includes eight parameters: global fractal dimension, multifractal spectral width, spectral symmetry parameter, maximum dimension, and the mean, standard deviation, skewness, and kurtosis of the local dimensions. Standardization is performed using the mean and standard deviation of each mode to eliminate the influence of dimensions.

[0085] Let the topological parameter feature set be T, and the fractal feature set be F. A mode-specific normalization method is used: .in, , and , These are the mean and standard deviation of the topological parameters and fractal features on the training set, respectively.

[0086] S4.2 Feature Weighting Based on Gated Attention. A gated attention mechanism is used to calculate feature importance weights. A gated network containing a learnable weight matrix and a bias vector is constructed. The attention score for each feature is calculated using the sigmoid activation function, and the weight coefficients are obtained after softmax normalization. These weight coefficients are then used to weight the standardized features, generating a weighted feature vector.

[0087] S4.3 Structured Feature Interaction Modeling. A decomposition machine model is used to extract higher-order interactions between features, and the inner product of the latent vectors between each pair of features is calculated as the interaction strength. Simultaneously, significant interactive feature terms are selected based on correlation coefficient and mutual information, retaining those with a correlation coefficient greater than 0.3 and mutual information greater than 0.2. The higher-order interactive features output by the decomposition machine are concatenated with the selected significant interactive terms into a weighted feature vector to form an enhanced feature vector.

[0088] S4.4 Manifold Learning Dimensionality Reduction. An isometric mapping manifold learning algorithm is used for nonlinear dimensionality reduction. The original manifold structure is preserved by calculating the geodesic distance between feature vectors. The final dimension is determined based on the cumulative contribution rate of the feature values, retaining principal components with a cumulative contribution rate exceeding 95%, generating the final multi-scale fused feature vector. This feature vector retains the main information of the original features while enhancing their representational power.

[0089] S5. Based on multi-scale feature vectors, the classification results of top coal fractures are obtained through machine learning model processing. The process of obtaining the classification results of top coal fractures through machine learning model processing includes: inputting multi-scale feature vectors into a pre-trained hybrid classification model for processing, the hybrid classification model including a parallel gradient boosting decision tree module and a deep neural network module; weighting and fusing the output results of the gradient boosting decision tree module and the deep neural network module to generate the final probability distribution output; and determining the development level classification result of top coal fractures based on the probability distribution output.

[0090] Furthermore, as a specific implementation of this embodiment, based on multi-scale feature vectors, the system integrates a multimodal machine learning model with an adaptive decision fusion mechanism to obtain accurate classification results of top coal fracture development levels. Specifically, a hybrid architecture of gradient boosting decision tree and deep neural network in parallel is adopted, combined with a confidence-based dynamic weighted fusion strategy, and transfer learning and uncertainty quantification are introduced to construct an intelligent classification system that can fully mine feature information and ensure classification reliability.

[0091] First, a hybrid classification model architecture is constructed. This model consists of a gradient boosting decision tree module and a deep neural network module connected in parallel. The gradient boosting decision tree module uses multiple regression trees for ensemble learning, iteratively optimizing the prediction results through the gradient boosting algorithm. The tree depth is set to 6 to 10 layers, the learning rate is 0.05, and the number of subtrees is 100. The deep neural network module adopts a three-layer fully connected network structure. The input layer dimension is the same as the multi-scale feature vector dimension, the dimensions of the two hidden layers are 64 and 32 respectively, and the output layer dimension is the same as the number of crack categories. The ReLU activation function is used, and Dropout regularization with a dropout rate of 0.2 is employed to prevent overfitting.

[0092] Next, a dynamic weighted fusion strategy is implemented. The outputs of the two modules are fused using a dynamic weighting method based on confidence. First, the probability distributions of the gradient boosting decision tree module and the deep neural network module are calculated separately. Then, weights are dynamically assigned based on the classification confidence of each module on the validation set. The final fusion result is a weighted sum of the outputs of the two modules, where the weights are dynamically adjusted based on the classification accuracy of each module's k most recent samples. The initial weights are set to 0.5, and the sum of the two weights is guaranteed to be 1.

[0093] Then, transfer learning and model optimization methods were employed. First, pre-training was performed using fracture data from other geological conditions, followed by fine-tuning using data from the target mining area to improve the model's generalization ability. The training process adopted a phased strategy: first, the deep neural network module was trained by fixing the gradient boosting decision tree parameters using the Adam optimizer with a learning rate of 0.001; then, the entire hybrid model was jointly trained, with the learning rate dynamically adjusted using cosine annealing. The model loss function was cross-entropy loss plus an L2 regularization term.

[0094] Uncertainty quantification and classification decision-making are then performed. Classification uncertainty is calculated based on the fused probability distribution, and entropy is used to measure prediction confidence, with a threshold of 0.3. When the entropy exceeds the threshold, an expert review mechanism is activated to ensure classification reliability. The final classification result is determined according to the maximum probability principle, and the corresponding confidence score is output.

[0095] Finally, the classification results are explained and visualized. The SHAP value is used to explain the classification decision, the contribution of each feature to the classification result is calculated, and a feature importance map is generated. Simultaneously, the t-SNE algorithm is used to reduce the high-dimensional features to two- or three-dimensional space, enabling the visualization of the classification results.

[0096] This embodiment also provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and a method implemented by the processor when executing the computer program.

[0097] This embodiment also provides a storage medium on which a computer program is stored, which implements the method when executed by a processor.

[0098] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A CT imaging-based top coal fracture fractal analysis method, characterized in that, Includes the following steps: Based on CT scan data of top coal samples, a three-dimensional digital model containing fracture information is obtained through three-dimensional reconstruction. The three-dimensional digital model is processed using image processing methods to extract the fracture network and calculate the topological parameters; The fracture network was processed using fractal analysis to calculate the global fractal dimension and local dimension statistical characteristics. The topological parameters, global fractal dimension, and local dimension statistical features are fused to form a multi-scale feature vector; Based on the multi-scale feature vectors, the classification results of top coal fractures are obtained through machine learning model processing. The process of obtaining a three-dimensional digital model containing fracture information includes: Based on CT scan data of top coal samples, a two-dimensional tomographic image sequence was obtained using a high-resolution X-ray scanning system; The two-dimensional tomographic image sequence is subjected to adaptive filtering and noise reduction processing to obtain the noise-reduced image; The denoised image is processed using a multi-threshold segmentation algorithm to obtain the segmented image; The segmented image is reconstructed using a region growing algorithm to generate a preliminary 3D model. Anisotropic diffusion smoothing processing is performed on the preliminary three-dimensional model to obtain an optimized three-dimensional digital model. The process of processing the three-dimensional digital model using image processing methods to extract the fracture network and calculate the topological parameters includes: The three-dimensional digital model is processed using a deep learning-based three-dimensional segmentation network to obtain a binarized crack segmentation result; The segmentation results are subjected to skeleton extraction and midline transformation to obtain a simplified fracture network skeleton. Based on graph theory, the skeleton of the fracture network is abstracted into a topological graph structure, where vertices represent fracture intersections or endpoints, and edges represent fracture branches. Based on the aforementioned topological graph structure, topological parameters including fracture rate, fracture density, average aperture, orientation anisotropy index, and connectivity coefficient are calculated. In the process of processing the three-dimensional digital model using a deep learning-based three-dimensional segmentation network, a slit segmentation neural network based on the 3D U-Net architecture is constructed to perform slit segmentation on the three-dimensional digital model; the slit segmentation neural network is an encoder-decoder structure, the encoder includes several downsampling modules, and the decoder includes several upsampling modules. The process of segmenting the three-dimensional digital model using a 3D U-Net-based slit segmentation neural network includes: The three-dimensional digital model containing the crack information is input into the downsampling module of the encoder to obtain a low-resolution high-dimensional feature map; wherein, several downsampling modules are connected in series, and the three-dimensional digital model undergoes two 3×3×3 convolution operations in sequence after passing through each downsampling module. After each convolution, the ReLU activation function is used for nonlinear transformation, and then a 2×2×2 max pooling operation is performed to obtain a feature map with half the size. The low-resolution high-dimensional feature map is input into the upsampling module of the decoder to obtain the decoder output feature map; wherein, several upsampling modules are connected in series, and the low-resolution high-dimensional feature map is transposed and convolved sequentially after passing through each upsampling module to enlarge the size of the low-resolution high-dimensional feature map by one time to obtain an enlarged feature map, and the enlarged feature map is concatenated with the feature map passed through the skip connection in the encoder path to obtain a concatenated feature map; The decoder output feature map is processed using a 1×1×1 convolutional layer to obtain a probability map with the same size as the input 3D digital model.

2. The CT imaging based top coal fracture fractal analysis method according to claim 1, characterized in that, The process of obtaining the statistical characteristics of the global fractal dimension and local dimension includes: The global fractal dimension of the fracture network was calculated using the three-dimensional box counting method. The three-dimensional model of the fracture network is traversed using the sliding window method, and the local box dimension is calculated at each window position to form a local dimension distribution matrix; Statistical features are calculated based on the local dimension distribution matrix; The global fractal dimension and the statistical feature quantity are combined to form a fractal feature vector.

3. The CT imaging based top coal fracture fractal analysis method according to claim 2, characterized in that, The statistical characteristics include: mean, standard deviation, skewness, and peak value.

4. The CT imaging based top coal fracture fractal analysis method according to claim 2, characterized in that, The process of fusing the topological parameters and the fractal feature vectors to form a multi-scale feature vector includes: The topological parameter feature set and the fractal feature set are respectively standardized and preprocessed to obtain the preprocessed topological parameter feature set and the preprocessed fractal feature set; The preprocessed topological parameter feature set and the preprocessed fractal feature set are concatenated to form an initial feature vector; Based on the initial feature vector, an attention mechanism is used to calculate the weight coefficient of each feature and generate a weighted feature vector; Construct a feature interaction network and extract feature interaction terms to obtain interaction features; The weighted feature vector is concatenated with the interaction feature to form an enhanced feature vector; The enhanced feature vector is then subjected to dimensionality reduction processing to generate the final multi-scale fused feature vector.

5. The fractal analysis method for top coal fractures based on CT imaging according to claim 1, characterized in that, The process of obtaining the top coal fracture classification result based on the multi-scale feature vector and through machine learning model processing includes: The multi-scale feature vectors are input into a pre-trained hybrid classification model for processing. The hybrid classification model includes a parallel gradient boosting decision tree module and a deep neural network module. The outputs of the gradient boosting decision tree module and the deep neural network module are weighted and fused to generate the final probability distribution output. The development level classification result of top coal fractures is determined based on the probability distribution output.

6. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in claim 1.

7. A storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in claim 1.

Images