Intelligent evaluation method for impact propensity of coal based on multi-task gating mechanism

Abstract

The application discloses an intelligent evaluation method for impact tendency of coal based on a multi-task gating mechanism, and belongs to the technical field of cross of electric digital data processing and geological engineering, and the steps are as follows: 1, collecting relevant data and constructing a data set; 2, training an intelligent evaluation model for impact tendency of coal body based on multi-task learning by using the constructed data set, so that the model learns the nonlinear mapping between the microstructure of the coal body and the macroscopic mechanical properties; 3, inputting the particle accumulation and joint structure information of the coal to be detected into the model trained in step 2, and performing regression analysis to obtain the impact tendency index of the coal, and the impact tendency of the coal can be judged according to the standard. The application provides an intelligent solution with low cost and high reliability for mine dynamic disaster prevention and control, and has technical innovation and engineering application value.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of electronic digital data processing and geological engineering, specifically relating to an intelligent assessment method for the impact tendency of coal based on multi-task gating mechanism, applicable to fields such as coal mine safety monitoring, geological engineering research, and mine dynamic disaster prevention and control. Background Technology

[0002] Rockbursts are a serious natural disaster during coal mining, characterized by the rapid destruction of coal and rock masses and the sudden release of elastic energy. They are dynamic phenomena marked by sudden, rapid, and violent destruction. Rockburst disasters not only threaten lives but can also damage mine equipment and disrupt production, resulting in enormous economic losses.

[0003] The shock susceptibility of coal refers to its property of accumulating deformation energy and causing impact damage. Its assessment is a prerequisite for evaluating the impact hazard of coal mine roadways and working faces; for coal mines where the assessment indicates a shock susceptibility, an impact hazard assessment must be conducted on the working faces. Therefore, with the deepening of mining operations, accurately determining the shock susceptibility of coal is of great significance for mine safety and disaster prevention.

[0004] Currently, traditional methods for determining the rockburst tendency of coal are laboratory-based. These methods rely on laboratory indicators, but deep mines, especially newly built ones, often struggle to obtain sufficient coal samples for testing according to these indicators. Furthermore, deep drilling and coring processes can easily damage the coal surface, leading to significant discrepancies between experimental and actual results. This makes accurate testing of the rockburst tendency of deep coal resources difficult. Some newly built mines are also constrained by the risk of rockbursts, making it difficult to accurately obtain information on the rockburst tendency of deep rock masses. This results in inaccurate coal property assessments and a lack of consideration for rockburst hazards during the mine construction phase, hindering the implementation of rockburst mitigation measures during the production phase and increasing the difficulty of disaster management. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent assessment method for the impact tendency of coal based on multi-task gating, which solves the technical problems of unstable measurement results and long measurement cycle in the existing technology.

[0006] The technical solution of this invention is: an intelligent assessment method for the impact tendency of coal based on multi-task gating, the steps of which are as follows:

[0007] Step 1: Use discrete element method (DEM) software to simulate uniaxial compression experiments on coal bodies, record the particle packing structure, joint shape, and simulation results in the coal body modeling, randomly model coal bodies with different particle packing and joint shapes and repeat the experiments to complete the construction of the dataset;

[0008] The specific method is as follows:

[0009] Step 1.1) Coal body mechanical parameters: elastic modulus E = 1130 MPa, Poisson's ratio ν =0.25, tensile strength Tu =4MPa, compressive strength Cu =40 MPa, coefficient of internal friction μi =0.7;

[0010] Step 1.2) Particle packing structure data: A 480 × 480 pixel color image is generated by cropping the cross-section of the 3D coal body model. Then, grayscale processing is applied to discard color information, and 20 slices are extracted at equal intervals along the z-axis. At the same time, the image data is downsampled to 90 × 90 pixels.

[0011] Step 1.3) Joint shape data: Add N internal joints of random location and size to the coal body model. The joint shape uses triangles in three-dimensional space, and the joint information is represented by N 9-dimensional vectors.

[0012] Step 1.4) Impact Tendency Index Acquisition: Run a discrete element uniaxial compression simulation experiment and record the dynamic failure time output by the experiment. Elasticity index Impact Energy Index Uniaxial compressive strength Four shock propensity indicators were used as labels for the dataset;

[0013] Step 1.5) Dataset partitioning: The 1200 repeated DEM simulation data are divided into training set, validation set and test set in a ratio of 7:2:1. The training set is used for training the model in step two, the validation set is used for optimizing the model hyperparameters, and the test set is used to evaluate the generalization performance of the model after training.

[0014] Step 2: Train the coal impact tendency intelligent assessment model based on multi-task learning using the constructed dataset, so that the model learns the nonlinear mapping between the coal microstructure and macroscopic mechanical properties;

[0015] Step 2.1) Multi-module fusion module Fusion Module

[0016] The discrete element method (DEM) model of particle packing structure images is initially fused with joint features. Then, an affine transformation is used to map the intersection of spatial joint surfaces and packing structure cross-sections onto the slice image, achieving spatial alignment between the microstructure and image data. First, the required encoding set is constructed:

[0017]

[0018]

[0019]

[0020] in, Indicates the first A slice in a normalized coordinate system The coordinates of the axes have a range of values. Used to determine the position of the slice in the 3D model. Number the slices. z is the total number of slices; min , z max These are the minimum and maximum values ​​of the ordinate of the joint surface, respectively. For joint The highest point of the axis, and the other two points of the joint are and , , , They are respectively the corresponding vertices coordinate; The linear interpolation parameters are calculated based on the proportional relationship between the intersection points of the slice plane and the joint edges. For the slice plane and joint edges The coordinates of the intersection point; For the slice plane and joint edges The coordinates of the intersection point; finally, the line segment is obtained. This is the projection encoding of the joint onto the current slice, denoted as... ,in Number the joints. Number the slices;

[0021] Next, image coordinate transformation is performed, utilizing image dimensions. To convert the scale coordinates into pixel coordinates, bilinear interpolation is then used on the pixel set. A continuous intensity distribution is generated around it;

[0022]

[0023]

[0024]

[0025] In formula (4), These are the pixel coordinates of the image; This represents the total number of joint projection points in the current slice image, i.e., the number of coded points generated by the intersection of all joints with this slice; For the first Image coordinates of each joint projection point; The standard deviation of the Gaussian kernel is used to control the width of the generated lines; This indicates that the pixel is formed by the combined effect of all joint projection points. The Gaussian response intensity generated at the point is superimposed to form a line image. ;

[0026] In formula (5), For the original sliced ​​image in pixels The grayscale value at that location; The line image calculated by formula (4) in pixels The value at; This is the mixing coefficient, used to balance the contributions of the original image and the line image; For the blended image at pixels Output value at;

[0027] In formula (6), Indicated by Centered on, with standard deviation of A single-channel image generated by the Gaussian kernel function; The symbol represents all The generated Gaussian images are collected and merged to form an image stack; This means that for each pixel position in the merged image stack, the maximum value among all images is taken, thereby generating a single-channel maximum response map; This represents a pixel-by-pixel multiplication operation; The original sliced ​​image; the final result The encoded image, incorporating joint information, is used as input for subsequent modules;

[0028] Step 2.2) Parallel processing by the expert module

[0029] Image processing expert Expert0 / 1: It uses two types of convolution kernels, 3×3 and 5×5, to capture local texture and regional structure patterns in parallel, combines global average pooling and fully connected layers to generate channel attention weights, and integrates complementary information from multi-scale feature maps;

[0030] Joint Processing Expert2: Introduces the JNR-GCN module to encode the node and edge features of joints, and combines it with the relation embedding method to realize graph structure topological relationship modeling;

[0031] Step 2.3) Dynamic Gating Mechanism (Gate Module)

[0032] Expert module weights are dynamically allocated using Swish activation functions and batch normalization techniques to achieve multi-task routing. The gating weights are normalized to dynamically balance the utilization of expert knowledge by each task. The formula is as follows:

[0033]

[0034]

[0035] in: This is the concatenation result of the output features of each expert module in the upper layer of the gating module, and the intermediate representation after processing by a fully connected layer and the Swish activation function; and These represent the feature vectors calculated in the current training batch. The mean and variance; It is a very small positive number used to prevent the denominator from being zero and to ensure numerical stability; and These are trainable scaling and translation parameters used to restore the network's expressive power, enabling the normalized features to adaptively adjust their distribution. The feature vectors after batch normalization are then input into the Softmax function to be converted into gating weights corresponding to each task. These weights are used to dynamically weight and fuse the outputs of multiple expert modules to optimize multi-task routing.

[0036] Step 3: Input the particle packing and joint structure information of the coal to be tested into the model trained in Step 2, perform regression analysis to obtain the coal's impact tendency index, and judge the coal's impact tendency according to the standard.

[0037] The specific method is as follows:

[0038] Step 3.1) Data Acquisition and Preprocessing

[0039] The microstructure information of the coal body to be tested can be obtained through the following two methods:

[0040] Approach 1: Numerical Simulation. Using discrete element method (DEM) software, a three-dimensional DEM model of the coal body under study is constructed according to the parameters described in step 1.1. Twenty cross-sectional images are extracted at equal intervals along the z-axis according to the method in step 1.2. After grayscale conversion and downsampling, a 90×90×20 particle packing structure image matrix is ​​generated. At the same time, all internal joint surfaces are extracted from the model. Each joint surface is fitted with a three-dimensional space triangle, and its 9-dimensional vector is recorded to form joint shape data.

[0041] Method 2: CT Scanning of Real Coal Samples: In underground coal mines or laboratories, coal samples are scanned using industrial CT scanning equipment to obtain three-dimensional CT image data. The CT images are preprocessed, including grayscale conversion and noise reduction, and 20 cross-sectional images are extracted at equal intervals along the axial direction and downsampled to 90×90 pixels to obtain the particle packing structure image. At the same time, internal joint surfaces are identified and extracted from the CT images using image segmentation and three-dimensional reconstruction algorithms. Each joint surface is fitted as a three-dimensional spatial triangle, and its 9-dimensional vector is recorded to obtain joint shape data.

[0042] Step 3.2) Model Inference

[0043] The preprocessed particle packing structure image and joint shape data are input into the impact tendency assessment model based on a multi-task gating mechanism trained in step two. The model performs forward calculations and outputs predicted values ​​for four impact tendency indicators: dynamic failure time. Elasticity index Impact Energy Index Uniaxial compressive strength ;

[0044] Step 3.3) Impact Tendency Determination and Early Warning

[0045] Based on the classification thresholds specified in standard GB / T 25217.2-2010, the predicted values ​​of the four indicators output by the model are compared with the thresholds to determine the tendency category corresponding to each indicator. Combining the categories of the four indicators, the final impact tendency level of the coal body is determined according to the principle of choosing the highest value or through a weighted comprehensive scoring method. When the final determination result is a strong impact tendency, the system triggers an early warning signal, prompting the implementation of corresponding anti-impact measures.

[0046] The beneficial effects of this invention are as follows:

[0047] This invention achieves a dual breakthrough in the accuracy and robustness of coal seam impact tendency discrimination through the deep integration of multi-task gating mechanisms and three-dimensional discrete element modeling: Expert module collaborative optimization based on dynamic weight allocation improves the accuracy of four core indicators; precise modeling of joint topological relationships through the JNR-GCN module enhances the model's tolerance to image noise and joint displacement, thus strengthening robustness; the end-to-end training strategy integrates discrete element simulation and deep learning frameworks, reducing single-sample inference time to 0.3 seconds while lowering physical experiment costs by 80%; this invention realizes cross-scale correlation modeling of microstructure and macromechanics, providing a low-cost, highly reliable intelligent solution for mine dynamic disaster prevention and control, demonstrating technological innovation and engineering application value. Attached Figure Description

[0048] Figure 1This is a flowchart of the dataset construction process in this invention.

[0049] Figure 2 This is a topology diagram of the coal impact tendency determination algorithm based on multi-task learning in this invention.

[0050] Figure 3 This is a diagram of a multi-scale feature extraction expert network for image processing experts in this invention.

[0051] Figure 4 This is a schematic diagram of the JNR-GCN module in this invention. Detailed Implementation

[0052] Example 1: An intelligent assessment method for the shock tendency of coal based on multi-task gating

[0053] Step 1. Data preprocessing and discrete element modeling, such as... Figure 1 As shown.

[0054] Step 1.1 Construction of Discrete Element Model of Coal Body A three-dimensional discrete element model of the coal body was established using discrete element software. The mechanical properties of the coal body were set as follows: elastic modulus E = 1130 MPa; Poisson's ratio ν = 0.25; tensile strength Tu = 4 MPa; compressive strength Cu = 40 MPa; internal friction coefficient μi = 0.7. N = 15 internal joints were randomly added to the model. The joints were three-dimensional spatial triangles, and their geometric parameters were encoded using 9-dimensional vectors.

[0055] Step 1.2 Generation of sliced ​​images

[0056] Twenty 2D slices of 480×480 pixels were extracted at equal intervals along the z-axis. Each slice was then processed into grayscale and downsampled to 90×90 pixels to form a 3D data matrix.

[0057] Step 2. Model Architecture and Module Implementation. The overall model architecture is as follows: Figure 2 As shown.

[0058] Step 2.1 Fusion Module Implementation

[0059] The particle packing structure image modeled by discrete element method is spatially aligned with the joint features. The intersection of the joint surface and the cross surface of the packing structure is mapped onto the slice image through affine transformation to generate a fused feature map.

[0060] Step 2.2 Expert Modules Design

[0061] Expert0 / 1 (Image Processing Expert) employs 3×3 and 5×5 convolutional kernels and depthwise separable convolutions in parallel to extract local texture and region structure features. Finally, it fuses complementary information from multi-scale feature maps, such as... Figure 3 As shown.

[0062] Expert2 (Joint Treatment Expert) constructs the JNR-GCN module. Node features encode the self-information of joints (area, centroid coordinates, dip angle), and edge features encode the spatial relationships between joints (intersection, minimum distance, difference in face dip angle). The GraphConstructor module is used to build a graph structure based on the joint interaction strength, preserving the relationships between adjacent joints, such as... Figure 4 As shown.

[0063] Step 2.3 Dynamic weight allocation of Gate Module

[0064] The gated network uses the Swish activation function and Softmax normalization, and the weights are initialized using a Xavier uniform distribution. The gate weight formula is as follows:

[0065]

[0066] in W For learnable parameter matrix, e i Features of the expert module b This is a bias term.

[0067] Step 3. Model Training and Hyperparameter Configuration

[0068] Step 3.1 Training Strategy

[0069] Optimizer: Adam optimizer, initial learning rate η = 0.001, learning rate decays by 5% every 10 training epochs. Loss function: Multi-task weighted MSE, total loss formula:

[0070]

[0071] in W k The task weights (dynamic failure time, elastic energy index, impact energy index, and uniaxial compressive strength are set to 0.3, 0.25, 0.25, and 0.2, respectively) are assigned. For the first The true value of each indicator For the model to the first Predicted values ​​for each indicator. Batch size: 32, training epochs: 200.

[0072] Step 3.2 Hardware Environment

[0073] GPU: NVIDIA Tesla V100 (32GB VRAM)

[0074] Software framework: PyTorch 1.7.0 + CUDA 11.0

[0075] Step 4. Actual Deployment and Application Process

[0076] Step 4.1 Downhole Data Interface Design

[0077] Input data: 3D scan data of coal or discrete element simulation data, in the format of a 90×90×20 3D matrix.

[0078] Preprocessing steps: grayscale processing, downsampling to 90×90 pixels, generating a 9-dimensional joint vector.

[0079] Step 4.2 Warning Trigger Logic

[0080] The model outputs four types of impact propensity indicators, which are compared with the standard:

[0081] Table 1: Classification of Coal Shock Tendency

[0082]

[0083] An early warning is triggered when the overall indicator shows a strong tendency to cause shocks.

[0084] Step 4.3 Implementation Effect Verification

[0085] After being deployed in a coal mine, the accuracy rate of the judgment of 127 sets of coal sample data reached 97.64%, the average early warning response time was 0.3 seconds, and it successfully issued early warnings for 3 potential rockburst events.

Claims

1. A method for intelligent assessment of the impact tendency of coal based on multi-task gating mechanism, characterized in that, The steps are as follows: Step 1: Use discrete element method (DEM) software to simulate uniaxial compression experiments on coal bodies, record the particle packing structure, joint shape, and simulation results in the coal body modeling, randomly model coal bodies with different particle packing and joint shapes and repeat the experiments to complete the construction of the dataset; Step 2: Train the coal impact tendency intelligent assessment model based on multi-task learning using the constructed dataset, so that the model learns the nonlinear mapping between the coal microstructure and macroscopic mechanical properties; Step 2.1) Multi-module fusion module Fusion Module The discrete element method (DEM) model of particle packing structure images is initially fused with joint features. Then, an affine transformation is used to map the intersection of spatial joint surfaces and packing structure cross-sections onto the slice image, achieving spatial alignment between the microstructure and image data. First, the required encoding set is constructed: in, Indicates the first A slice in a normalized coordinate system The coordinates of the axes have a range of values. Used to determine the position of the slice in the 3D model. Number the slices. z is the total number of slices; min , z max These are the minimum and maximum values ​​of the ordinate of the joint surface, respectively. For joint The highest point of the axis, and the other two points of the joint are and , , , They are the corresponding vertices coordinate; The linear interpolation parameters are calculated based on the proportional relationship between the intersection points of the slice plane and the joint edges. For the slice plane and joint edges The coordinates of the intersection point; For the slice plane and joint edges The coordinates of the intersection point; finally, the line segment is obtained. This is the projection encoding of the joint onto the current slice, denoted as... ,in Number the joints. Number the slices; Next, image coordinate transformation is performed, utilizing image dimensions. To convert the scale coordinates into pixel coordinates, bilinear interpolation is then used on the pixel set. A continuous intensity distribution is generated around it; In formula (4), These are the pixel coordinates of the image; This represents the total number of joint projection points in the current slice image, i.e., the number of coded points generated by the intersection of all joints with this slice; For the first Image coordinates of each joint projection point; The standard deviation of the Gaussian kernel is used to control the width of the generated lines; This indicates that the pixel is formed by the combined effect of all joint projection points. The Gaussian response intensity generated at the point is superimposed to form a line image. ; In formula (5), For the original sliced ​​image in pixels The grayscale value at that location; The line image calculated by formula (4) in pixels The value at; This is the mixing coefficient, used to balance the contributions of the original image and the line image; For the blended image at pixels Output value at; In formula (6), Indicates Centered on, with standard deviation of A single-channel image generated by the Gaussian kernel function; The symbol represents all The generated Gaussian images are collected and merged to form an image stack; This means that for each pixel position in the merged image stack, the maximum value among all images is taken, thereby generating a single-channel maximum response map; This represents a pixel-by-pixel multiplication operation; The original sliced ​​image; the final result The encoded image, incorporating joint information, is used as input for subsequent modules; Step 2.2) Parallel processing by the expert module Image processing expert Expert0 / 1: It uses two types of convolution kernels, 3×3 and 5×5, to capture local texture and regional structure patterns in parallel, combines global average pooling and fully connected layers to generate channel attention weights, and integrates complementary information from multi-scale feature maps; Joint Processing Expert2: Introduces the JNR-GCN module to encode the node and edge features of joints, and combines it with the relation embedding method to realize graph structure topological relationship modeling; Step 2.3) Dynamic Gating Mechanism (Gate Module) Expert module weights are dynamically allocated using Swish activation functions and batch normalization techniques to achieve multi-task routing. The gating weights are normalized to dynamically balance the utilization of expert knowledge by each task. The formula is as follows: in: This is the concatenation result of the output features of each expert module in the upper layer of the gating module, and the intermediate representation after processing by a fully connected layer and the Swish activation function; and These represent the feature vectors calculated in the current training batch. The mean and variance; It is a very small positive number used to prevent the denominator from being zero and to ensure numerical stability; and These are trainable scaling and translation parameters used to restore the network's expressive power, enabling the normalized features to adaptively adjust their distribution. The feature vectors after batch normalization are then input into the Softmax function to be converted into gating weights corresponding to each task. These weights are used to dynamically weight and fuse the outputs of multiple expert modules to optimize multi-task routing. Step 3: Input the particle packing and joint structure information of the coal to be tested into the model trained in Step 2, perform regression analysis to obtain the coal impact tendency index, and judge the coal impact tendency according to the standard.

2. The intelligent assessment method for the shock tendency of coal based on multi-task gating as described in claim 1, characterized in that, The specific method in step one is as follows: Step 1.1) Coal body mechanical parameters: elastic modulus E = 1130 MPa, Poisson's ratio ν =0.25, tensile strength Tu =4 MPa, compressive strength Cu =40 MPa, coefficient of internal friction μi =0.7; Step 1.2) Particle packing structure data: A 480 × 480 pixel color image is generated by cropping the cross-section of the 3D coal body model. Then, grayscale processing is applied to discard color information, and 20 slices are extracted at equal intervals along the z-axis. At the same time, the image data is downsampled to 90 × 90 pixels. Step 1.3) Joint shape data: Add N internal joints of random location and size to the coal body model. The joint shape uses triangles in three-dimensional space, and the joint information is represented by N 9-dimensional vectors. Step 1.4) Impact Tendency Index Acquisition: Run a discrete element uniaxial compression simulation experiment and record the dynamic failure time output by the experiment. Elasticity index Impact Energy Index Uniaxial compressive strength Four shock propensity indicators were used as labels for the dataset; Step 1.5) Dataset partitioning: The 1200 repeated DEM simulation data are divided into training set, validation set and test set in a ratio of 7:2:

1. The training set is used for training the model in step two, the validation set is used for optimizing the model hyperparameters, and the test set is used to evaluate the generalization performance of the model after training.

3. The intelligent assessment method for the shock tendency of coal based on multi-task gating as described in claim 2, characterized in that, The specific method in step three is as follows: Step 3.1) Data Acquisition and Preprocessing The microstructure information of the coal body to be tested can be obtained through the following two methods: Approach 1: Numerical Simulation. Using discrete element method (DEM) software, a three-dimensional DEM model of the coal body under study is constructed according to the parameters described in step 1.1). Following the method in step 1.2), 20 cross-sectional images are extracted at equal intervals along the z-axis. After grayscale conversion and downsampling, a 90×90×20 particle packing structure image matrix is ​​generated. Simultaneously, all internal joint surfaces are extracted from the model. Each joint surface is fitted with a three-dimensional spatial triangle, and its 9-dimensional vector is recorded to form joint shape data. Method 2: CT Scanning of Real Coal Samples: In underground coal mines or laboratories, coal samples are scanned using industrial CT scanning equipment to obtain three-dimensional CT image data. The CT images are preprocessed, including grayscale conversion and noise reduction, and 20 cross-sectional images are extracted at equal intervals along the axial direction and downsampled to 90×90 pixels to obtain the particle packing structure image. At the same time, internal joint surfaces are identified and extracted from the CT images using image segmentation and three-dimensional reconstruction algorithms. Each joint surface is fitted as a three-dimensional spatial triangle, and its 9-dimensional vector is recorded to obtain joint shape data. Step 3.2) Model Inference The preprocessed particle packing structure image and joint shape data are input into the impact tendency assessment model based on a multi-task gating mechanism trained in step two. The model performs forward calculations and outputs predicted values ​​for four impact tendency indicators: dynamic failure time. Elasticity index Impact Energy Index Uniaxial compressive strength ; Step 3.3) Impact Tendency Determination and Early Warning Based on the classification thresholds specified in standard GB / T 25217.2-2010, the predicted values ​​of the four indicators output by the model are compared with the thresholds to determine the tendency category corresponding to each indicator. Combining the categories of the four indicators, the final impact tendency level of the coal body is determined according to the principle of choosing the higher value or by using a weighted comprehensive scoring method. When the final judgment result is a strong impact tendency, the system triggers an early warning signal to prompt the implementation of corresponding anti-impact measures.

Images