A blasting explosive quantity fine design method based on rock mineral particle structure

Abstract

This invention provides a refined design method for explosive charge quantity based on rock mineral particle structure, relating to the field of geotechnical engineering blasting technology. This invention maps the mineral composition and particle size data of rock samples to mineral hardness distribution characteristics and particle structure characteristics, respectively, and generates a high-dimensional feature vector through nonlinear encoding fusion. This vector is input into a graph neural network, and a rock structure feature relationship map is formed through node representation and attention mechanisms. Using this map as a condition, under the constraint of baseline explosive consumption, a deep generative network learns and outputs a differentiated charge distribution matrix corresponding to spatial location. After alignment and transformation with blasting hole mesh parameters, a charge allocation design map is generated. Finally, after blasting based on this design map, actual effect indicators are collected to construct a multi-objective optimization function. An optimization algorithm is used to adjust the parameters of the graph attention mechanism and the generative network in reverse according to the effect differences, achieving iterative optimization of the model.

Description

Technical Field

[0001] This invention relates to the field of blasting technology in geotechnical engineering, specifically a method for refined design of explosive charge quantity based on the particle structure of rock minerals. Background Technology

[0002] In the field of geotechnical blasting, achieving precise allocation and efficient utilization of explosive energy is a core challenge for improving fragmentation effects, controlling blasting hazards, and reducing construction costs. Rock, as the object of blasting, fundamentally determines its resistance to blasting dynamics through its internal mineral composition, hardness, and the fineness and distribution of its particle structure. Traditionally, the design of explosive charge quantities for complex and variable rock types often relies on macroscopic methods based on field tests and empirical formulas. Specifically, engineers typically determine a preliminary explosive consumption based on macroscopic indicators such as the rock's strength coefficient, wave impedance, and uniaxial compressive strength, or by directly referring to empirical tables in industry regulations. To improve the design's specificity, some improvement methods correct empirical values ​​through small-scale field blasting funnel tests or attempt to introduce a blastability index calculated from several macroscopic parameters to adjust the charge quantity. Furthermore, with the development of detection technology, some studies have begun to utilize image recognition technology to observe macroscopic fracture networks in rock masses in order to optimize charge structure. These existing technologies collectively constitute the mainstream practical basis for current blasting parameter design.

[0003] However, the aforementioned existing technologies still have inherent limitations. First, their design logic is essentially macroscopically homogeneous, assuming that rock masses with similar macroscopic mechanical properties have similar crushing energy consumption. This ignores the essential influence of microstructure, such as differences in mineral hardness and particle interface effects, on the energy consumption for crack initiation and propagation. This leads to insufficient estimation of explosive dosage for difficult-to-blast rock masses with high quartz content and fine, dense particles, while potentially overdosing on relatively easy-to-blast rock masses. Second, existing methods fail to establish a direct, nonlinear mapping model between quantitatively obtained microscopic characteristics such as mineral composition ratios and particle size distribution and the final spatial explosive dosage distribution. Empirical formulas or simple exponential corrections cannot capture this complex relationship, making it difficult for the design results to truly achieve lithological adaptation. Finally, existing schemes generally lack a closed-loop learning and evolution mechanism. The effect data after a single blast is not systematically used to invert and calibrate the correlation model between microstructural parameters and explosive dosage. The design model is static and cannot utilize feedback data for self-iteration and optimization in continuous engineering practice, thus making it difficult to adapt to complex and changing geological conditions.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method for fine-tuning the design of explosive charge based on the particle structure of rock minerals, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for refined design of explosive charge quantity based on rock mineral particle structure, comprising the following steps: Step 1: Map the mineral composition mass percentage data of the rock sample in the area to be blasted to the mineral hardness distribution characteristics, and map its particle size distribution data to the particle structure characteristics. Perform nonlinear encoding and combination on the two features to generate a high-dimensional feature vector characterizing the microstructure of the rock. Step 2: Input the high-dimensional feature vector into the graph neural network model, and use the two feature vectors contained therein, which respectively represent the mineral hardness and grain structure, as nodes. Calculate the attention weights between nodes through the graph attention mechanism as the weights of the edges to construct a structural feature relationship map of the rock. Step 3: Using the structural feature relationship map as input and the benchmark explosive consumption as a constraint, the trained deep generative network learns the nonlinear mapping relationship from structural features to explosive distribution and outputs the explosive distribution matrix corresponding to the spatial location of the area to be detonated. The explosive distribution matrix is ​​then spatially aligned and scaled with the preset blasting hole network spatial layout parameters to generate the explosive distribution design map. Step 4: After blasting based on the charge distribution design, collect multi-dimensional evaluation indicators of the actual blasting effect to construct a multi-objective optimization function. Using an optimization algorithm, adjust the parameters in the graph attention mechanism and deep generative network in reverse according to the difference between the actual blasting effect and the expected target to achieve iterative optimization of the charge amount.

[0007] Furthermore, X-ray diffraction analysis was performed on the rock sample to obtain the mass percentage data of each mineral in the rock sample, and digital image processing analysis was performed on the scanning electron microscope image of the rock sample to obtain the particle size distribution data inside it. Mapping mineral composition mass percentage data to mineral hardness distribution characteristics involves the following steps: extracting the Mohs hardness scale value for each mineral and multiplying it by the corresponding content proportion of that mineral in the mass percentage data to obtain the hardness contribution value of each mineral; then arranging the hardness contribution values ​​of all minerals in the rock sample in descending order of their Mohs hardness values ​​and normalizing them to form a vector form of mineral hardness distribution characteristics; mapping particle size distribution data to particle structure characteristics involves statistical analysis of the particle size distribution data, extracting the average particle size and the standard deviation of the particle size distribution as key statistics, and constructing a vector from these key statistics to form particle structure characteristics. The mineral hardness distribution features are input into the first encoder to obtain the first encoded feature vector; the particle structure features are input into the second encoder to obtain the second encoded feature vector; wherein the first encoder and the second encoder each contain at least one fully connected layer and a nonlinear activation function; the first encoded feature vector and the second encoded feature vector are concatenated to generate a high-dimensional feature vector, wherein the high-dimensional feature vector is formed by sequentially concatenating the first encoded feature vector and the second encoded feature vector.

[0008] Furthermore, the high-dimensional feature vector is input into the graph neural network model. The specific process is as follows: extract the part corresponding to the first encoded feature vector from the high-dimensional feature vector as the initial feature vector of the first node, and extract the part corresponding to the second encoded feature vector as the initial feature vector of the second node; construct a complete graph structure containing two nodes, where each node is connected to another node and itself through an edge; calculate the attention coefficient between each node and all its neighboring nodes through a graph attention mechanism, where the neighboring nodes include the node itself and another node; normalize the attention coefficients using the softmax function to obtain normalized attention weights; for each node, sum the initial feature vectors of all its neighboring nodes based on the corresponding normalized attention weights to generate the updated feature vector of that node. The updated first node feature vector, the updated second node feature vector, and the normalized attention weights between adjacent nodes are used as edge weights, and the output is a structural feature relationship map of the rock.

[0009] Furthermore, the structural feature relationship map is used as input. The specific process is as follows: the trained deep generative network is pre-trained on historical data, which includes multiple sets of known structural feature relationship maps and corresponding charge distribution matrices determined by experts; during the training process, the benchmark explosive consumption is input into the deep generative network as a numerical constraint, ensuring that the total explosive consumption corresponding to the output charge distribution matrix is ​​not greater than the upper limit of the total charge calculated based on the benchmark explosive consumption; the structural feature relationship map of the area to be detonated is input into the trained deep generative network, which directly outputs the charge distribution matrix corresponding to the spatial location of the area to be detonated; The deep generative network outputs a charge distribution matrix corresponding to the spatial location of the area to be detonated. Each element in the charge distribution matrix represents the suggested charge consumption at a spatial location point within the area to be detonated.

[0010] Furthermore, the charge distribution matrix is ​​spatially aligned and scaled with the preset blasting hole network spatial layout parameters to generate a charge distribution design drawing. The specific process is as follows: based on the hole spacing, row spacing and step height in the preset blasting hole network spatial layout parameters, the design load volume of each blast hole is calculated. The design load volume is the product of the hole spacing, row spacing and step height. For each borehole, based on its center coordinates and charge section length, a corresponding influence area is defined in the charge distribution matrix. The average value of all elements in the charge distribution matrix within the influence area corresponding to each borehole is calculated. This average value is used as the suggested average explosive consumption per borehole. The product of the suggested average explosive consumption per borehole and the design load volume of the corresponding borehole is calculated as the charge per borehole. The charge per borehole of all boreholes is summarized to form the charge distribution design diagram.

[0011] Furthermore, multi-dimensional evaluation index data of actual blasting effects are collected to construct a multi-objective optimization function. The specific process is as follows: After blasting is carried out according to the charge distribution design, a drone is used to take aerial photos of the blast pile from multiple angles. A digital surface model of the blast pile is generated through three-dimensional reconstruction technology. The digital surface model is segmented and identified into rock blocks. The equivalent particle size of each rock block is counted. Based on the preset threshold particle size for large blocks and the threshold particle size for fine ore, the ratio of large blocks and the ratio of fine ore are calculated. The ratio of large blocks is the percentage of the total volume of rock blocks with an equivalent particle size greater than the threshold particle size to the total volume of the blast pile. The ratio of fine ore is the percentage of the total volume of rock blocks with an equivalent particle size less than the threshold particle size to the total volume of the blast pile. The optimization objectives are to minimize the bulk ore ratio and make the fine ore ratio fall within the preset target range. A multi-objective optimization function is constructed based on the optimization objectives.

[0012] Furthermore, an optimization algorithm is employed to iteratively optimize based on the difference between the actual blasting effect and the expected target. The specific process is as follows: the actual block rate and fine ore rate data collected after the actual blasting are substituted into the multi-objective optimization function to calculate the actual loss value that quantifies the deviation between the actual blasting effect and the expected target; using the backpropagation algorithm, starting from the actual loss value, the gradients of the actual loss value with respect to the weight parameters and bias parameters of each layer of the deep generative network are calculated sequentially, and the gradients of the actual loss value with respect to all training parameters in the graph neural network model are also calculated. The training parameters include the learned weight matrix that performs a linear transformation on the node feature vector in the graph attention mechanism, and the learned attention vector used to calculate the attention coefficients between nodes. The gradient descent algorithm is used to adjust the weights and bias parameters of the deep generative network and the training parameters in the graph neural network model in the opposite direction of the gradient, based on the calculated gradient. The updated graph neural network model and the deep generative network are then used for the next blast charge design. This process is repeated until the multi-objective optimization function converges or the block ratio and fine ore ratio of the actual blasting effect reach the expected target, thus realizing the iterative optimization of the blast charge design method.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention effectively overcomes the shortcomings of existing technologies, such as reliance on macroscopic homogenization assumptions, insufficient model mapping capabilities, and lack of adaptive learning mechanisms, by introducing quantitative characterization and intelligent mapping of rock microstructure features. This enables the scientific, targeted, and dynamic optimization of explosive charge design. Specifically, by analyzing the mineral composition mass percentage and particle size distribution data, mineral hardness distribution features and particle structure features are mapped to them respectively. A high-dimensional feature vector is generated through an innovative nonlinear coding and fusion process, realizing a microscale quantitative description of the intrinsic nature of rock resistance to blasting. This fundamentally changes the limitation of traditional methods that ignore the influence of microstructure, laying a data foundation for accurate energy demand assessment. This invention also constructs a graph neural network model with hardness and structural features as nodes, and uses a graph attention mechanism to capture their intrinsic relationships, forming a feature relationship map that can profoundly express the complex microstructure of rocks. Then, using this map as input, a deep generative network is trained to learn the nonlinear mapping from structural features to spatial charge distribution under the constraint of benchmark explosive consumption. This technical approach successfully establishes a direct and intelligent mapping model between precisely measured micro-features and differentiated charge design schemes, solving the problem that existing empirical formulas or simple exponential corrections cannot characterize this complex relationship, realizing true lithology-adaptive design, and significantly improving the rationality of charge allocation. This invention transforms the actual block size effect after blasting into a quantitative evaluation index and constructs a loss function. It can automatically inversely calculate the deviation between the actual effect and the design expectation, and adjust the internal parameters of the graph attention network and the deep generation network accordingly. This mechanism makes the entire design model no longer a one-time static calculation, but can use the feedback data of each blast for self-iteration and optimization in continuous engineering applications. This allows it to dynamically adapt to changes in the microstructure of rocks in different mining areas or even within the same blasting area, and eventually gradually converge to a better design state. This fundamentally solves the shortcomings of traditional methods that cannot continuously learn and evolve in practice, forming an intelligent blasting design with adaptive capabilities. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a bar chart showing the actual bulk ore ratio, fine ore ratio, and actual loss value of this invention. Figure 3 This is a scatter plot of the observed loss value versus the actual block size ratio in this invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0016] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0017] Example: Please see Figures 1-3 The present invention provides a technical solution: A method for refined design of explosive charge quantity based on rock mineral particle structure, comprising the following steps: Step 1: Map the mineral composition mass percentage data of the rock sample in the area to be blasted to the mineral hardness distribution characteristics, and map its particle size distribution data to the particle structure characteristics. Perform nonlinear encoding and combination on the two features to generate a high-dimensional feature vector characterizing the microstructure of the rock. X-ray diffraction analysis was performed on the rock sample to obtain the mass percentage data of each mineral in the rock sample. Digital image processing analysis was performed on the scanning electron microscope image of the rock sample to obtain the particle size distribution data of its interior. The reason for obtaining mineral composition mass percentage data through X-ray diffraction analysis is that different minerals have drastically different Mohs hardness, and the overall resistance of a rock to breakage is a comprehensive reflection of the hardness of its constituent minerals. X-ray diffraction analysis can accurately identify the types of minerals in a rock sample based on the diffraction pattern of the crystal structure, and calculate the relative mass percentage of each mineral based on the intensity of the diffraction peaks. This data is the only objective input for subsequent calculation of the overall hardness characteristics of the rock. The particle size distribution data is obtained by performing digital image processing on scanning electron microscope (SEM) images because the size, shape, and distribution of rock particles directly determine the path and difficulty of crack initiation and propagation. SEM can clearly present the micron to nanometer-scale particle morphology of rock fracture surfaces or polished surfaces. By performing digital image processing on the obtained SEM images, including grayscale conversion, median filtering for noise reduction, and Otsu thresholding, particles can be accurately distinguished from the background or cement. Subsequently, connected regions in the binary image are marked and measured, and the equivalent circle diameter of each particle is counted. Finally, a statistical set containing the particle sizes of all detected particles is obtained, i.e., particle size distribution data. Mapping mineral composition mass percentage data to mineral hardness distribution characteristics involves the following steps: extracting the Mohs hardness scale value for each mineral and multiplying it by the corresponding content proportion of that mineral in the mass percentage data to obtain the hardness contribution value of each mineral; then arranging the hardness contribution values ​​of all minerals in the rock sample in descending order of their Mohs hardness values ​​and normalizing them to form a vector form of mineral hardness distribution characteristics; mapping particle size distribution data to particle structure characteristics involves statistical analysis of the particle size distribution data, extracting the average particle size and the standard deviation of the particle size distribution as key statistics, and constructing a vector from these key statistics to form particle structure characteristics. The reason for mapping mineral composition mass percentage data to form mineral hardness distribution characteristics is that different minerals have different Mohs hardness, and their content ratio in rocks directly determines the overall compressive strength and fracture resistance of the rock. The core idea of ​​the mapping process in this method is to transform the abstract mineral content into a numerical distribution that can comprehensively reflect the overall hardness characteristics of the rock. It should be noted that mapping mineral composition mass percentage data to mineral hardness distribution characteristics is essentially transforming discrete mineral types and contents into a continuous, ordered, and calculable hardness weight spectrum. The specific process is as follows: first, based on mineralogy knowledge, each mineral appearing in the rock is assigned a standard Mohs hardness scale value; then... The Mohs hardness value of each mineral is multiplied by its mass percentage in the rock to obtain the mineral's contribution to the overall rock hardness. Then, all mineral hardness contributions are arranged in descending order of their corresponding Mohs hardness values. This sorting process gives the generated sequence a clear physical meaning, where the sequence position represents the hardness level. Finally, this ordered sequence is normalized, for example, using min-max normalization to scale each value in the sequence to between 0 and 1. The sequence obtained after these steps is defined as a mineral hardness distribution feature, which is a numerical vector with a dimension equal to the number of mineral types. The order of the elements in the vector represents the hardness level, and the value of the element represents the overall weight of that hardness level in the rock. It should be noted that mapping particle size distribution data to particle structure features is based on the core idea of ​​using a few statistical quantities with clear physical meaning to summarize the complex overall picture of particle size distribution. This invention selects the average particle size and the standard deviation of particle size distribution as key statistical quantities. The average particle size is calculated by summing the equivalent diameters of all particles in the particle size distribution data and dividing by the total number of particles. It represents the average coarseness of the rock particle system. The standard deviation of particle size distribution is calculated by first calculating the square of the difference between the particle size of each particle and the average particle size, summing the differences, dividing by the total number of particles, and then taking the square root. It characterizes the degree of dispersion of particle size around the average value, i.e., sorting. Combining the calculated average particle size and the standard deviation of particle size distribution in sequence forms a two-dimensional particle structure feature vector. This feature reflects the macroscopic structural properties of the particle system in the simplest mathematical form. The mineral hardness distribution features are input into the first encoder to obtain the first encoded feature vector; the particle structure features are input into the second encoder to obtain the second encoded feature vector; wherein the first encoder and the second encoder each contain at least one fully connected layer and a nonlinear activation function; the first encoded feature vector and the second encoded feature vector are concatenated to generate a high-dimensional feature vector, wherein the high-dimensional feature vector is formed by sequentially concatenating the first encoded feature vector and the second encoded feature vector; It should be noted that the purpose of nonlinearly encoding and combining mineral hardness distribution features and particle structure features is to map features from different physical sources, different dimensions, and different scales to a unified, higher-dimensional, linearly separable feature space through a neural network. This facilitates the subsequent model's ability to explore complex nonlinear relationships. Specifically, this involves two parallel encoders. The first encoder is dedicated to processing mineral hardness distribution features. Its structure includes a fully connected layer and a nonlinear activation function. The fully connected layer has multiple neurons, and its function is to project the input hardness distribution feature vector to a space of a predetermined dimension through a linear transformation. The result of the linear transformation is then input to a nonlinear activation function, such as a modified linear unit function. This function introduces nonlinear transformation capabilities into the model by setting negative values ​​to zero and retaining positive values, enabling it to learn more complex patterns. The output of the first encoder is called the first encoded feature vector. The second encoder independently processes particle structure features in exactly the same way and outputs the second encoded feature vector. The number of neurons in the two encoders can be the same or different, but their function is to perform deep nonlinear encoding of their respective input features. The output dimensions of the first encoder and the second encoder can be set to the same value, such as 128. This means that the hardness and structural features are mapped to a 128-dimensional high-dimensional space. The specific type of nonlinear activation function is the modified linear unit function, i.e., the ReLU function, which increases the nonlinear expressive power of the network. Finally, the first encoded feature vector and the second encoded feature vector are concatenated to generate the final high-dimensional feature vector. The specific method of the concatenation operation is to arrange all the elements of the second encoded feature vector in sequence after all the elements of the first encoded feature vector to form a new, longer single vector. This new vector is the high-dimensional feature vector that integrates both rock hardness and grain structure microscopic information. It will serve as the digital fingerprint representing the rock sample in the whole method and be input into the subsequent graph neural network model for deeper relationship modeling and knowledge mining. Features are encoded separately: , Final output: ,in This is a characteristic vector representing the mineral hardness distribution. It is an ordered vector calculated from the Mohs hardness value and its mass percentage of each mineral, sorted in descending order of hardness value and then normalized using a minimum-maximum normalization process. The value range of its elements is... , The feature vector of the particle structure is a two-dimensional vector composed of the average particle size (d50) and the standard deviation of the particle size distribution obtained from statistical analysis of particle size data. , and All are real numbers greater than 0; Here, represents the weight matrix of the first encoding layer. These are the learned parameters used to map the original features to a more suitable latent space for fusion. Their dimensions determine the length of the encoded vector. This is the corresponding bias vector; These are the weight matrix and bias vector of the fusion layer, used to learn the linear combination relationship between the two encoded features; Nonlinear activation functions (such as ReLU and Sigmoid) introduce nonlinear transformations, enabling the model to learn complex feature patterns. For vector concatenation operations, and Connecting the two ends together forms a longer one-dimensional vector, i.e., a high-dimensional feature vector.

[0018] Step 2: Input the high-dimensional feature vector into the graph neural network model, and use the two feature vectors contained therein, which respectively represent the mineral hardness and grain structure, as nodes. Calculate the attention weights between nodes through the graph attention mechanism as the weights of the edges to construct a structural feature relationship map of the rock. In a specific implementation, a structural feature relationship map of rocks is constructed. The aim is to explicitly model and quantify the intrinsic relationship between mineral hardness features and grain structure features through the advanced model of graph neural networks. This transforms the static microscopic properties of rocks into a dynamic, computable structured knowledge representation. The core of this method is that it not only preserves the independence of the two types of features, but also learns a relationship model that can reflect the synergistic influence of the internal structure of rocks against the fracturing process by simulating their physical interaction process. The process of inputting high-dimensional feature vectors into a graph neural network model is as follows: Extract the portion corresponding to the first encoded feature vector from the high-dimensional feature vector as the initial feature vector of the first node, and extract the portion corresponding to the second encoded feature vector as the initial feature vector of the second node; construct a complete graph structure containing two nodes, where each node is connected to another node and itself via edges; calculate the attention coefficients between each node and all its neighboring nodes using a graph attention mechanism, where neighboring nodes include the node itself and another node; normalize the attention coefficients using the softmax function to obtain normalized attention weights; for each node, sum the initial feature vectors of all its neighboring nodes based on the corresponding normalized attention weights to generate the updated feature vector of that node. The updated first node feature vector, the updated second node feature vector, and the normalized attention weights between adjacent nodes are used as edge weights, and the output is a structural feature relationship map of rocks. A complete graph structure containing two nodes is constructed. The self-loop is not an adjustable parameter, but a fixed graph structure setting. Its function is to ensure that when a node aggregates adjacent information, it always takes the features of its previous layer as a definite input source. This is a standard technique in graph neural networks. In the specific implementation, for a graph consisting of two nodes, the calculation process of the attention mechanism is as follows: For each node in the graph, the original attention coefficient between it and all its neighboring nodes (including the node itself and another node) is calculated. The original attention coefficient reflects the correlation strength between the two nodes. Specifically, the initial feature vector of one node is concatenated with the initial feature vector of the other node, and then the concatenated vector is input into a shared attention calculation module. This module consists of a linear transformation layer and a non-linear activation function (such as LeakyReLU), and finally outputs a scalar value as the original attention coefficient. For each node, the original attention coefficients corresponding to all neighboring nodes (including itself and another node) are normalized using the softmax function, ensuring that the sum of the normalized attention weights of each node with respect to all its neighboring nodes is 1. For a graph consisting of two nodes, each node will receive two normalized attention weights: for the first node, its normalized attention weight with respect to itself is denoted as... And its normalized attention weight for the second node is denoted as For the second node, its normalized attention weight over itself is denoted as... And its normalized attention weight for the first node is denoted as 1 is the index of the first node, 2 is the index of the second node, and these weight values ​​are all between 0 and 1, representing the proportion of the node's own features and the features of another node when updating the node features. The updated feature vector of each node is obtained by weighted summing of the initial feature vectors of all its neighboring nodes. Specifically, for node 1, its updated feature vector is equal to the initial feature vector of node 1 multiplied by... Add the initial feature vector of node 2 multiplied by Similarly, the updated feature vector of node 2 is equal to the initial feature vector of node 2 multiplied by... Add the initial feature vector of node 1 multiplied by Finally, the updated first node feature vector, the updated second node feature vector, and the normalized attention weights between adjacent nodes are used as edge weights and output together as the structural feature relationship map of the rock. The reason for dividing the high-dimensional feature vector into two parts and defining them as graph nodes is that the first part of the encoded and fused high-dimensional feature vector originates from the encoding of the mineral hardness distribution characteristics, carrying the hardness information of the rock's resistance to indentation and engraving; the second part originates from the encoding of the particle structure characteristics, carrying the geometric and statistical information of the rock particle system. These two parts of information are different in physical origin, but they have a strong mechanical coupling in the energy dissipation process of explosive fragmentation. Defining them as independent nodes, namely the first node and the second node, is to be able to explicitly express and measure this coupling relationship in subsequent graph calculations. When constructing the graph structure, a complete graph containing two nodes is used, where each node is connected to the other node and itself through an edge. This design aims to provide a complete and symmetrical interactive topology for subsequent attention calculations. Adding self-loops to each node is an important technique in graph neural networks. Its role is to ensure that when each node updates its own features, it can retain its previous feature information as an important reference and integrate it into the new feature representation. It should be noted that the core idea of ​​calculating the attention weights between nodes through the graph attention mechanism is to mimic the human attention allocation mechanism, allowing each node to adaptively and selectively aggregate information from its neighboring nodes. In the specific graph structure constructed in this invention, for any given node, its neighboring nodes are defined as nodes directly connected to it through edges. Specifically, due to the existence of self-loops, each node itself is also one of its neighboring nodes; at the same time, through that unique undirected edge, the first node and the second node are neighboring nodes to each other. The computation process begins with calculating the attention coefficient. For a node in the graph, the graph attention mechanism performs the following operations for each of its neighboring nodes: First, the initial feature vector of the node is concatenated with the initial feature vectors of the neighboring nodes to form a combined vector; second, this combined vector is input into a computation function (such as a small neural network) for operation, and finally a scalar output is obtained. This scalar is the original attention coefficient of the node to that neighboring node. The coefficient reflects the relative importance of the information obtained by the node from that neighboring node in the context of the current task. Then, normalization is performed. In order to make the sum of the attention weights of a single source node to all its neighboring nodes equal, so as to facilitate the representation of the probability distribution, the original attention coefficients corresponding to all neighboring nodes need to be normalized by applying the Softmax function. The value obtained after normalization is the normalized attention weight. This weight is a value between 0 and 1. The larger the weight, the greater the share or influence of the corresponding neighboring node features when updating the source node features. As mentioned earlier, normalized attention weights are a direct tool for quantifying coupling strength. Their calculation begins with the original attention coefficients, which are jointly determined by the features of the source and target nodes. Specifically, the feature vectors of the two nodes are first concatenated, and then this joint representation is applied to a computational function (e.g., a linear transformation followed by a non-linear activation function), ultimately generating a scalar coefficient. This coefficient reflects the original importance of neighboring nodes to a node without considering global relationships. Subsequently, the original coefficients of all neighboring nodes (including the node itself) of all nodes are normalized using the Softmax function. The Softmax function ensures that the sum of the attention weights of a single node to all its neighboring nodes is 1, making the weights probabilistically interpretable and stabilizing the training process. The resulting normalized attention weight is a value between 0 and 1, which is directly used as the coefficient for weighted summation to generate the updated feature vectors of the nodes. This entire mechanism evolves the correlation strength between nodes from a fixed connection to a set of dynamic weight coefficients adapted to specific rock samples. Finally, feature weighting aggregation and graph generation are performed. For each node in the graph, the updated feature vector is generated by multiplying the initial feature vectors of all its neighboring nodes by their corresponding normalized attention weights, and then summing these weighted feature vectors. The resulting vector is the updated feature vector of the node after incorporating the information of its neighboring nodes. Through this process, the first node's features incorporate the information of the second node, and the second node's features incorporate the information of the first node, while each retains its own core information. The updated first node feature vector, the updated second node feature vector, and the normalized attention weights connecting them are used as edge weights and output together to form the rock structural feature relationship graph defined in this invention. This graph is a complete data structure containing node attributes and connection relationships (represented in the form of attention weights).

[0019] Step 3: Using the structural feature relationship map as input and the benchmark explosive consumption as a constraint, the trained deep generative network learns the nonlinear mapping relationship from structural features to explosive distribution and outputs the explosive distribution matrix corresponding to the spatial location of the area to be detonated. The explosive distribution matrix is ​​then spatially aligned and scaled with the preset blasting hole network spatial layout parameters to generate the explosive distribution design map. In a specific implementation, the structural feature relationship map is used as input. The specific process is as follows: the trained deep generative network is pre-trained on historical data, which includes multiple sets of known structural feature relationship maps and corresponding charge distribution matrices determined by experts; during the training process, the benchmark explosive consumption is input into the deep generative network as a numerical constraint, ensuring that the total explosive consumption corresponding to the output charge distribution matrix is ​​not greater than the upper limit of the total charge calculated based on the benchmark explosive consumption; the structural feature relationship map of the area to be detonated is input into the trained deep generative network, which directly outputs the charge distribution matrix corresponding to the spatial location of the area to be detonated. The reason for using deep generative networks is that there is a highly nonlinear mapping relationship between complex structural features and fine spatial drug distribution. This relationship is difficult to fully describe using traditional physical formulas or empirical formulas. Deep generative networks have powerful feature extraction and complex function approximation capabilities, and can learn this implicit, high-dimensional mapping pattern from a large number of historical cases. It should be noted that the trained deep generative network is pre-trained on historical data. This historical data is the source of knowledge for the model's learning, and its quality directly determines the model's performance. Each set of historical data contains two parts: first, a known structural feature relationship map, which is a standardized map obtained after processing rock samples collected before a certain blasting operation in steps 1 and 2; second, a corresponding charge distribution matrix determined by experts. This refers to a charge distribution matrix designed by experienced blasting engineers based on factors such as the geological survey report of the area to be blasted, rock mass structure analysis, expected blasting effect, and economic cost. This matrix defines the recommended explosive consumption per unit area at each discrete location (such as a voxel grid). Its design goal is to achieve the expected blasting block size distribution and throwing effect. This charge distribution matrix generated by expert knowledge serves as the target output of the training network. During training, inputting the benchmark explosive consumption as a numerical constraint into the deep generative network is a key design feature. This aims to ensure that the model-generated schemes comply with the principles of engineering economy and safety. Specifically, this is achieved by adding a total explosive consumption constraint loss term to the network's loss function. This loss term is calculated by first summing the values ​​of all elements in the currently generated explosive distribution matrix to obtain the total explosive consumption predicted by the network. Simultaneously, based on the benchmark explosive consumption value determined by engineering geological analogy or small-scale experiments and applicable to the general lithology of the current mining area, multiplying it by the total volume of the blasting design, the theoretically allowed upper limit of the total explosive consumption is calculated. The loss function penalizes cases where the network predicts the total explosive consumption exceeding this upper limit. Through backpropagation, the network is forced to learn to allocate explosives while satisfying the total explosive consumption constraint during training. After training, the newly generated rock structure feature relationship map of the area to be blasted is input into the network, and the network can instantly infer and output a matching explosive distribution matrix based on the learned mapping rules. The baseline explosive consumption is determined by analogy based on long-term engineering experience in the mining area, or by an empirical value obtained through on-site small blasting funnel tests. It is a specific numerical input, such as 0.35. The depth generation network outputs a charge distribution matrix corresponding to the spatial location of the area to be blasted. Each element in the charge distribution matrix represents the suggested explosive consumption at a spatial location point within the area to be blasted. It should be noted that the charge distribution matrix is ​​a core output concept. It is a three-dimensional numerical array, with its three dimensions corresponding to the spatial division of the area to be detonated in the length, width, and depth directions, respectively. When creating this matrix, the three-dimensional space of the area to be detonated is first divided into voxels according to a fixed resolution (e.g., one grid cell per cubic meter). Each element in the charge distribution matrix uniquely corresponds to such a three-dimensional voxel grid. The value of this element represents the explosive consumption per unit volume of rock recommended by the model of this invention. This matrix intuitively reveals the differences in energy input based on the changes in the microstructure of the rock in a spatial manner. The charge distribution matrix is ​​spatially aligned and scaled with the preset blasting hole network spatial layout parameters to generate a charge distribution design drawing. The specific process is as follows: based on the hole spacing, row spacing and step height in the preset blasting hole network spatial layout parameters, the design load volume of each blast hole is calculated. The design load volume is the product of the hole spacing, row spacing and step height. For each borehole, based on its center coordinates and charge section length, a corresponding influence area is delineated in the charge distribution matrix. The average value of all elements in the charge distribution matrix within the influence area corresponding to each borehole is calculated. This average value is used as the suggested average explosive consumption per borehole. The product of the suggested average explosive consumption per borehole and the design load volume of the corresponding borehole is calculated as the charge per borehole. The charge per borehole of all boreholes is summarized to form the charge distribution design diagram. Calculate the design load volume for each borehole. This is a common term in blasting engineering design, specifically referring to the volume of rock that a single borehole needs to break and throw under its blasting action. The standard calculation formula is: the design load volume is equal to the product of the borehole spacing, row spacing, and step height of the area covered by the borehole. This is a deterministic calculation value based on the borehole spacing and the geometry of the blasting step. Subsequently, a spatial alignment operation is performed. For each borehole in the preset blasting scheme, based on its three-dimensional center coordinates and charge section length, a unique spatial influence area can be determined in the three-dimensional spatial coordinate system corresponding to the above charge distribution matrix. This area is a cylinder with the borehole as the center, its charge section as the height, and the equivalent range of its load area as the base. In the charge distribution matrix, this area covers a series of voxel grids. Then, a scale transformation operation is performed to calculate the arithmetic mean of the charge distribution matrix elements corresponding to all voxel grids within the influence area of ​​the borehole. This mean is the suggested average charge consumption per unit volume of the borehole, representing the average energy intensity that the model suggests to be injected per unit volume of rock around the borehole. Finally, the suggested average charge consumption per unit volume is multiplied by the calculated design load volume of the borehole to obtain the specific single-hole charge recommendation value for the borehole. Finally, the single-hole charge recommendation values ​​of all boreholes are summarized and organized and presented in the form of drawings or digital lists, thus generating a charge distribution design drawing that can be directly used for on-site drilling and charging operations. This design drawing realizes the final implementation of the blasting charge from rock structural characteristics to the spatial location of the borehole.

[0020] Step 4: After blasting based on the charge distribution design, collect multi-dimensional evaluation indicators of the actual blasting effect to construct a multi-objective optimization function. Use optimization algorithms to adjust the parameters in the graph attention mechanism and deep generation network in reverse according to the difference between the actual blasting effect and the expected target, so as to achieve iterative optimization of the charge amount. In a specific implementation, the core objective of this step is to establish a closed-loop self-evolution mechanism, enabling the explosive charge design method to automatically adjust its internal model parameters based on feedback from practical results, thereby continuously improving the accuracy and adaptability of the design. This step transforms blasting engineering from a one-time static design into a dynamic method that can continuously learn and optimize. The process of collecting multi-dimensional evaluation index data of actual blasting effects to construct a multi-objective optimization function is as follows: After blasting is carried out according to the explosive distribution design drawing, a drone is used to take aerial photos of the blast pile from multiple angles. A digital surface model of the blast pile is generated through three-dimensional reconstruction technology. The digital surface model is segmented and identified into rock blocks. The equivalent particle size of each rock block is counted. Based on the preset threshold particle size for large blocks and the threshold particle size for fine ore, the ratio of large blocks and the ratio of fine ore are calculated. The ratio of large blocks is the percentage of the total volume of rock blocks with an equivalent particle size greater than the threshold particle size to the total volume of the blast pile. The ratio of fine ore is the percentage of the total volume of rock blocks with an equivalent particle size less than the threshold particle size to the total volume of the blast pile. The key to collecting multi-dimensional evaluation indicators of actual blasting effects lies in transforming the blasting results in the physical world into quantitative data that can be understood and processed by computer models. After blasting is carried out according to the explosive distribution design, a high-definition camera mounted on a drone is used to take multi-angle, overlapping aerial photos of the blast pile to obtain a series of images covering the entire blast pile. Then, through motion recovery structure and multi-view stereo vision algorithms in computer vision, these two-dimensional images are reconstructed into a three-dimensional digital surface model of the blast pile. This model is a dense point cloud or grid that accurately records the three-dimensional coordinates of each point on the surface of the blast pile. The three-dimensional digital surface model is segmented and identified into rock blocks. This process uses a segmentation algorithm based on three-dimensional geometric features, such as a region growing algorithm based on abrupt changes in point cloud normals or curvature changes. The connected rock point cloud clusters are segmented into independent individuals. Each segmented point cloud cluster represents a rock block. For each identified rock block, its volume is calculated, and based on the principle of volume equivalence, its equivalent grain size is calculated, which is the diameter of a sphere with the same volume as the rock block. Based on the statistical calculation of equivalent particle size, two core evaluation indicators are used: large block ratio and fine ore ratio. Two key threshold parameters are introduced: the large block threshold particle size and the fine ore threshold particle size. The large block threshold particle size is a pre-set particle size value based on the process requirements of subsequent crushing equipment, such as the feed inlet size of a jaw crusher, for example, set to 1000 mm. The fine ore threshold particle size is a pre-set particle size value based on the control requirements for over-grinding during mineral processing, for example, set to 10 mm. The large block ratio is calculated as follows: the total volume of all rock blocks with an equivalent particle size larger than the large block threshold particle size is counted, divided by the total volume of the entire blast pile, and then multiplied by 100%. The resulting percentage is the large block ratio. The fine ore ratio is calculated as follows: the total volume of all rock blocks with an equivalent particle size smaller than the fine ore threshold particle size is counted, divided by the total volume of the entire blast pile, and then multiplied by 100%. The resulting percentage is the fine ore ratio. These two indicators quantify the quality of the blasting crushing effect from different dimensions: an excessively high large block ratio indicates insufficient crushing, while an excessively high fine ore ratio indicates energy waste and excessive dust generation. The optimization objectives are to minimize the bulk ore ratio and make the fine ore ratio fall within a preset target range. A multi-objective optimization function is constructed based on these objectives. An optimization algorithm is employed to iteratively optimize based on the difference between the actual blasting effect and the expected target. The specific process is as follows: the actual block rate and fine ore rate data collected after the actual blasting are substituted into a multi-objective optimization function to calculate the actual loss value that quantifies the deviation between the actual blasting effect and the expected target; using the backpropagation algorithm, starting from the actual loss value, the gradients of the actual loss value with respect to the weight parameters and bias parameters of each layer of the deep generative network are calculated sequentially, and the gradients of the actual loss value with respect to all training parameters in the graph neural network model are also calculated. The training parameters include the learned weight matrix that performs a linear transformation on the node feature vector in the graph attention mechanism, and the learned attention vector used to calculate the attention coefficients between nodes. The optimization objectives of this invention are twofold: first, to minimize the lump ratio; and second, to control the fines ratio within a preset target range, such as 5% to 15%. To express both objectives simultaneously in a single function, a weighted summation loss function is constructed. This function includes two sub-loss terms: the first is the lump ratio loss, which is, for example, directly taking the actual lump ratio value; the second is the fines ratio deviation loss, used to penalize deviations from the preset target range. This term is zero when the actual fines ratio is within the target range; if it is below the lower limit of the range, this term is the sum of the lower limit and the actual value. The difference; if it is higher than the upper limit of the interval, then this term is the difference between the actual value and the upper limit of the interval; finally, the first sub-loss and the second sub-loss are multiplied by a weight coefficient reflecting their engineering importance and then added together, which constitutes the multi-objective optimization function. The value of this function is a scalar, and its magnitude comprehensively represents the overall deviation between the current blasting effect and the ideal target, which is called the actual loss value; finally, the optimization algorithm is used to iteratively optimize the model parameters. This process is a standard gradient descent-based machine learning model fine-tuning process, but its application objects and purposes have clear engineering targets; The multi-objective optimization function is a mathematical expression whose input is an evaluation index of the actual blasting effect (such as the large lump rate and the fine ore rate), and whose output is a scalar value. The design of this function aims to combine the two objectives of minimizing the large lump rate and controlling the fine ore rate within a certain range into a single, optimized objective through weighted summation and other methods. The actual loss value is the specific scalar result calculated by the multi-objective optimization function after receiving the actual data of this blasting. It is a quantitative measure of the deviation of the blasting effect from the ideal state. The larger the value, the greater the deviation and the less ideal the effect; the smaller the value, the closer it is to the expectation. The specific process involves substituting the actual large block rate and fine ore rate calculated after the blasting into the multi-objective optimization function constructed above to calculate a specific actual loss value. This actual loss value is the quantitative score of the blasting design effect. Then, using the backpropagation algorithm, starting from the actual loss value, the algorithm proceeds backward along the computation graph. The algorithm will sequentially calculate the gradient of the actual loss value with respect to each training parameter in the deep generative network (including all weight parameters and bias parameters of each layer of the neural network). At the same time, it will also calculate the gradient of the actual loss value with respect to all training parameters in the graph neural network model. These training parameters specifically include the learned weight matrix used to linearly transform the node feature vectors in the graph attention mechanism, and the learned attention vector used to calculate the correlation score between nodes. in, The actual large block ratio was calculated using 3D reconstruction of the blast pile and rock block identification technology. It represents the powder ore ratio, obtained in the same way as the large block ratio, and reflects the degree to which the rock blocks are too fragmented and produce too much powder after blasting; The target lower limit for the fines ratio is determined based on the economic viability of mineral resource recovery (avoiding excessive crushing and loss of valuable minerals). The fines ratio should not be lower than this value; otherwise, it implies insufficient energy utilization or an overly conservative blasting design. It is determined based on mineral value and process, and is, for example, a constant value between 3% and 8%. ; The powder ratio should not exceed this value, as determined by workplace dust control standards, occupational health requirements, and subsequent screening efficiency. Otherwise, it could lead to increased dust hazards, environmental pollution, and blockages in subsequent processes. Therefore, it is determined according to environmental protection and safety regulations, for example, a constant value between 10% and 20%. ; This is the weighting coefficient for large block ratio loss, determined based on the importance attached to controlling the large block ratio in the project. It is used to balance the contribution of large block ratio loss to the total loss and is a positive real number. This is the weighting coefficient for the loss of fines ratio, determined based on the importance attached to controlling the fines ratio in the project. It is used to balance the contribution of the fines ratio loss to the total loss and is a positive real number. Defined as This is a linear rectification function used to implement conditional judgment: when the fines ratio is within the target range At that time, and The loss is zero; when the fines ratio is below the lower limit. hour, This results in losses; when the fines ratio exceeds the upper limit. hour, This results in losses; set up , , , The actual loss value is based on the theoretical loss value with added normal distribution noise. The region is divided according to the value of the fines ratio. Marked as the left-side penalty area. Marked as the right-side penalty area. Marked as a dead zone Marked as the upper boundary of the dead zone. The lower bound of the dead zone is marked; specific data on some sample numbers and loss values ​​are shown in Table 1.

[0021] Table 1. Statistics on Loss Values Through data analysis, a clear correlation was observed between the fines ratio and the corresponding loss value. The data shows that the fines ratio has a key impact on the overall loss value. When the fines ratio is in the middle range, the overall loss value mainly depends on the actual lump ratio. The fines ratio itself does not bring additional loss. For example, the fines ratios of samples 2, 4 and 6 are all in this middle range, and their theoretical loss values ​​are equal to the actual lump ratio values. However, when the fines ratio is below a certain level, the overall loss value will increase significantly, exceeding the loss represented by the actual large lump ratio. Samples 1, 5, and 8 are examples of this, as their fines ratios are all low, and their theoretical loss values ​​are all higher than the actual large lump ratios, indicating that an excessively low fines ratio will lead to significant additional losses. Conversely, when the fines ratio exceeds a certain level, it will also lead to an increase in overall losses. Samples 3, 7, and 13 have relatively high fines ratios, and their theoretical loss values ​​are correspondingly higher than their respective actual large lump ratios, indicating that an excessively high fines ratio is also not conducive to the comprehensive evaluation of blasting effects. Further analysis revealed that the actual loss value fluctuated slightly around the theoretical loss value, reflecting the randomness present in actual observations. However, the overall trend closely matched the theoretical expectations. For example, the actual loss value of sample nine was very close to the theoretical value, while the actual value of sample fifteen was slightly higher than the theoretical value. These subtle differences did not affect the manifestation of the overall pattern. In conclusion, when evaluating the blasting effect, it is necessary to pay attention not only to the large block ratio but also to control the fine ore ratio within a reasonable range. Both factors jointly determine the final loss assessment result, which is of great significance for optimizing blasting design.

[0022] Learn the dimension of the weight matrix, which is used to perform a linear transformation on the node features. If the dimension of the node feature vector is D (e.g., ...), then... Then, the typical shape of the learning weight matrix is: or ,in This refers to the transformed dimension, which is a parameter automatically learned by the model based on the data; it's the dimension of the learned attention vector, used to calculate the association score, and its length is equal to the length of the transformed node feature concatenation vector. If the transformed feature dimension is... Then the spliced ​​result is The length of the attention vector is... ; The gradient descent algorithm is adopted. Based on the calculated gradient, the weights and bias parameters of the deep generative network and the training parameters in the graph neural network model are numerically adjusted in the opposite direction of the gradient. The graph neural network model and the deep generative network with updated parameters are used for the next blasting charge design. This process is repeated until the multi-objective optimization function value converges or the large block rate and fine ore rate of the actual blasting effect reach the expected target, so as to realize the iterative optimization of the blasting charge design method. The gradient is a vector indicating how each parameter needs to be adjusted (increased or decreased) to most effectively reduce the actual loss value. After calculating the gradients of all parameters, the gradient descent algorithm or its variants, such as the Adam optimizer, are used. This algorithm performs a synchronous numerical adjustment on all parameters based on the gradient direction and magnitude of each parameter. The adjustment direction is opposite to the gradient direction, and the goal is to reduce the value of the multi-objective optimization function, i.e., the actual loss value, in the next calculation. After the parameter adjustment is completed, the updated graph neural network model and deep generative network have a new state that has learned from the experience of this blasting effect. Using these two updated models for the charge design of the next area to be blasted completes a complete iterative optimization loop. Repeating this process, the parameters will be continuously updated in multiple loops, eventually causing the value of the multi-objective optimization function to converge to a stable small value, or enabling the output blasting design scheme to stably achieve the expected target for both the block ratio and the fine ore ratio, thereby realizing the intelligent iterative optimization and performance improvement of the entire blasting charge design method. In iterative optimization, the model parameters are continuously updated using the gradient descent algorithm. Each update aims to make the actual loss value calculated after the next brute-force attack smaller than the previous one. After multiple (K) iterations, the change in the actual loss value will tend to level off. Therefore, a specific and quantifiable convergence criterion is set. For example, it is determined that in P consecutive iterations (such as 5 iterations), the decrease in the actual loss value is less than a preset minimum threshold (such as 0.001). When this condition is met, the optimization process is considered to have converged and the model parameters tend to stabilize. The expected objective is a pre-set specific engineering indicator for measuring the success of blasting. In this invention, the explicit objective is to minimize the proportion of large lumps while ensuring that the proportion of fine ore falls within a preset target range. Within, for example, the target interval is After a blast, directly check whether the actual large block rate collected and calculated is continuously decreasing and whether the actual fine ore rate is within the range of 5% to 15%. If both conditions are met, it is determined that the blast has achieved the expected goal. The ultimate goal of iterative optimization is to make the design output by the model stably meet these two conditions for the blasting effect.

[0023] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0024] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0025] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0026] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A method for refined design of explosive charge quantity based on rock mineral particle structure, characterized in that, The specific steps include: Step 1: Map the mineral composition mass percentage data of the rock sample in the area to be blasted to the mineral hardness distribution characteristics, and map its particle size distribution data to the particle structure characteristics. Perform nonlinear encoding and combination on the two features to generate a high-dimensional feature vector characterizing the microstructure of the rock. Step 2: Input the high-dimensional feature vector into the graph neural network model, and use the two feature vectors contained therein, which respectively represent the mineral hardness and grain structure, as nodes. Calculate the attention weights between nodes through the graph attention mechanism as the weights of the edges to construct a structural feature relationship map of the rock. Step 3: Using the structural feature relationship map as input and the benchmark explosive consumption as a constraint, the trained deep generative network learns the nonlinear mapping relationship from structural features to explosive distribution and outputs the explosive distribution matrix corresponding to the spatial location of the area to be detonated. The explosive distribution matrix is ​​then spatially aligned and scaled with the preset blasting hole network spatial layout parameters to generate the explosive distribution design map. Step 4: After blasting based on the charge distribution design, collect multi-dimensional evaluation indicators of the actual blasting effect to construct a multi-objective optimization function. Using an optimization algorithm, adjust the parameters in the graph attention mechanism and deep generative network in reverse according to the difference between the actual blasting effect and the expected target to achieve iterative optimization of the charge amount.

2. The method for refined design of explosive charge quantity based on rock mineral particle structure according to claim 1, characterized in that: X-ray diffraction analysis was performed on the rock sample to obtain the mass percentage data of each mineral in the rock sample. Digital image processing analysis was performed on the scanning electron microscope image of the rock sample to obtain the particle size distribution data of its interior. Mapping mineral composition mass percentage data to mineral hardness distribution characteristics involves the following steps: extracting the Mohs hardness scale value for each mineral and multiplying it by the corresponding content proportion of that mineral in the mass percentage data to obtain the hardness contribution value of each mineral; then arranging the hardness contribution values ​​of all minerals in the rock sample in descending order of their Mohs hardness values ​​and normalizing them to form a vector form of mineral hardness distribution characteristics; mapping particle size distribution data to particle structure characteristics involves statistical analysis of the particle size distribution data, extracting the average particle size and the standard deviation of the particle size distribution as key statistics, and constructing a vector from these key statistics to form particle structure characteristics. The mineral hardness distribution features are input into the first encoder to obtain the first encoded feature vector; the particle structure features are input into the second encoder to obtain the second encoded feature vector; wherein the first encoder and the second encoder each contain at least one fully connected layer and a nonlinear activation function; the first encoded feature vector and the second encoded feature vector are concatenated to generate a high-dimensional feature vector, wherein the high-dimensional feature vector is formed by sequentially concatenating the first encoded feature vector and the second encoded feature vector.

3. The method for refined design of explosive charge quantity based on rock mineral particle structure according to claim 2, characterized in that: The process of inputting high-dimensional feature vectors into a graph neural network model is as follows: extract the part corresponding to the first encoded feature vector from the high-dimensional feature vector as the initial feature vector of the first node, and extract the part corresponding to the second encoded feature vector as the initial feature vector of the second node; construct a complete graph structure containing two nodes, where each node is connected to another node and itself through an edge; calculate the attention coefficient between each node and all its neighboring nodes through a graph attention mechanism, wherein the neighboring nodes include the node itself and another node; and normalize the attention coefficients using the softmax function to obtain normalized attention weights. For each node, the initial feature vectors of all its neighboring nodes are weighted and summed based on the corresponding normalized attention weights to generate the updated feature vector of that node. The updated first node feature vector, the updated second node feature vector, and the normalized attention weights between adjacent nodes are used as edge weights, and the output is a structural feature relationship map of the rock.

4. The method for refined design of explosive charge quantity based on rock mineral particle structure according to claim 3, characterized in that: The structural feature relationship map is used as input. The specific process is as follows: the trained deep generative network is pre-trained on historical data, which includes multiple sets of known structural feature relationship maps and corresponding charge distribution matrices determined by experts; during the training process, the benchmark explosive consumption is input into the deep generative network as a numerical constraint, so that the total explosive consumption corresponding to the output charge distribution matrix is ​​not greater than the upper limit of the total charge calculated based on the benchmark explosive consumption. The structural feature relationship map of the area to be detonated is input into the trained deep generation network, and the deep generation network directly outputs the charge distribution matrix corresponding to the spatial location of the area to be detonated. The deep generative network outputs a charge distribution matrix corresponding to the spatial location of the area to be detonated. Each element in the charge distribution matrix represents the suggested charge consumption at a spatial location point within the area to be detonated.

5. The method for refined design of explosive charge quantity based on rock mineral particle structure according to claim 4, characterized in that: The charge distribution matrix is ​​spatially aligned and scaled with the preset blasting hole network spatial layout parameters to generate a charge distribution design drawing. The specific process is as follows: based on the hole spacing, row spacing and step height in the preset blasting hole network spatial layout parameters, the design load volume of each blast hole is calculated. The design load volume is the product of the hole spacing, row spacing and step height. For each borehole, based on its center coordinates and charge section length, a corresponding influence area is defined in the charge distribution matrix. The average value of all elements in the charge distribution matrix within the influence area corresponding to each borehole is calculated. This average value is used as the suggested average explosive consumption per borehole. The product of the suggested average explosive consumption per borehole and the design load volume of the corresponding borehole is calculated as the charge per borehole. The charge per borehole of all boreholes is summarized to form the charge distribution design diagram.

6. The method for refined design of explosive charge quantity based on rock mineral particle structure according to claim 5, characterized in that: The process of collecting multi-dimensional evaluation index data of actual blasting effects to construct a multi-objective optimization function is as follows: After blasting is carried out according to the explosive distribution design drawing, a drone is used to take aerial photos of the blast pile from multiple angles. A digital surface model of the blast pile is generated through three-dimensional reconstruction technology. The digital surface model is segmented and identified into rock blocks. The equivalent particle size of each rock block is counted. Based on the preset threshold particle size for large blocks and the threshold particle size for fine ore, the ratio of large blocks and the ratio of fine ore are calculated. The ratio of large blocks is the percentage of the total volume of rock blocks with an equivalent particle size greater than the threshold particle size to the total volume of the blast pile. The ratio of fine ore is the percentage of the total volume of rock blocks with an equivalent particle size less than the threshold particle size to the total volume of the blast pile. The optimization objectives are to minimize the bulk ore ratio and make the fine ore ratio fall within the preset target range. A multi-objective optimization function is constructed based on the optimization objectives.

7. The method for refined design of explosive charge quantity based on rock mineral particle structure according to claim 6, characterized in that: An optimization algorithm is adopted to achieve iterative optimization based on the difference between the actual blasting effect and the expected target. The specific process is as follows: the actual large block rate and fine ore rate data collected after the actual blasting are substituted into the multi-objective optimization function to calculate the actual loss value of the deviation between the actual blasting effect and the expected target. Using the backpropagation algorithm, starting from the actual loss value, the gradients of the weight parameters and bias parameters of each layer of the deep generative network are calculated sequentially, and the gradients of the actual loss value with respect to all training parameters in the graph neural network model are also calculated. The training parameters include the learned weight matrix that performs a linear transformation on the node feature vector in the graph attention mechanism, and the learned attention vector used to calculate the attention coefficients between nodes. The gradient descent algorithm is used to adjust the weights and bias parameters of the deep generative network and the training parameters in the graph neural network model in the opposite direction of the gradient, based on the calculated gradient. The updated graph neural network model and the deep generative network are then used for the next blast charge design. This process is repeated until the multi-objective optimization function converges or the block ratio and fine ore ratio of the actual blasting effect reach the expected target, thus realizing the iterative optimization of the blast charge design method.

Images