Mine geological structure safety prediction method and system fusing multi-dimensional features

Abstract

The application provides a mine geological structure safety prediction method and system fusing multi-dimensional features, relates to the field of coal mine safety production, and solves the technical problems that it is difficult to capture complex geological spatial relationships and the reliability of risk prediction results is insufficient in the prior art. The method comprises the following steps: collecting multi-source heterogeneous data of a mine, and performing spatio-temporal alignment and standardization processing on the multi-source heterogeneous data to obtain preprocessed data; based on the preprocessed data, abstracting a mine geological region into a graph structure, wherein the nodes of the graph structure represent geological entity units, and the edges represent the interaction relationship between the geological entities, and generating a geological graph comprising an adjacency matrix and a node feature matrix; inputting the geological graph into a geological graph neural network model constructed based on a deep learning algorithm, aggregating node features through multi-layer convolution operation, and combining physical constraint conditions to perform physical consistency verification and optimization on the model output to obtain the distribution of risk probability, so as to realize safety prediction of the geological structure.

Description

Technical Field

[0001] This invention belongs to the field of coal mine safety production and involves neural network technology, specifically a method and system for predicting the safety of mine geological structures by integrating multi-dimensional features. Background Technology

[0002] Predicting the safety of mine geological structures is a crucial aspect of ensuring safe coal mine production. As coal mining extends to deeper levels and geological conditions become more complex, traditional prediction methods face severe challenges. At the data level, traditional methods often rely on single types of geological or monitoring data, lacking effective integration of multi-source heterogeneous data and failing to comprehensively reflect the complex characteristics of the mine geological system. Secondly, in terms of model construction, existing prediction methods mostly employ simple statistical models or shallow machine learning algorithms, which are ineffective at representing the complex spatial relationships between geological entities. In particular, traditional models lack the ability to accurately characterize the topological connections between key geological structures such as faults and folds, as well as physical interactions such as hydraulic connectivity and stress transfer. These limitations severely affect the prediction models' understanding and simulation of the complex behavior of the mine geological system.

[0003] Furthermore, existing deep learning-based prediction methods often rely excessively on data-driven approaches and lack integration with hydrogeological laws and rock mechanics principles, which can easily lead to prediction results that contradict common sense in physics and result in obviously unreasonable safety risk assessments.

[0004] Therefore, existing mine geological structure safety prediction technologies have significant shortcomings in terms of data integration, model expression, and fusion of physical laws. There is an urgent need to develop a prediction method that can integrate multi-dimensional features, accurately express geological relationships, and conform to physical laws to meet the pressing needs of modern coal mine safety production. Summary of the Invention

[0005] This application provides a method and system for predicting the safety of mine geological structures by integrating multi-dimensional features, which solves the technical problems of existing technologies that make it difficult to capture complex geological spatial relationships and have insufficient reliability of risk prediction results.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides a method for predicting the safety of mine geological structures by integrating multi-dimensional features, including: Collect geological structure data, hydrological monitoring data, rock mass mechanical parameter data, and real-time monitoring data from the mine to obtain multi-source heterogeneous data. Then, perform spatiotemporal alignment and standardization on the multi-source heterogeneous data to obtain pre-processed data. Based on the preprocessed data, the geological area of ​​the mine is abstracted into a graph structure, where the nodes of the graph structure represent geological entity units and the edges represent the interaction relationships between geological entities, generating a geological map including an adjacency matrix and a node feature matrix. The geological map is input into a geological map neural network model built based on a deep learning algorithm. The node features are aggregated through multi-layer convolution operations, and the physical consistency of the model output is verified and optimized in combination with physical constraints to obtain the risk probability distribution. The risk probability distribution is used to realize the safe prediction of geological structures.

[0007] Based on the above technical solution, the mine geological structure safety prediction method integrating multi-dimensional features provided in this application integrates geological structure, hydrology, rock mechanics, and real-time monitoring data, and constructs a geological map structure that integrates spatial topology and physical relationships, thereby achieving a multi-dimensional and integrated characterization of the mine geological system. This application embeds physical constraints into the deep learning-based graph neural network prediction process, which not only fully utilizes data-driven feature learning capabilities but also ensures that the prediction results conform to the basic laws of hydrogeology and rock mechanics, thus significantly improving the accuracy and physical reliability of safety risk prediction.

[0008] Furthermore, the acquisition of the multi-source heterogeneous data includes: Geological structural data were collected through drilling core sampling and geophysical exploration equipment; the geological structural data included fault strike, fold axis location, joint density parameters, and recorded the three-dimensional coordinates and timestamp of each data point; Hydrological monitoring data is collected by deploying an underground sensor network; the hydrological monitoring data includes aquifer permeability coefficient, dynamic changes in water level, and mine water inflow. Rock mass mechanical parameter data are collected using in-situ testing instruments; the rock mass mechanical parameter data includes rock layer density, compressive strength, and in-situ stress distribution; Real-time monitoring data is acquired using microseismometers and gas sensors; the real-time monitoring data includes microseismic time series and gas concentration data. Metadata tags, including data source, collection time and location, are added to all data immediately after collection, and the data is integrated into a unified platform via a data bus.

[0009] Furthermore, the spatiotemporal alignment and standardization processing of the multi-source heterogeneous data includes: The raw data was cleaned using a sliding window anomaly detection algorithm, and the data with different sampling frequencies were unified to the same time interval using cubic spline interpolation to obtain a time-aligned data sequence. The discrete data points in the data sequence are mapped to grid cells using the Kriging spatial interpolation method; the continuous data in the data sequence are normalized, and the categorical data are one-hot encoded to obtain standardized feature data. Add timestamps and spatial coordinate labels to the standardized feature data to generate standardized feature data and a corresponding metadata index table.

[0010] Furthermore, the process of constructing the geological map includes: The mining area is divided into multiple grid units, each grid unit serving as a node on the geological map. The node feature vector is composed of preprocessed geological structure data, hydrological monitoring data, and rock mechanics parameter data. The edges of the geological map are constructed based on the spatial relationships and physical interactions between geological entities, and the edge types are set according to the hydraulic connectivity and stress transfer relationships; the edge types include hydraulic connectivity edges and stress transfer edges. An adjacency matrix A and a node feature matrix X are constructed based on node and edge information, and the topological relationships are stored in a graph database.

[0011] Furthermore, the edge weights of the geological map are determined based on the spatial relationships and physical interactions between geological entities, and the calculation formula is as follows: ;in, For edge weights, Let be the Euclidean distance between nodes. For scale parameters, This is the hydraulic pressure difference based on the permeability coefficient.

[0012] Furthermore, the geological map neural network model employs an attention map convolutional network, including an input module, a feature adaptation module, an attention map convolutional module, a hierarchical feature aggregation module, a regularization module, and an output optimization module; wherein, The input module is used to perform dimension calibration on the node feature matrix and remove redundant edges from the adjacency matrix to obtain the verified node feature matrix and the optimized adjacency matrix. The feature adaptation module is used to split, encode, and weightedly fuse the verified node matrix to obtain multimodal fusion features, and then superimpose the multimodal fusion features with the verified node matrix through residual connections to obtain a node feature matrix adapted from multi-source features. The attention graph convolution module is used to receive the adapted node feature matrix, the optimized adjacency matrix, and the edge type matrix. It introduces edge type factors and spatial attention factors to enhance feature extraction capabilities. Through two layers of graph attention convolution operations and activation processing, the enhanced feature matrix is ​​obtained. The edge type matrix is ​​constructed based on the edge types of the geological map and the hydraulic connectivity strength and stress transfer coefficient between corresponding nodes. It is obtained according to the recorded type representation and corresponding physical parameter values ​​of each edge. The hierarchical feature aggregation module is used to receive the enhanced feature matrix and the geological partition mask, and extract geological features at different scales through a multi-layer aggregation strategy to obtain risk embedding vectors and node aggregation features; wherein, the geological partition mask is obtained according to the geological partition category of each pre-labeled grid cell; The regularization module is used to perform regularization and noise reduction on the risk embedding vector and the node aggregation feature to obtain the regularized risk embedding vector and the regularized node aggregation feature. The output optimization module is used to perform probability mapping and calibration on the regularized risk embedding vector and the regularized node aggregation feature to obtain the node-level risk probability distribution and the full-scale risk probability distribution; wherein, the calibration corrects the risk probability that exceeds the reasonable range through physical constraints.

[0013] Furthermore, the internal workflow of the feature adaptation module is as follows: The verified node matrix is ​​split into geological feature sub-matrix, hydrological feature sub-matrix, rock mechanics feature sub-matrix and real-time monitoring feature sub-matrix according to data mode; By using four independent one-dimensional convolutional layers, feature extraction and dimensionality unification are performed on each feature submatrix to obtain the modality coding features. A trainable weight vector is introduced to adaptively weight and combine the modal coding features to obtain multimodal fusion features; wherein, the trainable weight vector is iteratively updated as the model is trained. A residual connection is constructed, and the multimodal fusion features are superimposed with the verified node matrix to offset the information loss during the feature encoding process, and the node feature matrix after multi-source feature adaptation is output.

[0014] Furthermore, the internal workflow of the attention map convolution module is as follows: Based on the adapted node feature matrix and the optimized adjacency matrix, a first-layer graph attention convolution operation with an edge type factor is performed, and the convolution output features are processed by the ReLU activation function to obtain the first-layer convolution feature matrix. Then, the first-layer convolution feature matrix is ​​superimposed with the adapted node feature matrix through residual connections to obtain an intermediate feature matrix; wherein, the edge type factor corresponding to the hydraulic connectivity edge is... The initial value is set to 0.6, which is the edge type factor for stress transfer edges. The initial value is set to 0.4, and Iterative updates are performed during model training; The spatial attention factor is calculated based on the three-dimensional coordinates of the nodes, and the spatial attention factor is weighted and calculated with the intermediate feature matrix to obtain the spatial association feature. The spatial correlation features are convolved using a second-layer graph attention convolution to obtain a second-layer convolution feature matrix. Then, an average pooling operation is performed on the second-layer convolution feature matrix to obtain an enhanced feature matrix.

[0015] Furthermore, the first layer graph attention convolution is configured with 8 attention heads, and the formula for calculating the attention weights of the first layer graph attention convolution is as follows: ;in, This represents the attention weight coefficient of node i to node j. and For model section training parameters, Let be the set of neighboring nodes of node i. It is an improved version of the ReLU activation function. Let i be the feature vector of node i. Let j be the feature vector of node j. Let k be the feature vector of node k. Let $\frac{i}{j}$ be the edge type factor between node $i$ and node $j$. The edge type factor between node i and node k; The second-layer graph attention convolution is configured with four attention heads, and the attention weights are calculated using a spatial attention factor parameter. The formula for calculating the attention weights of the second-layer graph attention convolution is as follows: ;in, represents the attention weight coefficients of node i to node j in the second-layer graph attention convolution.

[0016] Furthermore, the internal workflow of the hierarchical feature aggregation module is as follows: Based on the geological partitioning mask, the enhanced feature matrix is ​​split into local sub-graph feature matrices corresponding to fault zones, fold zones, and normal rock layers. Each local sub-graph feature matrix is ​​subjected to graph convolution operation to extract geological correlation features within the partition. Then, the dimension of the geological correlation features within the partition is compressed through average pooling operation to generate local feature vectors for each partition, thus obtaining a set of local feature vectors. Pooling operations of different scales are performed on the local feature vectors of each partition to obtain multi-scale pooling results including the first scale, the second scale, and the third scale; wherein the pooling kernel size of the first scale is 1×1, the pooling kernel size of the second scale is 3×3, and the pooling kernel size of the third scale is 5×5. The multi-scale pooling results of each partition are concatenated to obtain a cross-scale feature vector set. The cross-scale feature vector set is then concatenated with the local feature vector set through residual connection to obtain the fused features of each partition. Finally, the fused features of different scales are convolved and concatenated to obtain the cross-partition fused feature vector. The cross-region fusion feature vector is input into a fully connected layer for adaptive weighted fusion. The contribution of each cross-scale feature is adjusted by the parameters of the fully connected layer to obtain a global aggregated feature vector. The global aggregated feature vector is processed in two branches. One branch compresses the global aggregated feature vector into a fixed-dimensional vector through global average pooling to obtain the risk embedding vector. The other branch retains the feature information of the node dimension to obtain the node aggregated feature.

[0017] Furthermore, in the hierarchical feature aggregation module, the formula for generating the local feature vector of each partition is: Where m represents the geological zone category. For partition m, the enhanced feature matrix is... Let m be the adjacency matrix of partition m. For graph convolution operations, For average pooling operations, Let m be the local feature vector of partition m; The formula for generating the cross-scale feature vectors of each partition is as follows: ;in, This represents the result of the first-scale pooling. This represents the result of the second-scale pooling. This represents the result of the third-scale pooling. This represents the cross-scale feature vector of partition m, and the cross-scale feature vector set is... ; The formula for generating the cross-partition fusion feature vector is as follows: ; ;in, Let m be the fusion feature of partition m. This is a cross-region fusion feature vector; The global aggregated feature vector The formula for generating it is: ;in, This is the weight matrix of the fully connected layer. For bias terms; The risk embedding vector and node aggregation features The mathematical formula is: .

[0018] Furthermore, the step of performing physical consistency verification and optimization of the model output in conjunction with physical constraints includes: Physical constraints are defined based on hydrogeological equations and rock mechanics criteria. Physical constraints are incorporated as soft loss terms into the total loss function; After forward propagation of the model training, the physical constraint loss is calculated based on the total loss function, and the model parameters are optimized through backpropagation. During the model inference phase, the output probability of the model is physically consistent with the preset reasonable range of hydrogeology and rock mechanics. If the output probability is found to be inconsistent with the preset range, the probability is corrected according to the preset physical constraint threshold and correction rules.

[0019] Furthermore, the formula for calculating the total loss function is: ;in, Indicates the total loss. , These are the weighting coefficients, and their sum is less than 0.5; The task loss is calculated using the cross-entropy loss function, which reflects the deviation between the predicted risk probability and the actual geological risk label. The hydrological constraint loss is used to constrain the prediction results to conform to the hydrogeological seepage law, and the calculation formula is: , The number of samples for hydrological monitoring data. Let be the aquifer permeability coefficient of the i-th grid cell. Let be the water level gradient of the i-th grid cell; The rock mass mechanical constraint loss is calculated using the following formula: , Let i be the rock mass shear stress of the i-th grid element. Let i be the rock mass cohesion of the i-th grid cell. Let be the rock mass normal stress of the i-th grid cell. Let be the internal friction angle of the rock mass in the i-th grid cell, where i represents the grid cell number.

[0020] Secondly, this application provides a mine geological structure safety prediction system integrating multi-dimensional features, including: a data preprocessing module, a geological map analysis module, and a risk prediction module; wherein, The data preprocessing module is used to collect mine geological structure data, hydrological monitoring data, rock mass mechanical parameter data and real-time monitoring data to obtain multi-source heterogeneous data, and to perform spatiotemporal alignment and standardization processing on the multi-source heterogeneous data to obtain preprocessed data. The geological map analysis module is used to abstract the geological area of ​​the mine into a graph structure based on the preprocessed data, wherein the nodes of the graph structure represent geological entity units and the edges represent the interaction relationships between geological entities, and generate a geological map including an adjacency matrix and a node feature matrix. The risk prediction module is used to input the geological map into a geological map neural network model built based on a deep learning algorithm, aggregate node features through multi-layer convolution operations, and perform physical consistency verification and optimization on the model output in combination with physical constraints to obtain the risk probability distribution; the risk probability distribution is used to realize the safety prediction of geological structures.

[0021] Compared with the prior art, the beneficial effects of this application are: First, this application overcomes the limitations of traditional methods, such as single data sources and inconsistent spatiotemporal scales, by comprehensively collecting and standardizing multi-source heterogeneous data, thus improving the integrity and reliability of the data foundation. Second, this application abstracts the mine geological system into a geological map structure containing complex interactions. By introducing a physical mechanism-based edge weight calculation method and multiple edge types, it breaks through the bottleneck of traditional methods' inadequate characterization of the interaction relationships between geological entities, achieving a precise expression of the topological relationships of the mine geological system.

[0022] Specifically, in the attention graph convolutional network model of this application, the feature adaptation module solves the problem of collaborative utilization of heterogeneous data through the divide-and-conquer and fusion strategy of multimodal features; the graph convolution mechanism of introducing edge type factor and spatial attention factor enhances the model's ability to learn geological spatial correlation features; the hierarchical feature aggregation module realizes multi-level and multi-scale feature extraction through local, cross-scale and global multi-level aggregation strategies, which can capture local geological details and grasp the global risk distribution, thereby improving the precision and comprehensiveness of prediction.

[0023] Most importantly, this application deeply integrates physical constraints into the machine learning framework. By designing a loss function that incorporates hydrogeological and rock mechanics constraints and introducing a physical consistency verification mechanism during the inference stage, it ensures that the prediction results conform to both data patterns and physical laws. This dual guarantee mechanism of data-driven and physical guidance effectively overcomes the potential for physically unreasonable predictions from purely data-driven models. It can improve the accuracy and reliability of mine geological structure safety early warning, providing more scientific and effective technical support for coal mine safety production. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 System architecture diagram of the mine geological structure safety prediction system integrating multi-dimensional features provided in the embodiments of this application; Figure 2 A flowchart illustrating the mine geological structure safety prediction method integrating multi-dimensional features provided in this application embodiment; Figure 3 A flowchart illustrating another mine geological structure safety prediction method integrating multi-dimensional features provided in this application embodiment; Figure 4 A flowchart illustrating another mine geological structure safety prediction method integrating multi-dimensional features provided in this application embodiment; Figure 5 This is a flowchart illustrating another mine geological structure safety prediction method that integrates multi-dimensional features, provided as an embodiment of this application. Detailed Implementation

[0026] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0027] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0028] The mine geological structure safety prediction method integrating multi-dimensional features provided in this application embodiment can be applied to, for example... Figure 1 The mine geological structure safety prediction system shown integrates multi-dimensional features, such as Figure 1 As shown, the system includes: a data preprocessing module, a geological map analysis module, and a risk prediction module; among which, The data preprocessing module is used to collect mine geological structure data, hydrological monitoring data, rock mass mechanical parameter data and real-time monitoring data to obtain multi-source heterogeneous data, and to perform spatiotemporal alignment and standardization processing on the multi-source heterogeneous data to obtain preprocessed data; The geological map analysis module is used to abstract the geological area of ​​the mine into a graph structure based on the preprocessed data. The nodes of the graph structure represent geological entity units, and the edges represent the interaction relationships between geological entities, generating a geological map including an adjacency matrix and a node feature matrix. The risk prediction module is used to input geological maps into a geological map neural network model built based on deep learning algorithms. It aggregates node features through multi-layer convolution operations and performs physical consistency verification and optimization on the model output in combination with physical constraints to obtain the risk probability distribution, which is used to predict the safety of geological structures.

[0029] To address the technical problems of existing technologies that rely on a single type of data, struggle to depict complex spatial relationships and physical interactions between geological entities, and produce prediction results that easily violate physical principles, leading to insufficient reliability, this application provides a mine geological structure safety prediction method that integrates multi-dimensional features. This method includes: Collect geological structure data, hydrological monitoring data, rock mass mechanical parameter data and real-time monitoring data of the mine to obtain multi-source heterogeneous data, and perform spatiotemporal alignment and standardization processing on the multi-source heterogeneous data to obtain preprocessed data; Based on the preprocessed data, the geological area of ​​the mine is abstracted into a graph structure, where the nodes of the graph structure represent geological entity units and the edges represent the interaction relationships between geological entities, generating a geological map including an adjacency matrix and a node feature matrix; The geological map is input into a geological map neural network model built based on deep learning algorithms. The node features are aggregated through multi-layer convolution operations, and the physical consistency of the model output is checked and optimized in combination with physical constraints to obtain the risk probability distribution, which is used to achieve safe prediction of geological structures.

[0030] Based on this, the method achieves effective integration of multi-source data and accurate characterization of geological systems, while incorporating physical laws to ensure the rationality of predictions, thereby improving the comprehensiveness, accuracy and reliability of mine geological structure safety predictions.

[0031] The mine geological structure safety prediction method integrating multi-dimensional features provided in this application includes: S1. Collect geological structure data, hydrological monitoring data, rock mechanics parameter data and real-time monitoring data of the mine to obtain multi-source heterogeneous data, and perform spatiotemporal alignment and standardization processing on the multi-source heterogeneous data to obtain pre-processed data.

[0032] Among them, multi-source heterogeneous data includes geological structure data that can reflect the basic geological characteristics of the mine, hydrological monitoring data that reflects the hydrological environment, rock mechanics parameter data that characterizes the rock mechanics properties, and real-time monitoring data that captures real-time changes in working conditions. These data come from different sources and have different data types and formats.

[0033] In some implementations, spatiotemporal alignment and standardization methods can be adopted as follows: in the time dimension, linear interpolation or nearest neighbor interpolation can be used to unify data with different sampling frequencies; in the spatial dimension, inverse distance weighting or radial basis function can be used to achieve spatial mapping of discrete data; standardization can be carried out by using Z-score standardization, min-max normalization and other methods to process continuous data, and label coding and other methods to process categorical data.

[0034] It should be noted that this step is a fundamental step in safety prediction. Collecting multi-source heterogeneous data can comprehensively cover the key influencing factors of the mine's geological system, while spatiotemporal alignment and standardization processing solve the problems of inconsistent spatiotemporal scales, large differences in dimensions, and inconsistent formats of the data, thus avoiding data clutter from interfering with subsequent model building.

[0035] S2. Based on the preprocessed data, the geological area of ​​the mine is abstracted into a graph structure, where the nodes of the graph structure represent geological entity units and the edges represent the interaction relationships between geological entities, generating a geological map including an adjacency matrix and a node feature matrix.

[0036] Among them, the geological map is an abstract representation of the mine's geological system. It intuitively presents geological entity units and their interrelationships through the combination of nodes and edges. The adjacency matrix is ​​used to quantify the connection relationship between nodes, and the node feature matrix is ​​used to integrate the multi-source preprocessed data corresponding to each geological entity unit. Together, they provide structured input for the feature learning of the subsequent model.

[0037] In some implementations, the geological map construction process can divide the geological area of ​​the mine into several uniform or non-uniform geological entity units as nodes according to the scope and research accuracy; determine whether there is a connection relationship based on the spatial proximity relationship, geological structure correlation, physical transmission, etc. between geological entities to construct edges; integrate the multi-source preprocessed data corresponding to the nodes according to preset rules to form node feature vectors, and then construct node feature matrices; at the same time, construct adjacency matrices according to the edge connection situation.

[0038] It should be noted that abstracting the geological area of ​​a mine into a graph structure is the key to breaking through the bottleneck of traditional models in expressing complex geological relationships. The complex relationships such as spatial topological associations and physical interactions between geological entities in a geological system can be intuitively and accurately depicted through the nodes and edges of the graph structure, providing reasonable structural support for the model to capture the inherent laws of the geological system.

[0039] For example, a mining area of ​​a mine is divided into several geological entity units as nodes on a three-dimensional scale of 10m×10m×5m. If the geological entity units corresponding to two nodes are connected by faults, stress transfer, or hydraulic connection, then an edge is constructed between them. The preprocessed data such as fault parameters, water level data, rock mechanics parameters, and real-time monitoring data corresponding to each node are integrated into a multi-dimensional feature vector to form a node feature matrix. At the same time, 1 is used to indicate the existence of an edge between nodes and 0 is used to indicate the absence of an edge, thus constructing an adjacency matrix.

[0040] S3. Input the geological map into a geological map neural network model built based on deep learning algorithms. Aggregate node features through multi-layer convolution operations and combine physical constraints to perform physical consistency verification and optimization of the model output to obtain the risk probability distribution, which is used to achieve safe prediction of geological structures.

[0041] Among them, the geological map neural network model is a model designed based on deep learning algorithms and suitable for graph structure data. It can effectively aggregate node features and their associated features using convolution operations to uncover the complex patterns hidden in the geological system. It is used to learn features from structured geological map data and output risk-related prediction results.

[0042] In some implementations, physical constraints are requirements determined based on fundamental laws and criteria in fields such as hydrogeology and rock mechanics. These may include the continuity equation of water flow and seepage law constraints in hydrogeology, and the strength criteria and stress balance constraints in rock mechanics. The geological map neural network model can use common graph neural network models such as graph convolutional network (GCN) and graph attention network (GAT). Multi-layer convolution operations can be achieved by setting 2-5 convolutional layers to extract and aggregate features at different levels. Physical consistency verification and optimization can be achieved by incorporating physical constraints into the model loss function, or by setting a physical reasonable range threshold during the model inference stage to correct prediction results that exceed the range.

[0043] It should be noted that the risk probability distribution is used to reflect the likelihood of safety risks occurring in each geological entity unit within the mine's geological area, providing a precise basis for risk positioning and level assessment for coal mine safety production; while the incorporation of physical constraints effectively avoids prediction results that may occur in purely data-driven models that violate physical common sense, ensuring the reliability of the prediction.

[0044] For example, a graph convolutional neural network is selected as the geological map neural network model, with three convolutional layers. The first convolutional layer extracts local features of nodes, the second convolutional layer aggregates neighborhood correlation features, and the third convolutional layer captures global features. The physical constraints are the water flow continuity equation and the Mohr-Coulomb strength criterion, and their corresponding constraint loss terms are incorporated into the cross-entropy loss function to form the total loss function. The parameters are optimized during model training. The model outputs the probability of geological structural instability risk for each node, forming a risk probability distribution. The probability value is between 0 and 1, and the closer the probability is to 1, the higher the safety risk of the area.

[0045] Based on the above technical solutions, the mine geological structure safety prediction method integrating multi-dimensional features provided in this application ensures the comprehensiveness and reliability of the data foundation by collecting and preprocessing multi-source heterogeneous data; it achieves accurate representation of complex relationships between geological entities by abstracting geological regions into graph structures; and it fully leverages the feature learning capabilities of deep learning while ensuring that the prediction results conform to physical laws through feature aggregation of the geological map neural network model and fusion of physical constraints. This method effectively improves the accuracy, comprehensiveness, and reliability of mine geological structure safety prediction, providing scientific and effective technical support for coal mine safety production.

[0046] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 3As shown, the above S1 can be implemented through the following S101, S102 and S103, which are explained in detail below: S101. Collect mine geological structure data, hydrological monitoring data, rock mechanics parameter data, and real-time monitoring data. Add metadata tags to each data point and integrate them into a unified platform through a data bus to obtain raw multi-source heterogeneous data.

[0047] The geological structure data was acquired through drilling core sampling and geophysical exploration equipment, including parameters such as fault strike, fold axis location, and joint density, with each data point recording three-dimensional coordinates and a timestamp. Hydrological monitoring data was collected through a sensor network deployed underground, including aquifer permeability coefficients, dynamic water level changes, and mine water inflow. Rock mechanics parameter data was collected using in-situ testing instruments, covering indicators such as rock layer density, compressive strength, and geostress distribution. Real-time monitoring data was collected by microseismometers and gas sensors, primarily including microseismic time series and gas concentration data. Metadata tags were then added to the collected data, including information such as data source, collection time, and spatial location, ensuring data traceability.

[0048] In some implementation methods, geological structure data acquisition can utilize a combination of high-density seismic and electromagnetic exploration equipment to improve the accuracy of fault and fold identification; the hydrological monitoring sensor network can adopt a distributed deployment scheme, dividing monitoring areas according to the distribution characteristics of aquifers in the mine, with 3-5 monitoring points and several auxiliary monitoring points set up in each area; rock mass mechanical parameter acquisition can combine borehole penetration and acoustic testing technologies to simultaneously obtain multiple mechanical indicators; real-time monitoring equipment can use wireless self-organizing network technology to achieve real-time data transmission and breakpoint resume. Metadata tag generation can be automatically completed by the embedded module built into the acquisition equipment, and the data bus uses an industrial Ethernet bus, supporting multi-protocol data access.

[0049] It should be noted that the comprehensive collection of multi-source heterogeneous data is the foundation for subsequent safety prediction. Different types of data correspond to the structural characteristics of the mine geological system, hydrological environment, rock mechanical properties and real-time operating conditions, respectively. The coordinated collection of the four types of data can fully cover the key influencing factors of geological safety risks.

[0050] For example, a kilometer-deep drilling rig was used to collect core samples, and geological structure data of the No. 3 coal seam in a certain mine was collected using a two-dimensional seismic exploration device. The data recorded the fault strike of 315°, the three-dimensional coordinates of the fold axis (X=2568m, Y=1342m, Z=-890m), and the joint density of 2.3 joints / m. The timestamp was marked as 2024-05-20 14:30:00. Twelve hydrological monitoring sensors were deployed in the main aquifer area of ​​the mine, collecting water level data every 15 minutes and uploading the aquifer permeability coefficient of 0.03m / d and the mine water inflow of 25m³ / h in real time. The rock density of the rock layer was collected as 2.65g / cm³, the compressive strength as 68MPa, and the maximum principal stress as 28MPa through the in-hole rock mechanics testing system. The number of microseismic events was recorded once per hour using a microseismometer, and the gas concentration data was collected in real time by a gas sensor and converted into a time series. All data is automatically tagged with metadata, with data sources marked as "seismic exploration equipment" and "hydrological sensor-1# monitoring point", and integrated into the mine data center unified platform via Profinet industrial bus.

[0051] S102. Perform anomaly detection and cleaning on the original multi-source heterogeneous data, achieve time alignment of data with different sampling frequencies through time interpolation, and then map discrete data points to preset grid cells through spatial interpolation to obtain spatiotemporally aligned data.

[0052] Among them, anomaly detection and cleaning is to remove outliers caused by equipment failure or environmental interference during the acquisition process; time alignment unifies data from different sampling frequencies to the same time interval to ensure data consistency in the time dimension; and spatial interpolation converts scattered discrete data points into continuous grid cell data to achieve spatial structuring of the data.

[0053] In some implementations, S102 may include the following steps: Anomaly detection can employ an improved sliding window anomaly detection algorithm. The window size is dynamically adjusted according to the data type. For example, the window length for geological structural data is set to 10-20 data points, the window length for hydrological monitoring data is set to 5-10 sampling periods, and the window length for real-time monitoring data is set to 3-5 sampling periods. Then, by calculating the mean and standard deviation of the data within the window, data that deviate from the mean by more than 3 times the standard deviation are identified as outliers and removed.

[0054] In addition to cubic spline interpolation, piecewise Hermite interpolation can also be used for time interpolation to improve the interpolation accuracy for data with strong nonlinearity, such as dynamic changes in water level. The determination of the time interval needs to be combined with the timeliness of the data and the processing efficiency. For example, geological structure data is uniformly set to one time point per day, hydrological monitoring data is uniformly set to one time point per hour, and real-time monitoring data is uniformly set to one time point every 10 minutes.

[0055] Spatial interpolation can employ co-kriging interpolation, introducing geological zoning categories as auxiliary variables to improve the spatial correlation of interpolated data for special geological regions such as fault zones and fold zones.

[0056] S103. Standardize the spatiotemporally aligned data, add timestamps and spatial coordinate labels to the processed data, and generate standardized feature data and corresponding metadata index tables.

[0057] In some implementations, normalization of continuous data can employ an improved min-max normalization method. This method dynamically adjusts the normalization interval to adapt to the distribution characteristics of different data. For example, the normalization interval for geological structural data is set to [0.1, 0.9] to avoid feature suppression caused by extreme values; the normalization interval for hydrological monitoring and rock mechanics parameter data is set to [0.2, 0.8] to balance feature contributions. For categorical data with multi-level classifications, an embedded encoding method can be used to embed category information into a low-dimensional vector space, improving the model's feature learning efficiency. Timestamp labels use UTC time format, accurate to the second, and spatial coordinate labels use a mine-independent coordinate system, accurate to the centimeter level. The metadata index table uses a distributed storage structure, with index fields including data ID, metadata tag keywords, grid cell number, and standardized feature data storage address, supporting multi-dimensional retrieval by data type, spatiotemporal range, data source, and other dimensions.

[0058] It is important to note that standardization is a crucial step in eliminating data heterogeneity, preventing the model from overemphasizing certain features due to differences in data units and numerical ranges, and ensuring that all features have equal weight in model training. The addition of timestamps and spatial coordinate labels is fundamental to achieving accurate matching between data and geological map nodes, while the metadata index table improves data management efficiency and facilitates subsequent data retrieval, updates, and maintenance.

[0059] Based on the above technical solutions, S1 achieves the acquisition and preprocessing of multi-source heterogeneous data through multi-source data collaborative acquisition and integration in S101, spatiotemporal alignment processing in S102, and standardization and metadata management in S103. This not only solves the problems of single data, spatiotemporal inconsistency, and uneven quality in traditional data processing, but also provides high-quality and highly available structured data support for subsequent geological map construction and risk prediction models through refined acquisition strategies, efficient cleaning and interpolation methods, and standardized management, ensuring the reliability of the data foundation of the entire prediction method.

[0060] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S2 can be implemented through the following S201, S202 and S203, which are explained in detail below: S201. Divide the geological area of ​​the mine into multiple grid units. Each grid unit serves as a node on the geological map. Construct the node feature vector corresponding to each node. At the same time, generate a geological partition mask based on the geological exploration results and expert annotations, and establish a one-to-one correspondence between the mask and the node.

[0061] Among them, the grid cell is a structured mapping of the geological entity unit of the mine. When dividing it, the complexity of the geological structure and the computational efficiency of the model must be taken into account to ensure that each cell can accurately carry the geological characteristic information of the corresponding area. The node feature vector is the fusion carrier of multi-source preprocessed data and is used to reflect the core indicators of the geological foundation, hydrological environment and rock mechanical properties of the area where the node is located. The geological partition mask is the identification information used to mark the geological partition category to which each grid cell belongs, providing a partition basis for subsequent hierarchical feature aggregation, ensuring that the model can extract the features of different geological partitions in a targeted manner. Its category is consistent with the main geological structure type of the mine, including core categories such as fault zones, fold zones and normal strata.

[0062] In some implementations, an adaptive grid partitioning strategy can be adopted, adjusting the grid density based on the preliminary geological zoning results. For example, a fine grid of 5m×5m×3m is used in geologically complex areas such as fault zones and fold zones, while a coarse grid of 10m×10m×5m is used in areas with normal rock strata. During the partitioning process, it is necessary to ensure that the grid cells do not overlap, have no omissions, and completely cover the entire geological prediction area of ​​the mine. The construction of node feature vectors needs to be sorted according to the priority of data modes, and then the geological structure feature subset, hydrological monitoring feature subset, and rock mass mechanical parameter feature subset are spliced ​​in sequence. After splicing, the feature dimensions are uniformly processed (such as zero-filling or dimension compression) to ensure that the dimensions of all node feature vectors are consistent.

[0063] In some implementations, the generation of geological zoning masks requires combining high-precision geological exploration reports, borehole core analysis results, and the experience of geological experts. First, the boundaries of geological zoning in each region are determined using technologies such as ground-penetrating radar and seismic exploration. Then, each grid cell is labeled with a zoning category. The encoding method can be integer encoding or one-hot encoding. If integer encoding is used, 0 represents normal rock strata, 1 represents fault zones, and 2 represents folded areas. If one-hot encoding is used, the fault zones correspond to vectors [1,0,0], folded areas correspond to [0,1,0], and normal rock strata correspond to [0,0,1]. After the mask is generated, it needs to be associated with the grid cell number to form a mapping table of "node number - zoning category - mask vector".

[0064] It should be noted that the geological partition mask is a prerequisite for achieving hierarchical feature extraction, and its accuracy directly affects the subsequent model's ability to capture features of different geological structural regions. During the annotation process, cross-validation with multi-source geological data is required to avoid blurred partition boundaries or misclassification of categories. Grid cell division and geological partition mask generation should be carried out simultaneously to ensure that fine grid areas are accurately matched with high-priority geological partitions (such as fault zones) and to maximize the preservation of geological details in key areas. In addition, node feature vectors should be stored in association with the geological partition mask.

[0065] For example, the predicted area of ​​a certain mine is X∈[2000m,3000m], Y∈[1000m,2000m], Z∈[-1000m,-500m]. Seismic exploration determines the fault zone boundaries to be X∈[2400m,2600m], Y∈[1400m,1600m]. This area is divided using a fine 5m×5m×3m grid, resulting in 8000 nodes. The remaining normal rock strata area is divided using a coarse 10m×10m×5m grid, resulting in 5000 nodes. There are 13,000 nodes in the whole region. The feature vector of each node is spliced ​​in the order of “geological structure features (12-dimensional) + hydrological monitoring features (8-dimensional) + rock mass mechanical parameter features (10-dimensional)” and unified into 32-dimensional through zero filling. The geological partition mask adopts integer encoding. The mask value of fault zone node is 1, the mask value of normal rock layer node is 0, and the mask value of a certain fault zone node is 1. Its corresponding mapping table is recorded as “node number G00123-partition category fault zone-mask vector[1]”.

[0066] S202. Based on the spatial relationships, physical interactions, and geological zoning masks between geological entities, construct the edges of the geological map and define the edge types. Calculate the edge weights based on the Euclidean distance and hydraulic pressure difference between nodes.

[0067] Among them, edges are a concrete representation of the relationships between geological entities. Spatial relationships include the proximity between nodes. Physical interactions mainly refer to hydraulic connectivity and stress transfer effects. Geological partitioning masks are used to help determine the priority of edge types. Nodes within the same partition are more likely to form stress transfer edges, while nodes in different partitions but with hydraulic connectivity paths form hydraulic connectivity edges. Edge types are divided into hydraulic connectivity edges and stress transfer edges, which correspond to two types of physical interactions, respectively. Edge weights are used to quantify the strength of interactions between nodes. The larger the value, the closer the association and the more significant the mutual influence.

[0068] In some implementations, the rules for edge construction are as follows: An edge connection is constructed when the spatial distance between the grid cells corresponding to two nodes is less than a preset threshold (this threshold is determined based on the average porosity of mine geology and the stress transmission attenuation law, and is usually set to 50m), and any of the following conditions are met: First, based on the geological zoning mask, it is determined that two nodes belong to the same fault zone or fold zone, and the difference in geostress distribution is less than 10MPa, thus forming a stress transfer edge. Secondly, although the two nodes may belong to different zones, the product of the aquifer permeability coefficients is greater than the preset permeability threshold (taken as 0.3 times the average permeability coefficient of the region), and the water level difference is greater than 0.5m, forming a hydraulic connection edge.

[0069] Edge weights are expressed by the formula The calculation uses this formula to comprehensively consider the spatial distance attenuation effect and the influence of hydraulic pressure difference, accurately quantifying the overall interaction strength between nodes. Among these, Let be the edge weight between node i and node j, with a value range of (0,1]. The Euclidean distance between node i and node j is calculated using the 3D coordinates of the center points of the two node grid cells; σ is a scale parameter used to adjust the influence of spatial distance on edge weights, determined by statistically analyzing the mean and standard deviation of the Euclidean distances between all nodes in the mine, and the calculation formula is as follows. ,in Let be the mean of the Euclidean distances between all nodes. Standard deviation; The hydraulic pressure difference based on the permeability coefficient is calculated using the following formula: ,in and The water level heights at nodes i and j are respectively. and are the aquifer permeability coefficients for nodes i and j, respectively.

[0070] It should be noted that the introduction of geological zoning masks makes the judgment of edge types more targeted, ensuring that stress transmission edges are mainly concentrated in geologically active zones, and hydraulic connectivity edges accurately reflect the hydraulic connections across or within zones; edge types directly reflect the core physical interactions between geological entities, enabling geological maps to accurately depict the transmission paths of key physical effects such as hydraulics and stress; edge weights integrate spatial distance and physical parameters, ensuring that the weight values ​​of edges can truly reflect the strength of mutual influence between nodes, providing a quantitative basis for the model to learn the dynamic correlation laws of geological systems.

[0071] For example, the Euclidean distance between node A (node ​​number G00123, 3D coordinates (2500m, 1500m, -800m), geological zone mask value 1, belonging to a fault zone) and node B (node ​​number G00124, 3D coordinates (2505m, 1503m, -802m), geological zone mask value 1, belonging to a fault zone) is... The mean Euclidean distance between all nodes in the mine =45m, standard deviation =12m, then the scale parameter Water level at node A =120m, permeability coefficient =0.05m / d, water level height at node B =118m, permeability coefficient =0.04m / d, then the hydraulic pressure difference =|120×0.05-118×0.04|=|6-4.72|=1.28; Substituting into the edge weight formula, we get... Since both nodes belong to fault zones based on the geological zoning mask, and the stress difference is 8MPa (less than 10MPa), this edge is defined as a stress transfer edge.

[0072] S203. Based on node feature vectors, geological partition masks, edge information, and edge weights, construct node feature matrices, adjacency matrices, and edge type matrices, and store the topological relationships and mask information of the geological map through a graph database.

[0073] The node feature matrix is ​​a set of feature vectors representing all nodes, with a dimension of "number of nodes × feature vector dimension". Each row corresponds to the feature vector of one node, used to present the distribution of geological features across the entire region. The adjacency matrix is ​​a matrix mapping of edge information and edge weights, with a dimension of "number of nodes × number of nodes". The edge weight between node i and node j If the two nodes are not connected, then =0; The edge type matrix has the same dimension as the adjacency matrix, with elements having a value of 1 to represent a hydraulically connected edge, 2 to represent a stress-transfer edge, and 0 to represent no connection, corresponding one-to-one with the non-zero elements of the adjacency matrix; The geological partition mask is stored in vector form with a dimension of "number of nodes × 1", with each element corresponding to the partition code value of a node; The graph database is used to uniformly store node, edge, edge type, matrix information, and geological partition mask, supporting efficient topology query and data update.

[0074] In some implementations, the construction of the node feature matrix requires sequentially arranging the node feature vectors according to node number (in ascending order of grid cell XYZ coordinates). After construction, matrix normalization is performed to enhance the numerical stability of the matrix. The adjacency matrix adopts a sparse matrix storage format (such as CSR format), storing only non-zero elements, that is, only storing the weight values ​​of node pairs connected by edges, reducing storage overhead. The geological partition mask vector is arranged in node number order and strictly corresponds to the row index of the node feature matrix and adjacency matrix. If one-hot encoding is used, it is expanded into a matrix form of "number of nodes × number of partition categories". A distributed graph database is selected for the graph database. During storage, node number, 3D coordinates, partition category, and mask value are used as node attributes, and edge weight and edge type are used as edge attributes. The node feature matrix, adjacency matrix, edge type matrix, and mask vector are stored together in binary file format. An index mapping is established through node number, supporting fast querying of related information by node number, partition category, edge type, and other conditions.

[0075] Based on the above technical solution, S2, through node partitioning, feature vector construction and geological partition mask generation in S201, edge construction and weight calculation in S202, and matrix construction and topology and mask information storage in S203, completes the transformation of the mine geological system into a structured geological map containing partition information. This process not only achieves the structured integration of multi-source preprocessed data, but also provides a clear basis for distinguishing features of different geological structural regions through the introduction of geological partition masks. Combined with edge type definition and weight calculation, it characterizes the spatial topological relationships, physical interactions and partition attributes between geological entities, solving the technical problem that traditional models are unable to express complex geological relationships and distinguish partition features. The constructed geological map retains detailed information and partition attributes of the geological features of the entire region, and ensures efficient calling of subsequent models through matrixing and database storage, laying a solid structured foundation for hierarchical feature aggregation and risk prediction of the geological map neural network model.

[0076] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S3 can be implemented through the following S301, S302 and S303, which are explained in detail below: S301. Input the node feature matrix, adjacency matrix, edge type matrix and geological partition mask of the geological map into the geological map neural network model. The input module completes the data verification and optimization, and the feature adaptation module realizes multi-modal feature fusion to obtain the node feature matrix after multi-source feature adaptation.

[0077] The geological map neural network model is built on a Graph Attention Network (GAT) and includes an input module, a feature adaptation module, an attention graph convolution module, a hierarchical feature aggregation module, a regularization module, and an output optimization module. It is used to extract deep correlation features from structured geological maps and map them to risk probabilities. The input module is used for data preprocessing to ensure that the input matrix format is compliant and the information is effective. The feature adaptation module is used to solve the problem of collaborative utilization of multi-source heterogeneous data and to explore the complementary value of different modalities of data through divide-and-conquer and fusion strategies.

[0078] In some implementations, the operation flow of the input module includes: A feature dimension unification algorithm is used for dimension calibration, adjusting the column dimensions of the node feature matrix to the fixed dimensions preset by the model. If the original dimension is lower than the preset value, it is filled with zeros; if it is higher than the preset value, it is reduced by principal component analysis. Then, redundant edges are filtered out based on the edge weight threshold. The weight threshold is set to 0.1 times the average weight of all edges, and edges with weights lower than the threshold are removed. At the same time, the adjacency matrix and edge type matrix are updated.

[0079] In some implementations, the operation flow of the feature adaptation module includes: First, the verified node feature matrix is ​​split into geological feature sub-matrix, hydrological feature sub-matrix, rock mechanics feature sub-matrix and real-time monitoring feature sub-matrix according to data mode; Subsequently, each submatrix is ​​processed by four independent one-dimensional convolutional layers (Conv1D). The kernel size of each Conv1D is set to 3×1, the stride is 1, and the padding mode is "SAME", so as to achieve feature extraction and dimensionality unification and obtain the modality-coded features. Introducing trainable weight vectors to adaptively weight and fuse the coded features of each modality, the fusion formula is as follows: ,in For multimodal fusion features, These are the trainable weights for the k-th modality, and their initial values ​​are all set to 0.25, which are updated iteratively during model training. Let be the feature submatrix of the k-th mode, and Conv1D(·) be a one-dimensional convolution operation; Finally, residual connections are constructed, and the multimodal fusion features are superimposed with the verified node feature matrix to offset the information loss during the feature encoding process, and the node feature matrix after multi-source feature adaptation is output.

[0080] It should be noted that this step is the foundation of model feature learning. Redundant edge removal in the input module can reduce the interference of invalid associations on the model and improve computational efficiency. The divide-and-conquer strategy of the feature adaptation module enables targeted processing of data from different modalities, while weighted fusion and residual connection ensure the full utilization of multi-source information and the preservation of original information.

[0081] S302. Enhance feature representation capabilities through attention map convolution module, and extract multi-scale, cross-regional features by combining geological zoning mask with hierarchical feature aggregation module to obtain risk embedding vector and node aggregation features.

[0082] Among them, the attention map convolution module introduces edge type factor and spatial attention factor to enhance the model's ability to capture geological spatial correlation features; the hierarchical feature aggregation module realizes multi-level feature extraction of local, cross-scale and global based on geological partition mask, which not only preserves local geological details, but also grasps the global risk distribution; the risk embedding vector is a condensed expression of global risk features, while the node aggregation feature retains the refined feature information of each node.

[0083] In some implementations, the operation flow of the attention map convolution module includes: The first step involves computation using a first-layer graph attention convolution that incorporates an edge type factor. This layer is configured with 8 attention heads, and the attention weights are calculated using the following formula: ,in Let be the attention weight coefficient of node i to node j, a and W be the trainable parameters of the model, N(i) be the set of neighboring nodes of node i, and LeakyReLU() be the activation function. , , These are the adapted feature vectors for nodes i, j, and k, respectively. , The edge type factor is set to 0.6 for hydraulic connectivity edges and 0.4 for stress transfer edges, and is updated during training iterations. After convolution, the output undergoes LeakyReLU activation to obtain the first-layer convolution feature matrix. Then, it is superimposed with the adapted node feature matrix through residual connections to obtain the intermediate feature matrix.

[0084] The second step is to calculate the spatial attention factor based on the node's three-dimensional coordinates. The calculation formula is , where λ is the adjustment coefficient (initial value is set to 5). Let be the Euclidean distance between nodes i and j; the spatial attention factor and the intermediate feature matrix are weighted to obtain the spatial association features.

[0085] The third step involves using the second-layer graph attention convolution to process the spatial correlation features. This layer is configured with four attention heads, and the attention weights are calculated using the following formula: ,in These are the attention weight coefficients for the second layer. , , This is the node vector corresponding to the spatial association feature; the other parameters are the same as those in the first layer. The convolutional output is then subjected to average pooling to obtain the enhanced feature matrix.

[0086] In some implementations, the operation flow of the hierarchical feature aggregation module is as follows: First, based on the geological zoning mask, the enhanced feature matrix is ​​split into local sub-map feature matrices corresponding to fault zones, fold zones, and normal strata. Each local sub-map feature matrix undergoes graph convolution to extract geological correlation features within the zoning. Then, average pooling is used to compress the dimensions, generating local feature vectors for each zoning. The formula is as follows: , where m is the geological zone category (1 is the fault zone, 2 is the fold zone, and 3 is the normal rock strata). For partition m, the enhanced feature matrix is... Let m be the adjacency matrix of partition m, and let AvgPool() be the average pooling operation. Then, multi-scale pooling is performed on the local feature vectors of each partition, with the pooling kernels being: 1×1 for the first scale, 3×3 for the second scale, and 5×5 for the third scale, to obtain the multi-scale pooling results. After concatenation, the cross-scale feature vector is obtained. ; By concatenating cross-scale feature vectors with local feature vectors through residual connections, the fused features of each partition are obtained. Then, convolution operations are performed on the fusion features of different partitions and concatenated to obtain the cross-partition fusion feature vector. ; The cross-partition fused feature vector is input into the fully connected layer, through the weight matrix. and bias terms Adaptive weighted fusion is performed to obtain a global aggregated feature vector. ; Finally, the global aggregated feature vector is processed in two branches. One branch obtains the risk embedding vector through global average pooling. Another branch retains the node dimensional features to obtain the node aggregation features. .

[0087] S303. The regularization module reduces the risk of model overfitting, and the output results are validated and optimized in conjunction with physical constraints to obtain the node-level risk probability distribution and the full-scale risk probability distribution.

[0088] The regularization module is used for noise reduction and overfit suppression; the physical constraints are defined based on hydrogeological equations and rock mechanics criteria, and are divided into loss constraints in the training phase and output verification in the inference phase; the node-level risk probability distribution reflects the safety risk probability of each grid unit, and the full-scale risk probability distribution reflects the overall risk level of the entire mine area.

[0089] In some implementations, the regularization module uses a strategy combining L2 regularization and Dropout to process risk embedding vectors and node aggregation features. The Dropout probability is set to 0.15 to randomly deactivate some neurons to avoid overfitting. L2 regularization constrains the model parameter size by adding a weight decay term to the total loss function.

[0090] In some implementations, physical constraint optimization is divided into two stages: During the training phase, physical constraints are incorporated as a soft loss term into the total loss function, and the formula for the total loss function is as follows: ,in For the total loss, (Set to 0.2) (Set to 0.2) is the weighting coefficient and the sum is less than 0.5; For task loss, the cross-entropy loss function is used to calculate the deviation between the predicted risk and the actual label; To account for hydrological constraint losses and ensure that the prediction conforms to the seepage law, the formula is: Where N is the number of hydrological monitoring data samples, Let be the aquifer permeability coefficient of the i-th grid cell. For water level gradient; This represents the rock mass mechanical constraint loss, used to ensure that the prediction meets the strength criterion. The formula is: , For rock mass shear stress, For cohesion, It is normal stress. (where is the internal friction angle); during model training, the total loss is calculated and all parameters are updated through backpropagation.

[0091] During the inference phase, reasonable ranges for hydrogeology and rock mechanics are preset. For example, the reasonable range for water level gradient is 0-0.05, and the reasonable range for the ratio of rock mass shear stress to shear strength is 0-1. Then, the initial risk probability output by the model is verified. If the water level gradient corresponding to the risk probability of a certain grid cell exceeds the reasonable range, or the shear stress ratio is greater than 1, a correction rule is triggered—the probability value is adjusted according to the degree of deviation. When the deviation is within 10% of the upper limit of the reasonable range, the probability is corrected to the maximum probability value corresponding to the reasonable range. When the deviation exceeds 10%, the probability is remapped in combination with physical constraint thresholds (such as setting the hydrological constraint threshold to 0.05 and the mechanical constraint threshold to 1).

[0092] Finally, the output optimization module uses the Softmax function to perform probability mapping on the regularized risk embedding vector and node aggregation features to obtain the node-level risk probability distribution (dimension is the number of nodes × 1) and the full-scale risk probability distribution (dimension is 1 × 3, corresponding to the probability proportion of low, medium and high risk levels).

[0093] It should be noted that the regularization module can effectively improve the model's generalization ability and avoid overfitting caused by high feature dimensionality and complex data. The integration of physical constraints is key to solving the problem of physics-inappropriate pure data-driven models. Loss constraints during training ensure that the model's learning process follows physical laws, while verification and correction during inference further guarantee the reliability of the output. Its parameters (such as...) , All of these data (including those from S1) are derived from preprocessed data collected by S1, achieving a deep integration of data with physical laws. The two types of risk probability distributions output in the end not only meet the needs of accurate local early warning, but also support global security decision-making.

[0094] In some implementations, safety early warning of mine geological structures based on node-level risk probability distribution and full-scale risk probability distribution may include the following steps: First, based on historical mine safety accident data, geological structure complexity ratings, and industry safety standards, the grading thresholds for node-level and full-scale risks are determined. For example, node-level risk probabilities can be divided into four levels according to intervals: low risk (0-0.3), medium risk (0.3-0.6), high risk (0.6-0.8), and extremely high risk (0.8-1.0). The thresholds need to be determined by statistically analyzing the mean and standard deviation of risk probabilities corresponding to historical accidents; for example, the mean plus one standard deviation can be used as the starting threshold for high risk. Full-scale risks can be divided into three levels according to the proportion of high risks: low global risk (0-0.1), medium global risk (0.1-0.3), and high global risk (>0.3). The proportion thresholds are set with reference to the mine's annual safety targets and emergency response capabilities.

[0095] Then, the node-level risk probability distribution is traversed to filter out nodes with high and extremely high risk levels. Combining the 3D coordinates of the nodes with geological zoning masks, the spatial clustering regions of high-risk nodes are determined. The degree of clustering of high-risk nodes is analyzed using density clustering algorithms (such as DBSCAN). The clustering radius is set to three times the side length of the grid cell, and the minimum number of clustered nodes is five, forming high-risk clustering areas. Simultaneously, the geological zoning information of this area is correlated, marking high-risk clusters within special structures such as fault zones and fold zones, clarifying the geological background of local risks.

[0096] Next, based on the full-scale risk probability distribution, the weights of low, medium, and high risk levels are calculated, and the assessment weights are adjusted according to the mine production stage (such as the development phase and the mining phase). If the proportion of high risk in the mining phase is ≥0.25 (0.05 lower than the threshold in the development phase), the overall risk status is determined to be "requiring key monitoring"; if the proportion of high risk is ≥0.3 and the proportion of medium risk is ≥0.4, it is determined to be "overall early warning status"; at the same time, the dynamic change trend of each risk level is analyzed. If the proportion of high risk increases by more than 50% within 72 hours, the dynamic early warning mechanism is triggered.

[0097] Simultaneously, a two-dimensional early warning matrix of "local risk intensity - global risk status" can be established. Local risk intensity is quantified by the area and number of nodes of high-risk clusters (area accounting for 10% of the predicted area or number of nodes ≥500 is considered high intensity). Global risk status is divided into three levels: normal, key monitoring, and early warning. The two are combined to form a four-level early warning: blue warning (local low intensity + global normal), yellow warning (local high intensity + global normal or local low intensity + global key monitoring), orange warning (local high intensity + global key monitoring or local low intensity + global early warning), and red warning (local high intensity + global early warning).

[0098] Finally, the warning information is output, including: first, the three-dimensional coordinate range, geological zoning attributes, risk level and corresponding physical parameter anomalies of local high-risk areas, such as water level gradient exceeding the standard and ground stress concentration; second, the proportion of global risk levels, dynamic change trends and key influencing factors, such as fluctuations in hydrological monitoring data and stress transmission in fault zones; and third, targeted handling suggestions, such as routine inspections in low-risk areas, intensified monitoring in medium-risk areas, suspension of surrounding operations and initiation of special exploration in high-risk areas, and immediate evacuation of personnel and implementation of prevention and control measures such as grouting reinforcement in extremely high-risk areas.

[0099] Based on the above technical solutions, S3 utilizes input verification and multimodal feature adaptation in S301, attention map convolution enhancement and hierarchical feature aggregation in S302, and regularization and physical constraint optimization in S303. Through the modular design of the attention map convolutional network, it leverages the feature extraction capabilities of deep learning while ensuring the rationality of the prediction results through physical constraints. The final output risk probability distribution combines local accuracy with global comprehensiveness, effectively solving the technical problems of traditional methods' inability to capture complex geological relationships and insufficient prediction reliability. This provides a scientific and accurate decision-making basis for mine geological structure safety early warning.

[0100] Working principle of the invention: First, comprehensive multi-source heterogeneous data related to mine geological structure, hydrological monitoring, rock mechanics parameters and real-time monitoring are collected. Through anomaly detection, spatiotemporal alignment and standardization, the data format is regulated and the data quality is improved. At the same time, geological zoning masks are generated by combining geological exploration results and expert annotations to clarify the geological category attributes of each region. Subsequently, the geological area of ​​the mine is divided into several grid units as nodes. The pre-processed multi-source data is integrated to construct node feature vectors. Based on the spatial relationship, physical interaction and geological partition mask between geological entities, hydraulic connectivity edges and stress transmission edges are defined and edge weights are quantified. Then, a geological map containing node feature matrix, adjacency matrix, edge type matrix and geological partition mask is constructed to realize the structured representation of the mine geological system. Next, geological map data is input into a geological map neural network model built on an attention map convolutional network. The input module optimizes the data, the feature adaptation module fuses multimodal features, and the attention map convolutional module enhances feature representation. Combined with geological partitioning masks, the hierarchical feature aggregation module extracts local, cross-scale, and global features to obtain risk embedding vectors and node aggregation features. The regularization module reduces the risk of overfitting, while incorporating physical constraints related to hydrogeology and rock mechanics. The model learning process is constrained during the training phase, and the output results are verified and corrected during the inference phase. Finally, node-level and full-scale risk probability distributions are obtained. Finally, based on the probability distribution of the two types of risks, the system outputs early warning information and targeted disposal suggestions by calibrating risk thresholds, locating local high-risk areas, assessing the overall risk status, determining the comprehensive early warning level, and realizing the safety prediction and early warning of the mine's geological structure.

[0101] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0102] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of this application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and modifications.

Claims

1. A method for predicting the safety of mine geological structures by integrating multi-dimensional features, characterized in that, include: Collect geological structure data, hydrological monitoring data, rock mass mechanical parameter data, and real-time monitoring data from the mine to obtain multi-source heterogeneous data. Then, perform spatiotemporal alignment and standardization on the multi-source heterogeneous data to obtain pre-processed data. Based on the preprocessed data, the geological area of ​​the mine is abstracted into a graph structure, where the nodes of the graph structure represent geological entity units and the edges represent the interaction relationships between geological entities, generating a geological map including an adjacency matrix and a node feature matrix. The geological map is input into a geological map neural network model built based on a deep learning algorithm. The node features are aggregated through multi-layer convolution operations, and the physical consistency of the model output is verified and optimized in combination with physical constraints to obtain the risk probability distribution. The risk probability distribution is used to achieve safe prediction of geological structures.

2. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 1, characterized in that, The acquisition of the multi-source heterogeneous data includes: Geological structural data were collected through drilling core sampling and geophysical exploration equipment; the geological structural data included fault strike, fold axis location, joint density parameters, and recorded the three-dimensional coordinates and timestamp of each data point; Hydrological monitoring data is collected by deploying an underground sensor network; the hydrological monitoring data includes aquifer permeability coefficient, dynamic changes in water level, and mine water inflow. Rock mass mechanical parameter data are collected using in-situ testing instruments; the rock mass mechanical parameter data includes rock layer density, compressive strength, and in-situ stress distribution; Real-time monitoring data is acquired using microseismometers and gas sensors; the real-time monitoring data includes microseismic time series and gas concentration data. Metadata tags, including data source, collection time and location, are added to all data immediately after collection, and the data is integrated into a unified platform via a data bus.

3. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 1, characterized in that, The process of constructing the geological map includes: The mining area is divided into multiple grid units, each grid unit serving as a node on the geological map. The node feature vector is composed of preprocessed geological structure data, hydrological monitoring data, and rock mechanics parameter data. The edges of the geological map are constructed based on the spatial relationships and physical interactions between geological entities, and the edge types are set according to the hydraulic connectivity and stress transfer relationships; the edge types include hydraulic connectivity edges and stress transfer edges. Adjacency matrices and node feature matrices are constructed based on node and edge information, and topological relationships are stored in a graph database.

4. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 3, characterized in that, The geological map neural network model employs an attention map convolutional network, including an input module, a feature adaptation module, an attention map convolutional module, a hierarchical feature aggregation module, a regularization module, and an output optimization module; wherein... The input module is used to perform dimension calibration on the node feature matrix and remove redundant edges from the adjacency matrix to obtain the verified node feature matrix and the optimized adjacency matrix. The feature adaptation module is used to split, encode, and weightedly fuse the verified node matrix to obtain multimodal fusion features, and then superimpose the multimodal fusion features with the verified node matrix through residual connections to obtain a node feature matrix adapted from multi-source features. The attention graph convolution module is used to receive the adapted node feature matrix, the optimized adjacency matrix, and the edge type matrix. It introduces edge type factors and spatial attention factors to enhance feature extraction capabilities. Through two layers of graph attention convolution operations and activation processing, the enhanced feature matrix is ​​obtained. The edge type matrix is ​​constructed based on the edge types of the geological map and the hydraulic connectivity strength and stress transfer coefficient between corresponding nodes. It is obtained according to the recorded type representation and corresponding physical parameter values ​​of each edge. The hierarchical feature aggregation module is used to receive the enhanced feature matrix and the geological partition mask, and extract geological features at different scales through a multi-layer aggregation strategy to obtain risk embedding vectors and node aggregation features; wherein, the geological partition mask is obtained according to the geological partition category of each pre-labeled grid cell; The regularization module is used to perform regularization and noise reduction on the risk embedding vector and the node aggregation feature to obtain the regularized risk embedding vector and the regularized node aggregation feature. The output optimization module is used to perform probability mapping and calibration on the regularized risk embedding vector and the regularized node aggregation feature to obtain the node-level risk probability distribution and the full-scale risk probability distribution; wherein, the calibration corrects the risk probability that exceeds the reasonable range through physical constraints.

5. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 4, characterized in that, The internal workflow of the feature adaptation module is as follows: The verified node matrix is ​​split into geological feature sub-matrix, hydrological feature sub-matrix, rock mechanics feature sub-matrix and real-time monitoring feature sub-matrix according to data mode; By using four independent one-dimensional convolutional layers, feature extraction and dimensionality unification are performed on each feature submatrix to obtain the modality coding features. A trainable weight vector is introduced to adaptively weight and combine the modal coding features to obtain multimodal fusion features; wherein, the trainable weight vector is iteratively updated as the model is trained. A residual connection is constructed, and the multimodal fusion features are superimposed with the verified node matrix to offset the information loss during the feature encoding process, and the node feature matrix after multi-source feature adaptation is output.

6. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 5, characterized in that, The internal workflow of the attention map convolution module is as follows: Based on the adapted node feature matrix and the optimized adjacency matrix, a first-layer graph attention convolution operation with an edge type factor is performed, and the convolution output features are processed by the ReLU activation function to obtain the first-layer convolution feature matrix. Then, the first-layer convolution feature matrix is ​​superimposed with the adapted node feature matrix through residual connections to obtain an intermediate feature matrix; wherein, the edge type factor corresponding to the hydraulic connectivity edge is... The initial value is set to 0.6, which is the edge type factor for stress transfer edges. The initial value is set to 0.4, and Iterative updates are performed during model training; The spatial attention factor is calculated based on the three-dimensional coordinates of the nodes, and the spatial attention factor is weighted and calculated with the intermediate feature matrix to obtain the spatial association feature. The spatial correlation features are convolved using a second-layer graph attention convolution to obtain a second-layer convolution feature matrix. Then, an average pooling operation is performed on the second-layer convolution feature matrix to obtain an enhanced feature matrix.

7. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 4, characterized in that, The internal workflow of the hierarchical feature aggregation module is as follows: Based on the geological partitioning mask, the enhanced feature matrix is ​​split into local sub-graph feature matrices corresponding to fault zones, fold zones, and normal rock layers. Each local sub-graph feature matrix is ​​subjected to graph convolution operation to extract geological correlation features within the partition. Then, the dimension of the geological correlation features within the partition is compressed through average pooling operation to generate local feature vectors for each partition, thus obtaining a set of local feature vectors. Pooling operations of different scales are performed on the local feature vectors of each partition to obtain multi-scale pooling results including the first scale, the second scale, and the third scale; wherein the pooling kernel size of the first scale is 1×1, the pooling kernel size of the second scale is 3×3, and the pooling kernel size of the third scale is 5×5. The multi-scale pooling results of each partition are concatenated to obtain a cross-scale feature vector set. The cross-scale feature vector set is then concatenated with the local feature vector set through residual connection to obtain the fused features of each partition. Finally, the fused features of different scales are convolved and concatenated to obtain the cross-partition fused feature vector. The cross-region fusion feature vector is input into a fully connected layer for adaptive weighted fusion. The contribution of each cross-scale feature is adjusted by the parameters of the fully connected layer to obtain a global aggregated feature vector. The global aggregated feature vector is processed in two branches. One branch compresses the global aggregated feature vector into a fixed-dimensional vector through global average pooling to obtain the risk embedding vector. The other branch retains the feature information of the node dimension to obtain the node aggregated feature.

8. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 3, characterized in that, The physical consistency verification and optimization of the model output based on physical constraints includes: Physical constraints are defined based on hydrogeological equations and rock mechanics criteria. Physical constraints are incorporated as soft loss terms into the total loss function; After forward propagation of the model training, the physical constraint loss is calculated based on the total loss function, and the model parameters are optimized through backpropagation. During the model inference phase, the output probability of the model is physically consistent with the preset reasonable range of hydrogeology and rock mechanics. If the output probability is found to be inconsistent with the preset range, the probability is corrected according to the preset physical constraint threshold and correction rules.

9. The mine geological structure safety prediction method integrating multi-dimensional features according to claim 8, characterized in that, The formula for calculating the total loss function is: ;in, Indicates the total loss. , These are the weighting coefficients, and their sum is less than 0.5; The task loss is calculated using the cross-entropy loss function, which reflects the deviation between the predicted risk probability and the actual geological risk label. The hydrological constraint loss is used to constrain the prediction results to conform to the hydrogeological seepage law, and the calculation formula is: , The number of samples for hydrological monitoring data. Let be the aquifer permeability coefficient of the i-th grid cell. Let be the water level gradient of the i-th grid cell; The rock mass mechanical constraint loss is calculated using the following formula: , Let i be the rock mass shear stress of the i-th grid element. Let i be the rock mass cohesion of the i-th grid cell. Let be the rock mass normal stress of the i-th grid cell. Let be the internal friction angle of the rock mass in the i-th grid cell, where i represents the grid cell number.

10. A mine geological structure safety prediction system integrating multi-dimensional features, characterized in that, include: The module includes a data preprocessing module, a geological map analysis module, and a risk prediction module; among which, The data preprocessing module is used to collect mine geological structure data, hydrological monitoring data, rock mass mechanical parameter data and real-time monitoring data to obtain multi-source heterogeneous data, and to perform spatiotemporal alignment and standardization processing on the multi-source heterogeneous data to obtain preprocessed data. The geological map analysis module is used to abstract the geological area of ​​the mine into a graph structure based on the preprocessed data, wherein the nodes of the graph structure represent geological entity units and the edges represent the interaction relationships between geological entities, and generate a geological map including an adjacency matrix and a node feature matrix. The risk prediction module is used to input the geological map into a geological map neural network model built based on a deep learning algorithm, aggregate node features through multi-layer convolution operations, and perform physical consistency verification and optimization on the model output in combination with physical constraints to obtain the risk probability distribution; the risk probability distribution is used to realize the safety prediction of geological structures.

Images