A hidden disaster-causing factor risk determination method applied to a coal mine

Abstract

The application relates to the field of safety engineering and fault prediction technology, and discloses a hidden disaster-causing factor risk determination method applied to coal mines. The method comprises the following steps: acquiring geological, hydrological, gas, mining disturbance and real-time monitoring and other multi-source heterogeneous data; performing space-time alignment and normalization processing; constructing a disaster-causing factor correlation graph; extracting spatial topological features through a graph convolution network, combining a time series attention mechanism to capture dynamic evolution rules; adopting a gate fusion unit to adaptively fuse double-mode features to generate a comprehensive disaster-causing feature vector; outputting a four-grade risk determination result through a multilayer perception machine decoding, and triggering a visual early warning. The application realizes high-precision, interpretable and real-time intelligent risk determination by explicitly modeling the nonlinear coupling relationship between disaster-causing factors.

Description

Technical Field

[0001] This invention belongs to the field of safety engineering and fault prediction technology, specifically relating to a method for determining the hazard of hidden disaster-causing factors in coal mines. Background Technology

[0002] As coal mining depths increase and geological conditions become more complex in my country, the coupling and suddenness of hidden disaster-causing factors such as gas outbursts, mine water inrushes, and rock bursts are intensifying. The coal mine safety control system urgently needs to shift from passive response to proactive early warning. Hazard assessment, as a core component of risk prevention and control, directly impacts the timeliness and effectiveness of disaster prevention. Traditional assessment methods primarily rely on expert experience or static indicator systems, using techniques such as the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation for risk classification. While these methods have some applicability in shallow mining stages, they face challenges in complex deep mining conditions.

[0003] The assessment of the hazard of hidden disaster-causing factors focuses on the comprehensive evaluation of multi-source heterogeneous information such as gas occurrence state, hydrogeological structure, and geostress field distribution. This direction aims to scientifically determine the probability and severity of potential disasters by quantifying the activity level and interaction intensity of disaster-causing factors. However, existing technologies generally suffer from problems such as strong subjectivity in indicator weight allocation, weak data correlation mining capabilities, and poor model generalization performance, making it difficult to effectively characterize the nonlinear coupling mechanism and dynamic evolution law among disaster-causing factors.

[0004] Existing technologies typically employ fixed weights or single weighting methods to construct evaluation models, neglecting the fundamental differences in the causative mechanisms of different disaster types, resulting in a lack of specificity in the assessment results. Simultaneously, the deep features inherent in multi-source monitoring data are not fully extracted, causing key disaster-causing signals to be masked by noise. Furthermore, traditional two-dimensional evaluation frameworks only focus on risk probability and consequences, ignoring the "concealment" attribute unique to coal mine disasters, and failing to fully reflect the invisibility and suddenness of risks in deep mining environments. These shortcomings are particularly prominent in deep mines with high stress, high geothermal temperatures, and high gas content, restricting the accuracy of risk classification and the proactiveness of prevention and control measures. There is an urgent need to construct an intelligent hazard assessment system that integrates deep learning feature extraction, multi-model coupled judgment, and three-dimensional cloud assessment theory. Summary of the Invention

[0005] This invention provides a method for determining the hazard of hidden disaster-causing factors in coal mines, aiming to solve the problems of low accuracy in risk classification of disasters such as gas outbursts and water inrushes caused by reliance on human experience, strong subjectivity in weight allocation, and insufficient ability to model the coupling relationship of multi-source heterogeneous monitoring data in existing technologies. This invention constructs a multi-dimensional feature system integrating geological structure, hydrological conditions, gas occurrence state, mining disturbance effects, and real-time monitoring data. It introduces a cross-modal fusion model based on graph neural networks and attention mechanisms to achieve in-depth mining of the nonlinear and dynamic coupling relationship between hidden disaster-causing factors and generate physically interpretable hazard level determination results.

[0006] This invention provides a method for determining the hazard of hidden disaster-causing factors in coal mines, comprising: Acquire multi-source heterogeneous monitoring data of the coal mining area, including geological structure data, hydrogeological data, gas occurrence data, mining disturbance data, and real-time sensor monitoring data; The multi-source heterogeneous monitoring data are subjected to spatiotemporal alignment and feature normalization to generate a structured feature matrix in a unified spatiotemporal coordinate system. Based on the structured feature matrix, a disaster-causing factor correlation graph is constructed, where nodes represent various disaster-causing factors and edge weights represent the statistical correlation or physical coupling strength between factors. The disaster-causing factor association graph is input into the graph convolutional neural network module, and the high-order neighborhood aggregation features of the nodes are extracted through multi-layer graph convolution operations to generate a graph embedding representation of each disaster-causing factor. Simultaneously, the structured feature matrix is ​​input into the time series attention encoder, and the multi-head self-attention mechanism is used to capture the dynamic evolution of each disaster-causing factor in the time dimension, generating time series attention-weighted features; Cross-modal feature fusion is performed on the graph embedding representation and the temporal attention weighted features. A gated fusion unit is used to calculate the dynamic weight allocation of the two in the feature space to generate a fused comprehensive disaster-causing feature vector. The comprehensive disaster-causing feature vector is input into the multilayer perceptron decoder, and the output is a judgment result corresponding to the preset hazard level classification standard. The hazard level is divided into four levels: low hazard, medium hazard, high hazard and extremely high hazard. Based on the judgment results, a visual risk heat map is generated, and the corresponding early warning response mechanism is triggered.

[0007] Preferably, the multi-source heterogeneous monitoring data undergoes spatiotemporal alignment and feature normalization processing to generate a structured feature matrix in a unified spatiotemporal coordinate system, including: Based on the coordinates of the mine's geographic information system, all static data are mapped to a unified spatial grid. The dynamic monitoring data is segmented according to a preset time window, and the linear interpolation method is used to unify the data with different sampling frequencies to a sampling interval of once per minute; Minimum-maximum normalization is applied to all numerical features to compress the value range to the interval between 0 and 1. Perform one-hot encoding transformation on categorical features.

[0008] Preferably, based on the structured feature matrix, a disaster-causing factor correlation diagram is constructed, including: Calculate the Pearson correlation coefficient between any two types of disaster-causing factors in historical accident samples. If the absolute value is greater than 0.6, then establish an undirected edge between the corresponding nodes. For factor pairs with a clear physical causal relationship, a directed edge is forcibly established, with the edge weight set to 1; This forms a heterogeneous graph structure that includes node attributes and edge weights.

[0009] Preferably, the disaster-causing factor association graph is input into a graph convolutional neural network module, and high-order neighborhood aggregation features of nodes are extracted through multi-layer graph convolution operations to generate a graph embedding representation of each disaster-causing factor, including: A three-layer graph convolutional layer is used, and the neighborhood aggregation function of each layer is a weighted summation form. The aggregation weight is determined by both the edge weight and the node degree. Each layer's output is followed by a batch normalized layer and a modified linear unit activation function; The node embedding dimension of the third layer output is 128.

[0010] Preferably, the structured feature matrix is ​​input into a time-series attention encoder, and a multi-head self-attention mechanism is used to capture the dynamic evolution of each disaster-causing factor in the time dimension, generating time-series attention-weighted features, including: It contains two stacked encoder blocks, each consisting of a multi-head self-attention sublayer and a feedforward neural network sublayer; The number of multiple heads is set to 8, and the dimension of each head is 16; Position encoding uses a combination of sine and cosine functions; The output is a context-aware feature vector for each time step, and average pooling is performed on the time dimension to obtain temporal attention-weighted features.

[0011] Preferably, the graph embedding representation and the temporal attention-weighted features are fused across modalities, and a gated fusion unit is used to calculate the dynamic weight allocation of the two in the feature space to generate a fused comprehensive catastrophic feature vector, including: The gated fusion unit consists of two fully connected layers and a sigmoid activation function. It receives a concatenated vector of graph embedding representation and temporal attention weighted features as input and outputs fusion weights between 0 and 1. Comprehensive disaster-causing feature vector The calculation formula is: , For graph embedding representation, For temporal attention-weighted features, The first fully connected layer maps the input to sixty dimensions, the second fully connected layer maps it to one dimension, and the output is a scalar after activation by the Sigmoid function.

[0012] Preferably, the comprehensive disaster-causing feature vector is input into the multilayer perceptron decoder, and the output is a judgment result corresponding to a preset hazard level classification standard, including: The multilayer perceptron decoder consists of three fully connected layers with hidden layer dimensions of 256, 128, and 64, respectively, and the activation function is the modified linear unit. The output layer uses the Softmax function to generate probability distributions for four risk levels; The judgment result is the level with the highest probability.

[0013] Preferably, based on the determination result, a visual risk heatmap is generated, and a corresponding early warning response mechanism is triggered, including: Using the mine roadway topology map as the base map, the local risk index is calculated based on the comprehensive disaster-causing feature vector of each area, and the risk level is mapped by a four-color gradient of red-orange-yellow-green. When the assessment result is high risk, personnel entry is automatically restricted and enhanced local ventilation is activated. When the assessment result is extremely high danger, immediately cut off the power supply and start the emergency drainage and gas extraction system.

[0014] Preferably, the geological structural data includes fault distribution, fold morphology, geostress field direction, and rock mass integrity coefficient; The hydrogeological data includes aquifer thickness, water pressure gradient, permeability coefficient, and historical records of water inrush. The gas occurrence data includes gas pressure, gas content, adsorption and desorption characteristics, and coal seam permeability coefficient. The mining disturbance data includes mining progress, support method, roof collapse range, and frequency of microseismic events; The real-time sensing and monitoring data includes methane concentration, carbon monoxide concentration, temperature, humidity, surrounding rock displacement, microseismic energy release rate, and water level change rate.

[0015] Preferably, the aggregation function is a weighted summation formula: , Represents a node The neighborhood group, In the first Each node in the layer Graph embedding representation, For the first After layer graph convolution, nodes Embedded representation, edge weight matrix The elements in and They are nodes and The degree, For the first The learnable weight matrix of the layer, To modify the activation function of the linear unit.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention abandons the traditional subjective weight allocation mode based on expert scoring or analytic hierarchy process, and explicitly models the physical coupling and statistical dependence between various hidden disaster-causing factors by constructing a disaster-causing factor association diagram. 2. By utilizing graph convolutional neural networks, we can effectively capture high-order interaction features in spatial topology, overcoming the limitation of traditional machine learning methods that only process independent and identically distributed samples. 3. By introducing a time series attention mechanism, the dynamic evolution trend of disaster-causing factors with the mining process can be accurately depicted, avoiding the omission of instantaneous change risks by the static assessment model. 4. Adaptive weighted fusion of graph structure features and time series features is achieved through gated fusion units to ensure that the model maintains high discriminative ability under different geological conditions and mining stages; 5. The final output hazard level not only has a high accuracy rate, but its judgment process can also be backtracked through node embedding and attention weighting, which has good physical interpretability and provides scientific, reliable and real-time decision support for coal mine safety production. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the cross-modal fusion model based on graph neural networks and attention mechanisms in this invention; Figure 3 This is a flowchart illustrating the spatiotemporal alignment and feature normalization processing of multi-source heterogeneous monitoring data in this invention. Figure 4 This is a flowchart illustrating the logical flow of disaster-causing factor association graph construction and graph convolution feature extraction in this invention. Figure 5 This is a flowchart illustrating the logical flow of time-series attention encoding and cross-modal gating fusion in this invention. Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the central server in this invention. Detailed Implementation

[0018] refer to Figures 1 to 6 This invention provides a method for determining the hazard of hidden disaster-causing factors in coal mines. By integrating geological structure, hydrological conditions, gas occurrence status, mining disturbance effects, and real-time sensor monitoring data, a multi-source heterogeneous feature system with spatiotemporal consistency is constructed. Based on this, a cross-modal fusion model combining graph neural networks and time-series attention mechanisms is introduced to achieve deep modeling of the nonlinear and dynamic coupling relationships among hidden disaster-causing factors, ultimately outputting a physically interpretable hazard level determination result. This method strictly follows a pre-defined S-step process, executing S1 to S8 sequentially to ensure the integrity and reproducibility of the technical solution.

[0019] Step S1 involves acquiring multi-source heterogeneous monitoring data for the coal mining area. This multi-source heterogeneous monitoring data includes geological structural data, hydrogeological data, gas occurrence data, mining disturbance data, and real-time sensor monitoring data. Geological structural data covers the spatial distribution, strike and dip parameters of faults, the axial plane orientation and limb dip angles of folds, the direction and magnitude of the principal stresses in the regional stress field, and the rock mass integrity coefficient. The rock mass integrity coefficient is obtained through sonic logging or core drilling tests, and its value typically ranges from 0.2 to 0.9. Hydrogeological data includes the vertical thickness of major aquifers, inter-layer water pressure gradients, rock strata permeability coefficients, and records of the location, time, and volume of historical water inrush events. Gas occurrence data includes gas pressure in the coal seam, gas content per unit mass of coal, adsorption and desorption isotherm parameters of gas in the coal matrix, and the coal seam permeability coefficient, which reflects the flow capacity of gas within the coal seam fracture network.

[0020] Mining disturbance data is provided by the mine production scheduling system, including the current advance distance and direction of the mining face, the type and density of anchor bolts / cables used in roadway support, the range and height of roof collapse, and the frequency and total energy release of microseismic events recorded by the microseismic monitoring system. Real-time sensor monitoring data is continuously collected by sensor arrays deployed at key nodes in the roadway, including methane volume concentration, carbon monoxide volume concentration, ambient temperature, relative humidity, surrounding rock surface displacement, microseismic energy release rate, and water level change rate in water tanks or boreholes. All of the above data is accessed through the standard industrial Ethernet interface of the mine integrated automation platform, with a data update cycle of no more than 10 seconds to ensure the timeliness of information.

[0021] Step S2 involves performing spatiotemporal alignment and feature normalization on the multi-source heterogeneous monitoring data to generate a structured feature matrix in a unified spatiotemporal coordinate system. The process begins by mapping all static data onto a spatial grid of a preset resolution, using the three-dimensional spatial coordinate system established by the mine's geographic information system as a reference. The grid cell size is set to 10 meters × 10 meters × 5 meters based on the mine's scale. For dynamic monitoring data, due to their varying sampling frequencies, they need to be segmented according to a preset time window, with a time window length of 30 minutes. Within each time window, linear interpolation is used to unify the timestamps of all dynamic data to one sampling point per minute, thus forming a time series with a fixed time step. After completing the spatiotemporal alignment, all numerical features are subjected to min-max normalization, calculated using the following formula: , This is the original data. and These are the global minimum and maximum values ​​of all numerical features in the historical three-year data set, ensuring that all numerical features are compressed to a closed interval between 0 and 1. For categorical features, such as support methods, one-hot encoding is used to convert them into binary vectors with dimensions equal to the number of categories. Finally, all processed data is organized into a three-dimensional tensor with dimensions of [number of time steps, number of spatial grid cells, feature dimension], which is the structured feature matrix.

[0022] Step S3 involves constructing a disaster-causing factor correlation graph based on the structured feature matrix. This disaster-causing factor correlation graph is a directed heterogeneous graph. , Represents a set of nodes, each node This corresponds to a category of disaster-causing factors, such as "fault activity," "aquifer water pressure," and "gas pressure." Represents the set of edges; This is a node attribute matrix, where each row corresponds to the eigenvector of a node. The eigenvector is composed of the average eigenvalues ​​of the corresponding spatial grid cells in step S2 within the current time window. This is the edge weight matrix. Edge construction follows two rules: First, calculate the Pearson correlation coefficient between any two disaster-causing factors in the historical accident sample database. If its absolute value is greater than 0.6, then establish an undirected edge between the corresponding two nodes. Second, for factor pairs with a clear physical causal relationship, a directed edge is forcibly established, such as an edge from "aquifer water pressure" to "water inrush risk," and an edge from "gas pressure" to "outburst risk." The weights of such edges are... It is directly set to 1. For edges established through statistical correlation, the weight is... This is equal to the absolute value of the corresponding Pearson correlation coefficient. The resulting disaster-causing factor correlation diagram not only includes the attribute information of the nodes themselves, but also explicitly encodes the statistical dependencies and physical causal relationships between the factors.

[0023] Step S4 involves inputting the disaster-causing factor association graph into a graph convolutional neural network module. Through multi-layer graph convolution operations, high-order neighborhood aggregation features of nodes are extracted to generate a graph embedding representation for each disaster-causing factor. This graph convolutional neural network module employs three stacked graph convolutional layers. In the... Layer, each node Graph embedding representation Updates are performed by aggregating information from its first-order neighbors, and the aggregation function is a weighted summation form: , Represents a node The neighborhood group, For the first After layer graph convolution, nodes Embedded representation, edge weight matrix The elements in and They are nodes and The degree (i.e., the number of connecting edges). For the first The learnable weight matrix of the layer, To correct the linear unit activation function, the input of the first layer is the node attribute matrix. After each graph convolution operation, a batch normalization layer is applied to stabilize the training process and accelerate convergence. After three layers of graph convolution, the final embedding dimension of each node is 128-dimensional. This 128-dimensional vector is the graph embedding representation of the disaster-causing factor, which integrates its own attributes and its contextual information in the multi-hop neighborhood of the disaster-causing factor association graph.

[0024] Step S5 involves inputting the structured feature matrix into a time-series attention encoder. A multi-head self-attention mechanism is used to capture the dynamic evolution of each catastrophic factor over time, generating time-series attention-weighted features. This time-series attention encoder consists of two identical stacked encoder blocks. Each encoder block contains a multi-head self-attention sublayer and a feedforward neural network sublayer. For the input structured feature matrix, positional encoding is first added using a combination of sine and cosine functions to preserve the sequential information of the time series. The input sequence is then fed into the multi-head self-attention sublayer, which has 8 heads, each with a dimension of 16, for a total of 128 dimensions. The multi-head self-attention mechanism allows the model to learn the dependencies between time steps in parallel across different subspaces; for example, one head might focus on short-term mutations, while another focuses on long-term trends. The output of the multi-head self-attention sublayer is then nonlinearly transformed by a two-layer feedforward neural network. After processing by the two encoder blocks sequentially, the output is a sequence of context-aware feature vectors with the same number of input time steps, each time step being 128 dimensions. By performing average pooling on the time dimension of the context-aware feature vector sequence, a single 128-dimensional vector is obtained, which is the temporal attention-weighted feature, which condenses the dynamic evolution pattern of disaster-causing factors throughout the entire time window.

[0025] Step S6 involves cross-modal feature fusion of the graph embedding representation and the temporal attention-weighted features. A gated fusion unit is used to calculate the dynamic weight allocation of the two in the feature space, generating a fused comprehensive catastrophic feature vector. The gated fusion unit consists of two fully connected layers and a sigmoid activation function. Its input is a concatenated vector of the graph embedding representation and the temporal attention-weighted features, with a dimension of 256. The first fully connected layer maps the input to 60 dimensions, the second fully connected layer maps it to one dimension, and then the input is activated by the sigmoid function, outputting a scalar. Its value range is 0 to 1. This scalar g is the fusion weight of the graph embedding representation. The formula for calculating the comprehensive disaster-causing feature vector z is: , For graph embedding representation, This is a temporal attention-weighted feature. The gating mechanism can adaptively determine whether to rely more on the spatial topological relationship between disaster-causing factors or on their temporal evolution pattern based on the characteristics of the current input data, thereby achieving optimal fusion of the two modalities.

[0026] Step S7 involves inputting the comprehensive disaster-causing feature vector into a multilayer perceptron decoder and outputting a judgment result corresponding to a preset hazard level classification standard. This multilayer perceptron decoder comprises a three-layer fully connected network. The first hidden layer has a dimension of 256, the second 128, and the third 64, with each hidden layer using a modified linear unit as the activation function. The output layer contains four neurons, corresponding to four preset hazard levels: low, medium, high, and extremely high. The output layer is activated using the Softmax function, generating a four-dimensional probability distribution vector, where each element represents the probability of the corresponding hazard level. The judgment result is the level with the highest probability value in this probability distribution vector.

[0027] Step S8 involves generating a visualized risk heatmap based on the judgment results and triggering the corresponding early warning response mechanism. The visualized risk heatmap uses a two-dimensional or three-dimensional topological map of the mine roadways as its base map. The system calculates a local risk index based on the comprehensive disaster-causing feature vector z of each spatial grid unit. This local risk index is the sum of the probabilities of the "high risk" and "extremely high risk" levels output in step S7. This local risk index is mapped using a four-color gradient of red-orange-yellow-green, where red represents extremely high risk and green represents low risk. The risk heatmap is displayed in real-time on the operating terminal of the mine safety monitoring center, with a refresh frequency of no less than once every 30 seconds. The early warning response mechanism is strictly linked to the judgment result: when the judgment result is high risk, the system automatically sends an instruction to the personnel positioning system to restrict non-essential personnel from entering the high-risk area, and at the same time sends an instruction to the ventilation control system to start the local ventilation enhancement program to increase the air volume to dilute the gas; when the judgment result is extremely high risk, the system immediately sends an instruction to the power monitoring system to cut off the non-intrinsically safe power supply to the high-risk area, and at the same time starts the emergency drainage pump and gas extraction system to actively intervene and prevent disasters from occurring.

[0028] The complete implementation of the above method relies on a supporting hardware and software system architecture. The multi-source data acquisition module aggregates data streams in real time from geological exploration databases, hydrological monitoring stations, gas extraction monitoring systems, microseismic monitoring networks, and environmental sensor arrays through the standardized interface of the mine's integrated automation platform. The data preprocessing module is deployed on edge computing nodes close to the data sources, featuring local caching and breakpoint resume capabilities. Even in the event of network interruption, it can utilize locally stored historical extreme values ​​to perform basic normalization operations and write the preprocessed feature matrix into the central time-series database in an efficient columnar storage format. The graph feature extraction module and the time-series feature extraction module run on the deep learning inference engine of the central server. This deep learning inference engine employs a tensor parallel strategy, distributing the computational tasks of graph convolution and temporal encoding to multiple graphics processing units to accelerate model inference. The model parameters are obtained through offline training. The training dataset consists of over 5,000 historical samples from typical coal mines nationwide over the past 10 years, including accident cases and normal operating condition samples, ensuring the model has broad generalization capabilities. The early warning and visualization module, as the final interface for human-computer interaction, is integrated into the operation terminal of the mine safety monitoring center. It supports multi-screen linkage display and reliably outputs early warning commands to various actuators through intrinsically safe relays for mining.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for determining the hazard of hidden disaster-causing factors in coal mines, characterized in that, include: Acquire multi-source heterogeneous monitoring data of the coal mining area, including geological structure data, hydrogeological data, gas occurrence data, mining disturbance data, and real-time sensor monitoring data; The multi-source heterogeneous monitoring data are subjected to spatiotemporal alignment and feature normalization to generate a structured feature matrix in a unified spatiotemporal coordinate system. Based on the structured feature matrix, a disaster-causing factor correlation graph is constructed, where nodes represent various disaster-causing factors and edge weights represent the statistical correlation or physical coupling strength between factors. The disaster-causing factor association graph is input into the graph convolutional neural network module, and the high-order neighborhood aggregation features of the nodes are extracted through multi-layer graph convolution operations to generate a graph embedding representation of each disaster-causing factor. Simultaneously, the structured feature matrix is ​​input into the time series attention encoder, and the multi-head self-attention mechanism is used to capture the dynamic evolution of each disaster-causing factor in the time dimension, generating time series attention-weighted features; Cross-modal feature fusion is performed on the graph embedding representation and the temporal attention weighted features. A gated fusion unit is used to calculate the dynamic weight allocation of the two in the feature space to generate a fused comprehensive disaster-causing feature vector. The comprehensive disaster-causing feature vector is input into the multilayer perceptron decoder, and the output is a judgment result corresponding to the preset hazard level classification standard. The hazard level is divided into four levels: low hazard, medium hazard, high hazard and extremely high hazard. Based on the judgment results, a visual risk heat map is generated, and the corresponding early warning response mechanism is triggered.

2. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 1, characterized in that, The multi-source heterogeneous monitoring data undergoes spatiotemporal alignment and feature normalization to generate a structured feature matrix in a unified spatiotemporal coordinate system, including: Based on the coordinates of the mine's geographic information system, all static data are mapped to a unified spatial grid. The dynamic monitoring data is segmented according to a preset time window, and the linear interpolation method is used to unify the data with different sampling frequencies to a sampling interval of once per minute; Minimum-maximum normalization is applied to all numerical features to compress the value range to the interval between 0 and 1. Perform one-hot encoding transformation on categorical features.

3. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 2, characterized in that, Based on the structured feature matrix, a disaster-causing factor correlation diagram is constructed, including: Calculate the Pearson correlation coefficient between any two types of disaster-causing factors in historical accident samples. If the absolute value is greater than 0.6, then establish an undirected edge between the corresponding nodes. For factor pairs with a clear physical causal relationship, a directed edge is forcibly established, with the edge weight set to 1; This forms a heterogeneous graph structure that includes node attributes and edge weights.

4. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 3, characterized in that, The disaster-causing factor association graph is input into a graph convolutional neural network module. High-order neighborhood aggregation features of nodes are extracted through multi-layer graph convolution operations to generate a graph embedding representation of each disaster-causing factor, including: A three-layer graph convolutional layer is used, and the neighborhood aggregation function of each layer is a weighted summation form. The aggregation weight is determined by both the edge weight and the node degree. Each layer's output is followed by a batch normalized layer and a modified linear unit activation function; The node embedding dimension of the third layer output is 128.

5. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 4, characterized in that, The structured feature matrix is ​​input into a time-series attention encoder, and a multi-head self-attention mechanism is used to capture the dynamic evolution of each disaster-causing factor in the time dimension, generating time-series attention-weighted features, including: It contains two stacked encoder blocks, each consisting of a multi-head self-attention sublayer and a feedforward neural network sublayer; The number of multiple heads is set to 8, and the dimension of each head is 16; Position encoding uses a combination of sine and cosine functions; The output is a context-aware feature vector for each time step, and average pooling is performed on the time dimension to obtain temporal attention-weighted features.

6. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 5, characterized in that, Cross-modal feature fusion is performed on the graph embedding representation and the temporal attention-weighted features. A gated fusion unit is used to calculate the dynamic weight allocation of the two in the feature space to generate a fused comprehensive catastrophic feature vector, including: The gated fusion unit consists of two fully connected layers and a sigmoid activation function. It receives a concatenated vector of graph embedding representation and temporal attention weighted features as input and outputs fusion weights between 0 and 1. Comprehensive disaster-causing feature vector The calculation formula is: , For graph embedding representation, For temporal attention-weighted features, The first fully connected layer maps the input to sixty dimensions, the second fully connected layer maps it to one dimension, and the output is a scalar after activation by the Sigmoid function.

7. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 6, characterized in that, The comprehensive disaster-causing feature vector is input into the multilayer perceptron decoder, and the output is a judgment result corresponding to the preset hazard level classification standard, including: The multilayer perceptron decoder consists of three fully connected layers with hidden layer dimensions of 256, 128, and 64, respectively, and the activation function is the modified linear unit. The output layer uses the Softmax function to generate probability distributions for four risk levels; The judgment result is the level with the highest probability.

8. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 7, characterized in that, Based on the determination result, a visual risk heatmap is generated, and a corresponding early warning response mechanism is triggered, including: Using the mine roadway topology map as the base map, the local risk index is calculated based on the comprehensive disaster-causing feature vector of each area, and the risk level is mapped by a four-color gradient of red-orange-yellow-green. When the assessment result is high risk, personnel entry is automatically restricted and enhanced local ventilation is activated. When the assessment result is extremely high danger, immediately cut off the power supply and start the emergency drainage and gas extraction system.

9. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 8, characterized in that, The geological structural data includes fault distribution, fold morphology, geostress field direction, and rock mass integrity coefficient. The hydrogeological data includes aquifer thickness, water pressure gradient, permeability coefficient, and historical records of water inrush. The gas occurrence data includes gas pressure, gas content, adsorption and desorption characteristics, and coal seam permeability coefficient. The mining disturbance data includes mining progress, support method, roof collapse range, and frequency of microseismic events; The real-time sensing and monitoring data includes methane concentration, carbon monoxide concentration, temperature, humidity, surrounding rock displacement, microseismic energy release rate, and water level change rate.

10. The method for determining the hazard of hidden disaster-causing factors in coal mines according to claim 1, characterized in that, The aggregate function, in the form of a weighted summation, is as follows: , Represents a node The neighborhood group, In the first Each node in the layer Graph embedding representation, For the first After layer graph convolution, nodes Embedded representation, edge weight matrix The elements in and They are nodes and The degree, For the first The learnable weight matrix of the layer, To modify the activation function of the linear unit.

Images