Coal rock dynamic disaster intelligent early warning method and system

Abstract

This invention discloses an intelligent early warning method and system for coal and rock dynamic disasters, relating to the field of coal and rock dynamic disaster monitoring and early warning technology. The early warning method includes the following steps: constructing a spatiotemporally aligned multi-source heterogeneous fusion database; quantitatively constructing a composite dynamic coupling evolution index system, combining hierarchical analysis and the coefficient of variation method to determine the comprehensive weight of the index through combined weighting, and classifying the disaster early warning level into multiple levels from low to high; constructing a three-dimensional geomechanical model and discretizing it into grid cells, dynamically mapping the multi-source heterogeneous fusion database and the calculated real-time K value to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model; constructing a spatiotemporal graph neural network (STGCN) prediction model, using the dataset to supervise learning, training, and validation of the model, iteratively optimizing the model until the model is adjusted to the optimal parameters; inputting the newly collected data into the STGCN optimal model to predict the comprehensive danger threshold of the entire working face in the future time period.

Description

Technical Field

[0001] This invention relates to the field of coal and rock dynamic disaster monitoring and early warning technology, and in particular to an intelligent early warning method and system for coal and rock dynamic disasters. Background Technology

[0002] The mechanisms of coal and rock dynamic disasters are becoming increasingly complex, and their suddenness and destructiveness are intensifying. Therefore, accurate monitoring and early warning of coal and rock dynamic disaster risks within the mining face area has become a critical bottleneck restricting the safe and efficient mining of deep coal resources. Traditional early warning methods for coal and rock dynamic disasters often rely on threshold judgments of single physical quantities and simple fusion of multiple parameter variables. These methods lack a macroscopic spatiotemporal perspective on the entire process of disaster incubation, evolution, and occurrence, making it difficult to capture the deep nonlinear relationships between multi-source heterogeneous data, ignoring the correlation between data, resulting in significant lag in early warning. The corresponding indicator systems are also relatively simple, mostly based on empirical summaries, lacking in-depth consideration of the mechanical mechanisms of coal and rock mass instability and failure, particularly neglecting the dynamic coupling effects between multiple physical fields. This makes it difficult to comprehensively reflect the complete evolution process of the coal and rock mass system from steady state to instability, and also fails to accurately characterize the special laws governing disaster incubation under different geological conditions and mining processes.

[0003] A literature search of existing technologies revealed Chinese Patent Publication No. CN116467572A, published on July 21, 2023, entitled: "A Prediction Method for Coal-Rock-Gas Composite Dynamic Disasters Based on Deep Learning." The patent's technical solution is as follows: "First, prepare the impact index data of coal-rock-gas composite dynamic disasters, and clean the data using Boxplot analysis and MICE chain equation multiple interpolation method; second, use Grey Relational Analysis (GRA) to analyze the data and establish a prediction index system for coal-rock-gas composite dynamic disasters; third, use convolutional neural networks..." The method involves building a model using a CNN network and optimizing the model's hyperparameters with the Sparrow Search Algorithm (SSA) to improve prediction accuracy. Next, the model is trained to establish a prediction model for coal-rock-gas composite dynamic disasters based on GRA-SSA-CNN. Finally, a test set is used for prediction, and the predicted results are compared with actual results to determine the model's prediction accuracy. However, this method has limitations. While it comprehensively analyzes the multi-source influencing factors of coal-rock-gas composite dynamic disasters and constructs a prediction index system by introducing techniques such as grey relational analysis and deep learning, it has certain methodological limitations. First, this method is essentially a static risk rating rather than a dynamic prediction. It focuses on the hazard classification of specific geological blocks under certain parameters, failing to effectively integrate time-series data characteristics and unable to capture the dynamic evolution process before the disaster occurs. Second, the model lacks an effective closed-loop feedback and self-optimization mechanism, unable to learn from new monitoring data and early warning effectiveness. This leads to a gradual degradation of its early warning performance when facing constantly changing mine geological conditions and mining layouts, hindering continuous self-evolution and adaptation.

[0004] Therefore, there is an urgent need for an intelligent early warning method and system for coal and rock dynamic disasters that can solve the above problems. Summary of the Invention

[0005] This solution addresses the problems and needs raised above by proposing an intelligent early warning method and system for coal and rock dynamic disasters. Due to the adoption of the following technical features, it can achieve the above-mentioned technical objectives and bring about several other technical effects.

[0006] One objective of this invention is to provide an intelligent early warning method for coal and rock dynamic disasters, comprising the following steps: S10: Distribute multiple sensors in the mining face and surrounding area to synchronously and in real time collect acoustic and electrical monitoring data, microseismic data, gas concentration data and triaxial stress data, and preprocess the collected raw data to build a spatiotemporally aligned multi-source heterogeneous fusion database. S20: Based on a multi-source heterogeneous fusion database, key data fragments characterizing the evolution of mining-induced stress field, fracture field, and gas field are extracted. From the dual dimensions of intra-field evolution index and inter-field coupling index, a composite dynamic coupling evolution index system is quantitatively constructed. Combining the hierarchical analysis method and the coefficient of variation method, the comprehensive weight of the index is determined by combined weighting. Based on the comprehensive weight of the index and historical disaster data in the multi-source heterogeneous fusion database, the comprehensive danger threshold K value is calculated, and the disaster warning level is divided into multiple levels from low to high. S30: Based on the geological exploration data of the mining face and the roadway design drawing, a high-precision three-dimensional geomechanical model is constructed and discretized into grid units. The multi-source heterogeneous fusion database and the calculated real-time comprehensive hazard threshold K value are dynamically mapped to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model that is synchronized in real time with the geometry, physical properties and status of the physical working face. S40: Using 3D grid cells as graph nodes, edge connections are defined through spatial adjacency relationships to construct a weighted adjacency matrix. Based on the time series data of the indicator system, a node feature matrix is ​​formed. The dynamically evolving digital twin sequence is transformed into a spatiotemporal graph dataset. A spatiotemporal graph neural network STGCN prediction model is constructed. The model is supervised, trained, and validated using this dataset. The comprehensive danger threshold K value prediction result is output. The model is iteratively optimized until the model is adjusted to the optimal parameters. S50: The newly acquired data in real time is fed into the optimal model of the spatiotemporal graph neural network STGCN prediction model to predict the comprehensive danger threshold of the three-dimensional grid cells of the entire working face in the future time period, and the prediction results are fed back to the spatiotemporal graph neural network STGCN prediction model to realize the self-optimization of the spatiotemporal graph neural network STGCN prediction model.

[0007] In addition, the intelligent early warning method and system for coal and rock dynamic disasters according to the present invention may also have the following technical features: In one example of the present invention, in step S20, a composite dynamic coupling evolution index system is quantitatively constructed. By combining the hierarchical analysis method and the coefficient of variation method, the comprehensive weight of the index is determined through combined weighting. Based on the comprehensive weight and historical disaster data, the comprehensive danger threshold K value is calculated, and the disaster warning levels are divided from low to high. Specifically, the steps include the following: S21: Starting from two dimensions, namely, intra-field evolution indicators and inter-field coupling indicators, a composite dynamic coupling evolution indicator system is quantitatively constructed. Among them, intra-field evolution indicators include acoustic emission amplitude indicators, electromagnetic radiation amplitude indicators, micro-seismic amplitude indicators, gas concentration indicators, and maximum principal stress indicators. Inter-field coupling indicators include stress-energy coupling coefficient indicators, crack-gas coupling strength indicators, and stress-electromagnetic coupling sensitivity indicators. Appropriate time intervals are selected according to actual conditions. S22: First, Z-Score standardization is performed on the time series data of relevant indicators, and then Sigmoid normalization is performed. The subjective weight of the indicators is determined by the hierarchical analysis method, and the objective weight of the indicators is determined by the coefficient of variation method. The comprehensive weight of the indicators is determined by combining weights. Based on the comprehensive weight and historical disaster data, the comprehensive danger threshold K value is calculated, and the disaster warning level is divided into four levels, A, B, C and D, from low to high, using blue, green, yellow and red colors for identification.

[0008] In one example of the present invention, the specific expressions for the stress-energy coupling coefficient index, the crack-gas coupling strength index, and the stress-electromagnetic coupling sensitivity index in step S21 are as follows: Stress-energy coupling coefficient The expression is: Crack-gas coupling strength index The expression is: Stress-electromagnetic coupling sensitivity index The expression is: In the formula, Representing the A time interval, This represents the covariance between two variables. This represents the stress change sequence during the k-th time interval; This is the energy release sequence for the k-th time interval; This represents the standard deviation of X; This represents the standard deviation of Y; This represents the number of acoustic emission signal impacts after standardization and normalization. This represents the standardized and normalized value of the gas concentration. Let be the change in electromagnetic radiation intensity during the k-th time interval; Let be the stress change during the k-th time interval.

[0009] In one example of the present invention, in step S22, the comprehensive danger threshold K value is calculated, and the disaster warning level is divided into four levels, A, B, C, and D, from low to high, using four colors: blue, green, yellow, and red for identification. This includes the following steps: Overall hazard level A: 0≤K<0.3, blue, current status is normal, continue operation; Overall hazard level B: 0.3≤K<0.6, green, current status is low risk, intensify monitoring of the work area; Overall hazard level C: 0.6≤K<0.8, yellow, current status is medium risk, reduce mining speed and reduce number of workers; Overall hazard level D: 0.8≤K<1, red, current status is high risk. All personnel should immediately stop work and evacuate the work area, and the emergency plan should be activated.

[0010] In one example of the present invention, step S30 specifically includes the following steps: S31: Based on the relevant geological exploration data of the tunneling face and combined with the tunnel design drawings, professional mining software is used to create a three-dimensional solid model of the geological structure and construct a high-precision three-dimensional geomechanical model. S32: Perform spatial discretization operation to divide the entire three-dimensional geomechanical model into regular three-dimensional mesh units, determine the mesh size according to the actual situation, and number the three-dimensional mesh units in sequence. S33: The multi-source heterogeneous fusion database and the calculated real-time comprehensive danger threshold K value are dynamically mapped to the corresponding three-dimensional mesh unit through a spatial interpolation algorithm to construct a digital twin model that is synchronized in real time with the geometry, physical properties and status of the physical working surface. A professional three-dimensional visualization engine is used to develop a visualization interface for the digital twin.

[0011] In one example of the present invention, step S33, dynamically mapping data to the corresponding three-dimensional mesh element using a spatial interpolation algorithm, includes the following steps: Kriging spatial interpolation is used to fill in the grid cell data that cannot be covered by sensors. Based on limited sensor data, the dynamic properties of grid cells in areas where no sensors are deployed are calculated to form a continuous cloud map. The specific expression is as follows: Kriging interpolation formula: Unbiased estimation conditions: In the formula, For point The estimated value; n This represents the total number of known monitoring points in the surrounding area involved in the calculation; z is the weighting coefficient; i This represents the true value of the i-th known monitoring point distributed around the target point; For point The true value; Among them, the weighting coefficient The value of an unknown point is estimated by weighted summation of data from all known points in space, satisfying the condition that the unknown point... The estimated value at the location Compared with the true value The set of optimal coefficients that minimizes the difference, i.e. Meanwhile, the Kriging spatial interpolation method satisfies the unbiased estimation condition in its calculation process to ensure the macroscopic accuracy of spatial prediction.

[0012] In one example of the present invention, step S40 specifically includes the following steps: S41: Construct a spatial topology graph, using 3D mesh cells as graph nodes, and build a weighted adjacency matrix based on the spatial relationships between nodes. A node feature matrix is ​​formed based on the time series data of the indicator system. The feature matrix consists of normalized index data and a comprehensive risk threshold K. The dynamically evolving digital twin sequence is transformed into a spatiotemporal graph dataset, and divided into training, validation and test sets. S42: Construct an STGCN early warning model containing a spatiotemporal convolution module, and optimize the model using the mean squared error loss function and the RMSprop optimizer; S43: Import the training set into the model for supervised learning, training and validation, output the comprehensive danger threshold K value, iteratively optimize the model using the validation set, loss function and optimizer, evaluate the model performance using the test set, and determine the optimal model with the best parameters.

[0013] In one example of the present invention, in step S42, the STGCN early warning model comprises: two spatiotemporal convolutional modules and an output layer. Each spatiotemporal convolutional module is composed of alternating superpositions of temporal gate convolution, spatial graph convolution, and temporal gate convolution. The output layer comprises a temporal convolutional layer and a fully connected layer. The temporal gate convolution is configured to capture the dynamic changes of data in the time dimension. The spatial graph convolution is configured to capture the dependencies of data in the spatial dimension. The temporal gate convolution comprises: two one-dimensional causal convolutions, a Sigmoid function, and a Hadamard product. The two one-dimensional causal convolutions are connected in parallel. One of the one-dimensional causal convolutions is concatenated with the Sigmoid function. The other one-dimensional causal convolution is fused through residual links to alleviate the gradient degradation problem in deep network training. The signal output by the Sigmoid function is fused with the signal of the other one-dimensional causal convolution through residual links and then multiplied by Hadamard.

[0014] In one example of the present invention, in step S50, the actual disaster situation or confirmed safety status is compared with the prediction results of the model to form feedback data, and incremental learning is performed on the spatiotemporal graph neural network STGCN prediction model to achieve continuous self-optimization of the early warning model, including the following steps: S51: Visualize the predicted comprehensive threshold in the form of an early warning cloud map and output the danger level. At the same time, use the actual disaster situation or confirmed safety status as the ground truth label and compare it with the model's prediction results to form feedback data and build an incremental training database. The model prediction results include three situations: missed reporting, false reporting, or successful prediction. S52: The incremental training database is periodically imported into the Spatiotemporal Graph Neural Network (STGCN) prediction model for incremental training and fine-tuning, enabling the model to continuously adapt to the geological conditions and mining changes of specific mines and achieve self-evolution of early warning capabilities.

[0015] Another objective of this invention is to provide an intelligent early warning system for coal and rock dynamic disasters, comprising: The integrated database module is configured to distribute multiple sensors in and around the mining face to synchronously and in real time collect acoustic and electrical monitoring data, microseismic data, gas concentration data and triaxial stress data, and preprocess the collected raw data to build a spatiotemporally aligned multi-source heterogeneous integrated database. The early warning level classification module is configured to extract key data fragments characterizing the evolution of mining-induced stress field, fracture field, and gas field based on a multi-source heterogeneous fusion database. It quantitatively constructs a composite dynamic coupling evolution index system from two dimensions: intra-field evolution index and inter-field coupling index. Combining hierarchical analysis and the coefficient of variation method, it achieves combined weighting to determine the comprehensive weight of the index. Based on the comprehensive weight of the index and historical disaster data in the multi-source heterogeneous fusion database, it calculates the comprehensive danger threshold K value and classifies the disaster early warning level into multiple levels from low to high. The digital twin model module is configured to construct a high-precision three-dimensional geomechanical model based on the geological exploration data of the mining face and the tunnel design drawings, and discretize it into grid cells. The multi-source heterogeneous fusion database and the calculated real-time comprehensive hazard threshold K value are dynamically mapped to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model that is synchronized with the geometry, physical properties and status of the physical working face in real time. The optimal parameter adjustment module is configured to use three-dimensional grid cells as graph nodes, define edge connections through spatial adjacency relationships, construct a weighted adjacency matrix, and form a node feature matrix based on the time series data of the indicator system. It transforms the dynamically evolving digital twin sequence into a spatiotemporal graph dataset, constructs a spatiotemporal graph neural network STGCN prediction model, uses this dataset to perform supervised learning, training and validation of the model, outputs the prediction result of the comprehensive danger threshold K value, and iteratively optimizes the model until the model is adjusted to the optimal parameters. The integrated hazard prediction module is configured to input new data collected in real time into the optimal model of the spatiotemporal graph neural network STGCN prediction model, predict the integrated hazard threshold of the three-dimensional grid cells of the entire working face in the future time period, and feed the prediction results back to the spatiotemporal graph neural network STGCN prediction model to realize the self-optimization of the spatiotemporal graph neural network STGCN prediction model.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention collects and combines real-time acoustic, electrical, micro-seismic, gas concentration, and triaxial stress monitoring data from the mining face and surrounding areas to construct a spatiotemporally aligned multi-source heterogeneous database. Starting from multiple physical fields, it effectively suppresses external interference, more comprehensively depicts the disaster gestation process, and overcomes the limitations of a single data source and the simple integration of multi-parameter fusion early warning methods.

[0017] This invention constructs a composite dynamic coupling evolution index system from two aspects: intra-field evolution index and inter-field coupling index. It integrates multiple physical fields such as gas field, mining stress field, and fracture field. Through the collaborative perception and dynamic fusion of multiple physical fields, it achieves accurate characterization of the coal and rock dynamic disaster incubation process.

[0018] This invention introduces the concept of digital twins into early warning of coal and rock dynamic disasters by constructing a digital twin model of the mining face. The mining face and surrounding area are gridded in three dimensions to create a high-fidelity virtual digital model, which realizes a comprehensive and dynamic display of the mining face status and improves the intuitiveness and fidelity of the system.

[0019] This invention utilizes the Spatiotemporal Graph Neural Network (STGCN) prediction model, constructs a graph dataset to implement full-process supervised learning, training, and validation of the model, and uses loss functions and optimizers to adjust relevant model parameters. It profoundly reveals the evolution law of coal and rock dynamic disasters, effectively reduces the false alarm rate and false negative rate of traditional methods, improves the accuracy and reliability of disaster early warning, and provides strong support for disaster prevention and mitigation decision-making.

[0020] This invention establishes a closed-loop feedback and self-evolution mechanism, which compares the actual disaster situation or confirmed safety status as a truth label with the model's prediction results to form feedback data and build an incremental training database. The spatiotemporal graph neural network STGCN model is incrementally learned periodically to achieve continuous self-optimization of the early warning model, thereby enhancing the accuracy and reliability of long-term applications.

[0021] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0023] Figure 1 A flowchart of an intelligent early warning method for coal and rock dynamic disasters according to an embodiment of the present invention; Figure 2 This is a diagram illustrating the architecture of the composite dynamic coupling evolution index according to an embodiment of the present invention; Figure 3 This is a structural diagram of the spatiotemporal graph neural network STGCN prediction model according to an embodiment of the present invention; Figure 4 This is a structural diagram of the closed-loop feedback and model self-optimization mechanism according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0026] According to a first aspect of the present invention, a method for intelligent early warning of coal and rock dynamic disasters, such as Figure 1 As shown, it includes the following steps: S10: Distribute multiple sensors in and around the mining face. Specifically, integrate acoustic and electrical monitoring sensors, micro-vibration monitoring sensors, gas concentration monitoring sensors, and triaxial stress monitoring sensors to synchronously and in real time collect raw monitoring data of acoustic and electrical monitoring, micro-vibration, gas concentration, and triaxial stress. Perform preprocessing, cleaning, noise reduction, and outlier preprocessing on the raw data to construct a spatiotemporally aligned multi-source heterogeneous fusion database. S20: Based on a multi-source heterogeneous fusion database, key data fragments characterizing the evolution of mining-induced stress fields, fracture fields, and gas fields are extracted. A composite dynamic coupling evolution index system is quantitatively constructed from two dimensions: intra-field evolution indicators and inter-field coupling indicators. Combining hierarchical analysis and the coefficient of variation method, combined weighting is used to determine the comprehensive weight of the indicators. Based on the comprehensive weight of the indicators and historical disaster data in the multi-source heterogeneous fusion database, a comprehensive hazard threshold K value is calculated, and disaster warning levels are divided into multiple levels from low to high; for example, four levels: A, B, C, and D, identified using blue, green, yellow, and red colors. S30: Based on the geological exploration data of the mining face and the roadway design drawing, a high-precision three-dimensional geomechanical model is constructed and discretized into grid units. The multi-source heterogeneous fusion database and the calculated real-time comprehensive hazard threshold K value are dynamically mapped to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model that is synchronized in real time with the geometry, physical properties and status of the physical working face. S40: Using 3D grid cells as graph nodes, edge connections are defined through spatial adjacency relationships to construct a weighted adjacency matrix. Based on the time series data of the indicator system, a node feature matrix is ​​formed. The dynamically evolving digital twin sequence is transformed into a spatiotemporal graph dataset. A spatiotemporal graph neural network STGCN prediction model is constructed. The model is supervised, trained, and validated using this dataset. The comprehensive danger threshold K value prediction result is output. The model is iteratively optimized using a loss function and optimizer until the model is adjusted to the optimal parameters. S50: Real-time acquired new data is fed into the STGCN optimal model to predict the comprehensive hazard threshold of the entire working face's three-dimensional grid cells within a future time period. The prediction results are then fed back to the spatiotemporal graph neural network STGCN prediction model, enabling the STGCN prediction model to self-optimize. Specifically, the model is visualized in the form of an early warning cloud map on the digital twin model. At the same time, the actual disaster situation or confirmed safety status is used as the ground truth label and compared with the model's prediction results to form feedback data. An incremental training database is constructed, and the STGCN model is periodically subjected to incremental learning to achieve continuous self-optimization of the early warning model.

[0027] In short, this early warning method uses a high-fidelity digital twin as the data base and interaction carrier, and an embedded spatiotemporal deep learning model as the intelligent prediction kernel. It constructs a dynamic coupling evolution index system, and through deep integration of multi-dimensional monitoring data, it achieves dynamic and accurate prediction of the spatiotemporal evolution trend of disaster risks in mining space, and establishes a closed-loop feedback mechanism to promote the continuous self-evolution of the model.

[0028] This early warning method collects and combines real-time acoustic, electrical, micro-seismic, gas concentration, and triaxial stress monitoring data from the mining face and surrounding areas to construct a spatiotemporally aligned multi-source heterogeneous database. Starting from multiple physical fields, it effectively suppresses external interference, more comprehensively depicts the disaster gestation process, and overcomes the limitations of a single data source and the simple integration of multi-parameter fusion early warning methods.

[0029] This early warning method starts from two aspects: intra-field evolution indicators and inter-field coupling indicators. It constructs a composite dynamic coupling evolution indicator system, which integrates multiple physical fields such as gas field, mining stress field, and fracture field. Through the collaborative perception and dynamic fusion of multiple physical fields, it achieves accurate characterization of the coal and rock dynamic disaster incubation process.

[0030] This early warning method introduces the concept of digital twins into the early warning of coal and rock dynamic disasters by constructing a digital twin model of the mining face. The mining face and surrounding area are gridded in three dimensions to create a high-fidelity virtual digital model, which realizes a comprehensive and dynamic display of the mining face status and improves the intuitiveness and fidelity of the system.

[0031] This early warning method utilizes the Spatiotemporal Graph Neural Network (STGCN) prediction model, constructs a graph dataset to conduct full-process supervised learning, training, and validation of the model, and uses loss functions and optimizers to adjust relevant model parameters. It profoundly reveals the evolution law of coal and rock dynamic disasters, effectively reduces the false alarm rate and false negative rate of traditional methods, improves the accuracy and reliability of disaster early warning, and provides strong support for disaster prevention and mitigation decision-making.

[0032] This early warning method establishes a closed-loop feedback and self-evolution mechanism. It compares the actual disaster situation or confirmed safety status as a true value label with the model's prediction results to form feedback data and build an incremental training database. It also performs incremental learning on the spatiotemporal graph neural network STGCN model on a regular basis to achieve continuous self-optimization of the early warning model, thereby enhancing the accuracy and reliability of long-term applications.

[0033] In one example of the present invention, step S10 specifically includes the following steps: S11: Based on the direction of the mining face advance and considering the actual situation, scientifically plan the sensor layout spacing to ensure that the spatial coverage density matches the dynamic monitoring requirements; S12: Multiple sets of composite monitoring units are installed simultaneously in the upper and lower tunnels. Each unit integrates acoustic and electrical monitoring sensors, micro-vibration monitoring sensors, gas concentration monitoring sensors, and triaxial stress monitoring sensors to form a distributed sensing network. Time reference correction and synchronization are performed to collect raw data in real time. S13: Preprocess the original data, clean the data, fill in missing values ​​using the front-to-back average method, reduce random noise in the data by combining wavelet denoising technology, and construct a spatiotemporally aligned multi-source heterogeneous fusion database.

[0034] In one example of the present invention, such as Figure 2 As shown, in step S20, a composite dynamic coupling evolution index system is quantitatively constructed. Combining the hierarchical analysis method and the coefficient of variation method, a combined weighting is used to determine the comprehensive weight of the indicators. Based on the comprehensive weight and historical disaster data, the comprehensive danger threshold K is calculated, and disaster warning levels are divided from low to high. Specifically, the steps include: S21: Starting from two dimensions, namely, intra-field evolution indicators and inter-field coupling indicators, a composite dynamic coupling evolution indicator system is quantitatively constructed. Among them, intra-field evolution indicators include acoustic emission amplitude indicators, electromagnetic radiation amplitude indicators, micro-seismic amplitude indicators, gas concentration indicators, and maximum principal stress indicators. Inter-field coupling indicators include stress-energy coupling coefficient indicators, crack-gas coupling strength indicators, and stress-electromagnetic coupling sensitivity indicators. Appropriate time intervals are selected according to actual conditions. S22: First, Z-score standardization is applied to the time-series data of relevant indicators, followed by Sigmoid normalization. The subjective weights of the indicators are determined using the analytic hierarchy process (AHP), and the objective weights are determined using the coefficient of variation method. This combined weighting determines the comprehensive weights of the indicators. Based on the comprehensive weights and historical disaster data, the comprehensive hazard threshold K is calculated. Disaster warning levels are then divided into four grades—A, B, C, and D—from low to high, using blue, green, yellow, and red colors for identification. The specific expression is as follows: Z-Score standardization: Sigmoid normalization: Combination weighting: In the formula, It is the mean of the data. It is the standard deviation of the data. It is a number between 0 and 1, adjusted according to the actual situation. The weights of the analytic hierarchy process (AHP) are represented. This represents the weighting of the coefficient of variation method.

[0035] In one example of the present invention, the specific expressions for the stress-energy coupling coefficient index, the crack-gas coupling strength index, and the stress-electromagnetic coupling sensitivity index in step S21 are as follows: Stress-energy coupling coefficient The expression is: Crack-gas coupling strength index The expression is: Stress-electromagnetic coupling sensitivity index The expression is: In the formula, Representing the A time interval, This represents the covariance between two variables. The stress change sequence for the k-th time interval is expressed as: In the formula, t represents the length of the stress change sequence. This is the sequence of microseismic energy changes from time k-t+1 to the current time k. The stress change during the k-th time interval is expressed as: In the formula, This represents the stress monitoring value at the k-th time interval. Stress monitoring value at the (k-1)th time interval; The energy release sequence for the k-th time interval is expressed as: In the formula, The sequence of microseismic energy changes from time k-t+1 to the current time k; Let be the change in microseismic energy during the k-th time interval, and its expression is: In the formula, This represents the microseismic energy monitoring value at the k-th time interval. This represents the microseismic energy monitoring value at the (k-1)th time interval; This represents the standard deviation of X; This represents the standard deviation of Y; This represents the number of acoustic emission signal impacts after standardization and normalization. This represents the standardized and normalized value of the gas concentration. Let be the change in electromagnetic radiation intensity during the k-th time interval, and its expression is: In the formula, This represents the electromagnetic radiation intensity monitoring value at the k-th time interval. This represents the monitored electromagnetic radiation intensity value for the (k-1)th time interval. Let be the stress change during the k-th time interval.

[0036] In one example of the present invention, in step S22, the comprehensive danger threshold K value is calculated, and the disaster warning level is divided into four levels, A, B, C, and D, from low to high, using four colors: blue, green, yellow, and red for identification. This includes the following steps: Overall hazard level A: 0≤K<0.3, blue, current status is normal, work can continue; Overall hazard level B: 0.3≤K<0.6, green, current status is low risk, intensify monitoring of the work area; Overall hazard level C: 0.6≤K<0.8, yellow, current status is medium risk, reduce mining speed and reduce number of workers; Overall hazard level D: 0.8≤K<1, red, current status is high risk. All personnel should immediately stop work and evacuate the work area, and the emergency plan should be activated.

[0037] In one example of the present invention, step S30 specifically includes the following steps: S31: Based on the relevant geological exploration data of the tunneling face and combined with the tunnel design drawings, professional mining software is used to perform three-dimensional solid modeling of the main geological structures and construct a high-precision three-dimensional geomechanical model. S32: Perform spatial discretization operation to divide the entire three-dimensional geomechanical model into regular three-dimensional mesh units, determine the mesh size according to the actual situation, and number the three-dimensional mesh units in sequence. S33: The multi-source heterogeneous fusion database and the calculated real-time comprehensive danger threshold K value are dynamically mapped to the corresponding three-dimensional mesh unit through a spatial interpolation algorithm to construct a digital twin model that is synchronized in real time with the geometry, physical properties and status of the physical working surface. A professional three-dimensional visualization engine is used to develop a visualization interface for the digital twin.

[0038] In one example of the present invention, step S33, dynamically mapping data to the corresponding three-dimensional mesh element using a spatial interpolation algorithm, includes the following steps: Kriging spatial interpolation is used to fill in the grid cell data that cannot be covered by sensors. Based on limited sensor data, the dynamic properties of grid cells in areas where no sensors are deployed are calculated to form a continuous cloud map. The specific expression is as follows: Kriging interpolation formula: Unbiased estimation conditions: In the formula, For point The estimated value; n This represents the total number of known monitoring points in the surrounding area involved in the calculation; z is the weighting coefficient; i This represents the true value of the i-th known monitoring point distributed around the target point; For point The true value; Among them, the weighting coefficient The value of an unknown point is estimated by weighted summation of data from all known points in space, satisfying the condition that the unknown point... The estimated value at the location Compared with the true value The set of optimal coefficients that minimizes the difference, i.e. Meanwhile, the Kriging spatial interpolation method satisfies the unbiased estimation condition in its calculation process to ensure the macroscopic accuracy of spatial prediction.

[0039] In one example of the present invention, step S40 involves converting the dynamically evolving digital twin sequence into a spatiotemporal graph dataset, constructing a spatiotemporal graph neural network (STGCN) prediction model, using the dataset to learn, train, and validate the model, outputting the K-value prediction result, and iteratively optimizing the model until the model is adjusted to the optimal parameters. This includes the following steps: S41: Construct a spatial topology graph, using 3D mesh cells as graph nodes, and build a weighted adjacency matrix based on the spatial relationships between nodes. A node feature matrix is ​​formed based on the time series data of the indicator system. The feature matrix consists of normalized index data and a comprehensive risk threshold K. The dynamically evolving digital twin sequence is transformed into a spatiotemporal graph dataset and divided into training, validation, and test sets. For example, the ratio of the training, validation, and test sets is 7:2:1. The specific weighted adjacency matrix expression is as follows. In the formula, Represents a node and nodes The distance between them and Used for adjustment Distribution and sparsity.

[0040] S42: Construct an STGCN early warning model containing spatiotemporal convolutional modules. Select the sliding window size based on the actual situation, with a stride of 1. The model consists of two spatiotemporal convolutional modules and an output layer. Each spatiotemporal convolutional module is composed of alternating layers of temporal gated convolution, spatial graph convolution, and temporal gated convolution. The output layer includes a temporal convolutional layer and a fully connected layer. The model is optimized using the mean squared error loss function and the RMSprop optimizer, as shown in the following expressions: Mean squared error loss function: Spatiotemporal convolution formula: In the formula, For trainable parameters, It is the actual value. These are predicted values, where, and It is a piece The upper and lower time cores, It is the spectral kernel of graph convolution. It is an activation function, and the graph convolution formula is: ,in , These represent the sizes of the input and output feature maps, respectively. S43: Import the training set into the model for supervised learning, training, and validation. Output the comprehensive danger threshold K. Iteratively optimize the model using the validation set, loss function, and optimizer. Evaluate model performance using the test set and determine the optimal model parameters. For example, the sliding window length is 60, the variable dimension is 9, the number of temporal gate convolution channels is set to 64, the number of spatial graph convolution channels is set to 16, and the initial learning rate is set to 1×10⁻⁶. -3 Batch size is 50, distance scaling parameter sparse threshold .

[0041] In one example of the present invention, such as Figure 3 As shown, in step S42, the STGCN early warning model consists of two spatiotemporal convolutional modules and an output layer. Each spatiotemporal convolutional module is composed of alternating layers of temporal gate convolution, spatial graph convolution, and temporal gate convolution. The output layer includes a temporal convolutional layer and a fully connected layer. The temporal gate convolution is configured to capture the dynamic changes of data in the time dimension. The spatial graph convolution is configured to capture the dependencies of data in the spatial dimension. The temporal gate convolution includes two one-dimensional causal convolutions, a Sigmoid function, and a Hadamard product. The two one-dimensional causal convolutions are connected in parallel. One of the one-dimensional causal convolutions is concatenated with the Sigmoid function. The other one-dimensional causal convolution is fused through residual links to alleviate the gradient degradation problem in deep network training. The signal output by the Sigmoid function is fused with the signal of the other one-dimensional causal convolution through residual links and then multiplied by Hadamard. In other words, its input first goes through two independent one-dimensional causal convolutions, which generate two feature maps respectively. One feature map is transformed non-linearly by the Sigmoid function to output a gated signal with a value between 0 and 1. The other feature map is fused through residual links to alleviate the gradient degradation problem in deep network training. Then the two signals are multiplied by Hadamard, thereby achieving adaptive control and filtering of information flow in the time dimension.

[0042] The Hadamard product is expressed as: Temporal gate convolution formula: In the formula, , These are the GLU gate inputs and the Hadamard product, respectively. For the Sigmoid function; In one example of the present invention, such as Figure 4As shown, in step S50, the actual disaster situation or confirmed safety status is compared with the prediction results of the spatiotemporal graph neural network STGCN prediction model to form feedback data. Incremental learning is then performed on the STGCN model to achieve continuous self-optimization of the early warning model. This includes the following steps: S51: Visualize the predicted comprehensive threshold in the form of an early warning cloud map and output the danger level. At the same time, use the actual disaster situation or confirmed safety status as the ground truth label and compare it with the model's prediction results to form feedback data and build an incremental training database. The model prediction results include three situations: missed reporting, false reporting, or successful prediction. S52: The incremental training database is periodically imported into the Spatiotemporal Graph Neural Network (STGCN) prediction model for incremental training and fine-tuning, enabling the model to continuously adapt to the geological conditions and mining changes of specific mines and achieve self-evolution of early warning capabilities.

[0043] According to a second aspect of the present invention, an intelligent early warning system for coal and rock dynamic disasters includes: The integrated database module is configured to distribute multiple sensors in and around the mining face. Specifically, it integrates acoustic and electrical monitoring sensors, microseismic monitoring sensors, gas concentration monitoring sensors, and triaxial stress monitoring sensors to synchronously and in real time collect raw monitoring data such as acoustic and electrical monitoring data, microseismic data, gas concentration data, and triaxial stress data. The raw data is preprocessed, cleaned, denoised, and preprocessed for outliers to construct a spatiotemporally aligned multi-source heterogeneous fusion database. The early warning level classification module is configured to extract key data fragments characterizing the evolution of mining-induced stress fields, fracture fields, and gas fields based on a multi-source heterogeneous fusion database. It quantifies and constructs a composite dynamic coupling evolution index system from two dimensions: intra-field evolution indicators and inter-field coupling indicators. Combining hierarchical analysis and the coefficient of variation method, it determines the comprehensive weight of the indicators through combined weighting. Based on the comprehensive weight of the indicators and historical disaster data in the multi-source heterogeneous fusion database, it calculates the comprehensive hazard threshold K value and classifies disaster early warning levels into four levels (A, B, C, and D) from low to high, using blue, green, yellow, and red colors for identification. The digital twin model module is configured to construct a high-precision three-dimensional geomechanical model based on the geological exploration data of the mining face and the tunnel design drawings, and discretize it into grid cells. The multi-source heterogeneous fusion database and the calculated comprehensive hazard threshold K value are dynamically mapped to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model that is synchronized with the geometry, physical properties and status of the physical working face in real time. The optimal parameter adjustment module is configured to use three-dimensional grid cells as graph nodes, define edge connections through spatial adjacency relationships, construct a weighted adjacency matrix, and form a node feature matrix based on the time series data of the indicator system. It transforms the dynamically evolving digital twin sequence into a spatiotemporal graph dataset, constructs a spatiotemporal graph neural network STGCN prediction model, uses this dataset to perform supervised learning, training and validation of the model, outputs the prediction result of the comprehensive danger threshold K value, and uses a loss function and optimizer to iteratively optimize the model until the model is adjusted to the optimal parameters. The integrated hazard prediction module is configured to input new data collected in real time into the STGCN optimal model, predict the comprehensive hazard threshold of the entire working face's three-dimensional grid cells within a future time period, and feed the prediction results back to the spatiotemporal graph neural network STGCN prediction model to achieve self-optimization of the STGCN prediction model. Specifically, it visualizes the early warning cloud map on the digital twin model. At the same time, it uses the actual disaster situation or confirmed safety status as the ground truth label and compares it with the model's prediction results to form feedback data, build an incremental training database, and periodically perform incremental learning on the STGCN model to achieve continuous self-optimization of the early warning model.

[0044] This early warning system collects and combines real-time acoustic, electrical, micro-seismic, gas concentration, and triaxial stress monitoring data from the mining face and surrounding areas to construct a spatiotemporally aligned multi-source heterogeneous database. Starting from multiple physical fields, it effectively suppresses external interference, more comprehensively depicts the disaster gestation process, and overcomes the limitations of a single data source and the simplistic integration of multi-parameter fusion early warning methods.

[0045] This early warning system constructs a composite dynamic coupling evolution index system from two aspects: intra-field evolution index and inter-field coupling index. It integrates multiple physical fields such as gas field, mining stress field, and fracture field. Through the collaborative perception and dynamic fusion of multiple physical fields, it achieves accurate characterization of the coal and rock dynamic disaster incubation process.

[0046] This early warning system introduces the concept of digital twins into early warning of coal and rock dynamic disasters by constructing a digital twin model of the mining face. It creates a high-fidelity virtual digital model by three-dimensionally meshing the mining face and surrounding area, thus achieving a comprehensive and dynamic display of the mining face status and improving the intuitiveness and fidelity of the system.

[0047] This early warning system utilizes the Spatiotemporal Graph Neural Network (STGCN) prediction model, constructs a graph dataset to conduct full-process supervised learning, training, and validation of the model, and adjusts relevant model parameters using loss functions and optimizers. It profoundly reveals the evolution law of coal and rock dynamic disasters, effectively reduces the false alarm rate and false negative rate of traditional methods, improves the accuracy and reliability of disaster early warning, and provides strong support for disaster prevention and mitigation decision-making.

[0048] The early warning system establishes a closed-loop feedback and self-evolution mechanism, comparing the actual disaster situation or confirmed safety status as a true value label with the model's prediction results to form feedback data and build an incremental training database. It also performs incremental learning on the spatiotemporal graph neural network STGCN model on a regular basis, thereby achieving continuous self-optimization of the early warning model and enhancing the accuracy and reliability of long-term applications.

[0049] It should be noted that the intelligent early warning system for coal and rock dynamic disasters of the present invention can also perform any of the processing described in the previously described intelligent early warning method for coal and rock dynamic disasters, and the specific details are not repeated here.

[0050] The foregoing description, with reference to preferred embodiments, details an exemplary implementation of the intelligent early warning method and system for coal and rock dynamic disasters proposed in this invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of this invention, and various combinations can be made to the various technical features and structures proposed in this invention without exceeding the protection scope of this invention, which is determined by the appended claims.

Claims

1. A method for intelligent early warning of coal and rock dynamic disasters, characterized in that, Includes the following steps: S10: Distribute multiple sensors in the mining face and surrounding area to synchronously and in real time collect acoustic and electrical monitoring data, microseismic data, gas concentration data and triaxial stress data, and preprocess the collected raw data to build a spatiotemporally aligned multi-source heterogeneous fusion database. S20: Based on a multi-source heterogeneous fusion database, key data fragments characterizing the evolution of mining-induced stress field, fracture field, and gas field are extracted. From the dual dimensions of intra-field evolution index and inter-field coupling index, a composite dynamic coupling evolution index system is quantitatively constructed. Combining the hierarchical analysis method and the coefficient of variation method, the comprehensive weight of the index is determined by combined weighting. Based on the comprehensive weight of the index and historical disaster data in the multi-source heterogeneous fusion database, the comprehensive danger threshold K value is calculated, and the disaster warning level is divided into multiple levels from low to high. S30: Based on the geological exploration data of the mining face and the roadway design drawing, a high-precision three-dimensional geomechanical model is constructed and discretized into grid units. The multi-source heterogeneous fusion database and the calculated real-time comprehensive hazard threshold K value are dynamically mapped to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model that is synchronized in real time with the geometry, physical properties and status of the physical working face. S40: Using 3D grid cells as graph nodes, edge connections are defined through spatial adjacency relationships to construct a weighted adjacency matrix. Based on the time series data of the indicator system, a node feature matrix is ​​formed. The dynamically evolving digital twin sequence is transformed into a spatiotemporal graph dataset. A spatiotemporal graph neural network STGCN prediction model is constructed. The model is supervised, trained, and validated using this dataset. The comprehensive danger threshold K value prediction result is output. The model is iteratively optimized until the model is adjusted to the optimal parameters. S50: The newly acquired data in real time is fed into the optimal model of the spatiotemporal graph neural network STGCN prediction model to predict the comprehensive danger threshold of the three-dimensional grid cells of the entire working face in the future time period, and the prediction results are fed back to the spatiotemporal graph neural network STGCN prediction model to realize the self-optimization of the spatiotemporal graph neural network STGCN prediction model.

2. The intelligent early warning method for coal and rock dynamic disasters according to claim 1, characterized in that, In step S20, a composite dynamic coupling evolution index system is quantitatively constructed. Combining the hierarchical analysis method and the coefficient of variation method, the comprehensive weight of the index is determined by combined weighting. Based on the comprehensive weight and historical disaster data, the comprehensive danger threshold K value is calculated, and the disaster warning levels are divided from low to high. The specific steps include the following: S21: Starting from two dimensions, namely, intra-field evolution indicators and inter-field coupling indicators, a composite dynamic coupling evolution indicator system is quantitatively constructed. Among them, intra-field evolution indicators include acoustic emission amplitude indicators, electromagnetic radiation amplitude indicators, micro-seismic amplitude indicators, gas concentration indicators, and maximum principal stress indicators. Inter-field coupling indicators include stress-energy coupling coefficient indicators, crack-gas coupling strength indicators, and stress-electromagnetic coupling sensitivity indicators. Appropriate time intervals are selected according to actual conditions. S22: First, Z-Score standardization is performed on the time series data of relevant indicators, and then Sigmoid normalization is performed. The subjective weight of the indicators is determined by the hierarchical analysis method, and the objective weight of the indicators is determined by the coefficient of variation method. The comprehensive weight of the indicators is determined by combining weights. Based on the comprehensive weight and historical disaster data, the comprehensive danger threshold K value is calculated, and the disaster warning level is divided into four levels, A, B, C and D, from low to high, using blue, green, yellow and red colors for identification.

3. The intelligent early warning method for coal and rock dynamic disasters according to claim 2, characterized in that, In step S21, the specific expressions for the stress-energy coupling coefficient, the crack-gas coupling strength, and the stress-electromagnetic coupling sensitivity are as follows: Stress-energy coupling coefficient The expression is: Crack-gas coupling strength index The expression is: Stress-electromagnetic coupling sensitivity index The expression is: In the formula, Representing the A time interval, This represents the covariance between two variables. This represents the stress change sequence during the k-th time interval; This is the energy release sequence for the k-th time interval; This represents the standard deviation of X; This represents the standard deviation of Y; This represents the number of acoustic emission signal impacts after standardization and normalization. This represents the standardized and normalized value of the gas concentration. Let be the change in electromagnetic radiation intensity during the k-th time interval; This represents the stress change during the k-th time interval.

4. The intelligent early warning method for coal and rock dynamic disasters according to claim 2, characterized in that, In step S22, the comprehensive hazard threshold K value is calculated, and the disaster warning level is divided into four levels, A, B, C, and D, from low to high, using four colors: blue, green, yellow, and red. This includes the following steps: Overall hazard level A: 0≤K<0.3, blue, current status is normal, continue operation; Overall hazard level B: 0.3≤K<0.6, green, current status is low risk, intensify monitoring of the work area; Overall hazard level C: 0.6≤K<0.8, yellow, current status is medium risk, reduce mining speed and reduce number of workers; Overall hazard level D: 0.8≤K<1, red, current status is high risk. All personnel should immediately stop work and evacuate the work area, and the emergency plan should be activated.

5. The intelligent early warning method for coal and rock dynamic disasters according to claim 1, characterized in that, Step S30 specifically includes the following steps: S31: Based on the relevant geological exploration data of the tunneling face and combined with the tunnel design drawings, professional mining software is used to create a three-dimensional solid model of the geological structure and construct a high-precision three-dimensional geomechanical model. S32: Perform spatial discretization operation to divide the entire three-dimensional geomechanical model into regular three-dimensional mesh units, determine the mesh size according to the actual situation, and number the three-dimensional mesh units in sequence. S33: The multi-source heterogeneous fusion database and the calculated real-time comprehensive danger threshold K value are dynamically mapped to the corresponding three-dimensional mesh unit through a spatial interpolation algorithm to construct a digital twin model that is synchronized in real time with the geometry, physical properties and status of the physical working surface. A professional three-dimensional visualization engine is used to develop a visualization interface for the digital twin.

6. The intelligent early warning method for coal and rock dynamic disasters according to claim 1, characterized in that, In step S33, the data is dynamically mapped to the corresponding three-dimensional mesh element using a spatial interpolation algorithm, including the following steps: Kriging spatial interpolation is used to fill in the grid cell data that cannot be covered by sensors. Based on limited sensor data, the dynamic properties of grid cells in areas where no sensors are deployed are calculated to form a continuous cloud map. The specific expression is as follows: Kriging interpolation formula: Unbiased estimation conditions: In the formula, For point The estimated value; n This represents the total number of known monitoring points in the surrounding area involved in the calculation; z is the weighting coefficient; i This represents the true value of the i-th known monitoring point distributed around the target point; For point The true value; Among them, the weighting coefficient The value of an unknown point is estimated by weighted summation of data from all known points in space, satisfying the condition that the unknown point... The estimated value at the location Compared with the true value The set of optimal coefficients that minimizes the difference, i.e. Meanwhile, the Kriging spatial interpolation method satisfies the unbiased estimation condition in its calculation process to ensure the macroscopic accuracy of spatial prediction.

7. The intelligent early warning method for coal and rock dynamic disasters according to claim 1, characterized in that, Step S40 specifically includes the following steps: S41: Construct a spatial topology graph, using 3D mesh cells as graph nodes, and build a weighted adjacency matrix based on the spatial relationships between nodes. A node feature matrix is ​​formed based on the time series data of the indicator system. The feature matrix consists of normalized index data and a comprehensive risk threshold K. The dynamically evolving digital twin sequence is transformed into a spatiotemporal graph dataset, and divided into training, validation and test sets. S42: Construct an STGCN early warning model containing a spatiotemporal convolution module, and optimize the model using the mean squared error loss function and the RMSprop optimizer; S43: Import the training set into the model for supervised learning, training and validation, output the comprehensive danger threshold K value, iteratively optimize the model using the validation set, loss function and optimizer, evaluate the model performance using the test set, and determine the optimal model with the best parameters.

8. The intelligent early warning method for coal and rock dynamic disasters according to claim 7, characterized in that, In step S42, the STGCN early warning model comprises two spatiotemporal convolutional modules and an output layer. Each spatiotemporal convolutional module is composed of alternating layers of temporal gate convolution, spatial graph convolution, and temporal gate convolution. The output layer includes a temporal convolutional layer and a fully connected layer. The temporal gate convolution is configured to capture the dynamic changes of data in the time dimension. The spatial graph convolution is configured to capture the dependencies of data in the spatial dimension. The temporal gate convolution includes two one-dimensional causal convolutions, a Sigmoid function, and a Hadamard product. The two one-dimensional causal convolutions are connected in parallel. One of the one-dimensional causal convolutions is concatenated with the Sigmoid function. The other one-dimensional causal convolution is fused through residual links to alleviate the gradient degradation problem in deep network training. The signal output by the Sigmoid function is fused with the signal of the other one-dimensional causal convolution through residual links and then multiplied by Hadamard.

9. The intelligent early warning method for coal and rock dynamic disasters according to claim 1, characterized in that, In step S50, the actual disaster situation or confirmed safety status is compared with the model's prediction results to form feedback data. This data is then used to incrementally learn the spatiotemporal graph neural network (STGCN) prediction model, enabling continuous self-optimization of the early warning model. This includes the following steps: S51: Visualize the predicted comprehensive threshold in the form of an early warning cloud map and output the danger level. At the same time, use the actual disaster situation or confirmed safety status as the ground truth label and compare it with the model's prediction results to form feedback data and build an incremental training database. The model prediction results include three situations: missed reporting, false reporting, or successful prediction. S52: The incremental training database is periodically imported into the Spatiotemporal Graph Neural Network (STGCN) prediction model for incremental training and fine-tuning, enabling the model to continuously adapt to the geological conditions and mining changes of specific mines and achieve self-evolution of early warning capabilities.

10. An intelligent early warning system for coal and rock dynamic disasters, characterized in that, include: The integrated database module is configured to distribute multiple sensors in and around the mining face to synchronously and in real time collect acoustic and electrical monitoring data, microseismic data, gas concentration data and triaxial stress data, and preprocess the collected raw data to build a spatiotemporally aligned multi-source heterogeneous integrated database. The early warning level classification module is configured to extract key data fragments characterizing the evolution of mining-induced stress field, fracture field, and gas field based on a multi-source heterogeneous fusion database. It quantitatively constructs a composite dynamic coupling evolution index system from two dimensions: intra-field evolution index and inter-field coupling index. Combining hierarchical analysis and the coefficient of variation method, it achieves combined weighting to determine the comprehensive weight of the index. Based on the comprehensive weight of the index and historical disaster data in the multi-source heterogeneous fusion database, it calculates the comprehensive danger threshold K value and classifies the disaster early warning level into multiple levels from low to high. The digital twin model module is configured to construct a high-precision three-dimensional geomechanical model based on the geological exploration data of the mining face and the tunnel design drawings, and discretize it into grid cells. The multi-source heterogeneous fusion database and the calculated real-time comprehensive hazard threshold K value are dynamically mapped to the corresponding grid through a spatial interpolation algorithm to construct a digital twin model that is synchronized with the geometry, physical properties and status of the physical working face in real time. The optimal parameter adjustment module is configured to use three-dimensional grid cells as graph nodes, define edge connections through spatial adjacency relationships, construct a weighted adjacency matrix, and form a node feature matrix based on the time series data of the indicator system. It transforms the dynamically evolving digital twin sequence into a spatiotemporal graph dataset, constructs a spatiotemporal graph neural network STGCN prediction model, uses this dataset to perform supervised learning, training and validation of the model, outputs the prediction result of the comprehensive danger threshold K value, and iteratively optimizes the model until the model is adjusted to the optimal parameters. The integrated hazard prediction module is configured to input new data collected in real time into the optimal model of the spatiotemporal graph neural network STGCN prediction model, predict the integrated hazard threshold of the three-dimensional grid cells of the entire working face in the future time period, and feed the prediction results back to the spatiotemporal graph neural network STGCN prediction model to realize the self-optimization of the spatiotemporal graph neural network STGCN prediction model.

Images