Coal mine underground gas prevention and control method, device, equipment and medium

Abstract

The application discloses a coal mine underground gas prevention and treatment method, device, equipment and medium. Multi-source sensing data of a target area in a coal mine is acquired, spatio-temporal alignment and standardized fusion are performed to generate a spatio-temporal data matrix, a pre-trained GNN-LSTM hybrid model is input, and gas emission prediction data of each monitoring point within a preset time window is output. The prediction data and the running state of the ventilation system are jointly input into a pre-trained deep reinforcement learning decision model, the optimal ventilation control instruction is determined from a preset action set based on multi-objective optimization, and the instruction is sent to a ventilation control equipment to adjust the frequency of a fan or the opening degree of a damper. Through the fusion of multi-source information, dynamic and accurate prediction and intelligent decision, advanced warning and automatic control of gas prevention and treatment are realized, the frequency and duration of gas over-limit are effectively reduced, gas accumulation is prevented from the source, and the safety level of coal mine production is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety production technology, and in particular to a method, device, equipment and medium for preventing and controlling underground gas in coal mines. Background Technology

[0002] Gas hazards pose a significant threat to coal mine safety. Currently, relevant technologies for underground gas control in coal mines primarily involve deploying fixed-point gas sensors at key locations such as the working face and return airway to monitor gas concentration in real time. Upon triggering an alarm for excessive gas levels, ventilation fans and dampers are manually adjusted to manage the gas. However, existing methods only activate ventilation control after gas concentration exceeds limits, resulting in a significant delay in response and a lack of proactive and accurate prediction of gas outburst trends, leading to poor predictability. Furthermore, ventilation control strategies often rely on manual adjustments based on field experience, failing to achieve optimal control that matches the location of the gas outburst source and the airflow characteristics of the roadway network. This severely restricts the safety and production efficiency of underground coal mine operations and increases the risk of gas accumulation leading to disasters. Summary of the Invention

[0003] In view of this, the present invention provides a method, device, equipment and medium for the prevention and control of underground gas in coal mines, in order to solve the technical problem that existing gas prevention and control methods are seriously outdated and have poor correlation with the underground roadway network, resulting in high safety risks in coal mine production.

[0004] Firstly, a method for preventing and controlling underground gas in coal mines is provided, the method comprising: Acquire multi-source sensing data from multiple preset monitoring points in the target area of ​​a coal mine. The multi-source sensing data includes at least one of the following: environmental data, mining operation status data, geological condition data, and ventilation system operation status data. Alignment and standardization fusion of multi-source sensing data are performed to generate a spatiotemporal data matrix; The spatiotemporal data matrix is ​​input into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point in the target area within a preset time window. The gas outburst prediction data and the ventilation system operation status data are input into a pre-trained deep reinforcement learning decision model. Based on a preset multi-objective optimization function, the optimal action combination is determined from the preset action set of the ventilation system as the ventilation control instruction corresponding to the target area. Ventilation control commands are sent to ventilation control equipment to adjust the operating frequency of the ventilator and / or the opening of the damper.

[0005] Secondly, a coal mine underground gas control device is provided, the device comprising: The acquisition module is used to acquire multi-source sensing data from multiple preset monitoring points in the target area of ​​the coal mine. The multi-source sensing data includes at least one of the following: environmental data, mining operation status data, geological condition data, and ventilation system operation status data. The first generation module is used to align and standardize multi-source sensing data to generate a spatiotemporal data matrix. The second generation module is used to input the spatiotemporal data matrix into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point in the target area within a preset time window. The determination module is used to input gas outburst prediction data and ventilation system operation status data into a pre-trained deep reinforcement learning decision model. Based on a preset multi-objective optimization function, it determines the optimal action combination from the preset action set of the ventilation system as the ventilation control instruction corresponding to the target area. The sending module is used to send ventilation control commands to the ventilation control equipment to adjust the operating frequency of the fan and / or the opening degree of the damper.

[0006] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned coal mine underground gas prevention method.

[0007] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the aforementioned method for preventing and controlling underground gas in coal mines.

[0008] The aforementioned method, device, equipment, and medium for underground gas control in coal mines acquire multi-source sensing data from multiple preset monitoring points in the target area of ​​the coal mine. This multi-source sensing data undergoes spatiotemporal alignment and standardization fusion to generate a spatiotemporal data matrix. This matrix is ​​then input into a pre-trained gas emission prediction model to generate gas emission prediction data for each preset monitoring point within a preset time window. Subsequently, the predicted data, along with ventilation system operation status data, is input into a pre-trained deep reinforcement learning decision model. Based on a preset multi-objective optimization function, the optimal action combination is determined from the preset action set of the ventilation system as a ventilation control command. This command is then sent to the ventilation control equipment to adjust the fan frequency and / or damper opening. By fusing multi-source information and using a GNN-LSTM hybrid model to dynamically and accurately predict gas emission, and employing deep reinforcement learning for multi-objective optimization decision-making to automatically generate the optimal ventilation control strategy, advanced prediction and intelligent control of gas control are achieved. This effectively reduces the frequency and duration of gas exceedances, prevents gas accumulation at the source, and thus improves the safety level of coal mine production. Attached Figure Description

[0009] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic flowchart of a method for preventing and controlling underground gas in coal mines according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a coal mine underground gas control device in one embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the embodiments 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. It should be understood that the accompanying drawings in the present invention are only for illustrative and descriptive purposes and are not intended to limit the scope of protection of the present invention.

[0011] Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or performed simultaneously. Moreover, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0012] Furthermore, the embodiments described herein are merely some, not all, of the embodiments of the invention. The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0013] It should be noted that the term "comprising" will be used in the embodiments of the present invention to indicate the presence of a feature subsequently declared, but does not exclude the addition of other features. It should also be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0014] Gas hazards pose a significant threat to coal mine safety. Traditional gas control relies primarily on monitoring by fixed-point gas sensors and localized ventilation control based on human experience. However, traditional gas control methods cannot proactively and accurately predict gas outburst trends. Ventilation control typically only initiates after gas levels exceed limits, resulting in a delayed response. Furthermore, adjustment strategies largely depend on experience, failing to achieve dynamic, precise, and optimal control tailored to the gas outburst source and roadway network characteristics. In addition, multi-source data such as gas concentration, wind speed, equipment status, and geological information are independent of each other. Traditional fixed-point gas sensor monitoring lacks in-depth fusion analysis, making it difficult to form a holistic understanding of the underground safety situation.

[0015] Based on the above problems, this application proposes a method for underground gas control in coal mines. By integrating multi-source information, it uses a GNN-LSTM hybrid model to dynamically and accurately predict gas outbursts, and employs deep reinforcement learning for multi-objective optimization decision-making to automatically generate the optimal ventilation control strategy. This achieves advanced prediction and intelligent control of gas control, effectively reducing the frequency and duration of gas exceedances, preventing gas accumulation from the source, and thus improving the safety level of coal mine production.

[0016] The following is a detailed description of this case, in conjunction with the relevant accompanying drawings in the instruction manual.

[0017] Please see Figure 1 This description and embodiment provides a method for preventing and controlling underground gas in coal mines, specifically including the following steps: S10: Acquire multi-source sensing data from multiple preset monitoring points in the target area of ​​the underground coal mine; The multi-source sensing data includes at least one of the following: environmental data, mining operation status data, geological condition data, and ventilation system operation status data.

[0018] It is understood that the executing entity of this invention can be a coal mine underground gas control device, a terminal, or a server; no specific limitation is made here. This embodiment of the invention will be described using a server as an example.

[0019] In this step, the target area refers to the underground space within the coal mine mining environment that requires gas monitoring and ventilation control, dynamically defined based on the mining operation plan, historical gas emission patterns, or real-time risk assessment results. Examples include a specific mining face, return airway section, or local ventilation network area. Multiple pre-set monitoring points refer to the spatial locations of sensors or monitoring equipment pre-selected and fixed within the target area. Examples include gas sensor installation points at the mining face, wind speed monitoring points in the return airway, gas detection points in the upper corner, and temperature and pressure sensor deployment points in the main intake airway. Each monitoring point has a unique geospatial identifier (such as GIS coordinates or roadway mileage calibration values) and serves as a graph node in the subsequent graph neural network modeling.

[0020] During coal mine production and mining, multi-source sensing data is acquired in real time in the target area underground by sensors and monitoring equipment deployed at various preset monitoring points. The multi-source sensing data includes at least one of the following: environmental data (such as gas concentration, wind speed, temperature, and air pressure), mining operation status data (such as the location, operating speed, cutting status, and tunneling machine progress of the coal mining machine), geological condition data (such as coal seam thickness, fault and fold structure information measured while drilling), and ventilation system operation status data (such as the operating frequency, damper opening, and air volume of local ventilation fans and main ventilation fans).

[0021] Optionally, the target area can be dynamically delineated based on the mining operation plan, historical gas emission patterns, geological structure distribution, and ventilation network structure. This includes, but is not limited to, one or more of the following: operating mining faces (such as fully mechanized mining faces and tunneling faces), return airways, upper corners, main intake airways, local roadway sections with potential gas accumulation risks, and key areas affected by geological structures such as faults and folds. The target area can be a single independent monitoring unit (e.g., a single coal mining face) or a local ventilation network area composed of multiple roadways and several monitoring nodes.

[0022] In practical applications, an IoT sensing network comprised of distributed gas sensors, wind speed sensors, equipment status sensors, and ground-penetrating radar deployed in the target area collects multi-source sensing data in real time and transmits it via converged communication networks such as industrial Ethernet ring networks and 5G / WiFi 6. Environmental data includes, but is not limited to, gas concentration, wind speed, temperature, and air pressure; mining operation status data includes, but is not limited to, coal mining machine location, operating speed, cutting status, and tunneling machine progress; geological condition data includes, but is not limited to, coal seam thickness, fault and fold structure information obtained through drilling measurements; and ventilation system operation status data includes, but is not limited to, the operating frequency, damper opening, and air volume of local and main ventilation fans.

[0023] S20: Align and standardize the multi-source sensing data and fuse them to generate a spatiotemporal data matrix.

[0024] In this step, the sampling frequencies of different sensors vary, with some collecting data once per second, and others once every few seconds or minutes. This results in uneven distribution of data across the time axis and inability to directly align timestamps. Furthermore, the spatial reference systems of the sensors are inconsistent. For example, gas sensors are typically described by their equipment number or installation location, coal mining machine locations are represented by roadway mileage, and geological data uses a geodetic coordinate system. These data from different sources lack a unified spatial reference. In addition, environmental, production, geological, and ventilation data naturally exhibit complex spatiotemporal correlations, but the raw, isolated, discrete, and heterogeneous data formats cannot directly reflect these correlations. Directly using unprocessed raw data for subsequent analysis may fail to correctly understand the gas diffusion relationships between different monitoring points, leading to decreased accuracy in gas outburst prediction and decision failure. Therefore, this application proposes spatiotemporal alignment and fusion processing of the collected multi-source sensing data. This unifies the scattered, asynchronous, and heterogeneous raw data onto the same time reference and spatial grid, forming a structured spatiotemporal data matrix. This provides high-quality, consistent information rich in spatiotemporal correlations for subsequent gas outburst prediction.

[0025] In one embodiment of this application, a specific spatiotemporal data matrix generation scheme is provided. In S20, that is, aligning and standardizing the multi-source sensing data to generate a spatiotemporal data matrix, the following steps S21-S24 are specifically included: S21: Perform data cleaning on each data item in the multi-source sensing data.

[0026] In this step, due to the complex underground environment of coal mines, severe electromagnetic interference, and the potential drift or malfunction of the sensors themselves, the raw data inevitably contains outliers exceeding reasonable ranges (such as negative gas concentrations or values ​​far exceeding physical limits) as well as random noise caused by equipment vibration, communication errors, etc. To avoid invalid data adversely affecting subsequent spatiotemporal alignment, fusion processing, and model prediction, data cleaning is performed on all collected raw data to remove outliers and noise.

[0027] Optionally, reasonable threshold ranges for each physical quantity can be set based on domain knowledge, and data exceeding the threshold can be marked as anomalies and removed. Outlier detection algorithms based on statistical methods (such as interquartile range or standard deviation) can be used to identify and process data points that deviate significantly from the overall distribution. For isolated, unrelated abrupt noise, median filtering or moving average filtering can be used for smoothing or direct removal.

[0028] By employing the above methods, we can ensure that the multi-source sensing data participating in subsequent processing has high reliability and consistency, laying the foundation for generating a high-quality fused data matrix.

[0029] S22: Based on the geospatial location labels corresponding to each preset monitoring point, perform spatial registration on the cleaned multi-source sensing data, and map the data under different coordinate systems to the preset underground roadway network to obtain spatially aligned data.

[0030] In this step, each pre-set monitoring point is assigned a unique geospatial location label upon system deployment. This label can be coordinates in the roadway geographic information system, roadway mileage calibration values, or node numbers. Because data from different sources uses different spatial reference systems—for example, gas sensors are associated with pre-calibrated roadway locations via their device numbers, coal mining machine locations are represented by mileage values ​​from the roadway opening, and geological data uses absolute geodetic coordinates—these data cannot be directly analyzed jointly within the same spatial framework. Therefore, this application proposes to unify all data onto corresponding nodes or roadways in a pre-set underground roadway network by parsing the location label carried by each monitoring point and utilizing pre-set coordinate transformation relationships. This pre-set underground roadway network is a pre-constructed digital spatial model containing information such as the roadway topology, geometric dimensions, node connections, and airflow direction. After spatial registration, data originally scattered across different coordinate systems are aligned to the same spatial reference, forming spatially aligned data where each pre-set monitoring point corresponds one-to-one with its spatial location.

[0031] In practical applications, a spatial index for a pre-defined underground roadway network is first established. This index, based on roadway geographic information, includes the centerline coordinates, cross-sectional dimensions, node connections, and mileage calibration information for each roadway. Then, for each pre-defined monitoring point, the spatial information contained in its location label is parsed, and a pre-defined coordinate transformation algorithm (such as linear mapping or nonlinear registration algorithm) is used to map it to the corresponding location point in the pre-defined underground roadway network. For example, the device number of the gas sensor is associated with the pre-calibrated roadway coordinates, the mileage value of the coal mining machine is projected along the roadway centerline into geographic information system coordinates, and the geodetic coordinates of the geological data are transformed into the roadway network coordinate system. After mapping, the data of each pre-defined monitoring point is assigned unified roadway network coordinates, meaning each data point is associated with a unique spatial location node in the pre-defined underground roadway network, thus forming spatially aligned data. In this spatially aligned data, data from different sources and coordinate systems have been transformed to the same spatial reference, providing a unified spatial reference framework for subsequent feature fusion.

[0032] S23: Based on the timestamps of each data in the cleaned multi-source sensing data, perform time synchronization operation according to the preset time axis to align the data to the same time reference point and obtain the time series data of each preset monitoring point.

[0033] In this step, the sampling frequencies of various sensors and monitoring equipment in coal mines differ. For example, a gas sensor may collect data every 5 seconds, a wind speed sensor every 1 second, and the position of the coal mining machine is only updated when the equipment moves. Geological data may be static data at the hourly or even daily level, resulting in uneven distribution of timestamps and inconsistent time bases among the data points, making direct time-series correlation analysis impossible. Therefore, this application proposes first setting a unified time interval (e.g., 1 second, 5 seconds, or 1 minute) based on the actual needs of underground coal mine safety monitoring, as the basic resolution of the preset interval time axis. Then, for each preset monitoring point, the timestamps of all data points at that monitoring point are matched with the preset interval time axis: if measured data exists at a certain time base point, that data is directly used; otherwise, the approximate data value of the base point is estimated using measured data from adjacent times before and after the base point, employing methods such as linear interpolation or nearest neighbor interpolation. Through the above operations, the data from each preset monitoring point is mapped onto a unified time axis, forming a data sequence with equal time intervals. By performing the above time synchronization operation on all preset monitoring points, the time series data corresponding to each preset monitoring point can be obtained, providing a regular input for the subsequent extraction of time dynamic features by the Long Short-Term Memory Network.

[0034] S24: Link and integrate spatially aligned data and time-series data to generate a spatiotemporal data matrix.

[0035] In this step, spatial alignment data describes the fixed spatial location of each preset monitoring point in the underground roadway network and its multi-source sensing characteristic values ​​(such as gas concentration, wind speed, temperature, etc.), while time series data records the sequence of characteristic value changes for each preset monitoring point at different times. The spatial and temporal dimensions are unified into a single data structure: using the preset monitoring points as spatial indices and the preset time interval as temporal indices, the characteristic values ​​of each monitoring point at each time reference point are filled into the corresponding matrix cells, forming a three-dimensional spatiotemporal data matrix. In this matrix, each row represents the feature vector of all monitoring points at the same time, and each column represents the sequence of feature values ​​of the same monitoring point at different times, thus achieving the organic integration of spatial and temporal information. This spatiotemporal data matrix preserves both the topological connections of each monitoring point in the roadway network (mapped to the adjacency matrix through the monitoring point index) and the dynamic patterns of each monitoring point's evolution over time, providing standardized and structured input data for subsequent gas outburst prediction models.

[0036] S30: Input the spatiotemporal data matrix into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point in the target area within a preset time window.

[0037] In this step, the gas emission prediction model is used to characterize the nonlinear mapping relationship between multi-source sensing data and future changes in gas emission volume and concentration, thereby achieving advanced prediction of gas emission trends. The fused data is input into the pre-trained gas emission prediction model. After forward calculation by the model, the gas concentration change curves and emission volume numerical sequences of each preset monitoring point in the target area within a preset time period are output as gas emission prediction data.

[0038] Optionally, the preset time window can be configured according to the actual working conditions and safety production needs downhole. It is usually set to 15 minutes, 30 minutes or 60 minutes in the future, which ensures that the prediction results have sufficient advance warning time, while taking into account the real-time calculation and prediction accuracy of the model.

[0039] In one embodiment of this application, a specific scheme for generating gas outburst prediction data is provided. In S30, the spatiotemporal data matrix is ​​input into a pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point in the target area within a preset time window. This specifically includes the following steps S31-S34: S31: Input the spatiotemporal data matrix into the pre-trained gas outburst prediction model. Through graph neural network, based on the roadway topology connection relationship of the preset underground roadway network, perform spatial information aggregation on the spatially aligned data, extract the spatial features of gas diffusion at each preset monitoring point, and generate a spatial feature map containing the roadway topology connection relationship.

[0040] In this step, the spatiotemporal data matrix is ​​input into a pre-trained gas outburst prediction model. First, a Graph Neural Network (GNN) is used to aggregate spatial information from the spatially aligned data based on the topological connections of a pre-defined underground roadway network. The GNN is based on the topological structure of this network, which uses pre-defined monitoring points as nodes, roadway connections as edges, and includes attributes such as airflow direction, roadway length, and wind resistance. The GNN aggregates information from each node's neighbors through a message-passing mechanism: for each pre-defined monitoring point, the network collects feature data from its neighboring monitoring points and fuses them according to a pre-defined aggregation function (such as mean aggregation, attention-weighted aggregation, or gated aggregation) to update the current node's feature representation. Through multi-layer graph convolution operations, the feature representation of each node gradually integrates information over a larger spatial range, enabling monitoring points far from the gas outburst source to perceive the changing trends in upstream gas concentration. Finally, the GNN outputs a spatial feature map containing the roadway topological connections. This spatial feature map reflects the spatial propagation pattern of gas in the underground roadway network, such as the impact of upstream gas outbursts on downstream monitoring points, providing a spatial dimension representation for subsequent spatiotemporal feature fusion.

[0041] S32: Through the cyclic gating mechanism of the Long Short-Term Memory Network, extract the time dynamic features of the gas concentration evolution over time at each preset monitoring point in the time series data, and generate the time series hidden state of each preset monitoring point.

[0042] In this step, Long Short-Term Memory (LSTM) is a variant of recurrent neural network specifically designed for processing sequential data. Its core lies in introducing three gating units: an input gate, a forget gate, and an output gate, along with a memory cell state. For each preset monitoring point, its time-series data (e.g., gas concentration, wind speed, temperature, etc., over the past few minutes) is sequentially input into the LSTM network step by step. At each time step, the input gate controls whether new information from the current moment is written to the memory cell; the forget gate determines which information from the previous moment's memory cell needs to be discarded; and the output gate generates the hidden state output for the current moment based on the current memory cell state. Through this gating mechanism, the LSTM network can selectively retain long-term dependent information (such as the overall trend of a slow increase in gas concentration) while ignoring short-term random fluctuations (such as instantaneous jumps caused by electromagnetic interference). After multiple time steps of iteration, the network finally outputs a fixed-dimensional hidden state vector for each preset monitoring point. This hidden state contains the rising, falling, or fluctuating trend of the gas concentration at each preset monitoring point over time, providing a temporal dimension representation for subsequent spatiotemporal fusion.

[0043] S33: Perform feature fusion between the spatial feature map and the latent state of the time series to obtain the spatiotemporal fusion feature tensor.

[0044] In this step, after obtaining the spatial feature map and the time-series hidden state respectively, the two are fused to obtain a spatiotemporal fused feature tensor. The spatial feature map mainly depicts the spatial diffusion law of gas in the underground roadway network based on topological connections, reflecting the degree to which each preset monitoring point is affected by its neighboring nodes and upstream airflow. The time-series hidden state records the dynamic characteristics of the gas concentration evolution of each preset monitoring point over time. Since the gas emission and migration process is inherently spatiotemporally coupled—that is, the change in gas concentration at a certain location is affected by both the spatial propagation from the upstream gas emission source and the temporal continuity of the location's own historical state—neither spatial nor temporal features alone can fully describe this complex process. Feature fusion organically combines the information from the spatial and temporal dimensions, enabling the fused feature tensor to simultaneously possess both the spatial correlation of its origin and the temporal regularity of its changes over time.

[0045] In practical applications, methods such as splicing, element-wise addition, or attention-based weighted fusion are employed to align and merge the spatial feature map output by the graph neural network with the temporal series hidden states output by the long short-term memory network in terms of feature dimensions, resulting in a three-dimensional spatiotemporal fusion feature tensor. Each element of this tensor corresponds to the fusion feature vector of a specific monitoring point at a specific time step, providing a complete representation of both spatial topological dependence and temporal evolution patterns for subsequent fully connected regression layers, thereby supporting accurate prediction of gas outburst volume.

[0046] S34: Based on the spatiotemporal fusion feature tensor, determine the predicted values ​​of gas concentration and gas emission rate of each preset monitoring point in the target area within a preset time window, and use them as gas emission prediction data.

[0047] In this step, after obtaining the spatiotemporal fusion feature tensor, it is input into one or more fully connected regression layers (e.g., two- or three-layer fully connected neural networks). These regression layers perform nonlinear mapping and dimensionality transformation on the input spatiotemporal fusion features, gradually compressing the high-dimensional spatiotemporal features to an output dimension that matches the prediction target. For each preset monitoring point, the regression layer outputs the predicted gas concentration value at each discrete time step (e.g., every 1 minute or 5 minutes) within a preset future time window, thus forming a continuous time series prediction curve. Simultaneously, by integrating, averaging, or directly accumulating the predicted gas concentration values ​​within this time window, the predicted gas emission rate (e.g., in cubic meters or cubic meters per minute) for that monitoring point within the corresponding time window can be obtained. The aforementioned predicted gas concentration and gas emission rate together constitute the gas emission prediction data. This prediction data not only provides early warning of which monitoring points may reach or exceed safety thresholds at what time, but also quantifies the scale of the emission, providing quantitative risk information for subsequent deep reinforcement learning decision-making models to formulate refined ventilation control strategies.

[0048] Optionally, the preset time window can be set according to the actual needs of the dynamic response time of gas outburst, the response speed of ventilation control, and the early warning of safe production, such as the next 15 minutes, 30 minutes, or 60 minutes. This application does not make specific limitations on this.

[0049] By integrating multi-dimensional dynamic information such as geology and production data, and employing a GNN-LSTM hybrid model, the spatiotemporal prediction accuracy of gas outbursts under complex conditions has been significantly improved, realizing the transformation from passive monitoring to proactive early warning.

[0050] In one embodiment of this application, a specific gas outburst prediction model training scheme is provided. Before inputting the spatiotemporal data matrix into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point within a preset time window in the target area, the scheme further includes the following steps: A hybrid model of graph neural network and long short-term memory network is constructed, and the network parameters are randomly initialized to obtain the initialized hybrid model; Acquire historical multi-source sensing data and historical gas emission data of underground coal mine areas within a preset historical time period; Historical multi-source sensing data is aligned, standardized, and fused to generate a historical spatiotemporal data matrix. Using historical spatiotemporal data matrix as input and historical gas outburst data as supervision labels, the initialized hybrid model is trained through supervised learning. The network parameters are adjusted until the preset convergence condition is met, resulting in a pre-trained gas outburst prediction model.

[0051] In this embodiment, firstly, a hybrid model of graph neural network and long short-term memory network is constructed. This hybrid model uses graph neural network to model the topology of the underground tunnel network and the spatial dependencies between monitoring points, and uses long short-term memory network to extract the time-series dynamic features of each monitoring point. A feature fusion layer is set at the end of the network to concatenate or weightedly fuse spatial and temporal features to form a complete prediction network architecture. After construction, the parameters in the network (including the weight matrix of the graph neural network, the input gate, forget gate, output gate, and weights and biases of memory cells in the long short-term memory network, as well as the parameters of the fully connected regression layer) are randomly initialized, for example, using Xavier initialization or normal distribution random initialization, thus obtaining the initialized hybrid model.

[0052] Secondly, historical multi-source sensing data and corresponding historical gas emission data for multiple areas of the entire coal mine over a historical time period are acquired. The historical multi-source sensing data is consistent with the data type of the real-time multi-source sensing data, including at least one of environmental data, mining operation status data, geological condition data, and ventilation operation data. Its time span can be set according to actual needs, such as covering continuous monitoring data from the past month or quarter. Historical gas emission data serves as a monitoring label, which can be the actual measured gas concentration sequence and emission value at each monitoring point at the corresponding historical moment. Then, the historical multi-source sensing data undergoes alignment and standardization fusion processing, using the same spatiotemporal alignment and fusion methods as the real-time data (including data cleaning, spatial registration, time synchronization, and matrix integration) to generate a historical spatiotemporal data matrix. This historical spatiotemporal data matrix has the same structure as the real-time spatiotemporal data matrix, containing a time axis, monitoring point indexes, and multi-dimensional feature values.

[0053] Finally, using the historical spatiotemporal data matrix as input and historical gas outburst data as supervisory labels, supervised learning training is performed on the initialized hybrid model of graph neural network and long short-term memory network. During training, a mini-batch gradient descent algorithm (such as the Adam optimizer) is employed, using mean squared error or mean absolute error as the loss function, and the network parameters are iteratively updated through backpropagation. Training stops when the loss function value converges to a preset threshold or reaches a preset number of training epochs, and the current model parameters are saved as a pre-trained gas outburst prediction model.

[0054] In one embodiment of this application, a specific gas outburst prediction model parameter update scheme is provided, which involves inputting a spatiotemporal data matrix into a pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point within a preset time window in the target area, and further including the following steps: Acquire real-time multi-source sensing data and actual gas concentration change data fed back after the gas emission prediction model performs prediction; Compare the actual gas concentration change data with the gas emission prediction data to determine whether there is a discrepancy between the gas emission prediction data and the actual gas concentration change data; When there is a discrepancy between the predicted gas outburst data and the actual gas concentration change data, real-time multi-source sensing data and actual gas concentration change data are used as incremental training samples, and the backpropagation algorithm is used to update the network parameters of the gas outburst prediction model.

[0055] In this embodiment, real-time multi-source sensing data is acquired, along with actual gas concentration changes fed back by downhole sensors after the gas emission prediction model performs its predictions. Whenever new monitoring data arrives, it is determined whether there is a deviation between the model's predicted future gas emission data and subsequent actual monitoring values. If a deviation is confirmed, this input-output pair is used as a new training sample. The model calculates the prediction error based on this sample and uses the backpropagation algorithm to calculate the gradient, updating the parameters of the graph neural network and long short-term memory network with one or more gradient descent iterations at a relatively small learning rate.

[0056] Through the above methods, the gas outburst prediction model can adapt to dynamic factors such as the progress of mining, changes in geological conditions, and aging of the ventilation system, and continuously maintain prediction accuracy.

[0057] In one embodiment of this application, a specific early warning scheme for excessive gas concentration is provided. This scheme involves inputting a spatiotemporal data matrix into a pre-trained gas emission prediction model to generate gas emission prediction data for each preset monitoring point within a preset time window in the target area. The scheme further includes the following steps: Compare the predicted gas concentration values ​​of each preset monitoring point in the gas emission prediction data with the preset gas concentration threshold, and determine whether the predicted gas concentration value of each preset monitoring point exceeds the preset gas concentration threshold at at least one time point within the preset time window. For any preset monitoring point, when the predicted gas concentration value at at least one time point within its preset time window exceeds the preset gas concentration threshold, the location of the monitoring point exceeding the limit, the time point exceeding the limit, and the predicted value of the exceeding limit are obtained. Based on the location of the monitoring points exceeding the limit, the time of exceeding the limit, and the predicted value of the concentration exceeding the limit, an early warning message is generated; The warning information will be sent to the terminals of relevant personnel.

[0058] In this embodiment, after generating gas outburst prediction data, the predicted gas concentration values ​​of each preset monitoring point in the prediction data are compared with preset gas concentration thresholds to determine whether the predicted gas concentration value of each preset monitoring point exceeds the threshold at at least one time point within a preset time window. For each preset monitoring point, if the predicted gas concentration value exceeds the preset gas concentration threshold at at least one time point within the preset time window, the location of the exceeding monitoring point (e.g., roadway name, mileage marker, or GIS coordinates), the exceeding time point (i.e., the specific time predicted to occur), and the predicted exceeding concentration value are automatically obtained. Then, based on the obtained location of the exceeding monitoring point, the exceeding time point, and the predicted exceeding concentration value, a warning message is generated. Finally, the warning message is sent to the terminals of relevant personnel (e.g., the handheld terminals of underground miners, the computers of the safety monitoring center, or the mobile phones of on-duty personnel) through the underground broadcasting system, mobile communication network, or monitoring platform, and simultaneously triggers an audible and visual alarm device to remind relevant personnel to take timely countermeasures such as local ventilation enhancement, mining operation adjustment, or personnel evacuation.

[0059] Through the above methods, early warning of gas disasters was achieved, significantly improving the proactive prevention and control capabilities of coal mine gas disasters.

[0060] S40: Input the gas outburst prediction data and the ventilation system operation status data into the pre-trained deep reinforcement learning decision model. Based on the preset multi-objective optimization function, determine the optimal action combination from the preset action set of the ventilation system as the ventilation control instruction corresponding to the target area.

[0061] In this step, the generated gas outburst prediction data and the current operating status data of the ventilation system are input into a pre-trained Deep Reinforcement Learning (DRL) decision model. After receiving the input, the DRL decision model determines the optimal action combination from a preset action set of the ventilation system based on a pre-defined multi-objective optimization function. The preset action set refers to all control operation options available to the DRL decision model, including the adjustable frequency settings of each ventilator (e.g., 1Hz or 5Hz per setting) and the adjustable opening settings of each damper (e.g., 5% or 10% per setting). The model calculates the expected cumulative reward for each candidate action combination under the current input through forward propagation and selects the optimal action combination that maximizes the reward value (e.g., the combination of "increasing the frequency of local ventilator #3 by 15%" and "decreasing the opening of damper #106 by 10%"). Finally, the selected optimal action combination is used as the ventilation control command corresponding to the target area. The command format is compatible with the control protocol of the ventilation system actuators (such as frequency converters and intelligent damper controllers) arranged underground, and can be directly issued and executed, thereby achieving proactive and precise suppression of the risk of gas outburst.

[0062] Optionally, the preset multi-objective optimization function includes, but is not limited to: the gas concentration at each preset monitoring point underground does not exceed the preset safety threshold, the wind speed in key roadways meets the requirements of coal mine safety regulations, and the total energy consumption of the ventilation system is minimized.

[0063] By employing deep reinforcement learning for multi-objective optimization decision-making, the generated ventilation strategy not only meets safety constraints but also takes into account energy conservation and consumption reduction, thus achieving a leap from experience-based control to intelligent optimization.

[0064] In one embodiment of this application, a specific deep reinforcement learning decision model training scheme is provided. This involves inputting gas outburst prediction data and ventilation system operation status data into a pre-trained deep reinforcement learning decision model. Before determining the optimal action combination from a preset action set of the ventilation system based on a preset multi-objective optimization function as the ventilation control instruction corresponding to the target area, the scheme further includes the following steps: Construct a simulation environment for simulating underground ventilation networks in coal mines; Initialize the policy network and value network of the deep reinforcement learning decision model, and randomly initialize the network parameters. The policy network is used to output the action probability distribution based on the input environment state, and the value network is used to evaluate the expected reward of the current state or state-action pair. The current environmental status data is obtained from the simulation environment. The current environmental status data includes gas outburst simulation data, ventilation equipment operating parameters, and roadway network status within a preset historical time period. The current environmental state is input into the policy network, and the policy network outputs action decisions, which include wind turbine frequency adjustment commands and damper opening adjustment commands. The action decision is input into the simulation environment, and after the simulation environment executes the action decision, the reward value is calculated based on the preset multi-objective optimization function to update the environment state. Based on the current environment state, action decision, reward value, and the updated environment state, the policy gradient algorithm is used to update the parameters of the policy network and the value network until the preset convergence condition is met, thus obtaining a pre-trained deep reinforcement learning decision model.

[0065] In this embodiment, firstly, a simulation environment for simulating the ventilation network of an underground coal mine is constructed. This simulation environment is established based on the ventilation network topology of an actual mine, roadway air resistance parameters, fan characteristic curves, and historical patterns of gas outbursts. It can simulate airflow dynamics, gas migration processes, and state changes after ventilation equipment operation, providing a low-cost, risk-free training and interaction platform for the decision-making model.

[0066] Subsequently, the policy network and value network of the deep reinforcement learning decision model are initialized, and the network parameters are randomly initialized. The policy network is used to output the probability distribution of each action based on the input environmental state, that is, to determine which fan frequency adjustment or damper opening adjustment should be selected in the current state; the value network is used to evaluate the expected reward of the current state or state-action pair, providing a baseline for updating the policy network.

[0067] Furthermore, current environmental state data is acquired from the simulation environment. This data includes simulated gas emission data within a preset historical time period, ventilation equipment operating parameters (such as fan frequency and damper opening), and roadway network status (such as wind pressure and wind speed at each node). The current environmental state is input into the policy network, which outputs action decisions after forward computation. These action decisions specifically include frequency adjustment commands for each local ventilation fan and opening command commands for each damper. The action decisions are then input into the simulation environment. After executing these action decisions, the simulation environment calculates reward values ​​based on a preset multi-objective optimization function and updates the environmental state accordingly, thereby guiding the model to learn safe, compliant, and energy-saving control strategies.

[0068] Then, based on the current environmental state, action decisions, reward values, and the updated environmental state, the parameters of the policy network and value network are updated using a policy gradient algorithm (such as PPO or A3C). Specifically, the advantage function is calculated from the empirical trajectories generated by the interaction, and the expected cumulative reward is maximized through gradient ascent, while the output of the value network is used as a baseline to reduce variance. The above interaction, sampling, and parameter update steps are repeated until the model meets the preset convergence conditions (such as the cumulative reward tending to stabilize or the policy update magnitude being less than a threshold), resulting in a pre-trained deep reinforcement learning decision model. This model can then be directly used to generate optimal ventilation control instructions under real-time gas outburst prediction data without the need for exploratory trial and error in a real downhole environment.

[0069] S50: Send ventilation control commands to the ventilation control equipment to adjust the operating frequency of the fan and / or the opening of the damper.

[0070] In this step, after generating ventilation control commands, these commands are sent to the ventilation control equipment deployed underground via an industrial control network (such as a fieldbus, industrial Ethernet ring network, or wireless communication network). The ventilation control equipment includes actuators such as frequency converters and intelligent damper controllers. Upon receiving the command, the frequency converter adjusts the operating frequency of the corresponding fan in real time according to the frequency adjustment parameters (such as the target frequency value or frequency change step size) contained in the command, thereby changing the fan speed and air supply volume. Upon receiving the command, the intelligent damper controller drives the damper actuator to adjust the opening of the corresponding damper according to the opening adjustment parameters (such as the target opening percentage or opening change amount) contained in the command, thereby changing the airflow distribution path and air volume distribution.

[0071] The above methods enable dynamic adjustment of underground ventilation to proactively suppress the rising trend of methane concentration in the target area.

[0072] In practical applications, 50 smart gas sensors and 20 wind speed sensors are deployed in the fully mechanized mining face and related roadways underground in coal mines, and connected to the coal mining machine location monitoring system and geological exploration data. The collected multi-source sensing data is transmitted to the data center via an underground 10-gigabit ring network. At 1-minute intervals, the data from each preset monitoring point is spatiotemporally aligned with the roadway locations in the Geographic Information System (GIS) to form a real-time panoramic data map. Subsequently, a prediction is run every 5 minutes. The model identifies that the current coal mining machine is about to pass through a small fault (geological data) and that the wind speed in the return airway is showing a decreasing trend (environmental data). The GNN-LSTM model integrates this information and predicts that the gas concentration in the upper corner of the working face will rapidly increase within the next 20 minutes, potentially reaching the alarm threshold at the 18th minute, and generates a warning. Then, upon receiving this prediction warning, the DRL decision model immediately performs simulation calculations in the digital twin ventilation network model. Dozens of control schemes (such as increasing the airflow at the working face and adjusting the airflow diversion of adjacent roadways) can be evaluated within milliseconds. Ultimately, the optimal ventilation control scheme was selected: increasing the frequency of the local booster fan at this working face by 15%, while fine-tuning the isolation airflow dampers in adjacent abandoned roadways to optimize the airflow path. This scheme is expected to suppress peak methane concentration below the safe limit with minimal additional energy consumption. The ventilation control command was then automatically issued. After execution, the methane concentration rise curve in the upper corner flattened, and no alarm was triggered. The data from this successful control was recorded and used for online model fine-tuning to optimize future predictions and decisions.

[0073] As can be seen, in the above scheme, multi-source sensing data from multiple preset monitoring points in the target area of ​​the coal mine are acquired. This multi-source sensing data undergoes spatiotemporal alignment and standardization fusion to generate a spatiotemporal data matrix. This matrix is ​​then input into a pre-trained gas emission prediction model to generate gas emission prediction data for each preset monitoring point within a preset time window. Subsequently, the predicted data and ventilation system operation status data are input into a pre-trained deep reinforcement learning decision model. Based on a preset multi-objective optimization function, the optimal action combination is determined from the preset action set of the ventilation system as a ventilation control command, which is then sent to the ventilation control equipment to adjust the fan frequency and / or damper opening. By fusing multi-source information, using a GNN-LSTM hybrid model to dynamically and accurately predict gas emission, and employing deep reinforcement learning for multi-objective optimization decision-making to automatically generate the optimal ventilation control strategy, advanced prediction and intelligent control of gas prevention are achieved. This effectively reduces the frequency and duration of gas exceedances, prevents gas accumulation at the source, and thus improves the safety level of coal mine production.

[0074] In one embodiment, a coal mine underground gas control device is provided, which corresponds one-to-one with the coal mine underground gas control methods described in the above embodiments. For example... Figure 2As shown, the underground gas control device 100 in a coal mine includes: an acquisition module 101, a first generation module 102, a second generation module 103, a determination module 104, and a sending module 105. Detailed descriptions of each functional module are as follows: The acquisition module 101 is used to acquire multi-source sensing data from multiple preset monitoring points in the target area of ​​the coal mine. The multi-source sensing data includes at least one of the following: environmental data, mining operation status data, geological condition data, and ventilation system operation status data. The first generation module 102 is used to perform alignment and standardization fusion processing on multi-source sensing data to generate a spatiotemporal data matrix. The second generation module 103 is used to input the spatiotemporal data matrix into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point in the target area within a preset time window. The determination module 104 is used to input the gas outburst prediction data and the operation status data of the ventilation system into the pre-trained deep reinforcement learning decision model, and determine the optimal action combination from the preset action set of the ventilation system based on the preset multi-objective optimization function, as the ventilation control instruction corresponding to the target area. The sending module 105 is used to send ventilation control commands to the ventilation control equipment to adjust the operating frequency of the fan and / or the opening degree of the damper.

[0075] In one embodiment, the first generation module 102 is specifically used for: Data cleaning is performed on various data points in the multi-source sensing data. Based on the geospatial location labels corresponding to each preset monitoring point, spatial registration is performed on the cleaned multi-source sensing data to map the data under different coordinate systems to the preset underground roadway network, thus obtaining spatially aligned data. Based on the timestamps of each data in the cleaned multi-source sensing data, time synchronization is performed according to the preset time axis to align the data to the same time reference point and obtain the time series data of each preset monitoring point. Spatial aligned data and time series data are correlated and integrated to generate a spatiotemporal data matrix.

[0076] In one embodiment, the second generation module 103 is specifically used for: The spatiotemporal data matrix is ​​input into the pre-trained gas outburst prediction model. Through graph neural network, based on the roadway topology connection relationship of the preset underground roadway network, spatial information aggregation is performed on the spatially aligned data, and the spatial features of gas diffusion at each preset monitoring point are extracted to generate a spatial feature map containing the roadway topology connection relationship. By using the recurrent gating mechanism of the Long Short-Term Memory Network, the temporal dynamic characteristics of the gas concentration evolution over time at each preset monitoring point in the time series data are extracted, and the time series hidden state of each preset monitoring point is generated. The spatial feature map is fused with the latent state of the time series to obtain the spatiotemporal fusion feature tensor; Based on the spatiotemporal fusion feature tensor, the predicted values ​​of gas concentration and gas emission volume at each preset monitoring point within the target area are determined within a preset time window, and used as gas emission prediction data.

[0077] In one embodiment, the device further includes: The building module is used to construct a hybrid model of graph neural network and long short-term memory network, and to randomly initialize the network parameters to obtain the initialized hybrid model.

[0078] In one embodiment, the acquisition module 101 is further configured to acquire historical multi-source sensing data and historical gas outburst data of the underground coal mine area within a preset historical time period.

[0079] In one embodiment, the device further includes: The third generation module is used to align and standardize the historical multi-source sensing data to generate a historical spatiotemporal data matrix. The fourth generation module is used to train the initialized hybrid model under supervised learning by taking the historical spatiotemporal data matrix as input and the historical gas outburst data as supervision labels, adjusting the network parameters until the preset convergence condition is met, and obtaining the pre-trained gas outburst prediction model.

[0080] In one embodiment, the acquisition module 101 is further configured to acquire real-time multi-source sensing data and actual gas concentration change data fed back after the gas emission prediction model performs prediction.

[0081] In one embodiment, the device further includes: The judgment module is used to compare the actual gas concentration change data with the gas emission prediction data to determine whether there is a deviation between the gas emission prediction data and the actual gas concentration change data. The model update module is used to update the network parameters of the gas outburst prediction model by using real-time multi-source sensing data and actual gas concentration change data as incremental training samples when there is a discrepancy between the gas outburst prediction data and the actual gas concentration change data.

[0082] In one embodiment, the judgment module is further configured to compare the predicted gas concentration values ​​of each preset monitoring point in the gas emission prediction data with a preset gas concentration threshold, and determine whether there is at least one time point in the preset time window where the predicted gas concentration value of each preset monitoring point exceeds the preset gas concentration threshold.

[0083] In one embodiment, the acquisition module 101 is further configured to, for any preset monitoring point, when the predicted gas concentration at at least one time point within its preset time window exceeds the preset gas concentration threshold, acquire the location of the monitoring point exceeding the limit, the time point exceeding the limit, and the predicted value of the exceeding limit.

[0084] In one embodiment, the device further includes: The fifth generation module is used to generate early warning information based on the location of the monitoring point exceeding the limit, the time point of exceeding the limit, and the predicted concentration value exceeding the limit.

[0085] In one embodiment, the sending module 105 is further configured to send the warning information to the terminal of the relevant personnel.

[0086] In one embodiment, the construction module is further configured to: Construct a simulation environment for simulating underground ventilation networks in coal mines; Initialize the policy network and value network of the deep reinforcement learning decision model, and randomly initialize the network parameters. The policy network is used to output the action probability distribution based on the input environment state, and the value network is used to evaluate the expected reward of the current state or state-action pair.

[0087] In one embodiment, the acquisition module 101 is further configured to acquire current environmental status data from the simulation environment, wherein the current environmental status data includes gas outburst simulation data, ventilation equipment operating parameters, and roadway network status within a preset historical time period.

[0088] In one embodiment, the device further includes: The sixth generation module is used to input the current environmental state into the policy network, and the policy network outputs action decisions, including fan frequency adjustment commands and damper opening adjustment commands. The calculation module is used to input action decisions into the simulation environment. After the simulation environment executes the action decisions, it calculates the reward value based on a preset multi-objective optimization function to update the environment state. The seventh generation module is used to update the parameters of the policy network and the value network based on the current environment state, action decision, reward value and the updated environment state, using the policy gradient algorithm until the preset convergence condition is met, so as to obtain a pre-trained deep reinforcement learning decision model.

[0089] The coal mine underground gas control device 100 provided by this invention acquires multi-source sensing data from multiple preset monitoring points in a target area of ​​the coal mine. It performs spatiotemporal alignment and standardization fusion processing on the multi-source sensing data to generate a spatiotemporal data matrix. This matrix is ​​then input into a pre-trained gas emission prediction model to generate gas emission prediction data for each preset monitoring point within a preset time window. Subsequently, the predicted data and ventilation system operation status data are input into a pre-trained deep reinforcement learning decision model. Based on a preset multi-objective optimization function, the optimal action combination is determined from the preset action set of the ventilation system as a ventilation control command, which is then sent to the ventilation control equipment to adjust the fan frequency and / or damper opening. By fusing multi-source information, using a GNN-LSTM hybrid model to dynamically and accurately predict gas emission, and employing deep reinforcement learning for multi-objective optimization decision-making to automatically generate the optimal ventilation control strategy, this device achieves advanced prediction and intelligent control of gas control. It can effectively reduce the frequency and duration of gas exceedances, prevent gas accumulation at the source, and thus improve the safety level of coal mine production.

[0090] Specific limitations regarding underground gas control devices in coal mines can be found in the above-mentioned limitations on underground gas control methods, and will not be repeated here. Each module in the aforementioned underground gas control device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the corresponding operations of each module.

[0091] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for preventing underground gas in coal mines.

[0092] In one embodiment, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the above-described method for preventing underground gas in coal mines.

[0093] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or electronic device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0095] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0096] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for preventing and controlling underground gas in coal mines, characterized in that, include: Acquire multi-source sensing data from multiple preset monitoring points in a target area underground in a coal mine, wherein the multi-source sensing data includes at least one of the following: environmental data, mining operation status data, geological condition data, and ventilation system operation status data; The multi-source sensing data is aligned and standardized and fused to generate a spatiotemporal data matrix; The spatiotemporal data matrix is ​​input into a pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point within the target area within a preset time window. The gas outburst prediction data and the ventilation system operation status data are input into a pre-trained deep reinforcement learning decision model. Based on a preset multi-objective optimization function, the optimal action combination is determined from the preset action set of the ventilation system as the ventilation control instruction corresponding to the target area. The ventilation control command is sent to the ventilation control equipment to adjust the operating frequency of the fan and / or the opening degree of the damper.

2. The method for preventing and controlling underground gas in coal mines according to claim 1, characterized in that, The step of aligning and standardizing the multi-source sensing data to generate a spatiotemporal data matrix specifically includes: Data cleaning is performed on each data item in the multi-source sensing data; Based on the geospatial location labels corresponding to each preset monitoring point, spatial registration is performed on the cleaned multi-source sensing data to map the data under different coordinate systems to the preset underground roadway network, thus obtaining spatially aligned data. Based on the timestamps of each data in the cleaned multi-source sensing data, time synchronization is performed according to the preset time axis to align the data to the same time reference point and obtain the time series data of each preset monitoring point. The spatially aligned data and the time-series data are correlated and integrated to generate the spatiotemporal data matrix.

3. The method for preventing and controlling underground gas in coal mines according to claim 2, characterized in that, The gas emission prediction model includes a graph neural network and a long short-term memory network. The step of inputting the spatiotemporal data matrix into the pre-trained gas emission prediction model to generate gas emission prediction data for each preset monitoring point within the target area within a preset time window specifically includes: The spatiotemporal data matrix is ​​input into a pre-trained gas outburst prediction model. Through the graph neural network, based on the roadway topology connection relationship of the preset underground roadway network, spatial information aggregation is performed on the spatially aligned data, and the spatial features of gas diffusion at each preset monitoring point are extracted to generate a spatial feature map containing the roadway topology connection relationship. By using the recurrent gating mechanism of the Long Short-Term Memory Network, the temporal dynamic characteristics of the gas concentration evolution over time at each preset monitoring point in the time series data are extracted, and the time series hidden state of each preset monitoring point is generated. The spatial feature map is fused with the time series hidden state to obtain a spatiotemporal fusion feature tensor. Based on the spatiotemporal fusion feature tensor, the predicted values ​​of gas concentration and gas emission volume at each preset monitoring point within the target area are determined within a preset time window, and used as gas emission prediction data.

4. The method for preventing and controlling underground gas in coal mines according to claim 1, characterized in that, Before inputting the spatiotemporal data matrix into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point within the target area within a preset time window, the method further includes: A hybrid model of graph neural network and long short-term memory network is constructed, and the network parameters are randomly initialized to obtain the initialized hybrid model; Acquire historical multi-source sensing data and historical gas emission data of underground coal mine areas within a preset historical time period; The historical multi-source sensing data is aligned and standardized and fused to generate a historical spatiotemporal data matrix; Using the historical spatiotemporal data matrix as input and the historical gas outburst data as supervision labels, supervised learning training is performed on the initialized hybrid model, and the network parameters are adjusted until the preset convergence condition is met to obtain the pre-trained gas outburst prediction model.

5. The method for preventing and controlling underground gas in coal mines according to claim 1, characterized in that, After inputting the spatiotemporal data matrix into a pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point within the target area within a preset time window, the method further includes: Acquire real-time multi-source sensing data and actual gas concentration change data fed back after the gas emission prediction model performs prediction; The actual gas concentration change data is compared with the gas emission prediction data to determine whether there is a deviation between the gas emission prediction data and the actual gas concentration change data; In the event of a discrepancy between the predicted gas emission data and the actual gas concentration change data, the real-time multi-source sensing data and the actual gas concentration change data are used as incremental training samples, and the network parameters of the gas emission prediction model are updated using the backpropagation algorithm.

6. The method for preventing and controlling underground gas in coal mines according to claim 1, characterized in that, After inputting the spatiotemporal data matrix into a pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point within the target area within a preset time window, the method further includes: Compare the predicted gas concentration values ​​of each preset monitoring point in the gas emission prediction data with the preset gas concentration threshold, and determine whether the predicted gas concentration value of each preset monitoring point exceeds the preset gas concentration threshold at at least one time point within the preset time window. For any preset monitoring point, when the predicted gas concentration value at at least one time point within its preset time window exceeds the preset gas concentration threshold, the location of the monitoring point exceeding the limit, the time point exceeding the limit, and the predicted value of the exceeding limit are obtained. Based on the location of the monitoring point exceeding the limit, the time point of exceeding the limit, and the predicted concentration value exceeding the limit, an early warning message is generated; The warning information will be sent to the terminals of relevant personnel.

7. The method for preventing and controlling underground gas in coal mines according to claim 1, characterized in that, Before inputting the gas outburst prediction data and the ventilation system operation status data into a pre-trained deep reinforcement learning decision model, and determining the optimal action combination from a preset action set of the ventilation system based on a preset multi-objective optimization function as the ventilation control instruction corresponding to the target area, the method further includes: Construct a simulation environment for simulating underground ventilation networks in coal mines; Initialize the policy network and value network of the deep reinforcement learning decision model, and randomly initialize the network parameters. The policy network is used to output the action probability distribution according to the input environment state, and the value network is used to evaluate the expected reward of the current state or state-action pair. The current environmental status data is obtained from the simulation environment, wherein the current environmental status data includes gas outburst simulation data, ventilation equipment operating parameters and roadway network status within a preset historical time period; The current environmental state is input to the policy network, and the policy network outputs action decisions, wherein the action decisions include fan frequency adjustment commands and damper opening adjustment commands; The action decision is input into the simulation environment, and after the simulation environment executes the action decision, the reward value is calculated based on a preset multi-objective optimization function to update the environment state. Based on the current environment state, the action decision, the reward value, and the updated environment state, the parameters of the policy network and the value network are updated using the policy gradient algorithm until the preset convergence condition is met, thus obtaining a pre-trained deep reinforcement learning decision model.

8. A coal mine underground gas control device, characterized in that, include: The acquisition module is used to acquire multi-source sensing data from multiple preset monitoring points in a target area underground in a coal mine. The multi-source sensing data includes at least one of the following: environmental data, mining operation status data, geological condition data, and ventilation system operation status data. The first generation module is used to perform alignment and standardization fusion processing on the multi-source sensing data to generate a spatiotemporal data matrix. The second generation module is used to input the spatiotemporal data matrix into the pre-trained gas outburst prediction model to generate gas outburst prediction data for each preset monitoring point in the target area within a preset time window. The determination module is used to input the gas outburst prediction data and the operation status data of the ventilation system into a pre-trained deep reinforcement learning decision model, and determine the optimal action combination from the preset action set of the ventilation system based on a preset multi-objective optimization function, as the ventilation control instruction corresponding to the target area. The sending module is used to send the ventilation control command to the ventilation control equipment to adjust the operating frequency of the fan and / or the opening degree of the damper.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the coal mine underground gas prevention method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the coal mine underground gas prevention method as described in any one of claims 1 to 7.

Images