Mine gas regulation and control method and equipment based on multi-modal data and medium

Abstract

The embodiment of the invention discloses a mine gas regulation and control method and device based on multi-modal data and a medium, and relates to the technical field of gas regulation and control, and the method comprises the steps that multi-source real-time sensing data, roadway point cloud data and CFD airflow simulation grid data in a target mine roadway are obtained, space-time alignment is carried out to determine a real-time roadway data cube, and the real-time roadway data cube is obtained; determining gas concentration field distribution and causal contribution popularity data in a preset future time window through a multi-mode depth causal model and the real-time roadway data cube; generating a risk field gradient feature vector based on gas concentration field distribution and causal contribution popularity data, and matching a decision path according to the risk field gradient feature vector; and when the decision path is a causal decision path, a fan / air door adjusting target point is determined, the fan / air door adjusting target point is simulated to determine a dynamic safety constraint boundary, a preset optimization objective function is solved to generate a fan rotating speed adjusting instruction and an air door opening degree control instruction, and fan / air door control is achieved.

Description

Technical Field

[0001] This specification relates to the field of gas control technology, and in particular to a mine gas control method, equipment, and medium based on multimodal data. Background Art

[0002] In the field of coal mine safety, gas disaster prevention and control remains a core challenge in ensuring personnel safety and efficient resource development. Existing technologies primarily address this challenge through the following approaches: A passive monitoring system based on a discrete single-source sensor network deploys gas concentration sensors at key tunnel nodes, triggering audible and visual alarms based on fixed thresholds; a ventilation control mechanism driven by human experience, where dispatchers manually adjust fan speed and damper opening based on alarm signals and empirical rules; and an offline computational fluid dynamics (CFD) simulation-assisted ventilation design, optimizing static ventilation networks through periodic airflow simulation. However, these technologies are gradually exposing systemic flaws when addressing the dynamic risks of modern mines.

[0003] In the real-time risk perception process, single-point sensors can only capture localized transient concentration values ​​and are unable to characterize the spatiotemporal evolution of gas diffusion. This results in a short warning window and allows only the appearance of excessive gas levels, but cannot locate the source of sudden releases or predict the diffusion path, further weakening the effectiveness of these warnings. With the increasing deployment of the Internet of Things (IoT), mines have accumulated second-level time-series data on gas, temperature, humidity, and pressure differentials, as well as three-dimensional LiDAR tunnel point clouds. However, existing research has focused on machine learning predictions based on a single data source or at a single scale, lacking unified modeling of multi-source heterogeneous information and explicit characterization of physical causal laws. Furthermore, they have failed to integrate risk prediction with automatic ventilation control in a closed-loop control system. For example, traditional systems only issue alarms when concentration exceeds the limit, failing to predict the progression of risk accumulation in unexploded areas. This results in a significant lack of buffer time for underground personnel evacuation and emergency response. Furthermore, manual ventilation control relies on operator experience and lacks real-time physical guidance, leading to response lags and amplified operational errors during ventilation decision-making.

[0004] Therefore, in the process of mine gas disaster prevention and control, most people focus on machine learning predictions based on a single data source or a single scale, and respond passively when the concentration exceeds the standard. This results in a short warning window and low control efficiency, making it difficult to meet the rigid safety requirements of smart mines. Summary of the Invention

[0005] One or more embodiments of this specification provide a mine gas control method, equipment and medium based on multimodal data, which are used to solve the following technical problems: in the process of mine gas disaster prevention and control, the focus is on machine learning predictions of a single data source or a single scale, and passive response is performed when the concentration exceeds the standard, resulting in a short warning window and low control efficiency, which makes it difficult to meet the rigid safety requirements of smart mines.

[0006] One or more embodiments of this specification adopt the following technical solutions: One or more embodiments of the present specification provide a mine gas control method based on multimodal data, characterized in that the method includes: obtaining multi-source real-time sensor data, tunnel point cloud data and CFD airflow simulation grid data in the target mine tunnel, and performing spatiotemporal alignment to determine the real-time tunnel data cube, and determining the gas concentration field distribution and causal contribution heat data in a preset future time window through a preset multimodal deep causal model and the real-time tunnel data cube; generating a risk field gradient feature vector based on the gas concentration field distribution and the causal contribution heat data, so as to match a decision path according to the risk field gradient feature vector, wherein the decision path includes a Monte Carlo plan set and a causal decision path; when the decision path is a causal decision path, determining a fan / damper adjustment target, simulating the fan / damper adjustment target, determining a dynamic safety constraint boundary, solving a preset optimization objective function under the dynamic safety constraint boundary, generating a fan speed adjustment instruction and a damper opening control instruction, and realizing fan / damper control.

[0007] One or more embodiments of this specification provide a mine gas control device based on multimodal data, including: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the above method.

[0008] One or more embodiments of this specification provide a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above method.

[0009] At least one of the above technical solutions adopted in the embodiments of this specification can achieve the following beneficial effects: through the technical solutions of the embodiments of this specification, traditional single-point sensors can only perceive the instantaneous value of local concentration, while the embodiments of this specification construct a complete information field of gas diffusion at the fluid mechanics level through the spatiotemporal alignment mechanism of multimodal data cubes. The tunnel curvature characteristics extracted from the LiDAR point cloud are coupled with the turbulence prior of the CFD wind flow grid, so that the spatiotemporal Transformer model can capture the causal chain of pressure difference fluctuations, wind speed attenuation, and concentration accumulation, so that the early warning mechanism is transformed from passive concentration limit detection to active diffusion dynamics deduction. Traditional manual air adjustment needs to go through a multi-level delay chain of early warning, decision-making, and execution. However, the causal decision path of the embodiment of this specification realizes the integration of decision-making and execution at the edge computing end through real-time matching of the gradient field and the network topology. It selects the optimal control target based on the network transmission efficiency and directly drives the protocol-level instructions, effectively compressing the full-link reaction time of perception, decision-making, and execution. Conventional static safety thresholds cannot adapt to dynamic risks such as tunnel deformation. The ventilation network propagation model of the embodiment of this specification constructs a virtual test field for air adjustment instructions, quantifies structural risks into elastic protection parameters, and realizes the conversion from empirical safety margin to calculated safety boundary. For situations with high confidence in the gradient direction, the causal path is enabled to achieve the physical optimal solution. On the contrary, the group intelligent search of the Monte Carlo plan is adopted to directly realize the intelligent switching of the decision mode, effectively improving the control accuracy and control effect of mine gas, and meeting the rigid safety requirements of smart mines. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some of the embodiments described in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without inventive work. In the drawings: Figure 1 A flow chart of a mine gas control method based on multimodal data provided in an embodiment of this specification; Figure 2 A schematic diagram of the structure of a mine gas control device based on multimodal data provided in an embodiment of this specification. DETAILED DESCRIPTION

[0011] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this specification without creative work should fall within the scope of protection of this specification.

[0012] The embodiments of this specification provide a mine gas control method based on multimodal data. It should be noted that the execution subject in the embodiments of this specification can be a server or any device with data processing capabilities. Figure 1 A flow chart of a mine gas control method based on multimodal data is provided in the embodiment of this specification, such as Figure 1 As shown, it mainly includes the following steps: Step S101: Acquire multi-source real-time sensor data, tunnel point cloud data, and CFD airflow simulation grid data in the target mine tunnel, perform spatiotemporal alignment, determine the real-time tunnel data cube, and determine the gas concentration field distribution and causal contribution heat data within the preset future time window through the preset multimodal deep causal model and the real-time tunnel data cube.

[0013] In one embodiment of the present specification, multi-source real-time sensor data, roadway point cloud data, and CFD airflow simulation grid data are acquired within the target mine roadway. During the multi-source data acquisition and fusion process within the target mine roadway, a distributed, intrinsically safe sensor network is first used to capture the dynamic parameters of the roadway environment in real time. Infrared gas sensors installed at key nodes in the coal mining face, return airway, and main transport roadway generate concentration time-series curves at a frequency of seconds. Electrochemical temperature and humidity transmitters record changes in air state, and micro-differential pressure transmitters monitor airflow pressure fluctuations. A mining explosion-proof mobile LiDAR device periodically scans the three-dimensional shape of the roadway to generate a point cloud dataset with millimeter-level accuracy. The scanning trajectory covers the entire cross-section and support structure. Simultaneously, the latest roadway topology model is invoked based on the mine digital twin platform, and steady-state / transient fluid dynamics simulations are performed in ANSYS Fluent to output CFD structured grid data containing wind speed vectors, turbulent kinetic energy, and static pressure distribution.

[0014] Determining a real-time tunnel data cube specifically includes: acquiring multi-source real-time sensor data in a target mine tunnel in real time, wherein the multi-source real-time sensor data includes gas concentration time series data, temperature and humidity time series data, and pressure difference time series data; synchronizing the multi-source real-time sensor data with timestamps through a high-precision clock service for time alignment, and using a preset algorithm to spatially align the tunnel point cloud data and the CFD airflow simulation grid data in a unified three-dimensional coordinate system to determine the tunnel grid data, wherein the tunnel grid data includes tunnel grid coordinates and tunnel point cloud geometric features; encapsulating the time-space aligned data into real-time tunnel data cubes of multiple preset dimensions, wherein the preset dimensions include a time dimension, a space dimension, and a feature dimension, wherein the space dimension is mapped to the tunnel grid coordinate, and the feature dimension includes sensor measurement indicators and tunnel point cloud geometric features.

[0015] In the underground environment, the dynamics of gas concentration are affected by the coupling of multiple factors such as tunnel morphology, airflow field distribution and equipment status. If we only rely on isolated sensor data, we cannot capture the diffusion path of gas along the curved tunnel after it emerges from the working face; if we ignore the sudden change characteristics of the tunnel cross section scanned by LiDAR, the model will underestimate the risk of gas accumulation caused by local eddy currents; and if we deviate from the fluid mechanics prior of CFD simulation, the prediction results are likely to violate the law of conservation of mass and result in non-physical solutions such as diffusion against wind flow.

[0016] In one embodiment of this specification, gas concentration, temperature, humidity, and pressure differential time series data are aggregated to edge computing nodes via an industrial ring network. A data cube is then constructed through a three-layer fusion process. At the time layer, the IEEE 1588 precision clock protocol is used to synchronize multi-source sensor data, and sliding window aggregation is used to align the second-level streaming data with the minute-level point cloud / CFD time scale. The spatial layer performs SLAM mapping and outlier filtering on the LiDAR point cloud. The point cloud is registered to the mine's absolute coordinate system (based on underground mapping control points) using the ICP iterative closest point algorithm. CFD grid nodes are mapped to a unified coordinate system through bounding box matching. The feature layer voxelizes the registered spatial data, populating each voxel (e.g., a discrete grid segmented along the strike of a roadway) with feature channels such as sensor timing curves, local cross-sectional area and surface curvature extracted from the point cloud, and a baseline wind speed vector provided by CFD. This ultimately creates a data cube with a continuous time axis, consistent spatial indexing, and complete feature dimensions, which is stored in a columnar in-memory database for real-time model access.

[0017] To align the discrete time-series stream with the periodic scan data, bidirectional buffering is also implemented. Second-level sensor streams are batched using a 10-second sliding window. The LiDAR point cloud is downsampled to a 0.05-meter voxel grid and finely registered using ICP. The CFD simulation mesh (calculated offline using software such as Fluent) is aligned with the point cloud coordinate system using minimum bounding box coordinates, forming a unified 3D reference frame for subsequent spatiotemporal encoding. Second-level sensor batches first enter the stream processing engine (Apache Flink + Kafka), where denoising, null value compensation, and outlier screening are performed on the Edge Box. Spatial indexing is then applied to map each sensor sample to the nearest CFD grid cell and point cloud voxel. Simultaneously, normal vector estimation and surface reconstruction are performed on the point cloud data on the GPU, and geometric features such as roadway cross-sectional area and potential air velocity damping are calculated. The fused data cube is stored in an Arrow-format in-memory table according to the (time × space × feature) dimension, facilitating direct Tensor flow. Furthermore, column-based Delta Lake is used for version management to ensure replayability of historical data. During this phase, statistical priors (mean, variance, and quantiles) are calculated for key variables such as gas concentration, pressure differential, and wind speed, and written to the metadata warehouse for online normalization during model training and inference.

[0018] In one embodiment of this specification, a multimodal deep causal model includes a spatiotemporal Transformer-MoE architecture and a self-supervised causal masking mechanism. This model encodes time-series sensor data into temporal channel features and 3D point clouds and CFD meshes into spatial channel features. It relies on an expert gating network to dynamically select adaptively scaled attention experts, balancing long-range dependencies with local details. Through the self-supervised causal masking mechanism, the key sequence "gas-pressure differential-wind speed" is randomly masked during the training phase and its missing components are predicted, forcing the model to explicitly learn the physical causal chain. During the inference phase, the mask confidence level is used to determine the intensity of anomalies, enhancing sensitivity and interpretability to sudden gas leaks. Using a pre-set multimodal deep causal model and real-time roadway data cube, the gas concentration field distribution and causal contribution heat data within a pre-set future time window are determined.

[0019] The training center (a 4xA100 GPU server or a domestic computing card) reads the last three months' worth of data cubes and splits them into training, validation, and test sets at a 70% / 15% / 15% split. The spatiotemporal Transformer-MoE architecture employs eight layers of temporal attention, four layers of spatial attention, and six expert-gated branches. During the self-supervision phase, continuous segments of the "gas concentration-pressure differential-wind speed" graph are randomly masked, aiming to minimize reconstruction error in the masked regions while also minimizing physical residual loss (consistency between the CFD prior and the predicted wind speed). After training, the model is exported to an ONNX + TensorRT engine and deployed using a two-tiered deployment strategy: offline redundant inference at the central station and online low-latency inference at the wellhead Edge Box. The Edge Box uses Jetson Orin or a domestic AI chip, with an inference batch size of 1 and latency controlled to around 100 ms. The central station deployment version includes an additional first-order model capacity for nightly batch replay and accuracy verification.

[0020] Using a pre-set multimodal deep causal model and real-time tunnel data cube, the gas concentration field distribution and causal contribution heat data within a pre-set future time window are determined. In the multimodal deep causal model deployed on the downhole edge computing node, the real-time tunnel data cube is loaded as an input tensor into the spatiotemporal Transformer-MoE architecture for feature analysis. This architecture first processes the data through a separate attention channel. The temporal self-attention layer focuses on long-range dependencies in the sensor time series stream, capturing cross-period correlations between gas concentration and pressure differential fluctuations. The spatial self-attention layer fuses tunnel curvature features extracted from the LiDAR point cloud with the wind speed vector field from the CFD mesh to learn fluid motion patterns under spatial topological constraints. The expert-gated network dynamically routes cross-scale features. For example, when a sudden narrowing of the tunnel indicated by the point cloud is detected, high-resolution spatial expert-enhanced eddy flow modeling is automatically activated.

[0021] Through a self-supervised causal masking mechanism pre-embedded during model training, key segments of the "gas concentration-pressure differential-wind speed" sequence are randomly masked, forcing the model to reconstruct the masked physical quantities from the remaining features. This allows the network to explicitly grasp the causal chain of "gas sudden release from coal wall fissures → local pressure differential drop → wind speed reduction → concentration accumulation." During inference, the model outputs two channels: one is the three-dimensional gas concentration field distribution in the future time window, which is a tensor of the concentration prediction for each voxel within the roadway grid coordinates over time. This tensor visually displays high-risk concentration areas (such as the concentration gradient at the corner of the return airway); the other is causal contribution heat data. By quantifying the activation strength of the model's attention weights at key causal chain nodes, a heat map covering the roadway space is generated (for example, identifying the contribution of pressure differential monitoring points to the working face concentration prediction). A sudden increase in the heat value in a certain area indicates the presence of a hidden risk source not reflected in the sensor data. Ultimately, the concentration field distribution and causal heat data together constitute the digital twin of risk prediction, providing a situational awareness foundation with both physical interpretability and spatiotemporal accuracy for subsequent decision-making.

[0022] Through the above technical solution, multimodal dynamic coupling enables the prediction model to simultaneously perceive changes in the physical environment and airflow conditions. For example, when LiDAR detects delamination of the working face roof, it automatically correlates the porosity characteristics extracted from the point cloud in that area with the CFD eddy flow model. Combined with the sensor micro-pressure differences, it can often predict the risk of gas outburst from the delamination, significantly extending the warning window compared to traditional single-data source analysis. A unified spatiotemporal benchmark eliminates cross-source information bias. In complex ventilation networks, voxel indexing accurately locates the complete causal chain: "a sudden increase in concentration at sensor A → a point cloud indicating a roadway deflection downwind of that point → a CFD indication of a eddy flow zone forming at the deflection point." This guides targeted control to avoid secondary risks and reduces the number of ineffective equipment actions compared to manual ventilation strategies. The columnar storage and voxel-based feature extraction of data cubes enable the Transformer model to directly load tensors for inference, avoiding the latency associated with traditional methods that require repeated calls to multiple data sources from a central server, thus meeting the minute-level closed-loop control requirements of underground mines.

[0023] Step S102 : Based on the gas concentration field distribution and the causal contribution heat data, a risk field gradient feature vector is generated to match a decision path according to the risk field gradient feature vector.

[0024] The decision path includes a Monte Carlo plan set and a causal decision path; Based on the gas concentration field distribution and the causal contribution heat data, a risk field gradient characteristic vector is generated, specifically including: calculating the spatial partial derivative of the gas concentration field distribution along the three-dimensional coordinate direction of the tunnel to generate an initial gradient vector field; using the causal contribution heat data as a weight matrix, performing a Hadamard product operation with the initial gradient vector field to generate a weighted risk gradient field; extracting the main gradient direction vector, risk entropy value index and causal extreme point coordinates from the weighted risk gradient field, and encapsulating them into a structured characteristic vector to determine the risk field gradient characteristic vector.

[0025] Simply relying on the gas concentration field distribution can only provide a static risk snapshot and cannot reveal the dynamic mechanism and evolution direction of gas diffusion. For example, in the tunnel bifurcation area, the same concentration value may come from upstream outflow or local accumulation, and the control strategies required for the two are completely different.

[0026] In one embodiment of the present specification, the predicted gas concentration field distribution tensor is loaded through the GPU acceleration environment of the edge computing node, and the spatial partial derivatives are calculated along the three dimensions of the tunnel direction (X-axis), the horizontal and vertical direction (Y-axis), and the vertical direction (Z-axis). The Sobel-Feldman gradient operator is used to perform a convolution operation on the grid concentration values ​​to generate an initial gradient vector field, in which each grid point contains a direction vector and a modulus. The causal contribution heat data matrix is ​​read synchronously, and the heat value is mapped to the weight interval through Min-Max normalization to form a weight tensor with the same dimension as the gradient field. The Hadamard product operation is performed, and the gradient vector is multiplied by the heat weight of the corresponding position at each grid point, so that the gradient signal in the high-contribution area (such as near the pressure difference monitoring point) is significantly enhanced, and the gradient in the low-contribution area is attenuated. Feature extraction is performed within the weighted risk gradient field. First, the point with the maximum global gradient modulus is searched for, and its unit direction vector is taken as the main gradient direction. Next, the Shannon entropy is calculated based on the concentration probability distribution. The concentration values ​​are discretized into buckets and then calculated according to the probability distribution. Finally, a non-maximum suppression algorithm is used to locate extreme points where the causal heat weight exceeds the neighborhood mean, and their physical coordinates are recorded. Finally, the main gradient direction vector, entropy scalar, and extreme point coordinate sequence are packaged into a JSON structured data package.

[0027] Through the above technical solution, traditional gas risk analysis relies solely on concentration scalar thresholds to determine risk levels, and is unable to distinguish different risk modes at the same concentration, such as uniform diffusion and point source injection, resulting in insufficient targeting of air adjustment strategies. In addition, ignoring physical causal relationships makes the system blind to hidden risks in areas not covered by sensors. The embodiments of this specification achieve dynamic visualization of risk evolution through the innovative integration of gradient fields and causal heat, allowing the control system to break through the limitations of static thresholds. For example, when the concentration value at the working face does not exceed the standard, the weighted gradient field shows that the main gradient direction points to the abandoned tunnel. Combined with the extreme point coordinates, the hidden fissure gas source is located, and plugging measures are initiated in advance to avoid accumulation and explosion. However, due to the lack of gradient analysis, the traditional method does not respond until the concentration exceeds the limit, delaying the critical disposal opportunity. By embedding causal contributions quantitatively, decisions are given physical explainability. When the entropy value of a certain area in the weighted gradient field suddenly increases, the associated causal heat source is automatically traced. For example, if the temperature at the return air duct pressure difference monitoring point is abnormally increased, historical data is immediately linked to verify that dust accumulation on the fan blades at that point has caused a decrease in efficiency, thereby accurately locking in the equipment maintenance target and saving a lot of fault diagnosis time compared to manual experience-based troubleshooting.

[0028] According to the risk field gradient feature vector, the decision path is matched, specifically including: determining the risk entropy value index in the risk field gradient feature vector; when the risk entropy value index exceeds a preset dynamic threshold, activating the Monte Carlo plan generation engine to output a Monte Carlo plan set as the decision path; when the risk entropy value index does not exceed the preset dynamic threshold, determining that the decision path is a causal decision path.

[0029] In the intelligent control system of mine gas, the traditional control system adopts a single decision-making logic and cannot adapt to the dynamic changes of working conditions. When the sensor data is complete and the flow field is stable, precise control based on physical causality has significant advantages; but in high-uncertainty scenarios such as gas sudden release accompanied by equipment failure, the gradient direction that the causal model relies on may be distorted, and continuing to use a fixed decision path will lead to serious misjudgment.

[0030] In one embodiment of the present specification, after receiving the risk field gradient feature vector in real time, the embedded risk entropy index is first parsed, and the Shannon entropy calculated by the previous step is used. In the dynamic threshold configuration module preloaded on the edge computing node, the real-time ventilation network complexity, sensor health status and historical accident data are comprehensively considered to generate an adaptive decision threshold. For example, when the number of topological branches increases, the threshold is lowered. When the entropy index exceeds the threshold, the Monte Carlo plan engine is activated. When the entropy value does not exceed the threshold, the causal decision path is triggered. The risk entropy index exceeds the dynamic threshold, indicating that the system has entered the critical phase transition zone of multi-stable coexistence. At this time, the Monte Carlo plan engine is activated to avoid the risk of single-point decision failure through probability coverage and parallel simulation; and in a deterministic scenario where the entropy value is controllable, the causal decision path is maintained, which can give full play to the high efficiency and precision advantages of the physical model.

[0031] In step S103, when the decision path is a causal decision path, the fan / damper adjustment target is determined, the fan / damper adjustment target is simulated, and the dynamic safety constraint boundary is determined to solve the preset optimization objective function under the dynamic safety constraint boundary, generate the fan speed adjustment instruction and the damper opening control instruction, and realize the control of the fan / damper.

[0032] Traditional manual air regulation relies on experience to select the adjustment target, which easily ignores the coupling relationship between the gas diffusion dynamic direction and the ventilation network topology. For example, when the gradient direction and the main air flow angle are small, the wrong closing of the branch damper will block the airflow and cause local accumulation; in addition, the ability to locate hidden risk sources is insufficient, and the crack outburst points in the tunnel far away from the main air duct are often ignored; in the process of conventional manual regulation, the single adjustment target cannot cope with the multi-center diffusion of gas clouds. The embodiment of this specification uses the quantitative analysis of the main gradient direction and the main air flow angle to distinguish between the two control modes of source containment and path blocking from the essence of fluid mechanics, and the auxiliary target addition mechanism based on causal extreme points constructs a primary and secondary coordinated three-dimensional prevention and control network. Through dynamic target positioning, the control of main risks and the prevention of secondary disasters in complex ventilation networks are solved.

[0033] Determine the fan / damper adjustment target, specifically including: obtaining the main gradient direction vector and causal extreme point coordinates in the risk field gradient characteristic vector to calculate the angle between the main gradient direction vector and the main airflow direction in the pre-acquired real-time ventilation network topology; when the angle is less than a preset angle threshold, mark the three-level fan node upstream of the gradient as a source containment type main adjustment target; when the angle is greater than or equal to the preset angle threshold, mark the branch damper node in the gradient normal plane as a path blocking type main adjustment target; based on the gradient modulus length sorting result of the causal extreme point coordinates, add an auxiliary adjustment target at the extreme point that is more than a preset distance away from the main adjustment target.

[0034] In one embodiment of this specification, the main gradient direction vector (a three-dimensional unit vector) and the coordinate set of the causal extreme point (a three-dimensional position sequence) are extracted from the risk field gradient feature vector. The current main wind flow direction vector is obtained by accessing the real-time updated ventilation digital twin via the mine's 5G private network. This vector can be generated by integrating CFD simulation with wind speed sensors. The spatial angle between the two vectors is calculated, and the cosine value is obtained using the dot product formula, which is then inversely solved to obtain the angle value.

[0035] When the angle is less than a preset threshold, the gas is determined to be diffusing along the main air duct. It should be noted that the preset threshold here can be set empirically, for example, it can be set to 30°. When the gas diffuses along the main air duct, it is traced back to the third-level node upstream of the gradient. Specifically, starting from the risk source, the three-hop distance is traced back along the ventilation network, and the high-power fan at the entrance of the main air duct is located as the source containment target, that is, the source containment type main regulation target. In the ventilation topology map, fans that are three hops away from the risk source are identified. The first hop is the fan directly connected to the lane where the risk source is located, the second hop is the fan one level in the direction of the main air duct, and the third hop is the fan at the entrance of the main air duct (controlling the global air volume). When the angle is greater than or equal to the threshold, the gas is determined to be diffusing perpendicular to the main airflow. The nearest branch damper node is searched in the normal plane of the gradient direction as the path blocking target, that is, the path blocking type main regulation target is obtained. At the same time, the extreme point coordinate set is sorted by gradient modulus length, and the extreme points with the highest modulus length are selected. If the distance between the extreme point and the primary target exceeds the equivalent roadway length (e.g., 50 meters), an auxiliary target is added to the fan / damper device closest to the target. The final output is a structured instruction set containing the primary / auxiliary target type and device ID to determine at least one fan / damper adjustment target.

[0036] Two gas diffusion modes are distinguished by the gradient-windflow angle. When the gas diffuses along the main air duct, the three-stage fan in the main air duct is locked to control more than 60% of the total air volume, and the gas is diluted by adjusting the global wind pressure. When the gas diffuses perpendicular to the main air flow, the valve-type damper on the diffusion path is accurately closed, which can block the backflow of gas in the goaf and avoid the sharp increase in energy consumption caused by traditional full pressurization. The auxiliary target addition mechanism based on causal extreme points effectively solves the problem of multi-source risk prevention and control. The control equipment is deployed at the two extreme points with the largest gradient modulus, so that the dual-source gas cloud is suppressed at the early stage of diffusion, which effectively improves the wind control efficiency compared with the single-target strategy. The characteristics of the ventilation network topology are obtained in real time, and the target positioning rules are automatically updated in the event of tunnel through-renovation or fan failure, getting rid of the rigid defects of the traditional preset target solution and providing an all-weather safety barrier for smart mines.

[0037] The fan / damper adjustment target is simulated to determine the dynamic safety constraint boundary, specifically including: constructing a lightweight ventilation network propagation model, wherein the lightweight ventilation network propagation model is a graph network model, and the graph network model includes a fan node for representing a pressure source variable, a damper node for representing a resistance variable, and a tunnel edge, and the tunnel edge is a resistance function containing a cross-sectional change rate; based on the fan / damper adjustment target, an initial target adjustment action instruction is determined, and the initial target adjustment action instruction of the adjustment target is input into the lightweight ventilation network propagation model, and a graph traversal algorithm is used to calculate the pressure difference change distribution data of the nodes in the entire network, and the standard deviation of the pressure difference fluctuation of the entire network is determined; the standard deviation of the pressure difference fluctuation of the entire network is used as an offset to correct the basic wind pressure safety threshold in the pre-acquired mine safety regulations database to determine the dynamic safety constraint boundary.

[0038] The traditional air adjustment process uses a static safety threshold, which cannot adapt to the dynamic environment underground. When a certain air door is closed to dilute local gas, it may cause a sudden drop in the pressure difference in the remote tunnel, resulting in a reversal of the wind flow, causing gas backflow in the goaf to form a new explosion source; the embodiment of this specification uses a graph network topology to simulate the propagation effect of the target action in real time to predict secondary risks; and the standard deviation of the pressure difference fluctuation is used as the basic threshold for dynamic offset correction, which quantifies the elastic space of the safety boundary of the same air adjustment action under different tunnel deformation states, ensuring the global feasibility of the control instructions from a physical level, and providing an action boundary for subsequent optimization that complies with safety regulations and adapts to real-time working conditions.

[0039] In one embodiment of this specification, after determining the fan / damper adjustment target, baseline action parameters are loaded from a pre-set equipment control specification database. For fan targets, the minimum effective adjustment step length under rated operating conditions is indexed based on the equipment model. For example, the minimum speed adjustment of a certain type of mining counter-rotating fan is a fixed percentage of the rated value. The adjustment direction (increase or decrease) is determined by combining the main direction sign (positive / negative) of the risk field gradient eigenvector. For damper targets, the minimum opening change for a single-step action is extracted based on the mechanical design specifications of the pneumatic actuator, such as the minimum effective rotation angle of a certain type of louvered damper. The direction of opening increase or decrease is determined based on the topological relationship between the gradient direction and the damper position. Then, an initial action instruction draft is generated. Each target point corresponds to a structured instruction item, which includes the device's unique identification code, action type (speed regulation or opening control), reference action amount and direction flag. For example, the instruction for a three-stage fan node is "Device ID: FAN-201, Action type: Speed ​​regulation, Reference amount: +Minimum adjustment step", and the instruction for a branch damper is "Device ID: VALVE-35, Action type: Opening control, Reference amount: -Minimum effective angle".

[0040] A lightweight ventilation network model is constructed within the graph computing engine of the edge computing node. Based on the underground mapping database, the fan is abstracted as a pressure source node, whose attributes include the rated pressure-flow curve. The damper is abstracted as a variable resistance node, whose attributes include an opening-resistance coefficient mapping table. The roadway is modeled as an attributed edge, and the resistance coefficient function includes the cross-sectional change rate reconstructed in real time by LiDAR. After receiving the initial target adjustment action command, it is parsed into device action parameters, such as the fan speed adjustment percentage and the damper target opening, and converted into node attribute update values ​​for injection into the model. An improved Dijkstra graph traversal algorithm is initiated to calculate pressure propagation: Starting from the action target, the pressure differential change is iteratively solved along the roadway edge, and the cross-sectional change factor in the resistance function is dynamically corrected. For example, if the cross-sectional shrinkage exceeds the limit, a local resistance multiplication is triggered. After the traversal is completed, the pressure differential change matrix of all nodes in the network is output and its standard deviation is calculated. The standard deviation is used to characterize the intensity of system fluctuations.

[0041] The standard deviation of the pressure difference fluctuation across the entire network is used as an offset to correct the basic wind pressure safety threshold in the pre-obtained mine safety regulations database. The process of determining the dynamic safety constraint boundary is as follows: First, the topological complexity parameters of the tunnel, which are updated in real time, are loaded to determine the topological sensitivity coefficient. These parameters include the number of tunnel branches and the distribution of local resistance mutation points. Specifically, a tunnel branch node mapping table is established, and node connectivity relationships in the ventilation network diagram are automatically extracted within the digital twin platform. The number of directly connected tunnel branches is counted. For example, if a transport tunnel connects three coal mining faces and two return air tunnels, the number of branches is five. Nodes with high wind resistance mutations are also marked, such as support areas with excessive cross-sectional contraction and turbulent areas detected by wind speed sensors. The rate of change of cross-sectional area along the tunnel is calculated. If continuous cross-sectional contraction exceeds a safety threshold, such as a cross-sectional area reduction ratio exceeding 15% of the design value, a node is marked as a structural mutation node in a support area with excessive cross-sectional contraction. If abnormal pressure difference fluctuations at a node persist for longer than a preset period, such as if the standard deviation of the pressure difference fluctuations is continuously greater than twice the normal value, a node is marked as a fluid anomaly node in a turbulent area detected by wind speed sensors.

[0042] Based on topological features, the original coefficient is calculated. The initial scaling factor is calculated by dividing the total number of branches by the baseline number of branches. The baseline number of branches is the average of the historical safe operation samples for the mining area. High-resistance node weights are added, assigning each high-resistance node a fixed incremental value (e.g., 0.1) based on historical accident data. For example, a typical ventilation network contains twelve main roadway branches, three of which are located in high-resistance areas within a geologically fractured zone. High-resistance node identification results include one structural abrupt change node located in the fractured zone's cross-sectional contraction zone, as determined by LiDAR, and two fluid anomaly nodes located in an area of ​​abnormal pressure fluctuation, as determined by sensors. The incremental calculation is: number of high-resistance nodes = 3 × single-point increment of 0.1, for a total increment of 0.3. First, the branch scaling factor is calculated (current value 12 / baseline value 10 = 1.2), and then the high-resistance weight increment (3 × 0.1 = 0.3) is added, resulting in a topological sensitivity coefficient of 1.5. When the mine excavated a new working face, increasing the number of branches to fifteen and adding two new fault-induced high-wind resistance areas, the topological sensitivity coefficient was automatically updated to (15 / 10) + (5×0.1) = 2.0. At this point, if the pressure differential standard deviation reached a specific value, the generated dynamic offset automatically expanded, significantly tightening the safety constraint boundary relative to the stable working condition. This successfully intercepted the damper closing command in the fracture zone, preventing gas accumulation caused by airflow stagnation. This design, which quantifies abstract topological features into engineering parameters, enables the system to accurately perceive changes in network structure risks.

[0043] Afterwards, basic wind pressure thresholds are loaded from the mine safety regulations database, including the minimum wind pressure threshold P_min to prevent airflow stagnation and the maximum wind pressure threshold P_max to avoid confined wall rupture. When the standard deviation of the regional pressure difference fluctuation reaches a specific value set based on experience, this standard deviation is used as a dynamic offset to generate a safety envelope with synchronously contracting upper and lower bounds. Specifically, the standard deviation is multiplied by the topological sensitivity coefficient to generate a dynamic offset Δ. The specific correction logic is that when the standard deviation shows that the system fluctuates violently, that is, when it reaches a specific value set based on experience, Δ is added to the basic lower threshold P_min to increase safety redundancy, and Δ is subtracted from the basic upper threshold P_max to tighten the constraints; conversely, when the system is stable, the boundaries are appropriately relaxed. For example, during a tunnel penetration operation in the Huainan mining area, simulations revealed a significant increase in the standard deviation of the pressure differential at the moment of penetration. The system automatically calculated a positive offset Δ and generated a dynamic constraint boundary [P_min+Δ, P_max-Δ]. This tightened safety envelope successfully intercepted the original damper closing command, preventing a wind flow reversal accident caused by the penetration shock wave. However, the static threshold solution, which had not anticipated this risk, led to localized gas accumulation. This dynamic correction mechanism adheres to the baseline requirements of safety regulations while also providing the system with the flexibility to adapt to real-time operating conditions. The resulting dynamic safety constraint boundary is linked to the optimization objective function, forming a physical guardrail for command generation. For example, during a tunnel penetration operation, the basic threshold was fixed, but simulations revealed a significant increase in the standard deviation of the pressure differential fluctuation at the moment of penetration. This high standard deviation was automatically detected, and a stricter safety boundary was immediately generated to prevent air adjustment commands that could cause wind flow disturbances. This addresses the safety concerns of traditional methods that still use fixed thresholds in the face of sudden changes in operating conditions.

[0044] In one embodiment of the present specification, a preset optimization objective function is solved under the dynamic safety constraint boundary to generate a fan speed adjustment instruction and a damper opening control instruction to achieve fan / damper control. It should be noted that the optimization objective function here is as follows: ,in, is the equipment adjustment value of the i-th target point (fan / damper), such as the fan speed / damper opening adjustment value, m means there are m adjustment targets, is the equipment wear coefficient, which can be set according to the actual wear of the target equipment. β is the causal alignment weight, which can be set to 0.8. ▽R is the current risk gradient corresponding to the current risk field gradient feature vector. is the target gradient, that is, the ideal distribution after dilution.

[0045] After determining the optimization objective function (minimizing the weighted sum of the total number of equipment movements and the risk gradient deviation) and the dynamic safety constraint bounds (the pressure differential envelope output by the ventilation network propagation model), a hierarchical solution strategy was designed for different equipment types. For discrete variables such as damper opening, a branch-and-bound approach was used to traverse the feasible solution space, incorporating a taboo search strategy to avoid local optima. For continuous variables such as fan speed, a sequential quadratic programming approach was used for iterative solution, with dynamic constraint compliance verified after each iteration. After optimization, a structured result was output, including the global optimal solution that satisfies all constraints (the aggressive solution), the feasible solution with the smallest movement amplitude (the basic solution), and the historical optimal solution during convergence (the backup solution). Industrial command conversion was then performed. The fan speed solution was converted through linear mapping to a value written to a holding register in the Modbus-TCP protocol. For example, the speed percentage was converted to a hexadecimal function code 06 message. The damper opening angle solution was then encapsulated as a process data object location instruction word based on the CANopen protocol object dictionary address. The resulting instruction packet was appended with a timestamp and CRC checksum and transmitted to the PLC control cabinet via the industrial ring network.

[0046] Traditional mine air conditioning systems use manually set fixed parameters or PID control. When the PID controller is not loaded with airflow network constraints, continuously increasing the fan speed will cause gas backflow in the negative pressure area of ​​the tunnel; by optimizing the causal alignment term (risk gradient deviation) in the objective function in the embodiment of this specification, the control instructions are forced to comply with the fluid diffusion law; through the three types of output plans of radical, basic and backup, a tiered guarantee is formed. When the execution of the main plan is blocked (such as a fan failure), the backup plan is switched in milliseconds, which effectively improves the fault response speed tenfold compared with the traditional single-plan design; through the native protocol encapsulation of Modbus and CANopen, the traditional method avoids the need for secondary configuration of the configuration software, reducing the error rate of instruction conversion.

[0047] When the decision path is a Monte Carlo plan set, the method also includes: generating an initial set of Monte Carlo plans based on the risk field gradient eigenvector, simulating ventilation effects and screening confidence levels of the initial set of Monte Carlo plans through a preset evaluation model; and inputting high-confidence plans with confidence levels greater than a preset confidence threshold into a preset optimization objective function to solve optimal control instructions, so as to parse and execute the optimal control instructions. The ventilation effect simulation and confidence screening of the initial set of Monte Carlo plans are performed through a preset evaluation model, specifically including: determining a preset lightweight CFD proxy model, wherein the input of the lightweight CFD proxy model is the fan speed and damper opening parameters of each plan in the initial set of Monte Carlo plans, and the output of the lightweight CFD proxy model is the gas concentration distribution of the entire network and the node pressure difference change; through the lightweight CFD proxy model, the ventilation network fluid dynamics simulation is performed on each plan to determine the gas concentration distribution prediction data of the entire network and the node pressure difference change prediction data corresponding to each plan; according to the gas concentration distribution prediction data of the entire network and the node pressure difference change prediction data corresponding to each plan, the plan confidence of each plan is determined, so as to perform confidence screening based on the plan confidence and the preset confidence threshold.

[0048] In high-risk mine gas outbreak scenarios, when risk entropy indicators indicate a high level of system uncertainty, such as when sensor failure coincides with a gas release, the physical models that conventional causal decision-making relies on may completely fail. For example, if the gradient direction randomly drifts due to data distortion, blindly executing target adjustments can lead to catastrophic misjudgments.

[0049] In one embodiment of this specification, when the decision path is a Monte Carlo plan set, an initial plan set is generated based on the risk field gradient feature vector. Cases with high physical environment similarity (e.g., roadway topology matching and gradient modulus range) are retrieved from the historical operating condition database. These cases are expanded through constrained random sampling to include multiple fan / damper operation combinations (e.g., fan A speed-up and damper B closing combinations, global fan speed-down combinations, etc.). Specifically, the main gradient direction vector, risk entropy value, and causal extreme point coordinates from the risk field gradient feature vector are obtained. Based on the risk field gradient feature vector, a pre-constructed historical operating condition database is retrieved. A spatial similarity algorithm is used to match historical cases whose roadway topology (e.g., number of branches, distribution of high-drag nodes) matches the current gradient features. Sample sets with physical environment similarity exceeding a preset threshold are selected. For example, if the main gradient direction is detected as pointing toward an abandoned roadway and three high-drag nodes are present, one hundred effective disposal records from similar scenarios over the past six months are automatically retrieved. Based on these selected samples, an initial plan pool is generated, and constrained random sampling is performed on the fan / damper operating parameters in the samples. Using historical equipment actions as the central point, Gaussian distribution sampling is performed within the equipment's safe operating boundaries, such as the fan speed adjustment range ± a percentage of the rated value and the damper opening angle safety interval. Directional constraints are also applied to the current gradient vector, such as limiting the speed reduction sampling weight when the gradient points toward the air intake. For the causal extreme point coordinate set, additional equipment disturbance plans are added. A search is conducted for available fans / damper units within the extreme point radius, generating supplementary action combinations for localized boosting or isolation. The final output is an initial set of plans that integrate historical experience with real-time constraints.

[0050] Then the pre-deployed lightweight CFD proxy model is called. The lightweight CFD proxy model consists of a graph neural network trained with full-scale CFD simulation data. The input is the equipment parameters in the initial plan set, including the fan speed percentage and the damper opening angle. The output is the gas concentration distribution of the entire network and the predicted data of the node pressure difference change. According to the simulation results of each plan, the plan confidence is calculated, and the gas dilution efficiency (the concentration drop rate in the target area) and the pressure difference stability (node ​​fluctuation amplitude) are comprehensively evaluated. Specifically, the gas concentration drop rate of the key target area is first extracted, and the ratio of this rate to the maximum theoretical dilution rate is normalized into a gas dilution efficiency index; the node pressure difference fluctuation data is simultaneously analyzed, and the peak-to-average ratio of the pressure difference change of the entire network and the proportion of time exceeding the maximum allowable fluctuation amplitude of the safety threshold are calculated to generate a pressure difference stability coefficient. Pressure difference stability coefficient The calculation formula is as follows: ; in, α is the peak-to-average ratio weight, which can be set to 0.3. β is the weight of the excess ratio, which can be set to 0.5. Indicates the maximum absolute value of the node pressure difference change during the simulation cycle, Indicates the arithmetic mean of the pressure difference changes of all nodes in the network. Indicates the cumulative time that the node pressure difference exceeds the allowable fluctuation amplitude set in the mine safety regulations. Indicates the total simulation time, γ is the high wind resistance attenuation factor. When there is a high wind resistance node, ,λ is the attenuation coefficient, which can be set to 0.2, is the number of high-resistance areas obtained through topological analysis. When there are no high-resistance nodes, γ=1. The gas dilution efficiency index and the pressure difference stability coefficient are weighted and summed to generate a scalar confidence value. Finally, the confidence is compared with the preset confidence threshold. The preset confidence threshold here can be set according to the complexity of the tunnel topology network. Only multiple high-confidence plans with confidence greater than the preset confidence threshold are retained. After screening out the set of high-confidence plans, the optimization objective function is input for solution to obtain the optimal plan. The final plan is converted into the corresponding optimal control instruction. The optimization objective function here and the conversion of the optimal plan into the optimal control instruction have been explained in the previous steps of the causal decision path and will not be repeated here.

[0051] Through the above technical solution, by generating a set of plans covering multiple probabilities and implementing physical feasibility verification based on a lightweight CFD agent model, and using parallel simulation, options that violate the laws of fluid mechanics or trigger secondary risks are screened out, ensuring that the instructions finally executed have the dual guarantees of mathematical optimality and physical compliance.

[0052] Through the technical solutions of the embodiments of this specification, traditional single-point sensors can only perceive the instantaneous value of local concentration. However, the embodiments of this specification construct a complete information field of gas diffusion at the fluid mechanics level through the spatiotemporal alignment mechanism of multimodal data cubes. The tunnel curvature characteristics extracted from the LiDAR point cloud are coupled with the turbulence prior of the CFD wind flow grid, so that the spatiotemporal Transformer model can capture the causal chain of pressure difference fluctuations, wind speed attenuation, and concentration accumulation, so that the early warning mechanism is transformed from passive concentration limit detection to active diffusion dynamics deduction; traditional manual air adjustment requires early warning and decision-making. , and a multi-level delay chain of execution, while the causal decision path of the embodiment of this specification realizes the integration of decision execution at the edge computing end through real-time matching of the gradient field and the network topology, and the optimal control target selection based on the network transmission efficiency, in conjunction with the direct drive of the protocol-level instructions, effectively compresses the full-link reaction time of perception, decision-making, and execution; conventional static safety thresholds cannot adapt to dynamic risks such as tunnel deformation. The ventilation network propagation model of the embodiment of this specification constructs a virtual test field for air adjustment instructions, quantifies structural risks into elastic protection parameters, and realizes the conversion from empirical safety margin to calculated safety boundary. For situations with high confidence in the gradient direction, the causal path is enabled to achieve the physical optimal solution. On the contrary, the group intelligent search of the Monte Carlo plan is adopted to directly realize the intelligent switching of the decision mode, effectively improving the control accuracy and control effect of mine gas, and meeting the rigid safety requirements of smart mines.

[0053] The embodiment of this specification also provides a mine gas control device based on multimodal data, such as Figure 2 As shown, the device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above method.

[0054] The embodiments of this specification also provide a non-volatile computer storage medium storing computer executable instructions, wherein the computer executable instructions are configured to execute the above method.

[0055] The various embodiments in this specification are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simplified. For relevant details, refer to the descriptions of the method embodiments.

[0056] The foregoing description is merely one or more embodiments of this specification and is not intended to limit this specification. It will be apparent to those skilled in the art that various modifications and variations may be made to one or more embodiments of this specification. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of one or more embodiments of this specification are intended to be within the scope of the claims of this specification.

Claims

1. A mine gas control method based on multimodal data, characterized in that: The method comprises: Acquire multi-source real-time sensor data, roadway point cloud data, and CFD airflow simulation grid data within the target mine roadway, perform spatiotemporal alignment, determine a real-time roadway data cube, and use a preset multimodal deep causal model and the real-time roadway data cube to determine the gas concentration field distribution and causal contribution heat data within a preset future time window; generating a risk field gradient feature vector based on the gas concentration field distribution and the causal contribution heat data, and matching a decision path according to the risk field gradient feature vector, wherein the decision path includes a Monte Carlo plan set and a causal decision path; When the decision path is a causal decision path, the fan / damper adjustment target is determined, the fan / damper adjustment target is simulated, and the dynamic safety constraint boundary is determined to solve the preset optimization objective function under the dynamic safety constraint boundary, generate the fan speed adjustment instruction and the damper opening control instruction, and realize the control of the fan / damper.

2. A mine gas control method based on multimodal data according to claim 1, characterized in that: Determine the real-time lane data cube, including: Acquire multi-source real-time sensor data in the target mine tunnel in real time, wherein the multi-source real-time sensor data includes gas concentration time series data, temperature and humidity time series data, and pressure difference time series data; Time-stamp synchronization is performed on the multi-source real-time sensor data using a high-precision clock service to achieve time alignment, and a preset algorithm is used to spatially align the roadway point cloud data and the CFD wind flow simulation grid data in a unified three-dimensional coordinate system to determine roadway grid data, wherein the roadway grid data includes roadway grid coordinates and roadway point cloud geometric features; The time-space aligned data is encapsulated into real-time lane data cubes of multiple preset dimensions, wherein the preset dimensions include time dimension, space dimension and feature dimension. The space dimension is mapped to the lane grid coordinates, and the feature dimension includes sensor measurement indicators and lane point cloud geometric features.

3. The mine gas control method based on multimodal data according to claim 1, characterized in that: Based on the gas concentration field distribution and the causal contribution heat data, a risk field gradient feature vector is generated, specifically including: Calculating spatial partial derivatives of the gas concentration field distribution along the three-dimensional coordinate direction of the roadway to generate an initial gradient vector field; Using the causal contribution heat data as a weight matrix, performing a Hadamard product operation with the initial gradient vector field to generate a weighted risk gradient field; The main gradient direction vector, risk entropy index and causal extreme point coordinates are extracted from the weighted risk gradient field and encapsulated into a structured feature vector to determine the risk field gradient feature vector.

4. The mine gas control method based on multimodal data according to claim 1, characterized in that: Matching a decision path according to the risk field gradient feature vector specifically includes: Determining a risk entropy value index in the risk field gradient feature vector; When the risk entropy value indicator exceeds a preset dynamic threshold, the Monte Carlo plan generation engine is activated to output a Monte Carlo plan set as a decision path; When the risk entropy value indicator does not exceed the preset dynamic threshold, the decision path is determined to be a causal decision path.

5. The mine gas control method based on multimodal data according to claim 1, characterized in that: Determine the fan / damper adjustment target, including: Obtaining the main gradient direction vector and the causal extreme point coordinates in the risk field gradient characteristic vector to calculate the angle between the main gradient direction vector and the main wind flow direction in the pre-acquired real-time ventilation network topology; When the angle is less than a preset angle threshold, the three-level wind turbine node upstream of the gradient is marked as a source containment type main regulation target; When the angle is greater than or equal to a preset angle threshold, the branch damper node in the gradient normal plane is marked as a path blocking type main adjustment target point; Based on the gradient modulus length sorting result of the causal extreme point coordinates, an auxiliary adjustment target point is added at the extreme point that is more than a preset distance away from the main adjustment target point.

6. A mine gas control method based on multimodal data according to claim 5, characterized in that: Simulate the fan / damper adjustment target to determine the dynamic safety constraint boundary, specifically including: Constructing a lightweight ventilation network propagation model, wherein the lightweight ventilation network propagation model is a graph network model, the graph network model including a fan node for representing a pressure source variable, a damper node for representing a resistance variable, and a roadway edge, wherein the roadway edge is a resistance function including a cross-sectional change rate; Based on the fan / damper adjustment target, an initial target adjustment action instruction is determined, and the initial target adjustment action instruction of the adjustment target is input into the lightweight ventilation network propagation model, and a graph traversal algorithm is used to calculate the pressure difference change distribution data of the nodes in the entire network, and the standard deviation of the pressure difference fluctuation of the entire network is determined; The standard deviation of the pressure difference fluctuation of the entire network is used as an offset to correct the basic wind pressure safety threshold in the pre-acquired mine safety regulations database to determine the dynamic safety constraint boundary.

7. The mine gas control method based on multimodal data according to claim 1, characterized in that: When the decision path is a Monte Carlo plan set, the method further includes: Generate an initial set of Monte Carlo emergency plans based on the risk field gradient feature vector, and perform ventilation effect simulation and confidence screening on the initial set of Monte Carlo emergency plans through a preset evaluation model; The high-confidence plan with a confidence level greater than a preset confidence threshold is input into a preset optimization objective function to solve the optimal control instruction, so as to parse and execute the optimal control instruction.

8. The mine gas control method based on multimodal data according to claim 7, characterized in that: The ventilation effect simulation and confidence screening of the initial set of Monte Carlo plans are performed using a preset evaluation model, specifically including: Determine a preset lightweight CFD proxy model, wherein the input of the lightweight CFD proxy model is the fan speed and damper opening parameters of each plan in the initial set of Monte Carlo plans, and the output of the lightweight CFD proxy model is the gas concentration distribution of the entire network and the change in node pressure difference; Through the lightweight CFD proxy model, ventilation network fluid dynamics simulation is performed for each plan to determine the corresponding network gas concentration distribution prediction data and node pressure difference change prediction data for each plan; The plan confidence of each plan is determined according to the network-wide gas concentration distribution prediction data and the node pressure difference change prediction data corresponding to each plan, so as to perform confidence screening based on the plan confidence and the preset confidence threshold.

9. A mine gas control device based on multimodal data, characterized in that: The device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor. The instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions are configured to execute the method according to any one of claims 1 to 8.

Images