A method, equipment and medium for mine gas control based on multimodal data

Abstract

This specification discloses a method, equipment, and medium for mine gas control based on multimodal data, relating to the field of gas control technology. The method includes: acquiring multi-source real-time sensor data, roadway point cloud data, and CFD airflow simulation grid data in a target mine roadway; performing spatiotemporal alignment to determine a real-time roadway data cube; determining the gas concentration field distribution and causal contribution heat data within a preset future time window using a multimodal deep causal model and the real-time roadway data cube; generating a risk field gradient feature vector based on the gas concentration field distribution and causal contribution heat data; matching a decision path based on the risk field gradient feature vector; determining the fan / damper adjustment target point when the decision path is a causal decision path; simulating the fan / damper adjustment target point to determine the dynamic safety constraint boundary; solving a preset optimization objective function to generate fan speed adjustment commands and damper opening control commands to achieve fan / damper control.

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 Technology

[0002] In the field of coal mine safety production, gas disaster prevention and control has always been a core challenge in ensuring personnel safety and efficient resource development. Existing technologies mainly achieve this through the following methods: one is a passive monitoring system based on discrete single-source sensor networks, which involves deploying gas concentration sensors at key nodes in the roadway and triggering audible and visual alarms by setting fixed thresholds; another is a ventilation control mechanism driven by human experience, where dispatchers manually adjust fan speeds and damper openings based on alarm signals and empirical rules; and yet another is using offline computational fluid dynamics (CFD) simulation to assist ventilation design, optimizing the static ventilation network through periodic airflow simulation. However, these technologies are gradually revealing systemic flaws when dealing with the dynamic risks of modern mines.

[0003] In real-time risk perception, single-point sensors can only capture transient values ​​of local concentrations and cannot characterize the spatiotemporal evolution trend of gas diffusion, resulting in a short warning window. Furthermore, they can only identify the surface symptoms of excessive gas levels but cannot locate the source of the release or predict the diffusion path, further weakening the effectiveness of the warning. With the deepening 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 LiDAR 3D point clouds of roadways. However, existing research mostly focuses on machine learning predictions from single data sources or at single scales, lacking unified modeling of multi-source heterogeneous information and explicit characterization of physical causal laws. It also fails to form a closed-loop control system between risk prediction and automatic ventilation. For example, traditional systems only alarm when the concentration exceeds the limit, unable to predict the risk accumulation process in unexploded areas. This leads to a severe lack of buffer time for underground personnel evacuation and emergency response. Consequently, in the process of generating ventilation decisions, manual ventilation relies on operator experience and lacks real-time physical guidance, resulting in response lag and amplified operational errors.

[0004] Therefore, in the process of mine gas disaster prevention and control, the focus is often on machine learning prediction based on a single data source or a single scale, and passive response is made when the concentration exceeds the standard. This results in a short early warning window and low control efficiency, which makes it difficult to meet the rigid safety requirements of smart mines. Summary of the Invention

[0005] This specification provides one or more embodiments of a mine gas control method, device, and medium based on multimodal data to solve the following technical problem: In the process of mine gas disaster prevention and control, the focus is often on machine learning prediction based on a single data source or a single scale, and passive response is made when the concentration exceeds the standard, resulting in a short early 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 employ the following technical solutions:

[0007] This specification provides one or more embodiments of a mine gas control method based on multimodal data. The method includes: acquiring multi-source real-time sensor data, roadway point cloud data, and CFD airflow simulation grid data within a target mine roadway, performing spatiotemporal alignment to determine a real-time roadway data cube; determining the gas concentration field distribution and causal contribution heat data within a preset future time window using a preset multimodal depth causal model and the real-time roadway data cube; generating a risk field gradient feature vector based on the gas concentration field distribution and the causal contribution heat data; matching a decision path according to the risk field gradient feature vector, wherein the decision path includes a Monte Carlo contingency plan set and a causal decision path; when the decision path is a causal decision path, determining a fan / damper adjustment target point; simulating the fan / damper adjustment target point to determine a dynamic safety constraint boundary; solving a preset optimization objective function under the dynamic safety constraint boundary to generate a fan speed adjustment command and a damper opening control command to achieve fan / damper control.

[0008] This specification provides one or more embodiments of a mine gas control device based on multimodal data, comprising:

[0009] At least one processor; and,

[0010] A memory communicatively connected to the at least one processor; wherein,

[0011] The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the above-described method.

[0012] This specification provides one or more embodiments of a non-volatile computer storage medium storing computer-executable instructions configured to perform the above-described method.

[0013] The above-mentioned at least one technical solution 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 sense the instantaneous value of local concentration, while the embodiments of this specification, through the spatiotemporal alignment mechanism of multimodal data cubes, construct a complete information field of gas diffusion at the fluid dynamics level. The roadway curvature features extracted by LiDAR point cloud are coupled with the turbulence prior of CFD airflow grid, enabling the spatiotemporal Transformer model to capture the causal chain of pressure difference fluctuations, wind speed attenuation, and concentration accumulation, thus transforming the early warning mechanism from passive concentration exceedance detection to active diffusion dynamics deduction. Traditional manual ventilation control involves a multi-level delay chain of early warning, decision-making, and execution. In contrast, the causal decision-making path in this embodiment achieves integrated decision-making and execution at the edge computing end through real-time matching of gradient fields and network topology. Based on optimal control target selection for network transmission efficiency, coupled with protocol-level direct drive, it effectively compresses the entire chain reaction time of perception, decision-making, and execution. Conventional static safety thresholds cannot adapt to dynamic risks such as roadway deformation. The ventilation network propagation model in this embodiment constructs a virtual test field for ventilation control commands, quantifying structural risks into elastic protection parameters, and realizing the transformation from empirical safety margins to computational safety boundaries. For cases with high confidence in the gradient direction, a causal path is activated to achieve the physically optimal solution; conversely, a Monte Carlo scheme-based swarm intelligence search is used to directly achieve intelligent switching of decision-making modes, effectively improving the control accuracy and effect of mine gas and meeting the rigid safety requirements of intelligent mines. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments or prior art of this specification, the drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0015] Figure 1 A flowchart illustrating a mine gas control method based on multimodal data, provided as an embodiment of this specification;

[0016] Figure 2 This is a schematic diagram of a mine gas control device based on multimodal data, provided as an embodiment of this specification. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0018] This specification provides a method for controlling mine gas based on multimodal data. It should be noted that the execution entity in this specification can be a server or any device with data processing capabilities. Figure 1 A flowchart illustrating a mine gas control method based on multimodal data, provided as an embodiment of this specification, is shown below. Figure 1 As shown, the main steps include the following:

[0019] Step S101: Acquire multi-source real-time sensor data, roadway point cloud data, and CFD airflow simulation grid data in the target mine roadway, and perform spatiotemporal alignment to determine the real-time roadway data cube. Using the preset multimodal depth causal model and the real-time roadway data cube, determine the gas concentration field distribution and causal contribution heat data within the preset future time window.

[0020] In one embodiment of this 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 dynamic parameters of the roadway environment in real time. Infrared gas sensors installed at key nodes in the coal face, return airway, and main haulage roadway generate concentration time-series curves at a second-level frequency. Electrochemical temperature and humidity transmitters record changes in air state, and micro-differential pressure transmitters monitor airflow pressure fluctuations. A mine-use explosion-proof mobile LiDAR device periodically scans the three-dimensional morphology of the roadway to generate a point cloud dataset with millimeter-level precision; its scanning trajectory covers the entire cross-section and support structure. Simultaneously, based on the mine's digital twin platform, the latest roadway topology model is invoked, and steady-state / transient fluid dynamics simulations are performed in ANSYS Fluent, outputting CFD structured grid data containing wind speed vectors, turbulent kinetic energy, and static pressure distribution.

[0021] The determination of the real-time roadway data cube specifically includes: acquiring multi-source real-time sensor data within the target mine roadway in real time, including time-series data of gas concentration, temperature and humidity, and differential pressure; synchronizing the multi-source real-time sensor data with timestamps using a high-precision clock service for time alignment; and using a preset algorithm to spatially align the roadway point cloud data and the CFD airflow simulation grid data in a unified three-dimensional coordinate system to determine the roadway grid data, which includes roadway grid coordinates and roadway point cloud geometric features; and encapsulating the spatiotemporally aligned data into a real-time roadway data cube with multiple preset dimensions, including a time dimension, a spatial dimension, and a feature dimension. The spatial dimension is mapped to the roadway grid coordinates, and the feature dimension includes sensor measurement indicators and roadway point cloud geometric features.

[0022] In the underground environment, the dynamics of gas concentration are affected by multiple factors coupled together, including roadway morphology, airflow distribution, and equipment status. If only isolated sensor data is relied upon, it is impossible to capture the diffusion path of gas along the curved roadway after it surges out of the working face. If the abrupt changes in the roadway cross-section of LiDAR scanning are ignored, the model will underestimate the risk of gas accumulation caused by local eddies. Furthermore, without the prior knowledge of fluid mechanics from CFD simulation, the prediction results are prone to violating the law of conservation of mass and exhibiting non-physical phenomena such as backflow diffusion.

[0023] In one embodiment of this specification, time-series data on gas concentration, temperature and humidity, and differential pressure are aggregated to edge computing nodes via an industrial ring network. A data cube is then constructed through three-layer fusion processing. 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 timescale. At the spatial layer, SLAM mapping and outlier filtering are performed on the LiDAR point cloud. The ICP iterative nearest point algorithm is used to register the point cloud to the mine's absolute coordinate system (based on underground survey control points). Simultaneously, bounding box matching maps the CFD mesh nodes to a unified coordinate frame. At the feature layer, the registered spatial data is voxelized and encapsulated. Each voxel unit (e.g., a discrete grid segmented along the direction of the roadway) is filled with sensor time-series curves, local cross-sectional areas and surface curvatures extracted from the point cloud, and the reference wind speed vector provided by the CFD, among other feature channels. This ultimately forms a data cube with a continuous time axis, consistent spatial index, and complete feature dimensions, stored in a columnar in-memory database for real-time model access.

[0024] To align discrete temporal streams with periodic scan data, bidirectional buffering is implemented. Second-level sensor streams are batched using a 10-second sliding window. LiDAR point clouds are downsampled to voxel grids (0.05 m) and finely registered using ICP. The CFD simulation mesh (calculated offline by software such as Fluent) is aligned with the point cloud coordinate system using minimum bounding box coordinates, forming a unified 3D reference frame to prepare for subsequent spatiotemporal encoding. The second-level sensor batches first enter the streaming engine (Apache Flink + Kafka), where denoising, null value compensation, and outlier screening are performed on the edge boxes. Spatial indexes are then added, mapping each sensor sample to the nearest CFD grid cell and point cloud voxel. Simultaneously, point cloud data undergoes normal vector estimation and surface reconstruction on the GPU, calculating geometric features such as tunnel cross-sectional area and potential airflow damping. The fused data cube is stored in an Arrow format in-memory table according to the (time × space × feature) dimension, facilitating Tensor passthrough; in addition, columnar Delta Lake is used to manage versions, ensuring that historical data can be replayed. This stage also calculates statistical priors (mean, variance, quantiles) for key variables such as gas concentration, pressure difference, and wind speed and writes them to the metadata repository for online normalization during model training and inference.

[0025] In one embodiment of this specification, the multimodal deep causal model first includes a spatiotemporal Transformer-MoE architecture and a self-supervised causal masking mechanism. It encodes time-series sensor data into temporal channel features and 3D point clouds and CFD meshes into spatial channel features. An expert-gated network dynamically selects attention experts of appropriate scales, balancing long-range dependencies and local details. Through the self-supervised causal masking mechanism, the "gas-pressure difference-wind speed" key sequence is randomly masked during the training phase, and its missing parts 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 the sensitivity and interpretability of sudden gas escapes. Using the pre-defined multimodal deep causal model and real-time tunnel data cubes, the gas concentration field distribution and causal contribution heat data within a pre-defined future time window are determined.

[0026] Data cubes from the past three months are read from the training center (4×A100 GPU servers or domestic computing cards), and divided into training, validation, and test sets according to 70% / 15% / 15%. The spatiotemporal Transformer-MoE architecture employs 8 layers of temporal attention, 4 layers of spatial attention, and 6 expert-gated branches. During the self-supervised phase, continuous segments of "gas concentration-pressure difference-wind speed" are randomly occluded, aiming to minimize the reconstruction error of the occluded area and simultaneously minimize the physical residual loss (consistency between CFD prior and predicted wind speed). After training, the model is exported to the ONNX + TensorRT engine and deployed according to a two-tier deployment strategy: offline redundant inference at the central station and low-latency inference at the wellhead Edge Box. The Edge Box uses Jetson Orin or domestic AI chips, with an inference batch size of 1 and latency controlled at around 100 ms. The central station deployment version has an additional one-order model capacity for nighttime batch replay and accuracy verification.

[0027] By using a pre-defined multimodal deep causal model and a real-time roadway data cube, the gas concentration field distribution and causal contribution heat data within a pre-defined future time window are determined. In the multimodal deep causal model deployed on the underground edge computing node, the real-time roadway data cube is loaded as an input tensor into the spatiotemporal Transformer-MoE architecture for feature parsing. This architecture first processes the data through separate attention channels. The temporal self-attention layer focuses on the long-range dependencies in the sensor time-series flow, capturing the cross-period correlation between gas concentration and differential pressure fluctuations. The spatial self-attention layer integrates the roadway curvature features extracted from LiDAR point clouds with the wind speed vector field of the CFD mesh, learning the fluid motion pattern under spatial topological constraints. The expert-gated network dynamically routes cross-scale features; for example, when a sudden narrowing area of ​​the roadway is detected in the point cloud, high-resolution spatial expert-enhanced eddy current modeling is automatically activated.

[0028] By pre-embedding a self-supervised causal masking mechanism during model training, key sequence segments of "gas concentration-pressure difference-wind speed" are randomly masked, forcing the model to reconstruct the masked physical quantities through the remaining features. This allows the network to explicitly grasp the causal chain of "gas release from coal wall fissures → local pressure difference decrease → wind speed reduction → concentration accumulation." During inference, the model output includes dual-channel results: the first is the three-dimensional gas concentration field distribution in the future time window, i.e., the concentration prediction tensor of each voxel unit within the roadway grid coordinates over time, which can intuitively display high-risk accumulation areas (such as the concentration gradient change at the corner of the return airway); the second is causal contribution heat data. By quantifying the activation intensity of the model attention weights at key causal chain nodes, a heat map covering the roadway space is generated (e.g., indicating the contribution of pressure difference monitoring points to the concentration prediction of the working face). When the heat value of a certain area increases sharply, it indicates the existence of a hidden risk source not seen in the sensor data. Ultimately, the concentration field distribution and causal heat data together constitute a digital twin of risk prediction, providing a situational awareness foundation with both physical interpretability and spatiotemporal accuracy for subsequent decision-making.

[0029] Through the above technical solutions, multimodal dynamic coupling enables the prediction model to synchronously perceive changes in the physical environment and airflow conditions. For example, when LiDAR detects delamination of the working face roof, it automatically associates the porosity features extracted from the point cloud of that area with the CFD eddy current model. Combined with the abnormal micro-pressure difference of the sensor, it predicts the risk of gas outburst from the delamination in advance, significantly extending the warning window compared to traditional single-data source analysis. The unified spatiotemporal reference eliminates cross-source information bias. In complex ventilation networks, voxel indexing accurately locates the complete causal chain of "sudden increase in concentration at sensor A → point cloud shows a roadway change on the downwind side of that point → CFD indicates the formation of a eddy current zone at the change point," guiding targeted regulation to avoid secondary risks and reducing the number of ineffective equipment actions compared to manual experience-based ventilation strategies. The columnar storage and voxelized feature extraction of the data cube allow the Transformer model to directly load tensors for inference, avoiding the latency problem of repeatedly calling multi-source data on the central server in traditional methods, and meeting the minute-level closed-loop control timeliness requirements in underground mines.

[0030] Step S102: Based on the gas concentration field distribution and causal contribution heat data, generate a risk field gradient feature vector to match the decision path according to the risk field gradient feature vector.

[0031] This decision-making path includes the Monte Carlo contingency plan set and the causal decision-making path;

[0032] Based on the gas concentration field distribution and the causal contribution heat data, a risk field gradient feature vector is generated. Specifically, this includes: calculating the spatial partial derivative 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 and performing a Hadamard product operation with the initial gradient vector field to generate a weighted risk gradient field; and extracting the principal gradient direction vector, risk entropy index, and causal extreme point coordinates from the weighted risk gradient field, encapsulating them into a structured feature vector to determine the risk field gradient feature vector.

[0033] Simply relying on the distribution of gas concentration fields can only provide a static risk snapshot and cannot reveal the dynamic mechanism and evolution direction of gas diffusion. For example, in the bifurcation area of ​​a roadway, the same concentration value may originate from upstream outflow or local accumulation, and the control strategies required for the two are completely different.

[0034] In one embodiment of this specification, a predicted gas concentration field distribution tensor is loaded using a GPU-accelerated environment on an edge computing node. The tensor is calculated as time × spatial grid × concentration value, with spatial partial derivatives calculated along three dimensions: 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 convolve the grid concentration values ​​to generate an initial gradient vector field, where each grid point contains a direction vector and a magnitude. Simultaneously, the causal contribution heat data matrix is ​​read, and the heat values ​​are mapped to weight intervals using Min-Max normalization, forming a weight tensor of the same dimension as the gradient field. The Hadamard product operation is performed, multiplying the gradient vector by the corresponding heat weight at each grid point. This significantly enhances the gradient signal in high-contribution areas (such as near differential pressure monitoring points) and attenuates the gradient in low-contribution areas. Feature extraction is performed in the weighted risk gradient field. First, the point with the maximum global gradient magnitude is searched, and its unit direction vector is taken as the principal gradient direction. Second, the Shannon entropy is calculated based on the concentration probability distribution, and the concentration values ​​are discretized and binned before being calculated according to the probability distribution. Finally, the non-maximum suppression algorithm is used to locate the extreme points where the causal heat weight exceeds the neighborhood mean, and its physical coordinates are recorded. Finally, the principal gradient direction vector, entropy scalar, and extreme point coordinate sequence are encapsulated into a JSON structured data packet.

[0035] The above technical solutions address the limitations of traditional gas risk analysis, which relies solely on concentration scalar thresholds to determine risk levels. This fails to differentiate between different risk patterns at the same concentration, such as uniform diffusion versus point-source injection, leading to insufficient targeting of ventilation strategies. Furthermore, ignoring physical causal relationships renders the system blind to hidden risks in areas not covered by sensors. The embodiments in this specification, through the innovative fusion of gradient fields and causal thermals, visualize the dynamics of risk evolution, enabling the control system to overcome the limitations of static thresholds. For example, even when the concentration at the working face is within limits, the weighted gradient field shows the main gradient direction pointing to abandoned roadways. Combined with the coordinates of extreme points, this pinpoints a hidden fissure gas source, allowing for early activation of sealing measures to prevent accumulation and explosion. Traditional methods, lacking gradient analysis, only respond when the concentration exceeds limits, delaying crucial intervention opportunities. By embedding causal contribution quantification, the decision-making is made physically interpretable. When the entropy value of a certain region in the weighted gradient field suddenly increases, the associated causal heat source is automatically traced. For example, if the heat of the return air roadway differential pressure monitoring point is abnormally increased, historical data is immediately linked to verify that the dust accumulation on the fan blades at that point leads to a decrease in efficiency. This accurately pinpoints the equipment maintenance target, saving a lot of fault diagnosis time compared to manual experience-based troubleshooting.

[0036] Based on the gradient feature vector of the risk field, a decision path is matched, specifically including: determining the risk entropy value index in the gradient feature vector of the risk field; when the risk entropy value index exceeds a preset dynamic threshold, activating the Monte Carlo contingency plan generation engine to output a Monte Carlo contingency plan set as the decision path; when the risk entropy value index does not exceed the preset dynamic threshold, determining the decision path as a causal decision path.

[0037] In intelligent control systems for mine gas, traditional control systems employ a single decision-making logic, which cannot adapt to dynamic changes in operating conditions. When sensor data is complete and the flow field is stable, precise control based on physical causality has significant advantages. However, in high-uncertainty scenarios such as gas release accompanied by equipment failure, the gradient direction on which the causal model depends may be distorted, and continuing to use a fixed decision path will lead to serious misjudgments.

[0038] In one embodiment of this specification, after receiving the risk field gradient feature vector in real time, the embedded risk entropy index is first parsed, which is the Shannon entropy value calculated in the previous steps. In the dynamic threshold configuration module preloaded on the edge computing node, an adaptive decision threshold is generated by comprehensively considering the real-time ventilation network complexity, sensor health status, and historical accident data; the threshold is lowered when the number of topology branches increases. When the entropy index exceeds this threshold, the Monte Carlo contingency engine is activated. When the entropy value does not exceed the threshold, a causal decision path is triggered. The risk entropy index exceeding the dynamic threshold indicates that the system has entered the critical phase transition region of multiple stable states. Activating the Monte Carlo contingency engine at this time avoids the risk of single-point decision failure through probability coverage and parallel simulation; while maintaining the causal decision path in a deterministic scenario where the entropy value is controllable leverages the high efficiency and accuracy of the physical model.

[0039] Step S103: When the decision path is a causal decision path, determine the fan / damper adjustment target point, simulate the fan / damper adjustment target point, determine the dynamic safety constraint boundary, solve the preset optimization objective function under the dynamic safety constraint boundary, generate the fan speed adjustment command and damper opening control command, and realize the control of the fan / damper.

[0040] Traditional manual ventilation control relies on experience to select control targets, easily overlooking the coupling relationship between the direction of gas diffusion dynamics and the ventilation network topology. For example, when the angle between the gradient direction and the main airflow is small, incorrectly closing branch dampers can block airflow and lead to local accumulation. Furthermore, it lacks the ability to locate hidden risk sources; outflow points from fissures far from the main airflow in the roadway are often overlooked. In conventional manual control, the singular control target cannot cope with the multi-center diffusion of gas clouds. The embodiments in this specification, through quantitative analysis of the angle between the main gradient direction and the main airflow, distinguish between source containment and path blocking control modes from a fluid dynamics perspective. The auxiliary target addition mechanism based on causal extreme points constructs a three-dimensional prevention and control network with primary and secondary components working in tandem. Through dynamic target location, it solves the problem of controlling the main risks and preventing secondary disasters in complex ventilation networks.

[0041] Determining the fan / damper adjustment target points specifically includes: obtaining the principal gradient direction vector and causal extreme point coordinates in the gradient feature vector of the risk field to calculate the angle between the principal 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, marking the upstream three-stage fan node as a source containment type principal adjustment target point; when the angle is greater than or equal to the preset angle threshold, marking the branch damper node in the gradient normal plane as a path blocking type principal adjustment target point; based on the gradient magnitude sorting result of the causal extreme point coordinates, adding auxiliary adjustment target points at extreme points that are more than a preset distance away from the principal adjustment target point.

[0042] In one embodiment of this specification, the principal gradient direction vector (a three-dimensional unit vector) and the set of coordinates of causal extrema points (a three-dimensional position sequence) are extracted from the risk field gradient feature vector. The current principal airflow direction vector is obtained by accessing a real-time updated ventilation digital twin via a mine-use 5G private network; this vector can be generated by fusing CFD simulation with wind speed sensors. The spatial angle between the two vectors is calculated, the cosine value is obtained using the dot product formula, and then the angle value is obtained by inverse solving.

[0043] When the angle is less than a preset threshold, it is determined that the gas is spreading along the main air duct. It should be noted that this preset threshold can be set empirically, for example, to 30°. When the gas spreads along the main air duct, its source is traced back to the third-level node upstream of the gradient. Specifically, starting from the risk source, three hops are traced backward along the ventilation network to locate the high-power fan at the main air duct entrance as the source containment target, i.e., the source containment type main control target. In the ventilation topology map, fans three hops away from the risk source are identified: the first hop is the fan directly connected to the roadway where the risk source is located; the second hop is the fan one level above the main air duct; and the third hop is the fan at the main air duct entrance (controlling the global airflow). When the angle is greater than or equal to the threshold, it is determined that the gas is spreading perpendicular to the main airflow. Then, the nearest branch damper node is searched in the gradient direction normal plane as the path blocking target, i.e., the path blocking type main control target is obtained. Simultaneously, the extreme point coordinate set is sorted by gradient magnitude, and the extreme points with the highest magnitudes are selected. If the distance between the extreme point and the main target point exceeds the equivalent length of the roadway (e.g., 50 meters), an auxiliary target point is added to the nearest fan / damper device at that point. The final output is a structured instruction set containing the main / auxiliary target point type and device ID to determine at least one fan / damper adjustment target point.

[0044] Two gas diffusion modes are distinguished by the gradient-airflow angle. When gas diffuses along the main airway, the third-level fan in the main airway is locked, controlling more than 60% of the total airflow, and diluting the gas through global air pressure adjustment. When gas diffuses perpendicular to the main airflow, the valve-type air doors on the diffusion path are precisely closed, which can block the backflow of gas in the goaf and avoid the surge in energy consumption caused by traditional full pressurization. The auxiliary target point addition mechanism based on causal extreme points effectively solves the problem of multi-source risk prevention and control. Control devices are deployed at the two extreme points with the largest gradient modulus, so that the dual-source gas cloud is suppressed in the early stage of diffusion, which effectively improves the wind control efficiency compared with the single target point strategy. The characteristics of the ventilation network topology are acquired in real time, and the target point positioning rules are automatically updated when the roadway is modified or the fan fails, which gets rid of the rigidity of the traditional preset target point scheme and provides an all-weather safety barrier for smart mines.

[0045] The dynamic safety constraint boundary is determined by simulating the fan / damper adjustment target point. Specifically, this includes: constructing a lightweight ventilation network propagation model, which is a graph network model. This model includes fan nodes representing pressure source variables, damper nodes representing resistance variables, and roadway edges, where each roadway edge is a resistance function containing the rate of change of cross-section. Based on the fan / damper adjustment target point, an initial target point adjustment action command is determined and input into the lightweight ventilation network propagation model. A graph traversal algorithm is used to calculate the distribution data of pressure difference changes across all network nodes, and the standard deviation of the overall network pressure difference fluctuation is determined. This standard deviation of the overall network pressure difference fluctuation is used as an offset to correct the basic air pressure safety threshold in the pre-acquired mine safety regulations database, thus determining the dynamic safety constraint boundary.

[0046] Traditional ventilation control methods use static safety thresholds, which cannot adapt to the dynamic underground environment. When a ventilation door is closed to dilute local gas, it may cause a sudden drop in pressure differential in the distant roadway, leading to airflow reversal and causing gas backflow in the goaf, forming a new explosion source. The embodiments in this specification use graph network topology to simulate the propagation effect of target point actions in real time and predict secondary risks. By using the standard deviation of pressure differential fluctuation as the basic threshold for dynamic offset correction, the elastic space of the safety boundary of the same ventilation control action under different roadway deformation states is quantified. This ensures the global feasibility of control commands from a physical perspective and provides action boundaries that comply with safety regulations and adapt to real-time operating conditions for subsequent optimization.

[0047] In one embodiment of this specification, after determining the target point for fan / damper adjustment, baseline action parameters are loaded through a pre-set equipment control specification database. For fan-type targets, the minimum effective adjustment step size under rated operating conditions is indexed according to the equipment model, such as the minimum speed adjustment of a certain type of mine counter-rotating ventilator being a fixed percentage of the rated value. The adjustment direction (pressurization or depressurization) is determined by combining the principal direction sign (positive / negative) in the risk field gradient feature vector. For damper-type targets, the minimum opening change in a single step is extracted according to the mechanical design specifications of the pneumatic actuator, such as the minimum effective rotation angle of a certain type of louvered damper. At the same time, the direction of opening increase or decrease is determined based on the topological relationship between the gradient direction and the damper position. Subsequently, an initial draft of action instructions is generated. Each target point corresponds to a structured instruction item, which includes the equipment's unique identification code, action type (speed adjustment or opening control), reference action quantity, and direction indicator. For example, the instruction for a certain third-stage fan node is "Equipment ID: FAN-201, Action type: speed adjustment, Reference quantity: + minimum adjustment step size", while the instruction for a certain branch damper is "Equipment ID: VALVE-35, Action type: opening control, Reference quantity: - minimum effective rotation angle".

[0048] A lightweight ventilation network model is constructed within the graph computing engine of the edge computing nodes. Based on the downhole mapping database, fans are abstracted as pressure source nodes, with attributes including rated air pressure-flow rate curves. Damperes are abstracted as variable resistance nodes, with attributes including an opening-resistance coefficient mapping table. Roadways are modeled as edges with attributes, and the resistance coefficient function includes the cross-sectional change rate from real-time LiDAR reconstruction. Upon receiving the initial target point adjustment action command, it is parsed into equipment action parameters, such as the fan speed adjustment percentage and the target damper opening, and converted into node attribute update values ​​for injection into the model. An improved Dijkstra graph traversal algorithm is then used to calculate pressure propagation: starting from the action target point, the pressure difference change is iteratively solved along the roadway edges, dynamically correcting the cross-sectional change factor in the resistance function, such as triggering local resistance doubling when cross-sectional contraction exceeds the limit. After traversal, the pressure difference change matrix of all network nodes is output, and its standard deviation is calculated. The standard deviation is used to characterize the intensity of system fluctuations.

[0049] The process of using the standard deviation of the network-wide differential pressure fluctuation as an offset to correct the basic air pressure safety threshold in the pre-acquired mine safety regulations database, and determining the dynamic safety constraint boundary, is as follows:

[0050] First, real-time updated roadway topology complexity parameters are loaded to determine the topology sensitivity coefficient. These parameters include the number of roadway branches and the distribution of local resistance abrupt change points. Specifically, a roadway branch node mapping table is established, and the node connection relationships of the ventilation network diagram are automatically extracted in the digital twin platform. The number of directly connected roadway branches is counted. For example, if a main transport roadway connects three coal mining faces and two return air roadways, the number of branches is five. High wind resistance abrupt change nodes 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 roadway is calculated. If continuous cross-sectional contraction exceeds the safety threshold, such as the reduction ratio of the cross-sectional area of ​​a roadway section exceeding 15% of the design value, it is marked as a structural abrupt change node in the support area with excessive cross-sectional contraction. When the abnormal fluctuation of pressure difference at a node lasts for more than a preset period, such as the standard deviation of pressure difference fluctuation continuously exceeding twice the normal value, it is marked as a fluid anomaly node in the turbulent area detected by wind speed sensors.

[0051] The initial coefficients are calculated based on topological features. The total number of branches is divided by the baseline number of branches to obtain the initial scaling factor. The baseline number of branches is the average of the historical safe operation samples of the mining area. High-resistance node weights are then added, assigning a fixed incremental value (e.g., 0.1) to each high-resistance node. This value is based on historical accident data. For example, a typical ventilation network has twelve main roadway branches, three of which are located in high-resistance areas of geological fracture zones. The high-resistance node identification results include one structural abrupt change node in the fracture zone cross-sectional contraction area identified by LiDAR, and two fluid anomaly nodes in the pressure difference abnormal fluctuation area identified by sensors. The increment is calculated as: number of high-resistance nodes = 3 nodes × single-point increment 0.1, resulting in a total increment of 0.3. First, the branch scaling factor (current 12 / baseline 10 = 1.2) is calculated, then the high-resistance weight increment (3 × 0.1 = 0.3) is added, finally generating a topological sensitivity coefficient of 1.5. When the number of branches increases to fifteen due to the excavation of a new working face in the mine, and two new high-wind-resistance fault zones are added, the topology sensitivity coefficient is automatically updated to (15 / 10) + (5 × 0.1) = 2.0. At this time, if the standard deviation of the pressure difference is a specific value, the generated dynamic offset automatically expands, which significantly tightens the safety constraint boundary compared to stable working conditions, successfully intercepting the air door closing command in the fractured zone area and preventing gas accumulation induced by airflow stagnation. This design, which quantifies abstract topological features into engineering parameters, enables the system to accurately perceive changes in network structure risks.

[0052] Next, basic air pressure thresholds are loaded from the mine safety regulations database, including the minimum air pressure threshold P_min to prevent airflow stagnation and the maximum air pressure threshold P_max to prevent the sealing wall from rupturing. When the standard deviation of regional pressure difference fluctuations reaches a specific value set based on experience, this standard deviation is used as a dynamic offset to generate a safety envelope with synchronized upper and lower bounds. Specifically, the standard deviation is multiplied by a topology sensitivity coefficient to generate a dynamic offset Δ. The specific correction logic is as follows: when the standard deviation indicates severe system fluctuations, i.e., when it reaches the specific value set based on experience, Δ is added to the basic lower limit threshold P_min to increase safety redundancy, while Δ is subtracted from the basic upper limit threshold P_max to tighten constraints; conversely, when the system is stable, the boundaries are appropriately relaxed. For example, during a tunnel breakthrough operation in the Huainan mining area, simulations showed a surge in the standard deviation of the pressure difference at the moment of breakthrough. The system automatically calculated the positive offset Δ and generated a dynamic constraint boundary [P_min+Δ, P_max-Δ]. This tightened safety envelope successfully intercepted the original air door closing command, preventing an airflow reversal accident caused by the breakthrough shock wave. In contrast, the static threshold scheme, failing to anticipate this risk, led to local gas accumulation. This dynamic correction mechanism inherits the bottom-line requirements of safety regulations while giving the system the flexible protection capability to adapt to real-time operating conditions. The final output dynamic safety constraint boundary is linked with the optimization objective function to form a physical barrier for command generation. For example, during tunnel breakthrough operations, the basic threshold is a fixed value, but simulations show a surge in the standard deviation of the pressure difference fluctuation at the moment of breakthrough. At this time, the system automatically detects the high standard deviation and immediately generates a stricter safety boundary to prevent airflow adjustment commands that may cause airflow turbulence, thus solving the safety hazard of using a fixed threshold in traditional methods when operating conditions change abruptly.

[0053] In one embodiment of this specification, a preset optimization objective function is solved under dynamic safety constraints to generate fan speed adjustment commands and damper opening control commands, thereby achieving fan / damper control. It should be noted that the optimization objective function here is as follows: ,in, Let m represent the equipment adjustment amount at the i-th target point (fan / damper), such as the fan speed / damper opening adjustment amount, where m indicates that there are m adjustment target points. β 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. The target gradient is the ideal distribution after dilution.

[0054] After determining the objective function (minimizing the weighted sum of the total number of equipment actions and the deviation of the risk gradient) and the dynamic safety constraint boundary (the pressure envelope output by the ventilation network propagation model), a hierarchical solution strategy is designed for different types of equipment. For discrete variables such as damper opening, the branch and bound method is used to traverse the feasible solution space, combined with a tabu search strategy to avoid local optima; for continuous variables such as fan speed, iterative solution is based on sequential quadratic programming, and the compliance with dynamic constraints is checked after each iteration. After optimization, structured results are output, yielding the globally optimal solution that satisfies all constraints (aggressive solution), the feasible solution with the smallest action amplitude (basic solution), and the historical optimal solution during the convergence process (backup solution). Subsequently, industrial instruction conversion is performed. The fan speed solution value is converted into a Modbus-TCP protocol holding register write value through linear mapping, such as converting the speed percentage into a hexadecimal function code 06 message; the damper opening angle solution value is encapsulated into a process data object position instruction word according to the object dictionary address of the CANopen protocol. The finally generated instruction packet is appended with a timestamp and CRC check code and sent to the PLC control cabinet via the industrial ring network.

[0055] Traditional mine ventilation systems employ manually set fixed parameters or PID control. Without airflow network constraints applied to the PID controller, continuously increasing the fan speed leads to gas backflow in the negative pressure zone of the roadway. The embodiments in this specification optimize the causal alignment term (risk gradient deviation) in the objective function, forcing control commands to conform to fluid diffusion laws. A tiered protection system is formed through three types of output contingency plans: aggressive, basic, and backup. When the main plan is blocked (e.g., a fan fails), the backup plan switches within milliseconds, effectively improving fault response speed tenfold compared to traditional single-plan designs. Furthermore, the native protocol encapsulation of Modbus and CANopen avoids the secondary configuration required by traditional methods, reducing the error rate of command conversion.

[0056] When the decision path is a Monte Carlo contingency plan set, the method further includes: generating an initial set of Monte Carlo contingency plans based on the gradient feature vector of the risk field; performing ventilation effect simulation and confidence screening on the initial set of Monte Carlo contingency plans through a preset evaluation model; inputting high-confidence contingency plans with confidence scores greater than a preset confidence threshold into a preset optimization objective function to solve for the optimal control command, so as to parse and execute the optimal control command. The initial set of Monte Carlo contingency plans was simulated for ventilation effects and confidence level screening was performed using a pre-defined evaluation model. Specifically, this included: determining a pre-defined lightweight CFD proxy model, where the inputs to the lightweight CFD proxy model were the fan speed and damper opening parameters for each plan in the initial set of Monte Carlo contingency plans, and the outputs were the overall network gas concentration distribution and node pressure difference changes; using this lightweight CFD proxy model, ventilation network fluid dynamics simulation was performed for each plan to determine the predicted data for the overall network gas concentration distribution and node pressure difference changes for each plan; based on the predicted data for the overall network gas concentration distribution and node pressure difference changes for each plan, the confidence level of each plan was determined, and confidence level screening was performed based on this confidence level and a pre-set confidence threshold.

[0057] In high-risk scenarios of sudden gas leaks in mines, when risk entropy indicators show that the system is in a state of high uncertainty, such as when sensor failure occurs concurrently with a sudden gas release, the physical models upon which conventional causal decision-making relies may completely fail. For example, if the gradient direction drifts randomly due to data distortion, blindly adjusting the target point will lead to catastrophic misjudgments.

[0058] In one embodiment of this specification, when the decision path is a Monte Carlo contingency plan set, an initial contingency plan set is generated based on the risk field gradient feature vector. Cases with high physical environment similarity (such as tunnel topology matching degree, gradient magnitude range) are retrieved from the historical operating condition database. This is then expanded into a contingency plan pool containing multiple combinations of fan / door operation (such as the combination of fan A speed increase and door B closure, global fan speed decrease combination, etc.) through constrained random sampling. Specifically, the principal gradient direction vector, risk entropy value, and causal extreme point coordinates in 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. Historical cases matching the tunnel topology (such as the number of branches, distribution of high wind resistance nodes) with the current gradient features are matched using a spatial similarity algorithm, and a sample set with physical environment similarity exceeding a preset threshold is selected. For example, when the principal gradient direction points to an abandoned tunnel and there are three high wind resistance nodes, hundreds of valid disposal records from similar scenarios in the past six months are automatically retrieved. An initial contingency plan pool is generated based on the selected samples, and constrained random sampling is performed on the fan / door operation parameters in the samples. Centered on historical equipment actions, Gaussian distribution sampling is performed within the equipment's safe operating boundaries, such as the fan speed adjustment range ± percentage of rated value and the safe range of damper opening angle. Simultaneously, directional constraints are injected into the current gradient vector, such as limiting the sampling weight of speed reduction when the gradient points towards the inlet airway. For the causal extreme point coordinate set, additional equipment disturbance contingency plans are added. Available fans / dampers are searched within the radius of the extreme points to generate supplementary operation combinations for local pressurization or isolation. The final output is an initial contingency plan set that integrates historical experience and real-time constraints.

[0059] Subsequently, a pre-deployed lightweight CFD proxy model is invoked. This model consists of a graph neural network trained on full-scale CFD simulation data. The inputs are equipment parameters from the initial contingency plan set, including fan speed percentage and damper opening angle. The outputs are predicted data on the overall network gas concentration distribution and node differential pressure changes. Based on the simulation results of each contingency plan, the confidence level is calculated, and the gas dilution efficiency (target area concentration decrease rate) and differential pressure stability (node ​​fluctuation amplitude) are comprehensively evaluated. Specifically, the gas concentration decrease rate in key target areas is first extracted, and the ratio of this rate to the maximum theoretical dilution rate is normalized to obtain the gas dilution efficiency index. Simultaneously, node differential pressure fluctuation data is analyzed, and the peak-to-average ratio of the overall network differential pressure change and the percentage of time exceeding the maximum allowable fluctuation amplitude above the safety threshold are calculated to generate a differential pressure stability coefficient. The calculation formula is as follows:

[0060] ;

[0061] in, α The peak-to-average ratio (PAPR) weight can be set to 0.3. β The weight for the excess percentage can be set to 0.5. This represents the maximum absolute value of the nodal pressure difference change during the simulation period. This represents the arithmetic mean of the pressure difference changes across all network nodes. This indicates the cumulative duration during which the differential pressure at the nodes exceeds the allowable fluctuation range set in the mine safety regulations. This represents the total simulation time, and γ is the high drag attenuation factor, which is applied when high drag nodes exist. λ is the attenuation coefficient, which can be set to 0.2. To determine the number of high-resistance areas obtained through topology analysis, γ=1 when no high-resistance nodes exist. The gas dilution efficiency index and pressure differential stability coefficient are weighted and summed to generate a scalar confidence value. This confidence value is then compared to a pre-set confidence threshold, which can be set based on the complexity of the roadway topology. Only high-confidence plans with confidence values ​​greater than the pre-set threshold are retained. After selecting the set of high-confidence plans, an optimization objective function is input for solving to obtain the optimal plan. The final plan is then converted into the corresponding optimal control command. The optimization objective function and the conversion of the optimal plan into the optimal control command have been explained in the preliminary steps of the causal decision path and will not be repeated here.

[0062] By using the above technical solutions, a set of contingency plans covering multiple probabilities is generated, and physical feasibility is verified based on a lightweight CFD proxy model. Parallel simulation is used to screen out options that violate fluid dynamics laws or trigger secondary risks, ensuring that the final executed instructions have both mathematical optimality and physical compliance.

[0063] The technical solutions described in this specification demonstrate that traditional single-point sensors can only sense local instantaneous concentration values. However, the embodiments in this specification, through the spatiotemporal alignment mechanism of multimodal data cubes, construct a complete information field of gas diffusion at the fluid dynamics level. The roadway curvature features extracted from LiDAR point clouds are coupled with the turbulence priors of the CFD airflow grid, enabling the spatiotemporal Transformer model to capture the causal chain of pressure difference fluctuations, wind speed attenuation, and concentration accumulation. This transforms the early warning mechanism from passive concentration exceedance detection to active diffusion dynamics deduction. Traditional manual ventilation adjustment requires early warning and decision-making processes. The multi-level delay chain of execution is addressed by the causal decision-making path in the embodiments of this specification, which achieves integrated decision-making and execution at the edge computing end through real-time matching of gradient fields and network topology. Based on the optimal control target selection according to network transmission efficiency, and combined with protocol-level direct drive, the response time of the entire chain of perception, decision-making, and execution is effectively compressed. Conventional static safety thresholds cannot adapt to dynamic risks such as roadway deformation. The ventilation network propagation model in the embodiments of this specification constructs a virtual test field for ventilation adjustment commands, quantifying structural risks into elastic protection parameters, and realizing the transformation from empirical safety margin to computational safety boundary. For cases with high confidence in the gradient direction, causal paths are enabled to achieve the physically optimal solution; conversely, swarm intelligence search using the Monte Carlo scheme is adopted, directly realizing intelligent switching of decision-making modes, effectively improving the control accuracy and effect of mine gas, and meeting the rigid safety requirements of intelligent mines.

[0064] This specification also provides an embodiment of 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 executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described method.

[0065] This specification also provides a non-volatile computer storage medium storing computer-executable instructions configured to execute the above-described method.

[0066] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0067] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for controlling mine gas based on multimodal data, characterized in that, The method includes: Acquire multi-source real-time sensor data, roadway point cloud data, and CFD airflow simulation grid data in the target mine roadway, and perform spatiotemporal alignment to determine the real-time roadway data cube. Using a preset multimodal depth causal model and the real-time roadway data cube, determine the gas concentration field distribution and causal contribution heat data within a preset future time window. 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. The decision path includes the Monte Carlo contingency plan set and the causal decision path. When the decision path is a causal decision path, the fan / damper adjustment target point is determined, the fan / damper adjustment target point is simulated, the dynamic safety constraint boundary is determined, and the preset optimization objective function is solved under the dynamic safety constraint boundary to generate the fan speed adjustment command and the damper opening control command to realize the control of the fan / damper. Based on the gas concentration field distribution and the causal contribution heat data, a risk field gradient feature vector is generated, specifically including: Calculate the spatial partial derivative of the gas concentration field distribution along the three-dimensional coordinate direction of the roadway to generate an initial gradient vector field; The causal contribution heat data is used as a weight matrix and Hadamard product is performed with the initial gradient vector field to generate a weighted risk gradient field. The principal 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. Based on the risk field gradient feature vector, a decision path is matched, specifically including: Determine the risk entropy value index in the gradient feature vector of the risk field; When the risk entropy value exceeds the preset dynamic threshold, the Monte Carlo contingency plan generation engine is activated to output a set of Monte Carlo contingency plans as the decision path. When the risk entropy value does not exceed the preset dynamic threshold, the decision path is determined to be a causal decision path; Determine the target point for fan / damper adjustment, specifically including: Obtain the principal gradient direction vector and causal extreme point coordinates from the risk field gradient feature vector to calculate the angle between the principal gradient direction vector and the main airflow direction in the pre-acquired real-time ventilation network topology; When the included angle is less than a preset angle threshold, the upstream three-stage wind turbine node of the gradient is marked as the source-control type main regulation target point; When the included angle is greater than or equal to a preset angle threshold, the branch damper node in the gradient method plane is marked as a path blocking type main adjustment target point; Based on the gradient magnitude sorting result of the causal extreme point coordinates, auxiliary adjustment target points are added at extreme points that are more than a preset distance away from the main adjustment target point. Simulations were performed on the fan / damper adjustment target point to determine the dynamic safety constraint boundary, specifically including: A lightweight ventilation network propagation model is constructed, wherein the lightweight ventilation network propagation model is a graph network model, and the graph network model includes fan nodes for representing pressure source variables, damper nodes for representing resistance variables, and roadway edges, wherein the roadway edges are resistance functions containing the rate of change of cross section; Based on the fan / damper adjustment target point, an initial target point adjustment action command is determined, and the initial target point adjustment action command is input into the lightweight ventilation network propagation model. A graph traversal algorithm is used to calculate the distribution data of pressure difference change of all network nodes and determine the standard deviation of pressure difference fluctuation of the entire network. Using the standard deviation of the network pressure difference fluctuation as an offset, the basic wind pressure safety threshold in the pre-acquired mine safety regulations database is corrected to determine the dynamic safety constraint boundary.

2. The mine gas control method based on multimodal data according to claim 1, characterized in that, Determine the real-time roadway data cube, specifically including: Real-time acquisition of multi-source real-time sensor data within the target mine roadway, wherein the multi-source real-time sensor data includes time-series data of gas concentration, time-series data of temperature and humidity, and time-series data of pressure difference; The multi-source real-time sensor data is time-stamped and synchronized using a high-precision clock service for time alignment. A preset algorithm is used 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. The spatiotemporally aligned data is encapsulated into a real-time tunnel data cube with multiple preset dimensions. The preset dimensions include a time dimension, a spatial dimension, and a feature dimension. The spatial dimension is mapped to the tunnel grid coordinates, and the feature dimension includes sensor measurement indicators and tunnel point cloud geometric features.

3. The mine gas control method based on multimodal data according to claim 1, characterized in that, When the decision path is a Monte Carlo contingency plan set, the method further includes: An initial set of Monte Carlo contingency plans is generated based on the gradient feature vector of the risk field. The ventilation effect simulation and confidence screening of the initial set of Monte Carlo contingency plans are performed by a preset evaluation model. A high-confidence plan with a confidence level greater than a preset confidence threshold is input into a preset optimization objective function to solve for the optimal control command, so as to parse and execute the optimal control command.

4. The mine gas control method based on multimodal data according to claim 3, characterized in that, The ventilation effect simulation and confidence level screening of the initial set of Monte Carlo contingency plans were conducted using a pre-set evaluation model, specifically including: A preset lightweight CFD proxy model is determined, 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 of node pressure difference; Using the lightweight CFD proxy model, ventilation network fluid dynamics simulation is performed for each contingency plan to determine the predicted data of the whole network gas concentration distribution and the predicted data of node pressure difference change for each contingency plan. Based on the predicted data of the gas concentration distribution across the entire network and the predicted data of the pressure difference change at each node corresponding to each contingency plan, the confidence level of each contingency plan is determined, and confidence screening is performed based on the contingency plan confidence level and the preset confidence threshold.

5. A mine gas control device based on multimodal data, characterized in that, 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 to enable the at least one processor to perform the method as described in any one of claims 1-4.

6. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to execute the method as described in any one of claims 1-4.

Images