Coal mine disaster early warning method based on multi-source data fusion and time-space diagram network

Abstract

The invention discloses a coal mine disaster early warning method based on multi-source data fusion and a time-space diagram network. The method comprises the following steps: (a) collecting multi-source heterogeneous monitoring data; (b) constructing a geologic-physical dual constraint anisotropic space-time diagram; (c) training the graph by using a geologic-physical anisotropic space-time graph neural network, wherein a graph convolutional layer of the network is designed into a propagation mode which is learned and defined at the same time; and (d) predicting a disaster risk by using the trained model and generating an early warning. According to the method, geological and physical priori knowledge is taken as constraints to be incorporated into the topological structure of the GNN model, so that the model can follow physical laws, disaster propagation under anisotropic geological conditions is predicted, and the accuracy and timeliness of early warning can be improved.

Description

Technical Field

[0001] This invention relates to the field of disaster prediction technology, and more specifically, to a coal mine disaster early warning method based on multi-source data fusion and spatiotemporal graph networks. Background Technology

[0002] Coal mining is a high-risk industry, especially in mines with complex geological conditions, where safe production faces serious threats from multiple disasters such as gas, water, fire, roof collapse, and dust. Taking the Huoshaopu Mine of Guizhou Panjiang Refined Coal Co., Ltd. as an example, this mine is a typical complex geological coal mine. On the one hand, it has been identified as a "coal and gas outburst mine," with an absolute gas emission rate as high as 140.75 kJ / m³. The maximum gas emission at the mining face reached 25.1. On the other hand, the mine's geological structure is extremely complex, characterized by "developed fracture structures with steep dip angles." The roof and floor of the mineable coal seams are mostly composed of silty mudstone and silty mudstone, with well-developed bedding and joints, making "the roof of the coal seams extremely difficult to control" and requiring frequent roadway maintenance. Furthermore, the mine has moderate hydrogeological conditions, classifying it as a "fissure-filled water deposit," with each mining area facing varying degrees of water inrush threat, and each mineable coal seam posing a "coal dust explosion hazard." To address these hazards, modern coal mines have widely deployed various online monitoring systems. For example, Huoshaopu Mine has installed the KJ90X safety monitoring system (for monitoring gas, wind speed, CO, etc.), the KJ1580J coal mine underground personnel precision positioning system, the KJ428 mine distributed system laser fire monitoring system, the KJ1193 coal mine hydrological monitoring system, and a mine roof online monitoring system. These systems generate massive amounts of multi-dimensional data. However, existing technologies suffer from data silos and limitations in model application. First, the current situation at Huoshaopu Mine is that "big data integration analysis and dissemination have not been carried out." Each monitoring system (such as KJ90X, KJ1193, and KJ428) operates independently and issues alarms independently, failing to reveal the inherent mechanism of disasters (especially gas outbursts and water hazards) as the result of the coupled effects of multiple factors (such as ground stress, gas pressure, water pressure, and mining activities). Second, crucial geological data relies on manual processing. Although the geological support system is equipped with geophysical and drilling equipment, it "cannot achieve automatic data uploading," and currently still uses "local storage and copying of data for manual entry and analysis." This "manual" and "non-integrated" status quo leads to the lag, passivity, and one-sidedness of disaster early warning. For example, the safety monitoring system can only issue alarms when and where gas concentrations have already exceeded standards, but cannot predict where gas will accumulate. The hydrological monitoring system can only reflect the current water level, but cannot combine geological faults and mining stress to predict the risk of impending water inrush. In recent years, Spatiotemporal Graph Neural Networks (ST-GNNs) have demonstrated spatiotemporal data processing capabilities in areas such as traffic flow prediction and earthquake early warning. However, directly applying such standard models to complex geological coal mines presents technical challenges: The isotropic assumption is inapplicable: Standard GNN models are typically "isotropic," assuming relationships between nodes (sensors) are based on distance or data correlation. However, in coal mines, geology is "anisotropic." Two sensors only 10 meters apart might be completely disconnected from each other due to a closed fault. Standard GNNs struggle to handle this anisotropy determined by geological structures. The lack of physical field constraints: The spread of gas and fire (CO) is strictly governed by the "directivity" of the mine ventilation network. Airflow only flows from upstream (intake airway) to downstream (return airway). Standard GNNs typically use undirected graphs, which contradicts fundamental laws of fluid dynamics.Therefore, there is a need in this field for a new technical solution that can break through data silos and effectively integrate multi-source heterogeneous data from safety monitoring, hydrology, fire, geology, roof, etc. Furthermore, the solution must be able to incorporate complex geological conditions (such as faults and lithology) and physical laws (such as ventilation and airflow) as prior knowledge and hard constraints into the model in order to achieve effective prediction and early warning of complex geological coal mine disasters. Summary of the Invention

[0003] The main objective of this invention is to overcome the shortcomings of the prior art and provide a method for coordinated roof cutting and pressure relief in roadway excavation along the goaf of three soft coal seams, thereby solving the technical problems existing in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A coal mine disaster early warning method based on multi-source data fusion and spatiotemporal graph network includes the following steps:

[0006] a. Data acquisition steps: Real-time acquisition of multi-source heterogeneous monitoring data from multiple sensor nodes in the coal mine, including time-series disaster monitoring data and non-time-series geological environment data;

[0007] b. Graph construction steps: Based on the spatial locations of the multiple sensor nodes and the non-temporal geological environment data, construct a spatiotemporal map of mine disasters that characterizes the correlation between the multiple sensor nodes;

[0008] c. Model training steps: The time-series disaster monitoring data obtained in (a) is used as the node feature input. The topological structure of the mine disaster spatiotemporal map constructed in (b) is used to train the model through the spatiotemporal map neural network model to obtain the disaster early warning model.

[0009] d. Warning generation steps: Input the real-time acquired time-series disaster monitoring data into the disaster warning model obtained in (c), output the prediction of the disaster risk level within the future time window, and generate a warning signal when the risk level exceeds a preset threshold.

[0010] Furthermore, the time-series disaster monitoring data includes: gas concentration, carbon monoxide concentration, oxygen concentration, wind speed, and temperature data from the safety monitoring system; water level in the water tank, flow rate in the drainage ditch, and water level in the catchment channel from the water control system; fire monitoring data from the fire prevention system; and roof delamination and anchor stress data from the roof monitoring system.

[0011] Furthermore, in the graph construction step (b), the topological structure of the mine disaster spatiotemporal graph is composed of a geological adjacency matrix. Defined; the geological adjacency matrix The construction includes:

[0012] (b1) Obtain geological data from the non-time-series geological environment data, including mine geological structure map, fault distribution, rock strata permeability and coal seam thickness data;

[0013] (b2) For any two nodes among the plurality of sensor nodes and The nodes are characterized based on the geological data. and Edge weights of geological connectivity between ;

[0014] (b3) Based on the aforementioned edge weights Construct the geological adjacency matrix .

[0015] Furthermore, the edge weights The calculation formula is: ,

[0016] in, For nodes and Spatial distance between them; For nodes and Effective penetration rate along the path between; For the node and The influence factor of faults traversed by the path, when the path does not traverse a fault. When the path traverses a closed fault that acts as a barrier When the path traverses a conductive fault that serves as a passageway .

[0017] Furthermore, in the graph construction step (b), the topological structure of the mine disaster spatiotemporal graph is also composed of a physical adjacency matrix. Defined; the physical adjacency matrix The construction includes:

[0018] (b4) Obtain the mine physical field data from the non-time-series geological environment data, the mine physical field data including the mine ventilation network topology and ventilation dynamic parameters;

[0019] (b5) Construct the ventilation network as a directed graph. The edge represents the wind flow path;

[0020] (b6) For any two nodes among the plurality of sensor nodes and If and only if There exists a path from node To the node When calculating a directed path, the node is represented. and Edge weights of the correlation between airflow ,otherwise ;

[0021] (b7) Construct the physical adjacency matrix .

[0022] Furthermore, the spatiotemporal graph neural network model (c) is a geologically-physically constrained anisotropic spatiotemporal graph neural network (GP-ASTGNN); the graph convolutional layer of the GP-ASTGNN updates the first... Layer node features To obtain the first Layer node features :

[0023] in, It is the identity matrix; and They are respectively and The normalization degree matrix; , and For the first The trainable weight matrix of the layer; For the first The learnable adaptive adjacency matrix of the layer; This is the activation function.

[0024] Furthermore, the early warning generation step (d) is an early warning for gas outburst disasters. The output of the model (c) includes predicted values ​​of gas concentration, ground stress, and wind speed. The disaster risk level is determined based on the predicted values. The early warning signal is generated when the predicted gas concentration exceeds the gas outburst critical value.

[0025] Furthermore, the early warning generation step (d) is for mine water hazard early warning, the output of the model (c) includes the predicted values ​​of water level in the water tank and flow rate in the roadway, the disaster risk level is determined based on the predicted values, and the early warning signal is generated when the predicted rate of water level rise exceeds the critical value of water hazard.

[0026] Furthermore, the early warning generation step (d) is for mine fire early warning, and the output of the model (c) is a spatiotemporal diffusion prediction of carbon monoxide (CO) concentration; the early warning signal includes fire source location and a prediction based on the physical adjacency matrix. List of downstream nodes along the simulated CO diffusion path.

[0027] Compared with existing technologies, this invention has the following advantages: Addressing the problems of isolated monitoring data, inability to reveal disaster coupling mechanisms, and the inability of standard GNN models to handle geological anisotropy and physical field directionality in existing technologies, this invention distinguishes between the collection of "temporal disaster monitoring data" and "non-temporal geological environment data," providing separate inputs to solve the problems of data silos and model constraints. The topology of the "mine disaster spatiotemporal map" constructed from "non-temporal geological environment data" (such as fault distribution and ventilation networks) solidifies geological anisotropy and physical field directionality (such as airflow direction) as structured prior knowledge into the graph model, directly solving the limitations of standard GNNs in the background technology. "Temporal disaster monitoring data" (such as gas, stress, and water level) are used as node features for training. This design ensures that the spatiotemporal information propagation of the model (such as gas diffusion or water inrush) must follow the physical and geological laws defined in step (b) from the beginning, while also learning the complex coupling relationships between multi-source temporal features. Therefore, the "disaster risk prediction within the future time window" output by step (d) is not only based on data-driven inference, but also on spatiotemporal extrapolation based on physical and geological constraints, thus realizing the transformation from passive "alarm" to active "early warning", significantly improving the accuracy and timeliness of early warning. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the present invention. Detailed Implementation Specific Implementation Example 1:

[0031] Method of the present invention (reference) Figure 1 The required hardware and data infrastructure.

[0032] (1.1) System Hardware Architecture

[0033] The early warning method of this invention is deployed in a coal mine data center. In one embodiment, the data center may adopt a hyperconverged architecture, equipped with, for example, multiple hyperconverged node servers and 10 Gigabit switches, to meet the computational needs of massive data storage and model training. All underground monitoring systems can access the data center via a 10 Gigabit industrial Ethernet ring network.

[0034] (1.2) Multi-source heterogeneous data acquisition and preprocessing

[0035] The data acquisition step (1a) of the present invention is to fully access and integrate the data from multiple isolated monitoring systems mentioned in the background art.

[0036] (1.2.1) Time-series disaster monitoring data (node ​​dynamic characteristics) )

[0037] Temporal data is used to construct the input feature matrix of a spatiotemporal graphical neural network. ,in It is the time step. It is the number of sensor nodes. It is the feature dimension.

[0038] Safety monitoring data: Access the "safety monitoring system" described in the background technology (e.g., the KJ90X model from China Coal Technology & Engineering Group Chongqing Research Institute Co., Ltd.). Collect sensor data from all (e.g., 70) monitoring area controllers across the entire mine, with a sampling period of 30 seconds. This includes: methane (CH4) concentration, carbon monoxide (CO) concentration, oxygen (O2) concentration, wind speed, temperature, ventilation duct airflow, and equipment on / off status.

[0039] Hydrological monitoring data: Connect to the "water control system" described in the background section (e.g., the KJ1193 coal mine hydrological monitoring system of Xi'an Zhongyuan Zhicheng Information Measurement and Control Technology Co., Ltd.). Collect hydrological monitoring data from underground water tanks, pump rooms, main drainage ditches, and catchment channels at a sampling cycle of 5 minutes, including water level, water temperature, pipeline flow, and open channel flow.

[0040] Fire monitoring data: Access the "fire prevention system" described in the background section (e.g., the KJ428 mine distributed system laser fire monitoring system from Zhengzhou Guangli Technology Co., Ltd.). With a sampling period of 1 minute, fire monitoring data, including the concentrations of gases such as CO, O2, N2, and C2H2, are collected from the goaf of the coal mining face and some important rock passes.

[0041] Roof (surrounding rock) monitoring data: Connected to the "Roof (surrounding rock) monitoring system" as described in the background technology. Data such as fully mechanized mining pressure reduction, roof delamination, and anchor bolt (cable) stress are collected at a sampling cycle of 10 minutes.

[0042] Personnel positioning data: Access the "precise personnel positioning system" (e.g., KJ1580J) as described in the background technology. Obtain the real-time location coordinates of all personnel carrying cards throughout the mine with a sampling period of 5 seconds.

[0043] (1.2.2) Non-temporal geological environmental data (graph structure) (Based on the construction)

[0044] Non-temporal data is used to construct the geophysical anisotropy map (GP-AG) of this invention, i.e., defining... and .

[0045] Geological data (for) This involves connecting to the mine's "Geological Support System." It acquires and digitizes all electronic versions of basic geological data, including: topographic geological maps, contour maps of the minable coal seam floor, mine gas geological maps, mine water-bearing maps, comprehensive hydrogeological maps, borehole columnar sections, exploration line profiles, etc. This data is used to construct a geological adjacency matrix. The foundation.

[0046] Physical (ventilation) data (for) ): Connect to the mine's "ventilation subsystem". Obtain the operating parameters of all ventilation rooms in the entire mine and a complete mine ventilation network topology map, including the connection relationships, directions, and air resistance data of all roadways, air doors, air windows, and local fans.

[0047] (1.3) Data Preprocessing

[0048] Node definition: All fixed time-series sensors (such as gas probes, water level gauges, CO probes, and stress gauges) mentioned in (1.2.1) above are defined as part of the spatiotemporal diagram. nodes Let the total number of nodes be... .

[0049] Feature matrix construction: A series of time steps Each node Features (such as CH4, CO, wind speed, water level, stress, etc.) are used to construct a feature matrix. .

[0050] Data cleaning: for Perform missing value imputation (e.g., using spatiotemporal kriging interpolation or simple neighbor mean interpolation) and normalization (e.g., Z-Score normalization). ).

[0051] This invention does not employ a single adjacency matrix based on Euclidean distance, but instead constructs a dual graph structure, consisting of a geological adjacency matrix. and physical adjacency matrix Common definition.

[0052] (2.1) Geological adjacency matrix Construction

[0053] (2.1.1) The aim is to simulate the "seepage" and "diffusion" processes of disasters (especially water and gas) in underground rock strata and geological structures (such as faults and fissures). This is An undirected, weighted matrix This represents the total number of sensor nodes.

[0054] (2.1.2) Data source: Use the “geological data” (i.e., geological maps and exploration reports) in Example 1 (1.2.2).

[0055] (2.1.3) Construction steps:

[0056] First, a three-dimensional geological model of the mine is constructed by combining two-dimensional maps such as "contour map of the bottom plate of the mineable coal seam" and "topographic and geological map" with "drill columnar section map" and "exploration line profile map".

[0057] Voxelize the 3D model or create a finite element mesh. For each mesh element... Assigning geological properties, especially permeability and porosity.

[0058] Will sensor nodes Mapped to its location nearest grid cell .

[0059] For any two nodes and Calculate the geological connectivity weights between them. .

[0060] Step A: Calculate spatial distance :calculate and The Euclidean distance between them.

[0061] Step B: Calculate the effective penetration rate On the attribute grid, calculate from arrive Shortest geological path (For example, using Dijkstra's algorithm, where the path cost is...) ). It can be approximated as the harmonic average of the permeability along this path.

[0062] Step C: Calculate the fault influence factor :

[0063] The background technology mentions "developed fault structures." Geological studies indicate that faults can act as both barriers and pathways. Standard GNNs struggle to distinguish between these two. From the "Mine Gas Geological Map" and "Comprehensive Hydrogeological Map," each known fault in the mine is identified. Assign a "conduction multiplier" For example: if the geological report indicates a fault... A closed fault (barrier) filled with mudstone, then (That is, permeability is reduced by 99%). If the geological report indicates a fault. If it is a water-conducting / gas-conducting fault (channel) with fracture development, then (That is, permeability increases 10 times). If the path If it does not cross the fault, then . Defined as a path All intersecting faults of The product of:

[0064] Step D: Calculate the final weights:

[0065] Step E: Construct the matrix :

[0066] in It is an adjustable scale hyperparameter. This formula ensures... It is symmetric (undirected graph), and the weights are affected by geological anisotropy ( and The strong modulation of ).

[0067] (2.2) Physical Adjacency Matrix Construction

[0068] (2.2.1) Purpose: The aim is to simulate the "advection" and "convective" processes of hazards (especially gas and CO from fire) in a mine ventilation network. This is... Directed, weighted matrices (i.e.) ).

[0069] (2.2.2) Data source: Use the "physical (ventilation) data" (i.e., ventilation system diagram) in Example 1 (1.2.2).

[0070] (2.2.3) Construction steps (based on graph theory):

[0071] Ventilation network graph theory modeling: Representing the ventilation system diagram as a directed graph. . These are airflow nodes (intersections of alleyways, air doors, and fans). It is the airflow path (tunnel), which has direction (airflow direction) and weight (such as tunnel length or air resistance).

[0072] Node mapping: sensor nodes Mapped to The alleyway where it is located (side) ) or node ( ).

[0073] Directed path weight calculation: for any two nodes and :

[0074] Step A: In In this context, Dijkstra's algorithm or BFS is used to search for... arrive Directed path .

[0075] Step B:

[0076] if lie in Downstream of the wind flow (i.e., where there is) If the path length is not specified, then calculate the path length. .if lie in The upstream of the wind flow (i.e., there is no wind upstream) ),but .

[0077] Step C: Calculate the weights . ,

[0078] in It is a maximum influence distance (e.g., 500 meters), beyond which the advection effect is considered to be 0.

[0079] Step D: Constructing the matrix :

[0080] The matrix It is asymmetric because it strictly follows the directional physical constraints of airflow.

[0081] (c) Model training steps, which solves the problem in the background technology that the GNN model cannot simultaneously handle geological anisotropy and physical directionality.

[0082] (3.1) Model Architecture

[0083] This invention employs a multi-layer spatiotemporal graph neural network (GP-ASTGNN), the core of which is to stack multiple "geophysical graph convolution (GP-Conv)" layers and combine them with a temporal convolution module (such as gated TCN or GRU).

[0084] (3.2) Geological-physical map convolution (GP-Conv) layer

[0085] (3.2.1) Technical Principle: Convolution operation of traditional GCN It is to combine the features of neighboring nodes Perform a weighted average (by (Decision) and passed The transformation is performed. The technical problem this invention aims to solve is that the propagation mechanism of coal mine disasters is not singular (geological seepage + physical advection), thus requiring multiple adjacency matrices. The GP-Conv layer of this invention (the...) (Layer) addresses this issue by decoupling the propagation process into three parallel components: the geological diffusion term Physical advection term and adaptive learning items .

[0086] (3.2.2) Mathematical Derivation: Given the first... Layer node feature input ( 1 node Dimensional features). Layer node features Calculated using the following propagation rules: ,

[0087] in It is an activation function (e.g., ReLU).

[0088] Geological diffusion items : ,

[0089] This utilizes the technology constructed in Example 2.1 Perform graph convolution on the (geological adjacency matrix). It is the identity matrix, representing the self-join in GCN. yes The normalization degree matrix, . It is a trainable weight matrix along the geological propagation path. Function: This enables the model to learn how hazards (such as water and gas) diffuse and seep through anisotropic geological media (faults, fractures).

[0090] Physical advection term : ,

[0091] This utilizes the technology constructed in Example 2.2. Perform graph convolution on the (physical adjacency matrix). Note: Do not include the identity matrix here. ,because It is a directed graph, where the next state of a node should not (or not always) depend on itself, but rather on its upstream. yes The (out-degree or in-degree) normalized degree matrix. It is a trainable weight matrix along the physical propagation path. This enables the model to learn how hazards (such as gas and CO) strictly follow the directionality of the ventilation network for advection and convection.

[0092] Adaptive learning items : ,

[0093] This work draws inspiration from related spatiotemporal graph networks (such as Graph WaveNet). It is a learnable adaptive adjacency matrix. This matrix is ​​randomly initialized at the start of training and automatically learns potential, unlearned adjacency relationships between nodes through backpropagation. and Explicitly defined associations. This is a trainable weight matrix on the adaptive path. This term is used to capture hidden associations and data correlations that are not described by prior geological and physical knowledge.

[0094] (3.2.3) Working principle: This GP-Conv structure uses three parallel graph convolution kernels , This decouples the complex process of disaster propagation into geological seepage, physical advection, and unknown correlations. During training, the model automatically learns the relative importance of these three pathways (i.e., the three...). The matrix weights), thus targeting different disasters (such as floods, which mainly depend on...). Fire mainly depends Adaptively adjust the early warning logic.

[0095] (3.3) Model Training

[0096] Architecture: Stack the GP-Conv layer with a Temporal Convolutional Network (TCN) or a Gated Recurrent Unit (GRU) to form a complete spatiotemporal prediction model. For example, an Encoder-Decoder architecture can be used.

[0097] Input: Historical time series (For example, using data from the past 12 hours).

[0098] Output: Predict future time series (For example, predicting data for the next hour).

[0099] Loss function: Use mean absolute error (MAE) as the loss function. .

[0100] Training: Historical data from various systems (KJ90X, KJ1193, etc.) and known disaster records (such as gas outbursts and water inrush events) are used as the training set. The backpropagation algorithm and Adam optimizer are employed to train all trainable parameters of the model. and ).

[0101] Example 2: Application of the present invention for early warning of gas outburst disasters

[0102] (4.1) Scene

[0103] Gas outbursts are the result of the coupling of ground stress, gas pressure, and coal mechanical properties, and are strongly influenced by geological structures (such as "developed fault structures") and ventilation. Background technologies can only monitor, not provide early warning.

[0104] (4.2) Model Configuration

[0105] Node features The following parameters were selected: gas (CH4) concentration, roof stress, anchor bolt stress (from the roof monitoring system), and wind speed. Weight It is highly dependent on coal seam permeability and geological structure. The factor should be set to a higher value near the fault (gas-rich area) (e.g. This indicates that the gas is easy to flow out or connect. Weight The ventilation network must be strictly followed because gas diffuses with the airflow. Model: The GP-ASTGNN model as defined in claim 6 is adopted, which can simultaneously fuse... and Constraints.

[0106] (4.3) Early warning logic

[0107] Real-time input: Input the CH4, stress, and wind speed data of the most recent 12 hours into the GP-ASTGNN model trained in Example 3.

[0108] Predicted output: The model outputs the results in real time (e.g., every 5 minutes) for the next hour. CH4 concentration at each node location and stress The predicted value.

[0109] Level 1 Warning (Gas Accumulation):

[0110] Condition:IF > 1.0% AND Minimum wind speed threshold

[0111] Action: at node The department issued an "accumulation warning" and coordinated with the "ventilation subsystem" to adjust the air volume.

[0112] Level 2 Warning (Highlighted Risk):

[0113] Condition:IF Dramatic changes occur (predicted slope > threshold) AND The AND node then rose rapidly. lie in High-weight regions in (e.g.) (near the fault)

[0114] Actions: Issue a “high-risk warning”, coordinate with the “power supply and distribution electronic system” to cut off power to the area, and coordinate with the “KJ1580J” personnel positioning system to evacuate personnel from the area.

[0115] (4.4) Beneficial effects

[0116] Traditional methods only trigger an alarm when CH4 has already exceeded its limit. This invention can predict which area is about to be affected by (CH4). ) changes in geostress and ( This system can provide early warnings in case of ventilation failure leading to excessive CH4 levels.

[0117] Example 3: Application of Mine Water Hazard Early Warning Based on the Invention

[0118] (5.1) Scenario

[0119] Mine water hazards originate from fissures filled with water, and their path is highly dependent on geological structures, especially the water conductivity of faults.

[0120] (5.2) Model Configuration

[0121] Node features Water level, flow rate (from KJ1193), and roof stress (mining-induced fractures can lead to water conduction). (key): weight It must be based on the "comprehensive hydrogeological map". The factor must reflect that the fault is a water-conducting channel. ) or water barrier ( ). In this scenario, the ventilation network It has little to do with water damage. The model is under training. The weights adaptively approach 0. This demonstrates the decoupling and adaptive capabilities of the GP-Conv layer of this invention (as described in Embodiment 3). Model: Employs the spatiotemporal graph neural network in (c), whose topology is at least as defined in claim 3. constraint.

[0122] (5.3) Early warning logic

[0123] Real-time input: Input the water level, flow rate, and stress data of the most recent 24 hours into the GP-ASTGNN model.

[0124] Predictive output: The model outputs the water level for the next 6 hours in real time (e.g., every 15 minutes). and traffic The predicted value.

[0125] Level 1 Warning (Slowly Escalating):

[0126] Condition:IF Safety threshold. Action: The "drainage subsystem" will automatically start the water pump.

[0127] Level II Warning (Signs of Sudden Flooding):

[0128] Condition:IF The predicted slope ( Exceeding historical maximum AND node exist Above and known aquifers or water-rich faults ( It has high connectivity. (Large). Action: Issue a "sudden flood warning", activate the "KJ1580J" personnel positioning system, and evacuate personnel from the area and downstream tunnels.

[0129] (5.4) Beneficial effects

[0130] Traditional methods only consider a single water level gauge (drainage subsystem). This invention can... Understanding where the water comes from allows us to distinguish between "normal water inrush" and "precursors of water inrush from fissures or faults," leading to more timely and accurate early warnings.

[0131] Example 4: Mine Fire Early Warning and Rescue Assistance Application Based on the Invention

[0132] (6.1) Scenario

[0133] The deadly danger of mine fires (as monitored by the KJ428 system) is the spread of CO, the path of which is mainly determined by the ventilation network.

[0134] (6.2) Model Configuration

[0135] Node features CO concentration and temperature (from KJ428). In this scenario, geology There is little connection. It will approach 0. (key): It must be a high-precision directed graph reflecting real-time airflow (Example 2.2). Model: Employs the spatiotemporal graph neural network in (c), whose topology is at least as defined in claim 5. constraint.

[0136] (6.3) Early warning logic and disaster simulation

[0137] Fire source location: When KJ428 is at the node When a rapid increase in both CO concentration and temperature (slope > threshold) is detected simultaneously, a "fire source" warning is immediately triggered.

[0138] Disaster simulation:

[0139] Node The CO value is used as input. The GP-ASTGNN model (due to its...) Item and (Weighted dominance) will mainly follow CO diffusion is extrapolated from the directed edges (i.e., downstream of the airflow). Model output. Each node Moment Predicted value.

[0140] (6.4) Generate rescue routes

[0141] Downstream disaster list: All automatically generated Downstream nodes > 300ppm (high-risk threshold) A list.

[0142] Emergency Response Coordination:

[0143] Release: Immediately push this "Disaster Affected List" and "Spread Path Map" to the "Integrated Management and Control Platform" and the "Information Guidance and Release Subsystem." Location: Invoke the "KJ1580J" personnel location system to retrieve all underground personnel located in the "Disaster Affected List" area. Planning: Automatically plan for these personnel to avoid downstream nodes and travel in the opposite direction. The retreat route (against the wind) is indicated and guided via electronic displays and voice alarms in the information dissemination subsystem.

[0144] (6.5) Beneficial effects

[0145] Traditional fire alarms only indicate "where the fire is." This invention utilizes... The physical constraints allow for immediate communication with the rescue command center: "The fire is at point A, and point C will arrive at points B and C in 10 minutes." "Downstream), all personnel at points B and C must immediately evacuate from route D (upwind)." This is an improvement on the isolated fire and personnel location system in the background art.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A coal mine disaster early warning method based on multi-source data fusion and spatiotemporal graph networks, characterized in that, Includes the following steps: a. Data acquisition steps: Real-time acquisition of multi-source heterogeneous monitoring data from multiple sensor nodes in the coal mine, including time-series disaster monitoring data and non-time-series geological environment data; b. Graph construction steps: Based on the spatial locations of the multiple sensor nodes and the non-temporal geological environment data, construct a spatiotemporal map of mine disasters that characterizes the correlation between the multiple sensor nodes; c. Model training steps: The time-series disaster monitoring data obtained in (a) is used as the node feature input. The topological structure of the mine disaster spatiotemporal map constructed in (b) is used to train the model through the spatiotemporal map neural network model to obtain the disaster early warning model. d. Warning generation steps: Input the real-time acquired time-series disaster monitoring data into the disaster warning model obtained in (c), output the prediction of the disaster risk level within the future time window, and generate a warning signal when the risk level exceeds a preset threshold.

2. The method according to claim 1, characterized in that, The time-series disaster monitoring data includes: gas concentration, carbon monoxide concentration, oxygen concentration, wind speed, and temperature data from the safety monitoring system; water level in the water tank, flow rate in the drainage ditch, and water level in the catchment channel from the water control system; fire monitoring data from the fire prevention system; and roof delamination and anchor stress data from the roof monitoring system.

3. The method according to claim 1, characterized in that, In the graph construction step (b), the topological structure of the mine disaster spatiotemporal graph is composed of a geological adjacency matrix. Defined; the geological adjacency matrix The construction includes: (b1) Obtain geological data from the non-time-series geological environment data, including mine geological structure map, fault distribution, rock strata permeability and coal seam thickness data; (b2) For any two nodes among the plurality of sensor nodes and The nodes are characterized based on the geological data. and Edge weights of geological connectivity between ; (b3) Based on the aforementioned edge weights Construct the geological adjacency matrix .

4. The method according to claim 3, characterized in that, The edge weight The calculation formula is: ， in, For nodes and Spatial distance between them; For nodes and Effective penetration rate along the path between; For the node and The influence factor of faults traversed by the path, when the path does not traverse a fault. When the path traverses a closed fault that acts as a barrier When the path traverses a conductive fault that serves as a passageway .

5. The method according to claim 4, characterized in that, In the graph construction step (b), the topological structure of the mine disaster spatiotemporal graph is further composed of a physical adjacency matrix. Defined; the physical adjacency matrix The construction includes: (b4) Obtain the mine physical field data from the non-time-series geological environment data, the mine physical field data including the mine ventilation network topology and ventilation dynamic parameters; (b5) Construct the ventilation network as a directed graph. The edge represents the wind flow path; (b6) For any two nodes among the plurality of sensor nodes and If and only if There exists a path from node To the node When calculating a directed path, the node is represented. and Edge weights of the correlation between airflow ,otherwise ; (b7) Construct the physical adjacency matrix .

6. The method according to claim 5, characterized in that, The spatiotemporal graph neural network model (c) is a geo-physical dual-constraint anisotropic spatiotemporal graph neural network GP-ASTGNN; the graph convolutional layer of the GP-ASTGNN updates the first... Layer node features To obtain the first Layer node features : ， in, It is the identity matrix; and They are respectively and The normalization degree matrix; , and For the first The trainable weight matrix of the layer; For the first The learnable adaptive adjacency matrix of the layer; This is the activation function.

7. The method according to claim 6, characterized in that, The early warning generation step (d) is for early warning of gas outburst disasters. The output of the model (c) includes predicted values ​​of gas concentration, ground stress and wind speed. The disaster risk level is determined based on the predicted values. The early warning signal is generated when the predicted gas concentration exceeds the critical value for gas outburst.

8. The method according to claim 3, characterized in that, The early warning generation step (d) is for mine water hazard early warning. The output of the model (c) includes the predicted values ​​of water level in the water tank and flow rate in the roadway. The disaster risk level is determined based on the predicted values. The early warning signal is generated when the predicted rate of water level rise exceeds the critical value for water hazard.

9. The method according to claim 5, characterized in that, The early warning generation step (d) is for mine fire early warning; the output of the model (c) is the spatiotemporal diffusion prediction of carbon monoxide (CO) concentration; the early warning signal includes fire source location and data based on the physical adjacency matrix. List of downstream nodes along the simulated CO diffusion path.

Images