Mine ventilation dynamic control method and system based on spatiotemporal graph convolutional network

Abstract

The present invention relates to the field of mine ventilation technology, and more particularly to a method and system for dynamic mine ventilation control based on a spatiotemporal graph convolutional network. The control method comprises the following steps: Step 1: Ventilation system modeling and data acquisition; Step 2: Graph convolutional network modeling and anomaly prediction; Step 3: Generation and execution of collaborative control strategies for ventilation equipment; Step 4: Edge node calibration and model optimization; Step 5: Full-process control and emergency response. By integrating the ventilation network topology with multimodal time series data through a spatiotemporal graph convolutional network, a joint model comprising spatial feature extraction and time series modeling is constructed, enabling 30-minute advance warning of ventilation anomalies and propagation path prediction, overcoming the hysteresis response limitations of traditional systems that rely solely on a single parameter threshold, and improving risk prediction capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of mine ventilation, and in particular to a mine ventilation dynamic control method and system based on a spatiotemporal graph convolutional network. Background Art

[0002] Mine ventilation systems are crucial in mining operations, ensuring a safe and comfortable working environment by supplying sufficient fresh air to underground working areas and promptly exhausting contaminated air. Traditional mine ventilation systems typically utilize a large main fan design, resulting in a total air volume far exceeding actual requirements. As mining progresses deeper, simultaneous operations in multiple stopes and large-scale mining operations become increasingly common, resulting in an extremely complex underground ventilation network, with complex systems such as series, parallel, and angled connections forming one after another.

[0003] In this context, existing technologies have many defects. On the one hand, the rough ventilation mode is difficult to achieve fine dynamic control, and cannot meet the different air volume requirements generated by the dynamic changes in factors such as personnel, equipment, number of work points, process type, working status, blasting process, etc. in underground workplaces, resulting in extremely difficult calculation and adjustment of required air volume. On the other hand, the formation of complex ventilation networks has made ventilation safety issues more prominent, and the existing ventilation management level is unable to cope with this complex situation. The effective air volume rate is low, which seriously threatens the physical and mental health of underground workers. Therefore, we propose a dynamic control method and system for mine ventilation based on spatiotemporal graph convolutional networks. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a mine ventilation dynamic control method and system based on a spatiotemporal graph convolutional network, thereby solving the technical problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] The dynamic control method of mine ventilation based on spatiotemporal graph convolutional network includes the following steps:

[0007] Step 1: Ventilation system modeling and data collection: building a digital twin foundation for the mine ventilation system, generating a digital map of the ventilation network including the topological structure, and deploying multimodal sensor nodes to collect data;

[0008] Step 2: Graph Convolutional Network Modeling and Anomaly Prediction: Based on the generated digital topology graph and multimodal time series data, a spatiotemporal graph convolutional network is constructed to extract the spatiotemporal correlation characteristics of the ventilation system, and perform ventilation anomaly assessment and propagation path prediction.

[0009] Step 3: Generate and execute a coordinated control strategy for ventilation equipment. Based on the risk index and propagation path output by the spatiotemporal feature model, a reinforcement learning algorithm is used to build a decision model to achieve automated control of fan speed and damper opening.

[0010] Step 4: Edge node calibration and model optimization, designing a multi-level calibration and iteration mechanism to ensure the accuracy of the system in the long term;

[0011] Step 5: Full process control and emergency response. The system follows the closed-loop logic of "monitoring-prediction-decision-execution-feedback" and has a built-in four-level emergency response mechanism.

[0012] In one possible implementation, during the ventilation system modeling and data collection steps:

[0013] S1: Use LiDAR, inertial navigation data, and BIM technology to build a roadway point cloud model and convert it into a digital map of the ventilation network;

[0014] S2: Abstract the physical entities of tunnel intersections, fans, and dampers into a directed weighted graph to calculate the wind resistance matrix and the maximum ventilation volume matrix;

[0015] S3: Deploy multimodal sensor nodes at specific intervals in different areas to collect data such as gas concentration, wind speed, and vibration, and transmit them to the edge controller via the industrial ring network;

[0016] S4: Use the IEEE1588 precision time protocol to achieve clock synchronization and preprocess the data, including outlier filtering and normalization, to form a multidimensional time series dataset.

[0017] In one possible implementation, in the graph convolutional network modeling and anomaly prediction steps:

[0018] S1: Construct a spatiotemporal graph convolutional network, including two modules: spatial feature extraction and time series modeling;

[0019] S2: The spatial feature layer uses a two-layer graph convolutional network to learn the ventilation coupling relationship between node neighborhoods and cross-regions;

[0020] S3: The time series layer uses bidirectional LSTM and attention mechanism to capture temporal dependencies and highlight key temporal features;

[0021] S4: Quantify ventilation status through risk assessment models, trigger early warnings and predict abnormal transmission paths.

[0022] In one possible implementation, during the steps of generating and executing the ventilation equipment coordinated control strategy:

[0023] S1: Use the proximal policy optimization (PPO) algorithm to build a decision model and define the state space, action space, and reward function.

[0024] S2: Through simulation training, the agent learns the control strategy that complies with mining regulations;

[0025] S3: After the decision module outputs the optimal action, it is transmitted to the execution layer through the industrial bus to achieve fan speed and damper opening control.

[0026] In one possible implementation, during the edge node calibration and model optimization steps:

[0027] S1: Implement online sensor calibration and dynamic model correction to ensure that the topology model is consistent with the actual tunnel structure;

[0028] S2: Iteratively optimize the model using online learning and adversarial training;

[0029] S3: Organize expert reviews every quarter and adjust model parameters to match the latest safety standards.

[0030] In one possible implementation, during the full process control and emergency response steps:

[0031] S1: The system operates in a closed-loop logic of “monitoring-prediction-decision-execution-feedback”;

[0032] S2: Trigger the corresponding four-level emergency response mechanism for different risk levels, including automatic blocking of pollution paths, activation of backup ventilation systems, and intelligent escape guidance measures.

[0033] In one possible implementation, a mine ventilation dynamic control system based on a spatiotemporal graph convolutional network includes:

[0034] The data acquisition module is used to realize the real-time acquisition and preprocessing of mine ventilation environment parameters, equipment status and spatial topology data. The data acquisition module includes a sensor array unit, a data synchronization and preprocessing unit and a topology calibration unit;

[0035] The spatiotemporal modeling module, based on the spatiotemporal graph convolutional network, realizes the joint modeling of the spatial topological characteristics and time series characteristics of the ventilation network, and outputs the risk index and abnormal propagation path. The spatiotemporal modeling module includes a spatial feature extraction unit, a time series modeling unit, and a risk assessment and path prediction unit;

[0036] An intelligent decision-making module generates optimal control based on reinforcement learning, strategically balancing safety, energy consumption, and equipment loss to automate the execution of coordinated equipment control. The intelligent decision-making module includes a reinforcement learning training unit, a policy reasoning and safety assessment unit, and a multi-objective optimization unit.

[0037] The execution control module receives instructions from the intelligent decision-making module, drives the fan, damper and other equipment to perform control actions, and provides feedback on the execution effect. The execution control module includes an actuator control unit and a feedback data acquisition unit;

[0038] The system management module provides a human-computer interaction interface, system calibration and model iteration support to ensure the long-term stable operation of the system. The system management module includes a human-computer interaction unit, a calibration and iteration unit, and an emergency response unit.

[0039] Beneficial effects compared with existing technologies:

[0040] 1. This solution uses a spatiotemporal graph convolutional network to fuse ventilation network topology with multimodal time series data, building a joint model that combines spatial feature extraction and time series modeling. This enables 30-minute advance warning of ventilation anomalies and prediction of their propagation paths, overcoming the limitations of traditional systems that rely on a single parameter threshold for delayed response and improving risk prediction capabilities.

[0041] 2. This solution uses the Proximal Policy Optimization (PPO) reinforcement learning algorithm to build a decision-making model. This model defines a multidimensional state space encompassing risk index, air volume, and equipment status, as well as a discrete action space for speed / opening adjustment. A reward function balances safety, energy consumption, and equipment loss. After 30,000 simulation training runs, a control strategy that complies with mining regulations is generated. The control execution integrates Bayesian optimization and fuzzy logic algorithms, achieving a threefold increase in strategy generation speed and a 22% reduction in energy consumption.

[0042] 3. This solution incorporates a three-level iterative calibration system consisting of online sensor calibration, dynamic model correction, and online learning. This system, combined with quarterly expert reviews, ensures the system's long-term topological model consistency with physical lanes, data reliability, and anti-interference capabilities. This ensures practical and effective decision-making even in the face of data noise or equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention with reference to the accompanying drawings.

[0044] Figure 1 This is a schematic flow chart of the mine ventilation dynamic control method of the present invention;

[0045] Figure 2 This is a schematic diagram of the framework of the mine ventilation dynamic control system of the present invention. DETAILED DESCRIPTION

[0046] Preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the present invention can be implemented in various forms, and therefore the present invention is not limited to the embodiments described below. In addition, in order to more clearly describe the present invention, components that are not related to the present invention will be omitted from the drawings.

[0047] The technical solutions in the embodiments of the present application are to solve the problems of the above-mentioned background technology, which are generally as follows:

[0048] Example 1: A method for dynamic control of mine ventilation based on a spatiotemporal graph convolutional network, comprising the following steps:

[0049] Step 1: Ventilation system modeling and data collection

[0050] First, we need to build a digital twin foundation for the mine ventilation system. We use LiDAR to perform high-density 3D scanning of the tunnel (sampling point spacing 1m, accuracy 2cm), combined with inertial navigation data to generate coordinates The tunnel point cloud model is converted into a digital map of the ventilation network including the topological structure using BIM technology. Based on graph theory, physical entities such as tunnel intersections, fans, and dampers are abstracted into directed weighted graphs. , where the nodes Include The key position is the main fan. , intersection of return air lane Wait, side represents the ventilation path, and its wind resistance is calculated using the Darcy-Weisbach formula: ,in Take the corresponding value of the roadway roughness (such as 0.018 for bolting and shotcreting roadway), is the measured path length, the equivalent diameter of the non-circular tunnel ,in is the cross-sectional area of ​​the tunnel. During the non-production period each month, UAV LiDAR scanning is used to monitor tunnel deformation. When the path length change rate is When, according to Correct the topology and recalculate the wind resistance matrix and maximum ventilation matrix , ensuring the dynamic consistency between the digital model and the physical system.

[0051] In the data collection phase, for areas prone to gas accumulation, specifically, mining working faces and blind alley entrances, Multimodal sensor nodes are deployed at 8m intervals, and each node integrates:

[0052] Gas concentration sensor, infrared absorption type, detection range 0-4%LEL, accuracy 0.01%, response time <1s;

[0053] Hot film wind speed sensor, measuring range 0-30m / s, accuracy 0.1m / s, sampling frequency 10Hz;

[0054] Triaxial vibration accelerometer, range 10g, sampling frequency 1kHz, used to monitor the status of fan bearings.

[0055] Conventional areas, specifically, transport lanes and pit bottom parking lots are equipped with temperature and humidity sensors at 25m intervals, with an accuracy of 0.5℃ / 2%RH, and monitors ambient temperature, humidity, and fan power consumption through current transformers. All sensors are connected to the edge controller via an underground industrial ring network, using the IEEE1588 precision time protocol for clock synchronization. The data preprocessing process includes:

[0056] 1. Outlier filtering: Calculate the mean by sliding the window with a window size of 300s and standard deviation , remove Noise points and fill them with linear interpolation: , among which, when , data is used In addition, the data deal with.

[0057] 2. Normalization: Mapping multi-source data to The interval formula is: , which ultimately forms 、 、 、 、 Multidimensional time series dataset with equal parameters ,in is the total number of sensors.

[0058] By building a high-precision digital map and sensor network, the spatial structure of the ventilation system can be calculated and the operating parameters can be perceived in real time.

[0059] Step 2: Graph Convolutional Network Modeling and Anomaly Prediction

[0060] Based on the digital topology map generated in step 1 and the multimodal time series data, a spatiotemporal graph convolutional network (ST-GCN) is constructed to extract the spatiotemporal correlation features of the ventilation system. The network is divided into two modules: spatial feature extraction and time series modeling:

[0061] 1. Spatial feature layer: Graph Convolutional Network (GCN)

[0062] The first layer of GCN learns the ventilation coupling relationship of the node's immediate neighborhood. Input feature matrix Fusion node attributes, specifically, roadway cross-sectional area , Design air volume With real-time sensor data, by adding a self-loop adjacency matrix Sum degree matrix Perform normalized propagation: ,in Characterizes the ventilation correlation between a node and its direct neighbors, specifically, the direct impact intensity of increased gas concentration in a certain excavation face on the adjacent return air channel.

[0063] The second layer of GCN is through the second-order adjacency matrix To capture the cross-area ventilation impact, the propagation formula is: This layer can learn the indirect impact of the change in the main fan speed on the entire return air network, specifically, the wind speed change trend in the lane 200 meters away after the main fan air volume is adjusted.

[0064] 2. Time Series Layer: Bidirectional LSTM and Attention Mechanism

[0065] The spatial feature sequence output by GCN ,in, , corresponding to 10 minutes of data, input into the bidirectional LSTM network, capturing the temporal dependencies of the past and future at the same time. With the inverted hidden state Splice to , calculate the feature weights at each moment through the attention mechanism: , ,in, is the trainable weight vector, and the final output is Key time series features such as sudden changes in gas concentration and abnormal fluctuations in wind speed are highlighted.

[0066] 3. Ventilation abnormality assessment and transmission route prediction

[0067] Comprehensive spatiotemporal feature outputs are used to quantify ventilation status through risk assessment models: ,in, The regional standard wind speed is determined by the Coal Mine Safety Regulations. is the ideal temperature. When an alarm is triggered, the abnormal propagation path prediction is started. The improved Dijkstra algorithm is used to search for high-risk paths. The cost function integrates wind resistance, path length and real-time gas concentration: ,in, ,The path probability distribution is generated through 1000 Monte Carlo simulations, and the top three paths and their propagation probabilities are output, specifically “coal mining face → return air uphill → total return air lane”, with a probability of 78%.

[0068] Through joint spatiotemporal modeling, a 30-minute advance warning of ventilation anomalies was achieved with an accuracy of 89% and propagation path prediction. Compared with the traditional system that only relies on the delayed response of a single parameter threshold, it significantly improves the risk prediction ability.

[0069] Step 3: Generation and execution of ventilation equipment coordinated control strategy

[0070] Risk index based on the output of spatiotemporal feature model and propagation paths ,The proximal strategy is optimized by the PPO algorithm to build a decision model to realize the automatic control of the fan speed and the damper opening.

[0071] 1. Learning framework design

[0072] State Space: , including global risk, total air volume, equipment status and time and space characteristics, dimensions ,in is the number of fans, is the number of dampers;

[0073] Action Space: ,in , ;

[0074] Reward function: ,in, is the risk reduction, The amount of energy consumption reduction for fans; is the change in equipment loss index, based on vibration intensity calculate.

[0075] Through 30,000 simulation trainings (simulating scenarios such as gas exceeding the limit and fan failure), the intelligent agent learned a control strategy that complies with mining regulations. Specifically, when the gas concentration in a certain area rises to 1.2%LEL, it will prioritize turning on the fan in the air intake lane of that area. To 1450rpm, close the return air damper at the same time 20°.

[0076] 2. Regulation execution and equipment control

[0077] The decision module outputs the optimal action After that, it is transmitted to the execution layer via the industrial bus (Profibus-DP):

[0078] Fan speed control: Bayesian optimization algorithm is used to solve the optimal speed combination , in satisfying and Under the premise of minimizing the energy consumption function The closed-loop speed control is achieved through a PID controller: ,in, , the parameters are adjusted by Ziegler-Nichols method to , , , ensure that the speed adjustment overshoot is <5% and the adjustment time is <10s.

[0079] Air door opening control: Based on fuzzy logic algorithm, input risk deviation and rate of change , output opening adjustment amount Specifically, the error and When the air damper is closed, it closes 15° to block the pollution diffusion path. The action execution time is less than 8s and it is protected by mechanical limit.

[0080] Through reinforcement learning, intelligent generation of control strategies is achieved. Compared with traditional manual experience control, the strategy generation speed is increased by 3 times and energy consumption is reduced by 22%.

[0081] Step 4: Edge node calibration and model optimization

[0082] To ensure the accuracy of the system in long-term operation, a multi-level calibration and iteration mechanism is designed:

[0083] 1. Sensor and model calibration

[0084] Sensor online calibration: The self-calibration procedure is triggered at 3:00 am every day, and the gas sensor closes the gas path to measure the environmental background value. , the data correction formula is: ,when When LEL, it is determined that the sensor is faulty and an alarm is issued. The wind speed sensor passes the standard wind tunnel (accuracy) every quarter. 0.05m / s) to perform multi-point calibration and fit the linear equation , ensuring that the measurement error is less than 1.5%.

[0085] Dynamic model correction: Using the underground UWB personnel positioning system (accuracy <15cm), the coordinates of key points in the tunnel are collected weekly. , when the deviation from the digital map exceeds 20cm, the node coordinates are updated by the least squares method: Recalculate wind resistance after update Maximum flow rate of ventilation path , ensuring that the topological model is consistent with the actual tunnel structure.

[0086] 2. Model Iterative Optimization

[0087] Online learning: Incremental training is performed daily using the previous 24 hours of data (about 17,000 samples) using the momentum SGD optimizer, where the learning rate , momentum 0.9, update model parameters, and focus on optimizing the prediction accuracy of abnormal scenarios (such as sudden increase in gas after blasting).

[0088] Adversarial training: artificially inject simulated abnormal data, specifically, a 40% drop in wind speed and interruption of sensor signals, and then use the adversarial loss function to simulate abnormal data. Improve model robustness to ensure reliable decisions despite data noise or device failures.

[0089] Expert review: Ventilation experts and data scientists conduct joint reviews every quarter, analyzing the contribution of each feature to decision-making based on Shapley values. Specifically, they found that the gas concentration weight could be adjusted from 0.5 to 0.6 to strengthen the safety priority strategy, and simultaneously adjusted the reward function parameters to match the latest safety regulations.

[0090] Step 5: Full process control and emergency response

[0091] The system operation follows the closed-loop logic of "monitoring-prediction-decision-execution-feedback". The specific process is as follows:

[0092] Initialization phase: The edge controller loads the latest digital map after powering on. The trained ST-GCN+PPO model sends calibration instructions to all sensors and synchronizes the daily production plan with the ground dispatch center, specifically the coal mining face advancement speed, blasting time window, etc.

[0093] Real-time monitoring phase: The sensor collects data at a frequency of 10Hz, which is input into the ST-GCN model after preprocessing and outputs a risk index every 5 minutes. and the probability of abnormal propagation path .like , the system maintains the current ventilation state; if , triggering the reinforcement learning decision module.

[0094] Control execution phase: The PPO model generates the optimal action within 50ms After safety verification, specifically checking the balance of air volume supply and demand, the command is sent to the fan and damper actuators through redundant communication links (optical fiber + wireless Mesh), requiring command parsing to be completed within 0.3 seconds and equipment action to be completed within 10 seconds.

[0095] Effect feedback stage: After the action is executed, the sensor continues to collect data. If decline , then the regulation case is stored in the experience pool If the target is not met, a second decision is triggered and the strategy is adjusted and re-executed.

[0096] For extreme abnormal scenarios, the system has a built-in four-level emergency response mechanism:

[0097] Level 1 response (gas concentration > 1.5% LEL or risk index ):

[0098] Automatically block the pollution path: According to the high-risk transmission path predicted by the ST-GCN model, immediately close the upstream damper within 3 levels, specifically, close 、 , cut off the abnormal airflow diffusion channel, and the blocking efficiency must be completed within 15 seconds;

[0099] Start backup ventilation system: remotely activate backup fans To full speed (2000rpm), through the formula ,in The fan air volume coefficient can quickly increase the total air volume to 1.8 times the design value;

[0100] Intelligent escape guidance: Combined with UWB personnel positioning data, customized evacuation instructions are issued through the underground broadcast system. Specifically, "Personnel in Mining Area No. 3, please evacuate along the return air tunnel → East Wing Auxiliary Shaft." At the same time, the shortest disaster avoidance route is dynamically displayed on the tunnel LED screen.

[0101] Second level response (sudden equipment failure, fan vibration intensity> 5m / s 2 ):

[0102] Fault location and isolation: Using vibration accelerometer data and a fault knowledge graph, which includes 137 fault modes, the faulty device can be located within 2 seconds. Specifically, If the bearing is worn, the system will automatically switch to the backup fan and shut down the power supply of the faulty equipment;

[0103] Life prediction and maintenance scheduling: based on LSTM life prediction model ,in, m / s 2 As the threshold, if the remaining life is less than 14 days, a maintenance work order is automatically generated and pushed to the maintenance team, and the ventilation network topology is adjusted to bypass the faulty equipment.

[0104] Level 3 response (roadway deformation causes wind resistance to increase by more than 20%):

[0105] Automatic topology reconstruction: Lane deformation is detected through UWB positioning data, specifically, Increase by 30m, triggering the edge controller to update the ventilation network model in real time and recalculate the wind resistance and air volume distribution at each node;

[0106] Dynamic air pressure balancing: Reinforcement learning is used to re-optimize the fan speed and damper opening to ensure that the air volume fluctuation in the affected area is less than 10%. Specifically, the fan speed in the air inlet channel is increased by 100 rpm, and the central regulating damper is opened 15 degrees to balance the air pressure.

[0107] Level 4 response (system communication interruption):

[0108] Edge autonomous mode: The edge controller switches to offline mode and continues to operate based on the locally stored historical model and the last hour's data, attempting to reconnect to the central system every 10 minutes.

[0109] Preset strategy execution: If communication is not restored within 30 minutes, the preset conservative control strategy is automatically enabled. Specifically, the speed of all fans is maintained at 80% of the rated value and the damper opening is maintained at 50° until communication is restored or manual intervention is performed.

[0110] In order to ensure that operators can effectively monitor and intervene in the system, a multi-level human-computer interaction interface is designed:

[0111] Ground control center: A 3D digital twin screen displays the ventilation network status in real time, integrating risk heat maps, equipment operating parameters, and abnormal warning information. Specifically, high-risk areas are marked with flashing red. Experts can manually adjust the control strategy by dragging the fan / damper icon. The system automatically calculates the risk index change after manual intervention.

[0112] Underground mobile terminals: Patrol personnel are equipped with explosion-proof tablet computers that display ventilation parameters within a 50-meter radius of their current location in real time. They can also trigger local ventilation optimization with one click, specifically temporarily increasing the air volume at the tunneling face. This command is delivered to the edge controller within 2 seconds via the wireless mesh network.

[0113] Alarm and log system: A hierarchical alarm mechanism (yellow - warning, red - emergency) is adopted, and abnormal information is simultaneously pushed through sound and light alarms and SMS notifications (ground manager, safety director). Historical operation logs and control records are stored in the blockchain system to ensure traceability and clear accountability.

[0114] In addition, the following verification process is performed:

[0115] Laboratory verification: A 1:100 scale model was deployed in a simulated mine environment. By artificially injecting gas (concentration 0-2% LEL) and adjusting the fan speed (500-1500 rpm), the ST-GCN model's prediction accuracy for gas diffusion paths was verified. Specifically, the prediction accuracy was 87%, and the PPO strategy energy consumption optimization rate reached 20.1%.

[0116] Underground pilot test: A three-month trial run was conducted at a working face of Yangquan Coal Industry, completing the following verifications:

[0117] Sensor network stability: 90 days of continuous operation without node failure, and 99.8% data integrity rate;

[0118] Effectiveness of control strategy: Successfully handled 12 gas concentration exceeding the limit (0.85%-1.2% LEL), with an average handling time of 32 seconds, 70% shorter than manual control;

[0119] Edge computing performance: response delay of 0.28 seconds.

[0120] Example 2: A mine ventilation dynamic control system based on a spatiotemporal graph convolutional network, comprising the following modules:

[0121] 1. Data Acquisition Module

[0122] This system enables real-time collection and preprocessing of mine ventilation environment parameters, equipment status, and spatial topology data, providing a multi-dimensional data source for subsequent modeling and decision-making. It includes a sensor array unit, a data synchronization and preprocessing unit, and a topology calibration unit.

[0123] 1. Sensor array unit

[0124] The unit is deployed in key areas of the mine, with differentiated density deployment for different risk levels. In high-risk areas where gas is prone to accumulate, specifically the mining working face, return air lane, and Multimodal sensor nodes are densely deployed at 8-meter intervals. Each node integrates an infrared absorption gas concentration sensor with a detection range of 0-4%LEL and an accuracy of 0.01%, response time <1 second, hot film wind speed sensor, its measurement range is 0-30m / s, accuracy 0.1m / s, sampling frequency 10Hz and triaxial vibration accelerometer, its range 10g, sampling frequency 1kHz, used to capture key parameters such as gas concentration, wind speed fluctuation and fan bearing vibration in real time. In conventional areas such as transport lanes and pit bottom parking lots, temperature and humidity sensors are deployed at intervals of 25 meters. 0.5℃ / 2%RH, and monitor the ambient temperature, humidity and fan power consumption through current transformers. All sensors transmit data in real time through the underground industrial ring network, and the output includes gas concentration , wind speed , equipment speed , air door opening Time series data stream with equal parameters, with Unix timestamp to ensure time consistency.

[0125] 2. Data synchronization and preprocessing unit

[0126] This unit is based on the IEEE1588 time protocol to achieve nanosecond clock synchronization between distributed sensor nodes and edge controllers, with synchronization error controlled within 1ms, ensuring the spatiotemporal alignment of multi-source data. The data preprocessing phase uses a sliding window technology, with the window size consistent with the first embodiment, by calculating the mean of the data in the sliding window. and standard deviation ,based on In principle, outliers were eliminated and valid data were normalized.

[0127] 3. Topology calibration unit

[0128] The unit regularly monitors changes in the tunnel's spatial structure through UWB positioning base stations and drone LiDAR scanning. During non-production hours each week, the system automatically triggers tunnel deformation detection, compares current point cloud data with historical topology models, and calculates the path length change rate. .when When the topology correction formula and wind resistance update formula in the first embodiment are called, the wind resistance matrix and the maximum ventilation volume matrix of the ventilation network are recalculated to ensure the dynamic consistency between the digital model and the physical laneway.

[0129] 2. Spatiotemporal Modeling Module

[0130] Based on the spatiotemporal graph convolutional network (ST-GCN), the system jointly models the spatial topological and temporal characteristics of the ventilation network, outputting risk indices and abnormal propagation paths. The system includes a spatial feature extraction unit, a time series modeling unit, and a risk assessment and path prediction unit.

[0131] 1. Spatial feature extraction unit

[0132] This unit deploys a two-layer graph convolutional network (GCN) on the edge controller's GPU (NVIDIA Jetson XavierNX). The first layer inputs node attributes, specifically, the lane cross-sectional area. , Design air volume With the preprocessed sensor data, by adding the self-loop adjacency matrix Sum degree matrix , using the GCN propagation formula in Example 1 to extract the node neighborhood ventilation correlation characteristics, characterizing the direct impact of increased gas concentration on adjacent tunnels. The second layer is through the second-order adjacency matrix The cross-regional ventilation coupling relationship is learned, and the formula is similar to the second-order GCN formula in Example 1, which captures the indirect impact of the main fan speed change on the wind speed in the remote tunnel.

[0133] 2. Time Series Modeling Unit

[0134] This unit inputs the spatial feature sequence output by GCN into the bidirectional LSTM network through the forward hidden state With the inverted hidden state The attention mechanism is introduced, such as the attention weight formula in Example 1, which calculates the feature weights at each moment, highlights the key time series features such as sudden changes in gas concentration and abnormal fluctuations in wind speed, and generates a weighted time series feature vector. , for subsequent risk assessment.

[0135] 3. Risk Assessment and Path Prediction Unit

[0136] This unit is based on the spatiotemporal feature vector With the original sensor data, the comprehensive risk index is calculated by the risk assessment formula in Example 1. Anomaly prediction is triggered when the error is detected. The improved Dijkstra algorithm is combined with Monte Carlo simulation to search for high-risk propagation paths using the path cost formula in Example 1, and the top three paths and their propagation probabilities are output.

[0137] 3. Intelligent Decision-making Module

[0138] Generate optimal control strategies based on reinforcement learning, balance safety, energy consumption, and equipment loss, and automate the execution of coordinated equipment control. This system includes a reinforcement learning training unit, a policy reasoning and safety assessment unit, and a multi-objective optimization unit.

[0139] 1. Reinforcement Learning Training Unit

[0140] This unit constructed a virtual mine ventilation scenario, simulating abnormal scenarios such as gas overruns and fan failures. It collected 30,000 training samples and trained an intelligent agent using the Proximal Policy Optimization (PPO) algorithm. During training, the agent interacted with the environment to collect state-action-reward sequences. The PPO loss function formula, as described in Example 1, was used to optimize the policy network, learning a control strategy that complies with mining regulations. The trained model was then compressed and deployed on an edge controller.

[0141] 2. Policy Reasoning and Security Accuracy Unit

[0142] This unit receives the risk index output by the spatiotemporal modeling module in real time. , total air volume , device status and timing characteristics , call the trained PPO model to generate fan speed and damper opening adjustment instructions. The action space is defined as , The generated strategy must pass safety checks, including air volume constraints. The equipment is limited, specifically, the fan speed is 0-1500rpm and the damper opening is 10°-90°, to ensure that the control plan is feasible.

[0143] 3. Multi-objective optimization unit

[0144] This unit further refines the execution parameters of the control strategy. For fan speed optimization, the energy consumption minimization problem is solved based on the Bayesian optimization algorithm. Under the premise of meeting the risk control requirements, the optimal speed combination is calculated as shown in the energy consumption optimization formula in Example 1. For damper opening control, the fuzzy logic algorithm is used to calculate the optimal speed combination according to the risk deviation. and rate of change Dynamically adjust the opening, specifically, the error and When the air damper is closed, it closes 15° to block the pollution diffusion path.

[0145] 4. Execution Control Module

[0146] Receives instructions from the intelligent decision-making module, drives fans, dampers, and other equipment to perform control actions, and provides feedback on the execution results to form a closed-loop control. It includes an actuator control unit and a feedback data acquisition unit.

[0147] 1. Actuator control unit

[0148] This unit analyzes the control instructions and drives the equipment to execute them. For fan speed control, a PID controller is used to implement closed-loop regulation. For example, the PID control formula in Example 1 calculates the control quantity, where the error , ensuring that the speed adjustment overshoot is greater than 5% and the adjustment time is less than 10 seconds. For the damper opening control, the damper is adjusted by an electric actuator, and the built-in mechanical limit device ensures that the opening is within the range of 10°-90°.

[0149] 2. Feedback data collection unit

[0150] After the equipment executes control commands, this unit collects real-time feedback data, including actual speed, damper opening, gas concentration after control, and risk index, and transmits it back to the intelligent decision-making module via the industrial ring network. Based on this feedback data, the system calculates the actual reward value, as shown in the reward function formula in Example 1, updates the experience pool, and triggers fine-tuning of the model, forming a closed loop of "decision-execution-feedback-optimization."

[0151] 5. System Management Module

[0152] Provides human-computer interaction interface, system calibration, and model iteration support to ensure long-term stable operation of the system. Includes human-computer interaction unit, calibration and iteration unit, and emergency response unit.

[0153] 1. Human-computer interaction unit

[0154] This unit consists of a ground-based command center and an underground mobile terminal. The ground-based command center displays the ventilation network topology, risk heat map, and equipment status in real time through a 3D digital twin interface. Experts can manually adjust fan speed and damper opening by dragging and dropping, with the system automatically predicting changes in risk index after these adjustments. The underground mobile terminal, an explosion-proof tablet computer, displays ventilation parameters within a 50-meter radius of the inspector in real time and supports one-click triggering of local ventilation optimization commands, which are delivered to the edge controller within 2 seconds via the wireless mesh network.

[0155] 2. Calibration and Iteration Unit

[0156] This unit is responsible for sensor calibration and continuous model optimization. Every morning, the sensor zero point calibration is automatically triggered. Specifically, the gas sensor background value A fault alarm is triggered when the LEL is exceeded. Sensor sensitivity is calibrated quarterly using standard gas samples and wind tunnel equipment to ensure measurement error is less than 2%. For model iteration, incremental training is performed daily using the latest data, and the momentum SGD optimizer is used to update parameters. Expert reviews are conducted quarterly, and risk assessment formula weights are adjusted based on Shapley value analysis to improve the model's sensitivity to safety risks.

[0157] 3. Emergency Response Unit

[0158] The unit has a built-in four-level emergency response mechanism. In the first level response, the system automatically closes the shortest path to the abnormal node. The upstream damper of the emergency stop is activated, the backup fan is started to run at full speed, and a dynamic escape route is generated in combination with UWB personnel positioning. In the second-level response, the faulty equipment is located within 2 seconds and the backup unit is switched to. In the third-level response, the ventilation topology is updated in real time and the control strategy is re-optimized. In the fourth-level response, the edge controller switches to offline mode and executes the preset conservative strategy until communication is restored.

[0159] Finally, it should be noted that the above embodiments are merely examples for the purpose of illustrating the present invention and are not intended to limit the embodiments. Those skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. It is not necessary and impossible to provide an exhaustive list of all embodiments. However, obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A dynamic control method for mine ventilation based on spatiotemporal graph convolutional network, characterized in that: The following steps are involved: Step 1: Ventilation system modeling and data collection: building a digital twin foundation for the mine ventilation system, generating a digital map of the ventilation network including the topological structure, and deploying multimodal sensor nodes to collect data; Step 2: Graph Convolutional Network Modeling and Anomaly Prediction: Based on the generated digital topology graph and multimodal time series data, a spatiotemporal graph convolutional network is constructed to extract the spatiotemporal correlation characteristics of the ventilation system, and perform ventilation anomaly assessment and propagation path prediction. Step 3: Generate and execute a coordinated control strategy for ventilation equipment. Based on the risk index and propagation path output by the spatiotemporal feature model, a reinforcement learning algorithm is used to build a decision model to achieve automated control of fan speed and damper opening. Step 4: Edge node calibration and model optimization, designing a multi-level calibration and iteration mechanism to ensure the accuracy of the system in the long term; Step 5: Full process control and emergency response. The system follows the closed-loop logic of "monitoring-prediction-decision-execution-feedback" and has a built-in four-level emergency response mechanism.

2. The mine ventilation dynamic control method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: In the ventilation system modeling and data collection steps: S1: Use LiDAR, inertial navigation data, and BIM technology to build a roadway point cloud model and convert it into a digital map of the ventilation network; S2: Abstract the physical entities of tunnel intersections, fans, and dampers into a weighted graph, and calculate the wind resistance matrix and maximum ventilation volume matrix; S3: Deploy multimodal sensor nodes at specific intervals in different areas to collect gas concentration, wind speed, and vibration data, and transmit them to the edge controller through the industrial ring network; S4: Use the IEEE1588 precision time protocol to achieve clock synchronization and preprocess the data, including outlier filtering and normalization, to form a multidimensional time series dataset.

3. The mine ventilation dynamic control method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: In the graph convolutional network modeling and anomaly prediction steps: S1: Construct a spatiotemporal graph convolutional network, including two modules: spatial feature extraction and time series modeling; S2: The spatial feature layer uses a two-layer graph convolutional network to learn the ventilation coupling relationship between node neighborhoods and cross-regions; S3: The time series layer uses bidirectional LSTM and attention mechanism to capture temporal dependencies and highlight key temporal features; S4: Quantify ventilation status through risk assessment models, trigger early warnings and predict abnormal transmission paths.

4. The mine ventilation dynamic control method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: In the steps of generating and executing the coordinated control strategy for ventilation equipment: S1: Use the proximal policy optimization (PPO) algorithm to build a decision model and define the state space, action space, and reward function. S2: Through simulation training, the agent learns the control strategy that complies with mining regulations; S3: After the decision module outputs the optimal action, it is transmitted to the execution layer through the industrial bus to achieve fan speed and damper opening control.

5. The mine ventilation dynamic control method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: In the edge node calibration and model optimization steps: S1: Implement online sensor calibration and dynamic model correction to ensure that the topology model is consistent with the actual tunnel structure; S2: Iteratively optimize the model using online learning and adversarial training; S3: Organize expert reviews every quarter and adjust model parameters to match the latest safety standards.

6. The mine ventilation dynamic control method based on spatiotemporal graph convolutional network according to claim 1 is characterized in that: In the whole process control and emergency response steps: S1: The system operates in a closed-loop logic of "monitoring-prediction-decision-execution-feedback"; S2: Trigger the corresponding four-level emergency response mechanism for different risk levels, including automatic blocking of pollution paths, activation of backup ventilation systems, and intelligent escape guidance measures.

7. A control system for executing the mine ventilation dynamic control method based on spatiotemporal graph convolutional network according to any one of claims 1 to 6, characterized in that: include: The data acquisition module is used to realize the real-time acquisition and preprocessing of mine ventilation environment parameters, equipment status and spatial topology data. The data acquisition module includes a sensor array unit, a data synchronization and preprocessing unit and a topology calibration unit; The spatiotemporal modeling module, based on the spatiotemporal graph convolutional network, realizes the joint modeling of the spatial topological characteristics and time series characteristics of the ventilation network, and outputs the risk index and abnormal propagation path. The spatiotemporal modeling module includes a spatial feature extraction unit, a time series modeling unit, and a risk assessment and path prediction unit; An intelligent decision-making module generates optimal control based on reinforcement learning, strategically balancing safety, energy consumption, and equipment loss to automate the execution of coordinated equipment control. The intelligent decision-making module includes a reinforcement learning training unit, a policy reasoning and safety assessment unit, and a multi-objective optimization unit. The execution control module receives the instructions from the intelligent decision module, drives the fan and damper equipment to perform the control action, and provides feedback on the execution effect. The execution control module includes an actuator control unit and a feedback data acquisition unit; The system management module provides a human-computer interaction interface, system calibration and model iteration support to ensure the long-term stable operation of the system. The system management module includes a human-computer interaction unit, a calibration and iteration unit, and an emergency response unit.

Images