Mine ventilation dynamic regulation and control method and system based on space-time diagram convolutional network

Abstract

The invention relates to the technical field of mine ventilation, in particular to a mine ventilation dynamic regulation and control method and system based on a space-time diagram convolutional network, and the regulation and control method comprises the following steps: 1, carrying out the modeling and data collection of a ventilation system; step 2, modeling and anomaly prediction of the graph convolutional network; step 3, generation and execution of a ventilation equipment coordinated regulation and control strategy; step 4, edge node calibration and model optimization; and 5, performing full-process control and emergency response. A ventilation network topological structure and multi-modal time series data are fused through a space-time diagram convolutional network, a joint model containing spatial feature extraction and time series modeling is constructed, 30-minute early warning of ventilation abnormity and propagation path prediction are achieved, the lag response limitation that a traditional system only depends on a single parameter threshold value is broken through, and the risk pre-judgment capability is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mine ventilation, and particularly to a mine ventilation dynamic regulation method and system based on a spatio-temporal graph convolutional network. Background Art

[0002] In mine exploitation operations, the mine ventilation system is crucial. It is responsible for delivering sufficient fresh air to the underground working areas and timely exhausting the polluted air, thereby creating a good and safe working environment. Traditional mine ventilation systems usually adopt the large main fan mode design, and the total mine air volume far exceeds the actual required air volume. As mining extends deeper, the situation of multiple working faces operating simultaneously and large-scale mining becomes more and more common, and the underground ventilation network becomes extremely complex, and complex ventilation networks such as series, parallel, and diagonal connections are successively formed.

[0003] Under this background, there are many defects in the existing technologies. On the one hand, the rough ventilation mode is difficult to achieve fine dynamic regulation and cannot meet the different air volume requirements generated by the dynamic changes of factors such as personnel, equipment, the number of working points, process types, working states, and blasting operations in underground working places, resulting in extremely difficult calculation and adjustment of the required air volume. On the other hand, the formation of complex ventilation networks makes the ventilation safety problem more prominent, and the existing ventilation management level is difficult to cope with this complex situation, and the effective air volume rate is low, seriously threatening the physical and mental health of underground workers. Therefore, we propose a mine ventilation dynamic regulation method and system based on a spatio-temporal graph convolutional network. Summary of the Invention

[0004] Aiming at the deficiencies of the existing technologies, the present invention provides a mine ventilation dynamic regulation method and system based on a spatio-temporal graph convolutional network, thereby solving the technical problems mentioned in the background art.

[0005] To achieve the above purposes, the present invention is realized through the following technical solutions: A mine ventilation dynamic regulation method based on a spatio-temporal graph convolutional network, comprising the following steps: Step 1: Ventilation system modeling and data collection, constructing the digital twin foundation of the mine ventilation system, generating a ventilation network digital map including a topological structure, and deploying multi-modal sensor nodes to collect data; Step 2: Graph convolutional network modeling and anomaly prediction, based on the generated digital topological graph and multi-modal time series data, constructing a spatio-temporal graph convolutional network to extract the spatio-temporal correlation features of the ventilation system, and performing ventilation anomaly assessment and propagation path prediction; Step 3: Generation and execution of ventilation equipment collaborative regulation strategies, based on the risk index and propagation path output by the spatio-temporal feature model, using a reinforcement learning algorithm to construct a decision model to realize the automatic regulation of the fan speed and damper opening; Step 4: Edge node calibration and model optimization, design a multi-level calibration and iteration mechanism to ensure the accuracy of the system's long-term operation; Step 5: Full-process control and emergency response, the system operates according to the closed-loop logic of "monitoring - prediction - decision - execution - feedback", and is built-in with a four-level emergency response mechanism.

[0006] In a possible implementation, in the ventilation system modeling and data acquisition step: S1: Use lidar, inertial navigation data, and BIM technology to construct a roadway point cloud model and convert it into a ventilation network digital map; S2: Abstract the roadway intersection, fan, and air door physical entities into a directed weighted graph, and calculate the wind resistance matrix and the maximum ventilation volume matrix; S3: Deploy multi-modal sensor nodes at specific intervals for different regions, collect data such as gas concentration, wind speed, and vibration, and transmit it to the edge controller through the industrial ring network; S4: Use the IEEE1588 precise time protocol to achieve clock synchronization, preprocess the data, including outlier filtering and normalization, to form a multi-dimensional time series dataset.

[0007] In a possible implementation, in the graph convolutional network modeling and anomaly prediction step: S1: Construct a spatio-temporal 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 across regions; S3: The time series layer uses a bidirectional LSTM and an attention mechanism to capture temporal dependencies and highlight key temporal features; S4: Quantify the ventilation state through a risk assessment model, trigger an early warning, and predict the abnormal propagation path.

[0008] In a possible implementation, in the ventilation equipment collaborative control strategy generation and execution step: S1: Use the proximal policy optimization (PPO) algorithm to construct a decision-making model, and define the state space, action space, and reward function; S2: Through simulation training, enable the intelligent agent to learn a control strategy that complies with mining specifications; S3: After the decision-making module outputs the optimal action, it is transmitted to the execution layer through the industrial bus to achieve control of the fan speed and air door opening.

[0009] In a possible implementation, in the edge node calibration and model optimization step: S1: Implement online calibration of sensors and dynamic correction of the model to ensure that the topological model is consistent with the actual roadway structure; S2: Iteratively optimize the model using online learning and adversarial training; S3: Organize expert reviews quarterly and adjust the model parameters to match the latest safety specifications.

[0010] In a possible implementation, in the full - process control and emergency response steps: S1: The system operation follows the closed - loop logic of "monitoring - prediction - decision - execution - feedback"; S2: Trigger corresponding four - level emergency response mechanisms for different risk levels, including automatically blocking the pollution path, starting the standby ventilation system, and intelligent escape guidance measures.

[0011] In a possible implementation, the mine ventilation dynamic regulation system based on the spatio - temporal graph convolutional network includes: A data acquisition module, which is used to realize the real - time acquisition and pre - processing of mine ventilation environment parameters, equipment status, and spatial topology data. The data acquisition module includes a sensor array unit, a data synchronization and pre - processing unit, and a topology calibration unit; A spatio - temporal modeling module, based on the spatio - temporal graph convolutional network, realizes the joint modeling of the spatial topology features and time - series features of the ventilation network, and outputs a risk index and an abnormal propagation path. The spatio - temporal 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, which generates an optimal regulation based on reinforcement learning, balances safety, energy consumption, and equipment loss, and realizes the automation of equipment collaborative regulation execution. The intelligent decision - making module includes a reinforcement learning training unit, a policy reasoning and safety standard unit, and a multi - objective optimization unit; An execution control module, which receives the instructions from the intelligent decision - making module, drives devices such as fans and dampers to execute regulation actions, and feeds back the execution effect. The execution control module includes an actuator control unit and a feedback data acquisition unit; A system management module, which provides a human - machine 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 - machine interaction unit, a calibration and iteration unit, and an emergency response unit.

[0012] Beneficial effects compared with the prior art: 1. In this solution, by fusing the ventilation network topology structure and multi - modal time - series data through the spatio - temporal graph convolutional network, a joint model including spatial feature extraction and time - series modeling is constructed, realizing a 30 - minute early warning of ventilation anomalies and prediction of the propagation path, breaking through the lag response limitation of traditional systems that only rely on a single parameter threshold, and improving the risk prediction ability; 2. In this solution, the proximal policy optimization (PPO) reinforcement learning algorithm is used to construct a decision-making model. A multi-dimensional state space including a risk index, air volume, and equipment status, and a discrete action space for speed / opening adjustment are defined. The reward function is used to balance safety, energy consumption, and equipment loss. After 30,000 times of simulation training, a regulation strategy that complies with mining specifications is generated. The regulation execution integrates Bayesian optimization and fuzzy logic algorithms, achieving a three-fold increase in the strategy generation speed and a 22% reduction in energy consumption. 3. In this solution, a three-level calibration and iteration system for online sensor calibration, model dynamic correction, and online learning is designed. Combined with quarterly expert reviews, it ensures the consistency between the topological model and the physical roadway, data reliability, and anti-interference ability during the long-term operation of the system, and can still provide practical and effective decision-making solutions in the presence of data noise or equipment failures. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly and implement it in accordance with the content of the specification, the following describes the preferred embodiments of the present invention in detail in conjunction with the accompanying drawings.

[0014] Figure 1 It is a schematic flow chart of the mine ventilation dynamic regulation method of the present invention; Figure 2 It is a schematic framework diagram of the mine ventilation dynamic regulation system of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0015] The 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 different forms. Therefore, the present invention is not limited to the embodiments described below. In addition, in order to describe the present invention more clearly, components not connected to the invention will be omitted from the drawings; The technical solutions in the embodiments of the present application are to solve the problems in the above-mentioned background technology as follows: Embodiment 1: A mine ventilation dynamic regulation method based on a spatio-temporal graph convolutional network, including the following steps: Step 1: Ventilation system modeling and data collection First, it is necessary to construct the digital twin foundation of the mine ventilation system. The roadway is scanned three-dimensionally at high density by lidar (the sampling point spacing is 1 m, and the accuracy is 2 cm), and combined with inertial navigation data to generate a roadway point cloud model with coordinates . Using BIM technology, it is transformed into a ventilation network digital map including topological structures . Based on graph theory, physical entities such as roadway intersections, fans, and air doors are abstracted into a directed weighted graph , where the nodes include key positions. The key positions are specifically the main fan , the intersection point of the return airway and so on, along the representing the ventilation path, whose air resistance is calculated by the Darcy - Weisbach formula: , where take the corresponding value of the roadway roughness (such as 0.018 for shotcrete roadway), is the measured path length, and the equivalent diameter of non - circular roadway , where is the cross - sectional area of the roadway. During the non - production period every month, use UAV LiDAR scanning to monitor the roadway deformation. When the path length change rate is reached, according to correct the topology and recalculate the air resistance matrix and the maximum ventilation volume matrix to ensure the dynamic consistency between the digital model and the physical system.

[0016] In the data acquisition link, for areas where gas is prone to accumulate, specifically, the working face of mining and excavation, the entrance of blind roadway, deploy multi - modal sensor nodes at a spacing of 8m. Each node integrates: A gas concentration sensor, which is an infrared absorption type, with a detection range of 0 - 4%LEL and an accuracy of 0.01%, and a response time of <1s; A hot - film anemometer, with a measurement range of 0 - 30m / s and an accuracy of 0.1m / s and a sampling frequency of 10Hz; A three - axis vibration accelerometer, with a range of 10g and a sampling frequency of 1kHz, used to monitor the status of the fan bearing.

[0017] In the conventional area, specifically, in the transportation roadway and the bottom of the shaft yard, deploy temperature and humidity sensors at a spacing of 25m, with an accuracy of 0.5℃ / 2%RH, and monitor the ambient temperature, humidity and fan power consumption through a current transformer. All sensors are connected to the edge controller through the underground industrial ring network and achieve clock synchronization using the IEEE1588 precise time protocol. The data pre - processing process includes: 1. Outlier filtering: Through a sliding window with a window size of 300s, calculate the mean value and the standard deviation , eliminate the noise points where , and fill them with linear interpolation: , where, when , the data is used as . In addition, the data is processed .

[0018] 2. Normalization processing: Map the multi - source data to Interval, the formula is: , and finally form a multi-dimensional time series dataset containing , , , , and other parameters , where is the total number of sensors.

[0019] By constructing a high-precision digital map and a sensor network, the computability of the spatial structure of the ventilation system and the real-time perception of operating parameters are realized.

[0020] Step 2: Graph Convolutional Network Modeling and Anomaly Prediction Based on the digital topological map and multi-modal time series data generated in Step 1, a spatio-temporal graph convolutional network (ST-GCN) is constructed to extract the spatio-temporal correlation features of the ventilation system. The network is divided into two modules: spatial feature extraction and time series modeling: 1. Spatial Feature Layer: Graph Convolutional Network GCN The first layer of GCN learns the ventilation coupling relationship of the direct neighborhood of the nodes. The input feature matrix fuses the node attributes, specifically, the cross-sectional area of the roadway , the designed air volume and the real-time data of the sensors. Through adding a self-loop adjacency matrix and the degree matrix for normalized propagation: , where represents the ventilation correlation between the node and its direct neighbors, specifically, the direct influence intensity of the increase in the gas concentration at a certain heading face on the adjacent return airway.

[0021] The second layer of GCN captures the cross-regional ventilation influence through the second-order adjacency matrix , and the propagation formula is: , and this layer can learn the indirect influence of the main fan speed change on the entire return air network, specifically, the wind speed change trend in the roadway 200 meters away after the main fan air volume is adjusted.

[0022] 2. Time Series Layer: Bidirectional LSTM and Attention Mechanism The spatial feature sequence output by GCN, where , corresponding to 10-minute data, is input into the bidirectional LSTM network to capture the temporal dependencies of both the past and the future simultaneously. The forward hidden state and the backward hidden state are concatenated into , and the feature weights at each moment are calculated through the attention mechanism: , , where, is a trainable weight vector, and the finally output highlights key time series features such as sudden changes in gas concentration and abnormal fluctuations in wind speed.

[0023] 3. Ventilation Abnormality Assessment and Propagation Path Prediction Based on the comprehensive spatio-temporal feature output, the ventilation state is quantified through a risk assessment model: , where is the regional standard wind speed, which is specifically determined by the "Coal Mine Safety Regulations", is the ideal temperature. When is triggered, an early warning is issued, and the prediction of the abnormal propagation path is started. The improved Dijkstra algorithm is used to search for high-risk paths, and the cost function comprehensively considers wind resistance, path length, and real-time gas concentration: , where , the path probability distribution is generated through 1000 Monte Carlo simulations, and the top 3 paths and their propagation probabilities are output, specifically "coal mining face → return air rise → main return airway", with a probability of 78%.

[0024] Through spatio-temporal joint modeling, an early warning of ventilation abnormality 30 minutes in advance with an accuracy rate of 89% and the prediction of the propagation path are realized. Compared with the lag response of traditional systems that only rely on a single parameter threshold, the risk prediction ability is significantly improved.

[0025] Step 3: Generation and Execution of Ventilation Equipment Coordination Control Strategy Based on the risk index output by the spatio-temporal feature model and the propagation path

[0026] 1. Learning Framework Design State space: , which includes global risk, total air volume, equipment status, and spatio-temporal features, with a dimension , where is the number of fans, is the number of air doors; Action space: , where , ; Reward function: , where is the amount of risk reduction, is the amount of reduction in fan energy consumption; is the change in the equipment loss index, based on the vibration intensity calculated.

[0027] Through 30,000 times of simulation training (simulating scenarios such as gas overrun and fan failure), the intelligent agent learns the control strategies that comply with mining regulations. Specifically, when the gas concentration in a certain area rises to 1.2% LEL, the air inlet fan in this area is preferentially opened wider to 1450 rpm, and at the same time, the air return damper is closed by 20°.

[0028] 2. Regulation Execution and Equipment Control After the decision-making module outputs the optimal action , it is transmitted to the execution layer through the industrial bus (Profibus-DP): Fan speed control: The Bayesian optimization algorithm is used to solve the optimal speed combination , and on the premise of meeting and , the energy consumption function is minimized . The speed closed-loop control is realized through a PID controller: , where , the parameters are tuned to by the Ziegler-Nichols method, , , ensuring that the overshoot of speed regulation is <5% and the regulation time is <10 s.

[0029] Damper opening control: Based on the fuzzy logic algorithm, the risk deviation and the change rate are input, and the opening adjustment amount is output. Specifically, when the error and , the damper is closed by 15° to block the pollution diffusion path, and the action execution time is <8 s and is protected by mechanical limit.

[0030] The intelligent generation of regulation strategies is realized through reinforcement learning. Compared with the traditional manual experience regulation, the strategy generation speed is increased by 3 times, and the energy consumption is reduced by 22%.

[0031] Step 4: Edge Node Calibration and Model Optimization To ensure the accuracy of the long-term operation of the system, a multi-level calibration and iteration mechanism is designed: 1. Sensor and Model Calibration Sensor on-line calibration: The self-calibration program is triggered at 3 am every day. The gas sensor closes the gas path to measure the environmental background value , and the data correction formula is: , when LEL, it is determined that the sensor is faulty and an alarm is issued. The wind speed sensor is calibrated at multiple points through a standard wind tunnel (accuracy 0.05 m / s) every quarter, and the linear equation is fitted , ensure that the measurement error is less than 1.5%.

[0032] Model dynamic correction: Using the underground UWB personnel positioning system (accuracy < 15 cm), collect the coordinates of key points in the roadway every week , when the deviation from the digital map exceeds 20 cm, update the node coordinates by the least squares method: Recalculate the air resistance after the update and the maximum flow rate of the ventilation path , ensure that the topological model is consistent with the actual roadway structure.

[0033] 2. Model iterative optimization Online learning: Use the data of the previous 24 hours (about 17,000 samples) for incremental training every day, adopt the momentum SGD optimizer, where the learning rate , momentum 0.9, update the model parameters, and focus on optimizing the prediction accuracy of abnormal scenarios (such as sudden increase in gas after blasting).

[0034] Adversarial training: Artificially inject simulated abnormal data, specifically, the wind speed suddenly drops by 40% and the sensor signal is interrupted. Through the adversarial loss function Improve the robustness of the model to ensure that reliable decisions can still be output in the presence of data noise or equipment failures.

[0035] Expert review: Organize a joint review of ventilation experts and data scientists every quarter. Based on the Shapley value, analyze the contribution of each feature to the decision-making. Specifically, it is found that the weight of the gas concentration can be adjusted from 0.5 to 0.6 to strengthen the safety priority strategy, and the parameters of the reward function are synchronously adjusted to match the latest safety specifications.

[0036] Step Five: Full-process control and emergency response The system operation follows the closed-loop logic of "monitoring - prediction - decision - execution - feedback", and the specific process is as follows: Initialization phase: After the edge controller is powered on, load the latest digital map and the trained ST-GCN+PPO model, send calibration instructions to all sensors, and synchronize the daily production plan with the ground dispatching center at the same time. Specifically, the advancing speed of the coal mining face, the blasting time window, etc.

[0037] Real-time monitoring phase: The sensor collects data at a frequency of 10 Hz, and after preprocessing, inputs it into the ST-GCN model, and outputs the risk index every 5 minutes and the probability of the abnormal propagation path . If , the system maintains the current ventilation state; if , trigger the reinforcement learning decision-making module.

[0038] Regulation Execution Phase: The PPO model generates the optimal action within 50 ms , after safety verification, specifically after checking the balance between air supply and demand, it is sent to the fan and damper actuators through redundant communication links (fiber optic + wireless Mesh), requiring the instruction to be parsed within 0.3 seconds and the equipment to act within 10 seconds.

[0039] Effect Feedback Phase: After the execution action, the sensor continuously collects data. If within 5 minutes decreases , then this regulation case is stored in the experience pool ; if not up to standard, trigger a secondary decision, adjust the strategy and execute again.

[0040] For extreme abnormal scenarios, the system has a built-in four-level emergency response mechanism: Level 1 Response (gas concentration > 1.5% LEL or risk index ): Automatically block the pollution path: Through the high-risk propagation path predicted by the ST-GCN model, immediately close the dampers within 3 levels upstream. Specifically, close , , cut off the abnormal air flow diffusion channel, and the blocking efficiency needs to be completed within 15 seconds; Start the standby ventilation system: Remotely activate the standby fan to full speed operation (2000 rpm), through the formula , where is the fan air volume coefficient, quickly increase the total air volume to 1.8 times the design value; Intelligent escape guidance: Combine UWB personnel positioning data and issue customized evacuation instructions through the underground broadcast system. Specifically, "Personnel in Mining Area 3, please evacuate along the return air roadway → East Wing Auxiliary Shaft", and at the same time, dynamically display the shortest disaster avoidance route on the roadway LED screen.

[0041] Level 2 Response (sudden equipment failure, fan vibration intensity > 5 m / s 2 ): Fault location and isolation: Through the vibration accelerometer data and the fault knowledge graph, which contains 137 fault modes, locate the faulty equipment within 2 seconds. Specifically, bearing wear, automatically switch to the standby fan and turn off the power of the faulty equipment; Life prediction and maintenance scheduling: Based on the LSTM life prediction model , where m / s 2 is the threshold. If the remaining life < 14 days, automatically generate a maintenance work order and push it to the maintenance team, and at the same time adjust the ventilation network topology to bypass the faulty equipment.

[0042] Level 3 response (abrupt change in air resistance caused by roadway deformation > 20%): Topological automatic reconstruction: Detect roadway deformation through UWB positioning data. Specifically, When the roadway length increases by 30m, trigger the edge controller to update the ventilation network model in real time, recalculate the air resistance and the air volume distribution at each node; Dynamic air pressure balance: Use reinforcement learning to re-optimize the fan speed and damper opening to ensure that the air volume fluctuation in the affected area is < 10%. Specifically, increase the fan speed in the intake airway by 100 rpm, and at the same time open the middle regulating damper by 15° to balance the air pressure.

[0043] Level 4 response (system communication interruption): Edge autonomous mode: The edge controller switches to the offline mode and continues to run based on the locally stored historical model and the data in the last 1 hour, and tries to reconnect to the central system every 10 minutes; Preset strategy execution: If communication is not restored within 30 minutes, automatically enable the preset conservative control strategy. Specifically, maintain the fan speed of all fans at 80% of the rated value and the damper opening at 50° until communication is restored or manual intervention is carried out.

[0044] To ensure the effective monitoring and intervention of the system by operators, a multi-level human-computer interaction interface is designed: Ground command center: Real-time display the ventilation network status through a three-dimensional digital twin large screen, integrate the risk heat map, equipment operation parameters, and abnormal warning information. Specifically, high-risk areas are marked with red flashing, and experts are supported to manually adjust the control strategy by dragging the fan / damper icon, and the system automatically calculates the change in the risk index after manual intervention; Underground mobile terminal: Equip the inspection personnel with explosion-proof tablet computers to real-time display the ventilation parameters within 50m around the current location, and support one-key triggering of local ventilation optimization. Specifically, temporarily increase the air volume at the tunneling face, and the command is sent to the edge controller through the wireless Mesh network within 2 seconds; Alarm and log system: Adopt a hierarchical alarm mechanism (yellow - warning, red - emergency), and synchronously push abnormal information through audible and visual alarms and text messages (to the ground person in charge, safety director). The historical operation logs and control records are stored in the blockchain system to ensure traceability and clear responsibilities.

[0045] In addition, the following verification process is also carried out: Laboratory verification: Deploy a 1:100 scale model in a simulated mine environment, and verify the prediction accuracy of the ST-GCN model for the gas diffusion path by artificially injecting gas (concentration 0 - 2%LEL) and adjusting the fan speed (500 - 1500 rpm). Specifically, the prediction accuracy is 87%, and the energy consumption optimization rate of the PPO strategy reaches 20.1%; Underground pilot test: A three-month trial operation was carried out in a working face of Yangquan Coal Industry to complete the following verifications: Sensor network stability: Continuous operation for 90 days without node failures, and the data integrity rate is 99.8%; Effectiveness of the regulation strategy: Successfully handled 12 cases of gas concentration exceeding the limit (0.85%-1.2%LEL), with an average handling time of 32 seconds, which is 70% shorter than manual regulation; Edge computing performance: Response delay is 0.28 seconds.

[0046] Example 2: A mine ventilation dynamic regulation system based on a spatio-temporal graph convolutional network, including the following modules: I. Data acquisition module Realize the real-time acquisition and preprocessing of mine ventilation environment parameters, equipment status and spatial topology data, and provide multi-dimensional data sources for subsequent modeling and decision-making. It includes a sensor array unit, a data synchronization and preprocessing unit, and a topology calibration unit.

[0047] 1. Sensor array unit This unit is deployed in key areas of the mine, and differential density deployment is implemented according to different risk levels. In high-risk areas where gas is prone to accumulate, specifically in the excavation and working faces and the return air roadway, Multi-modal sensor nodes are densely deployed at an 8-meter interval. Each node integrates an infrared absorption type gas concentration sensor with a detection range of 0-4%LEL and an accuracy of 0.01%, a response time of <1 second, a hot film type wind speed sensor with a measurement range of 0-30m / s and an accuracy of 0.1m / s, a sampling frequency of 10Hz, and a three-axis vibration accelerometer with a range of 10g and a sampling frequency of 1kHz, which are used to capture key parameters such as gas concentration, wind speed fluctuations, and fan bearing vibrations in real time. In conventional areas such as the transportation roadway and the bottom of the shaft, temperature and humidity sensors are deployed at a 25-meter interval, with an accuracy of 0.5℃ / 2%RH, and the ambient temperature, humidity, and fan power consumption are monitored 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 、damper opening and other parameter time series data streams, with Unix timestamps attached to ensure time consistency.

[0048] 2. Data synchronization and preprocessing unit This unit is based on the IEEE 1588 time protocol to achieve nanosecond-level clock synchronization between distributed sensor nodes and edge controllers, with the synchronization error controlled within 1 ms to ensure the spatio-temporal alignment of multi-source data. In the data preprocessing section, the sliding window technique is adopted. The window size is the same as that in the first embodiment, and the mean value of the data within the sliding window is calculated and the standard deviation . Based on the principle, outliers are removed, and the valid data is normalized.

[0049] 3. Topology Calibration Unit This unit regularly monitors the change of the roadway spatial structure through UWB positioning base stations and UAV LiDAR scans. During the non-production period every week, the system automatically triggers the roadway deformation detection, compares the current point cloud data with the historical topology model, and calculates the path length change rate . When the condition is met, the topology correction formula and the air resistance update formula in the first embodiment are called to recalculate the air resistance matrix and the maximum ventilation volume matrix of the ventilation network to ensure the dynamic consistency between the digital model and the physical roadway.

[0050] II. Spatio-Temporal Modeling Module Based on the spatio-temporal graph convolutional network ST-GCN, the joint modeling of the spatial topology features and time series features of the ventilation network is realized, and the risk index and the abnormal propagation path are output. It includes a spatial feature extraction unit, a time series modeling unit, and a risk assessment and path prediction unit.

[0051] 1. Spatial Feature Extraction Unit This unit deploys a two-layer graph convolutional network (GCN) on the GPU (NVIDIA Jetson Xavier NX) of the edge controller. The first layer inputs the node attributes, specifically, the cross-sectional area of the roadway , the designed air volume and the preprocessed sensor data. Through the self-loop added adjacency matrix and the degree matrix , the ventilation correlation features of the node neighborhood are extracted using the GCN propagation formula in the first embodiment to characterize the direct impact of the increase in gas concentration on adjacent roadways. The second layer learns the cross-regional ventilation coupling relationship through the second-order adjacency matrix , and the formula is like the second-order GCN formula in the first embodiment to capture the indirect impact of the main fan speed change on the wind speed of the far-end roadway.

[0052] 2. Time Series Modeling Unit This unit inputs the spatial feature sequence output by the GCN into a bidirectional LSTM network, and through the forward hidden state and the backward hidden state The splicing captures the historical change patterns and future trends of ventilation parameters. The attention mechanism is introduced. For example, in the first embodiment, the attention weight formula is used to calculate the feature weights at each moment, highlighting key time series features such as sudden changes in gas concentration and abnormal fluctuations in wind speed, and generating a weighted time series feature vector for subsequent risk assessment.

[0053] 3. Risk Assessment and Path Prediction Unit This unit is based on the spatio-temporal feature vector and the original sensor data. For example, in the first embodiment, the comprehensive risk index is calculated using the risk assessment formula. When it triggers abnormal prediction. The improved Dijkstra algorithm is adopted, combined with Monte Carlo simulation. For example, in the first embodiment, the path cost formula is used to search for high-risk propagation paths and output the top 3 paths and their propagation probabilities.

[0054] III. Intelligent Decision-making Module Based on reinforcement learning, the optimal regulation strategy is generated to balance safety, energy consumption, and equipment loss, and realize the automation of equipment collaborative regulation execution. It includes a reinforcement learning training unit, a policy inference and safety criterion unit, and a multi-objective optimization unit.

[0055] 1. Reinforcement Learning Training Unit This unit constructs a virtual mine ventilation scenario, simulates abnormal scenarios such as gas overrun and fan failure, collects 30,000 training samples, and trains the agent based on the proximal policy optimization (PPO) algorithm. During the training process, the agent collects state-action-reward sequences through interaction with the environment, and uses the PPO loss function formula in the first embodiment to optimize the policy network, learning regulation strategies that conform to mining specifications. The trained model is compressed and deployed on the edge controller.

[0056] 2. Policy Inference and Safety Criterion Unit This unit receives the risk index , total air volume , equipment status and time series features output by the spatio-temporal modeling module in real time, and calls the trained PPO model to generate adjustment instructions for the fan speed and damper opening. The action space is defined as , . The generated policy needs to pass safety verification, including air volume constraints and equipment limits. Specifically, the fan speed is 0 - 1500 rpm, and the damper opening is 10° - 90°, ensuring the feasibility of the regulation plan.

[0057] 3. Multi-objective Optimization Unit This unit further refines the execution parameters of the control strategy. For the optimization of the fan speed, the Bayesian optimization algorithm is used to solve the problem of minimizing energy consumption. On the premise of meeting the risk control requirements, the optimal speed combination is calculated according to the energy consumption optimization formula in Embodiment 1. For the control of the damper opening, the fuzzy logic algorithm is adopted, and the opening is dynamically adjusted according to the risk deviation and the rate of change . Specifically, when the error and , the damper closes by 15° to block the pollution diffusion path.

[0058] IV. Execution Control Module Receives the instructions from the intelligent decision-making module, drives devices such as fans and dampers to execute control actions, and feeds back the execution effect to form a closed-loop control. It includes an actuator control unit and a feedback data acquisition unit.

[0059] 1. Actuator Control Unit This unit analyzes the control instructions and drives the devices to execute. For the control of the fan speed, a PID controller is used to achieve closed-loop regulation. The control quantity is calculated according to the PID control formula in Embodiment 1, where the error , ensuring that the overshoot of the speed regulation is greater than 5% and the regulation time is less than 10 seconds. For the control of the damper opening, the damper is driven to adjust by an electric actuator, and the built-in mechanical limit device ensures that the opening is within the range of 10° - 90°.

[0060] 2. Feedback Data Acquisition Unit This unit, after the device executes the control instructions, real-time collects feedback data such as the actual speed, damper opening, gas concentration after regulation, and risk index, and transmits them back to the intelligent decision-making module through the industrial ring network. The system calculates the actual reward value based on the feedback data, updates the experience pool, and triggers the model micro-adjustment according to the reward function formula in Embodiment 1, forming a closed-loop of "decision - execution - feedback - optimization".

[0061] V. System Management Module Provides a human - machine interaction interface, system calibration, and model iteration support to ensure the long - term stable operation of the system. It includes a human - machine interaction unit, a calibration and iteration unit, and an emergency response unit.

[0062] 1. Human - Machine Interaction Unit This unit includes a ground command center and an underground mobile terminal. The ground command center real - time displays the ventilation network topology, risk heat map, and device status through a three - dimensional digital twin interface, supports experts to manually adjust the fan speed and damper opening through drag - and - drop operations, and the system automatically predicts the change of the risk index after adjustment. The underground mobile terminal is an explosion - proof tablet computer, which real - time displays the ventilation parameters within 50 meters around the patrol personnel and supports one - key triggering of local ventilation optimization instructions. The instructions are sent to the edge controller through the wireless Mesh network within 2 seconds.

[0063] 2. Calibration and Iteration Unit This unit is responsible for sensor calibration and continuous optimization of the model. The zero-point calibration of the sensor is automatically triggered at midnight every day. Specifically, when the background value of the gas sensor reaches the LEL, a fault alarm is triggered; every quarter, the sensitivity of the sensor is calibrated using a standard gas sample and a wind tunnel device to ensure that the measurement error is less than 2%. In terms of model iteration, incremental training is carried out daily using the latest data, and the parameters are updated using the momentum SGD optimizer; every quarter, an expert review is organized, and the weights of the risk assessment formula are adjusted based on the Shapley value analysis to improve the sensitivity of the model to safety risks.

[0064] 3. Emergency Response Unit This unit has a four-level emergency response mechanism built in. In the first-level response, the system automatically closes the upstream air door on the shortest path to the abnormal node and starts the standby fan to run at full speed, and generates a dynamic escape route in combination with UWB personnel positioning; in the second-level response, the faulty equipment is located within 2 seconds and switched to the standby unit; 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 the offline mode and executes the preset conservative strategy until the communication is restored.

[0065] Finally, it should be noted that: Obviously, the above embodiments are only examples for clearly illustrating the present invention, and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or modifications derived therefrom are still within the protection scope of the present invention.

Claims

1. A mine ventilation dynamic regulation method based on a spatio-temporal graph convolutional network, characterized in that, It includes the following steps: Step 1: Ventilation system modeling and data collection, constructing the digital twin foundation of the mine ventilation system, generating a ventilation network digital map containing the topological structure, and deploying multi-modal sensor nodes to collect data; Step 2: Graph convolutional network modeling and anomaly prediction, based on the generated digital topological map and multi-modal time series data, constructing a spatio-temporal graph convolutional network to extract the spatio-temporal correlation features of the ventilation system, and conducting ventilation anomaly assessment and propagation path prediction; Step 3: Generation and execution of collaborative control strategies for ventilation equipment, based on the risk index and propagation path output by the spatio-temporal feature model, using a reinforcement learning algorithm to construct a decision-making model to achieve automatic control of the 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's long-term operation; Step 5: Full-process control and emergency response, the system operates according to the closed-loop logic of "monitoring - prediction - decision-making - execution - feedback", and is built-in with a four-level emergency response mechanism.

2. The method for dynamically regulating mine ventilation based on a spatio-temporal graph convolutional network according to claim 1, characterized in that In the ventilation system modeling and data collection step: S1: Use lidar, inertial navigation data, and BIM technology to construct a roadway point cloud model and convert it into a ventilation network digital map; S2: Abstract the physical entities of roadway intersections, fans, and dampers into a weighted graph, and calculate the air resistance matrix and maximum ventilation volume matrix; S3: Deploy multi-modal sensor nodes at specific intervals for different regions to collect data such as gas concentration, wind speed, and vibration, and transmit the data to the edge controller through the industrial ring network; S4: Use the IEEE1588 precise time protocol to achieve clock synchronization, preprocess the data, including outlier filtering and normalization, to form a multi-dimensional time series dataset.

3. The mine ventilation dynamic regulation method based on the spatio-temporal graph convolutional network according to claim 1, characterized in that In the graph convolutional network modeling and anomaly prediction step: S1: Construct a spatio-temporal 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 regions; S3: The time series layer uses a bidirectional LSTM and an attention mechanism to capture time series dependencies and highlight key time series features; S4: Quantify the ventilation state through a risk assessment model, trigger early warnings, and predict the abnormal propagation path.

4. The method for dynamically regulating mine ventilation based on a spatio-temporal graph convolutional network according to claim 1, wherein In the generation and execution of collaborative control strategies for ventilation equipment step: S1: Use the proximal policy optimization (PPO) algorithm to construct a decision-making model, defining the state space, action space, and reward function; S2: Through simulation training, enable the intelligent agent to learn control strategies that comply with mining specifications; S3: After the decision-making module outputs the optimal action, transmit it to the execution layer through the industrial bus to achieve control of the fan speed and damper opening.

5. The mine ventilation dynamic regulation method based on a spatio-temporal graph convolutional network according to claim 1, wherein, In the edge node calibration and model optimization step: S1: Implement online calibration of sensors and dynamic correction of the model to ensure that the topological model is consistent with the actual roadway structure; S2: Use online learning and adversarial training to iteratively optimize the model; S3: Organize expert reviews every quarter to adjust the model parameters to match the latest safety specifications.

6. The method for dynamically regulating mine ventilation based on a spatio-temporal graph convolutional network according to claim 1, wherein In the full-process control and emergency response step: S1: The system operates according to the closed-loop logic of "monitoring - prediction - decision-making - execution - feedback"; S2: Trigger the corresponding four-level emergency response mechanism according to different risk levels, including automatically blocking the pollution path, starting the standby ventilation system, and intelligent escape guidance measures.

7. A regulation system for implementing the mine ventilation dynamic regulation method based on a spatio-temporal graph convolutional network according to claims 1 to 6 above, characterized in that, Including: 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 spatio-temporal modeling module, based on the spatio-temporal graph convolutional network, realizes the joint modeling of the spatial topology features and time series features of the ventilation network, and outputs the risk index and the abnormal propagation path. The spatio-temporal modeling module includes a spatial feature extraction unit, a time series modeling unit, and a risk assessment and path prediction unit; The intelligent decision-making module generates the optimal regulation based on reinforcement learning, balances safety, energy consumption and equipment loss, and realizes the automation of equipment collaborative regulation execution. The intelligent decision-making module includes a reinforcement learning training unit, a policy reasoning and safety standard unit, and a multi-objective optimization unit; The execution control module receives the instructions of the intelligent decision-making module, drives devices such as fans and dampers to execute regulation actions, and feeds back 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