Coal mine production ventilation optimization control method and system based on adaptive learning

Abstract

This invention discloses a method and system for optimizing ventilation control in coal mines based on adaptive learning. The method involves real-time acquisition of initial multi-source heterogeneous data; construction of a spatiotemporal prediction model, inputting the initial multi-source heterogeneous data into the model to output the mine's airflow field distribution and gas concentration field distribution; designing a strategy optimizer, using the airflow field distribution and gas concentration field distribution as input, and safety, energy consumption, and airflow stability as reward functions, to output action commands for coordinated control of multiple ventilation devices; and online adaptive updates of the spatiotemporal prediction model and strategy optimizer based on the actual ventilation effect data after the action commands are executed. This improves the utilization efficiency and accuracy of the coal mine ventilation optimization control system.

Description

Technical Field

[0001] This invention relates to the field of coal mining technology, and in particular to a method and system for optimizing ventilation control in coal mining based on adaptive learning. Background Technology

[0002] Coal mining environments are complex and highly variable. Ventilation systems are core facilities for ensuring underground operational safety and reducing the risk of gas accumulation. Their control effectiveness directly affects the safety of workers and mining efficiency. Currently, coal mine ventilation control relies heavily on traditional manual experience or semi-automated systems, which are prone to delays due to information fragmentation. Existing prediction models struggle to adapt to non-Euclidean spatial relationships and long- and short-term temporal dynamics underground, resulting in insufficient accuracy in predicting airflow and gas concentration fields. Control strategies often focus on single objectives, easily leading to imbalances between safety, energy consumption, and stability. Furthermore, they lack adaptive updating capabilities and struggle to cope with fluctuations in ventilation conditions caused by geological changes and mining progress, resulting in low efficiency of ventilation optimization control in coal mines. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method and system for optimizing ventilation control in coal mines based on adaptive learning.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned adaptive learning-based ventilation optimization and control method for coal mines: Real-time acquisition of underground multi-source heterogeneous data during coal mining process; preprocessing of underground multi-source heterogeneous data to obtain initial multi-source heterogeneous data. A spatiotemporal prediction model of GCN-STAN is constructed. The non-Euclidean dependencies between nodes are learned by using a 2-layer AGCN adaptive graph convolution, the long-term and short-term dynamics are captured by bidirectional BI-LSTM, and the features are focused by a dual-channel attention mechanism. The initial multi-source heterogeneous data is input into the GCN-STAN spatiotemporal prediction model, which outputs the airflow field distribution and gas concentration field distribution of the mine. Design a DDPG-Proximal strategy optimizer, taking airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption and airflow stability as reward functions, and outputting action commands for coordinated control of multiple ventilation devices; Based on the actual ventilation effect data after the execution of the action command, the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer are updated online adaptively.

[0005] Furthermore, in the aforementioned adaptive learning-based ventilation optimization and control method for coal mine mining, the real-time acquisition of underground multi-source heterogeneous data during the coal mine mining process, and the preprocessing of the underground multi-source heterogeneous data to obtain initial multi-source heterogeneous data, include: Real-time acquisition of multi-source heterogeneous data from underground coal mining processes, including at least sensor data, image data, ventilation equipment data, and geological dynamic data; The extreme data caused by sensor failure and equipment interference is removed by using the 3σ criterion. The image data is denoised by using Gaussian filtering to remove noise caused by downhole dust and light changes. Morphological operations are used to repair image edge defects. Missing data is filled by linear interpolation. Based on timestamps, data from different acquisition frequencies are uniformly aligned to a 1Hz time scale. A weighted fusion algorithm is used to integrate heterogeneous data from the same source. Image data is transformed into numerical data by extracting feature vectors through CNN. It is then uniformly mapped to the same feature space with sensor data, equipment data, and geological data to generate a multi-source data matrix with consistent dimensions. The multi-source data matrix is ​​subjected to Min-Max normalization to map parameters of different dimensions to the [0,1] interval. At the same time, Z-Score normalization is used to eliminate data distribution bias and obtain the initial multi-source heterogeneous data.

[0006] Furthermore, in the aforementioned adaptive learning-based ventilation optimization and control method for coal mine operations, the construction of the GCN-STAN spatiotemporal prediction model utilizes a two-layer AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, captures long-term and short-term dynamics through bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism, including: The first layer AGCN is used for local spatial feature extraction. Based on the underground roadway topology, a dynamic graph adjacency matrix is ​​constructed. The matrix elements represent the ventilation correlation strength between different spatial nodes. The local spatial dependency relationship between nodes and neighboring nodes is learned through adaptive graph convolution kernels. The spatial distribution features of airflow and gas in the roadway are extracted, and the feature map is output. The second layer AGCN is used for global spatial feature fusion. Based on the feature map output by the first layer, the adjacency matrix is ​​dynamically updated. By expanding the receptive field, it learns global spatial dependencies across regions and outputs spatiotemporal fusion features.

[0007] Furthermore, in the aforementioned adaptive learning-based ventilation optimization and control method for coal mine operations, the construction of the GCN-STAN spatiotemporal prediction model utilizes a two-layer AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, captures long-term and short-term dynamics through bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism, including: The spatiotemporal fusion features output by AGCN are expanded according to the time series and input into the bidirectional BI-LSTM module. The forward LSTM captures the temporal dynamic features from the past to the present, and the backward LSTM captures the temporal correlation features from the present to the future. By concatenating the bidirectional output results, the time feature vector is obtained.

[0008] Furthermore, in the aforementioned adaptive learning-based ventilation optimization and control method for coal mine operations, the construction of the GCN-STAN spatiotemporal prediction model utilizes a two-layer AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, captures long-term and short-term dynamics through bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism, including: By utilizing spatial attention channels to calculate attention weights, different weights are assigned to spatial features, focusing on the spatial features of key areas; Based on the feature attention channel, the fully connected layer is used to calculate the feature attention weight, which enhances the contribution of the feature, suppresses the interference of redundant features, and outputs the spatiotemporal fusion feature of the target after attention focusing.

[0009] Furthermore, in the aforementioned adaptive learning-based ventilation optimization and control method for coal mine operations, the DDPG-Proximal strategy optimizer is designed, taking the airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption, and airflow stability as reward functions, to output action commands for coordinated control of multiple ventilation devices, including: A dual-network architecture of Actor-Critic is used to construct the DDPG-Proximal policy optimizer. The Actor network is responsible for generating coordinated control action commands for ventilation equipment, while the Critic network is responsible for evaluating the value of the action commands and optimizing the policy. To prevent system instability caused by excessive policy updates, a Proximal policy constraint is introduced. The network learning rate is set to 0.001 and the discount factor is set to 0.95.

[0010] Furthermore, in the aforementioned adaptive learning-based ventilation optimization and control method for coal mine operations, the step of online adaptively updating the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer based on the actual ventilation effect data after the execution of action commands includes: Using the error between the predicted and actual values ​​as the loss function, the parameters of the GCN-STAN spatiotemporal prediction model are adjusted online using the SGD stochastic gradient descent algorithm. The actual value of the reward function is calculated based on actual ventilation effect data. It is then compared with the value function predicted by the Critic network to update the Critic network parameters and optimize the value assessment accuracy. At the same time, the action generation strategy of the Actor network is adjusted in conjunction with the Proximal policy constraint so that the KL divergence between the new strategy and the old strategy is ≤0.01.

[0011] Furthermore, an adaptive learning-based ventilation optimization and control system for coal mines is adopted, which includes the following modules: The multi-source data acquisition module is used to acquire multi-source heterogeneous data in the coal mining process in real time, and to preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. The prediction model building module is used to construct the GCN-STAN spatiotemporal prediction model. It uses a 2-layer AGCN adaptive graph convolution to learn the non-Euclidean dependencies between nodes, captures long-term and short-term dynamics through bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism. The coal mine condition prediction module is used to input the initial multi-source heterogeneous data into the GCN-STAN spatiotemporal prediction model and output the airflow field distribution and gas concentration field distribution of the mine. The coal mine real-time control module is used to design the DDPG-Proximal strategy optimizer. It takes the airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption and airflow stability as reward functions, and outputs action commands for coordinated control of multiple ventilation devices. The adaptive learning module is used to perform online adaptive updates to the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer based on the actual ventilation effect data after the execution of the action command.

[0012] Furthermore, in the adaptive learning-based coal mine ventilation optimization and control system, the real-time coal mine control module includes the following sub-modules: A submodule is constructed to build the DDPG-Proximal policy optimizer using an Actor-Critic dual-network architecture. The Actor network is responsible for generating coordinated control action instructions for ventilation equipment, while the Critic network is responsible for evaluating the value of the action instructions and optimizing the policy. A submodule is introduced to introduce Proximal policy constraints to prevent the system from becoming unstable due to excessive policy updates. The network learning rate is set to 0.001 and the discount factor is set to 0.95.

[0013] Furthermore, in the adaptive learning-based coal mine ventilation optimization and control system, the real-time coal mine control module includes the following sub-modules: The adjustment submodule is used to adjust the parameters of the GCN-STAN spatiotemporal prediction model online using the SGD stochastic gradient descent algorithm, with the error between the predicted and actual values ​​as the loss function. The update submodule is used to calculate the actual value of the reward function based on the actual ventilation effect data, compare it with the value function predicted by the Critic network, update the Critic network parameters, optimize the value assessment accuracy, and adjust the action generation strategy of the Actor network in combination with the Proximal policy constraint so that the KL divergence between the new strategy and the old strategy is ≤0.01.

[0014] Its beneficial effects are as follows: First, it eliminates heterogeneous data interference through standardized preprocessing, providing high-quality data support for subsequent prediction and control, and breaking down information silos. Second, it improves the accuracy of spatiotemporal prediction. The GCN-STAN model, combined with adaptive graph convolution and bidirectional LSTM, accurately captures the spatial correlation and temporal dynamics of the ventilation system, providing a reliable basis for control decisions. Third, it achieves multi-objective optimal control. The optimizer constructs a reward function based on safety, energy consumption, and stability to achieve coordinated equipment control, balancing safety baselines and operational efficiency. Fourth, it possesses dynamic adaptive capabilities, updating the model and strategy online based on actual results to adapt to changes in underground working conditions and avoid control failure. In summary, this method significantly improves the intelligence and precision of ventilation control, effectively reduces gas risk and energy consumption, and provides technical support for safe and efficient coal mining. It also improves the utilization efficiency and accuracy of ventilation optimization and control systems in coal mines. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the adaptive learning-based ventilation optimization and control method for coal mine operations according to the present invention. Figure 2 This is a schematic diagram of the second embodiment of the adaptive learning-based ventilation optimization and control method for coal mine operations in this invention. Figure 3 This is a schematic diagram of the first embodiment of the adaptive learning-based ventilation optimization and control system for coal mine operations according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, the ventilation optimization and control method for coal mines based on adaptive learning includes the following steps: Step 101: Collect multi-source heterogeneous data from underground mines in real time during the coal mining process, and preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. Specifically, in this embodiment, multi-source heterogeneous data from underground coal mining processes are collected in real time, including at least sensor data, image data, ventilation equipment data, and geological dynamic data. Extreme data caused by sensor malfunctions and equipment interference are removed using the 3σ criterion. Noise reduction is applied to the image data, and Gaussian filtering is used to remove noise caused by downhole dust and light variations. Morphological operations are used to repair image edge defects, and linear interpolation is used to fill in missing data. Based on the timestamp, data from different acquisition frequencies are aligned to a 1Hz time scale. A weighted fusion algorithm is used to integrate homogeneous and heterogeneous data. Feature vectors are extracted from the image data using CNN and transformed into numerical data. These are then mapped to the same feature space along with sensor data, equipment data, and geological data to generate a multi-source data matrix with consistent dimensions. Min-Max normalization is applied to the multi-source data matrix to map parameters of different dimensions to the [0,1] interval. Z-Score standardization is used to eliminate data distribution bias, resulting in the initial multi-source heterogeneous data.

[0020] Sensor data acquisition involves deploying wind speed sensors, gas concentration sensors, temperature sensors, humidity sensors, and pressure sensors in key areas such as underground mining faces, return airways, intake airways, and gas extraction points. Utilizing a dual-mode 4G / 5G + industrial Ethernet transmission method, data on wind speed, gas volume fraction, ambient temperature and humidity, and roadway air pressure are collected in real-time at a frequency of 1Hz. The acquisition accuracy meets the requirements of ±0.01% for gas concentration and ±0.1m / s for wind speed, with data transmission latency controlled within 500ms. Data transmission can also resume from where it left off to prevent data loss due to unstable underground signals. Image data acquisition involves installing explosion-proof high-definition cameras at roadway intersections, near ventilation equipment, and in key mining areas. Video stream data is acquired at 30 frames per second, and frame images are extracted synchronously. The focus is on capturing visual information such as the status of air doors, the operating posture of fans, and abnormal obstructions within the roadway, such as coal falls and equipment misalignment. The image resolution is set to 1920×1080, and H.265 encoding is used for compression and storage, balancing transmission efficiency and image clarity. Ventilation Equipment Data Acquisition: Through the built-in communication interface of the equipment controller, using the RS485 / Modbus protocol, real-time operating parameters of equipment such as main ventilators, local ventilators, damper actuators, and air window regulators are acquired. These parameters include fan speed, motor power, damper opening, air window adjustment angle, equipment operating time, and fault codes. The acquisition frequency is synchronized with the equipment control cycle at 0.5Hz, enabling real-time monitoring of equipment operation status. Geological Dynamic Data Acquisition: In conjunction with the coal mine geological monitoring system, real-time data on roadway surrounding rock deformation, coal seam thickness changes, fault movement trajectories, and roof pressure are obtained. Surrounding rock deformation data is acquired through fiber optic grating sensors with an accuracy of ±0.01mm, updated every 5 minutes. Simultaneously, dynamic supplementary data from geological exploration departments is integrated to reflect the impact of real-time changes in underground geological conditions on the ventilation system.

[0021] Step 102: Construct the GCN-STAN spatiotemporal prediction model, use 2 layers of AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, capture long-term and short-term dynamics through bidirectional BI-LSTM, and focus features through a dual-channel attention mechanism. Specifically, in this embodiment, the first layer AGCN is used for local spatial feature extraction. Based on the underground tunnel topology, a dynamic graph adjacency matrix is ​​constructed, and the matrix elements represent the ventilation correlation strength between different spatial nodes. Adaptive graph convolution kernels are used to learn the local spatial dependencies between nodes and their neighbors, extracting the spatial distribution features of airflow and gas within the tunnel, and outputting a feature map. The second layer AGCN is used for global spatial feature fusion. Based on the feature map output from the first layer, the adjacency matrix is ​​dynamically updated. By expanding the receptive field, global spatial dependencies across regions are learned, and spatiotemporal fusion features are output. The spatiotemporal fusion features output by AGCN are expanded according to the time series and input into a bidirectional BI-LSTM module. The forward LSTM captures the temporal dynamic features from the past to the present, and the backward LSTM captures the temporal correlation features from the present to the future. By concatenating the bidirectional output results, a temporal feature vector is obtained. Spatial attention channels are used to calculate attention weights, assigning different weights to spatial features to focus on the spatial features of key areas. Based on the feature attention channels, fully connected layers are used to calculate feature attention weights, strengthening the contribution of features and suppressing the interference of redundant features, outputting the target spatiotemporal fusion features after attention focusing.

[0022] The model adopts a four-layer structure of "input layer - spatiotemporal feature extraction layer - attention focusing layer - output layer". The input layer receives the preprocessed initial multi-source heterogeneous data, and the output layer outputs the predicted results of wind flow field and gas concentration field distribution. The hidden layer dimension of the model is set to 256, the activation function is ReLU, and the dropout coefficient is set to 0.3 to avoid overfitting.

[0023] The first layer, AGCN, extracts local spatial features: Based on the underground roadway topology, a dynamic graph adjacency matrix is ​​constructed. The matrix elements represent the ventilation correlation strength between different spatial nodes and sensor deployment points, such as the influence of roadway length and cross-sectional area on airflow transmission. Through an adaptive graph convolution kernel with a kernel size of 3, the local spatial dependencies between nodes and their neighboring nodes are learned, and the spatial distribution features of airflow and gas in the roadway are extracted. The output dimension is a feature map with the number of samples × the number of spatial nodes × 128.

[0024] The second layer AGCN, global spatial feature fusion: Based on the output features of the first layer, the adjacency matrix is ​​dynamically updated. The influence of geological dynamics on the roadway ventilation relationship is considered. By expanding the receptive field and setting the kernel size to 5, the global spatial dependency relationship across regions is learned, which solves the problem that traditional GCN cannot adapt to the dynamic topology of the mine. It fuses local and global spatial features, and the output dimension is the spatiotemporal fusion feature of sample number × number of spatial nodes × 256.

[0025] The spatial features output by AGCN are expanded according to the time series and input into the bidirectional BI-LSTM module. The number of hidden layer units is set to 128 and the number of layers is 2. The forward LSTM captures the temporal dynamic features from the past to the present, such as the cumulative change trend of gas concentration, while the backward LSTM captures the temporal correlation features from the present to the future, such as the predicted trend of airflow velocity. By splicing the bidirectional output results, a feature vector containing long-term and short-term temporal information is obtained, which effectively solves the problem of strong temporal dependence and rapid dynamic changes in ventilation systems.

[0026] Spatial Attention Channel: Based on the differences in importance of different spatial nodes, such as the mining face node having a greater impact on safety than the return airway node, attention weights are calculated to assign different weights to spatial features, focusing on key areas, such as areas where gas is prone to accumulate and the spatial features of the mining face, thereby improving the model's prediction accuracy for core areas.

[0027] Feature Attention Channel: To address the differences in feature importance among multi-source data, such as gas concentration data being more critical for safety control than humidity data, feature attention weights are calculated through a fully connected layer to enhance effective features, such as the contribution of gas concentration and wind speed, while suppressing redundant features, such as the interference of temperature, humidity, and equipment operating time. The final output is a spatiotemporal fusion feature after attention focusing, providing accurate input for subsequent predictions.

[0028] Step 103: Input the initial multi-source heterogeneous data into the GCN-STAN spatiotemporal prediction model and output the airflow field distribution and gas concentration field distribution of the mine. Specifically, in this embodiment, the data input and model forward propagation involve inputting the preprocessed initial multi-source heterogeneous data into the GCN-STAN model in batches, with a batch size of 32. After being standardized by the input layer, the data passes through the AGCN module, the bidirectional BI-LSTM module, and the dual-channel attention module in sequence. Finally, it is mapped to the output space through a fully connected layer, outputting the predicted values ​​of airflow velocity, airflow direction, and gas concentration for each spatial node in the next 5 minutes.

[0029] The distribution of airflow and gas concentration fields is generated based on the node prediction values ​​output by the model. Combining the underground roadway topology and grid division, the underground area is divided into 1m×1m grid cells. The Kriging interpolation algorithm is used to complete the airflow and gas parameters in non-node areas, generating a continuous airflow field distribution map, including wind speed and direction distribution and gas concentration field distribution maps, including gas concentration values ​​for each grid. At the same time, areas with gas concentration exceeding the threshold (≥0.8%) and areas with abnormal airflow velocity (<0.2m / s or >4m / s) are marked, providing an intuitive basis for control decisions.

[0030] Step 104: Design the DDPG-Proximal strategy optimizer, taking the airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption and airflow stability as reward functions, and outputting action commands for coordinated control of multiple ventilation devices. Specifically, in this embodiment, an Actor-Critic dual-network architecture is used to construct the DDPG-Proximal policy optimizer. The Actor network is responsible for generating coordinated control action commands for ventilation equipment, while the Critic network is responsible for evaluating the value of the action commands and optimizing the policy. Proximal policy constraints are introduced to avoid system instability caused by excessive policy updates. The network learning rate is set to 0.001, and the discount factor is set to 0.95.

[0031] The overall architecture of the optimizer adopts an "Actor-Critic" dual-network architecture to build the DDPG-Proximal policy optimizer. The Actor network is responsible for generating action instructions for the coordinated control of ventilation equipment, while the Critic network is responsible for evaluating the value of the action instructions and optimizing the policy. Proximal policy constraints are introduced to avoid system instability caused by excessive policy updates. The network learning rate is set to 0.001 and the discount factor is set to 0.95 to ensure optimal long-term control effect.

[0032] A multi-objective reward function is designed, with safety, energy consumption, and airflow stability as the core objectives. A weighted summation reward function R = ω1R1 + ω2R2 + ω3R3, where ω1 = 0.5, ω2 = 0.3, and ω3 = 0.2, is constructed. The weights are set based on the safety priority of coal mines. The specific design is as follows: Safety reward R1: Based on the distribution of gas concentration field, if the gas concentration is below the threshold of 0.8%, R1=1; if there is a region exceeding the threshold, it is calculated according to the proportion of the region exceeding the threshold, R1=1-λ×S, where λ is the penalty coefficient, set to 0.02, and S is the proportion of the region exceeding the threshold. When the proportion of the region exceeding the threshold is >20%, R1=-1, with strong penalty.

[0033] Energy consumption incentive R2: Based on the energy consumption of ventilation equipment, R2 = E0 - E / E0, where E0 is the baseline energy consumption and E is the current energy consumption. If E≤E0, R2 is positive; if E>E0, R2 is negative, encouraging the reduction of equipment energy consumption.

[0034] The airflow stability bonus R3 is calculated based on the airflow field distribution. The airflow velocity fluctuation coefficient σ is calculated, where σ = standard deviation / average value, and R3 = 1 - σ. If σ ≤ 0.1, the stable range is reached, and R3 = 1; if σ > 0.3, the unstable range is reached, and R3 = -0.5, ensuring the stability of the airflow field.

[0035] The collaborative control action command generation uses the airflow field and gas concentration field distribution as inputs, combined with the layout of underground ventilation equipment, including main ventilators, local ventilators, dampers, and air windows, to output multi-equipment collaborative control action commands. These commands include fan speed adjustment (±0-50 r / min), damper opening adjustment (±0-30°), and air window angle adjustment (±0-20°). The command output uses continuous value encoding to ensure control accuracy. At the same time, equipment linkage constraints are introduced, such as the coordinated speed of main ventilators and local ventilators, and the matching of damper and air window openings, to avoid system imbalance caused by the control of a single device.

[0036] Step 105: Based on the actual ventilation effect data after the execution of the action command, perform online adaptive updates to the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer.

[0037] Specifically, in this embodiment, the error between the predicted value and the actual value is used as the loss function, and the parameters of the GCN-STAN spatiotemporal prediction model are adjusted online using the SGD stochastic gradient descent algorithm. The actual value of the reward function is calculated based on the actual ventilation effect data, and compared with the value function predicted by the Critic network. The parameters of the Critic network are updated to optimize the value assessment accuracy. At the same time, combined with the Proximal policy constraint, the action generation strategy of the Actor network is adjusted so that the KL divergence between the new strategy and the old strategy is ≤0.01.

[0038] Actual ventilation effect data collection: After the action command is executed, the actual ventilation effect data within 10 minutes is collected in real time through the downhole data acquisition network. This includes the adjusted airflow velocity, gas concentration, equipment energy consumption, airflow fluctuation coefficient, etc., in each area. The data is compared with the prediction results, and error indices, mean absolute error (MAE) and root mean square error (RMSE), are calculated as the basis for updates.

[0039] The GCN-STAN model is updated online, using the error between predicted and actual values ​​as the loss function. The MSE loss function is adopted, and the model parameters are adjusted online through stochastic gradient descent and SGD algorithms. The AGCN convolution kernel weights, BI-LSTM hidden layer parameters, and attention weights are updated. The update frequency is synchronized with the action command execution cycle, once every 10 minutes. At the same time, an incremental learning strategy is adopted, which only updates local parameters on newly collected data to avoid system delays caused by full training and ensure that the model adapts to the dynamic changes of the downhole ventilation system.

[0040] The DDPG-Proximal policy optimizer updates online, calculating the actual value of the reward function based on actual ventilation effect data, comparing it with the value function predicted by the Critic network, updating the Critic network parameters, and optimizing the value assessment accuracy. Simultaneously, combined with Proximal policy constraints, it adjusts the action generation strategy of the Actor network to ensure that the KL divergence between the new and old strategies is ≤0.01, avoiding ventilation system oscillations caused by sudden policy changes. If the actual effect does not meet expectations, the reward value is <0, triggering an emergency update mechanism to accelerate parameter adjustment and ensure the safe and stable operation of the system.

[0041] Its beneficial effects are as follows: First, it eliminates heterogeneous data interference through standardized preprocessing, providing high-quality data support for subsequent prediction and control, and breaking down information silos. Second, it improves the accuracy of spatiotemporal prediction. The GCN-STAN model, combined with adaptive graph convolution and bidirectional LSTM, accurately captures the spatial correlation and temporal dynamics of the ventilation system, providing a reliable basis for control decisions. Third, it achieves multi-objective optimal control. The optimizer constructs a reward function based on safety, energy consumption, and stability to achieve coordinated equipment control, balancing safety baselines and operational efficiency. Fourth, it possesses dynamic adaptive capabilities, updating the model and strategy online based on actual results to adapt to changes in underground working conditions and avoid control failure. In summary, this method significantly improves the intelligence and precision of ventilation control, effectively reduces gas risk and energy consumption, and provides technical support for safe and efficient coal mining. It also improves the utilization efficiency and accuracy of ventilation optimization and control systems in coal mines.

[0042] Please see Figure 2 In the adaptive learning-based ventilation optimization and control method for coal mine mining, real-time collection of multi-source heterogeneous data during the coal mining process and preprocessing of this data to obtain initial multi-source heterogeneous data include the following steps. Step 201: Collect multi-source heterogeneous data in real time during the coal mining process, including at least sensor data, image data, ventilation equipment data, and geological dynamic data; Step 202: Use the 3σ criterion to remove extreme data caused by sensor failure and equipment interference, perform noise reduction on the image data, remove noise caused by downhole dust and light changes by Gaussian filtering, repair image edge defects by morphological operations, and complete missing data by linear interpolation. Step 203: Based on the timestamp, align data from different acquisition frequencies to a 1Hz time scale, use a weighted fusion algorithm to integrate heterogeneous data from the same source, extract feature vectors from image data through CNN, convert them into numerical data, and map them to the same feature space along with sensor data, equipment data, and geological data to generate a multi-source data matrix with consistent dimensions. Step 204: Perform Min-Max normalization on the multi-source data matrix to map parameters of different dimensions to the [0,1] interval. At the same time, use Z-Score normalization to eliminate data distribution bias and obtain the initial multi-source heterogeneous data.

[0043] The above describes embodiments of the adaptive learning-based ventilation optimization and control method for coal mine operations according to the present invention. Please refer to [link / reference]. Figure 3 In the adaptive learning-based ventilation optimization and control system for coal mines, the system includes the following modules: The multi-source data acquisition module is used to collect multi-source heterogeneous data in the coal mining process in real time, and to preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. The prediction model building module is used to construct the GCN-STAN spatiotemporal prediction model. It uses a 2-layer AGCN adaptive graph convolution to learn the non-Euclidean dependencies between nodes, captures long-term and short-term dynamics through bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism. The coal mine condition prediction module is used to input initial multi-source heterogeneous data into the GCN-STAN spatiotemporal prediction model and output the airflow field distribution and gas concentration field distribution of the mine. The coal mine real-time control module is used to design the DDPG-Proximal strategy optimizer. It takes the airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption and airflow stability as reward functions, and outputs action commands for coordinated control of multiple ventilation devices. The adaptive learning module is used to adaptively update the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer online based on the actual ventilation effect data after the execution of the action instructions.

[0044] Its beneficial effects are as follows: First, it eliminates heterogeneous data interference through standardized preprocessing, providing high-quality data support for subsequent prediction and control, and breaking down information silos. Second, it improves the accuracy of spatiotemporal prediction. The GCN-STAN model, combined with adaptive graph convolution and bidirectional LSTM, accurately captures the spatial correlation and temporal dynamics of the ventilation system, providing a reliable basis for control decisions. Third, it achieves multi-objective optimal control. The optimizer constructs a reward function based on safety, energy consumption, and stability to achieve coordinated equipment control, balancing safety baselines and operational efficiency. Fourth, it possesses dynamic adaptive capabilities, updating the model and strategy online based on actual results to adapt to changes in underground working conditions and avoid control failure. In summary, this method significantly improves the intelligence and precision of ventilation control, effectively reduces gas risk and energy consumption, and provides technical support for safe and efficient coal mining. It also improves the utilization efficiency and accuracy of ventilation optimization and control systems in coal mines.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and equivalents.

Claims

1. A ventilation optimization and control method for coal mines based on adaptive learning, characterized in that, The coal mine ventilation optimization and control method includes the following steps: Real-time acquisition of underground multi-source heterogeneous data during coal mining process; preprocessing of underground multi-source heterogeneous data to obtain initial multi-source heterogeneous data. A spatiotemporal prediction model of GCN-STAN is constructed. The non-Euclidean dependencies between nodes are learned by using a 2-layer AGCN adaptive graph convolution, the long-term and short-term dynamics are captured by bidirectional BI-LSTM, and the features are focused by a dual-channel attention mechanism. The initial multi-source heterogeneous data is input into the GCN-STAN spatiotemporal prediction model, which outputs the airflow field distribution and gas concentration field distribution of the mine. Design a DDPG-Proximal strategy optimizer, taking airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption and airflow stability as reward functions, and outputting action commands for coordinated control of multiple ventilation devices; Based on the actual ventilation effect data after the execution of the action command, the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer are updated online adaptively.

2. The adaptive learning-based ventilation optimization and control method for coal mines as described in claim 1, characterized in that, The real-time acquisition of underground multi-source heterogeneous data during the coal mining process, followed by preprocessing of the underground multi-source heterogeneous data to obtain initial multi-source heterogeneous data, including: Real-time acquisition of multi-source heterogeneous data from underground coal mining processes, including at least sensor data, image data, ventilation equipment data, and geological dynamic data; The extreme data caused by sensor failure and equipment interference is removed by using the 3σ criterion. The image data is denoised by using Gaussian filtering to remove noise caused by downhole dust and light changes. Morphological operations are used to repair image edge defects. Missing data is filled by linear interpolation. Based on timestamps, data from different acquisition frequencies are uniformly aligned to a 1Hz time scale. A weighted fusion algorithm is used to integrate heterogeneous data from the same source. Image data is transformed into numerical data by extracting feature vectors through CNN. It is then uniformly mapped to the same feature space with sensor data, equipment data, and geological data to generate a multi-source data matrix with consistent dimensions. The multi-source data matrix is ​​subjected to Min-Max normalization to map parameters of different dimensions to the [0,1] interval. At the same time, Z-Score normalization is used to eliminate data distribution bias and obtain the initial multi-source heterogeneous data.

3. The adaptive learning-based ventilation optimization and control method for coal mines as described in claim 1, characterized in that, The construction of the GCN-STAN spatiotemporal prediction model utilizes a 2-layer AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, captures long- and short-term dynamics through a bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism, including: The first layer AGCN is used for local spatial feature extraction. Based on the underground roadway topology, a dynamic graph adjacency matrix is ​​constructed. The matrix elements represent the ventilation correlation strength between different spatial nodes. The local spatial dependency relationship between nodes and neighboring nodes is learned through adaptive graph convolution kernels. The spatial distribution features of airflow and gas in the roadway are extracted, and the feature map is output. The second layer AGCN is used for global spatial feature fusion. Based on the feature map output by the first layer, the adjacency matrix is ​​dynamically updated. By expanding the receptive field, it learns global spatial dependencies across regions and outputs spatiotemporal fusion features.

4. The adaptive learning-based ventilation optimization and control method for coal mines as described in claim 1, characterized in that, The construction of the GCN-STAN spatiotemporal prediction model utilizes a 2-layer AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, captures long- and short-term dynamics through a bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism, including: The spatiotemporal fusion features output by AGCN are expanded according to the time series and input into the bidirectional BI-LSTM module. The forward LSTM captures the temporal dynamic features from the past to the present, and the backward LSTM captures the temporal correlation features from the present to the future. By concatenating the bidirectional output results, the time feature vector is obtained.

5. The adaptive learning-based ventilation optimization and control method for coal mines as described in claim 1, characterized in that, The construction of the GCN-STAN spatiotemporal prediction model utilizes a 2-layer AGCN adaptive graph convolution to learn non-Euclidean dependencies between nodes, captures long- and short-term dynamics through a bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism, including: By utilizing spatial attention channels to calculate attention weights, different weights are assigned to spatial features, focusing on the spatial features of key areas; Based on the feature attention channel, the fully connected layer is used to calculate the feature attention weight, which enhances the contribution of the feature, suppresses the interference of redundant features, and outputs the spatiotemporal fusion feature of the target after attention focusing.

6. The adaptive learning-based ventilation optimization and control method for coal mines as described in claim 1, characterized in that, The DDPG-Proximal strategy optimizer is designed to take airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption, and airflow stability as reward functions. It outputs action commands for the coordinated control of multiple ventilation devices, including: A dual-network architecture of Actor-Critic is used to construct the DDPG-Proximal policy optimizer. The Actor network is responsible for generating coordinated control action commands for ventilation equipment, while the Critic network is responsible for evaluating the value of the action commands and optimizing the policy. To prevent system instability caused by excessive policy updates, a Proximal policy constraint is introduced. The network learning rate is set to 0.001 and the discount factor is set to 0.

95.

7. The adaptive learning-based ventilation optimization and control method for coal mines as described in claim 1, characterized in that, The step of performing online adaptive updates to the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer based on the actual ventilation effect data after the execution of the action command includes: Using the error between the predicted and actual values ​​as the loss function, the parameters of the GCN-STAN spatiotemporal prediction model are adjusted online using the SGD stochastic gradient descent algorithm. The actual value of the reward function is calculated based on actual ventilation effect data. It is then compared with the value function predicted by the Critic network to update the Critic network parameters and optimize the value assessment accuracy. At the same time, the action generation strategy of the Actor network is adjusted in conjunction with the Proximal policy constraint so that the KL divergence between the new strategy and the old strategy is ≤0.

01.

8. A coal mine ventilation optimization and control system based on adaptive learning, characterized in that, The coal mine ventilation optimization and control system includes the following modules: The multi-source data acquisition module is used to acquire multi-source heterogeneous data in the coal mining process in real time, and to preprocess the multi-source heterogeneous data to obtain initial multi-source heterogeneous data. The prediction model building module is used to construct the GCN-STAN spatiotemporal prediction model. It uses a 2-layer AGCN adaptive graph convolution to learn the non-Euclidean dependencies between nodes, captures long-term and short-term dynamics through bidirectional BI-LSTM, and focuses features through a dual-channel attention mechanism. The coal mine condition prediction module is used to input the initial multi-source heterogeneous data into the GCN-STAN spatiotemporal prediction model and output the mine's airflow field distribution and gas concentration field distribution. The coal mine real-time control module is used to design the DDPG-Proximal strategy optimizer. It takes the airflow field distribution and gas concentration field distribution as inputs, and safety, energy consumption and airflow stability as reward functions, and outputs action commands for coordinated control of multiple ventilation devices. The adaptive learning module is used to perform online adaptive updates to the GCN-STAN spatiotemporal prediction model and the DDPG-Proximal policy optimizer based on the actual ventilation effect data after the execution of the action command.

9. The coal mine ventilation optimization and control system based on adaptive learning as described in claim 8, characterized in that, The real-time control module for coal mines includes the following sub-modules: A submodule is constructed to build the DDPG-Proximal policy optimizer using an Actor-Critic dual-network architecture. The Actor network is responsible for generating coordinated control action instructions for ventilation equipment, while the Critic network is responsible for evaluating the value of the action instructions and optimizing the policy. A submodule is introduced to introduce Proximal policy constraints to prevent the system from becoming unstable due to excessive policy updates. The network learning rate is set to 0.001 and the discount factor is set to 0.

95.

10. The coal mine ventilation optimization and control system based on adaptive learning as described in claim 8, characterized in that, The real-time control module for coal mines includes the following sub-modules: The adjustment submodule is used to adjust the parameters of the GCN-STAN spatiotemporal prediction model online using the SGD stochastic gradient descent algorithm, with the error between the predicted and actual values ​​as the loss function. The update submodule is used to calculate the actual value of the reward function based on the actual ventilation effect data, compare it with the value function predicted by the Critic network, update the Critic network parameters, optimize the value assessment accuracy, and adjust the action generation strategy of the Actor network in combination with the Proximal policy constraint so that the KL divergence between the new strategy and the old strategy is ≤0.01.

Images