Mine ventilation control method and system, electronic equipment and storage medium
By combining multi-task learning with MPPO reinforcement learning, a mine ventilation control system was constructed, which solved the problems of poor adaptability and low efficiency of traditional methods. It achieved a multi-objective balance between safety, decontamination efficiency and energy consumption, and improved the accuracy and adaptability of mine ventilation control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING BEIKUANG INTELLIGENT TECH CO LTD
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-21
AI Technical Summary
Traditional mine ventilation control methods are difficult to adapt to the dynamic changes in pollutant concentration after blasting, and cannot achieve a balance between multiple objectives such as safety, decontamination efficiency and energy consumption cost. In addition, data acquisition is difficult, and computational fluid dynamics methods have long calculation times and cannot meet the needs of real-time control.
By combining multi-task learning with MPPO reinforcement learning, a multi-task learning prediction network and an MPPO decision model are constructed. High-quality datasets are generated using fluid mechanics theory to achieve intelligent control of mine ventilation.
It achieves a multi-objective intelligent balance between safety, decontamination efficiency and energy consumption in mine ventilation, improves control accuracy and adaptability, reduces calculation time, lowers energy consumption and increases control success rate.
Smart Images
Figure CN121897393A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mine ventilation technology, and more specifically, to a mine ventilation control method, system, electronic device, and storage medium. Background Technology
[0002] In metal mine tunneling operations, blasting generates large amounts of harmful substances such as carbon monoxide (CO) and dust, posing a serious threat to the safety of workers. Traditional ventilation control methods mainly rely on fixed rules or manual experience, which are difficult to adapt to the dynamic changes in pollutant concentrations after blasting. They also fail to achieve a balance between multiple objectives such as safety, decontamination efficiency, and energy costs, resulting in low control precision. Summary of the Invention
[0003] In view of the above, the purpose of this application is to overcome the shortcomings of the prior art and provide a mine ventilation control method, system, electronic device, and storage medium. This application provides the following technical solution: In a first aspect, the present invention provides a mine ventilation control method, the method comprising: Obtain the mine ventilation dataset; A multi-task learning and prediction network is constructed, comprising an input layer, a shared feature layer, and a task decoding layer. The shared feature layer comprises a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer comprises a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. The multi-task learning prediction network is trained based on the mine ventilation dataset to obtain the target learning prediction network. An MPPO decision model is constructed, which includes a policy network and a value network, both of which adopt a dual-hidden-layer fully connected structure. Based on the mine ventilation dataset and the target learning prediction network, the MPPO decision model is trained to obtain the target MPPO decision model. Multiple environmental parameters of the mine to be ventilated are obtained, and each of the environmental parameters is input into the target MPPO decision model to obtain the ventilation parameter adjustment strategy; The ventilation of the mine to be ventilated is controlled according to the ventilation parameter adjustment strategy.
[0004] In one embodiment, acquiring the mine ventilation dataset includes: A ventilation simulation model was constructed based on the pollutant exponential decay law. A jet velocity decay equation is constructed based on Prandtl's jet theory; Based on the ventilation simulation model and the jet velocity attenuation equation, the Latin hypercube sampling method is used to generate multiple sets of ventilation data within a preset parameter space. Physical consistency verification is performed on each of the ventilation data, and the ventilation data that pass the verification constitute the mine ventilation dataset.
[0005] In one embodiment, the ventilation data includes: time elapsed since the blast, carbon monoxide concentration, ventilation duct installation distance, wind speed, fan flow rate, and fan power. The physical consistency verification of each of the ventilation data includes: Verify whether the negative correlation between the time elapsed after the explosion and the carbon monoxide concentration meets the first preset correlation requirement; Verify whether the negative correlation between the installation distance of the wind duct and the wind speed meets the second preset correlation requirement; Verify whether the positive correlation between the fan flow rate and the fan power meets the third preset correlation requirement; Verify whether the goodness of fit between the ventilation simulation model and the jet velocity attenuation equation for each ventilation data meets the preset fitting requirements; If all conditions are met, then the ventilation data verification is deemed successful.
[0006] In one embodiment, training the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network includes: Set the initial learning rate and weight decay coefficient of the AdamW optimizer, and define the loss function of the multi-task learning prediction network; Based on the mine ventilation dataset, the multi-task learning prediction network is trained for a preset number of rounds using the AdamW optimizer and the loss function. During the training process, the learning rate of the multi-task learning prediction network dynamically decays from the initial learning rate according to a preset rule. The multi-task learning prediction network after training is determined as the target learning prediction network.
[0007] In one embodiment, training the multi-task learning prediction network based on the mine ventilation dataset using the AdamW optimizer and the loss function includes: The mine ventilation dataset is divided into a training set, a validation set, and a test set according to a preset ratio. Based on the training set, the multi-task learning and prediction network is trained iteratively for a preset number of rounds. In each round of training, sample data is extracted from the training set according to a preset batch size and input into the multi-task learning and prediction network. The mixed loss value is calculated through forward propagation, and the network parameters of the multi-task learning and prediction network are updated by backpropagation in combination with the AdamW optimizer. After each preset number of training rounds, the performance of the trained multi-task learning prediction network is evaluated using the validation set. After training reaches the preset number of rounds, the network performance of the trained multi-task learning prediction network is verified using a test set to see if it meets the preset requirements.
[0008] In one embodiment, constructing the MPPO decision model includes: Define a Markov decision quintuple, which includes a state space, an action space, a state transition rule, a reward function, and a discount factor. The state transition rule is determined based on the target learning prediction network. Construct the policy network and the value network; By integrating the Markov decision quintuple, the multi-objective hierarchical reward function, the MPPO algorithm logic, the policy network, and the value network, the MPPO decision model is obtained. The MPPO algorithm logic uses generalized advantage to determine the action advantage value and performs dynamic weight adjustment through a sigmoid function.
[0009] In one embodiment, training the MPPO decision model based on the mine ventilation dataset and the target learning prediction network to obtain the target MPPO decision model includes: Based on the mine ventilation dataset and the target learning prediction network, a reinforcement learning training environment is constructed, and the reinforcement learning training environment outputs the state transition results through the target learning prediction network. Start training of the MPPO decision model. The training process is divided into an exploration phase, an improvement phase, and a stabilization phase, with each phase divided according to a preset number of steps. After each sampling step of training is completed, the parameters of the MPPO decision model are updated using the optimized number of rounds; The average reward and task success rate are monitored in real time during the training process. When the average reward reaches a preset reward threshold and the task success rate reaches a preset success rate threshold, the MPPO decision model is determined to have converged. The converged MPPO decision model is determined as the target MPPO decision model.
[0010] Secondly, the present invention provides a mine ventilation control system, the system comprising: The dataset building module is used to obtain mine ventilation datasets; A network construction module is used to construct a multi-task learning and prediction network. The multi-task learning and prediction network includes an input layer, a shared feature layer, and a task decoding layer. The shared feature layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer includes a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. The network training module is used to train the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network; The model building module is used to build an MPPO decision model, which includes a policy network and a value network, both of which adopt a double hidden layer fully connected structure. The model training module is used to train the MPPO decision model based on the mine ventilation dataset and the target learning prediction network to obtain the target MPPO decision model. The decision module is used to acquire multiple environmental parameters of the mine to be ventilated, input each of the environmental parameters into the target MPPO decision model, and obtain the ventilation parameter adjustment strategy. The ventilation control module is used to control the ventilation of the mine to be ventilated according to the ventilation parameter adjustment strategy.
[0011] Thirdly, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when run on the processor, executes the mine ventilation control method described in any of the foregoing embodiments.
[0012] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the mine ventilation control method described in any of the foregoing embodiments.
[0013] This application combines multi-task learning with MPPO reinforcement learning, and relies on fluid mechanics theory to generate high-quality datasets and lightweight prediction networks. It achieves a multi-objective intelligent balance between safety, decontamination efficiency and energy consumption in mine ventilation, effectively solving the problems of poor adaptability, low efficiency and difficulty in data acquisition of traditional control methods.
[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 A schematic flowchart of a mine ventilation control method provided in an embodiment of this application is shown; Figure 2 This paper illustrates another schematic flowchart of the mine ventilation control method provided in an embodiment of this application; Figure 3 The graph showing the exponential decay analysis of carbon monoxide concentration is shown. Figure 4 The graph shows the attenuation of wind speed with distance. Figure 5 The graph showing the relationship between fan flow rate and fan power is shown. Figure 6 A network structure diagram of a multi-task learning prediction network is shown; Figure 7 A schematic diagram of the MPPO decision model is shown. Figure 8 A schematic diagram of a mine ventilation control system provided in an embodiment of this application is shown; Figure 9 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown.
[0017] Explanation of key component symbols: 800 - Mine ventilation control system; 810 - Dataset construction module; 820 - Network construction module; 830 - Network training module; 840 - Model construction module; 850 - Model training module; 860 - Decision module; 870 - Ventilation control module; 900 - Electronic equipment; 901 - Transceiver; 902 - Processor; 903 - Memory. Detailed Implementation
[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0019] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the template description is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0021] Example 1 In metal mine tunneling operations, blasting generates large amounts of carbon monoxide (CO) and dust, posing a serious threat to worker safety. Traditional ventilation control methods rely primarily on fixed rules or manual experience, making it difficult to adapt to the dynamic changes in pollutant concentrations after blasting. They fail to achieve a multi-objective balance between safety, decontamination efficiency, and energy costs, and lack adaptive adjustment capabilities for different operating conditions. While computational fluid dynamics-based ventilation simulation methods exist, their computation time is long, requiring hours to days for a single solution, which cannot meet real-time control requirements. Some studies employ simplified constant pollution source assumptions, which do not match the actual attenuation patterns of pollutants after blasting, resulting in insufficient control accuracy. Therefore, there is an urgent need for an intelligent control method capable of rapidly and accurately predicting ventilation effects and achieving multi-objective adaptive optimization. For more information, please refer to [link to relevant documentation]. Figure 1 This application provides a mine ventilation control method, which includes steps S110 to S170.
[0022] Step S110: Obtain the mine ventilation dataset.
[0023] It is understandable that the dynamic changes in pollutant concentration after blasting at the metal mine face present challenges. Traditional field experiments are costly, have long sampling cycles, and pose significant safety risks, making it difficult to obtain high-quality data. Existing ventilation simulation methods based on computational fluid dynamics have long computation times, requiring hours to days for a single solution, which cannot meet the needs of real-time control. Furthermore, the constant pollution source assumption used in some studies does not match the actual attenuation law of pollutants after blasting, further affecting the reliability of the data.
[0024] In this embodiment, the dataset is obtained by combining theoretical simulation and physical verification, providing high-quality and reliable data support for subsequent training of multi-task learning prediction networks and MPPO decision models.
[0025] In one implementation, please refer to Figure 2 The acquisition of the mine ventilation dataset includes steps S111 to S114.
[0026] Step S111: Construct a ventilation simulation model based on the pollutant index decay law.
[0027] Based on fluid mechanics theory, a ventilation simulation model considering the exponential decay law of pollutants after blasting is constructed:
[0028] in, Indicates pollutant concentration. This indicates the initial peak concentration after the blast. Represents the attenuation constant. Indicates the time elapsed after the explosion. Indicates the time elapsed after the explosion. The concentration of pollutants afterward.
[0029] Step S112: Construct the jet velocity decay equation based on Prandtl's jet theory.
[0030] Based on Prandtl's jet theory, a jet velocity decay equation considering the coupled effects of temperature, Reynolds coefficient, and jet angle is established:
[0031] in, Indicates temperature as The jet angle is At that time, the distance from the jet outlet The centerline velocity at that point; Indicates the jet exit velocity. Indicates the diameter of the jet outlet. Indicates the jet-corrected diffusion coefficient. , Represents the Reynolds coefficient. This represents the angle correction function.
[0032] Step S113: Based on the ventilation simulation model and the jet velocity attenuation equation, the Latin hypercube sampling method is used to generate multiple sets of ventilation data within a preset parameter space.
[0033] In this embodiment, the Latin hypercube sampling method is used to generate multiple sets of ventilation data within a preset parameter space. The input parameters of the preset parameter space include the distance between the ventilation duct and the working face (3-12m), the installation angle of the ventilation duct (0-30°), the diameter of the ventilation duct (0.6-1.5m), the time elapsed after the blast (300-7200s), the initial pollutant concentration and volumetric flow rate (1-20m³ / s), and the output parameters include: carbon monoxide concentration, dust concentration, wind speed and safety level.
[0034] It is understandable that by coupling the two theoretical models, the generated data is ensured to conform to the actual physical operation law of mine ventilation, avoiding the problem that the traditional constant pollution source assumption does not match the actual attenuation law. The Latin hypercube sampling method can uniformly cover the entire parameter range with a limited sample size, efficiently obtain physical rationality data under multiple working conditions, avoid the defects of high cost, long cycle and high safety risk of field experiments, and significantly improve the data generation efficiency compared with traditional computational fluid dynamics methods, providing high-quality, high-coverage and high-reliability data support for the training of subsequent multi-task learning prediction networks.
[0035] Step S114: Perform physical consistency verification on each of the ventilation data, and the ventilation data that pass the verification constitute the mine ventilation dataset.
[0036] In this embodiment, the generated data is physically consistent to ensure that the time elapsed after the blast is negatively correlated with the carbon monoxide concentration, the distance from the ventilation duct is negatively correlated with the wind speed, and the fan flow rate is positively correlated with the fan power. The model fit is also verified.
[0037] In one embodiment, the ventilation data includes: time elapsed since the blast, carbon monoxide concentration, ventilation duct installation distance, wind speed, fan flow rate, and fan power. The physical consistency verification of each of the ventilation data includes: Verify whether the negative correlation between the time elapsed after the explosion and the carbon monoxide concentration meets the first preset correlation requirement; Verify whether the negative correlation between the installation distance of the wind duct and the wind speed meets the second preset correlation requirement; Verify whether the positive correlation between the fan flow rate and the fan power meets the third preset correlation requirement; Verify whether the goodness of fit between the ventilation simulation model and the jet velocity attenuation equation for each ventilation data meets the preset fitting requirements; If all conditions are met, then the ventilation data verification is deemed successful.
[0038] In this embodiment, based on the physical operation law of mine ventilation and the previously constructed jet velocity attenuation equation (reflecting the coupling relationship between the distance to the ventilation duct and the wind speed), multi-dimensional physical consistency verification is carried out on the generated ventilation data. Specifically, this includes verifying whether the negative correlation between the time elapsed after blasting and the carbon monoxide concentration meets the first preset correlation requirement (Pearson correlation coefficient not greater than -0.320, significance level less than 0.001), verifying whether the negative correlation between the distance between the ventilation duct and the working face and the wind speed conforms to the jet theory derivation results (Pearson correlation coefficient not greater than -0.440, significance level less than 0.001), verifying whether the positive correlation between the fan flow rate and the fan power meets the law of energy conservation (Pearson correlation coefficient not less than 0.925), and verifying whether the goodness of fit between the ventilation simulation model and the jet velocity attenuation equation on the data reaches the preset standard (goodness of fit greater than 0.98). All ventilation data that meet the above verification requirements are selected, and these verified ventilation data together constitute the mine ventilation dataset.
[0039] For details, please see Figure 3 , Figure 3 The graph of exponential decay of carbon monoxide concentration is shown, which demonstrates the negative correlation between time and CO concentration (r=-0.320), and verifies the goodness of fit of the exponential decay model (R²=0.987).
[0040] Please see Figure 4 , Figure 4 The analysis diagram of wind speed attenuation with distance is shown, demonstrating the negative correlation between wind speed and duct distance (r=-0.440), which verifies Prandtl's jet theory (R²=0.990).
[0041] Please see Figure 5 , Figure 5 A graph showing the relationship between fan flow rate and fan power is presented, demonstrating a positive correlation between fan flow rate and fan power (r=0.925), verifying the law of energy conservation (R²=0.997).
[0042] It is understandable that by verifying the physical laws in multiple dimensions, data noise that does not conform to the actual ventilation operation logic of a mine can be accurately eliminated, ensuring the physical rationality and reliability of the dataset and avoiding deviations in the subsequent training of the multi-task learning and prediction network due to data distortion.
[0043] Step S120: Construct a multi-task learning prediction network. The multi-task learning prediction network includes an input layer, a shared feature layer, and a task decoding layer. The shared feature layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer includes a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head.
[0044] In this embodiment, please refer to Figure 6 , Figure 6 A network structure diagram of a multi-task learning prediction network is shown. The network input layer receives seven physical parameters: distance to the ventilation duct, installation angle of the ventilation duct, diameter of the ventilation duct, volumetric flow rate of pollutants, initial carbon monoxide concentration, initial dust concentration, and time elapsed after the blast, providing comprehensive raw data support for feature extraction. The shared feature layer consists of three fully connected layers, each configured with batch normalization, linear rectification activation function, and dropout regularization, respectively, to extract nonlinear coupling features of the ventilation system through deep mining while avoiding overfitting. The task decoding layer sets up four output heads in parallel. The carbon monoxide concentration regression head, dust concentration regression head, and wind speed regression head use linear outputs to accurately predict three types of key ventilation state parameters, respectively. The safety classification head outputs the probability of safety compliance through an S-shaped activation function, realizing binary classification judgment of the safety state.
[0045] It is understandable that by using an architecture of shared feature extraction and multi-task parallel decoding, parameter redundancy can be reduced (the total number of parameters is controlled within 3×10). 4 While achieving a scale of [number], it enables simultaneous prediction of multi-dimensional ventilation conditions, improving training and inference efficiency compared to traditional single-task networks; combining batch normalization and dropout mechanisms enhances the network's generalization ability, ensuring high-precision predictions under different mine ventilation conditions (carbon monoxide concentration prediction R²=0.953, wind speed prediction R²=0.972, safety classification accuracy 93.3%); single inference time is less than 1ms, meeting the real-time decision-making requirements of reinforcement learning.
[0046] Step S130: Train the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network.
[0047] In one embodiment, training the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network includes: Set the initial learning rate and weight decay coefficient of the AdamW optimizer, and define the loss function of the multi-task learning prediction network; Based on the mine ventilation dataset, the multi-task learning prediction network is trained for a preset number of rounds using the AdamW optimizer and the loss function. During the training process, the learning rate of the multi-task learning prediction network dynamically decays from the initial learning rate according to a preset rule. The multi-task learning prediction network after training is determined as the target learning prediction network.
[0048] In this embodiment, a mine ventilation dataset divided into training, validation, and test sets according to a preset ratio is used as the training basis. Systematic training is conducted on the constructed multi-task learning and prediction network consisting of an input layer, a shared feature layer, and a task decoding layer. Before training, the AdamW optimization algorithm is configured, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001. A hybrid loss function (mean squared error loss corresponding to regression prediction tasks for carbon monoxide concentration, dust concentration, and wind speed, 0.5) is used. (A binary cross-entropy loss is used for classification tasks related to safety compliance probability). During training, the network is trained for 200 iterations with a preset batch size of 32. After every 20 iterations, the network performance is evaluated using a validation set. In the 100th and 150th iterations, the learning rate is decayed to a preset intermediate value of 0.0001 and a preset final value of 0.00001, respectively. The network parameters are continuously updated through backpropagation. After training, the network's generalization ability is verified using a test set. When the goodness of fit R² for carbon monoxide concentration prediction reaches 0.953, the goodness of fit R² for dust concentration prediction reaches 0.931, the goodness of fit R² for wind speed prediction reaches 0.972, and the safety classification accuracy reaches 93.3%, the network training is considered complete, and the target learning prediction network is obtained.
[0049] It is understandable that by reasonably dividing the dataset, optimizing training parameters, and adopting a hybrid loss function, the network can accurately learn the nonlinear relationships between various parameters of the ventilation system, thereby achieving high-precision prediction of multiple tasks simultaneously while controlling the number of network parameters.
[0050] In one embodiment, training the multi-task learning prediction network based on the mine ventilation dataset using the AdamW optimizer and the loss function includes: The mine ventilation dataset is divided into a training set, a validation set, and a test set according to a preset ratio. Based on the training set, the multi-task learning and prediction network is trained iteratively for a preset number of rounds. In each round of training, sample data is extracted from the training set according to a preset batch size and input into the multi-task learning and prediction network. The mixed loss value is calculated through forward propagation, and the network parameters of the multi-task learning and prediction network are updated by backpropagation in combination with the AdamW optimizer. After each preset number of training rounds, the performance of the trained multi-task learning prediction network is evaluated using the validation set. After training reaches the preset number of rounds, the network performance of the trained multi-task learning prediction network is verified using a test set to see if it meets the preset requirements.
[0051] In this embodiment, the multi-task learning prediction network is first trained iteratively for 200 rounds based on the training set. In each round, sample data (including seven-dimensional input parameters such as duct distance and installation angle, and corresponding output parameters such as CO concentration and dust concentration) is extracted from the training set in a preset batch size of 32 and input into the network. A mixed loss value is calculated through forward propagation (mean squared error loss corresponds to the regression tasks of CO concentration, dust concentration, and wind speed; 0.5 times binary cross-entropy loss corresponds to the safety classification task). Then, the weights and bias parameters of the network are dynamically updated using the backpropagation algorithm combined with the AdamW optimizer (initial learning rate 0.001, weight decay coefficient 0.0001). After every 20 rounds of training at a preset interval, the prediction accuracy and classification accuracy of the network are evaluated using a validation set. Based on the evaluation results, the learning rate is decayed to 0.0001 and 0.00001 respectively in the 100th and 150th rounds according to preset rules, adjusting the training direction in a timely manner. After training reaches the preset number of 200 rounds, the network performance is verified using a test set. When the concentration prediction R² is ≥0.953, the dust concentration prediction R² is ≥0.931, the wind speed prediction R² is ≥0.972, and the safety classification accuracy is ≥93.3%, the preset requirements are met.
[0052] Understandably, by scientifically dividing the dataset and employing a batch training mechanism, training efficiency is improved while overfitting is avoided. The validation through interval rounds and the dynamic learning rate decay strategy ensure that the network converges quickly and stabilizes at its optimal state. Final validation on the test set guarantees the generalization ability and prediction reliability of the target learning and prediction network, with its parameter count controlled at 3×102. 4 With a scale of less than 1ms per inference, it achieves both high-precision prediction across multiple tasks and meets the real-time decision-making requirements of reinforcement learning.
[0053] Step S140: Construct an MPPO decision model, which includes a policy network and a value network, both of which adopt a double hidden layer fully connected structure.
[0054] In this embodiment, please refer to Figure 7 , Figure 7A schematic diagram of the MPPO decision model is shown. Both the policy network and value network in the MPPO decision model adopt a double-hidden-layer fully connected structure. The input layers each receive 8-dimensional state-space parameters (ventilation duct distance, installation angle, ventilation duct diameter, volumetric flow rate, carbon monoxide concentration, dust concentration, wind speed, and safety level). Each layer has 64 neurons, and each hidden layer uses the Rectified Linear Activation Function (ReLU) to enhance nonlinear feature extraction capabilities. The policy network output layer is divided into two parallel branches: action mean and action standard deviation, each outputting a 4-dimensional result (corresponding to changes in ventilation duct distance, installation angle, etc.). The action parameters for changes in duct diameter and volumetric flow rate are mapped to the legal action range (matching preset constraints such as duct distance of 3-12m and installation angle of 0-30°) using the hyperbolic tangent function (Tanh) to ensure compliance of the action parameters. The output layer of the value network is a 1-dimensional state value, used to quantitatively evaluate the long-term benefits in the current state and provide a valuable reference for strategy optimization. During the network initialization phase, the action mean of the policy network is initialized to 0 and the action standard deviation is initialized to 0.1. The parameters of the value network are initialized using a random normal distribution to ensure a stable training starting point.
[0055] It is understandable that the policy network is responsible for outputting the optimal control action, while the value network provides reliable state value support for generalized advantage estimation (GAE), helping the MPPO algorithm to achieve stable iterative updates of the policy, and ultimately achieving a multi-objective balance of "safety, decontamination efficiency, and energy consumption" in mine ventilation, thereby improving the accuracy and robustness of ventilation control.
[0056] In one embodiment, constructing the MPPO decision model includes: Define a Markov decision quintuple, which includes a state space, an action space, a state transition rule, a reward function, and a discount factor. The state transition rule is determined based on the target learning prediction network. Construct the policy network and the value network; By integrating the Markov decision quintuple, the multi-objective hierarchical reward function, the MPPO algorithm logic, the policy network, and the value network, the MPPO decision model is obtained. The MPPO algorithm logic uses generalized advantage to determine the action advantage value and performs dynamic weight adjustment through a sigmoid function.
[0057] In this embodiment, a Markov decision quintuple is first defined, which includes an 8-dimensional state space (distance from the duct, installation angle, duct diameter, volumetric flow rate, carbon monoxide concentration, dust concentration, wind speed, and safety level), a 4-dimensional action space (changes in duct distance, installation angle, duct diameter, and volumetric flow rate), a state transition rule based on a target learning prediction network (inputting the current state and action parameters, outputting the state at the next moment), a multi-objective hierarchical reward function (including safety reward, efficiency reward, energy consumption penalty, and time penalty), and a preset discount factor (0.99), fully covering the core elements of reinforcement learning decision-making. Next, a policy network and a value network are constructed, both using a double-hidden-layer fully connected structure with 64 neurons per layer and a ReLU activation function. The policy network outputs the action mean and standard deviation, mapped to a valid interval using Tanh, while the value network outputs the state value. Finally, the aforementioned quintuple, multi-objective hierarchical reward function, and MPPO are integrated. The algorithm logic (using Generalized Advantage Estimation (GAE) to calculate the action advantage value, with a GAE discount factor of 0.99 and a decay coefficient of 0.95; and using a S-shaped function to achieve dynamic weight adjustment, with a carbon monoxide safety sensitivity parameter of 0.1 and an energy consumption sensitivity parameter of 0.05) clarifies the data flow and algorithm execution order, forming a complete MPPO decision model.
[0058] The quintuple definition aligns with the physical scenario of mine ventilation, and the state transition rules rely on a target learning prediction network to ensure accuracy. The dual-network architecture works in conjunction with the MPPO core algorithm to improve the accuracy of advantage value calculation through generalized advantage estimation, and achieve a multi-objective balance between safety, decontamination efficiency, and energy consumption through dynamic weight adjustment. The network structure is lightweight and computationally efficient, and together with the algorithm optimization logic, it ensures that the model has real-time decision-making capabilities, providing an adaptive and high-precision intelligent control strategy for mine ventilation systems, significantly improving control performance and engineering practicality.
[0059] Step S150: Based on the mine ventilation dataset and the target learning prediction network, train the MPPO decision model to obtain the target MPPO decision model.
[0060] In one embodiment, training the MPPO decision model based on the mine ventilation dataset and the target learning prediction network to obtain the target MPPO decision model includes: Based on the mine ventilation dataset and the target learning prediction network, a reinforcement learning training environment is constructed, and the reinforcement learning training environment outputs the state transition results through the target learning prediction network. Start training of the MPPO decision model. The training process is divided into an exploration phase, an improvement phase, and a stabilization phase, with each phase divided according to a preset number of steps. After each sampling step of training is completed, the parameters of the MPPO decision model are updated using the optimized number of rounds; The average reward and task success rate are monitored in real time during the training process. When the average reward reaches a preset reward threshold and the task success rate reaches a preset success rate threshold, the MPPO decision model is determined to have converged. The converged MPPO decision model is determined as the target MPPO decision model.
[0061] In this embodiment, a reinforcement learning training environment is constructed based on the target learning prediction network and the 8-dimensional state space and 4-dimensional action space of the Markov decision quintuple. This environment directly outputs the state parameters of the next moment by inputting the current state parameters and action parameters into the target learning prediction network to complete the state transition. The state transition result is consistent with the jet velocity decay law and pollutant exponential decay characteristics based on Prandtl jet theory. The MPPO decision model training is initiated, divided into three phases according to a preset step range: the exploration phase (0-5000 steps) uses a high-entropy strategy to fully explore the action space; the rapid improvement phase (5000-12000 steps) rapidly optimizes the strategy parameters based on reward feedback; and the stable optimization phase (12000-20000 steps) fine-tunes the parameters to stabilize the strategy. After sampling every 1024 steps, training data is extracted in batches of 64. The action advantage value is calculated using generalized advantage estimation (GAE, discount factor 0.99, decay coefficient 0.95). The strategy network parameters are iteratively updated in conjunction with 10 optimization rounds. At the same time, the value network is updated according to the value loss weight of 0.5. The entropy coefficient of 0.01 is used to maintain the exploratory nature of the strategy, and the maximum norm of gradient pruning of 0.5 is used to avoid training divergence. During training, the average reward and task success rate are monitored in real time. When the average reward reaches the preset threshold of 107 and the task success rate reaches 98%, the model is considered to have converged. Finally, the converged MPPO decision model is determined as the target MPPO decision model.
[0062] Step S160: Obtain multiple environmental parameters of the mine to be ventilated, input each of the environmental parameters into the target MPPO decision model, and obtain the ventilation parameter adjustment strategy.
[0063] In this embodiment, real-time environmental parameters of the mine to be ventilated are acquired, including equipment parameters such as the distance between the ventilation duct and the working face, installation angle, ventilation duct diameter, and current volumetric flow rate, as well as operating condition parameters such as the elapsed time after blasting, initial carbon monoxide concentration, initial dust concentration, and real-time wind speed, forming 8-dimensional state input data. This data is then input into a training convergent target MPPO decision model. The model extracts nonlinear features through a policy network and outputs 4-dimensional action parameters (change in ventilation duct distance). Change in installation angle Change in duct diameter Volumetric flow rate change The action parameters are mapped to the valid range via Tanh, such as... Corresponding to a range of 3-12m with a step size of 0.5m, and combined with the dynamic weight logic of a multi-objective hierarchical reward function, a ventilation parameter adjustment strategy adapted to the current working conditions is finally generated.
[0064] The strategy generation relies on precise state-action association learning, which can adapt to different pollution conditions in the mine to be ventilated, and the single strategy output response time is short, meeting the real-time control requirements; the adjustment strategy not only conforms to physical laws such as Prandtl jet theory, but also balances multiple control objectives, avoiding the limitations of traditional fixed rules or human experience.
[0065] Step S170: Control the ventilation of the mine to be ventilated according to the ventilation parameter adjustment strategy.
[0066] In this embodiment, the intelligent ventilation control equipment for mines adjusts the ventilation parameters according to the ventilation parameter adjustment strategy output by the target MPPO decision model and the parameter update rules, and precisely adjusts the distance of the ventilation duct, the installation angle, the diameter of the ventilation duct, and the volumetric flow rate of the fan. During the control process, the real-time ventilation status is continuously fed back and the adjustment range is dynamically corrected until the terminal carbon monoxide concentration of the mine to be ventilated is <30mg / m³, the dust concentration is <15mg / m³, the wind speed is >0.5m / s and the safety level is >0.8, thus achieving the standard control.
[0067] Compared to traditional fixed parameter control and manual experience adjustment, it can achieve standard control within 30 steps (about 13 minutes), which is more than 3 times faster; energy consumption is reduced by 32% compared to traditional methods, the average pollutant removal rate reaches 26%-28%, and the control success rate is 100% in multiple scenarios; the control process does not require manual intervention, adapts to the dynamic change characteristics of pollutants after blasting in mine tunneling faces, and significantly improves ventilation safety and the cost-effectiveness of engineering applications.
[0068] The mine ventilation control method provided in this application involves: acquiring a mine ventilation dataset; constructing a multi-task learning prediction network, which includes an input layer, a shared feature layer, and a task decoding layer. The shared feature layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer includes a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. The multi-task learning prediction network is trained based on the mine ventilation dataset to obtain a target learning prediction network. An MPPO decision model is constructed, which includes a policy network and a value network, both of which employ a double-hidden-layer fully connected structure. The MPPO decision model is trained based on the mine ventilation dataset and the target learning prediction network to obtain a target MPPO decision model. Multiple environmental parameters of the mine to be ventilated are acquired, and each environmental parameter is input into the target MPPO decision model to obtain a ventilation parameter adjustment strategy. Ventilation of the mine to be ventilated is controlled according to the ventilation parameter adjustment strategy. This application combines multi-task learning with MPPO reinforcement learning, and relies on fluid mechanics theory to generate high-quality datasets and lightweight prediction networks. It achieves a multi-objective intelligent balance between safety, decontamination efficiency and energy consumption in mine ventilation, effectively solving the problems of poor adaptability, low efficiency and difficulty in data acquisition of traditional control methods.
[0069] Example 2 In addition, please see Figure 8 This application also provides a mine ventilation control system 800, comprising: Dataset building module 810 is used to obtain mine ventilation datasets; The network construction module 820 is used to construct a multi-task learning and prediction network. The multi-task learning and prediction network includes an input layer, a shared feature layer, and a task decoding layer. The shared feature layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer includes a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. Network training module 830 is used to train the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network; The model building module 840 is used to build an MPPO decision model, which includes a policy network and a value network, both of which adopt a double hidden layer fully connected structure. The model training module 850 is used to train the MPPO decision model based on the mine ventilation dataset and the target learning prediction network to obtain the target MPPO decision model. The decision module 860 is used to acquire multiple environmental parameters of the mine to be ventilated, input each of the environmental parameters into the target MPPO decision model, and obtain the ventilation parameter adjustment strategy. The ventilation control module 870 is used to control the ventilation of the mine to be ventilated according to the ventilation parameter adjustment strategy.
[0070] The mine ventilation control system 800 provided in this application embodiment can execute the mine ventilation control method provided in the above-described method embodiment 1. To avoid repetition, it will not be described again here.
[0071] Example 3 Furthermore, this embodiment of the invention provides an electronic device 900, including a memory 903 and a processor 902. The memory 903 stores a computer program, and the computer program executes the mine ventilation control method provided in Embodiment 1 when it runs on the processor 902.
[0072] For details, please see Figure 9 The electronic device 900 includes: a transceiver 901, a bus interface, and a processor 902. The processor 902 is used to acquire a mine ventilation dataset; construct a multi-task learning prediction network, which includes an input layer, a shared feature layer, and a task decoding layer. The shared feature layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer includes a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. The multi-task learning prediction network is trained based on the mine ventilation dataset to obtain a target learning prediction network. An MPPO decision model is constructed, which includes a policy network and a value network, both of which adopt a double hidden layer fully connected structure. The MPPO decision model is trained based on the mine ventilation dataset and the target learning prediction network to obtain a target MPPO decision model. Multiple environmental parameters of the mine to be ventilated are acquired, and each environmental parameter is input into the target MPPO decision model to obtain a ventilation parameter adjustment strategy. The ventilation of the mine to be ventilated is controlled according to the ventilation parameter adjustment strategy.
[0073] In this embodiment of the invention, the electronic device 900 further includes a memory 903. Figure 9In this context, the bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors 902 (represented by processor 902) and memory 903 (represented by memory 903). The bus architecture can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides an interface. The transceiver 901 can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium. The processor 902 is responsible for managing the bus architecture and general processing, and the memory 903 can store data used by the processor 902 during operation.
[0074] The electronic device 900 provided in this embodiment of the invention can execute the mine ventilation control method provided in the above-described method embodiment 1. To avoid repetition, it will not be described again here.
[0075] Example 4 Furthermore, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the mine ventilation control method provided in Embodiment 1.
[0076] In this embodiment, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.
[0077] The computer-readable storage medium provided in this embodiment can implement the mine ventilation control method provided in Embodiment 1. To avoid repetition, it will not be described again here.
[0078] In all examples shown and described herein, any specific values should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.
[0079] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0080] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.
Claims
1. A method for controlling ventilation in a mine, characterized in that, The method includes: Obtain the mine ventilation dataset; A multi-task learning and prediction network is constructed, comprising: an input layer, a shared feature layer, and a task decoding layer. The shared feature layer comprises: a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer comprises: a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. The multi-task learning prediction network is trained based on the mine ventilation dataset to obtain the target learning prediction network. An MPPO decision model is constructed, which includes a policy network and a value network, both of which adopt a dual-hidden-layer fully connected structure. Based on the mine ventilation dataset and the target learning prediction network, the MPPO decision model is trained to obtain the target MPPO decision model. Multiple environmental parameters of the mine to be ventilated are obtained, and each of the environmental parameters is input into the target MPPO decision model to obtain the ventilation parameter adjustment strategy; The ventilation of the mine to be ventilated is controlled according to the ventilation parameter adjustment strategy.
2. The mine ventilation control method according to claim 1, characterized in that, The acquisition of the mine ventilation dataset includes: A ventilation simulation model was constructed based on the pollutant exponential decay law. A jet velocity decay equation is constructed based on Prandtl's jet theory; Based on the ventilation simulation model and the jet velocity attenuation equation, the Latin hypercube sampling method is used to generate multiple sets of ventilation data within a preset parameter space. Physical consistency verification is performed on each of the ventilation data, and the ventilation data that pass the verification constitute the mine ventilation dataset.
3. The mine ventilation control method according to claim 2, characterized in that, The ventilation data includes: time elapsed since the blast, carbon monoxide concentration, ventilation duct installation distance, wind speed, fan flow rate, and fan power. The physical consistency verification of each of the ventilation data includes: Verify whether the negative correlation between the time elapsed after the explosion and the carbon monoxide concentration meets the first preset correlation requirement; Verify whether the negative correlation between the installation distance of the wind duct and the wind speed meets the second preset correlation requirement; Verify whether the positive correlation between the fan flow rate and the fan power meets the third preset correlation requirement; Verify whether the goodness of fit between the ventilation simulation model and the jet velocity attenuation equation for each ventilation data meets the preset fitting requirements; If all conditions are met, then the ventilation data verification is deemed successful.
4. The mine ventilation control method according to claim 1, characterized in that, The process of training the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network includes: Set the initial learning rate and weight decay coefficient of the AdamW optimizer, and define the loss function of the multi-task learning prediction network; Based on the mine ventilation dataset, the multi-task learning prediction network is trained for a preset number of rounds using the AdamW optimizer and the loss function. During the training process, the learning rate of the multi-task learning prediction network dynamically decays from the initial learning rate according to a preset rule. The multi-task learning prediction network after training is determined as the target learning prediction network.
5. The mine ventilation control method according to claim 4, characterized in that, The process of training the multi-task learning prediction network based on the mine ventilation dataset, using the AdamW optimizer and the loss function, includes: The mine ventilation dataset is divided into a training set, a validation set, and a test set according to a preset ratio. Based on the training set, the multi-task learning and prediction network is iteratively trained for a preset number of rounds. In each round of training, sample data is extracted from the training set according to a preset batch size and input into the multi-task learning and prediction network. The mixed loss value is calculated through forward propagation, and the network parameters of the multi-task learning and prediction network are updated by backpropagation in combination with the AdamW optimizer. After each preset number of training rounds, the performance of the trained multi-task learning prediction network is evaluated using the validation set. After training reaches the preset number of rounds, the network performance of the trained multi-task learning prediction network is verified using a test set to see if it meets the preset requirements.
6. The mine ventilation control method according to claim 1, characterized in that, The construction of the MPPO decision model includes: Define a Markov decision quintuple, which includes a state space, an action space, a state transition rule, a reward function, and a discount factor. The state transition rule is determined based on the target learning prediction network. Construct the policy network and the value network; By integrating the Markov decision quintuple, the multi-objective hierarchical reward function, the MPPO algorithm logic, the policy network, and the value network, the MPPO decision model is obtained. The MPPO algorithm logic uses generalized advantage to determine the action advantage value and uses a sigmoid function for dynamic weight adjustment.
7. The mine ventilation control method according to claim 6, characterized in that, The process of training the MPPO decision model based on the mine ventilation dataset and the target learning prediction network to obtain the target MPPO decision model includes: Based on the mine ventilation dataset and the target learning prediction network, a reinforcement learning training environment is constructed, and the reinforcement learning training environment outputs the state transition results through the target learning prediction network. Start training of the MPPO decision model. The training process is divided into an exploration phase, an improvement phase, and a stabilization phase, with each phase divided according to a preset number of steps. After each sampling step of training is completed, the parameters of the MPPO decision model are updated using the optimized number of rounds; The average reward and task success rate are monitored in real time during the training process. When the average reward reaches a preset reward threshold and the task success rate reaches a preset success rate threshold, the MPPO decision model is determined to have converged. The converged MPPO decision model is determined as the target MPPO decision model.
8. A mine ventilation control system, characterized in that, The system includes: The dataset building module is used to obtain mine ventilation datasets; A network construction module is used to construct a multi-task learning and prediction network. The multi-task learning and prediction network includes an input layer, a shared feature layer, and a task decoding layer. The shared feature layer includes a first fully connected layer, a second fully connected layer, and a third fully connected layer. The task decoding layer includes a CO concentration regression head, a dust concentration regression head, a wind speed regression head, and a safety classification head. The network training module is used to train the multi-task learning prediction network based on the mine ventilation dataset to obtain the target learning prediction network; The model building module is used to build an MPPO decision model, which includes a policy network and a value network, both of which adopt a double hidden layer fully connected structure. The model training module is used to train the MPPO decision model based on the mine ventilation dataset and the target learning prediction network to obtain the target MPPO decision model. The decision module is used to acquire multiple environmental parameters of the mine to be ventilated, input each of the environmental parameters into the target MPPO decision model, and obtain the ventilation parameter adjustment strategy. The ventilation control module is used to control the ventilation of the mine to be ventilated according to the ventilation parameter adjustment strategy.
9. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that, when executed on the processor, performs the mine ventilation control method according to any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the mine ventilation control method according to any one of claims 1-7.