Intelligent ventilation and poison filtration dynamic regulation and control method and system for mine

Abstract

The invention relates to the technical field of mine intelligent ventilation and poison filtration, and discloses a mine intelligent ventilation and poison filtration dynamic regulation and control method and system.The mine intelligent ventilation and poison filtration dynamic regulation and control method comprises the steps that a neural physical hybrid model is constructed, physical constraint conditions are combined with a deep neural network, and mine airflow field distribution is predicted; decomposing the neural physical hybrid model into three hierarchical sub-models, and deploying the three hierarchical sub-models to corresponding computing nodes to realize a hierarchical computing architecture; establishing a multi-level cooperation mechanism based on federal learning, and realizing distributed model training and optimization; constructing a task allocation decision model, and dynamically determining an optimal execution hierarchy of an airflow field calculation task according to task characteristics, resource states and real-time requirements; dynamically adjusting the precision and performance balance of airflow field calculation; according to the method, the key technical problems in the mine ventilation and poison filtration system are solved through the neural physical hybrid model, the multi-level distributed computing architecture and the federal learning collaboration mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent mine ventilation and gas filtration, and more specifically, it relates to a dynamic regulation method and system for intelligent mine ventilation and gas filtration. Background Art

[0002] The mine ventilation and gas filtration system is an important technical means to ensure the safe production of mines. Its main function is to control the concentration and distribution of harmful gases in the mine through reasonable ventilation design and gas filtration device configuration, ensuring the safety of the working environment. With the increase in the depth of mine exploitation and the improvement of exploitation intensity, the mine ventilation and gas filtration system is facing increasingly complex technical challenges.

[0003] Traditional mine ventilation and gas filtration systems mainly rely on empirical design and regular inspection and maintenance, lacking the ability of real-time dynamic regulation. In the system design stage, the steady-state analysis method is usually adopted. Based on the geometric structure of the mine, the expected gas generation amount, and ventilation requirements, a fixed ventilation network and gas filtration device configuration are designed. In the operation stage, the system mainly ensures normal operation through regular manual inspection and maintenance. The regulation means are relatively simple, mainly including operations such as adjusting the fan speed and opening and closing air doors.

[0004] In recent years, with the development of computer technology and sensor technology, the mine ventilation and gas filtration system has started to develop towards the intelligent direction. The existing intelligent ventilation systems mainly have the following problems: CFD-based numerical simulation technology: Using computational fluid dynamics methods to establish a mathematical model of the mine ventilation system, and simulating the airflow field distribution by solving control equations such as the Navier-Stokes equation. This method can provide high computational accuracy, but the computational complexity is high. A single simulation usually takes several hours or even several days, making it difficult to meet the requirements of real-time control; Sensor network-based monitoring technology: Deploying gas concentration sensors, wind speed sensors and other devices at key positions in the mine to monitor the mine environmental parameters in real time. This technology can provide real-time environmental information, but lacks the ability to predict future states and can only respond passively; Expert system-based decision-making technology: Encoding the experience and knowledge of mine ventilation experts into a rule base and making decisions based on current monitoring data and preset rules. This method has a fast response speed, but poor adaptability and is difficult to handle complex and changeable mine environments; Machine learning-based prediction technology: Using historical data to train a machine learning model to predict the change trend of gas concentration. This method has a certain prediction ability, but lacks physical constraints, and the prediction results may violate physical laws, with limited reliability. The main limitations of the existing technology are as follows: First, it is difficult to balance calculation accuracy and efficiency. High-precision CFD methods are time-consuming, while fast simplified methods lack accuracy. Second, the system architecture mostly adopts a central centralized design, with problems such as large communication delays, heavy computing burdens, and poor reliability. Third, there is a lack of an effective coordination mechanism for multi-level computing resources, and distributed computing capabilities cannot be fully utilized. Finally, it is difficult to dynamically balance calculation accuracy and response speed according to actual needs.

[0005] Therefore, there is an urgent need to develop a new dynamic regulation method for intelligent mine ventilation and gas filtration, which can improve calculation efficiency while ensuring calculation accuracy, make full use of multi-level distributed computing resources, achieve dynamic balance between accuracy and performance, and meet the actual needs of mine safety production. Summary of the Invention

[0006] The present invention provides a dynamic regulation method and system for intelligent mine ventilation and gas filtration, which solves the technical problems of the contradiction between the calculation efficiency and accuracy of the airflow field, the performance bottleneck of the central centralized architecture, the insufficient coordination of multi-level computing resources, and the difficulty in dynamically balancing accuracy and performance in related technologies.

[0007] The present invention provides a dynamic regulation method for intelligent mine ventilation and gas filtration, including: Construct a neuro-physical hybrid model, combine physical constraint conditions with a deep neural network to predict the distribution of the mine airflow field; Based on the neuro-physical hybrid model, decompose the neuro-physical hybrid model into three hierarchical sub-models and deploy them to corresponding computing nodes to achieve a hierarchical computing architecture; Based on the hierarchical computing architecture, establish a multi-level coordination mechanism based on federated learning, enabling each computing node to share model updates without sharing original data, and achieving distributed model training and optimization; Based on the resource status information of the multi-level coordination mechanism, construct a task allocation decision model, and dynamically determine the optimal execution level of the airflow field calculation task according to task characteristics, resource status, and real-time requirements; Based on the execution results of the task allocation decision model, dynamically adjust the accuracy and performance balance of the airflow field calculation according to the urgency and accuracy requirements of the ventilation and gas filtration decision, and achieve adaptive services in different scenarios.

[0008] Furthermore, the steps of constructing the neuro-physical hybrid model include: Construct physical constraint conditions describing the mine airflow field, where the physical constraint conditions are based on the basic equations of fluid mechanics and consider the special environment of the mine; Construct a deep neural network structure for airflow field prediction, including an input layer, an encoder layer, a physical information fusion layer, a decoder layer, and an output layer; Construct a composite loss function including a data fitting loss and a physical constraint loss; Based on computational fluid dynamics simulation data and measured data, a deep neural network is trained using a composite loss function to enable the model to simultaneously meet the requirements of data fitting and physical constraints.

[0009] Furthermore, the physical constraint conditions include: The mass conservation equation to ensure the incompressibility of the fluid; The momentum conservation equation to describe the relationship between the change in fluid velocity and the pressure gradient, viscous diffusion, and external force; The geometric constraints of the mine roadway, representing the flow velocity and pressure conditions of the solid wall, inlet, and outlet through boundary conditions.

[0010] Furthermore, the steps of decomposing the neuro-physical hybrid model into three hierarchical sub-models include: Based on the characteristics and dependencies of the computing tasks, the functions of the neuro-physical hybrid model are decomposed into an edge layer model, a fog layer model, and a cloud layer model; According to the physical layout of the mine and the structure of the ventilation system, a spatial partitioning strategy is formulated to determine the spatial scope responsible for each computing level; According to the hardware capabilities of the computing nodes at each level, the model is classified by complexity to make each hierarchical model adapt to the corresponding hardware capabilities.

[0011] Furthermore, the edge layer model is responsible for the rapid estimation of the local area airflow field with a spatial scope of less than 50 meters and is deployed on the edge computing devices at the monitoring points; The fog layer model is responsible for the fusion and coordination of the medium area airflow field with a spatial scope of 200 to 500 meters and is deployed on the computing devices at the regional control stations; The cloud layer model is responsible for the global airflow field optimization and long-term prediction and is deployed on the high-performance servers or cloud computing platforms in the central control room.

[0012] Furthermore, the steps of establishing a multi-level collaborative mechanism based on federated learning include: Construct a three-layer federated learning architecture, including edge layer federated learning, fog layer federated learning, and cloud layer global coordination; Establish a local model update mechanism to enable each computing node to update the model parameters based on local data; Establish a model parameter aggregation strategy, and calculate the weighted average value based on the data volume and data quality of each node to form an updated global model; Establish a cross-level model synchronization mechanism to achieve the synchronization and sharing of model parameters between different levels.

[0013] Furthermore, the model parameter aggregation strategy includes: Calculate the aggregation weight based on the data volume of each node, sensor confidence, data coefficient of variation, and historical accuracy; Introduce a parameter anomaly detection mechanism to identify and filter existing malicious and abnormal updates, ensuring the correctness of the aggregation results.

[0014] Further, the steps of constructing the task assignment decision model include: Analyze the characteristics of the airflow field calculation task, including the spatio-temporal range, real-time requirements, accuracy requirements, and computational complexity; Real-time monitor the resource status of each computing level, including computing resources, communication resources, and power status; Construct a comprehensive cost function considering computing cost, communication cost, latency cost, and energy consumption cost; Based on the comprehensive cost function, construct a task assignment decision function, and select the computing level with the minimum comprehensive cost as the task execution level.

[0015] Further, the steps of dynamically adjusting the accuracy and performance balance of the airflow field calculation include: Evaluate the urgency of the current mine environment, including the concentration and change rate of harmful gases, personnel distribution, ventilation equipment status, and abnormal event detection; Based on the scenario urgency, set the accuracy configuration parameters for the airflow field calculation, including spatial resolution, time step, iterative convergence threshold, and model complexity; Implement an asynchronous collaborative computing and result update mechanism between different levels, enabling the edge layer to quickly generate preliminary results based on the local model, and the fog layer and cloud layer to perform parallel computing for more accurate results; Based on the airflow field calculation results, generate and execute ventilation and gas filtration decisions, and simultaneously monitor the execution effect in real time.

[0016] The present invention provides a mine intelligent ventilation and gas filtration dynamic regulation system for implementing the above-mentioned mine intelligent ventilation and gas filtration dynamic regulation method, including: A neuro-physical hybrid modeling module for constructing an airflow field prediction model that integrates physical constraints and a deep neural network; A hierarchical computing architecture module for decomposing the hybrid model into multi-level sub-models and deploying them to corresponding computing nodes; A federated learning collaboration module for establishing a multi-level collaboration mechanism to achieve model parameter sharing and distributed optimization between computing nodes; An adaptive task scheduling module for constructing a task assignment decision model and dynamically allocating computing tasks according to task characteristics and resource status; An accuracy-performance balance module for dynamically adjusting the accuracy and performance balance strategy according to the scenario urgency and accuracy requirements.

[0017] The beneficial effects of the present invention are as follows: Achieve an effective balance between computational efficiency and accuracy: The neuro-physical hybrid model combines the interpretability of physical constraints and the learning ability of neural networks, significantly improving the computational speed compared to traditional CFD methods while maintaining high prediction accuracy, meeting the requirements of real-time ventilation and gas filtration control; Reduce communication burden and system latency: Adopting an edge-fog-cloud three-layer computing architecture and a federated learning mechanism reduces the amount of data transmission and communication latency, shortening the system response time from seconds to milliseconds; Enhance system robustness and reliability: The multi-layer distributed architecture eliminates the risk of single-point failures. Even if some devices fail, the system can still maintain basic functions, improving system availability and reliability; Optimize the utilization of computing resources: The adaptive task scheduling mechanism makes full use of the computing resources at all levels of the mine, improving resource utilization efficiency, achieving dynamic balance of computing loads, and avoiding local resource overload or idleness; Improve decision-making quality and adaptability: The accuracy-performance dynamic balance mechanism enables the system to adaptively adjust the computing strategy according to the urgency of the scenario, ensuring accuracy in normal monitoring states and prioritizing response speed in emergencies, making the ventilation and gas filtration strategy more precise and effective; Reduce energy consumption and improve processing efficiency: Through intelligent dynamic regulation, the system reduces energy consumption, improves the processing efficiency of harmful gases, and improves the mine safety production environment; Enhance predictive control capabilities: Based on the prediction of the future evolution of the airflow field, the system can make control adjustments in advance, avoiding the lag effect of passive responses and achieving preventive control, which is particularly suitable for scenarios with response lags such as large fan speed regulation. Brief Description of the Drawings

[0018] Figure 1 is the flowchart of the intelligent ventilation and gas filtration dynamic regulation method for mines in the present invention; Figure 2 is the bar chart of the comparison results of the calculation times of the method of the present invention, traditional CFD methods, and simplified analytical models under different tasks; Figure 3 is the line chart of the changing trend of the prediction accuracy of the models at the edge layer, fog layer, and cloud layer over time; Figure 4 is the scatter chart of the relationship between the computing load and response time of the method of the present invention; Figure 5 is the radar chart of the comparison of multiple performance indicators between the system of the present invention and traditional systems; Figure 6 is the area chart of the computing resource allocation ratios at the edge layer, fog layer, and cloud layer under scenarios of different urgencies. Detailed Embodiments

[0019] Reference will now be made to example embodiments to discuss the subject matter described herein. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the content of this specification. Each example may omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples may also be combined in other examples.

[0020] In at least one embodiment of the present invention, a method for dynamically regulating intelligent ventilation and gas filtration in a mine is disclosed. As Figure 1 shown, it includes: Step 1, construct a neuro-physical hybrid model, combine physical constraint conditions with a deep neural network to predict the distribution of the mine airflow field; This step uses the neuro-physical hybrid modeling method to analyze mine airflow field data and physical laws, and generate an airflow field prediction model with both high efficiency and high accuracy. Specifically, it includes the following sub-steps: Step 1.1, physical constraint construction; Construct physical constraint conditions describing the mine airflow field, mainly based on the Navier-Stokes equations. This system of equations includes the momentum conservation equation and the mass conservation equation. The momentum conservation equation describes the relationship between the change of fluid velocity over time, convection, pressure gradient, viscous diffusion, and external force, while the mass conservation equation ensures the incompressibility of the fluid.

[0021] In view of the particularity of the mine environment, the standard equations are simplified and corrected, and factors such as the geometric constraints of the mine roadway, the fan characteristic curve, and the resistance characteristic of the air door are additionally considered to form physical constraint conditions suitable for the mine environment. Specifically, the geometric constraints of the mine roadway are represented by boundary conditions: the flow velocity is zero on the solid wall surface (no-slip condition), the flow velocity and pressure at the air inlet are determined according to the fan characteristic curve, and the pressure at the air outlet satisfies the atmospheric pressure condition. The resistance characteristic of the air door is represented by the local resistance coefficient, and its value has a non-linear relationship with the opening of the air door. In practical applications, the mine roadway network can be simplified to a one-dimensional pipe network model for preliminary calculation, and a three-dimensional model is used for fine simulation in key areas.

[0022] Optionally, in some embodiments with complex mine working conditions, the coupled equations of the temperature field and the concentration field can be introduced to consider the effects of thermodynamic effects and gas diffusion effects on the airflow field. For example, in high-temperature areas or high-gas areas, the energy conservation equation and the component conservation equation can be added to form a more complete physical constraint system. The energy conservation equation describes the relationship between temperature change over time, convection, heat diffusion, and heat sources, while the component conservation equation describes the relationship between gas concentration change over time, convection, mass diffusion, and gas sources. These equations together constitute a complete physical constraint system for describing airflow, heat, and gas diffusion in a complex mine environment.

[0023] According to the embodiments of the present application, in some other embodiments, for edge devices with very limited computing resources, the physical constraints can be further simplified, and a quasi-one-dimensional or network model is used to describe the mine ventilation system, simplifying complex partial differential equations into ordinary differential equation systems, greatly reducing the computational complexity while maintaining a reasonable description of the main physical phenomena.

[0024] Step 1.2, Deep neural network structure design; Construct a deep neural network structure for airflow field prediction, adopting an encoder-decoder architecture, specifically including: Input layer: Receive spatial coordinates , time and boundary condition parameters (such as fan speed, damper opening, etc.) as inputs; Encoder layer: Consists of multiple layers of fully connected neural networks, extracting abstract representations of input features layer by layer; Physical information fusion layer: Fuse physical constraint information with intermediate neural network features; Decoder layer: Consists of multiple layers of fully connected neural networks, mapping the fused features to the target output; Output layer: Predict the fluid state variables at a given spatio-temporal position, including velocity components and pressure .

[0025] According to an embodiment of the present application, the encoder layer specifically includes 4 fully connected layers, each layer containing 128, 256, 512, and 512 neurons respectively, and uses LeakyReLU as the activation function. The physical information fusion layer adopts an attention mechanism to convert physical constraint information into attention weights and perform weighted fusion with the features output by the encoder. The decoder layer includes 4 fully connected layers, each layer containing 512, 256, 128, and 64 neurons respectively, and also uses LeakyReLU (Leaky Rectified Linear Unit) as the activation function. It should be noted that in the applications in different regions of the mine, the width and depth of the network structure can be adjusted according to the regional complexity. For example, in complex regions such as the mining face, the structure can be increased to 6 layers to improve the model's expressive ability.

[0026] When applied to the main ventilation roadway of the mine, the input layer receives spatial coordinates , time , the rotational speed of the main ventilator, the opening degree of the adjacent air door, etc. as input features; when applied to the mining face, additional features such as the advancing speed of the working face and the gas content of the coal seam are added as input parameters to improve the prediction accuracy.

[0027] Optionally, when dealing with more complex non-linear fluid dynamics problems, a convolutional neural network (Convolutional Neural Network, CNN) or a graph neural network (Graph Neural Network, GNN) structure can be used to replace the fully connected network. For example, for the mine roadway with a complex spatial structure, the roadway network can be represented as a graph structure, each node represents a roadway intersection or a key position, and the edge represents the connecting roadway. A graph convolutional network is used to process this non-Euclidean spatial structure to better capture the spatial topological relationship.

[0028] In some embodiments, to handle the temporal dynamic characteristics, a recurrent neural network (Recurrent Neural Network, RNN) unit, such as a long short-term memory (Long Short-Term Memory, LSTM) or a gated recurrent unit (Gated Recurrent Unit, GRU), can also be introduced into the network structure to capture the dynamic characteristics of the airflow field evolving over time. In addition, this spatio-temporal hybrid neural network structure is particularly suitable for predicting the diffusion process of gas after an emergency (such as a gas outburst).

[0029] Step 1.3, construction of the physical constraint loss function; Construct a composite loss function containing a data fitting term and a physical constraint term for neural network training. It should be noted that the expression of this loss function is: ; Where, represents the total loss function value; represents the weight coefficient of the data fitting loss; represents the data fitting loss, which measures the difference between the predicted value of the model and the measured data; represents the weight coefficient of the physical constraint loss; represents the physical constraint loss, which measures the degree to which the predicted result of the model violates the physical laws.

[0030] According to the embodiments of the present application, the data fitting loss can be measured by calculating the average of the squared differences between the predicted value and the measured value. The physical constraint loss is measured by calculating the sum of the squared residuals after substituting the predicted result into the physical equation. The smaller the residual, the more the predicted result conforms to the physical laws.

[0031] In actual implementation, the physical constraint loss function is further divided into three parts: mass conservation loss, momentum conservation loss, and energy conservation loss, which respectively correspond to the three basic conservation laws in fluid mechanics. For the special environment of the mine, a boundary condition loss is also added to ensure that the predicted result meets the geometric constraints of the mine roadway and the equipment characteristic constraints. It should be noted that the weight coefficients of each part of the loss function can be dynamically adjusted according to the specific application scenario. For example, in the high gas area, the weight of the mass conservation loss will increase accordingly to ensure the accuracy of gas concentration prediction.

[0032] Optionally, to enhance the sensitivity of the model to different types of errors, the data fitting loss can adopt a more complex form, such as Huber loss or Log-Cosh loss, to reduce the impact of outliers on training. In addition, to improve the prediction accuracy of the model for key regions and dangerous situations, a weighted sampling strategy can be introduced to assign higher weights to samples in high-risk regions (such as near the mining face) or abnormal states (such as sudden increase in gas concentration), so that the model pays more attention to these key situations during training.

[0033] Step 1.4, hybrid model training and optimization; Based on the CFD (Computational Fluid Dynamics) simulation data and the measured data, train the neuro-physical hybrid model using the constructed loss function. The training process adopts an adaptive learning rate optimization algorithm to reduce the composite loss function value through iterative optimization, so that the model meets both the data fitting and physical constraint requirements.

[0034] During training, a dynamic weight adjustment strategy is adopted. Increase the weight of the data fitting loss at the initial stage of training , to promote the model to quickly approach the data distribution; increase the weight of the physical constraint loss at the later stage of training , to ensure that the predicted result of the model conforms to the physical laws.

[0035] Specifically, at the initial stage of training (the first 30% of the training cycles), set , ; at the middle stage of training (30% to 70% of the training cycles), set , ; at the final stage of training (the last 30% of the training cycles), set , . This dynamic adjustment strategy ensures that the model training can balance data fitting and physical constraints.

[0036] In addition, model pruning and knowledge distillation techniques are used to compress and optimize the trained model, reduce the model's computational complexity, and improve the inference efficiency to meet the subsequent deployment requirements on edge devices. During the model pruning process, an importance scoring method is used to identify neuron connections with less contribution, and after removing these connections, fine-tuning is performed to maintain the model performance. Knowledge distillation transfers the knowledge of a complex model to a lightweight model by training a small student network to mimic the behavior of a large teacher network.

[0037] Optionally, in some implementation manners with higher requirements for the model convergence speed, a transfer learning strategy can be adopted. First, pre-train a basic model on a large amount of CFD simulation data, and then fine-tune it on the measured data of a specific mine to accelerate the model convergence speed and reduce the required amount of measured data. This method is particularly suitable for newly built mines or scenarios after the transformation of the ventilation system, and can quickly build an effective model when the data accumulation is insufficient.

[0038] In some other implementation manners, to improve the generalization ability and robustness of the model, data augmentation techniques can be introduced to expand the training dataset by adding noise, geometric transformation, parameter perturbation, etc. For example, adding random perturbations that conform to the actual noise distribution to the sensor data to simulate sensor drift and measurement errors; making small random adjustments to the ventilation system parameters (such as fan speed, damper opening) to enhance the model's adaptability to parameter changes. These data augmentation techniques can improve the stability and reliability of the model in the actual complex environment.

[0039] Step 2: Based on the neuro-physical hybrid model, decompose the neuro-physical hybrid model into three hierarchical sub-models and deploy them to the corresponding computing nodes to implement a hierarchical computing architecture; This step decomposes the neuro-physical hybrid model constructed in Step 1 into sub-models suitable for different computing levels and deploys them to the corresponding computing nodes. It specifically includes the following sub-steps: Step 2.1: Model function decomposition; Based on the characteristics and dependencies of the computing tasks, decompose the functions of the neuro-physical hybrid model into the following three levels: Edge layer model: Responsible for the rapid estimation of the airflow field in a local small area (less than 50 meters), mainly dealing with tasks with high real-time requirements and low computational complexity, such as predicting the airflow state in a small area around a specific monitoring point; Fog layer model: Responsible for the fusion and coordination of the airflow field in a medium area (200 to 500 meters), dealing with tasks with medium complexity and medium real-time requirements, such as the overall estimation and short-term prediction of the airflow field within the working area; Cloud layer model: Responsible for the global optimization of the airflow field and long-term prediction, dealing with tasks with high complexity, high accuracy requirements but relatively low real-time requirements, such as simulating and optimizing the airflow field in the entire mine.

[0040] Specifically, when implemented, the edge layer model focuses on tasks such as predicting gas diffusion within 10 meters around a gas sensor and analyzing airflow changes near a ventilation door, with a response time within 100 milliseconds; the fog layer model is responsible for calculating the airflow field of the entire mining area or a branch of the ventilation system, such as evaluating the overall ventilation effect of a coal mining face, with a response time of about 1 second; the cloud layer model deals with complex tasks such as optimizing the entire mine ventilation network and comparing and evaluating different ventilation schemes, with an acceptable response time of more than 10 seconds.

[0041] Optionally, in some special mine environments, the hierarchical structure can be adjusted according to actual needs. For example, in a small mine or a mine with a relatively simple ventilation system, a two-layer architecture can be adopted, only retaining the edge layer and the cloud layer to simplify the system structure; in an extra-large mine or a mine with an extremely complex ventilation system, a four-layer architecture can be introduced, adding a super cloud layer above the cloud layer to be responsible for collaborative optimization between multiple mines, or adding a deep edge layer below the edge layer, deployed on micro-devices closer to the monitoring points to further improve the local response speed.

[0042] In some other implementation manners, the model function decomposition can adopt a task-driven method and be specifically optimized according to different types of prediction tasks. For example, dedicated models can be constructed respectively for different tasks such as gas concentration prediction, wind speed prediction, and pressure distribution prediction. Each model focuses on a specific prediction target to improve the prediction accuracy and efficiency of specific tasks.

[0043] Step 2.2, formulating a spatial partitioning strategy; According to the physical layout of the mine and the structure of the ventilation system, formulate a spatial partitioning strategy to determine the spatial scope responsible for each calculation level: Edge layer: Centered on each monitoring point, covering a local area with a radius of about 10 to 30 meters; Fog layer: Covering a complete functional area, such as a mining area or a group of related roadways, with an area about 10 to 100 times that of the edge layer; Cloud layer: Covering the entire mine area.

[0044] The partitioning strategy ensures an appropriate overlapping area between different levels, facilitating boundary data exchange and result fusion.

[0045] In an actual mine, taking a coal mine as an example, the edge layer model usually covers the area around a single sensor node, such as the area around a certain gas sensor in the return airway; the fog layer model covers a complete coal mining face and its intake and return airways, with an area of about 200 to 500 square meters; the cloud layer model covers all the roadways and working areas of the entire mine, reaching thousands to tens of thousands of square meters. The overlapping area between different levels is usually set to a range of 5 to 10 meters outside the edge area to ensure a smooth transition of the calculation results of different levels.

[0046] Optionally, in a mine with complex and variable airflows, the spatial partitioning can adopt an adaptive dynamic partitioning strategy, dynamically adjusting the partitioning boundary according to the real-time change characteristics of the airflow field. For example, when a sharp change in gas concentration is detected in a certain area, the coverage range of the edge layer model in this area is automatically expanded to improve the local prediction accuracy; when the airflow fields of multiple adjacent areas show strong correlation, the fog layer models of these areas are automatically merged to ensure integrated processing and improve the prediction consistency.

[0047] In some embodiments, the spatial partitioning can also be combined with the time dimension to form a spatio-temporal partitioning strategy. For short-term predictions (such as the change of the airflow field within the next few minutes), mainly rely on the edge layer model; for medium-term predictions (such as the change trend within the next few hours), mainly rely on the fog layer model; for long-term predictions (such as the evolution law within the next day or longer), mainly rely on the cloud layer model. This spatio-temporal joint partitioning strategy can more efficiently allocate computing resources to ensure the accuracy and real-time performance of various prediction tasks.

[0048] Step 2.3, grading the computational complexity of the model; Grade the model according to the hardware capabilities of the computing nodes at each level: Edge layer model: Adopt a lightweight network structure, reduce the number of network layers and parameters, and focus on retaining network components sensitive to local features; Fog layer model: Adopt a network structure with medium complexity to balance computational efficiency and prediction accuracy; Cloud layer model: Retain the complete network structure, focusing on high accuracy and global consistency.

[0049] In specific implementation, the edge layer model usually adopts a 2- to 3-layer structure, with the number of neurons in each layer not exceeding 64, and the total amount of model parameters controlled within 50KB to adapt to the storage and computing power limitations of edge devices; the fog layer model adopts a 3- to 4-layer structure, with the number of neurons in each layer between 64 and 256, and the total amount of model parameters controlled within 500KB; the cloud layer model retains a complete 4- to 6-layer structure, with the number of neurons in each layer reaching 256 to 512, and the total amount of model parameters reaching 2MB. The edge layer model focuses on retaining network components sensitive to local gas concentration and wind speed characteristics, while the cloud layer model enhances the perception ability of the global ventilation network topology.

[0050] Optionally, on edge devices with particularly limited hardware resources, binary neural network or quantized neural network technologies can be adopted to convert model parameters and calculation processes from floating-point numbers to integers or even binary values, greatly reducing the computing and storage requirements. For example, a binary neural network limits weights and activation values to two values, +1 and -1, which can compress the model size to 1 / 32 of the original, while increasing the computing speed, and is particularly suitable for deployment on low-power microcontrollers.

[0051] Step 2.4, hierarchical model deployment; Deploy the decomposed model to the corresponding computing nodes: The edge layer model is deployed to edge computing devices at each monitoring point, such as intelligent sensor nodes, edge gateways, etc.; The fog layer model is deployed to computing devices at the regional control station, such as industrial control computers, edge servers, etc.; The cloud layer model is deployed to high-performance servers or cloud computing platforms in the central control room.

[0052] During the deployment process, necessary format conversion and optimization are performed on the model according to the characteristics of different hardware platforms to ensure that the model can run efficiently.

[0053] Optionally, in some application scenarios that require high reliability, a redundant deployment strategy can be adopted, where the same model is deployed and run in parallel on multiple devices, and the final output result is determined through a voting mechanism to improve the fault tolerance and reliability of the system. For example, in the critical mining face area, the same model can be deployed on multiple edge nodes. When one node fails, other nodes can immediately take over the prediction task to ensure the continuous operation of the system.

[0054] In some other implementation manners, to adapt to the network instability problem in the mine environment, a progressive deployment and update strategy can be adopted, where the model is split into a core part and an extended part. First, the core function is deployed to ensure the basic prediction ability, and then the extended function is gradually deployed when network conditions permit, improving the adaptability and stability of the system.

[0055] Step 3, based on the hierarchical computing architecture, establish a multi-level collaborative mechanism based on federated learning, enabling each computing node to share model updates without sharing the original data, and achieving distributed model training and optimization; This step establishes a multi-level collaborative mechanism based on federated learning, enabling each computing node to share model updates without the need to share the original data, and achieving distributed model training and optimization. It specifically includes the following sub-steps: Step 3.1, Construction of the federated learning architecture; Construct a three-layer federated learning architecture, including edge-layer federated learning, fog-layer federated learning, and cloud-layer global coordination: Edge-layer federated learning: Multiple edge nodes under the jurisdiction of the same fog-layer node form a federated learning group to jointly optimize the model in this area; Fog-layer federated learning: Multiple fog-layer nodes form a higher-level federated learning group to coordinate and optimize the model in a larger area; Cloud-layer global coordination: The cloud-layer node aggregates the model updates of all fog-layers to complete the optimization of the global model.

[0056] In actual mine applications, the federated learning architecture is organized according to the management level and physical layout of the mine. Taking a coal mine as an example, multiple intelligent sensors in a mining area form an edge-layer federated learning group, which is coordinated by the control station of this mining area; the control stations of multiple mining areas in the mine form a fog-layer federated learning group, which is coordinated by the central control system of the mine; the central control systems of multiple mines can further form a higher-level federated learning network, which is coordinated by the group control center to achieve cross-mine knowledge sharing and model optimization.

[0057] Optionally, in a mine environment with complex and changeable network conditions, a federated learning architecture with an adaptive topology structure can be adopted to dynamically adjust the organizational structure of federated learning according to network connection quality, computing node status, and data distribution characteristics. For example, when it is detected that the network connection quality between certain nodes decreases, the division of the federated learning group is automatically adjusted to ensure stable and reliable network connections within the group; when certain nodes have particularly valuable data (such as abnormal event data), the participation and weight of these nodes in federated learning are increased.

[0058] In some embodiments, the federated learning architecture can adopt a hierarchical asynchronous update mechanism, where different levels update and synchronize the model at different frequencies. For example, edge-layer nodes may update the local model once an hour, fog-layer nodes update the regional model once a day, and cloud-layer nodes update the global model once a week. This hierarchical asynchronous mechanism can balance communication overhead and the timeliness of model updates, adapting to the network constraint conditions in the mine environment.

[0059] Step 3.2, Local model update mechanism; Mechanism for each computing node to update model parameters based on local data: The edge node collects local sensing data, combines physical constraints, and calculates the gradient of the local model parameters; Local update adopts the mini-batch stochastic gradient descent algorithm to update the local model parameters. According to the embodiments of the present application, the update process can be expressed as: ; Among them, represents the model parameters of node at the th round; represents the model parameters of node at the th round; represents the learning rate; represents the gradient operator; represents the value of the loss function calculated based on the local data of node ; represents the gradient calculated based on the local data.

[0060] Optionally, on edge devices with limited computing resources, a model partial update strategy can be adopted, where only some of the model parameters are updated each time, rather than all of them. For example, a block update method can be used, where the model parameters are divided into multiple blocks, and only one block is updated each time, and the process is repeated; or an importance update method can be used, where according to the importance of the parameters to the current prediction task, important parameters are updated first, and secondary parameters are updated later. These strategies can reduce the computational complexity of a single update, enabling edge devices to continuously optimize the model with limited computing resources.

[0061] In some other embodiments, local model update can adopt an incremental learning strategy, focusing on the differences between newly collected data and historical data, and only updating the parameters related to these differences in the model, avoiding unnecessary repeated learning of well-learned knowledge, and improving learning efficiency and model stability.

[0062] Step 3.3, model parameter aggregation strategy; How the model parameters of multiple nodes are aggregated to form an updated global model: Weighted average aggregation: Calculate the weighted average according to the data volume and data quality of each node. It can be seen that the aggregation process can be expressed as: ; Among them, represents the aggregated global model parameters; represents the summation symbol; represents the weighted coefficient of node ; represents node The effective data volume; Represents the total data volume; Represents the node In the Model parameters in the round; Represents the number of nodes participating in the aggregation.

[0063] Anomaly detection and defense: Introduce a parameter anomaly detection mechanism to identify and filter existing malicious and abnormal updates, ensuring the correctness of the aggregation result.

[0064] Optionally, during the model aggregation process, a more complex aggregation strategy can be adopted, such as adaptive aggregation based on model performance. Specifically, before each round of aggregation, use the validation dataset to evaluate the performance of each node's model, and dynamically adjust the aggregation weights according to the performance metrics, so that models with better performance have a higher influence in the aggregation. The performance evaluation can be based on a comprehensive score of metrics such as prediction accuracy and physical constraint satisfaction.

[0065] In some embodiments, to address the problem of possible uneven data distribution in the mine environment, a fair aggregation strategy can be adopted to ensure that different types of data (such as normal working condition data and abnormal working condition data) are equally emphasized in model training. For example, for nodes with a large amount of normal working condition data but very little abnormal working condition data, their weights can be appropriately reduced during aggregation to avoid the model overfitting common scenarios and ignoring key abnormal scenarios.

[0066] Step 3.4, cross-level model synchronization; How to synchronize and share model parameters between different levels: Bottom-up transmission: The model updates of lower-level nodes are regularly transmitted to upper-level nodes for aggregation; Top-down distribution: The optimized global model of upper-level nodes is regularly distributed to lower-level nodes as the basis for further optimization of lower-level nodes; Differentiated update: Only transmit the changed parts of the model parameters to reduce communication overhead.

[0067] The synchronization process adopts an asynchronous communication mechanism to avoid performance bottlenecks caused by waiting for synchronization.

[0068] Optionally, in the mine environment with unstable network conditions, an opportunistic synchronization strategy can be adopted. According to the network quality and bandwidth status, dynamically determine when to perform model synchronization. For example, when detecting good network quality, actively initiate model synchronization; when the network quality is poor, delay non-critical synchronization operations and wait until the network conditions improve before proceeding. This strategy can avoid a large amount of data transmission during network congestion, reducing the communication failure rate and network burden.

[0069] In some embodiments, to address possible network interruptions in the mine, a model incremental synchronization mechanism can be implemented, which decomposes a large-scale synchronization into multiple small-scale incremental synchronizations, and only transmits a part of the model parameters each time. Even if a network interruption occurs during the synchronization process, it is only necessary to continue from the breakpoint, rather than restarting the entire synchronization process. This incremental synchronization mechanism greatly improves the reliability and robustness of model synchronization, and is particularly suitable for environments with unstable network conditions such as mines.

[0070] Step 4: Based on the resource status information of the multi-level collaboration mechanism, construct a task allocation decision model, and dynamically determine the optimal execution level of the airflow field calculation task according to the task characteristics, resource status, and real-time requirements. This step realizes dynamically determining the optimal execution level of the airflow field calculation task according to the task characteristics, resource status, and real-time requirements, and improves the overall computing efficiency of the system. It specifically includes the following sub-steps: Step 4.1: Task characteristic analysis; Analyze the characteristics of the airflow field calculation task, including: Spatial and temporal scope: The size of the spatial scope and the time prediction length involved in the task; Real-time requirements: The timeliness requirements of the task result, that is, the maximum acceptable delay from the request being sent to obtaining the result; Accuracy requirements: The requirements for the accuracy of the calculation result, usually related to the application scenario of the task; Computational complexity: The estimated computational resources required to complete the task.

[0071] During the task characteristic analysis process, the system will establish a task characteristic database based on historical task execution data to automatically classify and identify the requirements of new tasks. In particular, for tasks triggered by emergencies (such as abnormal increase in gas concentration), the system will automatically increase their real-time requirement level to ensure a quick response.

[0072] Optionally, in some complex and changeable mine environments, the task characteristic analysis can adopt a finer-grained classification method, dividing the task characteristic space into multiple regions, and each region corresponds to a different processing strategy. For example, the spatial and temporal scope can be further divided into four categories: small-scale short-term, small-scale long-term, large-scale short-term, and large-scale long-term; the real-time requirements can be divided into four levels: extremely fast response (<100 milliseconds), fast response (<1 second), normal response (<10 seconds), and batch processing (>10 seconds); the accuracy requirements can be divided into four grades: low accuracy (error <20%), medium accuracy (error <10%), high accuracy (error <5%), and ultra-high accuracy (error <1%). This finer-grained classification can more precisely match task requirements and system resources, and further improve resource utilization efficiency.

[0073] In some other embodiments, task characteristic analysis can combine the mine safety status and production status, and dynamically adjust the task priority and resource allocation strategy according to the overall operation status of the current mine. For example, in the normal production state, energy consumption optimization tasks can obtain a higher priority; in the safety alert state, the priority of gas concentration monitoring and warning tasks is greatly improved; in the emergency evacuation state, evacuation route planning and gas diffusion prediction tasks obtain the highest priority. This state-aware task characteristic analysis mechanism enables the system to flexibly adjust the resource allocation strategy according to actual needs, ensuring the smooth execution of critical tasks.

[0074] Step 4.2, resource status monitoring; Monitor the resource status of each computing level in real time, including: Computing resources: CPU / GPU utilization rate, memory usage, etc.; Communication resources: network bandwidth, latency, congestion level, etc.; Power status: For battery-powered edge devices, the remaining battery power also needs to be monitored.

[0075] In actual implementation, the system adopts a distributed resource monitoring architecture, and resource monitoring agent programs are deployed on each node to collect the local resource status regularly (usually every 5 to 10 seconds) and report it. To reduce communication overhead, reporting is only triggered when the resource status change exceeds a preset threshold (such as the CPU utilization rate change exceeds 10%). The resource status information of each level is summarized to form a system resource status map for task allocation decision-making.

[0076] For the special mine environment, resource monitoring also includes equipment health status monitoring, such as sensor drift detection, communication link quality assessment, etc. The system will automatically reduce the task allocation priority of abnormally operating devices to avoid assigning critical tasks to unstable nodes. The power status monitoring of edge nodes is particularly important. When the detected battery power is lower than 20%, the system will automatically reduce the computing tasks assigned to this node to give priority to ensuring the basic monitoring function.

[0077] Optionally, in a mine environment with large fluctuations in resource status, a predictive resource monitoring mechanism can be adopted to predict the available resource situation in the future period based on the historical resource usage pattern. For example, through time series analysis methods, according to the resource usage trend in the past few hours, predict the resource load change in the next few tens of minutes, and make preparations for task scheduling in advance to avoid allocating a large number of computing tasks when resources are tense, resulting in a decline in system performance.

[0078] In some embodiments, resource status monitoring can be combined with environmental perception functions to consider the potential impact of external environmental factors on computing resources. For example, computing nodes in high-temperature areas may face heat dissipation challenges, and the system will actively reduce their resource utilization rate to avoid performance degradation or equipment damage caused by overheating; communication devices in high-humidity areas may face signal attenuation problems, and the system will accordingly adjust the communication resource evaluation results to avoid arranging a large number of data transmission tasks when the communication conditions deteriorate.

[0079] Step 4.3, constructing a task allocation decision model; Construct a task allocation decision model to determine the optimal execution level of the task based on task characteristics and resource status. The decision model uses a comprehensive cost function. It should be noted that this function can be expressed as: ; Where, represents the total cost function value; represents task characteristics; represents resource status; represents the computing level (edge, fog, or cloud); represents the weight coefficient of the computing cost; represents the computing cost function value; represents the weight coefficient of the communication cost; represents the communication cost function value; represents the weight coefficient of the latency cost; represents the latency cost function value; represents the weight coefficient of the energy consumption cost; represents the energy consumption cost function value.

[0080] According to the embodiments of the present application, the task allocation decision function is to select the computing level that minimizes the comprehensive cost as the task execution level.

[0081] In specific implementations, each cost function is modeled according to the actual situation of the mine: Computing cost function value : Consider the matching degree between the task computing volume and the computing capacity at the level, as well as the current computing load at the level; Communication cost function value : Consider the relationship between the task input data volume, result data volume and the network bandwidth and latency at the level; Latency cost function value : Consider the matching degree between the estimated time for the task to be executed at the level and the timeliness requirements of the task; Energy consumption cost function value : Consider the task at the level The energy consumption during execution is particularly important for battery-powered edge devices.

[0082] The weight coefficients are dynamically adjusted according to the mine operation mode: in the normal monitoring mode, the system balances various factors; in the emergency event handling mode, the weight of the latency cost is significantly increased, and the response speed is given priority; in the low-power mode, the weight of the energy consumption cost is increased to reduce the overall energy consumption of the system.

[0083] Optionally, in a mine where the computing environment is complex and changeable, the decision-making model can be adaptively optimized using reinforcement learning methods. The system records each task allocation decision and its execution results (such as actual completion time, resource consumption, computing accuracy, etc.) as training data to continuously optimize the decision-making model. Therefore, through the reward mechanism, the reinforcement learning algorithm encourages the system to make decisions that can balance performance, energy consumption, and reliability. As the system runs for a longer time, the performance of the decision-making model continuously improves and gradually adapts to the operating characteristics of a specific mine.

[0084] In some embodiments, to handle the problems of sudden task occurrences and resource fluctuations that may occur in the mine, the decision-making model can incorporate a risk assessment function to calculate a risk index for each possible allocation decision, avoiding making decisions that may cause system instability. Additionally, the system will avoid allocating a large number of computationally intensive tasks to the same computing node, even if the node currently has sufficient resources, because this may cause the node to run out of resources in a short period of time; the system will also avoid allocating critical tasks to nodes with unstable communication conditions, even if the node has strong computing capabilities, because this may cause the results to not be transmitted back in a timely manner.

[0085] Step 4.4, task execution and result aggregation; Execute the task according to the allocation decision and aggregate the results: When the task is allocated to the edge layer for execution, the relevant edge nodes independently complete the calculation and upload the results to the fog layer or directly provide them to the user; When the task is allocated to the fog layer for execution, the fog layer nodes coordinate the relevant edge nodes to provide the necessary input data, and after completing the calculation, upload the results to the cloud layer or provide them to the user; When the task is allocated to the cloud layer for execution, the cloud layer nodes coordinate to obtain the necessary input data, and after completing the calculation, distribute the results to the relevant fog layer and edge layer nodes, or provide them to the user.

[0086] If the task needs to be completed through cross-layer collaboration, the subtask decomposition strategy is adopted to decompose the original task into subtasks suitable for different layers to execute, and then summarize the results after separate execution.

[0087] In actual mine applications, the task execution process includes multiple stages such as task scheduling, data preparation, calculation execution, result processing, and distribution. The system uses a task queue mechanism to manage computing tasks at all levels and dynamically adjusts the execution order according to task priorities and resource conditions. For high-priority tasks (such as abnormal gas concentration warnings), they can preempt low-priority tasks that are being executed.

[0088] For complex tasks that require cross-level collaboration, such as predicting the airflow field of the entire mine, the system automatically decomposes the task into multiple subtasks: the edge layer is responsible for local high-precision calculations, the fog layer is responsible for regional integration, and the cloud layer is responsible for global optimization and coordination. The calculation results at each level are integrated through a hierarchical data fusion algorithm to obtain a globally consistent and accurate prediction result. In the process of handling emergencies, the system will quickly generate preliminary results at the edge layer for emergency response, while calculating more accurate results at higher levels, and then gradually replace the preliminary results with the accurate results to achieve a balance between accuracy and response speed.

[0089] Optionally, during the task execution process, an adaptive checkpoint mechanism can be adopted to regularly save the intermediate calculation state to cope with possible computing node failures or communication interruptions. The checkpoint saving frequency is dynamically adjusted according to task importance, computational complexity, and node reliability. For critical tasks and tasks running on nodes with lower reliability, the checkpoint frequency is increased; for secondary tasks and tasks running on high-reliability nodes, the checkpoint frequency is reduced to reduce additional overhead. When a node failure is detected, the system can resume the calculation from the nearest checkpoint instead of starting from scratch, improving the system's reliability and fault tolerance.

[0090] In some embodiments, the result aggregation process can adopt a progressive refinement strategy, first generating low-precision preliminary results, and then gradually improving the precision. This strategy is particularly suitable for interactive analysis scenarios, where users can first view the preliminary results and decide whether to wait for more accurate results or adjust the analysis parameters based on the preliminary results to improve the system's responsiveness and user experience.

[0091] Step 5: Based on the execution results of the task assignment decision model, dynamically adjust the balance between the accuracy and performance of the airflow field calculation according to the urgency and accuracy requirements of the ventilation and gas filtration decision, so as to achieve adaptive services in different scenarios; This step dynamically adjusts the balance between the accuracy and performance of the airflow field calculation according to the urgency and accuracy requirements of the ventilation and gas filtration decision, ensuring that the system can adaptively provide the most suitable services in different scenarios. It specifically includes the following sub-steps: Step 5.1: Evaluate the scene urgency; Evaluate the urgency of the current mine environment, including: Concentration and change rate of harmful gases: Monitor the concentration values and change trends of various harmful gases; Personnel distribution: Monitor the location distribution of the operating personnel in the mine; Ventilation equipment status: Monitor the operating status of ventilation equipment such as fans and air doors; Abnormal event detection: Identify possible abnormal events, such as gas leakage, equipment failure, etc.

[0092] Calculate the scene emergency degree index based on the above factors , where 0 represents the normal state and 1 represents the extreme emergency state.

[0093] According to the embodiments of the present application, the personnel distribution information is obtained through the mine personnel positioning system. When personnel are concentrated in high-risk areas, the emergency degree index is correspondingly increased. The ventilation equipment status monitoring includes monitoring of fan operation parameters, air door opening degree, wind speed, etc. When key equipment malfunctions, the emergency degree index is increased. Therefore, the system continuously analyzes various monitoring data through an anomaly detection algorithm. When potential abnormal events are identified, the emergency degree is triggered to increase. The final emergency degree index is calculated through a weighted fusion algorithm, and the weights are set according to mine safety management experience and can be dynamically adjusted according to the actual operation situation.

[0094] Optionally, in some special mine environments, the scene emergency degree assessment can be combined with historical data and prediction models to achieve forward-looking assessment. The system analyzes historical accident data to identify precursor patterns leading to emergencies. When similar patterns are detected, even if the current monitoring indicators have not reached the warning threshold, the emergency degree index is increased in advance. In addition, by analyzing the monitoring data before historical gas outburst events, the system may find that certain subtle parameter combination change patterns often indicate the imminent occurrence of emergencies. By identifying these patterns, early warning is achieved, and more time is gained for preventive measures.

[0095] In some other embodiments, the scene emergency degree assessment can be combined with a multi-level collaborative judgment method to avoid false alarms caused by single-point misjudgment. When abnormal data is detected by sensors in a certain area, the system automatically activates the high-frequency sampling mode of sensors in the surrounding areas to cross-verify the authenticity of the abnormal situation. Only when data from multiple independent sources all point to the abnormal situation can the emergency degree index be increased. This collaborative judgment method reduces the false alarm rate and improves the reliability of the system and user trust.

[0096] Step 5.2, setting the accuracy configuration parameters; Based on the scene emergency degree, set the accuracy configuration parameters for the airflow field calculation, including: Spatial resolution: The density of the calculation grid, which can be appropriately reduced in case of emergency; Time step: The time accuracy of the calculation, which can be appropriately increased in case of emergency; Iterative convergence threshold: The convergence accuracy of numerical calculations, which can be appropriately relaxed in case of emergency; Model complexity: The complexity of the model used, and a simpler model can be selected in case of emergency.

[0097] It should be noted that the precision configuration parameter is inversely proportional to the emergency index. The higher the emergency level, the lower the precision parameter, in order to improve the calculation speed.

[0098] According to the embodiments of the present application, adjusting the model complexity is the most effective means of precision-performance balance. The complete neuro-physical hybrid model is used under normal conditions, and a simplified model (such as removing some physical constraint terms, reducing the number of network layers, etc.) can be switched to in case of emergency. The adjustment of each parameter uses a piecewise linear or sigmoid function to achieve a smooth transition and avoid calculation instability caused by parameter mutations. Therefore, the system will automatically optimize the adjustment strategy of the precision configuration parameter according to historical operation data, and maintain the calculation precision as much as possible on the premise of ensuring the response speed.

[0099] Optionally, in some mine environments with uneven distribution of computing resources, the precision configuration parameter setting can adopt a regional differentiation strategy, and different precision configuration parameters are set according to the computing resource situation and safety importance of each region. For example, for the mining face area with a higher safety risk, a higher calculation precision is maintained even in case of emergency; for the auxiliary roadway area with a lower safety risk, the calculation precision can be more aggressively reduced to improve the response speed.

[0100] In some other embodiments, the precision configuration parameter setting can be combined with the priority classification method, and different precision adjustment strategies are adopted for different types of calculation tasks. In addition, for the prediction of key parameters (such as gas concentration), a high precision is maintained even in case of emergency, while for the prediction of secondary parameters (such as temperature distribution), the precision can be greatly reduced to save computing resources. This differentiation strategy ensures that the system can focus on the most critical prediction tasks and improve the overall system efficiency under resource constraints.

[0101] Step 5.3, Asynchronous calculation and result update; Implement the asynchronous collaborative calculation and result update method between different levels: The edge layer quickly generates preliminary results based on the local model for emergency response; The fog layer and the cloud layer perform parallel calculations for more accurate results and gradually update the preliminary results of the edge layer; Establish a result confidence evaluation method to clearly indicate the reliability level of the current result.

[0102] Optionally, in a mine environment with fluctuating network conditions, asynchronous computing can adopt a result caching and pre-computation strategy to improve the system's response speed and stability. The system analyzes historical request patterns, predicts possible future requests, performs calculations in advance, and caches the results. Therefore, when an actual request arrives, if it is consistent with the prediction, the cached result can be directly returned to reduce the response latency; if it does not match the prediction, real-time computing is immediately initiated. The cached results are updated regularly according to the latest data to ensure their accuracy.

[0103] In some embodiments, asynchronous computing can be combined with a method for adaptive allocation of computing resources. According to the importance and urgency of the current task, the resource ratio allocated to each computing level is dynamically adjusted. Additionally, when potential risks are detected, the system can temporarily reduce the resource allocation for routine monitoring tasks and allocate more computing resources to high-precision prediction tasks in the risk area to ensure the most accurate prediction results at critical moments and support risk prevention and control decisions.

[0104] Step 5.4, Decision optimization and execution; Based on the airflow field calculation results, generate and execute a ventilation and gas filtration decision: Construct a ventilation and gas filtration objective function, comprehensively considering factors such as safety, energy consumption, system response speed, etc.; Adopt a multi-objective optimization algorithm to generate an optimal ventilation and gas filtration strategy; Send control instructions to each ventilation and gas filtration device to execute the optimization strategy; Real-time monitor the execution effect and make strategy adjustments if necessary.

[0105] In the actual application of the mine, the ventilation and gas filtration objective function consists of multiple components: a safety objective (minimizing the area where harmful gas concentrations exceed the standard), an energy consumption objective (minimizing the fan power consumption and the operating cost of the gas filtration device), a system response speed objective (minimizing the control adjustment time), an equipment life objective (minimizing the start-stop frequency and load fluctuations of the equipment), etc. The weights between the objectives are dynamically adjusted according to the current operating state of the mine, and the weight of the safety objective is increased in an emergency state.

[0106] The optimization algorithm adopts an improved genetic algorithm combined with a particle swarm optimization method, which can quickly find a solution close to the global optimum under complex constraint conditions. The optimization process considers various actual constraints, such as the fan characteristic curve, damper opening limit, equipment adjustment rate limit, etc. In the case of limited computing resources, the system will automatically reduce the optimization accuracy to ensure a feasible solution within an acceptable time.

[0107] Control instructions are sent to each execution device, such as variable-frequency fans, intelligent air dampers, automatic gas filtration devices, etc., through industrial fieldbuses or wireless communication networks. The instruction sending adopts a hierarchical control strategy. First, key adjustments (such as the speed adjustment of the main ventilation fan) are executed, and then secondary adjustments (such as the fine adjustment of the opening of local air dampers) are executed. The system continuously monitors the execution effect and adjusts the control parameters in a timely manner through a closed-loop feedback mechanism to ensure the achievement of the control target.

[0108] Under special working conditions, such as the advancement of the mining and excavation working face, the change of the temporary ventilation system, etc., the system will automatically identify the change of the working condition, recalculate the ventilation network parameters, and adjust the control strategy to ensure the stability and continuity of the ventilation and gas filtration effect.

[0109] Optionally, in some complex and changeable mine environments, robust optimization methods can be used for decision optimization. Considering the uncertainty and variability of system parameters, control strategies with good performance in various possible situations are generated. Specifically, when implementing, the system performs perturbation analysis on key parameters (such as air resistance coefficient, air leakage rate, etc.), evaluates the impact of different parameter changes on the control effect, and selects the control strategy with the optimal overall performance within the parameter fluctuation range to improve the robustness and reliability of the system in the actual environment.

[0110] In some other embodiments, decision execution can be combined with predictive control strategies. Based on the prediction results of the future evolution of the airflow field, control adjustments are made in advance to avoid the lag effect caused by passive response. For example, when it is predicted that the gas concentration in a certain area will rise to the warning value in a short time, the system increases the ventilation volume in that area in advance to prevent the gas concentration from actually reaching the warning value, realizing preventive control. Predictive control is particularly suitable for scenarios where the system response has a lag. For example, it takes a certain time for the airflow field to stabilize after the speed adjustment of a large fan. Advance control can effectively make up for this lag effect and improve the overall control effect of the system.

[0111] The mine intelligent ventilation and gas filtration dynamic regulation system is used to execute the above-mentioned mine intelligent ventilation and gas filtration dynamic regulation method, including: A neural-physical hybrid modeling module for constructing an airflow field prediction model that integrates physical constraints and deep neural networks; A hierarchical computing architecture module for decomposing the hybrid model into multi-level sub-models and deploying them to corresponding computing nodes; A federated learning collaboration module for establishing a multi-level collaboration mechanism to realize the sharing of model parameters and distributed optimization among computing nodes; An adaptive task scheduling module for constructing a task allocation decision model and dynamically allocating computing tasks according to task characteristics and resource status; An accuracy-performance balance module for dynamically adjusting the calculation accuracy and performance balance strategy according to the urgency of the scenario and the accuracy requirements.

[0112] As shown Figures 2 to 6 in the figure, there are respectively bar charts of the comparison results of the calculation time of the method of the present invention, the traditional CFD method, and the simplified analytical model under different tasks; line charts of the changing trend of the prediction accuracy of the edge layer, fog layer, and cloud layer models over time; scatter plots of the relationship between the calculation load and response time of the method of the present invention; radar charts of the comparison of multiple performance indicators between the system of the present invention and the traditional system; area charts of the calculation resource allocation ratios of the edge layer, fog layer, and cloud layer under different emergency scenarios.

[0113] Herein, the present invention provides an implementation example: This implementation has been applied to the ventilation system of a fully mechanized coal mining face in a large coal mine. The depth of the mine is about 800 meters, the length of the fully mechanized coal mining face is about 250 meters, the length of the return airway is about 2,000 meters, and the ventilation system adopts the "U-shaped" ventilation method. The main risks are relatively high coal seam gas content (average 8 m³ / t) and prone to local gas accumulation. The traditional ventilation control system has problems such as insufficient real-time performance, low system reliability, and poor energy utilization efficiency. Especially during the advancement of the working face, the gas distribution characteristics change greatly, and fixed ventilation parameters are difficult to adapt to the dynamic environmental changes.

[0114] The existing equipment environment of the mine includes: Sensor network: 120 gas sensors, 85 carbon monoxide sensors, 60 wind speed sensors, and 30 temperature and humidity sensors are arranged; Computing equipment: 180 edge computing nodes are deployed in the edge layer (based on an ARM architecture microprocessor, main frequency 1.2 GHz, memory 512 MB), 12 regional control stations (industrial computers, Intel i5 processors, 8 GB memory) are deployed in the fog layer, and a central control server cluster (dual Xeon processors, 64 GB memory, equipped with NVIDIA T4 GPU acceleration cards) is deployed in the cloud layer; Ventilation equipment: 2 main ventilators (power 750 kW), 28 local ventilators (power 15 to 45 kW), 18 intelligent air doors, and 35 adjustable air windows; Communication network: A hybrid network based on the combination of industrial Ethernet and wireless communication. The communication bandwidth from the edge layer to the fog layer is 100 Mbps, and the communication bandwidth from the fog layer to the cloud layer is 1 Gbps.

[0115] The main objectives of applying this implementation are: improving the real-time response ability of the ventilation system to gas distribution changes, reducing ventilation energy consumption, improving system reliability, realizing automatic dynamic adjustment of ventilation parameters, and at the same time ensuring that the system can quickly respond in case of emergencies to provide safety protection for miners.

[0116] For the specific environment of this mine, two types of data were first collected: Based on the refined simulation data generated by computational fluid dynamics software, it covers the airflow field distribution under 15 different working conditions, and each working condition contains approximately 12,000 spatial grid points; The measured data obtained from the mine sensor network, including parameters such as gas concentration, wind speed, and pressure at different times within three months, totaling approximately 2.7 million records.

[0117] The construction process of the neuro-physical hybrid model is as follows: First, based on the geometric structure and ventilation system characteristics of the mine, physical constraint conditions were constructed. The main ventilation roadway adopts a one-dimensional simplified model, and the working face area adopts a three-dimensional grid model. The physical constraints mainly consider three aspects: mass conservation, momentum conservation, and gas component conservation, and at the same time introduce dynamic boundary conditions considering the influence of working face advancement.

[0118] Next, a 6-layer deep neural network structure was constructed. The input layer receives 13 feature parameters, including spatial coordinates , time , the rotational speed of the main ventilation fan, the opening degree of the air door, the advancing distance of the working face, etc. The attention mechanism is used in the middle layer of the network to fuse physical information, and the output layer predicts 4 variables: three-dimensional wind speed components and gas concentration.

[0119] In the model training stage, the weights of the composite loss function are dynamically adjusted. In the initial stage, the weight of the data fitting loss is 0.8, and the weight of the physical constraint loss is 0.2, and it is gradually adjusted to 0.4 and 0.6 as the training progresses. The training uses the Adam optimizer with an initial learning rate of 0.001, and decays by 10% every 50 epochs. The training is carried out on the cloud server and takes about 16 hours. The average absolute error of the final model on the validation set is: wind speed 0.12 m / s, gas concentration 0.08%.

[0120] After the model training is completed, knowledge distillation technology is applied to construct three model versions with different complexities, which are respectively deployed on the edge, fog, and cloud. The number of parameters of the edge layer model is about 0.3 MB, and the inference time < 20 ms; the number of parameters of the fog layer model is about 1.2 MB, and the inference time < 200 ms; the cloud retains the complete model, with the number of parameters about 4.8 MB and the inference time < 2 s.

[0121] According to the physical layout of the mine, the entire mine space is divided into 38 regions, and the working face area is further divided into 12 sub-regions. The spatial partitioning strategy is as follows: Edge layer: Each sensor node is responsible for a local area with a radius of about 15 meters, and the coverage range of key sensors (such as working face gas sensors) is extended to 25 meters; Fog layer: Each regional control station is responsible for a complete functional area, such as a mining area or a group of related roadways, with an average coverage area of about 2,000 square meters; Cloud layer: covers the entire mine area, with a total area of about 80,000 square meters.

[0122] An overlap area of 5 to 10 meters is set between adjacent areas to ensure a smooth transition of boundary data. For high-risk areas (such as near coal mining machines at the working face), a higher grid density is set, and calculation accuracy is given higher priority.

[0123] The model deployment adopts a hierarchical deployment strategy: The edge layer deploys a simplified model, retaining only a two-layer network structure, which is mainly responsible for the rapid prediction of the local area around the sensor, such as the gas diffusion prediction within a 10-meter range around the gas sensor; The fog layer deploys a medium-complexity model, retains a 4-layer network structure, and is responsible for the calculation of the airflow field of the entire working face or mining area; The cloud layer deploys a complete complexity model, retains all 6 layers of network structure, and is responsible for the optimization and coordination of the airflow field throughout the entire mine.

[0124] To adapt to different hardware platforms, the edge layer model uses 8-bit integer quantization, the fog layer model uses 16-bit floating point representation, and the cloud layer model maintains 32-bit floating point precision. During the deployment process, in order to improve system reliability, the edge nodes in key areas are deployed with dual redundancy to ensure that the system can continue to operate even if a single node fails.

[0125] In the actual application of this mine, the federated learning architecture is organized according to the mine's management hierarchy and physical layout. The 180 edge computing nodes are divided into 12 federated learning groups, each corresponding to an area governed by a fog layer node. The 12 fog layer nodes form a higher-level federated learning network, coordinated by the cloud layer.

[0126] The local model update mechanism is set up as follows: Edge layer nodes: Model parameters are updated every 10 minutes based on local sensor data, with the batch size set to 32 and the initial learning rate value set to 0.008, which is gradually reduced using the cosine annealing strategy. The update frequency of sensor nodes in high-risk areas near the mining face is increased to once every 5 minutes; Fog layer nodes: summarize the model updates of the edge nodes under their jurisdiction every 2 hours and perform regional model optimization; Cloud node: Perform global model optimization every 12 hours and distribute the updated global model parameters to nodes at each level.

[0127] During the model parameter aggregation process, a weighted strategy based on data quality is applied. The data quality assessment is based on three metrics: sensor health (rated according to the sensor self-diagnosis results), data coefficient of variation (measuring data stability), and historical prediction accuracy. The model update weight of each edge node is calculated by weighting these three metrics. In addition, an anomaly detection mechanism is implemented. When the cosine similarity between the model update submitted by a certain node and the average update is lower than 0.65, this update will be excluded from the aggregation process.

[0128] During the operation of the system, it is found that the local updates in the edge layer occasionally show oscillation phenomena, especially when the working face advances and causes environmental mutations. To solve this problem, a momentum update strategy is introduced, which mixes the current update and the historical update in a ratio of 0.7:0.3, effectively suppressing the parameter oscillation.

[0129] The actual operation data shows that after the collaborative optimization using federated learning, the amount of model training data transmission is reduced by 92%, from the original approximately 3.8TB of raw data that needed to be transmitted to only about 304MB of model parameter updates that need to be transmitted. At the same time, the distributed training makes full use of the computing resources of each node, and the model convergence speed is increased by about 3.5 times.

[0130] In the actual application of the mine ventilation system, according to the different characteristics of tasks, the system automatically assigns the computing tasks to the most suitable computing level. The task classification and assignment strategy are as follows: Local monitoring tasks: Such as monitoring the gas concentration distribution within 10 meters around the sensor, with a required response time < 100ms, are assigned to the edge layer for execution. A typical example is when the gas sensor near the shearer detects an increase in concentration, immediately start the gas distribution calculation in the surrounding area to evaluate whether local ventilation parameters need to be adjusted.

[0131] Regional prediction tasks: Such as predicting the overall airflow field change trend of the working face, with a required response time < 2s, are assigned to the fog layer for execution. For example, when the working face advances to a new position, it is necessary to recalculate the ventilation effect of the entire working face and predict the gas distribution changes in the next few hours.

[0132] Global optimization tasks: Such as optimizing the ventilation network parameters of the entire mine, with an acceptable response time < 30s, are assigned to the cloud layer for execution. A typical example is that before the start of each shift, the system calculates the optimal main ventilator parameters and damper opening configurations according to the production plan of the day and the current ventilation conditions.

[0133] The task allocation decision model adopts a comprehensive cost function, considering four factors: computational complexity, data transmission volume, latency requirement, and energy consumption. In the normal monitoring state, the weights of the four factors are 0.25, 0.2, 0.25, and 0.3 respectively; in the warning state where the gas concentration is close to the warning value, the weight of the latency factor is increased to 0.6, and the weights of other factors are correspondingly reduced; in the low-load period at night, the weight of the energy consumption factor is increased to 0.5, and energy conservation is given priority.

[0134] Actual operation data shows that the adaptive task scheduling improves the response speed and resource utilization rate of the system. Compared with the fixed task allocation, the average response time is reduced by 62%, and the computational resource utilization rate is increased by 58%. Especially in the high-load situation, the system can still maintain a stable response without task accumulation.

[0135] During the daily operation of the mine, the system dynamically adjusts the balance between computational accuracy and performance according to the scene urgency. The scene urgency assessment is based on the following index weight settings: gas concentration (40%), gas concentration change rate (25%), personnel distribution (20%), and equipment status (15%).

[0136] The accuracy configuration parameters under different urgencies are set as follows: Normal state ( is from 0 to 0.3): Spatial resolution is 5 m / point, time step is 10 s, iteration convergence threshold , and the complete model is used; Warning state ( is from 0.3 to 0.6): Spatial resolution is 8 m / point, time step is 20 s, iteration convergence threshold , and the medium-complexity model is used; Warning state ( is from 0.6 to 0.8): Spatial resolution is 12 m / point, time step is 30 s, iteration convergence threshold , and the simplified model is used; Emergency state ( is from 0.8 to 1.0): Spatial resolution is 20 m / point, time step is 60 s, iteration convergence threshold , and the highly simplified model is used.

[0137] During the application of the system, a typical emergency handling process was recorded: When the coal shearer cut into a gas-rich area, the nearby sensors detected a rapid increase in the gas concentration, and the scene urgency quickly rose from 0.2 to 0.85. The system immediately switched to the emergency state mode. The edge layer gave a preliminary prediction result within 82 milliseconds, indicating the gas diffusion direction and scope, and triggering the automatic adjustment of the nearby air doors. At the same time, the fog layer and the cloud layer started more precise calculations in parallel and provided more precise prediction results after 1.6 seconds and 12 seconds respectively, gradually optimizing the control strategy. During the whole process, the system response time was 7.8 times faster than that of the traditional fixed-precision system, successfully avoiding gas overrun incidents.

[0138] The asynchronous calculation and result update mechanism performed excellently in practical applications. In the above emergency, when the edge layer gave the preliminary result, the control system immediately responded; as the more precise results of the fog layer and the cloud layer were successively generated, the control strategy was continuously optimized, forming a smooth control process. This progressive result update mechanism ensured that the system could provide the best available result at any time without waiting for the final precise calculation to be completed.

[0139] Six months after the application of this implementation method in this mine, the system operation data was collected for technical effect verification, focusing on two core technical effects: the balance between calculation efficiency and accuracy and system reliability.

[0140] By comparing the performance of this implementation method with that of the traditional CFD method and the simplified analytical model, the following data was obtained: Local airflow field prediction (10m×10m×3m area): Traditional CFD method: The calculation time was about 45 minutes, and the average error was about 2%; Simplified analytical model: The calculation time was about 0.8 seconds, and the average error was about 12%; This implementation method: The calculation time was about 0.016 seconds, and the average error was about 4.2%.

[0141] Working face airflow field prediction (250m×20m×3m area): Traditional CFD method: The calculation time was about 4 hours, and the average error was about 2.5%; Simplified analytical model: The calculation time was about 4 seconds, and the average error was about 15%; This implementation method: The calculation time was about 0.6 seconds, and the average error was about 5.8%.

[0142] Full mine airflow network optimization: Traditional CFD method: The calculation time was about 18 hours, and the average error was about 3%; Simplified analytical model: The calculation time was about 25 seconds, and the average error was about 18%; This embodiment: The calculation time is about 8 seconds, and the average error is about 6.5%.

[0143] Data shows that the calculation speed of this embodiment is about 160 to 10,800 times faster than the traditional CFD method. At the same time, the prediction error is controlled within a reasonable range, which is better than the simplified analytical model. Especially in the local airflow field prediction task, this embodiment achieves a response speed in milliseconds, meeting the requirements of real-time control.

[0144] In actual production, this high-efficiency airflow field prediction ability enables the system to detect abnormal trends and respond at the early stage of gas concentration changes, detecting potential risks about 8 to 15 minutes earlier than traditional systems and winning valuable time for preventive measures.

[0145] By comparing the reliability indicators of this embodiment with those of traditional centralized systems, the following data are obtained: System availability: Traditional centralized system: The cumulative unavailable time within 6 months is about 263 minutes, and the availability is about 99.9%; This embodiment: The cumulative unavailable time within 6 months is about 18 minutes, and the availability is about 99.993%.

[0146] Single-point failure recovery ability: Traditional centralized system: A failure of the central server causes the entire system to collapse, and the average recovery time is about 45 minutes; This embodiment: The failure of any computing node only affects local functions. The average recovery time of edge nodes is about 3 minutes, that of fog layer nodes is about 8 minutes, and that of cloud layer nodes is about 15 minutes.

[0147] Network interruption adaptation ability: Traditional centralized system: Network interruption causes the system functions to be completely lost; This embodiment: When the network is interrupted, the edge layer still maintains 80% of its basic functions, the fog layer maintains 60% of its functions, and only the advanced optimization functions are temporarily unavailable.

[0148] Data shows that this embodiment improves the reliability of the system, especially performing excellently in dealing with various faults and abnormal situations. In a main network interruption event, the traditional system completely failed, while the edge layer and fog layer of this system continued to operate, successfully handling a local gas concentration anomaly event and avoiding possible safety risks.

[0149] In addition, the distributed architecture also improves the scalability of the system. During the system operation, 28 new sensor nodes were added. The traditional system needed to be shut down for reconfiguration, while this system only needed to complete the deployment and configuration of edge nodes locally without affecting the normal operation of other areas.

[0150] Generally speaking, this embodiment has achieved good technical effects in actual mine applications. In particular, the improvements in key indicators such as the balance between calculation efficiency and accuracy, and system reliability provide a strong guarantee for mine safety production.

[0151] The embodiments of the present invention have been described above, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. Intelligent ventilation and gas filtration dynamic regulation method for mine, characterized by Including: Construct a neuro-physical hybrid model by combining physical constraint conditions with a deep neural network to predict the distribution of the mine airflow field; Based on the neuro-physical hybrid model, decompose the neuro-physical hybrid model into three hierarchical sub-models and deploy them to corresponding computing nodes to implement a hierarchical computing architecture; Based on the hierarchical computing architecture, establish a multi-level collaborative mechanism based on federated learning, enabling each computing node to share model updates without sharing raw data, and realizing distributed model training and optimization; Based on the resource status information of the multi-level collaborative mechanism, construct a task allocation decision model, and dynamically determine the optimal execution level of the airflow field calculation task according to task characteristics, resource status, and real-time requirements; Based on the execution results of the task allocation decision model, dynamically adjust the accuracy and performance balance of the airflow field calculation according to the urgency and accuracy requirements of the ventilation and gas filtration decision, and realize adaptive services in different scenarios.

2. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 1, wherein, The steps of constructing the neuro-physical hybrid model include: Construct physical constraint conditions describing the mine airflow field, where the physical constraint conditions are based on the basic equations of fluid mechanics and consider the special mine environment; Construct a deep neural network structure for airflow field prediction, including an input layer, an encoder layer, a physical information fusion layer, a decoder layer, and an output layer; Construct a composite loss function containing data fitting loss and physical constraint loss; Based on computational fluid dynamics simulation data and measured data, use the composite loss function to train the deep neural network to make the model meet both data fitting and physical constraint requirements.

3. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 2, wherein The physical constraint conditions include: The mass conservation equation to ensure the incompressibility of the fluid; The momentum conservation equation to describe the relationship between the change in fluid velocity and the pressure gradient, viscous diffusion, and external forces; The geometric constraints of the mine roadway, representing the flow velocity and pressure conditions of the solid wall, air inlet, and air outlet through boundary conditions.

4. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 1, characterized in that The steps of decomposing the neuro-physical hybrid model into three hierarchical sub-models include: Based on the characteristics and dependencies of the computing tasks, decompose the functions of the neuro-physical hybrid model into an edge layer model, a fog layer model, and a cloud layer model; According to the physical layout of the mine and the structure of the ventilation system, formulate a spatial partitioning strategy to determine the spatial scope responsible for each computing level; According to the hardware capabilities of the computing nodes at each level, classify the complexity of the model to make each level of model adapt to the corresponding hardware capabilities.

5. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 4, characterized in that, The edge layer model is responsible for quickly estimating the airflow field in a local area with a spatial range of less than 50 meters and is deployed on the edge computing devices at the monitoring points; The fog layer model is responsible for the fusion and coordination of the airflow field in a medium area with a spatial range of 200 to 500 meters and is deployed on the computing devices at the regional control stations; The cloud layer model is responsible for global airflow field optimization and long-term prediction and is deployed on the high-performance servers or cloud computing platforms in the central control room.

6. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 1, characterized in that The steps of establishing a multi-level collaborative mechanism based on federated learning include: Construct a three-layer federated learning architecture, including edge layer federated learning, fog layer federated learning, and cloud layer global coordination; Establish a local model update mechanism to enable each computing node to update model parameters based on local data; Establish a model parameter aggregation strategy to calculate the weighted average value according to the data volume and data quality of each node to form an updated global model. Establish a cross - level model synchronization mechanism to achieve model parameter synchronization and sharing between different levels.

7. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 6, characterized in that The model parameter aggregation strategy includes: Calculate the aggregation weight based on the data volume of each node, sensor confidence, data variation coefficient, and historical accuracy; Introduce a parameter anomaly detection mechanism to identify and filter malicious updates and abnormal updates, ensuring the correctness of the aggregation result.

8. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 1, wherein The steps for constructing the task assignment decision model include: Analyze the characteristics of the airflow field calculation task, including the spatio - temporal range, real - time requirement, accuracy requirement, and computational complexity; Real - time monitor the resource status of each computing level, including computing resources, communication resources, and power status; Construct a comprehensive cost function considering computing cost, communication cost, latency cost, and energy consumption cost; Based on the comprehensive cost function, construct a task assignment decision function, and select the computing level that minimizes the comprehensive cost as the task execution level.

9. The intelligent ventilation and gas filtration dynamic regulation method for mine according to claim 1, wherein The steps for dynamically adjusting the accuracy - performance balance of airflow field calculation include: Evaluate the urgency of the current mine environment, including harmful gas concentration and its change rate, personnel distribution, ventilation equipment status, and abnormal event detection; Based on the scene urgency, set the accuracy configuration parameters for airflow field calculation, including spatial resolution, time step, iteration convergence threshold, and model complexity; Implement an asynchronous collaborative computing and result update mechanism between different levels, enabling the edge layer to quickly generate preliminary results based on the local model, and the fog layer and cloud layer to perform parallel computing for more accurate results; Based on the airflow field calculation result, generate and execute the ventilation and gas filtration decision, and at the same time, real - time monitor the execution effect.

10. The intelligent ventilation and gas filtration dynamic regulation system for mine is characterized in that, Used to execute the dynamic regulation method for intelligent mine ventilation and gas filtration according to any one of claims 1 - 9, including: A neuro - physical hybrid modeling module for constructing an airflow field prediction model that integrates physical constraints and deep neural networks; A hierarchical computing architecture module for decomposing the hybrid model into multi - level sub - models and deploying them to corresponding computing nodes; A federated learning collaboration module for establishing a multi - level collaboration mechanism to achieve model parameter sharing and distributed optimization between computing nodes; An adaptive task scheduling module for constructing a task assignment decision model and dynamically allocating computing tasks according to task characteristics and resource status; An accuracy - performance balance module for dynamically adjusting the accuracy - performance balance strategy according to the scene urgency and accuracy requirements.

Images