Method for optimizing distribution of air volume of graded cooling ventilation system for deep well heat damage treatment

Abstract

The application discloses a method for optimizing air volume distribution of a graded cooling ventilation system for deep well heat treatment, which comprises the following steps: S1, constructing a physical model of a deep well ventilation network, and deploying a multi-stage cooling device and an intelligent air door adjusting system; S2, collecting real-time heat hazard image data of a roadway and a working face in a well; S3, constructing an improved U-Net++ image segmentation network based on a mixed attention mechanism, and performing pixel-level segmentation on the real-time heat hazard image; S4, calculating a spatial heat load weight coefficient of a heat source according to the image segmentation result; S5, solving an optimal air volume distribution scheme by using an improved multi-objective particle swarm optimization algorithm; and S6, adjusting air volume and cooling capacity according to a control signal by an actuator. The application ensures continuous optimization of treatment effect by a closed-loop feedback correction mechanism, significantly improves the accuracy, real-time performance and energy efficiency of deep well heat treatment, and effectively overcomes the core problems of heat source identification ambiguity, ventilation adjustment dependence on experience, high energy consumption and difficulty in guaranteeing thermal comfort in the traditional method.

Description

Technical Field

[0001] This invention relates to the field of deep well heat hazard control technology, and in particular to a method for optimizing the air volume distribution of a graded cooling and ventilation system for deep well heat hazard control. Background Technology

[0002] In the process of deep well mining, the problem of heat hazards is becoming increasingly prominent. Traditional deep well heat hazard control technologies generally have shortcomings such as low accuracy of heat source identification, high energy consumption due to reliance on manual experience for ventilation and cooling system adjustment, and difficulty in ensuring thermal comfort in the working environment. Especially when facing complex and ever-changing underground thermal environments, existing technologies lack the ability to accurately perceive and dynamically quantify the spatial distribution characteristics of heat sources, making it difficult to achieve precise allocation of ventilation cooling capacity on demand. This leads to a contradictory situation where inadequate local heat hazard control coexists with overall energy waste. It is urgent to solve the core contradiction between heat hazard control and energy efficiency improvement through technological innovation. Summary of the Invention

[0003] The technical problem to be solved by this invention is to overcome the shortcomings of existing technologies that easily lead to the contradiction of insufficient local heat damage control and overall energy waste, and to provide a method for optimizing the air volume distribution of a graded cooling and ventilation system for deep well heat damage control.

[0004] The technical solution adopted by this invention to solve its technical problem is a method for optimizing the air volume distribution of a graded cooling and ventilation system for deep well heat hazard control, comprising the following steps:

[0005] S1: Construct a physical model of the deep well ventilation network, deploy multi-level cooling devices and intelligent damper adjustment system, and arrange infrared thermal imaging and visible light binocular acquisition equipment at key nodes;

[0006] S2: Collect real-time thermal hazard image data of underground roadways and working faces, construct a deep well thermal hazard image dataset containing the features of "heat source-background-airflow disturbance", and perform preprocessing and enhancement.

[0007] S3: Construct an improved U-Net++ image segmentation network based on a hybrid attention mechanism to perform pixel-level segmentation on real-time heat hazard images and extract the precise contours, temperature field distribution features, and geometric center coordinates of the heat source region;

[0008] S4: Based on the image segmentation results, calculate the spatial heat load weight coefficient of the heat source, and combine it with the ventilation network solution model to establish an air volume optimization allocation function with the dual objectives of "lowest energy consumption and highest thermal comfort".

[0009] S5: The improved multi-objective particle swarm optimization algorithm is used to solve the optimal air volume distribution scheme, and generate the cooling output command and damper opening control signal of the staged cooling device.

[0010] S6: The actuator adjusts the air volume and cooling capacity according to the control signal and verifies the treatment effect through the feedback sensor. If the standard is not met, it returns to S2 for iterative correction.

[0011] Furthermore, the image preprocessing in step S2 includes an enhancement algorithm for low illumination and noise in downhole environments, namely, "dark channel prior and thermal radiation compensation":

[0012] The acquired original image I(x) is processed to obtain the enhanced image. Represented as:

[0013] ;

[0014] in, Here is the dark channel plot, and t(x) is the transmittance plot. This is the ambient light compensation term based on infrared thermal radiation calibration; the algorithm solves for the optimal transmittance by minimizing the cost function E(t):

[0015] ;

[0016] in, For gradient operators, γ is the input source for the algorithm, and γ is the fidelity coefficient used to strike a balance between denoising and preserving thermal boundary details.

[0017] By introducing a dark channel prior, the visibility in low-light areas is effectively improved, and the ambient light compensation term of infrared thermal radiation calibration is combined to correct the thermal radiation characteristics of the heat source, thus avoiding the thermal information deviation caused by simply relying on visible light enhancement.

[0018] Furthermore, the specific algorithm process of the hybrid attention mechanism in step S3 is as follows:

[0019] In U-Net++, a parallel structure of convolutional block attention modules and Transformer self-attention modules is embedded in the skip connections and bottleneck layers. The feature weighted fusion formula is as follows:

[0020] ;

[0021] in, For the input feature map, To output the feature map, For the input feature map, ⊕ denotes dot product, and ⊕ denotes residual join; The channel attention weights are generated from a shared MLP of max pooling and average pooling:

[0022] ;

[0023] in, and These are average pooling and max pooling, respectively. Spatial attention weights are generated based on channel dimension statistics of the feature map:

[0024]

[0025] Where σ is the Sigmoid activation function. , , , The weight matrix is ​​a learnable matrix. This represents a 7×7 convolution operation. and These are the average and maximum values, respectively, calculated along the channel dimension.

[0026] The hybrid attention mechanism suppresses downhole dust and water mist noise, highlighting the characteristics of high-temperature heat sources.

[0027] Furthermore, the calculation of the heat load weighting coefficient of the heat source space and the heat source classification strategy in step S4 are as follows:

[0028] The heat source regions segmented from the image are mapped to ventilation network nodes, and the heat load weight coefficient of node i is defined. for:

[0029] ;

[0030] in, This refers to the physical area corresponding to the segmented heat source pixel area. Let J be the physical area of ​​the j-th heat source region. This represents the average overheat value for the region. The heat dissipation efficiency coefficient of the heat source. This is the gas concentration correction factor associated with this node, where N is the total number of nodes. This is the gas concentration correction factor;

[0031] according to The heat source is divided into three levels: >0.15 indicates a Class 1 strong heat source, 0.05< ≤0.15 indicates a secondary heat source. ≤0.05 indicates a level 3 weak heat source.

[0032] Furthermore, the airflow optimization allocation function established in step S4 is a constrained nonlinear programming model:

[0033] Objective function:

[0034] ;

[0035] in, Let be the air volume of the k-th fan; The actual temperature of the i-th monitoring point;

[0036] Constraints:

[0037] ;

[0038] in, Let be the energy consumption function of the k-th wind turbine. ,in, This usually corresponds to the main resistance term. For the corresponding secondary resistance term, This corresponds to the basic loss item; Standard operating temperature; Penalties for violating safety constraints; , , It is a dynamic weighting coefficient that is adaptively adjusted based on the current severity of heat damage.

[0039] Furthermore, the improved multi-objective particle swarm optimization algorithm in step S5 introduces an adaptive inertia weight based on the heat source gradient:

[0040] The particle velocity update formula is revised as follows:

[0041] ;

[0042] Among them, inertia weight Dynamic changes with heat source concentration:

[0043] ;

[0044] in, and These are learning factors; and They are random numbers; The speed at the current moment; and These are the individual's historical best position and the global best position, respectively. α is the temperature gradient vector of the image segmentation region corresponding to the particle's current position, guiding the particle to search for the optimal solution in the high-temperature region; α is the thermal gradient guidance coefficient. and These represent the maximum and minimum values ​​set for parameter ω, respectively, and β is the slope coefficient of the Sigmoid function. Here, is the heat source threshold, and e is the natural constant.

[0045] Furthermore, a "virtual-reality" two-way correction mechanism based on digital twins is introduced:

[0046] A digital twin of the downhole ventilation network is established, and heat source boundary conditions obtained from image segmentation are received in real time. A computational fluid dynamics solver is used to pre-simulate the airflow distribution scheme in the twin and predict the temperature field distribution. Define the visual-physical consistency loss function. :

[0047] ;

[0048] in, The first term represents the true temperature field retrieved from the infrared image, and the second term represents the mass conservation residual. This refers to the physical loss weighting coefficient; The density of air; if Then, the boundary condition parameters of the physical model are corrected using the residuals until the error converges.

[0049] Furthermore, it also includes a feedforward-feedback composite control strategy under unsteady thermal disturbances:

[0050] When the image segmentation network detects a new heat source, it triggers feedforward control, directly based on the area of ​​the heat source. With strength Look up the table to obtain the initial air volume increment , For ambient temperature:

[0051] ;

[0052] Subsequently, the PID feedback controller adjusts the temperature deviation based on the measured values. Make fine adjustments, among which, and Let these represent the set temperature and the measured temperature, respectively. The final control variable u(t) is:

[0053] ;

[0054] in, This is the feedforward control variable. The time integral of the error, The rate of change of error, , , The PID parameters are tuned based on the response time of the ventilation system to achieve millisecond-level response to thermal disturbances and steady-state zero-static-error regulation.

[0055] Furthermore, the tiered cooling and ventilation system includes a primary cooling station, secondary local coolers, and intelligent damper assemblies; the primary cooling station is located at the bottom of the air intake shaft and is responsible for reducing the intake air temperature to [a lower value]. The secondary local cooler, based on the heat source location identified in step S3, is moved to a position furthest from the center of the heat source using a hydraulic automatic shifting device. Among them:

[0056] ;

[0057] in, Let the cooling efficiency decrease with distance as a function. The airflow at this location. For mobile costs;

[0058] The intelligent damper assembly adopts a variable cross-section structure, with the opening S corresponding to the target air volume. The relationship is non-linear: , where k is the damper characteristic coefficient and ΔP is the pressure difference across the damper.

[0059] Furthermore, the tiered cooling and ventilation system possesses a self-learning and model update mechanism:

[0060] Regularly compare actual governance results with model predictions to build an experience playback buffer.

[0061] The parameters θ of the attention mechanism network are fine-tuned using a deep deterministic gradient algorithm, with the loss function being:

[0062] ;

[0063] in, s represents the current thermal damage image status, a represents the airflow allocation action, r represents the negative energy consumption and over-temperature penalty reward, and D represents the experience buffer.

[0064] Through continuous iteration, the system adapts to changes in mine geological conditions and characteristic drift caused by equipment aging.

[0065] In summary, the present invention has the following beneficial technical effects:

[0066] This invention constructs a physical model of a deep well ventilation network and deploys multi-stage cooling devices and an intelligent damper adjustment system. By combining infrared thermal imaging and advanced image segmentation technology, it achieves accurate identification and quantitative classification of heat hazard sources. Then, it uses an improved multi-objective optimization algorithm to dynamically solve for the optimal airflow distribution scheme that balances the lowest energy consumption and the highest thermal comfort. A closed-loop feedback correction mechanism ensures continuous optimization of the treatment effect, significantly improving the accuracy, real-time performance, and energy efficiency of deep well heat hazard treatment. It effectively overcomes the core problems of traditional methods, such as fuzzy heat source identification, reliance on experience for ventilation adjustment, high energy consumption, and difficulty in guaranteeing thermal comfort. Attached Figure Description

[0067] Figure 1 This is a flowchart of an embodiment of the air volume optimization allocation method for the graded cooling and ventilation system for deep well heat hazard control of the present invention;

[0068] Figure 2 This is a framework diagram of the air volume optimization distribution system of the graded cooling and ventilation system for deep well heat hazard control according to the present invention. Detailed Implementation

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0070] Reference Figure 2 This embodiment includes an airflow optimization and distribution system for a graded cooling and ventilation system for deep well heat hazard control, and an airflow optimization and distribution method for a graded cooling and ventilation system applicable to deep well heat hazard control. The airflow optimization and distribution system includes the following modules:

[0071] The deep well ventilation network physical model construction module is used to build a deep well ventilation network physical model, deploy multi-level cooling devices and intelligent damper adjustment systems, and arrange infrared thermal imaging and visible light binocular acquisition devices at key nodes;

[0072] The thermal hazard image data acquisition and preprocessing module is used to acquire real-time thermal hazard image data of underground roadways and working faces, construct a deep well thermal hazard image dataset containing the features of "heat source-background-airflow disturbance", and perform preprocessing and enhancement. The preprocessing includes an enhancement algorithm for "dark channel prior and thermal radiation compensation" for low illumination and noise in underground mines.

[0073] The thermal hazard image segmentation module is used to construct an improved U-Net++ image segmentation network based on the Hybrid Attention Mechanism (HAM) to perform pixel-level segmentation on real-time thermal hazard images and extract the precise contours, temperature field distribution features, and geometric center coordinates of the heat source region.

[0074] The heat source load calculation and classification module is used to calculate the spatial heat load weight coefficient of the heat source based on the image segmentation results. Combined with the ventilation network solution model, it establishes an air volume optimization allocation function with the dual objectives of "lowest energy consumption and highest thermal comfort". The heat source is divided into three levels according to the spatial heat load weight coefficient of the heat source.

[0075] The air volume optimization and allocation solution module is used to solve the optimal air volume allocation scheme using an improved multi-objective particle swarm optimization algorithm (MOPSO) to generate the cooling output command and damper opening control signal of the staged cooling device. The improved multi-objective particle swarm optimization algorithm (MOPSO) introduces an adaptive inertial weight based on the heat source gradient.

[0076] The execution and feedback correction module is used to adjust the air volume and cooling capacity according to the control signal, and verify the treatment effect through the feedback sensor. If the standard is not met, it returns to the heat damage image data acquisition and preprocessing module for iterative correction.

[0077] Reference Figure 1 An airflow optimization allocation method for a graded cooling and ventilation system based on deep well heat hazard control includes the following steps:

[0078] S1: Construct a physical model of the deep well ventilation network, deploy multi-level cooling devices and intelligent damper adjustment system, and arrange infrared thermal imaging and visible light binocular acquisition equipment at key nodes;

[0079] S2: Collect real-time thermal hazard image data of underground roadways and working faces, construct a deep well thermal hazard image dataset containing the features of "heat source-background-airflow disturbance", and perform preprocessing and enhancement.

[0080] S3: Construct an improved U-Net++ image segmentation network based on the Hybrid Attention Mechanism (HAM) to perform pixel-level segmentation on real-time heat hazard images and extract the precise contours, temperature field distribution features, and geometric center coordinates of the heat source region.

[0081] S4: Based on the image segmentation results, calculate the spatial heat load weight coefficient of the heat source, and combine it with the ventilation network solution model to establish an air volume optimization allocation function with the dual objectives of "lowest energy consumption and highest thermal comfort".

[0082] S5: The improved multi-objective particle swarm optimization algorithm (MOPSO) is used to solve the optimal air volume distribution scheme and generate the cooling output command and damper opening control signal of the staged cooling device.

[0083] S6: The actuator adjusts the air volume and cooling capacity according to the control signal and verifies the treatment effect through the feedback sensor. If the standard is not met, it returns to S2 for iterative correction.

[0084] Based on the above complete steps, this paper addresses the core problems in existing deep well heat hazard management, such as ambiguous heat source identification, reliance on experience for ventilation regulation leading to high energy consumption and difficulty in ensuring thermal comfort, and lack of real-time closed-loop correction. This is achieved by constructing a physical model of the deep well ventilation network and deploying multi-stage cooling devices, intelligent dampers, and infrared-visible binocular acquisition equipment (S1), enabling real-time perception of underground heat hazards and accurate construction of a "heat source-background-airflow disturbance" feature dataset (S2). Furthermore, a U-Net++ network improved with a hybrid attention mechanism is used to perform pixel-level segmentation of heat hazard images, accurately extracting heat source contours. The system identifies the temperature field distribution and geometric center coordinates (S3), addressing the pain point of "coarse generalization" in traditional heat source identification. Based on the segmentation results, it calculates the spatial heat load weight of the heat source and establishes a dual-objective airflow optimization function (S4) combining ventilation network calculations to achieve "lowest energy consumption and highest thermal comfort." An improved multi-objective particle swarm optimization (MOPSO) algorithm is then used to solve for the optimal airflow allocation scheme (S5). Finally, the system adjusts the airflow and cooling capacity through actuators, and the effect is verified and iteratively corrected by feedback sensors (S6), forming a closed-loop intelligent governance process of "perception-segmentation-optimization-execution-feedback." Ultimately, this achieves accurate heat source identification, dynamic intelligent allocation of ventilation and cooling capacity, and a balance between energy consumption and thermal comfort, effectively improving the real-time performance, accuracy, and energy efficiency of deep well heat hazard management, overcoming the shortcomings of traditional methods such as "experience dependence, slow response, high energy consumption, and poor comfort."

[0085] In step S2, image preprocessing includes an enhancement algorithm for low downhole illumination and noise: "dark channel prior and thermal radiation compensation".

[0086] The acquired original image I(x) is processed to obtain the enhanced image. Represented as:

[0087] ;

[0088] in, Here is the dark channel plot, and t(x) is the transmittance plot. This is the ambient light compensation term based on infrared thermal radiation calibration; the algorithm solves for the optimal transmittance by minimizing the cost function E(t):

[0089] ;

[0090] in, For gradient operators, γ is the input source for the algorithm, and γ is the fidelity coefficient used to strike a balance between denoising and preserving thermal boundary details.

[0091] Based on the "dark channel prior and thermal radiation compensation" enhancement algorithm, this paper addresses the problems of image detail loss and severe noise interference caused by low-light environments in downhole mines, and the tendency of traditional enhancement algorithms to cause blurred thermal boundaries or temperature distortion. By introducing a dark channel prior, the visibility of low-light areas is effectively improved. Combined with an ambient light compensation term based on infrared thermal radiation calibration, the thermal radiation characteristics of the heat source are corrected, avoiding the thermal information deviation caused by relying solely on visible light enhancement. At the same time, by constructing a cost function that includes gradient constraints and data fidelity terms, an adaptive balance is achieved between denoising and smoothing and preserving the original thermal boundary details using the fidelity coefficient γ. This effectively suppresses the noise of complex downhole environments and accurately preserves the key features of heat source edges and temperature change areas, significantly improving the signal-to-noise ratio and contrast of the image. This provides a high-quality, high-fidelity data foundation for pixel-level heat source segmentation and temperature field extraction based on the hybrid attention mechanism of the U-Net++ network in the subsequent S3 step, overcoming the technical bottleneck of "denoising is blurring and enhancement is distortion" under severe lighting conditions.

[0092] In step S3, the specific algorithm process of the hybrid attention mechanism is as follows:

[0093] In U-Net++, a parallel structure is constructed by embedding convolutional block attention modules (CBAM) and Transformer self-attention modules within the skip connections and bottleneck layers. The feature weighted fusion formula is as follows:

[0094] ;

[0095] in, For the input feature map, To output the feature map, For the input feature map, ⊕ denotes dot product, and ⊕ denotes residual join; The channel attention weights are generated from a shared MLP of max pooling and average pooling:

[0096] ;

[0097] in, and These are average pooling and max pooling, respectively. Spatial attention weights are generated based on channel dimension statistics of the feature map:

[0098]

[0099] Where σ is the Sigmoid activation function. , , , The weight matrix is ​​a learnable matrix. This represents a 7×7 convolution operation. and These are the average and maximum values, respectively, calculated along the channel dimension.

[0100] This mechanism suppresses downhole dust and water mist noise, highlighting the characteristics of high-temperature heat sources.

[0101] This invention addresses the technical challenges of strong noise interference from dust and water mist in downhole environments, the easy obscuring of heat source features by the background, and the difficulty of traditional U-Net and single attention mechanisms in simultaneously considering global context and local details. It employs a parallel structure embedding CBAM and Transformer self-attention modules into the skip connections and bottleneck layers of U-Net++. Channel attention (based on a shared MLP with max / average pooling) is used to accurately calibrate the importance of different feature channels, while Transformer self-attention captures long-distance dependencies to strengthen global thermal field correlation. The two are weighted and fused through residual connections and dot multiplication, effectively suppressing irrelevant noise caused by dust and water mist and significantly enhancing the feature response of high-temperature heat source regions. This achieves synergistic optimization of noise suppression and heat source feature prominence, greatly improving the robustness of heat source feature extraction and pixel-level segmentation accuracy in complex downhole environments. It solves the pain points of incomplete heat source contour extraction and blurred edges in existing technologies under low signal-to-noise ratio conditions.

[0102] In step S4, the calculation of the heat load weighting coefficient of the heat source space and the heat source classification strategy are as follows:

[0103] The heat source regions segmented from the image are mapped to ventilation network nodes, and the heat load weight coefficient of node i is defined. for:

[0104] ;

[0105] in, This refers to the physical area corresponding to the segmented heat source pixel area. Let J be the physical area of ​​the j-th heat source region. This represents the average overheat value for the region. The heat dissipation efficiency coefficient of the heat source. This is the gas concentration correction factor associated with this node (μ>1 to increase weight when gas exceeds the limit), where N is the total number of nodes. This is the gas concentration correction factor;

[0106] according to The heat source is divided into three levels: >0.15 indicates a strong heat source (requiring localized strong cooling), 0.05< ≤0.15 indicates a secondary heat source (requires conventional cooling). ≤0.05 indicates a level 3 weak heat source (only requires ventilation for dilution).

[0107] Based on the above calculations and heat source classification strategy, this strategy addresses the shortcomings of existing deep well heat hazard management methods, such as a single dimension of heat load assessment (considering only temperature), neglecting the differences in rock heat dissipation characteristics and the coupling effect of gas safety, leading to a "one-size-fits-all" approach to airflow allocation or mismatch with actual needs. It constructs a multi-dimensional heat load weight coefficient model that includes the physical area of ​​the heat source, average over-temperature value, rock heat dissipation efficiency coefficient, and gas concentration correction factor, achieving a quantitative and accurate evaluation of node heat intensity. In particular, it introduces a dynamic gas concentration correction mechanism (μ>1), automatically increasing the heat load weight when gas levels exceed limits, effectively coupling the safety requirements of heat hazard management and gas management. Combined with a three-level classification standard (strong heat source, secondary heat source, weak heat source), it achieves differentiated and accurate classification of heat sources, providing a scientific weight basis for the multi-objective optimization of "lowest energy consumption and highest thermal comfort" in the subsequent step S5. This avoids energy waste caused by excessive cooling of weak heat sources and safety hazards from insufficient management of strong heat sources, significantly improving the targeting, safety, and energy efficiency of the ventilation and cooling system.

[0108] In step S4, the established airflow optimization allocation function is a constrained nonlinear programming model:

[0109] Objective function:

[0110] ;

[0111] in, Let be the air volume of the k-th fan; The actual temperature of the i-th monitoring point;

[0112] Constraints:

[0113] ;

[0114] in, Let be the energy consumption function of the k-th wind turbine. ,in, This usually corresponds to the main resistance term. For the corresponding secondary resistance term, This corresponds to the basic loss item; The standard operating temperature is 26°C. Penalties for violating safety constraints (such as exceeding gas limits or wind speed limits); , , It is a dynamic weighting coefficient that is adaptively adjusted based on the current severity of heat damage.

[0115] Based on the above model, this paper addresses the technical bottlenecks in existing deep well ventilation control, such as a single objective (pursuing only energy consumption or only cooling), difficulty in balancing thermal comfort and safety constraints, and lack of dynamic adaptive capabilities. By constructing a multi-objective nonlinear programming model that includes fan energy consumption, the sum of squared temperature deviations, and a safety penalty term, a three-in-one synergistic optimization of "lowest energy consumption, highest thermal comfort, and lowest safety risk" is achieved. Specifically, the temperature deviation term strictly constrains the operating environment to approach the standard temperature control target of 26℃, the safety penalty term effectively avoids the risks of exceeding gas and wind speed limits, and the dynamic weighting coefficients are adaptively adjusted based on the severity of heat damage. , , This gives the system the ability to switch strategies under different operating conditions (such as prioritizing cooling in high-temperature areas and prioritizing safety in high-gas areas). Combined with the physical constraints of the ventilation network, it ensures the mathematical feasibility and engineering effectiveness of the optimization scheme, providing a scientific goal orientation for subsequent intelligent algorithm solutions and significantly improving the overall energy efficiency and intelligence level of the ventilation system.

[0116] In step S5, the improved multi-objective particle swarm optimization algorithm introduces an adaptive inertia weight based on the heat source gradient:

[0117] The particle velocity update formula is revised as follows:

[0118] ;

[0119] Among them, inertia weight Dynamic changes with heat source concentration:

[0120] ;

[0121] in, and These are learning factors; and They are random numbers; The speed at the current moment; and These are the individual's historical best position and p's global best position, respectively; α is the temperature gradient vector of the image segmentation region corresponding to the particle's current position, guiding the particle to search for the optimal solution in the high-temperature region; α is the thermal gradient guidance coefficient. and These represent the maximum and minimum values ​​set for parameter ω, respectively, and β is the slope coefficient of the Sigmoid function. Here, is the heat source threshold, and e is the natural constant.

[0122] Based on the above, the improved Multi-Objective Particle Swarm Optimization (MOPSO) algorithm addresses the shortcomings of existing standard particle swarm optimization algorithms, such as getting trapped in local optima, slow convergence speed, strong blindness, and inability to effectively utilize the spatial distribution characteristics of heat sources when optimizing complex ventilation networks. It introduces an adaptive inertial weight mechanism based on heat source gradients, using temperature gradient vectors to guide particles to actively search towards high-temperature regions, achieving directional optimization where "the search focuses on the heat source." By dynamically adjusting the inertial weights using the Sigmoid function and heat source thresholds, the algorithm automatically enhances its local development capabilities for fine-tuning in areas with concentrated heat sources, while maintaining strong global exploration capabilities in areas with dispersed heat sources to avoid premature convergence. This significantly improves the convergence speed and solution quality of the algorithm in multi-peak, high-dimensional ventilation networks. Ultimately, it overcomes the pain points of traditional optimization algorithms, such as "blind search, low efficiency, and difficulty in considering multiple objectives," providing efficient and accurate computational support for finding the optimal airflow allocation scheme that achieves "lowest energy consumption and highest thermal comfort."

[0123] In another embodiment, a "virtual-reality" bidirectional correction mechanism based on digital twins is introduced:

[0124] A digital twin of the downhole ventilation network is established, receiving heat source boundary conditions obtained from image segmentation in real time. A computational fluid dynamics (CFD) solver (such as OpenFOAM) is used to pre-simulate the airflow distribution scheme within the twin, predicting the temperature field distribution. Define the visual-physical consistency loss function. :

[0125] ;

[0126] in, The first term represents the true temperature field retrieved from the infrared image, and the second term represents the mass conservation residual. This refers to the physical loss weighting coefficient; The density of air; if Then, the boundary condition parameters of the physical model (such as the heat dissipation coefficient of the surrounding rock) are corrected using the residuals until the error converges.

[0127] This mechanism addresses the core challenges in existing deep well thermal hazard management systems, such as inaccurate boundary conditions and distorted predictions in ventilation network physical models, as well as the disconnect between virtual simulation and real-world data. It achieves dynamic synchronization between virtual and real-world data by establishing a digital twin of the downhole ventilation network and receiving real-time image-segmented heat source boundary conditions. Using a CFD solver to pre-determine the airflow scheme and defining a visual-physical consistency loss function, it introduces the infrared-inverted real temperature field and mass conservation residuals into the model calibration process. When errors exceed limits, it automatically corrects key parameters such as the surrounding rock heat dissipation coefficient, forming a closed-loop evolutionary process of "data-driven - simulation prediction - residual feedback - parameter correction." This not only significantly improves the fidelity and prediction accuracy of the physical model and solves the distortion problem caused by fixed parameters in traditional models, but also enhances the system's adaptability to changes in the downhole environment through deep integration of visual and physical information. This provides a more reliable computational foundation for solving the optimal airflow in step S5, ultimately achieving a leap from "empirical estimation" to "high-fidelity simulation verification" of the management scheme.

[0128] It should also be noted that the air volume optimization allocation method also includes a feedforward-feedback composite control strategy under unsteady thermal disturbances:

[0129] When the image segmentation network detects a new heat source, it triggers feedforward control, directly based on the area of ​​the heat source. With strength Look up the table to obtain the initial air volume increment , For ambient temperature:

[0130] ;

[0131] Subsequently, the PID feedback controller adjusts the temperature deviation based on the measured values. Make fine adjustments, among which, and Let these represent the set temperature and the measured temperature, respectively. The final control variable u(t) is:

[0132] ;

[0133] in, This is the feedforward control variable. The time integral of the error, The rate of change of error, , , The PID parameters are tuned based on the response time of the ventilation system to achieve millisecond-level response to thermal disturbances and steady-state zero-static-error regulation.

[0134] This strategy addresses the challenges of existing ventilation control systems, such as slow response to non-steady-state thermal disturbances like underground blasting and equipment start-up / shutdown, and poor dynamic performance due to the reliance on error accumulation in simple PID control. It employs a composite control mode of "feedforward prediction + feedback correction": utilizing the real-time sensing capability of an image segmentation network, the initial airflow increment is directly output via a lookup table at the moment of thermal disturbance, achieving a millisecond-level rapid feedforward response and effectively overcoming the time delay of traditional feedback control. Subsequently, a PID controller is introduced to fine-tune based on the measured temperature deviation, eliminating steady-state errors and ensuring control accuracy. The combination of these two approaches achieves full-cycle coverage control of thermal disturbances, significantly improving the system's resistance to sudden thermal loads and dynamic response speed, while ensuring zero steady-state error performance and significantly suppressing temperature fluctuations. This solves the pain points of ventilation systems in complex deep well environments, characterized by "slow response, large oscillations, and poor stability."

[0135] In the method for optimizing the airflow distribution of the graded cooling and ventilation system for deep well heat hazard control, the graded cooling and ventilation system includes a primary cooling station, a secondary local cooler, and an intelligent damper assembly; the primary cooling station is located at the bottom of the air intake shaft and is responsible for reducing the intake air temperature to [a lower value]. The secondary local cooler, based on the heat source location identified in step S3, is moved to a position furthest from the center of the heat source using a hydraulic automatic shifting device. Among them:

[0136] ;

[0137] in, Let the cooling efficiency decrease with distance as a function. The airflow at this location. For mobile costs;

[0138] The intelligent damper assembly adopts a variable cross-section structure, with the opening degree S corresponding to the target air volume. The relationship is non-linear: , where k is the damper characteristic coefficient and ΔP is the pressure difference across the damper.

[0139] This tiered cooling and ventilation system addresses the shortcomings of existing technologies, such as fixed and rigid cooling equipment that cannot be dynamically deployed with the movement of heat sources, low flow control accuracy due to linear damper adjustment, and energy waste caused by unified cooling across the entire network. It establishes a primary cooling station to lower the inlet air temperature to 18°C ​​as the basic cold source, combined with a secondary local cooler that automatically relocates using a hydraulic device based on the heat source's location. This is achieved by applying a balance formula between the cooling efficiency decay function and the relocation cost (…). It precisely locates the center of the heat source, achieving dynamic, on-demand cooling that "follows the heat source"; simultaneously, the intelligent damper assembly adopts a variable cross-section structure, utilizing the non-linear relationship between opening and airflow. By precisely matching the flow resistance characteristics of the ventilation network, the problem of poor linear adjustment accuracy of traditional dampers was solved. This ultimately resulted in a three-dimensional treatment system of "whole-network pre-cooling + localized enhancement + precise air distribution," significantly improving cooling capacity utilization and airflow control accuracy, while greatly reducing system energy consumption while ensuring thermal comfort.

[0140] More specifically, the tiered cooling and ventilation system has a self-learning and model update mechanism:

[0141] Regularly compare actual governance results with model predictions to build an experience playback buffer.

[0142] The parameters θ of the attention mechanism network are fine-tuned using a deep deterministic gradient algorithm, with the loss function being:

[0143] ;

[0144] in, s represents the current thermal damage image status, a represents the airflow allocation action, r represents the negative energy consumption and over-temperature penalty reward, and D represents the experience buffer.

[0145] Through continuous iteration, the system adapts to changes in mine geological conditions and characteristic drift caused by equipment aging.

[0146] The self-learning and model update mechanism addresses the challenges of existing deep well thermal hazard control systems, such as fixed model parameters that cannot adapt to changes in mine geological conditions (e.g., altered geothermal gradients) and equipment aging (e.g., reduced fan efficiency and drifting of wind resistance characteristics), leading to long-term performance degradation. By constructing an experience replay buffer and introducing the Deep Deterministic Policy Gradient (DDPG) algorithm, the deviation between actual control data and model predictions is transformed into negative reward signals (energy consumption and over-temperature penalties), continuously fine-tuning the parameters θ of the hybrid attention mechanism network. This "data-driven-reinforcement learning" online evolution strategy endows the system with the "vitality" to adapt to environmental changes: it can not only automatically compensate for time-varying drift in the physical characteristics of the ventilation network but also continuously optimize the airflow allocation strategy through accumulated historical experience. This overcomes the shortcomings of traditional models that are "one-time calibration and rigid throughout their entire lifecycle," significantly improving the system's robustness, control accuracy, and energy efficiency maintenance capabilities during long-term operation, achieving a technological leap from "passive response" to "active evolution."

[0147] In this application, the airflow optimization allocation system of the graded cooling and ventilation system for deep well heat hazard control constructs an accurate physical model of the deep well ventilation network and deploys multi-level cooling and intelligent adjustment devices. Combined with infrared-visible light acquisition and advanced image segmentation technology, it achieves accurate identification of heat hazard sources and quantitative classification of spatial heat load. Then, by using an improved multi-objective optimization algorithm, it quickly solves the optimal airflow allocation scheme that balances the lowest energy consumption and the highest thermal comfort. Through execution and feedback correction mechanisms, it ensures continuous optimization of the treatment effect, significantly improving the accuracy, real-time performance, and energy efficiency of deep well heat hazard control.

[0148] This application also provides an electronic device applicable to the above-mentioned method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control, comprising:

[0149] Memory, used to store computer programs and data;

[0150] A processor is used to run system programs.

[0151] This application provides a computer storage medium applicable to the airflow optimization allocation method of the graded cooling and ventilation system for deep well heat hazard control, and performs graded confidentiality management of the system and data in accordance with confidentiality management requirements.

[0152] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Identical components are represented by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, while the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component. Therefore, all equivalent changes made to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control, characterized in that, Includes the following steps: S1: Construct a physical model of the deep well ventilation network, deploy multi-level cooling devices and intelligent damper adjustment system, and arrange infrared thermal imaging and visible light binocular acquisition equipment at key nodes; S2: Collect real-time thermal hazard image data of underground roadways and working faces, construct a deep well thermal hazard image dataset containing the features of "heat source-background-airflow disturbance", and perform preprocessing and enhancement. S3: Construct an improved U-Net++ image segmentation network based on a hybrid attention mechanism to perform pixel-level segmentation on real-time heat hazard images and extract the precise contours, temperature field distribution features, and geometric center coordinates of the heat source region; S4: Based on the image segmentation results, calculate the spatial heat load weight coefficient of the heat source, and combine it with the ventilation network solution model to establish an air volume optimization allocation function with the dual objectives of "lowest energy consumption and highest thermal comfort". S5: The improved multi-objective particle swarm optimization algorithm is used to solve the optimal air volume distribution scheme, and generate the cooling output command and damper opening control signal of the staged cooling device. S6: The actuator adjusts the air volume and cooling capacity according to the control signal and verifies the treatment effect through the feedback sensor. If the standard is not met, it returns to S2 for iterative correction.

2. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The image preprocessing in step S2 includes an enhancement algorithm for low-light and noise conditions downhole, called "dark channel prior and thermal radiation compensation": The acquired original image I(x) is processed to obtain the enhanced image. Represented as: ； in, Here is the dark channel plot, and t(x) is the transmittance plot. This is the ambient light compensation term based on infrared thermal radiation calibration; the algorithm solves for the optimal transmittance by minimizing the cost function E(t): ； in, For gradient operators, γ is the input source for the algorithm, and γ is the fidelity coefficient used to strike a balance between denoising and preserving thermal boundary details. By introducing a dark channel prior, the visibility in low-light areas is effectively improved, and the ambient light compensation term of infrared thermal radiation calibration is combined to correct the thermal radiation characteristics of the heat source, thus avoiding the thermal information deviation caused by simply relying on visible light enhancement.

3. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The specific algorithm process of the hybrid attention mechanism in step S3 is as follows: In U-Net++, a parallel structure is constructed by embedding convolutional block attention modules and Transformer self-attention modules within the skip connections and bottleneck layers. Its feature-weighted fusion formula is: ； in, For the input feature map, To output the feature map, For the input feature map, ⊕ denotes dot product, and ⊕ denotes residual join; The channel attention weights are generated from a shared MLP of max pooling and average pooling: ； in, and These are average pooling and max pooling, respectively. Spatial attention weights are generated based on channel dimension statistics of the feature map: ；； Where σ is the Sigmoid activation function. , , , The weight matrix is ​​a learnable matrix. This represents a 7×7 convolution operation. and These are the average and maximum values, respectively, calculated along the channel dimension. The hybrid attention mechanism suppresses downhole dust and water mist noise, highlighting the characteristics of high-temperature heat sources.

4. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The calculation of the heat load weighting coefficient of the heat source space and the heat source classification strategy in step S4 are as follows: The heat source regions segmented from the image are mapped to ventilation network nodes, and the heat load weight coefficient of node i is defined. for: ； in, This refers to the physical area corresponding to the segmented heat source pixel area. Let J be the physical area of ​​the j-th heat source region. This represents the average overheat value for the region. The heat dissipation efficiency coefficient of the heat source. This is the gas concentration correction factor associated with this node, where N is the total number of nodes. This is the gas concentration correction factor; according to The heat source is divided into three levels: >0.15 indicates a Class 1 strong heat source, 0.05< ≤0.15 indicates a secondary heat source. ≤0.05 indicates a level 3 weak heat source.

5. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 4, characterized in that, The air volume optimization allocation function established in step S4 is a constrained nonlinear programming model: Objective function: ； in, Let be the air volume of the k-th fan; The actual temperature of the i-th monitoring point; Constraints: ； in, Let be the energy consumption function of the k-th wind turbine. ,in, This usually corresponds to the main resistance term. For the corresponding secondary resistance term, This corresponds to the basic loss item; Standard operating temperature; Penalties for violating safety constraints; , , It is a dynamic weighting coefficient that is adaptively adjusted based on the current severity of heat damage.

6. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The improved multi-objective particle swarm optimization algorithm in step S5 introduces an adaptive inertia weight based on the heat source gradient: The particle velocity update formula is revised as follows: ； Among them, inertia weight Dynamic changes with heat source concentration: ； in, and These are learning factors; and They are random numbers; The speed at the current moment; and These are the individual's historical best position and the global best position, respectively. α is the temperature gradient vector of the image segmentation region corresponding to the particle's current position, guiding the particle to search for the optimal solution in the high-temperature region; α is the thermal gradient guidance coefficient. and These represent the maximum and minimum values ​​set for parameter ω, respectively, and β is the slope coefficient of the Sigmoid function. Here, is the heat source threshold, and e is the natural constant.

7. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The introduction of a "virtual-reality" two-way correction mechanism based on digital twins: A digital twin of the downhole ventilation network is established, and heat source boundary conditions obtained from image segmentation are received in real time. A computational fluid dynamics solver is used to pre-simulate the airflow distribution scheme in the twin and predict the temperature field distribution. Define the visual-physical consistency loss function. : ; in, The first term represents the true temperature field retrieved from the infrared image, and the second term represents the mass conservation residual. This refers to the physical loss weighting coefficient; The density of air; if Then, the boundary condition parameters of the physical model are corrected using the residuals until the error converges.

8. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, It also includes feedforward-feedback composite control strategies under unsteady thermal disturbances: When the image segmentation network detects a new heat source, it triggers feedforward control, directly based on the area of ​​the heat source. With strength Look up the table to obtain the initial air volume increment , For ambient temperature: ； Subsequently, the PID feedback controller adjusts the temperature deviation based on the measured values. Make fine adjustments, among which, and Let these represent the set temperature and the measured temperature, respectively. The final control variable u(t) is: ； in, This is the feedforward control variable. The time integral of the error, The rate of change of error, , , The PID parameters are tuned based on the response time of the ventilation system to achieve millisecond-level response to thermal disturbances and steady-state zero-static-error regulation.

9. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The tiered cooling and ventilation system includes a primary cooling station, secondary local coolers, and intelligent damper assemblies; the primary cooling station is located at the bottom of the air intake shaft and is responsible for reducing the intake air temperature to [a lower value]. ; The secondary local cooler is moved to a position relative to the center of the heat source by a hydraulic automatic shifting device, based on the heat source location identified in step S3. Among them: ； in, Let the cooling efficiency decrease with distance as a function. The airflow at this location. For mobile costs; The intelligent damper assembly adopts a variable cross-section structure, with the opening S corresponding to the target air volume. The relationship is non-linear: , where k is the damper characteristic coefficient and ΔP is the pressure difference across the damper.

10. The method for optimizing airflow distribution in a graded cooling and ventilation system for deep well heat hazard control according to claim 1, characterized in that, The tiered cooling and ventilation system has a self-learning and model update mechanism: Regularly compare actual governance results with model predictions to build an experience playback buffer. The parameters θ of the attention mechanism network are fine-tuned using a deep deterministic gradient algorithm, with the loss function being: ； in, s represents the current thermal damage image status, a represents the airflow allocation action, r represents the negative energy consumption and over-temperature penalty reward, and D represents the experience buffer. Through continuous iteration, the system adapts to changes in mine geological conditions and characteristic drift caused by equipment aging.

Images