Intelligent water mist dedusting and cooling control method and system fused with machine vision

Abstract

The invention relates to the technical field of dust removal, temperature reduction and cooling, and discloses an intelligent water mist dust removal, temperature reduction and cooling control method and system fused with machine vision, and the method comprises the steps: collecting and fusing image data through a multi-mode visual perception module, and recognizing a key region; constructing a neural ejection field by using an implicit neural representation technology; developing a water mist diffusion prediction model based on the conditional generative adversarial network; pID control and reinforcement learning are combined to realize parameter adaptive optimization; visualizing a preview effect by adopting a neural rendering technology; according to the system, accurate identification and positioning of a dust high-temperature area are achieved, the water resource utilization rate and the cooling speed are improved, centimeter-level spraying precision is achieved, control delay is reduced, energy consumption is reduced, and the technical problems of blind spraying, inaccurate control and response lag of a traditional system are solved.

Description

Technical Field

[0001] The present invention relates to the field of dust removal, cooling and other technical fields, and more specifically, to an intelligent fine water mist dust removal, cooling and other control method and system integrating machine vision. Background Art

[0002] Fine water mist dust removal and cooling technology is widely used due to its high efficiency and environmental protection. It produces micron-sized water mist particles to absorb dust in the air and reduce the ambient temperature. However, the traditional fine water mist dust removal and cooling system mainly relies on fixed nozzles to work according to preset parameters, and there are three main problems: lack of precise positioning ability, unable to identify areas that need to be focused on dust removal and cooling, resulting in waste of water resources and low cooling efficiency; using discrete parameters to control fine water mist injection, it is difficult to achieve continuous and smooth spatial control effects, especially in high-precision directional injection scenarios; it is impossible to accurately predict the diffusion trajectory of fine water mist in complex spaces, resulting in control strategy lag and reduced dust removal and cooling efficiency.

[0003] With the development of computer vision and artificial intelligence technology, the integration of machine vision and intelligent control technology and its application in water mist dust removal and cooling systems is expected to solve the above problems and improve system performance and resource utilization efficiency. Summary of the invention

[0004] The present invention provides an intelligent fine water mist dust removal, temperature reduction and cooling control method and system integrating machine vision, which solves the technical problems in the related technology of lack of precise positioning capability, difficulty in achieving continuous and smooth spatial control effect, and inability to accurately predict the diffusion trajectory of fine water mist in complex space.

[0005] The present invention provides an intelligent water mist dust removal and cooling control method integrating machine vision, comprising: Through the multimodal visual perception module, RGB images, depth maps and thermal imaging data are collected and integrated to segment the target area and establish a temperature distribution hotspot map to identify key areas that require dust removal and cooling. According to the identified key areas, implicit neural representation technology is used to construct neural ejection fields for these areas, mapping three-dimensional spatial coordinate points into continuous ejection parameters, including pressure, direction vector and flow rate; Generate the initial water mist distribution based on the constructed neural jet field, and build a water mist diffusion prediction model through the conditional generative adversarial network in combination with environmental conditions to predict the water mist distribution at future time points; Combining PID control algorithm with reinforcement learning method, the nozzle angle, pressure and flow rate are pre-adjusted according to the water mist diffusion prediction results to achieve active control of fine water mist injection; Using the adjusted injection parameters, a real-time visual preview of the injection effect is achieved through the volume rendering model, the difference metric between the actual effect and the expected effect is calculated, and the results are fed back to the parameter optimization module for closed-loop control and continuous optimization.

[0006] In a preferred embodiment, the multimodal visual perception module adopts an improved U-Net network structure. On the basis of the standard U-Net, an attention mechanism module is added to each encoder block. The module improves the perception accuracy of high temperature areas and high dust concentration areas by calculating the importance weights of each area in the feature map.

[0007] In a preferred embodiment, the multimodal data fusion in the multimodal visual perception module uses an attention mechanism to perform weighted fusion of different modal features: ; in, represents the fused features, Indicates The characteristics of a mode, Indicates The attention weights of the modalities satisfy , is the number of modes, and the weight coefficient is dynamically adjusted according to environmental changes.

[0008] In a preferred embodiment, the implicit neural representation technology implements the mapping function from spatial coordinates to injection parameters through a multi-layer perceptron network: ; in, Represented by parameter set Determine the mapping function, represents the horizontal coordinate, represents the vertical coordinate, represents the height coordinate, are the corresponding injection parameters, including pressure , direction vector and flow , Represents a neural network parameter set.

[0009] In a preferred embodiment, the multilayer perceptron network includes 8 fully connected layers, the middle layer uses the ReLU activation function, the output layer uses the Sigmoid activation function for normalization, and the network input uses position encoding for high-dimensional feature mapping: ; in, is the high-dimensional feature vector after position encoding, is the input coordinate, Indicates that the coordinate value Multiply The angle value obtained later is is the encoding dimension, represents the sine function, represents the cosine function, represents the frequency multiplication factor, as As the value increases, the corresponding frequency gradually increases.

[0010] In a preferred embodiment, the water mist diffusion prediction model introduces a physical constraint loss function: ; in, is the total loss function of physical constraints, is the mass conservation loss, is the momentum conservation loss, is the energy conservation loss, , , Control the weights of the mass conservation, momentum conservation, and energy conservation constraints respectively.

[0011] In a preferred embodiment, the reinforcement learning method uses a dual-time deep Q network algorithm to construct a state space Including current temperature distribution, dust concentration distribution and nozzle status, action space Including nozzle angle adjustment, pressure adjustment and flow adjustment, the reward function is: ; in, is the total reward function of the reinforcement learning algorithm, Reward for temperature control, Reward for dust removal effect, Rewards for water use, For energy consumption rewards, , , , They are temperature control, dust removal effect, water resource utilization, and energy consumption weight coefficient respectively.

[0012] In a preferred embodiment, the volume rendering model uses neural radiation field technology to represent the water mist volume: ; in, Represented by parameter set Determine the mapping function, represents the neural network parameter set, represents the horizontal coordinate, represents the vertical coordinate, represents the height coordinate, represents the viewing angle parameter, is the azimuth, is the pitch angle, Indicates the color and density of corresponding points, is an RGB color vector, is the bulk density value.

[0013] In a preferred embodiment, the difference metric is calculated by the following formula: ; in, is the difference measure, To predict the image, For the actual image, and are the image height and width, represents the row index of the image, Represents the column index of the image.

[0014] In a preferred embodiment, an intelligent fine water mist dust removal and cooling control system integrating machine vision is used to execute an intelligent fine water mist dust removal and cooling control method integrating machine vision, characterized in that it includes: Multimodal visual perception module, used to collect and fuse RGB images, depth maps and thermal imaging data to identify key areas that require dust removal and cooling; Implicit neural jet field building module for mapping 3D spatial coordinates into continuous jet parameters; Physical constraint diffusion prediction module, used to predict the diffusion trajectory of water mist in complex space; Parameter adaptive optimization module, used to optimize the nozzle control parameters according to the prediction results; The real-time effect visualization and feedback module is used to provide a visual preview of the injection effect and form a closed-loop control.

[0015] The beneficial effects of the present invention are: Improve water resource utilization, increase cooling speed, and reduce secondary pollution caused by excessive humidity; Realize continuous and precise control of water mist in three-dimensional space, break through the limitations of traditional discrete parameter control, and improve the spraying accuracy to centimeter level; The system changes from passive response to active predictive control, reducing control delays and significantly enhancing its ability to adapt to environmental changes; Automatically adjust the injection strategy for different areas to solve the blind injection problem of traditional systems and reduce energy consumption; Provides visual prediction of water mist diffusion behavior, allowing operators to intuitively understand the system working status and expected effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of an intelligent fine water mist dust removal and cooling control method integrating machine vision of the present invention; Figure 2 It is a detailed flow chart of the present invention for collecting and fusing image data to identify key areas through a multimodal visual perception module; Figure 3 is a detailed flow chart of the present invention for constructing a neural ejection field using implicit neural representation technology; Figure 4 It is a detailed flow chart of the water mist diffusion prediction model developed based on the conditional generative adversarial network of the present invention; Figure 5 It is a detailed flow chart of the present invention for combining PID control with reinforcement learning to achieve parameter adaptive optimization; Figure 6 It is a detailed flow chart of the visual preview effect using neural rendering technology of the present invention. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and 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 contents of this specification. Various examples may omit, replace or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0018] At least one embodiment of the present invention discloses an intelligent water mist dust removal and cooling control method integrating machine vision, such as Figures 1 to 6 As shown, the following steps are included: Step 1: Through the multimodal visual perception module, RGB images, depth maps and thermal imaging data are collected and integrated to segment the target area and establish a temperature distribution hotspot map to identify key areas that require dust removal and cooling; The specific steps include: Step 1.1, construct an improved U-Net network structure; It consists of two parts: the encoder and the decoder. The encoder consists of 4 downsampling blocks, each of which contains two 3×3 convolutional layers and a maximum pooling layer; the decoder consists of 4 upsampling blocks, each of which contains a transposed convolutional layer and two 3×3 convolutional layers. Compared with the standard U-Net, an attention mechanism module is added to each encoder block to improve the sensitivity to the target area. Specifically, the attention mechanism module calculates the importance weights of each area in the feature map, and can effectively identify areas with abnormal temperature and high dust concentration in high-temperature dust environments such as steelmaking workshops in steel plants, thereby improving the perception accuracy of key areas.

[0019] Step 1.2, fusion of multimodal data input; include: RGB image data: collected by industrial-grade cameras with a resolution of 1920×1080 pixels, used to capture visible light information of dust removal and cooling scenes; Depth map data: obtained through structured light or TOF depth camera, used to provide three-dimensional structure information of the scene, with a resolution of 640×480 pixels and a depth accuracy of ±1cm. For example, dust sources at different heights and distances can be accurately identified at a mining site; Thermal imaging data: collected by infrared thermal imager with a resolution of 640×480 pixels and a temperature accuracy of ±0.5°C, which can accurately locate heat source areas in construction site environments.

[0020] Step 1.3, construct a multimodal feature fusion layer; The attention mechanism is used to perform weighted fusion of different modal features: ; in, represents the fused features, Indicates The characteristics of a mode, Indicates The attention weights of the modalities satisfy , is the number of modes.

[0021] In practical applications, such as in an industrial boiler room environment, when the ambient light is insufficient, the system will automatically increase the weight of the thermal imaging data ( ); when the light is sufficient but the temperature change is not obvious, the weight of RGB image data is increased ( ) to dynamically adapt to different monitoring environments.

[0022] Step 1.4, training the target area segmentation model; Use a mixed loss function: ; in, is the total loss function for the segmentation task, is the Dice loss function, is the Focal loss function, is the boundary loss function, , , They are the weight coefficients of Dice, Focal, and boundary loss functions respectively.

[0023] In actual application, by adjusting the steel smelting workshop environment Increasing the weight of boundary loss improves the accuracy of heat source area boundary identification and achieves accurate segmentation of high temperature areas.

[0024] Step 1.5, generating a temperature distribution hotspot map based on the thermal imaging data and the segmentation results; The temperature anomaly areas in the scene are graded and labeled according to the temperature values ​​from high to low to form a hot spot map : ; in, Representing coordinates The hotspot value at Indicates the temperature value at that location. Indicates the temperature threshold, is the mapping function.

[0025] In practical applications, for example in a hot rolling workshop, the area with a temperature exceeding 250°C is marked as the highest priority, and the 200-250°C area is the second highest priority, to achieve differentiated dust removal and cooling treatment for different temperature areas.

[0026] Step 2: Based on the identified key areas, the implicit neural representation technology is used to construct the neural ejection field for these areas, mapping the three-dimensional spatial coordinate points into continuous ejection parameters, including pressure, direction vector and flow rate; The specific steps include: Step 2.1, construct a multi-layer perceptron (MLP) network structure; It contains 8 fully connected layers, with the number of nodes in each layer being [256, 512, 512, 256, 256, 128, 64, 3], the middle layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function for normalization. In the application of dust suppression in coal mines, this network structure can meet the needs of fine injection control in complex spatial environments, and ensures complex nonlinear mapping capabilities through deep networks.

[0027] Step 2.2, implement the implicit neural representation function: ; in, Represented by parameter set Determine the mapping function, represents the horizontal coordinate, represents the vertical coordinate, represents the height coordinate, are the corresponding injection parameters, including pressure , direction vector and flow , Represents a neural network parameter set.

[0028] In practical applications, for example, for cement plant dust control scenarios, the plant space can be established as a three-dimensional coordinate system, and the system can Directly output the optimal injection parameters, such as pressure 2.4MPa, direction vector and flow rate of 0.6L / min, achieving precise control of dust sources.

[0029] Step 2.3, construct the position encoding module; Map the input coordinates into high-dimensional features: ; in, is the high-dimensional feature vector after position encoding, is the input coordinate, Indicates that the coordinate value Multiply The angle value obtained later is is the encoding dimension, represents the sine function, represents the cosine function, represents the frequency multiplication factor, as As the value increases, the corresponding frequency gradually increases.

[0030] Step 2.4, training the neural ejection field network; Use a combined loss function: ; in, is the total loss function, is the mean square error loss, To smooth the loss and ensure the continuous change of parameters of adjacent spatial points, For physical constraint loss, ensure that the injection parameters meet the fluid dynamics constraints, , , are the weight coefficients of mean square error loss, smoothing loss, and physical constraint loss respectively.

[0031] Step 2.5, realize the fast query and interpolation function of neural ejection field; For any query point , directly obtain the corresponding injection parameters through the trained network .

[0032] Step 3: Generate the initial water mist distribution based on the constructed neural jet field, and build a water mist diffusion prediction model through a conditional generative adversarial network in combination with environmental conditions to predict the water mist distribution at future time points; The specific steps include: Step 3.1, construct the generator network G; The U-Net structure is adopted, the input is the initial water mist distribution and environmental conditions, and the output is the water mist distribution prediction at a future time point.

[0033] In practical applications, such as hot rolling workshops in the metallurgical industry, the generator can receive the initial injection state, workshop airflow and temperature distribution data, predict the diffusion path of water mist in the next 30 seconds, and provide a basis for optimizing injection parameters.

[0034] Step 3.2, build the discriminator network D; The PatchGAN structure is used to distinguish between the real water mist distribution and the generated water mist distribution.

[0035] In practical applications, by collecting a large number of real water mist diffusion images as training data, the discriminator can identify water mist distribution predictions that do not conform to physical laws. For example, in a construction site environment, if the prediction results show that water mist diffuses against the wind, the discriminator will give a low score, prompting the generator to optimize the prediction results.

[0036] Step 3.3, construct the conditional input module; Embed environment parameters as conditions into the build process: Air temperature distribution; Air flow velocity field; Air humidity distribution; Spatial constraint boundaries.

[0037] Step 3.4, introduce the physical constraint loss function; Ensure that the prediction results conform to the laws of fluid mechanics: ; in, is the total loss function of physical constraints, is the mass conservation loss, is the momentum conservation loss, is the energy conservation loss, , , Control the weights of the mass conservation, momentum conservation, and energy conservation constraints respectively.

[0038] In practical applications, such as in open-pit mines, the system predicts the trajectory of the water mist sprayed from high altitudes through momentum conservation loss. Consider the influence of gravity and air resistance to ensure that the predicted water mist motion trajectory conforms to the laws of physics and improve prediction accuracy.

[0039] Step 3.5, realize the time series prediction function; Build a recursive prediction framework based on the current state Predicting future states , and use the prediction results as new input to continue prediction , achieving multi-time step prediction.

[0040] Step 4, combining the PID control algorithm with the reinforcement learning method, pre-adjusting the nozzle angle, pressure and flow rate according to the water mist diffusion prediction results, and realizing active control of the fine water mist injection; The specific steps include: Step 4.1, construct state space and action space: State Space Including current temperature distribution, dust concentration distribution, nozzle status, etc.; Action Space Including nozzle angle adjustment, pressure adjustment and flow adjustment.

[0041] Step 4.2, design reward function: ; in, is the total reward function of the reinforcement learning algorithm, Reward for temperature control, Reward for dust removal effect, Rewards for water use, For energy consumption rewards, , , , They are temperature control, dust removal effect, water resource utilization, and energy consumption weight coefficient respectively.

[0042] Step 4.3, implement the dual temporal deep Q network (DDQN) algorithm; include: Construct the main Q network and the target Q network, and the network structure is a 4-layer fully connected network; Construct an experience replay buffer to store state transition samples ,in, Indicates the current state. Indicates the action to be performed. Indicates the reward received, Indicates the new state after executing the action; accomplish The greedy strategy makes a trade-off between exploration and exploitation, where represents the probability of random exploration; In practical applications, such as in the steelmaking workshop of a large steel plant, the DDQN algorithm can simultaneously control multiple nozzles to form a collaborative operation. When a sudden increase in temperature in a certain area is detected, the system can quickly learn and optimize the parameter combination of multiple nozzles to form the optimal cooling strategy. The value (0.8) is used for extensive exploration, and gradually reduced to 0.1 as experience accumulates, focusing on exploiting known high-quality strategies.

[0043] Step 4.4, integrated PID controller for fast response adjustment: ; in, To control the output, is the error signal, , , are proportional, integral and differential coefficients respectively, From the initial time to the current time The error integral of is the rate of change of error.

[0044] Step 4.5, implement the hybrid optimization strategy of reinforcement learning and PID control; Reinforcement learning methods are used for global optimization during the system stability phase, and PID control is used for rapid response in the event of sudden changes.

[0045] Step 5: Use the adjusted injection parameters to achieve real-time visual preview of the injection effect through the volume rendering model, calculate the difference between the actual effect and the expected effect, and feed the results back to the parameter optimization module for closed-loop control and continuous optimization; The specific steps include: Step 5.1, constructing a volume rendering model; Neural Radiant Field (NeRF) technology is used to represent the volume of water mist: ; in, Represented by parameter set Determine the mapping function, represents the neural network parameter set, represents the horizontal coordinate, represents the vertical coordinate, represents the height coordinate, represents the viewing angle parameter, is the azimuth, is the pitch angle, Indicates the color and density of corresponding points, is an RGB color vector, is the bulk density value.

[0046] In practical applications, the model can visualize the water mist distribution in a mine crushing station in the form of a three-dimensional volume. Operators can directly observe the water mist coverage effect on the control room screen and judge the dust removal effect.

[0047] Step 5.2, implement the ray casting algorithm; Sample points along the line of sight and compute the cumulative transparency and color: ; in, Represents light Color, represents cumulative transparency, Indicates the opacity of the current point. Indicates the color of the current point. is the number of sampling points.

[0048] Step 5.3, build a module for comparing actual effects with expected effects; Compute the difference measure between the two: ; in, is the difference measure, To predict the image, For the actual image, and are the image height and width, represents the row index of the image, Represents the column index of the image.

[0049] Step 5.4, build an interactive visualization interface; Provides multi-view preview function, allowing operators to view the prediction effect from different angles.

[0050] Step 5.5, implement parameter adjustment feedback mechanism; The effect comparison results are fed back to the parameter optimization module to form a closed-loop control.

[0051] Real-world application examples of this implementation: In the hot rolling workshop of a large steel enterprise, a large amount of heat and metal dust are generated during the rolling process of high-temperature steel. The ambient temperature is maintained at 40-60℃ all year round, and the temperature in some local areas can reach 200-300℃. The dust concentration is about 5 to 15mg / m³. The traditional fixed fine water mist dust removal and cooling system has problems such as waste of water resources, poor cooling effect, and uneven dust control. This implementation method was applied in the hot rolling workshop to realize intelligent dust removal and cooling control.

[0052] System Configuration: In this hot rolling workshop, the system is equipped with the following hardware equipment: Multimodal sensing system: 8 industrial-grade high-definition cameras (resolution 1920×1080); 6 infrared thermal imagers (resolution 640×480, temperature accuracy ±0.5℃); 4 TOF depth cameras (depth accuracy ±1cm); Intelligent injection system: 24 adjustable angle nozzles (angle adjustment range: horizontal ±180°, vertical ±90°); Control accuracy: angle ±0.5°, pressure ±0.05MPa, flow rate ±0.02L / min; Water mist particle diameter: 5-25μm; Edge computing unit: Processor: Intel i7-11800H Memory: 32GB GPU: NVIDIA RTX3080 (16GB); Storage: 1TB NVMe SSD; Monitoring display terminal: 55-inch 4K touch screen; Resolution: 3840×2160; Refresh rate: 60Hz; Implementation process example: Multimodal visual perception and implicit neural jet field example: In the application of this system in the hot rolling workshop, the multimodal visual perception module achieves accurate identification of high temperature areas and dust areas by fusing three data sources. Table 1 shows the weight distribution of different modal data under different working conditions.

[0053] Table 1: Example of adaptive adjustment of multimodal data weights;

[0054] The system dynamically adjusts the weight according to environmental changes. For example, when steam interference is severe, the depth map weight is increased ( ), reduce the weight of RGB images that are easily affected by steam ( ); in low light conditions at night, increase the thermal imaging weight ( ) to ensure the detection effect. In actual application, this adaptive weight mechanism improves the target area recognition accuracy from 78% of the traditional single mode to 95%.

[0055] In the hot rolling workshop application, the system established a three-dimensional coordinate system for the workshop space and constructed an accurate neural injection field through implicit neural representation technology. Table 2 shows an example of injection parameter mapping for a local high-temperature area.

[0056] Table 2: Example of implicit neural jet field parameter mapping (high temperature area at the mill exit);

[0057] Compared with traditional discrete parameter control, the implicit neural injection field can generate the optimal injection parameters for any point in space, achieving centimeter-level injection accuracy. Through this continuous parameter field control, the system can accurately adjust the injection parameters according to the changes in the temperature distribution of the steel plate, achieving directional and precise cooling.

[0058] Examples of physical constraint diffusion prediction and parameter adaptive optimization: In the actual application of the hot rolling workshop, the system uses the physical constraint diffusion prediction model to accurately predict the diffusion trajectory of water mist under different environmental conditions. At the same time, through the parameter adaptive optimization technology combining reinforcement learning and PID, rapid response and optimized control of complex working conditions are achieved. Tables 3 and 4 show the relevant experimental data respectively.

[0059] Table 3: Performance of the physical constraint diffusion prediction model under different environmental conditions;

[0060] Note: IoU (Intersection over Union) represents the intersection over union ratio between the predicted area and the actual area. The higher the value, the more accurate the prediction.

[0061] Table 4: Performance of parameter adaptive optimization at different stages; As the system runs longer, the control strategy is continuously optimized, and the temperature control accuracy is improved from ±3.5°C to ±1.0°C. At the same time, the water consumption is reduced from 48.5L / min to 26.8L / min, a reduction of about 45%. When unexpected operating conditions (such as sudden changes in rolling mill speed and steel temperature) occur, the system can respond quickly through the PID controller and then use reinforcement learning for fine-tuning and optimization.

[0062] Technical effect verification: In order to verify the technical effect of this implementation method, a six-month comparative test was conducted in the hot rolling workshop, and the workshop was divided into an experimental area (using this system) and a control area (using a traditional fixed parameter injection system). Tables 5 and 6 show the comparative data of water resource utilization efficiency and control response performance, respectively.

[0063] Table 5: Comparison of water resource utilization efficiency and energy consumption;

[0064] Table 6: Comparison of control response performance and dust suppression effect;

[0065] It can be seen from the above data that the application of this embodiment in the hot rolling workshop has achieved remarkable results: water resource utilization rate increased by 67.2%, cooling speed increased by 42.9%, control response delay reduced by 85.2%, and energy consumption reduced by 30.1%. These effects directly solve the problems of water resource waste, low control accuracy and response lag in traditional fine water mist dust removal and cooling systems, and verify the technical effectiveness and practical value of the present invention.

[0066] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are protected by this embodiment.

Claims

1. An intelligent water mist dust removal and cooling control method integrating machine vision, characterized in that: The following steps are involved: Through the multimodal visual perception module, RGB images, depth maps and thermal imaging data are collected and integrated to segment the target area and establish a temperature distribution hotspot map to identify key areas that require dust removal and cooling. According to the identified key areas, the implicit neural representation technology is used to construct the neural ejection field for these areas, mapping the three-dimensional spatial coordinate points into continuous ejection parameters, including pressure, direction vector and flow rate; Generate the initial water mist distribution based on the constructed neural jet field, and build a water mist diffusion prediction model through the conditional generative adversarial network in combination with environmental conditions to predict the water mist distribution at future time points; Combining PID control algorithm with reinforcement learning method, the nozzle angle, pressure and flow rate are pre-adjusted according to the water mist diffusion prediction results to achieve active control of fine water mist injection; Using the adjusted injection parameters, a real-time visual preview of the injection effect is achieved through the volume rendering model, the difference metric between the actual effect and the expected effect is calculated, and the results are fed back to the parameter optimization module for closed-loop control and continuous optimization.

2. According to the intelligent water mist dust removal and cooling control method integrating machine vision in claim 1, it is characterized in that: The multimodal visual perception module adopts an improved U-Net network structure. On the basis of the standard U-Net, an attention mechanism module is added to each encoder block. The module improves the perception accuracy of high-temperature areas and high dust concentration areas by calculating the importance weights of each area in the feature map.

3. According to the intelligent water mist dust removal and cooling control method integrating machine vision as described in claim 1, it is characterized in that: The multimodal data fusion in the multimodal visual perception module adopts the attention mechanism to perform weighted fusion of different modal features: ； in, represents the fused features, Indicates The characteristics of a mode, Indicates The attention weights of the modalities satisfy , is the number of modes, and the weight coefficient is dynamically adjusted according to environmental changes.

4. According to the intelligent water mist dust removal and cooling control method integrating machine vision as described in claim 1, it is characterized in that: The implicit neural representation technology realizes the mapping function from spatial coordinates to injection parameters through a multi-layer perceptron network: ； in, Represented by parameter set Determine the mapping function, represents the horizontal coordinate, represents the vertical coordinate, represents the height coordinate, are the corresponding injection parameters, including pressure , direction vector and flow , Represents a neural network parameter set.

5. According to the intelligent water mist dust removal and cooling control method integrating machine vision as described in claim 4, it is characterized in that: The multi-layer perceptron network contains 8 fully connected layers, the middle layer uses the ReLU activation function, the output layer uses the Sigmoid activation function for normalization, and the network input uses position encoding for high-dimensional feature mapping: ； in, is the high-dimensional feature vector after position encoding, is the input coordinate, Indicates that the coordinate value Multiply The angle value obtained later is is the encoding dimension, represents the sine function, represents the cosine function, represents the frequency multiplication factor, as As the value increases, the corresponding frequency gradually increases.

6. The intelligent water mist dust removal and cooling control method integrating machine vision according to claim 1 is characterized in that: The water mist diffusion prediction model introduces a physical constraint loss function: ； in, is the total loss function of physical constraints, is the mass conservation loss, is the momentum conservation loss, is the energy conservation loss, , , Control the weights of the mass conservation, momentum conservation, and energy conservation constraints respectively.

7. The intelligent water mist dust removal and cooling control method integrating machine vision according to claim 1 is characterized in that: The reinforcement learning method adopts a dual-sequence deep Q network algorithm to construct a state space Including current temperature distribution, dust concentration distribution and nozzle status, action space Including nozzle angle adjustment, pressure adjustment and flow adjustment, the reward function is: ； in, is the total reward function of the reinforcement learning algorithm, Reward for temperature control, Reward for dust removal effect, Rewards for water use, For energy consumption rewards, , , , They are temperature control, dust removal effect, water resource utilization, and energy consumption weight coefficient respectively.

8. The intelligent water mist dust removal and cooling control method integrating machine vision according to claim 1 is characterized in that: The volume rendering model uses neural radiation field technology to represent the water fog volume: ； in, Represented by parameter set Determine the mapping function, represents the neural network parameter set, represents the horizontal coordinate, represents the vertical coordinate, represents the height coordinate, represents the viewing angle parameter, is the azimuth, is the pitch angle, Indicates the color and density of corresponding points, is an RGB color vector, is the bulk density value.

9. The intelligent water mist dust removal and cooling control method integrating machine vision according to claim 1 is characterized in that: The difference metric is calculated by the following formula: ； in, is the difference measure, To predict the image, For the actual image, and are the image height and width, represents the row index of the image, Represents the column index of the image.

10. An intelligent fine water mist dust removal and cooling control system integrating machine vision, used to execute an intelligent fine water mist dust removal and cooling control method integrating machine vision as claimed in any one of claims 1 to 9, characterized in that: include: Multimodal visual perception module, used to collect and fuse RGB images, depth maps and thermal imaging data to identify key areas that require dust removal and cooling; Implicit neural jet field building module for mapping 3D spatial coordinates into continuous jet parameters; Physical constraint diffusion prediction module, used to predict the diffusion trajectory of water mist in complex space; Parameter adaptive optimization module, used to optimize the nozzle control parameters according to the prediction results; The real-time effect visualization and feedback module is used to provide a visual preview of the injection effect and form a closed-loop control.

Images