A radiation air conditioning system energy-saving control method based on human body infrared image recognition

Abstract

This invention discloses an energy-saving control method for a radiant air conditioning system based on human infrared image recognition, comprising the following steps: Step 1: acquiring human infrared images, indoor environmental parameters, and radiant air conditioning system operating parameters of the target indoor space; Step 2: performing human thermal target segmentation, human region localization, and human thermal state instance recognition using an improved YOLOv8-Seg infrared human body recognition network; Step 3: constructing an indoor human thermal state distribution map using region mapping and thermal state rasterization methods; Step 4: calculating the radiant air conditioning demand representation vector for each air conditioning control zone; Step 5: performing energy-saving control decisions using a thermally constrained zoning MPC algorithm; and Step 6: performing differentiated energy-saving control on the radiant air conditioning system. This invention improves the zoning control adaptability and energy-saving operation stability of the radiant air conditioning system through the improved YOLOv8-Seg infrared human body recognition network and the thermally constrained zoning MPC algorithm.

Description

Technical Field

[0001] This invention relates to the field of air conditioning and building energy-saving control technology, and in particular to an energy-saving control method for a radiant air conditioning system based on human infrared image recognition. Background Technology

[0002] With the increasing demand for building energy conservation and indoor thermal comfort control, zoned operation control and personnel status sensing technology for radiant air conditioning systems have received widespread attention. Existing radiant air conditioning systems mainly rely on indoor air temperature, return water temperature, supply water flow rate, or the opening of radiant terminal valves for operation regulation. Some systems also use ordinary human body sensors to determine whether someone is in the room. However, in practical applications, the following problems are commonly encountered: The collected ambient temperature and equipment operating parameters can only reflect the overall thermal environment of the target indoor space, making it difficult to accurately characterize the number, location, outline area, and thermal radiation intensity of human targets within each air conditioning control zone. This results in a lack of refined control basis for radiant air conditioning systems between occupied, unoccupied, and high-heat-load zones. Ordinary human presence detection methods can only output a simple "occupied" or "unoccupied" result, failing to accurately correlate human targets with air conditioning control zone numbers and making it difficult to eliminate short-lived or incompletely obscured human targets. This easily leads to frequent adjustments at the radiant terminals and ineffective cooling and heating. Existing radiant air conditioning control methods mostly employ fixed thresholds or single temperature deviation control strategies, failing to combine indoor human thermal state distribution maps, human heat load characteristics, environmental deviation characteristics, radiant surface heat transfer characteristics, and historical operating energy consumption characteristics for energy-saving control decisions. This results in a mismatch between target water supply temperature, target water supply flow rate, target valve opening, and target operating mode adjustments and actual human thermal needs, affecting the energy efficiency, zone adaptability, and thermal comfort stability of the radiant air conditioning system.

[0003] Therefore, how to provide an energy-saving control method for radiant air conditioning systems based on human infrared image recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an energy-saving control method for radiant air conditioning systems based on human infrared image recognition. This invention fully utilizes an improved YOLOv8-Seg infrared human body recognition network, a region mapping and thermal state rasterization method, and a thermal constraint zoning MPC algorithm. It describes in detail the energy-saving control process of constructing an indoor thermal state distribution map based on human infrared recognition results and generating radiant air conditioning zoning control quantities according to the radiant air conditioning demand representation vector. It has the advantages of high accuracy in human thermal state recognition, strong adaptability of air conditioning control zoning, and good energy-saving effect of radiant air conditioning systems.

[0005] An energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to an embodiment of the present invention includes the following steps: Step 1: Collect infrared images of the human body, indoor environmental parameters, and operating parameters of the radiant air conditioning system in the target indoor space; Step 2: Input the human infrared image into the improved YOLOv8-Seg infrared human body recognition network to perform human thermal target segmentation, human region localization, and human thermal state instance recognition to obtain the human infrared recognition results for each human target; the improved YOLOv8-Seg infrared human body recognition network introduces temperature difference hierarchical residual convolution units, partitioned position embedding fusion units, and human thermal state instance output heads. Step 3: Based on the results of human infrared recognition, construct an indoor thermal distribution map of people using region mapping and thermal state rasterization methods; Step 4: Based on the indoor thermal state distribution map of people, indoor environmental parameters, and radiant air conditioning system operating parameters, calculate the radiant air conditioning demand representation vector for each air conditioning control zone; Step 5: Based on the radiant air conditioning demand representation vector, energy-saving control decisions are made for each air conditioning control zone using the thermal constraint partitioning MPC algorithm to generate radiant air conditioning zone control quantities; Step Six: Based on the radiant air conditioning zoning control values ​​and the indoor thermal distribution map of occupants, implement differentiated energy-saving control for the radiant air conditioning system.

[0006] Optionally, the indoor environmental parameters include indoor air temperature, indoor relative humidity, radiant surface temperature, indoor carbon dioxide concentration, and indoor area location identification. The radiant air conditioning system operating parameters include supply water temperature, return water temperature, supply water flow rate, radiant terminal valve opening, radiant air conditioning operating mode, and air conditioning control zone number.

[0007] Optionally, the improved YOLOv8-Seg infrared human body recognition network includes a basic convolutional module, a backbone feature extraction module, a Neck multi-scale feature fusion module, and a Head multi-task output module; The Backbone feature extraction module includes a temperature difference hierarchical residual convolution unit, a C2f unit, and an SPPF unit; the Neck multi-scale feature fusion module includes a standard multi-scale feature fusion path and a partition position embedding fusion unit; the Head multi-task output module includes a detection branch, a segmentation branch, and a human thermal state instance output head.

[0008] Optionally, step two specifically includes: A shallow infrared feature map is obtained by performing a 3×3 convolution mapping on the human infrared image through the basic convolution module. The shallow infrared feature map is input into the temperature difference hierarchical residual convolution unit to generate the temperature difference hierarchical enhanced feature map. The temperature difference hierarchical enhancement feature map is input into the C2f unit, and channel splitting, bottleneck convolution, cross-layer feature splicing and channel fusion are performed to obtain the infrared human body intermediate feature map. The infrared human body intermediate feature map is input into the SPPF unit, and multi-level max pooling, pooling feature stitching and convolution fusion are performed to obtain multi-scale infrared human body feature maps. The multi-scale infrared human body feature map is input into the standard multi-scale feature fusion path, and multi-scale fused feature map is obtained through upsampling, downsampling, channel stitching and C2f fusion processing. The multi-scale fused feature map is input into the partition location embedding fusion unit. For the multi-scale fused feature map at each scale, the horizontal and vertical pixel coordinates of each pixel position are normalized by the feature map height and feature map width respectively to obtain the position coordinate matrix. The location coordinate matrix is ​​convolved with a 1×1 convolution to generate a partition location embedding feature map; Multi-scale fused feature maps at various scales and partition location embedded feature maps are channel-separated and convolutionally fused to obtain partition-aware fused feature maps. The partition-aware fusion feature maps at various scales are fused to generate the Head input fusion feature map; The Head input fused feature map is output as a human target detection result through the detection branch. The human target detection result includes human target confidence and human target location. The human target location includes the coordinates of the human target center point, the width of the human target detection box, and the height of the human target detection box. Based on the confidence level of human targets, determine the number of human targets in the target indoor space; The Head input is fused into the feature map input to the segmentation branch, and R prototype masks are generated through 3×3 convolution, batch normalization, SiLU activation function, upsampling and 1×1 convolution; Based on the coordinates of the center point of the human target, the target position feature vector is extracted from the fused feature map of the Head input and linear prediction is performed to obtain the mask coefficient vector. Based on the mask coefficient vector, the prototype masks are linearly combined, and the human instance mask is generated by the Sigmoid function. The human body instance mask is binarized, and the pixel positions with a binarized pixel value of 1 are determined as the human body target contour region corresponding to the human body target. The Head input is fused with the feature map and the human instance mask, and then mask pooling is performed to obtain the human instance feature vector. The feature vector of the human body instance is input into the output header of the human body thermal state instance. Through thermal intensity regression, zoning classification and control effectiveness determination, the human body thermal radiation intensity value, air conditioning control zoning number and control effectiveness identifier of each human body target are obtained.

[0009] Optionally, the step of inputting the shallow infrared feature map into the temperature difference hierarchical residual convolution unit to generate the temperature difference hierarchical enhanced feature map specifically includes: The temperature difference hierarchical residual convolutional unit includes a basic temperature difference response branch, a hierarchical residual correction branch, an original structure recharge branch, and a hierarchical fusion layer; The shallow infrared feature map is subjected to convolution operations with kernel sizes of 3×3 and 5×5 respectively by the basic temperature difference response branch, and the kernels are spliced ​​in the channel dimension. Channel fusion is performed by 1×1 convolution to obtain the basic temperature difference response feature map. The shallow infrared feature map is aligned by channel through 1×1 convolution to obtain the shallow aligned feature map. The first-level structural residual between the basic temperature difference response feature map and the shallow alignment feature map is calculated by using a hierarchical residual correction branch, and convolution correction is performed using 1×1 convolution to obtain the first-level residual correction feature map. The basic temperature difference response feature map and the first-level residual correction feature map are subjected to residual compensation to obtain the first hierarchical temperature difference feature map. Calculate the second-level structural residual between the first hierarchical temperature difference feature map and the shallow alignment feature map, and perform convolution correction on the second-level structural residual through 1×1 convolution to obtain the second-level residual correction feature map; The first hierarchical temperature difference feature map and the second-level residual correction feature map are subjected to residual compensation to obtain the second hierarchical temperature difference feature map. The shallow alignment feature map and the second hierarchical temperature difference feature map are spliced ​​together by the original structure re-injection branch, and the overall human body structure information is re-injected through 1×1 convolution to obtain the structure re-injection feature map. By performing residual fusion of the second-order temperature difference feature map and the structural recharge feature map through a hierarchical fusion layer, a temperature difference hierarchical enhancement feature map is obtained.

[0010] Optionally, step three specifically includes: Human targets with a control effectiveness flag of 1 are selected to obtain the set of human targets participating in radiation air conditioning zoning control; Based on the air conditioning control zone number, the human targets participating in the radiation air conditioning zone control are mapped to the corresponding air conditioning control zone to obtain the zone human target set of each air conditioning control zone. For each air conditioning control zone, initialize the zone grid matrix, and map the human target position, human target contour region and human thermal radiation intensity value in the zone human target set to the zone grid matrix to obtain the zone human thermal state grid matrix. The partitioned human thermal state grid matrix includes several grid cells; If a grid cell overlaps with the human target contour area, the human occupancy flag of the current grid cell is set to 1; otherwise, it is set to 0. If there is a grid cell in the air conditioning control zone with a human occupancy identifier of 1, then the human presence status of the corresponding air conditioning control zone is determined to be 1; otherwise, it is determined to be 0. The number of human targets in the human target set of each air conditioning control zone is counted, and the ratio of the number of human targets to the number of grid cells in the human thermal state grid matrix of the zone is calculated to obtain the human density state of each air conditioning control zone. Based on the human body thermal radiation intensity value and the number of grid cells covered by the human body target outline area in each air conditioning control zone, the human body thermal radiation intensity status of each air conditioning control zone is calculated. Based on the human thermal state grid matrix of the current and previous data collection times, the grid state change of each air conditioning control zone is calculated, and the change of human activity in each air conditioning control zone is determined according to the grid state change. Initialize the partition state matrix, and write the human presence status, human density status, human thermal radiation intensity status, and human activity change status of each air conditioning control zone into the partition state matrix according to the air conditioning control zone number. By associating the human thermal state grid matrix of each air conditioning control zone with the zone state matrix, an indoor human thermal state distribution map is obtained.

[0011] Optionally, step four specifically includes: The thermal load characteristics of personnel are calculated based on the product of human body density state and human body thermal radiation intensity state. The environmental deviation characteristics are calculated based on the absolute value of the difference between the indoor air temperature and the target set temperature. The heat transfer characteristics of the radiant surface are calculated based on the absolute value of the difference between the indoor air temperature and the radiant surface temperature. The state of human presence is used as a partitioning feature; The historical operating energy consumption characteristics are calculated by multiplying the absolute value of the difference between the supply water temperature and the return water temperature, the supply water flow rate, and the opening degree of the radiant terminal valve. By sequentially concatenating the characteristics of personnel heat load, environmental deviation, radiant surface heat transfer, zoning occupancy, and historical operating energy consumption, a radiant air conditioning demand characterization vector for the corresponding air conditioning control zone is obtained.

[0012] Optionally, the thermally constrained partitioning MPC algorithm is specifically as follows: Based on the radiant air conditioning demand representation vector, the initial state of thermal constraint prediction for each air conditioning control zone in the current control cycle is constructed. Set the prediction step size and construct the zone control sequence for each air conditioning control zone within the prediction step size. The zone control sequence includes the target water supply temperature sequence, the target water supply flow rate sequence, the target valve opening sequence, and the target operating mode sequence. Using the zoned control sequence, the radiative air conditioning regulation response sequence of each air conditioning control zone within the prediction step size is calculated; Based on the initial state of thermal constraint prediction and the sequence of radiant air conditioning regulation response, the thermal demand state sequence of each air conditioning control zone within the prediction step is predicted by a linear recursive prediction method. The thermal comfort deviation term for each air conditioning control zone is calculated based on the difference between the predicted zone thermal demand state sequence and the target thermal demand state. Based on the target supply water temperature sequence, return water temperature, target supply water flow sequence, and target valve opening sequence, calculate the operating energy consumption constraints for each air conditioning control zone; Based on the changes between adjacent prediction steps in the target water supply temperature sequence, the changes between adjacent prediction steps in the target water supply flow sequence, and the changes between adjacent prediction steps in the target valve opening sequence, the control smoothing constraint terms for each air conditioning control zone are calculated. By combining thermal comfort deviation, operating energy consumption constraint, and control smoothness constraint, a thermal constraint partitioning MPC objective function is constructed. Under the constraints of the water supply temperature range, water supply flow range, radiant terminal valve opening range, and radiant air conditioning operation mode, the MPC objective function of the thermal constraint zone is solved by rolling optimization to obtain the optimal control sequence for each air conditioning control zone. Extract the target water supply temperature, target water supply flow rate, target valve opening degree, and target operating mode corresponding to the current control cycle from the optimal control sequence; Based on the characteristics of personnel heat load, environmental deviation, and zoning occupancy, the target adjustment priority of each air conditioning control zone is calculated. The target water supply temperature, target water supply flow rate, target valve opening, target operating mode, and target adjustment priority of each air conditioning control zone are combined to form the radiant air conditioning zone control quantities.

[0013] Optionally, step six specifically includes: Calculate the average human body density status of each air conditioning control zone to obtain the human body density judgment threshold. Calculate the average value of the human body thermal radiation intensity state in each air conditioning control zone to obtain the thermal radiation intensity judgment threshold. If the presence of a human body is 0, the corresponding air conditioning control zone will be designated as an unmanned zone, and the unmanned zone will be controlled to enter the heat preservation operation state according to the target operation mode. If the human body exists in state 1, and the human body density is less than the human body density judgment threshold, and the human body heat radiation intensity is less than the heat radiation intensity judgment threshold, then the corresponding air conditioning control zone will be determined as a low heat load zone, and the low heat load zone will be controlled to operate according to the target water supply flow, target valve opening and target operating mode. If the human body existence state is 1, and the human body density state is greater than or equal to the human body density judgment threshold, or the human body thermal radiation intensity state is greater than or equal to the thermal radiation intensity judgment threshold, then the corresponding air conditioning control zone will be determined as a high heat load zone, and the target water supply temperature, target water supply flow rate, target valve opening and target operating mode will be executed in the high heat load zone according to the target adjustment priority. For high heat load zones, calculate the average value of personnel activity changes in each air conditioning control zone to obtain activity change judgment values; If the change in personnel activity in a high heat load zone exceeds the activity change judgment value, then increase the target water supply flow and target valve opening for the corresponding high heat load zone. If the personnel activity change status of the high heat load zone is less than or equal to the activity change judgment value, then the target water supply flow and target valve opening of the corresponding high heat load zone shall be maintained.

[0014] The beneficial effects of this invention are: This invention inputs human infrared images into an improved YOLOv8-Seg infrared human body recognition network, and introduces temperature difference hierarchical residual convolutional units, partitioned position embedding fusion units, and human thermal state instance output heads into the network. This enhances the ability to extract local temperature difference features from human infrared images, the ability to perceive spatial partitions of human targets, and the ability to instantiate and output human thermal states, thereby improving the accuracy of human thermal state recognition in scenarios with weak infrared textures, low temperature differences, and unclear boundaries. Furthermore, through region mapping and thermal state rasterization methods, human targets participating in radiant air conditioning zoning control are mapped to corresponding air conditioning control zones, forming a zoned human thermal state raster matrix and a zoned state matrix. This allows the indoor personnel thermal state distribution map to simultaneously represent the human presence state, human density state, human thermal radiation intensity state, and personnel activity change state, enhancing the correlation between human body recognition results and radiant air conditioning zoning control. In the energy-saving control decision-making stage, a radiant air conditioning demand representation vector is calculated based on the indoor personnel thermal state distribution map, indoor environmental parameters, and radiant air conditioning system operating parameters. Energy-saving control decisions are then executed for each air conditioning control zone using the thermal constraint zoning MPC algorithm, improving the zoning adaptability, energy-saving control accuracy, and operational stability of the radiant air conditioning system. During the implementation phase, differentiated energy-saving control is applied to unoccupied zones, low-heat-load zones, and high-heat-load zones based on human density thresholds, thermal radiation intensity thresholds, and changes in human activity. This reduces ineffective cooling and heating in unoccupied and low-heat-load areas and improves the responsiveness of high-heat-load areas. Therefore, this invention is of great significance for reducing the energy consumption of radiant air conditioning systems, maintaining indoor thermal comfort, and improving the level of intelligent control of building air conditioning. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of an energy-saving control method for a radiant air conditioning system based on human infrared image recognition proposed in this invention; Figure 2 This is a flowchart of the temperature difference hierarchical residual convolution unit structure in the energy-saving control method for a radiant air conditioning system based on human infrared image recognition proposed in this invention. Figure 3 This is a flowchart of the thermal constraint partitioning MPC algorithm in an energy-saving control method for a radiant air conditioning system based on human infrared image recognition proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 An energy-saving control method for a radiant air conditioning system based on human infrared image recognition includes the following steps: Step 1: Collect infrared images of the human body, indoor environmental parameters, and operating parameters of the radiant air conditioning system in the target indoor space; Step 2: Input the human infrared image into the improved YOLOv8-Seg infrared human body recognition network to perform human thermal target segmentation, human region localization, and human thermal state instance recognition to obtain the human infrared recognition results for each human target; the improved YOLOv8-Seg infrared human body recognition network introduces temperature difference hierarchical residual convolution units, partitioned position embedding fusion units, and human thermal state instance output heads. Step 3: Based on the results of human infrared recognition, construct an indoor thermal distribution map of people using region mapping and thermal state rasterization methods; Step 4: Based on the indoor thermal state distribution map of people, indoor environmental parameters, and radiant air conditioning system operating parameters, calculate the radiant air conditioning demand representation vector for each air conditioning control zone; Step 5: Based on the radiant air conditioning demand representation vector, energy-saving control decisions are made for each air conditioning control zone using the thermal constraint partitioning MPC algorithm to generate radiant air conditioning zone control quantities; Step Six: Based on the radiant air conditioning zoning control values ​​and the indoor thermal distribution map of occupants, implement differentiated energy-saving control for the radiant air conditioning system.

[0018] In this embodiment, the indoor environmental parameters include indoor air temperature, indoor relative humidity, radiant surface temperature, indoor carbon dioxide concentration, and indoor area location identification. The operating parameters of the radiant air conditioning system include supply water temperature, return water temperature, supply water flow rate, radiant terminal valve opening, radiant air conditioning operating mode, and air conditioning control zone number.

[0019] In this embodiment, the improved YOLOv8-Seg infrared human body recognition network includes a basic convolutional module, a backbone feature extraction module, a Neck multi-scale feature fusion module, and a Head multi-task output module. The Backbone feature extraction module includes a temperature difference hierarchical residual convolution unit, a C2f unit, and an SPPF unit; the Neck multi-scale feature fusion module includes a standard multi-scale feature fusion path and a partition position embedding fusion unit; the Head multi-task output module includes a detection branch, a segmentation branch, and a human thermal state instance output head.

[0020] In this embodiment, step two specifically includes: A shallow infrared feature map is obtained by performing a 3×3 convolution mapping on the human infrared image through the basic convolution module. The shallow infrared feature map is input into the temperature difference hierarchical residual convolution unit to generate the temperature difference hierarchical enhanced feature map. The temperature difference hierarchical enhancement feature map is input into the C2f unit, and channel splitting, bottleneck convolution, cross-layer feature splicing and channel fusion are performed to obtain the infrared human body intermediate feature map. The infrared human body intermediate feature map is input into the SPPF unit, and multi-level max pooling, pooling feature concatenation and convolution fusion are performed to obtain multi-scale infrared human body feature maps; among them, the multi-scale infrared human body feature maps include shallow infrared human body feature maps, mid-level infrared human body feature maps and deep infrared human body feature maps. The multi-scale infrared human body feature map is input into the standard multi-scale feature fusion path, and multi-scale fused feature map is obtained through upsampling, downsampling, channel stitching and C2f fusion processing. The multi-scale fused feature map is input into the partition location embedding fusion unit. For the multi-scale fused feature map at each scale, the horizontal and vertical pixel coordinates of each pixel position are normalized by the feature map height and feature map width respectively to obtain the position coordinate matrix. The location coordinate matrix is ​​convolved with a 1×1 convolution to generate a partition location embedding feature map; The multi-scale fused feature maps at various scales are concatenated with the partition location embedding feature maps in the channel dimension and then fused through 1×1 convolution to obtain the partition-aware fused feature map. Among them, the partition-aware fusion feature map includes shallow partition-aware fusion feature map, mid-level partition-aware fusion feature map and deep partition-aware fusion feature map; The partition-aware fusion feature maps at various scales are fused to generate the Head input fusion feature map. Specifically, the shallow partition-aware fusion feature map, the middle partition-aware fusion feature map, and the deep partition-aware fusion feature map are scale-aligned and concatenated in the channel dimension. They are then fused using a 1×1 convolution to obtain the Head input fusion feature map. The Head input fused feature map is output as a human target detection result through the detection branch. The human target detection result includes human target confidence and human target location. The human target location includes the coordinates of the human target center point, the width of the human target detection box, and the height of the human target detection box. Based on the confidence level of human targets, the number of human targets in the target indoor space is determined, specifically by counting the number of human targets whose confidence level exceeds the confidence level threshold. The Head input is fused into the feature map input to the segmentation branch, and R prototype masks are generated through 3×3 convolution, batch normalization, SiLU activation function, upsampling and 1×1 convolution; Based on the coordinates of the center point of the human target, the target position feature vector is extracted from the Head input fusion feature map and linear prediction is performed to obtain the mask coefficient vector; wherein, the mask coefficient vector includes R components. Based on the mask coefficient vector, the prototype masks are linearly combined, and the human instance mask is generated by the Sigmoid function. The human instance mask is binarized, and the pixel positions with a binarized pixel value of 1 are determined as the human target contour region of the corresponding human target. In the binarization process, the pixel positions in the human instance mask that are greater than or equal to 0.5 are mapped to 1, and the pixel positions in the human instance mask that are less than 0.5 are mapped to 0. The Head input is fused with the feature map and the human instance mask, and then mask pooling is performed to obtain the human instance feature vector. The feature vector of a human instance is input into the output header of the human thermal state instance. Through thermal intensity regression, zoning classification, and control effectiveness determination, the human thermal radiation intensity value, air conditioning control zoning number, and control effectiveness identifier of each human target are obtained, specifically: Human body instance feature vectors are processed through fully connected mapping, SiLU activation function, and fully connected mapping to generate human body thermal radiation intensity values; The human instance feature vector is mapped using a fully connected method and normalized using the Softmax function to obtain the partition assignment probability vector; based on the partition assignment probability vector, the vector index corresponding to the maximum probability value is used as the air conditioning control partition number of the corresponding human target. The feature vector of a human instance is processed through a fully connected mapping, SiLU activation function, and Sigmoid activation function to output the control effectiveness probability. Based on the control effectiveness probability, the control effectiveness indicator for each human target is determined; the control effectiveness indicator includes 1: the human target participates in the radiation air conditioning zoning control, and 0: the human target does not participate in the radiation air conditioning zoning control.

[0021] In this embodiment, the shallow infrared feature map is input into the temperature difference hierarchical residual convolution unit to generate the temperature difference hierarchical enhanced feature map, specifically including: The temperature difference hierarchical residual convolutional unit includes a basic temperature difference response branch, a hierarchical residual correction branch, an original structure recharge branch, and a hierarchical fusion layer; The shallow infrared feature map is subjected to convolution operations with kernel sizes of 3×3 and 5×5 respectively by the basic temperature difference response branch, and the kernels are spliced ​​in the channel dimension. Channel fusion is performed by 1×1 convolution to obtain the basic temperature difference response feature map. The shallow infrared feature map is aligned by channel through 1×1 convolution to obtain the shallow aligned feature map. The first-level structural residual between the basic temperature difference response feature map and the shallow alignment feature map is calculated by using a hierarchical residual correction branch, and convolution correction is performed using 1×1 convolution to obtain the first-level residual correction feature map. The basic temperature difference response feature map and the first-level residual correction feature map are subjected to residual compensation to obtain the first hierarchical temperature difference feature map. Calculate the second-level structural residual between the first hierarchical temperature difference feature map and the shallow alignment feature map, and perform convolution correction on the second-level structural residual through 1×1 convolution to obtain the second-level residual correction feature map; The first hierarchical temperature difference feature map and the second-level residual correction feature map are subjected to residual compensation to obtain the second hierarchical temperature difference feature map. The shallow alignment feature map and the second hierarchical temperature difference feature map are spliced ​​together by the original structure re-injection branch, and the overall human body structure information is re-injected through 1×1 convolution to obtain the structure re-injection feature map. By performing residual fusion of the second-order temperature difference feature map and the structural recharge feature map through a hierarchical fusion layer, a temperature difference hierarchical enhancement feature map is obtained.

[0022] In this invention, the improved YOLOv8-Seg infrared human body recognition network maintains the same basic framework as the standard YOLOv8-Seg network, both consisting of a basic convolutional module, a backbone feature extraction module, a Neck multi-scale feature fusion module, and a Head multi-task output module. In the standard YOLOv8-Seg network, the basic convolutional module performs convolutional mapping on the input image and generates shallow feature maps. The backbone feature extraction module extracts target semantic features and spatial context features at different levels through C2f and SPPF units. The Neck multi-scale feature fusion module achieves the fusion of shallow detail features, mid-level structural features, and deep semantic features through upsampling, downsampling, channel concatenation, and C2f fusion processing. The Head multi-task output module outputs target confidence and target location through the detection branch, and generates prototype masks and predicted mask coefficient vectors through the segmentation branch, thereby obtaining the human instance mask and the human target contour region.

[0023] Based on the standard YOLOv8-Seg network structure, the improved YOLOv8-Seg infrared human body recognition network has made the following structural improvements to meet the needs of human infrared image recognition and radiant air conditioning zone control. A temperature difference hierarchical residual convolution unit is added to the Backbone feature extraction module. This unit generates a temperature difference hierarchical enhanced feature map through basic temperature difference response, hierarchical residual correction, original structure refeedback, and hierarchical fusion. A zone location embedding fusion unit is added to the Neck multi-scale feature fusion module. This unit embeds the spatial location representation of the human target into the multi-scale fusion feature map through the location coordinate matrix, zone location embedding feature map, and zone perception fusion feature map, forming the Head input fusion feature map. A human thermal state instance output header is added to the Head multi-task output module. This header obtains the human instance feature vector through mask pooling and outputs the human thermal radiation intensity value, air conditioning control zone number, and control effectiveness identifier after thermal intensity regression, zone classification, and control effectiveness determination.

[0024] The temperature difference hierarchical residual convolutional unit enhances the representation of human thermal target features under conditions of weak texture, low temperature difference, and unclear boundaries in human infrared images through basic temperature difference response, hierarchical residual correction, and original structure refeeding, thereby improving the stability of human thermal target segmentation and human region localization. The partitioned location embedding fusion unit enables the network to form a spatially related partitioned perception fusion feature map during multi-scale feature fusion, improving the matching accuracy between human targets and air conditioning control partition numbers. The human thermal state instance output header extends the detection and segmentation results of the standard YOLOv8-Seg network into human infrared recognition results that can be directly used for radiant air conditioning partition control, jointly supporting the subsequent construction of indoor human thermal state distribution maps and energy-saving control of radiant air conditioning systems.

[0025] In this embodiment, step three specifically includes: Human targets with a control effectiveness flag of 1 are selected to obtain the set of human targets participating in radiation air conditioning zoning control; Based on the air conditioning control zone number, the human targets participating in the radiation air conditioning zone control are mapped to the corresponding air conditioning control zone to obtain the zone human target set of each air conditioning control zone. For each air conditioning control zone, initialize the zone grid matrix, and map the human target position, human target contour region and human thermal radiation intensity value in the zone human target set to the zone grid matrix to obtain the zone human thermal state grid matrix. The partitioned human thermal state grid matrix includes several grid cells; If a grid cell overlaps with the human target contour area, the human occupancy flag of the current grid cell is set to 1; otherwise, it is set to 0. If there is a grid cell in the air conditioning control zone with a human occupancy identifier of 1, then the human presence status of the corresponding air conditioning control zone is determined to be 1; otherwise, it is determined to be 0. The number of human targets in the human target set of each air conditioning control zone is counted, and the ratio of the number of human targets to the number of grid cells in the human thermal state grid matrix of the zone is calculated to obtain the human density state of each air conditioning control zone. Based on the human body thermal radiation intensity value and the number of grid cells covered by the human body target outline area in each air conditioning control zone, the human body thermal radiation intensity status of each air conditioning control zone is calculated. The calculation method for the human body thermal radiation intensity state is as follows: For the m-th air conditioning control zone, obtain the human body thermal radiation intensity value of each human target in the corresponding human target set, and count the number of grid cells covered by the human target contour region of each human target; multiply the human body thermal radiation intensity value of each human target by the number of grid cells covered by the corresponding human target contour region to obtain the thermal radiation contribution value of the corresponding human target; sum the thermal radiation contribution values ​​of each human target in the m-th air conditioning control zone to obtain the total thermal radiation contribution value of the zone; calculate the ratio of the total thermal radiation contribution value of the zone to the total number of grid cells covered by the human target contour region in the m-th air conditioning control zone to obtain the human body thermal radiation intensity state of the m-th air conditioning control zone. Based on the human thermal state grid matrix of the current and previous data collection times, the grid state change of each air conditioning control zone is calculated, and the change of human activity in each air conditioning control zone is determined according to the grid state change. The specific calculation method for changes in personnel activity status is as follows: For the m-th air conditioning control zone, obtain the zone's human thermal state grid matrix at the current acquisition time and the zone's human thermal state grid matrix at the previous acquisition time; calculate the average of the absolute values ​​of the differences between the human thermal radiation intensity values ​​of each grid cell at the current acquisition time and the human thermal radiation intensity values ​​of the corresponding grid cells at the previous acquisition time to obtain the thermal radiation change judgment value; compare the human occupancy markers and human thermal radiation intensity values ​​of the same grid cells in the two zone human thermal state grid matrices one by one. If the human occupancy markers of the same grid cells change, or if the absolute difference in the human thermal radiation intensity values ​​of the same grid cells is greater than the thermal radiation change judgment value, then the corresponding grid cell is determined as a state change grid cell. The system counts the number of state change grid cells and calculates the ratio of this number to the number of grid cells in the human thermal state grid matrix of the m-th air conditioning control zone, thus obtaining the grid state change amount for the m-th air conditioning control zone. Based on the grid state change amount, the system determines the personnel activity change state of the m-th air conditioning control zone. The system calculates the average of the grid state change amounts for each air conditioning control zone to obtain the activity change judgment threshold. When the grid state change amount is 0, the personnel activity change state is static; when the grid state change amount is greater than 0 and less than the activity change judgment threshold, the personnel activity change state is low activity; and when the grid state change amount is greater than or equal to the activity change judgment threshold, the personnel activity change state is high activity. Initialize the partition state matrix, and write the human presence status, human density status, human thermal radiation intensity status, and human activity change status of each air conditioning control zone into the partition state matrix according to the air conditioning control zone number. By associating the human thermal state grid matrix of each air conditioning control zone with the zone state matrix, an indoor human thermal state distribution map is obtained.

[0026] In this embodiment, step four specifically includes: The thermal load characteristics of personnel are calculated based on the product of human body density state and human body thermal radiation intensity state. The environmental deviation characteristics are calculated based on the absolute value of the difference between the indoor air temperature and the target set temperature. The heat transfer characteristics of the radiant surface are calculated based on the absolute value of the difference between the indoor air temperature and the radiant surface temperature. The state of human presence is used as a partitioning feature; The historical operating energy consumption characteristics are calculated by multiplying the absolute value of the difference between the supply water temperature and the return water temperature, the supply water flow rate, and the opening degree of the radiant terminal valve. By sequentially concatenating the characteristics of personnel heat load, environmental deviation, radiant surface heat transfer, zoning occupancy, and historical operating energy consumption, a radiant air conditioning demand characterization vector for the corresponding air conditioning control zone is obtained.

[0027] In this embodiment, the thermally constrained partitioning MPC algorithm is specifically as follows: Based on the radiant air conditioning demand representation vector, the initial state of thermal constraint prediction for each air conditioning control zone in the current control cycle is constructed. Set the prediction step size and construct the zone control sequence for each air conditioning control zone within the prediction step size. The zone control sequence includes the target water supply temperature sequence, the target water supply flow rate sequence, the target valve opening sequence, and the target operating mode sequence. Using the zone control sequence, the radiant air conditioning regulation response sequence of each air conditioning control zone within the prediction step is calculated. Specifically, for each air conditioning control zone, the target water supply temperature, target water supply flow rate, target valve opening, and target operating mode are read sequentially from the zone control sequence within the prediction step. The operating mode symbol is determined according to the target operating mode, where the operating mode symbol corresponding to the cooling operating mode is −1, the operating mode symbol corresponding to the heating operating mode is 1, and the operating mode symbol corresponding to the heat preservation operating mode is 0. The absolute value of the difference between the target water supply temperature and the indoor air temperature is calculated to obtain the water supply temperature difference response value. The water supply temperature difference response value, the target water supply flow rate, and the target valve opening are multiplied to obtain the basic regulation response value. The basic regulation response value is multiplied by the operating mode symbol to obtain the radiant air conditioning regulation response quantity for the corresponding prediction step. The radiant air conditioning regulation response quantities within the prediction step are arranged in order of the prediction step to obtain the radiant air conditioning regulation response sequence for the corresponding air conditioning control zone. Based on the initial state of thermal constraint prediction and the sequence of radiant air conditioning regulation response, the predicted zone heat demand state sequence of each air conditioning control zone within the prediction step is predicted by a linear recursive prediction method. Specifically, the initial state of thermal constraint prediction for each air conditioning control zone in the current control cycle is taken as the predicted zone heat demand state of the 0th prediction step, and the radiant air conditioning regulation response is read sequentially according to the prediction step order. Within each prediction step, the predicted zone heat demand state of the previous prediction step is multiplied by the thermal inertia retention coefficient to obtain the thermal inertia continuation state, and the radiant air conditioning regulation response of the current prediction step is multiplied by the regulation response coefficient to obtain the air conditioning regulation reduction state. The difference between the thermal inertia continuation state and the air conditioning regulation reduction state is taken as the predicted zone heat demand state of the current prediction step. The above linear recursive process is repeated until the state prediction of all prediction steps is completed. The predicted zone heat demand states of each predicted zone are combined into the predicted zone heat demand state sequence of the corresponding air conditioning control zone according to the prediction step order. Based on the difference between the predicted zone heat demand state sequence and the target heat demand state, the thermal comfort deviation term of each air conditioning control zone is calculated, where the target heat demand state is used to characterize the heat demand state when the air conditioning control zone meets the thermal comfort requirements. Based on the target supply water temperature sequence, return water temperature, target supply water flow rate sequence, and target valve opening sequence, the operating energy consumption constraint term for each air conditioning control zone is calculated. Specifically, for each air conditioning control zone, within the prediction step, the target supply water temperature in the target supply water temperature sequence, the target supply water flow rate in the target supply water flow rate sequence, and the target valve opening in the target valve opening sequence are read sequentially, and the return water temperature of the corresponding air conditioning control zone is read; the absolute value of the difference between the target supply water temperature and the return water temperature within each prediction step is calculated to obtain the predicted supply and return water temperature difference; the predicted supply and return water temperature difference, the target supply water flow rate, and the target valve opening are multiplied to obtain the predicted operating energy consumption characterization value for the corresponding prediction step; the predicted operating energy consumption characterization values ​​within the prediction step are accumulated to obtain the operating energy consumption constraint term for the corresponding air conditioning control zone. Based on the changes between adjacent prediction steps in the target water supply temperature sequence, the changes between adjacent prediction steps in the target water supply flow sequence, and the changes between adjacent prediction steps in the target valve opening sequence, the control smoothing constraint terms for each air conditioning control zone are calculated. By combining thermal comfort deviation, operating energy consumption constraint, and control smoothness constraint, a thermal constraint partitioning MPC objective function is constructed. Under the constraints of the water supply temperature range, water supply flow range, radiant terminal valve opening range, and radiant air conditioning operation mode, the MPC objective function of the thermal constraint zone is solved by rolling optimization to obtain the optimal control sequence for each air conditioning control zone. Extract the target water supply temperature, target water supply flow rate, target valve opening degree, and target operating mode corresponding to the current control cycle from the optimal control sequence; Based on personnel heat load characteristics, environmental deviation characteristics, and zone occupancy characteristics, the target adjustment priority of each air conditioning control zone is calculated. Specifically, for each air conditioning control zone, the corresponding personnel heat load characteristics, environmental deviation characteristics, and zone occupancy characteristics are read. If the zone occupancy characteristic is 0, the target adjustment priority of the corresponding air conditioning control zone is determined as the lowest priority. If the zone occupancy characteristic is 1, the personnel heat load characteristics and environmental deviation characteristics are summed to obtain the zone adjustment demand value. The zone adjustment demand values ​​of each air conditioning control zone are sorted from largest to smallest, and the target adjustment priority of each air conditioning control zone is determined according to the sorting result. Among them, the air conditioning control zone with a larger zone adjustment demand value has a higher target adjustment priority. The target water supply temperature, target water supply flow rate, target valve opening, target operating mode, and target adjustment priority of each air conditioning control zone are combined to form the radiant air conditioning zone control quantities.

[0028] In this embodiment, step six specifically includes: Calculate the average human body density status of each air conditioning control zone to obtain the human body density judgment threshold. Calculate the average value of the human body thermal radiation intensity state in each air conditioning control zone to obtain the thermal radiation intensity judgment threshold. If the presence of a human body is 0, the corresponding air conditioning control zone will be designated as an unmanned zone, and the unmanned zone will be controlled to enter the heat preservation operation state according to the target operation mode. If the human body exists in state 1, and the human body density is less than the human body density judgment threshold, and the human body heat radiation intensity is less than the heat radiation intensity judgment threshold, then the corresponding air conditioning control zone will be determined as a low heat load zone, and the low heat load zone will be controlled to operate according to the target water supply flow, target valve opening and target operating mode. If the human body existence state is 1, and the human body density state is greater than or equal to the human body density judgment threshold, or the human body thermal radiation intensity state is greater than or equal to the thermal radiation intensity judgment threshold, then the corresponding air conditioning control zone will be determined as a high heat load zone, and the target water supply temperature, target water supply flow rate, target valve opening and target operating mode will be executed in the high heat load zone according to the target adjustment priority. For high heat load zones, calculate the average value of personnel activity changes in each air conditioning control zone to obtain activity change judgment values; If the change in personnel activity in a high heat load zone exceeds the activity change judgment value, then increase the target water supply flow and target valve opening for the corresponding high heat load zone. If the personnel activity change status of the high heat load zone is less than or equal to the activity change judgment value, then the target water supply flow and target valve opening of the corresponding high heat load zone shall be maintained.

[0029] Example 1: To verify the feasibility of this invention in practice, the method of this invention was applied to an energy-saving control scenario of a radiant air conditioning system in an office building. The office building has six air conditioning control zones, including window-side office areas, inner office areas, meeting areas, corridor areas, reception areas, and rest areas. Each air conditioning control zone is equipped with radiant terminals, supply and return water branch valves, and zone temperature and humidity acquisition units. On weekday mornings, the office building frequently experiences situations such as rapid entry of people, short-term gatherings in meeting areas, significant impact of solar radiation on window-side areas, and frequent short-term passage of people in corridor areas. Traditional control methods mainly adjust based on indoor air temperature and the opening of radiant terminal valves, which easily leads to problems such as continuous cooling in unoccupied areas, delayed response when meeting areas are crowded, and frequent adjustments triggered by short-term passage of people. This results in high energy consumption of the radiant air conditioning system and unstable zone thermal comfort.

[0030] In the above implementation scenario, infrared cameras for each air conditioning control zone are installed on the ceiling or in the corners of the room to continuously collect infrared images of the human body in the target indoor space. Simultaneously, they collect indoor air temperature, relative humidity, radiant surface temperature, indoor carbon dioxide concentration, indoor area location markers, as well as supply water temperature, return water temperature, supply water flow rate, radiant terminal valve opening, radiant air conditioning operating mode, and air conditioning control zone number. The collected infrared images of the human body are input into an improved YOLOv8-Seg infrared human body recognition network. A temperature difference hierarchical residual convolution unit enhances the representation of local temperature difference features in the infrared images of the human body. A zone location embedding fusion unit associates the location of the human target with the air conditioning control zone number. The human thermal state instance output header outputs the number of human targets, their location, their contour area, their thermal radiation intensity value, the air conditioning control zone number, and the control validity indicator. Human targets that are briefly passing through the corridor or whose obstruction is incomplete are excluded based on the control validity indicator to avoid ineffective regulation of the radiant air conditioning system.

[0031] By employing region mapping and thermal state rasterization methods, human targets participating in radiant air conditioning zone control are mapped to their corresponding air conditioning control zones, and a zone-specific human thermal state raster matrix is ​​constructed. Each zone raster matrix records human occupancy identifiers, human thermal radiation intensity values, and raster changes, thereby obtaining the human presence status, human density status, human thermal radiation intensity status, and human activity change status. Combining indoor environmental parameters and radiant air conditioning system operating parameters, human heat load characteristics, environmental deviation characteristics, radiant surface heat transfer characteristics, zone occupancy characteristics, and historical operating energy consumption characteristics are calculated and concatenated to form a radiant air conditioning demand characterization vector. Furthermore, based on the radiant air conditioning demand characterization vector, an initial state for thermal constraint prediction is constructed. Within the prediction step, a zone control sequence is generated, and rolling optimization is performed using thermal comfort deviation terms, operating energy consumption constraint terms, and control smoothness constraint terms. Finally, the target water supply temperature, target water supply flow rate, target valve opening, target operating mode, and target adjustment priority for each air conditioning control zone are output. During the execution phase, based on the indoor human thermal state distribution map, zones are divided into unoccupied zones, low heat load zones, and high heat load zones, and differentiated energy-saving control is implemented.

[0032] To further verify the practical effectiveness of this invention, a comparative experiment was conducted with the method of this invention, a traditional temperature threshold control scheme, and a standard YOLOv8-Seg network control scheme. The standard YOLOv8-Seg network control scheme involves using a standard YOLOv8-Seg network to identify the location and contour region of the human target, and then combining this with zoning rules for radiant air conditioning control. The comparison results are shown in Table 1.

[0033] Table 1. Comparison of the implementation effects of different radiant air conditioning control schemes

[0034] As shown in Table 1, the method of this invention outperforms the comparative schemes in all comparative indicators. The human target recognition accuracy of the method of this invention reaches 96.7%, which is 12.1 percentage points higher than the traditional temperature threshold control scheme and 8.5 percentage points higher than the standard YOLOv8-Seg network control scheme. This indicates that the temperature difference hierarchical residual convolutional unit can enhance the representation ability of human thermal targets in infrared weak texture and low temperature difference scenarios. The air conditioning control zone matching accuracy of the method of this invention reaches 95.8%, significantly higher than the two comparative schemes, indicating that the zone position embedding fusion unit can improve the correspondence accuracy between human targets and air conditioning control zone numbers.

[0035] Furthermore, in terms of energy saving, the method of this invention reduces the ineffective cooling time of unattended zones to 0.6 h / day, and the average daily radiant air conditioning energy consumption to 221.3 kWh, achieving a comprehensive energy saving rate of 22.7%. This demonstrates that the thermal constraint zoning MPC algorithm can reduce ineffective cooling and heating in unattended zones and low heat load zones. Simultaneously, the method of this invention reduces the average response time of high heat load zones to 8.7 min, increases the indoor thermal comfort compliance rate to 95.2%, and significantly reduces the frequency of valve opening fluctuations and the average daily fluctuation range of water supply flow. This indicates that the invention can maintain the operational stability and indoor thermal comfort of the radiant air conditioning system while reducing energy consumption.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An energy-saving control method for a radiant air conditioning system based on human infrared image recognition, characterized in that, include: Step 1: Collect infrared images of the human body, indoor environmental parameters, and operating parameters of the radiant air conditioning system in the target indoor space; Step 2: Input the human infrared image into the improved YOLOv8-Seg infrared human body recognition network to perform human thermal target segmentation, human region localization, and human thermal state instance recognition to obtain the human infrared recognition results for each human target; the improved YOLOv8-Seg infrared human body recognition network introduces temperature difference hierarchical residual convolution units, partitioned position embedding fusion units, and human thermal state instance output heads. Step 3: Based on the results of human infrared recognition, construct an indoor thermal distribution map of people using region mapping and thermal state rasterization methods; Step 4: Based on the indoor thermal state distribution map of people, indoor environmental parameters, and radiant air conditioning system operating parameters, calculate the radiant air conditioning demand representation vector for each air conditioning control zone; Step 5: Based on the radiant air conditioning demand representation vector, energy-saving control decisions are made for each air conditioning control zone using the thermal constraint partitioning MPC algorithm to generate radiant air conditioning zone control quantities; Step Six: Based on the radiant air conditioning zoning control values ​​and the indoor thermal distribution map of occupants, implement differentiated energy-saving control for the radiant air conditioning system.

2. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, The indoor environmental parameters include indoor air temperature, indoor relative humidity, radiant surface temperature, indoor carbon dioxide concentration, and indoor area location identification. The radiant air conditioning system operating parameters include supply water temperature, return water temperature, supply water flow rate, radiant terminal valve opening, radiant air conditioning operating mode, and air conditioning control zone number.

3. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, The improved YOLOv8-Seg infrared human body recognition network includes a basic convolutional module, a backbone feature extraction module, a Neck multi-scale feature fusion module, and a Head multi-task output module; The Backbone feature extraction module includes a temperature difference hierarchical residual convolution unit, a C2f unit, and an SPPF unit; the Neck multi-scale feature fusion module includes a standard multi-scale feature fusion path and a partition position embedding fusion unit; the Head multi-task output module includes a detection branch, a segmentation branch, and a human thermal state instance output head.

4. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, Step two specifically includes: A shallow infrared feature map is obtained by performing a 3×3 convolution mapping on the human infrared image through the basic convolution module. The shallow infrared feature map is input into the temperature difference hierarchical residual convolution unit to generate the temperature difference hierarchical enhanced feature map. The temperature difference hierarchical enhancement feature map is input into the C2f unit, and channel splitting, bottleneck convolution, cross-layer feature splicing and channel fusion are performed to obtain the infrared human body intermediate feature map. The infrared human body intermediate feature map is input into the SPPF unit, and multi-level max pooling, pooling feature stitching and convolution fusion are performed to obtain multi-scale infrared human body feature maps. The multi-scale infrared human body feature map is input into the standard multi-scale feature fusion path, and multi-scale fused feature map is obtained through upsampling, downsampling, channel stitching and C2f fusion processing. The multi-scale fused feature map is input into the partition location embedding fusion unit. For the multi-scale fused feature map at each scale, the horizontal and vertical pixel coordinates of each pixel position are normalized by the feature map height and feature map width respectively to obtain the position coordinate matrix. The location coordinate matrix is ​​convolved with a 1×1 convolution to generate a partition location embedding feature map; Multi-scale fused feature maps at various scales and partition location embedded feature maps are channel-separated and convolutionally fused to obtain partition-aware fused feature maps. The partition-aware fusion feature maps at various scales are fused to generate the Head input fusion feature map; The Head input fused feature map is output as a human target detection result through the detection branch. The human target detection result includes human target confidence and human target location. The human target location includes the coordinates of the human target center point, the width of the human target detection box, and the height of the human target detection box. Based on the confidence level of human targets, determine the number of human targets in the target indoor space; The Head input is fused into the feature map input to the segmentation branch, and R prototype masks are generated through 3×3 convolution, batch normalization, SiLU activation function, upsampling and 1×1 convolution; Based on the coordinates of the center point of the human target, the target position feature vector is extracted from the fused feature map of the Head input and linear prediction is performed to obtain the mask coefficient vector. Based on the mask coefficient vector, the prototype masks are linearly combined, and the human instance mask is generated by the Sigmoid function. The human body instance mask is binarized, and the pixel positions with a binarized pixel value of 1 are determined as the human body target contour region corresponding to the human body target. The Head input is fused with the feature map and the human instance mask, and then mask pooling is performed to obtain the human instance feature vector. The feature vector of the human body instance is input into the output header of the human body thermal state instance. Through thermal intensity regression, zoning classification and control effectiveness determination, the human body thermal radiation intensity value, air conditioning control zoning number and control effectiveness identifier of each human body target are obtained.

5. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 4, characterized in that, The step of inputting the shallow infrared feature map into the temperature difference hierarchical residual convolution unit to generate the temperature difference hierarchical enhanced feature map specifically includes: The temperature difference hierarchical residual convolutional unit includes a basic temperature difference response branch, a hierarchical residual correction branch, an original structure recharge branch, and a hierarchical fusion layer; The shallow infrared feature map is subjected to convolution operations with kernel sizes of 3×3 and 5×5 respectively by the basic temperature difference response branch, and the kernels are spliced ​​in the channel dimension. Channel fusion is performed by 1×1 convolution to obtain the basic temperature difference response feature map. The shallow infrared feature map is aligned by channel through 1×1 convolution to obtain the shallow aligned feature map. The first-level structural residual between the basic temperature difference response feature map and the shallow alignment feature map is calculated by using a hierarchical residual correction branch, and convolution correction is performed using 1×1 convolution to obtain the first-level residual correction feature map. The basic temperature difference response feature map and the first-level residual correction feature map are subjected to residual compensation to obtain the first hierarchical temperature difference feature map. Calculate the second-level structural residual between the first hierarchical temperature difference feature map and the shallow alignment feature map, and perform convolution correction on the second-level structural residual through 1×1 convolution to obtain the second-level residual correction feature map; The first hierarchical temperature difference feature map and the second-level residual correction feature map are subjected to residual compensation to obtain the second hierarchical temperature difference feature map. The shallow alignment feature map and the second hierarchical temperature difference feature map are spliced ​​together by the original structure re-injection branch, and the overall human body structure information is re-injected through 1×1 convolution to obtain the structure re-injection feature map. By performing residual fusion of the second-order temperature difference feature map and the structural recharge feature map through a hierarchical fusion layer, a temperature difference hierarchical enhancement feature map is obtained.

6. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, Step three specifically includes: Human targets with a control effectiveness flag of 1 are selected to obtain the set of human targets participating in radiation air conditioning zoning control; Based on the air conditioning control zone number, the human targets participating in the radiation air conditioning zone control are mapped to the corresponding air conditioning control zone to obtain the zone human target set of each air conditioning control zone. For each air conditioning control zone, initialize the zone grid matrix, and map the human target position, human target contour region and human thermal radiation intensity value in the zone human target set to the zone grid matrix to obtain the zone human thermal state grid matrix. The partitioned human thermal state grid matrix includes several grid cells; If a grid cell overlaps with the human target contour area, the human occupancy flag of the current grid cell is set to 1; otherwise, it is set to 0. If there is a grid cell in the air conditioning control zone with a human occupancy identifier of 1, then the human presence status of the corresponding air conditioning control zone is determined to be 1; otherwise, it is determined to be 0. The number of human targets in the human target set of each air conditioning control zone is counted, and the ratio of the number of human targets to the number of grid cells in the human thermal state grid matrix of the zone is calculated to obtain the human density state of each air conditioning control zone. Based on the human body thermal radiation intensity value and the number of grid cells covered by the human body target outline area in each air conditioning control zone, the human body thermal radiation intensity status of each air conditioning control zone is calculated. Based on the human thermal state grid matrix of the current and previous data collection times, the grid state change of each air conditioning control zone is calculated, and the change of human activity in each air conditioning control zone is determined according to the grid state change. Initialize the partition state matrix, and write the human presence status, human density status, human thermal radiation intensity status, and human activity change status of each air conditioning control zone into the partition state matrix according to the air conditioning control zone number. By associating the human thermal state grid matrix of each air conditioning control zone with the zone state matrix, an indoor human thermal state distribution map is obtained.

7. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, Step four specifically includes: The thermal load characteristics of personnel are calculated based on the product of human body density state and human body thermal radiation intensity state. The environmental deviation characteristics are calculated based on the absolute value of the difference between the indoor air temperature and the target set temperature. The heat transfer characteristics of the radiant surface are calculated based on the absolute value of the difference between the indoor air temperature and the radiant surface temperature. The state of human presence is used as a partitioning feature; The historical operating energy consumption characteristics are calculated by multiplying the absolute value of the difference between the supply water temperature and the return water temperature, the supply water flow rate, and the opening degree of the radiant terminal valve. By sequentially concatenating the characteristics of personnel heat load, environmental deviation, radiant surface heat transfer, zoning occupancy, and historical operating energy consumption, a radiant air conditioning demand characterization vector for the corresponding air conditioning control zone is obtained.

8. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, The thermally constrained partitioning MPC algorithm is specifically as follows: Based on the radiant air conditioning demand representation vector, the initial state of thermal constraint prediction for each air conditioning control zone in the current control cycle is constructed. Set the prediction step size and construct the zone control sequence for each air conditioning control zone within the prediction step size. The zone control sequence includes the target water supply temperature sequence, the target water supply flow rate sequence, the target valve opening sequence, and the target operating mode sequence. Using the zoned control sequence, the radiative air conditioning regulation response sequence of each air conditioning control zone within the prediction step size is calculated; Based on the initial state of thermal constraint prediction and the radiant air conditioning regulation response sequence, the thermal demand state sequence of each air conditioning control zone within the prediction step is predicted by a linear recursive prediction method. The thermal comfort deviation term for each air conditioning control zone is calculated based on the difference between the predicted zone heat demand state sequence and the target heat demand state. Based on the target supply water temperature sequence, return water temperature, target supply water flow sequence, and target valve opening sequence, calculate the operating energy consumption constraints for each air conditioning control zone; Based on the changes between adjacent prediction steps in the target water supply temperature sequence, the changes between adjacent prediction steps in the target water supply flow sequence, and the changes between adjacent prediction steps in the target valve opening sequence, the control smoothing constraint terms for each air conditioning control zone are calculated. By combining thermal comfort deviation, operating energy consumption constraint, and control smoothness constraint, a thermal constraint partitioning MPC objective function is constructed. Under the constraints of the water supply temperature range, water supply flow range, radiant terminal valve opening range, and radiant air conditioning operation mode, the MPC objective function of the thermal constraint zone is solved by rolling optimization to obtain the optimal control sequence for each air conditioning control zone. Extract the target water supply temperature, target water supply flow rate, target valve opening degree, and target operating mode corresponding to the current control cycle from the optimal control sequence; Based on the characteristics of personnel heat load, environmental deviation, and zoning occupancy, the target adjustment priority of each air conditioning control zone is calculated. The target water supply temperature, target water supply flow rate, target valve opening, target operating mode, and target adjustment priority of each air conditioning control zone are combined to form the radiant air conditioning zone control quantities.

9. The energy-saving control method for a radiant air conditioning system based on human infrared image recognition according to claim 1, characterized in that, Step six specifically includes: Calculate the average human density status of each air conditioning control zone to obtain the human density judgment threshold. Calculate the average value of the human body thermal radiation intensity state in each air-conditioning control zone to obtain the thermal radiation intensity judgment threshold. If the presence of a human body is 0, the corresponding air conditioning control zone will be designated as an unmanned zone, and the unmanned zone will be controlled to enter the heat preservation operation state according to the target operation mode. If the human body exists in state 1, and the human body density is less than the human body density judgment threshold, and the human body heat radiation intensity is less than the heat radiation intensity judgment threshold, then the corresponding air conditioning control zone will be determined as a low heat load zone, and the low heat load zone will be controlled to operate according to the target water supply flow, target valve opening and target operating mode. If the human body existence state is 1, and the human body density state is greater than or equal to the human body density judgment threshold, or the human body thermal radiation intensity state is greater than or equal to the thermal radiation intensity judgment threshold, then the corresponding air conditioning control zone will be determined as a high heat load zone, and the target water supply temperature, target water supply flow rate, target valve opening and target operating mode will be executed in the high heat load zone according to the target adjustment priority. For high heat load zones, calculate the average value of personnel activity changes in each air conditioning control zone to obtain activity change judgment values; If the change in personnel activity in a high heat load zone exceeds the activity change judgment value, then increase the target water supply flow and target valve opening for the corresponding high heat load zone. If the personnel activity change status of the high heat load zone is less than or equal to the activity change judgment value, then the target water supply flow and target valve opening of the corresponding high heat load zone shall be maintained.

Images